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/ADT/SmallBitVector.h"
15#include "llvm/Analysis/CmpInstAnalysis.h"
16#include "llvm/Analysis/FloatingPointPredicateUtils.h"
17#include "llvm/Analysis/InstructionSimplify.h"
18#include "llvm/IR/ConstantRange.h"
19#include "llvm/IR/DerivedTypes.h"
20#include "llvm/IR/Instructions.h"
21#include "llvm/IR/Intrinsics.h"
22#include "llvm/IR/PatternMatch.h"
23#include "llvm/IR/ProfDataUtils.h"
24#include "llvm/Transforms/InstCombine/InstCombiner.h"
25#include "llvm/Transforms/Utils/Local.h"
26
27using namespace llvm;
28using namespace PatternMatch;
29
30#define DEBUG_TYPE "instcombine"
31
32namespace llvm {
33extern cl::opt<bool> ProfcheckDisableMetadataFixes;
34}
35
36/// This is the complement of getICmpCode, which turns an opcode and two
37/// operands into either a constant true or false, or a brand new ICmp
38/// instruction. The sign is passed in to determine which kind of predicate to
39/// use in the new icmp instruction.
40static Value *getNewICmpValue(unsigned Code, bool Sign, Value *LHS, Value *RHS,
41 InstCombiner::BuilderTy &Builder) {
42 ICmpInst::Predicate NewPred;
43 if (Constant *TorF = getPredForICmpCode(Code, Sign, OpTy: LHS->getType(), Pred&: NewPred))
44 return TorF;
45 return Builder.CreateICmp(P: NewPred, LHS, RHS);
46}
47
48/// This is the complement of getFCmpCode, which turns an opcode and two
49/// operands into either a FCmp instruction, or a true/false constant.
50static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS,
51 InstCombiner::BuilderTy &Builder, FMFSource FMF) {
52 FCmpInst::Predicate NewPred;
53 if (Constant *TorF = getPredForFCmpCode(Code, OpTy: LHS->getType(), Pred&: NewPred))
54 return TorF;
55 return Builder.CreateFCmpFMF(P: NewPred, LHS, RHS, FMFSource: FMF);
56}
57
58/// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
59/// (V < Lo || V >= Hi). This method expects that Lo < Hi. IsSigned indicates
60/// whether to treat V, Lo, and Hi as signed or not.
61Value *InstCombinerImpl::insertRangeTest(Value *V, const APInt &Lo,
62 const APInt &Hi, bool isSigned,
63 bool Inside) {
64 assert((isSigned ? Lo.slt(Hi) : Lo.ult(Hi)) &&
65 "Lo is not < Hi in range emission code!");
66
67 Type *Ty = V->getType();
68
69 // V >= Min && V < Hi --> V < Hi
70 // V < Min || V >= Hi --> V >= Hi
71 ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
72 if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) {
73 Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred;
74 return Builder.CreateICmp(P: Pred, LHS: V, RHS: ConstantInt::get(Ty, V: Hi));
75 }
76
77 // V >= Lo && V < Hi --> V - Lo u< Hi - Lo
78 // V < Lo || V >= Hi --> V - Lo u>= Hi - Lo
79 Value *VMinusLo =
80 Builder.CreateSub(LHS: V, RHS: ConstantInt::get(Ty, V: Lo), Name: V->getName() + ".off");
81 Constant *HiMinusLo = ConstantInt::get(Ty, V: Hi - Lo);
82 return Builder.CreateICmp(P: Pred, LHS: VMinusLo, RHS: HiMinusLo);
83}
84
85/// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns
86/// that can be simplified.
87/// One of A and B is considered the mask. The other is the value. This is
88/// described as the "AMask" or "BMask" part of the enum. If the enum contains
89/// only "Mask", then both A and B can be considered masks. If A is the mask,
90/// then it was proven that (A & C) == C. This is trivial if C == A or C == 0.
91/// If both A and C are constants, this proof is also easy.
92/// For the following explanations, we assume that A is the mask.
93///
94/// "AllOnes" declares that the comparison is true only if (A & B) == A or all
95/// bits of A are set in B.
96/// Example: (icmp eq (A & 3), 3) -> AMask_AllOnes
97///
98/// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all
99/// bits of A are cleared in B.
100/// Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes
101///
102/// "Mixed" declares that (A & B) == C and C might or might not contain any
103/// number of one bits and zero bits.
104/// Example: (icmp eq (A & 3), 1) -> AMask_Mixed
105///
106/// "Not" means that in above descriptions "==" should be replaced by "!=".
107/// Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes
108///
109/// If the mask A contains a single bit, then the following is equivalent:
110/// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
111/// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
112enum MaskedICmpType {
113 AMask_AllOnes = 1,
114 AMask_NotAllOnes = 2,
115 BMask_AllOnes = 4,
116 BMask_NotAllOnes = 8,
117 Mask_AllZeros = 16,
118 Mask_NotAllZeros = 32,
119 AMask_Mixed = 64,
120 AMask_NotMixed = 128,
121 BMask_Mixed = 256,
122 BMask_NotMixed = 512
123};
124
125/// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C)
126/// satisfies.
127static unsigned getMaskedICmpType(Value *A, Value *B, Value *C,
128 ICmpInst::Predicate Pred) {
129 const APInt *ConstA = nullptr, *ConstB = nullptr, *ConstC = nullptr;
130 match(V: A, P: m_APInt(Res&: ConstA));
131 match(V: B, P: m_APInt(Res&: ConstB));
132 match(V: C, P: m_APInt(Res&: ConstC));
133 bool IsEq = (Pred == ICmpInst::ICMP_EQ);
134 bool IsAPow2 = ConstA && ConstA->isPowerOf2();
135 bool IsBPow2 = ConstB && ConstB->isPowerOf2();
136 unsigned MaskVal = 0;
137 if (ConstC && ConstC->isZero()) {
138 // if C is zero, then both A and B qualify as mask
139 MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed)
140 : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed));
141 if (IsAPow2)
142 MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed)
143 : (AMask_AllOnes | AMask_Mixed));
144 if (IsBPow2)
145 MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed)
146 : (BMask_AllOnes | BMask_Mixed));
147 return MaskVal;
148 }
149
150 if (A == C) {
151 MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed)
152 : (AMask_NotAllOnes | AMask_NotMixed));
153 if (IsAPow2)
154 MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed)
155 : (Mask_AllZeros | AMask_Mixed));
156 } else if (ConstA && ConstC && ConstC->isSubsetOf(RHS: *ConstA)) {
157 MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed);
158 }
159
160 if (B == C) {
161 MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed)
162 : (BMask_NotAllOnes | BMask_NotMixed));
163 if (IsBPow2)
164 MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed)
165 : (Mask_AllZeros | BMask_Mixed));
166 } else if (ConstB && ConstC && ConstC->isSubsetOf(RHS: *ConstB)) {
167 MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed);
168 }
169
170 return MaskVal;
171}
172
173/// Convert an analysis of a masked ICmp into its equivalent if all boolean
174/// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
175/// is adjacent to the corresponding normal flag (recording ==), this just
176/// involves swapping those bits over.
177static unsigned conjugateICmpMask(unsigned Mask) {
178 unsigned NewMask;
179 NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros |
180 AMask_Mixed | BMask_Mixed))
181 << 1;
182
183 NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros |
184 AMask_NotMixed | BMask_NotMixed))
185 >> 1;
186
187 return NewMask;
188}
189
190// Adapts the external decomposeBitTest for local use.
191static bool decomposeBitTest(Value *Cond, CmpInst::Predicate &Pred, Value *&X,
192 Value *&Y, Value *&Z) {
193 auto Res =
194 llvm::decomposeBitTest(Cond, /*LookThroughTrunc=*/true,
195 /*AllowNonZeroC=*/true, /*DecomposeAnd=*/true);
196 if (!Res)
197 return false;
198
199 Pred = Res->Pred;
200 X = Res->X;
201 Y = ConstantInt::get(Ty: X->getType(), V: Res->Mask);
202 Z = ConstantInt::get(Ty: X->getType(), V: Res->C);
203 return true;
204}
205
206/// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E).
207/// Return the pattern classes (from MaskedICmpType) for the left hand side and
208/// the right hand side as a pair.
209/// LHS and RHS are the left hand side and the right hand side ICmps and PredL
210/// and PredR are their predicates, respectively.
211static std::optional<std::pair<unsigned, unsigned>>
212getMaskedTypeForICmpPair(Value *&A, Value *&B, Value *&C, Value *&D, Value *&E,
213 Value *LHS, Value *RHS, ICmpInst::Predicate &PredL,
214 ICmpInst::Predicate &PredR) {
215
216 // Here comes the tricky part:
217 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
218 // and L11 & L12 == L21 & L22. The same goes for RHS.
219 // Now we must find those components L** and R**, that are equal, so
220 // that we can extract the parameters A, B, C, D, and E for the canonical
221 // above.
222
223 // Check whether the icmp can be decomposed into a bit test.
224 Value *L1, *L11, *L12, *L2, *L21, *L22;
225 if (decomposeBitTest(Cond: LHS, Pred&: PredL, X&: L11, Y&: L12, Z&: L2)) {
226 L21 = L22 = L1 = nullptr;
227 } else {
228 auto *LHSCMP = dyn_cast<ICmpInst>(Val: LHS);
229 if (!LHSCMP)
230 return std::nullopt;
231
232 // Don't allow pointers. Splat vectors are fine.
233 if (!LHSCMP->getOperand(i_nocapture: 0)->getType()->isIntOrIntVectorTy())
234 return std::nullopt;
235
236 PredL = LHSCMP->getPredicate();
237 L1 = LHSCMP->getOperand(i_nocapture: 0);
238 L2 = LHSCMP->getOperand(i_nocapture: 1);
239 // Look for ANDs in the LHS icmp.
240 if (!match(V: L1, P: m_And(L: m_Value(V&: L11), R: m_Value(V&: L12)))) {
241 // Any icmp can be viewed as being trivially masked; if it allows us to
242 // remove one, it's worth it.
243 L11 = L1;
244 L12 = Constant::getAllOnesValue(Ty: L1->getType());
245 }
246
247 if (!match(V: L2, P: m_And(L: m_Value(V&: L21), R: m_Value(V&: L22)))) {
248 L21 = L2;
249 L22 = Constant::getAllOnesValue(Ty: L2->getType());
250 }
251 }
252
253 // Bail if LHS was a icmp that can't be decomposed into an equality.
254 if (!ICmpInst::isEquality(P: PredL))
255 return std::nullopt;
256
257 Value *R11, *R12, *R2;
258 if (decomposeBitTest(Cond: RHS, Pred&: PredR, X&: R11, Y&: R12, Z&: R2)) {
259 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
260 A = R11;
261 D = R12;
262 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
263 A = R12;
264 D = R11;
265 } else {
266 return std::nullopt;
267 }
268 E = R2;
269 } else {
270 auto *RHSCMP = dyn_cast<ICmpInst>(Val: RHS);
271 if (!RHSCMP)
272 return std::nullopt;
273 // Don't allow pointers. Splat vectors are fine.
274 if (!RHSCMP->getOperand(i_nocapture: 0)->getType()->isIntOrIntVectorTy())
275 return std::nullopt;
276
277 PredR = RHSCMP->getPredicate();
278
279 Value *R1 = RHSCMP->getOperand(i_nocapture: 0);
280 R2 = RHSCMP->getOperand(i_nocapture: 1);
281 bool Ok = false;
282 if (!match(V: R1, P: m_And(L: m_Value(V&: R11), R: m_Value(V&: R12)))) {
283 // As before, model no mask as a trivial mask if it'll let us do an
284 // optimization.
285 R11 = R1;
286 R12 = Constant::getAllOnesValue(Ty: R1->getType());
287 }
288
289 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
290 A = R11;
291 D = R12;
292 E = R2;
293 Ok = true;
294 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
295 A = R12;
296 D = R11;
297 E = R2;
298 Ok = true;
299 }
300
301 // Avoid matching against the -1 value we created for unmasked operand.
302 if (Ok && match(V: A, P: m_AllOnes()))
303 Ok = false;
304
305 // Look for ANDs on the right side of the RHS icmp.
306 if (!Ok) {
307 if (!match(V: R2, P: m_And(L: m_Value(V&: R11), R: m_Value(V&: R12)))) {
308 R11 = R2;
309 R12 = Constant::getAllOnesValue(Ty: R2->getType());
310 }
311
312 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
313 A = R11;
314 D = R12;
315 E = R1;
316 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
317 A = R12;
318 D = R11;
319 E = R1;
320 } else {
321 return std::nullopt;
322 }
323 }
324 }
325
326 // Bail if RHS was a icmp that can't be decomposed into an equality.
327 if (!ICmpInst::isEquality(P: PredR))
328 return std::nullopt;
329
330 if (L11 == A) {
331 B = L12;
332 C = L2;
333 } else if (L12 == A) {
334 B = L11;
335 C = L2;
336 } else if (L21 == A) {
337 B = L22;
338 C = L1;
339 } else if (L22 == A) {
340 B = L21;
341 C = L1;
342 }
343
344 unsigned LeftType = getMaskedICmpType(A, B, C, Pred: PredL);
345 unsigned RightType = getMaskedICmpType(A, B: D, C: E, Pred: PredR);
346 return std::optional<std::pair<unsigned, unsigned>>(
347 std::make_pair(x&: LeftType, y&: RightType));
348}
349
350/// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single
351/// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros
352/// and the right hand side is of type BMask_Mixed. For example,
353/// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8).
354/// Also used for logical and/or, must be poison safe.
355static Value *foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
356 Value *LHS, Value *RHS, bool IsAnd, Value *A, Value *B, Value *D, Value *E,
357 ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
358 InstCombiner::BuilderTy &Builder) {
359 // We are given the canonical form:
360 // (icmp ne (A & B), 0) & (icmp eq (A & D), E).
361 // where D & E == E.
362 //
363 // If IsAnd is false, we get it in negated form:
364 // (icmp eq (A & B), 0) | (icmp ne (A & D), E) ->
365 // !((icmp ne (A & B), 0) & (icmp eq (A & D), E)).
366 //
367 // We currently handle the case of B, C, D, E are constant.
368 //
369 const APInt *BCst, *DCst, *OrigECst;
370 if (!match(V: B, P: m_APInt(Res&: BCst)) || !match(V: D, P: m_APInt(Res&: DCst)) ||
371 !match(V: E, P: m_APInt(Res&: OrigECst)))
372 return nullptr;
373
374 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
375
376 // Update E to the canonical form when D is a power of two and RHS is
377 // canonicalized as,
378 // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or
379 // (icmp ne (A & D), D) -> (icmp eq (A & D), 0).
380 APInt ECst = *OrigECst;
381 if (PredR != NewCC)
382 ECst ^= *DCst;
383
384 // If B or D is zero, skip because if LHS or RHS can be trivially folded by
385 // other folding rules and this pattern won't apply any more.
386 if (*BCst == 0 || *DCst == 0)
387 return nullptr;
388
389 // If B and D don't intersect, ie. (B & D) == 0, try to fold isNaN idiom:
390 // (icmp ne (A & FractionBits), 0) & (icmp eq (A & ExpBits), ExpBits)
391 // -> isNaN(A)
392 // Otherwise, we cannot deduce anything from it.
393 if (!BCst->intersects(RHS: *DCst)) {
394 Value *Src;
395 if (*DCst == ECst && match(V: A, P: m_ElementWiseBitCast(Op: m_Value(V&: Src))) &&
396 !Builder.GetInsertBlock()->getParent()->hasFnAttribute(
397 Kind: Attribute::StrictFP)) {
398 Type *Ty = Src->getType()->getScalarType();
399 if (!Ty->isIEEELikeFPTy())
400 return nullptr;
401
402 APInt ExpBits = APFloat::getInf(Sem: Ty->getFltSemantics()).bitcastToAPInt();
403 if (ECst != ExpBits)
404 return nullptr;
405 APInt FractionBits = ~ExpBits;
406 FractionBits.clearSignBit();
407 if (*BCst != FractionBits)
408 return nullptr;
409
410 return Builder.CreateFCmp(P: IsAnd ? FCmpInst::FCMP_UNO : FCmpInst::FCMP_ORD,
411 LHS: Src, RHS: ConstantFP::getZero(Ty: Src->getType()));
412 }
413 return nullptr;
414 }
415
416 // If the following two conditions are met:
417 //
418 // 1. mask B covers only a single bit that's not covered by mask D, that is,
419 // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of
420 // B and D has only one bit set) and,
421 //
422 // 2. RHS (and E) indicates that the rest of B's bits are zero (in other
423 // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0
424 //
425 // then that single bit in B must be one and thus the whole expression can be
426 // folded to
427 // (A & (B | D)) == (B & (B ^ D)) | E.
428 //
429 // For example,
430 // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9)
431 // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8)
432 if ((((*BCst & *DCst) & ECst) == 0) &&
433 (*BCst & (*BCst ^ *DCst)).isPowerOf2()) {
434 APInt BorD = *BCst | *DCst;
435 APInt BandBxorDorE = (*BCst & (*BCst ^ *DCst)) | ECst;
436 Value *NewMask = ConstantInt::get(Ty: A->getType(), V: BorD);
437 Value *NewMaskedValue = ConstantInt::get(Ty: A->getType(), V: BandBxorDorE);
438 Value *NewAnd = Builder.CreateAnd(LHS: A, RHS: NewMask);
439 return Builder.CreateICmp(P: NewCC, LHS: NewAnd, RHS: NewMaskedValue);
440 }
441
442 auto IsSubSetOrEqual = [](const APInt *C1, const APInt *C2) {
443 return (*C1 & *C2) == *C1;
444 };
445 auto IsSuperSetOrEqual = [](const APInt *C1, const APInt *C2) {
446 return (*C1 & *C2) == *C2;
447 };
448
449 // In the following, we consider only the cases where B is a superset of D, B
450 // is a subset of D, or B == D because otherwise there's at least one bit
451 // covered by B but not D, in which case we can't deduce much from it, so
452 // no folding (aside from the single must-be-one bit case right above.)
453 // For example,
454 // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding.
455 if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst))
456 return nullptr;
457
458 // At this point, either B is a superset of D, B is a subset of D or B == D.
459
460 // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict
461 // and the whole expression becomes false (or true if negated), otherwise, no
462 // folding.
463 // For example,
464 // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false.
465 // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding.
466 if (ECst.isZero()) {
467 if (IsSubSetOrEqual(BCst, DCst))
468 return ConstantInt::get(Ty: LHS->getType(), V: !IsAnd);
469 return nullptr;
470 }
471
472 // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B ==
473 // D. If B is a superset of (or equal to) D, since E is not zero, LHS is
474 // subsumed by RHS (RHS implies LHS.) So the whole expression becomes
475 // RHS. For example,
476 // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
477 // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
478 if (IsSuperSetOrEqual(BCst, DCst)) {
479 // We can't guarantee that samesign hold after this fold.
480 if (auto *ICmp = dyn_cast<ICmpInst>(Val: RHS))
481 ICmp->setSameSign(false);
482 return RHS;
483 }
484 // Otherwise, B is a subset of D. If B and E have a common bit set,
485 // ie. (B & E) != 0, then LHS is subsumed by RHS. For example.
486 // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
487 assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code");
488 if ((*BCst & ECst) != 0) {
489 // We can't guarantee that samesign hold after this fold.
490 if (auto *ICmp = dyn_cast<ICmpInst>(Val: RHS))
491 ICmp->setSameSign(false);
492 return RHS;
493 }
494 // Otherwise, LHS and RHS contradict and the whole expression becomes false
495 // (or true if negated.) For example,
496 // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false.
497 // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false.
498 return ConstantInt::get(Ty: LHS->getType(), V: !IsAnd);
499}
500
501/// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single
502/// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side
503/// aren't of the common mask pattern type.
504/// Also used for logical and/or, must be poison safe.
505static Value *foldLogOpOfMaskedICmpsAsymmetric(
506 Value *LHS, Value *RHS, bool IsAnd, Value *A, Value *B, Value *C, Value *D,
507 Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
508 unsigned LHSMask, unsigned RHSMask, InstCombiner::BuilderTy &Builder) {
509 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
510 "Expected equality predicates for masked type of icmps.");
511 // Handle Mask_NotAllZeros-BMask_Mixed cases.
512 // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or
513 // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E)
514 // which gets swapped to
515 // (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C).
516 if (!IsAnd) {
517 LHSMask = conjugateICmpMask(Mask: LHSMask);
518 RHSMask = conjugateICmpMask(Mask: RHSMask);
519 }
520 if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) {
521 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
522 LHS, RHS, IsAnd, A, B, D, E, PredL, PredR, Builder)) {
523 return V;
524 }
525 } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) {
526 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
527 LHS: RHS, RHS: LHS, IsAnd, A, B: D, D: B, E: C, PredL: PredR, PredR: PredL, Builder)) {
528 return V;
529 }
530 }
531 return nullptr;
532}
533
534/// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
535/// into a single (icmp(A & X) ==/!= Y).
536static Value *foldLogOpOfMaskedICmps(Value *LHS, Value *RHS, bool IsAnd,
537 bool IsLogical,
538 InstCombiner::BuilderTy &Builder,
539 const SimplifyQuery &Q) {
540 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
541 ICmpInst::Predicate PredL, PredR;
542 std::optional<std::pair<unsigned, unsigned>> MaskPair =
543 getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR);
544 if (!MaskPair)
545 return nullptr;
546 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
547 "Expected equality predicates for masked type of icmps.");
548 unsigned LHSMask = MaskPair->first;
549 unsigned RHSMask = MaskPair->second;
550 unsigned Mask = LHSMask & RHSMask;
551 if (Mask == 0) {
552 // Even if the two sides don't share a common pattern, check if folding can
553 // still happen.
554 if (Value *V = foldLogOpOfMaskedICmpsAsymmetric(
555 LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask,
556 Builder))
557 return V;
558 return nullptr;
559 }
560
561 // In full generality:
562 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
563 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
564 //
565 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
566 // equivalent to (icmp (A & X) !Op Y).
567 //
568 // Therefore, we can pretend for the rest of this function that we're dealing
569 // with the conjunction, provided we flip the sense of any comparisons (both
570 // input and output).
571
572 // In most cases we're going to produce an EQ for the "&&" case.
573 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
574 if (!IsAnd) {
575 // Convert the masking analysis into its equivalent with negated
576 // comparisons.
577 Mask = conjugateICmpMask(Mask);
578 }
579
580 if (Mask & Mask_AllZeros) {
581 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
582 // -> (icmp eq (A & (B|D)), 0)
583 if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(V: D))
584 return nullptr; // TODO: Use freeze?
585 Value *NewOr = Builder.CreateOr(LHS: B, RHS: D);
586 Value *NewAnd = Builder.CreateAnd(LHS: A, RHS: NewOr);
587 // We can't use C as zero because we might actually handle
588 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
589 // with B and D, having a single bit set.
590 Value *Zero = Constant::getNullValue(Ty: A->getType());
591 return Builder.CreateICmp(P: NewCC, LHS: NewAnd, RHS: Zero);
592 }
593 if (Mask & BMask_AllOnes) {
594 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
595 // -> (icmp eq (A & (B|D)), (B|D))
596 if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(V: D))
597 return nullptr; // TODO: Use freeze?
598 Value *NewOr = Builder.CreateOr(LHS: B, RHS: D);
599 Value *NewAnd = Builder.CreateAnd(LHS: A, RHS: NewOr);
600 return Builder.CreateICmp(P: NewCC, LHS: NewAnd, RHS: NewOr);
601 }
602 if (Mask & AMask_AllOnes) {
603 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
604 // -> (icmp eq (A & (B&D)), A)
605 if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(V: D))
606 return nullptr; // TODO: Use freeze?
607 Value *NewAnd1 = Builder.CreateAnd(LHS: B, RHS: D);
608 Value *NewAnd2 = Builder.CreateAnd(LHS: A, RHS: NewAnd1);
609 return Builder.CreateICmp(P: NewCC, LHS: NewAnd2, RHS: A);
610 }
611
612 const APInt *ConstB, *ConstD;
613 if (match(V: B, P: m_APInt(Res&: ConstB)) && match(V: D, P: m_APInt(Res&: ConstD))) {
614 if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) {
615 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
616 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
617 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
618 // Only valid if one of the masks is a superset of the other (check "B&D"
619 // is the same as either B or D).
620 APInt NewMask = *ConstB & *ConstD;
621 if (NewMask == *ConstB)
622 return LHS;
623 if (NewMask == *ConstD) {
624 if (IsLogical) {
625 if (auto *RHSI = dyn_cast<Instruction>(Val: RHS))
626 RHSI->dropPoisonGeneratingFlags();
627 }
628 return RHS;
629 }
630 }
631
632 if (Mask & AMask_NotAllOnes) {
633 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
634 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
635 // Only valid if one of the masks is a superset of the other (check "B|D"
636 // is the same as either B or D).
637 APInt NewMask = *ConstB | *ConstD;
638 if (NewMask == *ConstB)
639 return LHS;
640 if (NewMask == *ConstD)
641 return RHS;
642 }
643
644 if (Mask & (BMask_Mixed | BMask_NotMixed)) {
645 // Mixed:
646 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
647 // We already know that B & C == C && D & E == E.
648 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
649 // C and E, which are shared by both the mask B and the mask D, don't
650 // contradict, then we can transform to
651 // -> (icmp eq (A & (B|D)), (C|E))
652 // Currently, we only handle the case of B, C, D, and E being constant.
653 // We can't simply use C and E because we might actually handle
654 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
655 // with B and D, having a single bit set.
656
657 // NotMixed:
658 // (icmp ne (A & B), C) & (icmp ne (A & D), E)
659 // -> (icmp ne (A & (B & D)), (C & E))
660 // Check the intersection (B & D) for inequality.
661 // Assume that (B & D) == B || (B & D) == D, i.e B/D is a subset of D/B
662 // and (B & D) & (C ^ E) == 0, bits of C and E, which are shared by both
663 // the B and the D, don't contradict. Note that we can assume (~B & C) ==
664 // 0 && (~D & E) == 0, previous operation should delete these icmps if it
665 // hadn't been met.
666
667 const APInt *OldConstC, *OldConstE;
668 if (!match(V: C, P: m_APInt(Res&: OldConstC)) || !match(V: E, P: m_APInt(Res&: OldConstE)))
669 return nullptr;
670
671 auto FoldBMixed = [&](ICmpInst::Predicate CC, bool IsNot) -> Value * {
672 CC = IsNot ? CmpInst::getInversePredicate(pred: CC) : CC;
673 const APInt ConstC = PredL != CC ? *ConstB ^ *OldConstC : *OldConstC;
674 const APInt ConstE = PredR != CC ? *ConstD ^ *OldConstE : *OldConstE;
675
676 if (((*ConstB & *ConstD) & (ConstC ^ ConstE)).getBoolValue())
677 return IsNot ? nullptr : ConstantInt::get(Ty: LHS->getType(), V: !IsAnd);
678
679 if (IsNot && !ConstB->isSubsetOf(RHS: *ConstD) &&
680 !ConstD->isSubsetOf(RHS: *ConstB))
681 return nullptr;
682
683 APInt BD, CE;
684 if (IsNot) {
685 BD = *ConstB & *ConstD;
686 CE = ConstC & ConstE;
687 } else {
688 BD = *ConstB | *ConstD;
689 CE = ConstC | ConstE;
690 }
691 Value *NewAnd = Builder.CreateAnd(LHS: A, RHS: BD);
692 Value *CEVal = ConstantInt::get(Ty: A->getType(), V: CE);
693 return Builder.CreateICmp(P: CC, LHS: NewAnd, RHS: CEVal);
694 };
695
696 if (Mask & BMask_Mixed)
697 return FoldBMixed(NewCC, false);
698 if (Mask & BMask_NotMixed) // can be else also
699 return FoldBMixed(NewCC, true);
700 }
701 }
702
703 // (icmp eq (A & B), 0) | (icmp eq (A & D), 0)
704 // -> (icmp ne (A & (B|D)), (B|D))
705 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0)
706 // -> (icmp eq (A & (B|D)), (B|D))
707 // iff B and D is known to be a power of two
708 if (Mask & Mask_NotAllZeros &&
709 isKnownToBeAPowerOfTwo(V: B, /*OrZero=*/false, Q) &&
710 isKnownToBeAPowerOfTwo(V: D, /*OrZero=*/false, Q)) {
711 // If this is a logical and/or, then we must prevent propagation of a
712 // poison value from the RHS by inserting freeze.
713 if (IsLogical)
714 D = Builder.CreateFreeze(V: D);
715 Value *Mask = Builder.CreateOr(LHS: B, RHS: D);
716 Value *Masked = Builder.CreateAnd(LHS: A, RHS: Mask);
717 return Builder.CreateICmp(P: NewCC, LHS: Masked, RHS: Mask);
718 }
719 return nullptr;
720}
721
722/// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
723/// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
724/// If \p Inverted is true then the check is for the inverted range, e.g.
725/// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
726Value *InstCombinerImpl::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
727 bool Inverted) {
728 // Check the lower range comparison, e.g. x >= 0
729 // InstCombine already ensured that if there is a constant it's on the RHS.
730 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Val: Cmp0->getOperand(i_nocapture: 1));
731 if (!RangeStart)
732 return nullptr;
733
734 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
735 Cmp0->getPredicate());
736
737 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
738 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
739 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
740 return nullptr;
741
742 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
743 Cmp1->getPredicate());
744
745 Value *Input = Cmp0->getOperand(i_nocapture: 0);
746 Value *Cmp1Op0 = Cmp1->getOperand(i_nocapture: 0);
747 Value *Cmp1Op1 = Cmp1->getOperand(i_nocapture: 1);
748 Value *RangeEnd;
749 if (match(V: Cmp1Op0, P: m_SExtOrSelf(Op: m_Specific(V: Input)))) {
750 // For the upper range compare we have: icmp x, n
751 Input = Cmp1Op0;
752 RangeEnd = Cmp1Op1;
753 } else if (match(V: Cmp1Op1, P: m_SExtOrSelf(Op: m_Specific(V: Input)))) {
754 // For the upper range compare we have: icmp n, x
755 Input = Cmp1Op1;
756 RangeEnd = Cmp1Op0;
757 Pred1 = ICmpInst::getSwappedPredicate(pred: Pred1);
758 } else {
759 return nullptr;
760 }
761
762 // Check the upper range comparison, e.g. x < n
763 ICmpInst::Predicate NewPred;
764 switch (Pred1) {
765 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
766 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
767 default: return nullptr;
768 }
769
770 // This simplification is only valid if the upper range is not negative.
771 KnownBits Known = computeKnownBits(V: RangeEnd, CxtI: Cmp1);
772 if (!Known.isNonNegative())
773 return nullptr;
774
775 if (Inverted)
776 NewPred = ICmpInst::getInversePredicate(pred: NewPred);
777
778 return Builder.CreateICmp(P: NewPred, LHS: Input, RHS: RangeEnd);
779}
780
781// (or (icmp eq X, 0), (icmp eq X, Pow2OrZero))
782// -> (icmp eq (and X, Pow2OrZero), X)
783// (and (icmp ne X, 0), (icmp ne X, Pow2OrZero))
784// -> (icmp ne (and X, Pow2OrZero), X)
785static Value *
786foldAndOrOfICmpsWithPow2AndWithZero(InstCombiner::BuilderTy &Builder,
787 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
788 const SimplifyQuery &Q) {
789 CmpPredicate Pred = IsAnd ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
790 // Make sure we have right compares for our op.
791 if (LHS->getPredicate() != Pred || RHS->getPredicate() != Pred)
792 return nullptr;
793
794 // Make it so we can match LHS against the (icmp eq/ne X, 0) just for
795 // simplicity.
796 if (match(V: RHS->getOperand(i_nocapture: 1), P: m_Zero()))
797 std::swap(a&: LHS, b&: RHS);
798
799 Value *Pow2, *Op;
800 // Match the desired pattern:
801 // LHS: (icmp eq/ne X, 0)
802 // RHS: (icmp eq/ne X, Pow2OrZero)
803 // Skip if Pow2OrZero is 1. Either way it gets folded to (icmp ugt X, 1) but
804 // this form ends up slightly less canonical.
805 // We could potentially be more sophisticated than requiring LHS/RHS
806 // be one-use. We don't create additional instructions if only one
807 // of them is one-use. So cases where one is one-use and the other
808 // is two-use might be profitable.
809 if (!match(V: LHS, P: m_OneUse(SubPattern: m_ICmp(Pred, L: m_Value(V&: Op), R: m_Zero()))) ||
810 !match(V: RHS, P: m_OneUse(SubPattern: m_c_ICmp(Pred, L: m_Specific(V: Op), R: m_Value(V&: Pow2)))) ||
811 match(V: Pow2, P: m_One()) ||
812 !isKnownToBeAPowerOfTwo(V: Pow2, DL: Q.DL, /*OrZero=*/true, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT))
813 return nullptr;
814
815 Value *And = Builder.CreateAnd(LHS: Op, RHS: Pow2);
816 return Builder.CreateICmp(P: Pred, LHS: And, RHS: Op);
817}
818
819/// General pattern:
820/// X & Y
821///
822/// Where Y is checking that all the high bits (covered by a mask 4294967168)
823/// are uniform, i.e. %arg & 4294967168 can be either 4294967168 or 0
824/// Pattern can be one of:
825/// %t = add i32 %arg, 128
826/// %r = icmp ult i32 %t, 256
827/// Or
828/// %t0 = shl i32 %arg, 24
829/// %t1 = ashr i32 %t0, 24
830/// %r = icmp eq i32 %t1, %arg
831/// Or
832/// %t0 = trunc i32 %arg to i8
833/// %t1 = sext i8 %t0 to i32
834/// %r = icmp eq i32 %t1, %arg
835/// This pattern is a signed truncation check.
836///
837/// And X is checking that some bit in that same mask is zero.
838/// I.e. can be one of:
839/// %r = icmp sgt i32 %arg, -1
840/// Or
841/// %t = and i32 %arg, 2147483648
842/// %r = icmp eq i32 %t, 0
843///
844/// Since we are checking that all the bits in that mask are the same,
845/// and a particular bit is zero, what we are really checking is that all the
846/// masked bits are zero.
847/// So this should be transformed to:
848/// %r = icmp ult i32 %arg, 128
849static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1,
850 Instruction &CxtI,
851 InstCombiner::BuilderTy &Builder) {
852 assert(CxtI.getOpcode() == Instruction::And);
853
854 // Match icmp ult (add %arg, C01), C1 (C1 == C01 << 1; powers of two)
855 auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X,
856 APInt &SignBitMask) -> bool {
857 const APInt *I01, *I1; // powers of two; I1 == I01 << 1
858 if (!(match(V: ICmp, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_ULT,
859 L: m_Add(L: m_Value(V&: X), R: m_Power2(V&: I01)),
860 R: m_Power2(V&: I1))) &&
861 I1->ugt(RHS: *I01) && I01->shl(shiftAmt: 1) == *I1))
862 return false;
863 // Which bit is the new sign bit as per the 'signed truncation' pattern?
864 SignBitMask = *I01;
865 return true;
866 };
867
868 // One icmp needs to be 'signed truncation check'.
869 // We need to match this first, else we will mismatch commutative cases.
870 Value *X1;
871 APInt HighestBit;
872 ICmpInst *OtherICmp;
873 if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit))
874 OtherICmp = ICmp0;
875 else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit))
876 OtherICmp = ICmp1;
877 else
878 return nullptr;
879
880 assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)");
881
882 // Try to match/decompose into: icmp eq (X & Mask), 0
883 auto tryToDecompose = [](ICmpInst *ICmp, Value *&X,
884 APInt &UnsetBitsMask) -> bool {
885 CmpPredicate Pred = ICmp->getPredicate();
886 // Can it be decomposed into icmp eq (X & Mask), 0 ?
887 auto Res = llvm::decomposeBitTestICmp(
888 LHS: ICmp->getOperand(i_nocapture: 0), RHS: ICmp->getOperand(i_nocapture: 1), Pred,
889 /*LookThroughTrunc=*/false, /*AllowNonZeroC=*/false,
890 /*DecomposeAnd=*/true);
891 if (Res && Res->Pred == ICmpInst::ICMP_EQ) {
892 X = Res->X;
893 UnsetBitsMask = Res->Mask;
894 return true;
895 }
896
897 return false;
898 };
899
900 // And the other icmp needs to be decomposable into a bit test.
901 Value *X0;
902 APInt UnsetBitsMask;
903 if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask))
904 return nullptr;
905
906 assert(!UnsetBitsMask.isZero() && "empty mask makes no sense.");
907
908 // Are they working on the same value?
909 Value *X;
910 if (X1 == X0) {
911 // Ok as is.
912 X = X1;
913 } else if (match(V: X0, P: m_Trunc(Op: m_Specific(V: X1)))) {
914 UnsetBitsMask = UnsetBitsMask.zext(width: X1->getType()->getScalarSizeInBits());
915 X = X1;
916 } else
917 return nullptr;
918
919 // So which bits should be uniform as per the 'signed truncation check'?
920 // (all the bits starting with (i.e. including) HighestBit)
921 APInt SignBitsMask = ~(HighestBit - 1U);
922
923 // UnsetBitsMask must have some common bits with SignBitsMask,
924 if (!UnsetBitsMask.intersects(RHS: SignBitsMask))
925 return nullptr;
926
927 // Does UnsetBitsMask contain any bits outside of SignBitsMask?
928 if (!UnsetBitsMask.isSubsetOf(RHS: SignBitsMask)) {
929 APInt OtherHighestBit = (~UnsetBitsMask) + 1U;
930 if (!OtherHighestBit.isPowerOf2())
931 return nullptr;
932 HighestBit = APIntOps::umin(A: HighestBit, B: OtherHighestBit);
933 }
934 // Else, if it does not, then all is ok as-is.
935
936 // %r = icmp ult %X, SignBit
937 return Builder.CreateICmpULT(LHS: X, RHS: ConstantInt::get(Ty: X->getType(), V: HighestBit),
938 Name: CxtI.getName() + ".simplified");
939}
940
941/// Fold (icmp eq ctpop(X) 1) | (icmp eq X 0) into (icmp ult ctpop(X) 2) and
942/// fold (icmp ne ctpop(X) 1) & (icmp ne X 0) into (icmp ugt ctpop(X) 1).
943/// Also used for logical and/or, must be poison safe if range attributes are
944/// dropped.
945static Value *foldIsPowerOf2OrZero(ICmpInst *Cmp0, ICmpInst *Cmp1, bool IsAnd,
946 InstCombiner::BuilderTy &Builder,
947 InstCombinerImpl &IC) {
948 CmpPredicate Pred0, Pred1;
949 Value *X;
950 if (!match(V: Cmp0, P: m_ICmp(Pred&: Pred0, L: m_Ctpop(Op0: m_Value(V&: X)), R: m_SpecificInt(V: 1))) ||
951 !match(V: Cmp1, P: m_ICmp(Pred&: Pred1, L: m_Specific(V: X), R: m_ZeroInt())))
952 return nullptr;
953
954 auto *CtPop = cast<Instruction>(Val: Cmp0->getOperand(i_nocapture: 0));
955 if (IsAnd && Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_NE) {
956 // Drop range attributes and re-infer them in the next iteration.
957 CtPop->dropPoisonGeneratingAnnotations();
958 IC.addToWorklist(I: CtPop);
959 return Builder.CreateICmpUGT(LHS: CtPop, RHS: ConstantInt::get(Ty: CtPop->getType(), V: 1));
960 }
961 if (!IsAnd && Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_EQ) {
962 // Drop range attributes and re-infer them in the next iteration.
963 CtPop->dropPoisonGeneratingAnnotations();
964 IC.addToWorklist(I: CtPop);
965 return Builder.CreateICmpULT(LHS: CtPop, RHS: ConstantInt::get(Ty: CtPop->getType(), V: 2));
966 }
967
968 return nullptr;
969}
970
971/// Reduce a pair of compares that check if a value has exactly 1 bit set.
972/// Also used for logical and/or, must be poison safe if range attributes are
973/// dropped.
974static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd,
975 InstCombiner::BuilderTy &Builder,
976 InstCombinerImpl &IC) {
977 // Handle 'and' / 'or' commutation: make the equality check the first operand.
978 if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE)
979 std::swap(a&: Cmp0, b&: Cmp1);
980 else if (!JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_EQ)
981 std::swap(a&: Cmp0, b&: Cmp1);
982
983 // (X != 0) && (ctpop(X) u< 2) --> ctpop(X) == 1
984 Value *X;
985 if (JoinedByAnd &&
986 match(V: Cmp0, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, L: m_Value(V&: X), R: m_ZeroInt())) &&
987 match(V: Cmp1, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_ULT, L: m_Ctpop(Op0: m_Specific(V: X)),
988 R: m_SpecificInt(V: 2)))) {
989 auto *CtPop = cast<Instruction>(Val: Cmp1->getOperand(i_nocapture: 0));
990 // Drop range attributes and re-infer them in the next iteration.
991 CtPop->dropPoisonGeneratingAnnotations();
992 IC.addToWorklist(I: CtPop);
993 return Builder.CreateICmpEQ(LHS: CtPop, RHS: ConstantInt::get(Ty: CtPop->getType(), V: 1));
994 }
995 // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1
996 if (!JoinedByAnd &&
997 match(V: Cmp0, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Value(V&: X), R: m_ZeroInt())) &&
998 match(V: Cmp1, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_UGT, L: m_Ctpop(Op0: m_Specific(V: X)),
999 R: m_SpecificInt(V: 1)))) {
1000 auto *CtPop = cast<Instruction>(Val: Cmp1->getOperand(i_nocapture: 0));
1001 // Drop range attributes and re-infer them in the next iteration.
1002 CtPop->dropPoisonGeneratingAnnotations();
1003 IC.addToWorklist(I: CtPop);
1004 return Builder.CreateICmpNE(LHS: CtPop, RHS: ConstantInt::get(Ty: CtPop->getType(), V: 1));
1005 }
1006 return nullptr;
1007}
1008
1009/// Try to fold (icmp(A & B) == 0) & (icmp(A & D) != E) into (icmp A u< D) iff
1010/// B is a contiguous set of ones starting from the most significant bit
1011/// (negative power of 2), D and E are equal, and D is a contiguous set of ones
1012/// starting at the most significant zero bit in B. Parameter B supports masking
1013/// using undef/poison in either scalar or vector values.
1014static Value *foldNegativePower2AndShiftedMask(
1015 Value *A, Value *B, Value *D, Value *E, ICmpInst::Predicate PredL,
1016 ICmpInst::Predicate PredR, InstCombiner::BuilderTy &Builder) {
1017 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
1018 "Expected equality predicates for masked type of icmps.");
1019 if (PredL != ICmpInst::ICMP_EQ || PredR != ICmpInst::ICMP_NE)
1020 return nullptr;
1021
1022 if (!match(V: B, P: m_NegatedPower2()) || !match(V: D, P: m_ShiftedMask()) ||
1023 !match(V: E, P: m_ShiftedMask()))
1024 return nullptr;
1025
1026 // Test scalar arguments for conversion. B has been validated earlier to be a
1027 // negative power of two and thus is guaranteed to have one or more contiguous
1028 // ones starting from the MSB followed by zero or more contiguous zeros. D has
1029 // been validated earlier to be a shifted set of one or more contiguous ones.
1030 // In order to match, B leading ones and D leading zeros should be equal. The
1031 // predicate that B be a negative power of 2 prevents the condition of there
1032 // ever being zero leading ones. Thus 0 == 0 cannot occur. The predicate that
1033 // D always be a shifted mask prevents the condition of D equaling 0. This
1034 // prevents matching the condition where B contains the maximum number of
1035 // leading one bits (-1) and D contains the maximum number of leading zero
1036 // bits (0).
1037 auto isReducible = [](const Value *B, const Value *D, const Value *E) {
1038 const APInt *BCst, *DCst, *ECst;
1039 return match(V: B, P: m_APIntAllowPoison(Res&: BCst)) && match(V: D, P: m_APInt(Res&: DCst)) &&
1040 match(V: E, P: m_APInt(Res&: ECst)) && *DCst == *ECst &&
1041 (isa<PoisonValue>(Val: B) ||
1042 (BCst->countLeadingOnes() == DCst->countLeadingZeros()));
1043 };
1044
1045 // Test vector type arguments for conversion.
1046 if (const auto *BVTy = dyn_cast<VectorType>(Val: B->getType())) {
1047 const auto *BFVTy = dyn_cast<FixedVectorType>(Val: BVTy);
1048 const auto *BConst = dyn_cast<Constant>(Val: B);
1049 const auto *DConst = dyn_cast<Constant>(Val: D);
1050 const auto *EConst = dyn_cast<Constant>(Val: E);
1051
1052 if (!BFVTy || !BConst || !DConst || !EConst)
1053 return nullptr;
1054
1055 for (unsigned I = 0; I != BFVTy->getNumElements(); ++I) {
1056 const auto *BElt = BConst->getAggregateElement(Elt: I);
1057 const auto *DElt = DConst->getAggregateElement(Elt: I);
1058 const auto *EElt = EConst->getAggregateElement(Elt: I);
1059
1060 if (!BElt || !DElt || !EElt)
1061 return nullptr;
1062 if (!isReducible(BElt, DElt, EElt))
1063 return nullptr;
1064 }
1065 } else {
1066 // Test scalar type arguments for conversion.
1067 if (!isReducible(B, D, E))
1068 return nullptr;
1069 }
1070 return Builder.CreateICmp(P: ICmpInst::ICMP_ULT, LHS: A, RHS: D);
1071}
1072
1073/// Try to fold ((icmp X u< P) & (icmp(X & M) != M)) or ((icmp X s> -1) &
1074/// (icmp(X & M) != M)) into (icmp X u< M). Where P is a power of 2, M < P, and
1075/// M is a contiguous shifted mask starting at the right most significant zero
1076/// bit in P. SGT is supported as when P is the largest representable power of
1077/// 2, an earlier optimization converts the expression into (icmp X s> -1).
1078/// Parameter P supports masking using undef/poison in either scalar or vector
1079/// values.
1080static Value *foldPowerOf2AndShiftedMask(ICmpInst *Cmp0, ICmpInst *Cmp1,
1081 bool JoinedByAnd,
1082 InstCombiner::BuilderTy &Builder) {
1083 if (!JoinedByAnd)
1084 return nullptr;
1085 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
1086 ICmpInst::Predicate CmpPred0, CmpPred1;
1087 // Assuming P is a 2^n, getMaskedTypeForICmpPair will normalize (icmp X u<
1088 // 2^n) into (icmp (X & ~(2^n-1)) == 0) and (icmp X s> -1) into (icmp (X &
1089 // SignMask) == 0).
1090 std::optional<std::pair<unsigned, unsigned>> MaskPair =
1091 getMaskedTypeForICmpPair(A, B, C, D, E, LHS: Cmp0, RHS: Cmp1, PredL&: CmpPred0, PredR&: CmpPred1);
1092 if (!MaskPair)
1093 return nullptr;
1094
1095 const auto compareBMask = BMask_NotMixed | BMask_NotAllOnes;
1096 unsigned CmpMask0 = MaskPair->first;
1097 unsigned CmpMask1 = MaskPair->second;
1098 if ((CmpMask0 & Mask_AllZeros) && (CmpMask1 == compareBMask)) {
1099 if (Value *V = foldNegativePower2AndShiftedMask(A, B, D, E, PredL: CmpPred0,
1100 PredR: CmpPred1, Builder))
1101 return V;
1102 } else if ((CmpMask0 == compareBMask) && (CmpMask1 & Mask_AllZeros)) {
1103 if (Value *V = foldNegativePower2AndShiftedMask(A, B: D, D: B, E: C, PredL: CmpPred1,
1104 PredR: CmpPred0, Builder))
1105 return V;
1106 }
1107 return nullptr;
1108}
1109
1110/// Commuted variants are assumed to be handled by calling this function again
1111/// with the parameters swapped.
1112static Value *foldUnsignedUnderflowCheck(ICmpInst *ZeroICmp,
1113 ICmpInst *UnsignedICmp, bool IsAnd,
1114 const SimplifyQuery &Q,
1115 InstCombiner::BuilderTy &Builder) {
1116 Value *ZeroCmpOp;
1117 CmpPredicate EqPred;
1118 if (!match(V: ZeroICmp, P: m_ICmp(Pred&: EqPred, L: m_Value(V&: ZeroCmpOp), R: m_Zero())) ||
1119 !ICmpInst::isEquality(P: EqPred))
1120 return nullptr;
1121
1122 CmpPredicate UnsignedPred;
1123
1124 Value *A, *B;
1125 if (match(V: UnsignedICmp,
1126 P: m_c_ICmp(Pred&: UnsignedPred, L: m_Specific(V: ZeroCmpOp), R: m_Value(V&: A))) &&
1127 match(V: ZeroCmpOp, P: m_c_Add(L: m_Specific(V: A), R: m_Value(V&: B))) &&
1128 (ZeroICmp->hasOneUse() || UnsignedICmp->hasOneUse())) {
1129 auto GetKnownNonZeroAndOther = [&](Value *&NonZero, Value *&Other) {
1130 if (!isKnownNonZero(V: NonZero, Q))
1131 std::swap(a&: NonZero, b&: Other);
1132 return isKnownNonZero(V: NonZero, Q);
1133 };
1134
1135 // Given ZeroCmpOp = (A + B)
1136 // ZeroCmpOp < A && ZeroCmpOp != 0 --> (0-X) < Y iff
1137 // ZeroCmpOp >= A || ZeroCmpOp == 0 --> (0-X) >= Y iff
1138 // with X being the value (A/B) that is known to be non-zero,
1139 // and Y being remaining value.
1140 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE &&
1141 IsAnd && GetKnownNonZeroAndOther(B, A))
1142 return Builder.CreateICmpULT(LHS: Builder.CreateNeg(V: B), RHS: A);
1143 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ &&
1144 !IsAnd && GetKnownNonZeroAndOther(B, A))
1145 return Builder.CreateICmpUGE(LHS: Builder.CreateNeg(V: B), RHS: A);
1146 }
1147
1148 return nullptr;
1149}
1150
1151struct IntPart {
1152 Value *From;
1153 unsigned StartBit;
1154 unsigned NumBits;
1155};
1156
1157/// Match an extraction of bits from an integer.
1158static std::optional<IntPart> matchIntPart(Value *V) {
1159 Value *X;
1160 if (!match(V, P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: X)))))
1161 return std::nullopt;
1162
1163 unsigned NumOriginalBits = X->getType()->getScalarSizeInBits();
1164 unsigned NumExtractedBits = V->getType()->getScalarSizeInBits();
1165 Value *Y;
1166 const APInt *Shift;
1167 // For a trunc(lshr Y, Shift) pattern, make sure we're only extracting bits
1168 // from Y, not any shifted-in zeroes.
1169 if (match(V: X, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: Y), R: m_APInt(Res&: Shift)))) &&
1170 Shift->ule(RHS: NumOriginalBits - NumExtractedBits))
1171 return {{.From: Y, .StartBit: (unsigned)Shift->getZExtValue(), .NumBits: NumExtractedBits}};
1172 return {{.From: X, .StartBit: 0, .NumBits: NumExtractedBits}};
1173}
1174
1175/// Materialize an extraction of bits from an integer in IR.
1176static Value *extractIntPart(const IntPart &P, IRBuilderBase &Builder) {
1177 Value *V = P.From;
1178 if (P.StartBit)
1179 V = Builder.CreateLShr(LHS: V, RHS: P.StartBit);
1180 Type *TruncTy = V->getType()->getWithNewBitWidth(NewBitWidth: P.NumBits);
1181 if (TruncTy != V->getType())
1182 V = Builder.CreateTrunc(V, DestTy: TruncTy);
1183 return V;
1184}
1185
1186/// (icmp eq X0, Y0) & (icmp eq X1, Y1) -> icmp eq X01, Y01
1187/// (icmp ne X0, Y0) | (icmp ne X1, Y1) -> icmp ne X01, Y01
1188/// where X0, X1 and Y0, Y1 are adjacent parts extracted from an integer.
1189Value *InstCombinerImpl::foldEqOfParts(Value *Cmp0, Value *Cmp1, bool IsAnd) {
1190 if (!Cmp0->hasOneUse() || !Cmp1->hasOneUse())
1191 return nullptr;
1192
1193 CmpInst::Predicate Pred = IsAnd ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1194 auto GetMatchPart = [&](Value *CmpV,
1195 unsigned OpNo) -> std::optional<IntPart> {
1196 assert(CmpV->getType()->isIntOrIntVectorTy(1) && "Must be bool");
1197
1198 Value *X, *Y;
1199 // icmp ne (and x, 1), (and y, 1) <=> trunc (xor x, y) to i1
1200 // icmp eq (and x, 1), (and y, 1) <=> not (trunc (xor x, y) to i1)
1201 if (Pred == CmpInst::ICMP_NE
1202 ? match(V: CmpV, P: m_Trunc(Op: m_Xor(L: m_Value(V&: X), R: m_Value(V&: Y))))
1203 : match(V: CmpV, P: m_Not(V: m_Trunc(Op: m_Xor(L: m_Value(V&: X), R: m_Value(V&: Y))))))
1204 return {{.From: OpNo == 0 ? X : Y, .StartBit: 0, .NumBits: 1}};
1205
1206 auto *Cmp = dyn_cast<ICmpInst>(Val: CmpV);
1207 if (!Cmp)
1208 return std::nullopt;
1209
1210 if (Pred == Cmp->getPredicate())
1211 return matchIntPart(V: Cmp->getOperand(i_nocapture: OpNo));
1212
1213 const APInt *C;
1214 // (icmp eq (lshr x, C), (lshr y, C)) gets optimized to:
1215 // (icmp ult (xor x, y), 1 << C) so also look for that.
1216 if (Pred == CmpInst::ICMP_EQ && Cmp->getPredicate() == CmpInst::ICMP_ULT) {
1217 if (!match(V: Cmp->getOperand(i_nocapture: 1), P: m_Power2(V&: C)) ||
1218 !match(V: Cmp->getOperand(i_nocapture: 0), P: m_Xor(L: m_Value(), R: m_Value())))
1219 return std::nullopt;
1220 }
1221
1222 // (icmp ne (lshr x, C), (lshr y, C)) gets optimized to:
1223 // (icmp ugt (xor x, y), (1 << C) - 1) so also look for that.
1224 else if (Pred == CmpInst::ICMP_NE &&
1225 Cmp->getPredicate() == CmpInst::ICMP_UGT) {
1226 if (!match(V: Cmp->getOperand(i_nocapture: 1), P: m_LowBitMask(V&: C)) ||
1227 !match(V: Cmp->getOperand(i_nocapture: 0), P: m_Xor(L: m_Value(), R: m_Value())))
1228 return std::nullopt;
1229 } else {
1230 return std::nullopt;
1231 }
1232
1233 unsigned From = Pred == CmpInst::ICMP_NE ? C->popcount() : C->countr_zero();
1234 Instruction *I = cast<Instruction>(Val: Cmp->getOperand(i_nocapture: 0));
1235 return {{.From: I->getOperand(i: OpNo), .StartBit: From, .NumBits: C->getBitWidth() - From}};
1236 };
1237
1238 std::optional<IntPart> L0 = GetMatchPart(Cmp0, 0);
1239 std::optional<IntPart> R0 = GetMatchPart(Cmp0, 1);
1240 std::optional<IntPart> L1 = GetMatchPart(Cmp1, 0);
1241 std::optional<IntPart> R1 = GetMatchPart(Cmp1, 1);
1242 if (!L0 || !R0 || !L1 || !R1)
1243 return nullptr;
1244
1245 // Make sure the LHS/RHS compare a part of the same value, possibly after
1246 // an operand swap.
1247 if (L0->From != L1->From || R0->From != R1->From) {
1248 if (L0->From != R1->From || R0->From != L1->From)
1249 return nullptr;
1250 std::swap(lhs&: L1, rhs&: R1);
1251 }
1252
1253 // Make sure the extracted parts are adjacent, canonicalizing to L0/R0 being
1254 // the low part and L1/R1 being the high part.
1255 if (L0->StartBit + L0->NumBits != L1->StartBit ||
1256 R0->StartBit + R0->NumBits != R1->StartBit) {
1257 if (L1->StartBit + L1->NumBits != L0->StartBit ||
1258 R1->StartBit + R1->NumBits != R0->StartBit)
1259 return nullptr;
1260 std::swap(lhs&: L0, rhs&: L1);
1261 std::swap(lhs&: R0, rhs&: R1);
1262 }
1263
1264 // We can simplify to a comparison of these larger parts of the integers.
1265 IntPart L = {.From: L0->From, .StartBit: L0->StartBit, .NumBits: L0->NumBits + L1->NumBits};
1266 IntPart R = {.From: R0->From, .StartBit: R0->StartBit, .NumBits: R0->NumBits + R1->NumBits};
1267 Value *LValue = extractIntPart(P: L, Builder);
1268 Value *RValue = extractIntPart(P: R, Builder);
1269 return Builder.CreateICmp(P: Pred, LHS: LValue, RHS: RValue);
1270}
1271
1272/// Reduce logic-of-compares with equality to a constant by substituting a
1273/// common operand with the constant. Callers are expected to call this with
1274/// Cmp0/Cmp1 switched to handle logic op commutativity.
1275static Value *foldAndOrOfICmpsWithConstEq(ICmpInst *Cmp0, ICmpInst *Cmp1,
1276 bool IsAnd, bool IsLogical,
1277 InstCombiner::BuilderTy &Builder,
1278 const SimplifyQuery &Q,
1279 Instruction &I) {
1280 // Match an equality compare with a non-poison constant as Cmp0.
1281 // Also, give up if the compare can be constant-folded to avoid looping.
1282 CmpPredicate Pred0;
1283 Value *X;
1284 Constant *C;
1285 if (!match(V: Cmp0, P: m_ICmp(Pred&: Pred0, L: m_Value(V&: X), R: m_Constant(C))) ||
1286 !isGuaranteedNotToBeUndefOrPoison(V: C) || isa<Constant>(Val: X))
1287 return nullptr;
1288 if ((IsAnd && Pred0 != ICmpInst::ICMP_EQ) ||
1289 (!IsAnd && Pred0 != ICmpInst::ICMP_NE))
1290 return nullptr;
1291
1292 // The other compare must include a common operand (X). Canonicalize the
1293 // common operand as operand 1 (Pred1 is swapped if the common operand was
1294 // operand 0).
1295 Value *Y;
1296 CmpPredicate Pred1;
1297 if (!match(V: Cmp1, P: m_c_ICmp(Pred&: Pred1, L: m_Value(V&: Y), R: m_Specific(V: X))))
1298 return nullptr;
1299
1300 // Replace variable with constant value equivalence to remove a variable use:
1301 // (X == C) && (Y Pred1 X) --> (X == C) && (Y Pred1 C)
1302 // (X != C) || (Y Pred1 X) --> (X != C) || (Y Pred1 C)
1303 // Can think of the 'or' substitution with the 'and' bool equivalent:
1304 // A || B --> A || (!A && B)
1305 Value *SubstituteCmp = simplifyICmpInst(Pred: Pred1, LHS: Y, RHS: C, Q);
1306 if (!SubstituteCmp) {
1307 // If we need to create a new instruction, require that the old compare can
1308 // be removed.
1309 if (!Cmp1->hasOneUse())
1310 return nullptr;
1311 SubstituteCmp = Builder.CreateICmp(P: Pred1, LHS: Y, RHS: C);
1312 }
1313 if (IsLogical) {
1314 Instruction *MDFrom =
1315 ProfcheckDisableMetadataFixes && isa<SelectInst>(Val: I) ? nullptr : &I;
1316 return IsAnd ? Builder.CreateLogicalAnd(Cond1: Cmp0, Cond2: SubstituteCmp, Name: "", MDFrom)
1317 : Builder.CreateLogicalOr(Cond1: Cmp0, Cond2: SubstituteCmp, Name: "", MDFrom);
1318 }
1319 return Builder.CreateBinOp(Opc: IsAnd ? Instruction::And : Instruction::Or, LHS: Cmp0,
1320 RHS: SubstituteCmp);
1321}
1322
1323/// Fold (icmp Pred1 V1, C1) & (icmp Pred2 V2, C2)
1324/// or (icmp Pred1 V1, C1) | (icmp Pred2 V2, C2)
1325/// into a single comparison using range-based reasoning.
1326/// NOTE: This is also used for logical and/or, must be poison-safe!
1327Value *InstCombinerImpl::foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1,
1328 ICmpInst *ICmp2,
1329 bool IsAnd) {
1330 // Return (V, CR) for a range check idiom V in CR.
1331 auto MatchExactRangeCheck =
1332 [](ICmpInst *ICmp) -> std::optional<std::pair<Value *, ConstantRange>> {
1333 const APInt *C;
1334 if (!match(V: ICmp->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)))
1335 return std::nullopt;
1336 Value *LHS = ICmp->getOperand(i_nocapture: 0);
1337 CmpPredicate Pred = ICmp->getPredicate();
1338 Value *X;
1339 // Match (x & NegPow2) ==/!= C
1340 const APInt *Mask;
1341 if (ICmpInst::isEquality(P: Pred) &&
1342 match(V: LHS, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_NegatedPower2(V&: Mask)))) &&
1343 C->countr_zero() >= Mask->countr_zero()) {
1344 ConstantRange CR(*C, *C - *Mask);
1345 if (Pred == ICmpInst::ICMP_NE)
1346 CR = CR.inverse();
1347 return std::make_pair(x&: X, y&: CR);
1348 }
1349 ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred, Other: *C);
1350 // Match (add X, C1) pred C
1351 // TODO: investigate whether we should apply the one-use check on m_AddLike.
1352 const APInt *C1;
1353 if (match(V: LHS, P: m_AddLike(L: m_Value(V&: X), R: m_APInt(Res&: C1))))
1354 return std::make_pair(x&: X, y: CR.subtract(CI: *C1));
1355 return std::make_pair(x&: LHS, y&: CR);
1356 };
1357
1358 auto RC1 = MatchExactRangeCheck(ICmp1);
1359 if (!RC1)
1360 return nullptr;
1361
1362 auto RC2 = MatchExactRangeCheck(ICmp2);
1363 if (!RC2)
1364 return nullptr;
1365
1366 auto &[V1, CR1] = *RC1;
1367 auto &[V2, CR2] = *RC2;
1368 if (V1 != V2)
1369 return nullptr;
1370
1371 // For 'and', we use the De Morgan's Laws to simplify the implementation.
1372 if (IsAnd) {
1373 CR1 = CR1.inverse();
1374 CR2 = CR2.inverse();
1375 }
1376
1377 Type *Ty = V1->getType();
1378 Value *NewV = V1;
1379 std::optional<ConstantRange> CR = CR1.exactUnionWith(CR: CR2);
1380 if (!CR) {
1381 if (!(ICmp1->hasOneUse() && ICmp2->hasOneUse()) || CR1.isWrappedSet() ||
1382 CR2.isWrappedSet())
1383 return nullptr;
1384
1385 // Check whether we have equal-size ranges that only differ by one bit.
1386 // In that case we can apply a mask to map one range onto the other.
1387 APInt LowerDiff = CR1.getLower() ^ CR2.getLower();
1388 APInt UpperDiff = (CR1.getUpper() - 1) ^ (CR2.getUpper() - 1);
1389 APInt CR1Size = CR1.getUpper() - CR1.getLower();
1390 if (!LowerDiff.isPowerOf2() || LowerDiff != UpperDiff ||
1391 CR1Size != CR2.getUpper() - CR2.getLower())
1392 return nullptr;
1393
1394 CR = CR1.getLower().ult(RHS: CR2.getLower()) ? CR1 : CR2;
1395 NewV = Builder.CreateAnd(LHS: NewV, RHS: ConstantInt::get(Ty, V: ~LowerDiff));
1396 }
1397
1398 if (IsAnd)
1399 CR = CR->inverse();
1400
1401 CmpInst::Predicate NewPred;
1402 APInt NewC, Offset;
1403 CR->getEquivalentICmp(Pred&: NewPred, RHS&: NewC, Offset);
1404
1405 if (Offset != 0)
1406 NewV = Builder.CreateAdd(LHS: NewV, RHS: ConstantInt::get(Ty, V: Offset));
1407 return Builder.CreateICmp(P: NewPred, LHS: NewV, RHS: ConstantInt::get(Ty, V: NewC));
1408}
1409
1410/// Matches canonical form of isnan, fcmp ord x, 0
1411static bool matchIsNotNaN(FCmpInst::Predicate P, Value *LHS, Value *RHS) {
1412 return P == FCmpInst::FCMP_ORD && match(V: RHS, P: m_AnyZeroFP());
1413}
1414
1415/// Matches fcmp u__ x, +/-inf
1416static bool matchUnorderedInfCompare(FCmpInst::Predicate P, Value *LHS,
1417 Value *RHS) {
1418 return FCmpInst::isUnordered(predicate: P) && match(V: RHS, P: m_Inf());
1419}
1420
1421/// and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf
1422///
1423/// Clang emits this pattern for doing an isfinite check in __builtin_isnormal.
1424static Value *matchIsFiniteTest(InstCombiner::BuilderTy &Builder, FCmpInst *LHS,
1425 FCmpInst *RHS) {
1426 Value *LHS0 = LHS->getOperand(i_nocapture: 0), *LHS1 = LHS->getOperand(i_nocapture: 1);
1427 Value *RHS0 = RHS->getOperand(i_nocapture: 0), *RHS1 = RHS->getOperand(i_nocapture: 1);
1428 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1429
1430 if (!matchIsNotNaN(P: PredL, LHS: LHS0, RHS: LHS1) ||
1431 !matchUnorderedInfCompare(P: PredR, LHS: RHS0, RHS: RHS1))
1432 return nullptr;
1433
1434 return Builder.CreateFCmpFMF(P: FCmpInst::getOrderedPredicate(Pred: PredR), LHS: RHS0, RHS: RHS1,
1435 FMFSource: FMFSource::intersect(A: LHS, B: RHS));
1436}
1437
1438Value *InstCombinerImpl::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS,
1439 bool IsAnd, bool IsLogicalSelect) {
1440 Value *LHS0 = LHS->getOperand(i_nocapture: 0), *LHS1 = LHS->getOperand(i_nocapture: 1);
1441 Value *RHS0 = RHS->getOperand(i_nocapture: 0), *RHS1 = RHS->getOperand(i_nocapture: 1);
1442 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1443
1444 if (LHS0 == RHS1 && RHS0 == LHS1) {
1445 // Swap RHS operands to match LHS.
1446 PredR = FCmpInst::getSwappedPredicate(pred: PredR);
1447 std::swap(a&: RHS0, b&: RHS1);
1448 }
1449
1450 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1451 // Suppose the relation between x and y is R, where R is one of
1452 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for
1453 // testing the desired relations.
1454 //
1455 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1456 // bool(R & CC0) && bool(R & CC1)
1457 // = bool((R & CC0) & (R & CC1))
1458 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency
1459 //
1460 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1461 // bool(R & CC0) || bool(R & CC1)
1462 // = bool((R & CC0) | (R & CC1))
1463 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)
1464 if (LHS0 == RHS0 && LHS1 == RHS1) {
1465 unsigned FCmpCodeL = getFCmpCode(CC: PredL);
1466 unsigned FCmpCodeR = getFCmpCode(CC: PredR);
1467 unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR;
1468
1469 // Intersect the fast math flags.
1470 // TODO: We can union the fast math flags unless this is a logical select.
1471 return getFCmpValue(Code: NewPred, LHS: LHS0, RHS: LHS1, Builder,
1472 FMF: FMFSource::intersect(A: LHS, B: RHS));
1473 }
1474
1475 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1476 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1477 if (LHS0->getType() != RHS0->getType())
1478 return nullptr;
1479
1480 // FCmp canonicalization ensures that (fcmp ord/uno X, X) and
1481 // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0).
1482 if (match(V: LHS1, P: m_PosZeroFP()) && match(V: RHS1, P: m_PosZeroFP())) {
1483 // Ignore the constants because they are obviously not NANs:
1484 // (fcmp ord x, 0.0) & (fcmp ord y, 0.0) -> (fcmp ord x, y)
1485 // (fcmp uno x, 0.0) | (fcmp uno y, 0.0) -> (fcmp uno x, y)
1486 Value *Y = RHS0;
1487 FastMathFlags FMF = LHS->getFastMathFlags() & RHS->getFastMathFlags();
1488 if (IsLogicalSelect) {
1489 Y = Builder.CreateFreeze(V: Y, Name: Y->getName() + ".fr");
1490 FMF.setNoNaNs(false);
1491 FMF.setNoInfs(false);
1492 }
1493 return Builder.CreateFCmpFMF(P: PredL, LHS: LHS0, RHS: Y, FMFSource: FMF);
1494 }
1495 }
1496
1497 // This transform is not valid for a logical select.
1498 if (!IsLogicalSelect && IsAnd &&
1499 stripSignOnlyFPOps(Val: LHS0) == stripSignOnlyFPOps(Val: RHS0)) {
1500 // and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf
1501 // and (fcmp ord x, 0), (fcmp u* fabs(x), inf) -> fcmp o* x, inf
1502 if (Value *Left = matchIsFiniteTest(Builder, LHS, RHS))
1503 return Left;
1504 if (Value *Right = matchIsFiniteTest(Builder, LHS: RHS, RHS: LHS))
1505 return Right;
1506 }
1507
1508 // Turn at least two fcmps with constants into llvm.is.fpclass.
1509 //
1510 // If we can represent a combined value test with one class call, we can
1511 // potentially eliminate 4-6 instructions. If we can represent a test with a
1512 // single fcmp with fneg and fabs, that's likely a better canonical form.
1513 if (LHS->hasOneUse() && RHS->hasOneUse()) {
1514 auto [ClassValRHS, ClassMaskRHS] =
1515 fcmpToClassTest(Pred: PredR, F: *RHS->getFunction(), LHS: RHS0, RHS: RHS1);
1516 if (ClassValRHS) {
1517 auto [ClassValLHS, ClassMaskLHS] =
1518 fcmpToClassTest(Pred: PredL, F: *LHS->getFunction(), LHS: LHS0, RHS: LHS1);
1519 if (ClassValLHS == ClassValRHS) {
1520 unsigned CombinedMask = IsAnd ? (ClassMaskLHS & ClassMaskRHS)
1521 : (ClassMaskLHS | ClassMaskRHS);
1522 return Builder.CreateIntrinsic(
1523 ID: Intrinsic::is_fpclass, OverloadTypes: {ClassValLHS->getType()},
1524 Args: {ClassValLHS, Builder.getInt32(C: CombinedMask)});
1525 }
1526 }
1527 }
1528
1529 // Canonicalize the range check idiom:
1530 // and (fcmp olt/ole/ult/ule x, C), (fcmp ogt/oge/ugt/uge x, -C)
1531 // --> fabs(x) olt/ole/ult/ule C
1532 // or (fcmp ogt/oge/ugt/uge x, C), (fcmp olt/ole/ult/ule x, -C)
1533 // --> fabs(x) ogt/oge/ugt/uge C
1534 // TODO: Generalize to handle a negated variable operand?
1535 const APFloat *LHSC, *RHSC;
1536 if (LHS0 == RHS0 && LHS->hasOneUse() && RHS->hasOneUse() &&
1537 FCmpInst::getSwappedPredicate(pred: PredL) == PredR &&
1538 match(V: LHS1, P: m_APFloatAllowPoison(Res&: LHSC)) &&
1539 match(V: RHS1, P: m_APFloatAllowPoison(Res&: RHSC)) &&
1540 LHSC->bitwiseIsEqual(RHS: neg(X: *RHSC))) {
1541 auto IsLessThanOrLessEqual = [](FCmpInst::Predicate Pred) {
1542 switch (Pred) {
1543 case FCmpInst::FCMP_OLT:
1544 case FCmpInst::FCMP_OLE:
1545 case FCmpInst::FCMP_ULT:
1546 case FCmpInst::FCMP_ULE:
1547 return true;
1548 default:
1549 return false;
1550 }
1551 };
1552 if (IsLessThanOrLessEqual(IsAnd ? PredR : PredL)) {
1553 std::swap(a&: LHSC, b&: RHSC);
1554 std::swap(a&: PredL, b&: PredR);
1555 }
1556 if (IsLessThanOrLessEqual(IsAnd ? PredL : PredR)) {
1557 FastMathFlags NewFlag = LHS->getFastMathFlags();
1558 if (!IsLogicalSelect)
1559 NewFlag |= RHS->getFastMathFlags();
1560
1561 Value *FAbs = Builder.CreateFAbs(V: LHS0, FMFSource: NewFlag);
1562 return Builder.CreateFCmpFMF(
1563 P: PredL, LHS: FAbs, RHS: ConstantFP::get(Ty: LHS0->getType(), V: *LHSC), FMFSource: NewFlag);
1564 }
1565 }
1566
1567 return nullptr;
1568}
1569
1570/// Match an fcmp against a special value that performs a test possible by
1571/// llvm.is.fpclass.
1572static bool matchIsFPClassLikeFCmp(Value *Op, Value *&ClassVal,
1573 uint64_t &ClassMask) {
1574 auto *FCmp = dyn_cast<FCmpInst>(Val: Op);
1575 if (!FCmp || !FCmp->hasOneUse())
1576 return false;
1577
1578 std::tie(args&: ClassVal, args&: ClassMask) =
1579 fcmpToClassTest(Pred: FCmp->getPredicate(), F: *FCmp->getParent()->getParent(),
1580 LHS: FCmp->getOperand(i_nocapture: 0), RHS: FCmp->getOperand(i_nocapture: 1));
1581 return ClassVal != nullptr;
1582}
1583
1584/// or (is_fpclass x, mask0), (is_fpclass x, mask1)
1585/// -> is_fpclass x, (mask0 | mask1)
1586/// and (is_fpclass x, mask0), (is_fpclass x, mask1)
1587/// -> is_fpclass x, (mask0 & mask1)
1588/// xor (is_fpclass x, mask0), (is_fpclass x, mask1)
1589/// -> is_fpclass x, (mask0 ^ mask1)
1590Instruction *InstCombinerImpl::foldLogicOfIsFPClass(BinaryOperator &BO,
1591 Value *Op0, Value *Op1) {
1592 Value *ClassVal0 = nullptr;
1593 Value *ClassVal1 = nullptr;
1594 uint64_t ClassMask0, ClassMask1;
1595
1596 // Restrict to folding one fcmp into one is.fpclass for now, don't introduce a
1597 // new class.
1598 //
1599 // TODO: Support forming is.fpclass out of 2 separate fcmps when codegen is
1600 // better.
1601
1602 bool IsLHSClass =
1603 match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::is_fpclass>(
1604 Op0: m_Value(V&: ClassVal0), Op1: m_ConstantInt(V&: ClassMask0))));
1605 bool IsRHSClass =
1606 match(V: Op1, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::is_fpclass>(
1607 Op0: m_Value(V&: ClassVal1), Op1: m_ConstantInt(V&: ClassMask1))));
1608 if ((((IsLHSClass || matchIsFPClassLikeFCmp(Op: Op0, ClassVal&: ClassVal0, ClassMask&: ClassMask0)) &&
1609 (IsRHSClass || matchIsFPClassLikeFCmp(Op: Op1, ClassVal&: ClassVal1, ClassMask&: ClassMask1)))) &&
1610 ClassVal0 == ClassVal1) {
1611 unsigned NewClassMask;
1612 switch (BO.getOpcode()) {
1613 case Instruction::And:
1614 NewClassMask = ClassMask0 & ClassMask1;
1615 break;
1616 case Instruction::Or:
1617 NewClassMask = ClassMask0 | ClassMask1;
1618 break;
1619 case Instruction::Xor:
1620 NewClassMask = ClassMask0 ^ ClassMask1;
1621 break;
1622 default:
1623 llvm_unreachable("not a binary logic operator");
1624 }
1625
1626 if (IsLHSClass) {
1627 auto *II = cast<IntrinsicInst>(Val: Op0);
1628 II->setArgOperand(
1629 i: 1, v: ConstantInt::get(Ty: II->getArgOperand(i: 1)->getType(), V: NewClassMask));
1630 return replaceInstUsesWith(I&: BO, V: II);
1631 }
1632
1633 if (IsRHSClass) {
1634 auto *II = cast<IntrinsicInst>(Val: Op1);
1635 II->setArgOperand(
1636 i: 1, v: ConstantInt::get(Ty: II->getArgOperand(i: 1)->getType(), V: NewClassMask));
1637 return replaceInstUsesWith(I&: BO, V: II);
1638 }
1639
1640 Value *NewClass =
1641 Builder.CreateIntrinsic(ID: Intrinsic::is_fpclass, OverloadTypes: {ClassVal0->getType()},
1642 Args: {ClassVal0, Builder.getInt32(C: NewClassMask)});
1643 return replaceInstUsesWith(I&: BO, V: NewClass);
1644 }
1645
1646 return nullptr;
1647}
1648
1649/// Look for the pattern that conditionally negates a value via math operations:
1650/// cond.splat = sext i1 cond
1651/// sub = add cond.splat, x
1652/// xor = xor sub, cond.splat
1653/// and rewrite it to do the same, but via logical operations:
1654/// value.neg = sub 0, value
1655/// cond = select i1 neg, value.neg, value
1656Instruction *InstCombinerImpl::canonicalizeConditionalNegationViaMathToSelect(
1657 BinaryOperator &I) {
1658 assert(I.getOpcode() == BinaryOperator::Xor && "Only for xor!");
1659 Value *Cond, *X;
1660 // As per complexity ordering, `xor` is not commutative here.
1661 if (!match(V: &I, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Value()), R: m_Value())) ||
1662 !match(V: I.getOperand(i_nocapture: 1), P: m_SExt(Op: m_Value(V&: Cond))) ||
1663 !Cond->getType()->isIntOrIntVectorTy(BitWidth: 1) ||
1664 !match(V: I.getOperand(i_nocapture: 0), P: m_c_Add(L: m_SExt(Op: m_Specific(V: Cond)), R: m_Value(V&: X))))
1665 return nullptr;
1666 return createSelectInstWithUnknownProfile(
1667 C: Cond, S1: Builder.CreateNeg(V: X, Name: X->getName() + ".neg"), S2: X);
1668}
1669
1670/// This a limited reassociation for a special case (see above) where we are
1671/// checking if two values are either both NAN (unordered) or not-NAN (ordered).
1672/// This could be handled more generally in '-reassociation', but it seems like
1673/// an unlikely pattern for a large number of logic ops and fcmps.
1674static Instruction *reassociateFCmps(BinaryOperator &BO,
1675 InstCombiner::BuilderTy &Builder) {
1676 Instruction::BinaryOps Opcode = BO.getOpcode();
1677 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1678 "Expecting and/or op for fcmp transform");
1679
1680 // There are 4 commuted variants of the pattern. Canonicalize operands of this
1681 // logic op so an fcmp is operand 0 and a matching logic op is operand 1.
1682 Value *Op0 = BO.getOperand(i_nocapture: 0), *Op1 = BO.getOperand(i_nocapture: 1), *X;
1683 if (match(V: Op1, P: m_FCmp(L: m_Value(), R: m_AnyZeroFP())))
1684 std::swap(a&: Op0, b&: Op1);
1685
1686 // Match inner binop and the predicate for combining 2 NAN checks into 1.
1687 Value *BO10, *BO11;
1688 FCmpInst::Predicate NanPred = Opcode == Instruction::And ? FCmpInst::FCMP_ORD
1689 : FCmpInst::FCMP_UNO;
1690 if (!match(V: Op0, P: m_SpecificFCmp(MatchPred: NanPred, L: m_Value(V&: X), R: m_AnyZeroFP())) ||
1691 !match(V: Op1, P: m_BinOp(Opcode, L: m_Value(V&: BO10), R: m_Value(V&: BO11))))
1692 return nullptr;
1693
1694 // The inner logic op must have a matching fcmp operand.
1695 Value *Y;
1696 if (!match(V: BO10, P: m_SpecificFCmp(MatchPred: NanPred, L: m_Value(V&: Y), R: m_AnyZeroFP())) ||
1697 X->getType() != Y->getType())
1698 std::swap(a&: BO10, b&: BO11);
1699
1700 if (!match(V: BO10, P: m_SpecificFCmp(MatchPred: NanPred, L: m_Value(V&: Y), R: m_AnyZeroFP())) ||
1701 X->getType() != Y->getType())
1702 return nullptr;
1703
1704 // and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z
1705 // or (fcmp uno X, 0), (or (fcmp uno Y, 0), Z) --> or (fcmp uno X, Y), Z
1706 // Intersect FMF from the 2 source fcmps.
1707 Value *NewFCmp =
1708 Builder.CreateFCmpFMF(P: NanPred, LHS: X, RHS: Y, FMFSource: FMFSource::intersect(A: Op0, B: BO10));
1709 return BinaryOperator::Create(Op: Opcode, S1: NewFCmp, S2: BO11);
1710}
1711
1712/// Match variations of De Morgan's Laws:
1713/// (~A & ~B) == (~(A | B))
1714/// (~A | ~B) == (~(A & B))
1715static Instruction *matchDeMorgansLaws(BinaryOperator &I,
1716 InstCombiner &IC) {
1717 const Instruction::BinaryOps Opcode = I.getOpcode();
1718 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1719 "Trying to match De Morgan's Laws with something other than and/or");
1720
1721 // Flip the logic operation.
1722 const Instruction::BinaryOps FlippedOpcode =
1723 (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;
1724
1725 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
1726 Value *A, *B;
1727 if (match(V: Op0, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: A)))) &&
1728 match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: B)))) &&
1729 !IC.isFreeToInvert(V: A, WillInvertAllUses: A->hasOneUse()) &&
1730 !IC.isFreeToInvert(V: B, WillInvertAllUses: B->hasOneUse())) {
1731 Value *AndOr =
1732 IC.Builder.CreateBinOp(Opc: FlippedOpcode, LHS: A, RHS: B, Name: I.getName() + ".demorgan");
1733 return BinaryOperator::CreateNot(Op: AndOr);
1734 }
1735
1736 // The 'not' ops may require reassociation.
1737 // (A & ~B) & ~C --> A & ~(B | C)
1738 // (~B & A) & ~C --> A & ~(B | C)
1739 // (A | ~B) | ~C --> A | ~(B & C)
1740 // (~B | A) | ~C --> A | ~(B & C)
1741 Value *C;
1742 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_BinOp(Opcode, L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B))))) &&
1743 match(V: Op1, P: m_Not(V: m_Value(V&: C)))) {
1744 Value *FlippedBO = IC.Builder.CreateBinOp(Opc: FlippedOpcode, LHS: B, RHS: C);
1745 return BinaryOperator::Create(Op: Opcode, S1: A, S2: IC.Builder.CreateNot(V: FlippedBO));
1746 }
1747
1748 return nullptr;
1749}
1750
1751bool InstCombinerImpl::shouldOptimizeCast(CastInst *CI) {
1752 Value *CastSrc = CI->getOperand(i_nocapture: 0);
1753
1754 // Noop casts and casts of constants should be eliminated trivially.
1755 if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(Val: CastSrc))
1756 return false;
1757
1758 // If this cast is paired with another cast that can be eliminated, we prefer
1759 // to have it eliminated.
1760 if (const auto *PrecedingCI = dyn_cast<CastInst>(Val: CastSrc))
1761 if (isEliminableCastPair(CI1: PrecedingCI, CI2: CI))
1762 return false;
1763
1764 return true;
1765}
1766
1767/// Fold {and,or,xor} (cast X), C.
1768static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast,
1769 InstCombinerImpl &IC) {
1770 Constant *C = dyn_cast<Constant>(Val: Logic.getOperand(i_nocapture: 1));
1771 if (!C)
1772 return nullptr;
1773
1774 auto LogicOpc = Logic.getOpcode();
1775 Type *DestTy = Logic.getType();
1776 Type *SrcTy = Cast->getSrcTy();
1777
1778 // Move the logic operation ahead of a zext or sext if the constant is
1779 // unchanged in the smaller source type. Performing the logic in a smaller
1780 // type may provide more information to later folds, and the smaller logic
1781 // instruction may be cheaper (particularly in the case of vectors).
1782 Value *X;
1783 auto &DL = IC.getDataLayout();
1784 if (match(V: Cast, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))))) {
1785 PreservedCastFlags Flags;
1786 if (Constant *TruncC = getLosslessUnsignedTrunc(C, DestTy: SrcTy, DL, Flags: &Flags)) {
1787 // LogicOpc (zext X), C --> zext (LogicOpc X, C)
1788 Value *NewOp = IC.Builder.CreateBinOp(Opc: LogicOpc, LHS: X, RHS: TruncC);
1789 auto *ZExt = new ZExtInst(NewOp, DestTy);
1790 ZExt->setNonNeg(Flags.NNeg);
1791 ZExt->andIRFlags(V: Cast);
1792 return ZExt;
1793 }
1794 }
1795
1796 if (match(V: Cast, P: m_OneUse(SubPattern: m_SExtLike(Op: m_Value(V&: X))))) {
1797 if (Constant *TruncC = getLosslessSignedTrunc(C, DestTy: SrcTy, DL)) {
1798 // LogicOpc (sext X), C --> sext (LogicOpc X, C)
1799 Value *NewOp = IC.Builder.CreateBinOp(Opc: LogicOpc, LHS: X, RHS: TruncC);
1800 return new SExtInst(NewOp, DestTy);
1801 }
1802 }
1803
1804 return nullptr;
1805}
1806
1807/// Fold {and,or,xor} (cast X), Y.
1808Instruction *InstCombinerImpl::foldCastedBitwiseLogic(BinaryOperator &I) {
1809 auto LogicOpc = I.getOpcode();
1810 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding");
1811
1812 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
1813
1814 // fold bitwise(A >> BW - 1, zext(icmp)) (BW is the scalar bits of the
1815 // type of A)
1816 // -> bitwise(zext(A < 0), zext(icmp))
1817 // -> zext(bitwise(A < 0, icmp))
1818 auto FoldBitwiseICmpZeroWithICmp = [&](Value *Op0,
1819 Value *Op1) -> Instruction * {
1820 Value *A;
1821 bool IsMatched =
1822 match(V: Op0,
1823 P: m_OneUse(SubPattern: m_LShr(
1824 L: m_Value(V&: A),
1825 R: m_SpecificInt(V: Op0->getType()->getScalarSizeInBits() - 1)))) &&
1826 match(V: Op1, P: m_OneUse(SubPattern: m_ZExt(Op: m_ICmp(L: m_Value(), R: m_Value()))));
1827
1828 if (!IsMatched)
1829 return nullptr;
1830
1831 auto *ICmpL =
1832 Builder.CreateICmpSLT(LHS: A, RHS: Constant::getNullValue(Ty: A->getType()));
1833 auto *ICmpR = cast<ZExtInst>(Val: Op1)->getOperand(i_nocapture: 0);
1834 auto *BitwiseOp = Builder.CreateBinOp(Opc: LogicOpc, LHS: ICmpL, RHS: ICmpR);
1835
1836 return new ZExtInst(BitwiseOp, Op0->getType());
1837 };
1838
1839 if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op0, Op1))
1840 return Ret;
1841
1842 if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op1, Op0))
1843 return Ret;
1844
1845 CastInst *Cast0 = dyn_cast<CastInst>(Val: Op0);
1846 if (!Cast0)
1847 return nullptr;
1848
1849 // This must be a cast from an integer or integer vector source type to allow
1850 // transformation of the logic operation to the source type.
1851 Type *DestTy = I.getType();
1852 Type *SrcTy = Cast0->getSrcTy();
1853 if (!SrcTy->isIntOrIntVectorTy())
1854 return nullptr;
1855
1856 if (Instruction *Ret = foldLogicCastConstant(Logic&: I, Cast: Cast0, IC&: *this))
1857 return Ret;
1858
1859 CastInst *Cast1 = dyn_cast<CastInst>(Val: Op1);
1860 if (!Cast1)
1861 return nullptr;
1862
1863 // Both operands of the logic operation are casts. The casts must be the
1864 // same kind for reduction.
1865 Instruction::CastOps CastOpcode = Cast0->getOpcode();
1866 if (CastOpcode != Cast1->getOpcode())
1867 return nullptr;
1868
1869 // Can't fold it profitably if no one of casts has one use.
1870 if (!Cast0->hasOneUse() && !Cast1->hasOneUse())
1871 return nullptr;
1872
1873 Value *X, *Y;
1874 if (match(V: Cast0, P: m_ZExtOrSExt(Op: m_Value(V&: X))) &&
1875 match(V: Cast1, P: m_ZExtOrSExt(Op: m_Value(V&: Y)))) {
1876 // Cast the narrower source to the wider source type.
1877 unsigned XNumBits = X->getType()->getScalarSizeInBits();
1878 unsigned YNumBits = Y->getType()->getScalarSizeInBits();
1879 if (XNumBits != YNumBits) {
1880 // Cast the narrower source to the wider source type only if both of casts
1881 // have one use to avoid creating an extra instruction.
1882 if (!Cast0->hasOneUse() || !Cast1->hasOneUse())
1883 return nullptr;
1884
1885 // If the source types do not match, but the casts are matching extends,
1886 // we can still narrow the logic op.
1887 if (XNumBits < YNumBits) {
1888 X = Builder.CreateCast(Op: CastOpcode, V: X, DestTy: Y->getType());
1889 } else if (YNumBits < XNumBits) {
1890 Y = Builder.CreateCast(Op: CastOpcode, V: Y, DestTy: X->getType());
1891 }
1892 }
1893
1894 // Do the logic op in the intermediate width, then widen more.
1895 Value *NarrowLogic = Builder.CreateBinOp(Opc: LogicOpc, LHS: X, RHS: Y, Name: I.getName());
1896 auto *Disjoint = dyn_cast<PossiblyDisjointInst>(Val: &I);
1897 auto *NewDisjoint = dyn_cast<PossiblyDisjointInst>(Val: NarrowLogic);
1898 if (Disjoint && NewDisjoint)
1899 NewDisjoint->setIsDisjoint(Disjoint->isDisjoint());
1900 return CastInst::Create(CastOpcode, S: NarrowLogic, Ty: DestTy);
1901 }
1902
1903 // If the src type of casts are different, give up for other cast opcodes.
1904 if (SrcTy != Cast1->getSrcTy())
1905 return nullptr;
1906
1907 Value *Cast0Src = Cast0->getOperand(i_nocapture: 0);
1908 Value *Cast1Src = Cast1->getOperand(i_nocapture: 0);
1909
1910 // fold logic(cast(A), cast(B)) -> cast(logic(A, B))
1911 if (shouldOptimizeCast(CI: Cast0) && shouldOptimizeCast(CI: Cast1)) {
1912 Value *NewOp = Builder.CreateBinOp(Opc: LogicOpc, LHS: Cast0Src, RHS: Cast1Src,
1913 Name: I.getName());
1914 return CastInst::Create(CastOpcode, S: NewOp, Ty: DestTy);
1915 }
1916
1917 return nullptr;
1918}
1919
1920static Instruction *foldAndToXor(BinaryOperator &I,
1921 InstCombiner::BuilderTy &Builder) {
1922 assert(I.getOpcode() == Instruction::And);
1923 Value *Op0 = I.getOperand(i_nocapture: 0);
1924 Value *Op1 = I.getOperand(i_nocapture: 1);
1925 Value *A, *B;
1926
1927 // Operand complexity canonicalization guarantees that the 'or' is Op0.
1928 // (A | B) & ~(A & B) --> A ^ B
1929 // (A | B) & ~(B & A) --> A ^ B
1930 if (match(V: &I, P: m_BinOp(L: m_Or(L: m_Value(V&: A), R: m_Value(V&: B)),
1931 R: m_Not(V: m_c_And(L: m_Deferred(V: A), R: m_Deferred(V: B))))))
1932 return BinaryOperator::CreateXor(V1: A, V2: B);
1933
1934 // (A | ~B) & (~A | B) --> ~(A ^ B)
1935 // (A | ~B) & (B | ~A) --> ~(A ^ B)
1936 // (~B | A) & (~A | B) --> ~(A ^ B)
1937 // (~B | A) & (B | ~A) --> ~(A ^ B)
1938 if (Op0->hasOneUse() || Op1->hasOneUse())
1939 if (match(V: &I, P: m_BinOp(L: m_c_Or(L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B))),
1940 R: m_c_Or(L: m_Not(V: m_Deferred(V: A)), R: m_Deferred(V: B)))))
1941 return BinaryOperator::CreateNot(Op: Builder.CreateXor(LHS: A, RHS: B));
1942
1943 return nullptr;
1944}
1945
1946static Instruction *foldOrToXor(BinaryOperator &I,
1947 InstCombiner::BuilderTy &Builder) {
1948 assert(I.getOpcode() == Instruction::Or);
1949 Value *Op0 = I.getOperand(i_nocapture: 0);
1950 Value *Op1 = I.getOperand(i_nocapture: 1);
1951 Value *A, *B;
1952
1953 // Operand complexity canonicalization guarantees that the 'and' is Op0.
1954 // (A & B) | ~(A | B) --> ~(A ^ B)
1955 // (A & B) | ~(B | A) --> ~(A ^ B)
1956 if (Op0->hasOneUse() || Op1->hasOneUse())
1957 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
1958 match(V: Op1, P: m_Not(V: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B)))))
1959 return BinaryOperator::CreateNot(Op: Builder.CreateXor(LHS: A, RHS: B));
1960
1961 // Operand complexity canonicalization guarantees that the 'xor' is Op0.
1962 // (A ^ B) | ~(A | B) --> ~(A & B)
1963 // (A ^ B) | ~(B | A) --> ~(A & B)
1964 if (Op0->hasOneUse() || Op1->hasOneUse())
1965 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
1966 match(V: Op1, P: m_Not(V: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B)))))
1967 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: A, RHS: B));
1968
1969 // (A & ~B) | (~A & B) --> A ^ B
1970 // (A & ~B) | (B & ~A) --> A ^ B
1971 // (~B & A) | (~A & B) --> A ^ B
1972 // (~B & A) | (B & ~A) --> A ^ B
1973 if (match(V: Op0, P: m_c_And(L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B)))) &&
1974 match(V: Op1, P: m_c_And(L: m_Not(V: m_Specific(V: A)), R: m_Specific(V: B))))
1975 return BinaryOperator::CreateXor(V1: A, V2: B);
1976
1977 return nullptr;
1978}
1979
1980/// Return true if a constant shift amount is always less than the specified
1981/// bit-width. If not, the shift could create poison in the narrower type.
1982static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) {
1983 APInt Threshold(C->getType()->getScalarSizeInBits(), BitWidth);
1984 return match(V: C, P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_ULT, Threshold));
1985}
1986
1987/// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and
1988/// a common zext operand: and (binop (zext X), C), (zext X).
1989Instruction *InstCombinerImpl::narrowMaskedBinOp(BinaryOperator &And) {
1990 // This transform could also apply to {or, and, xor}, but there are better
1991 // folds for those cases, so we don't expect those patterns here. AShr is not
1992 // handled because it should always be transformed to LShr in this sequence.
1993 // The subtract transform is different because it has a constant on the left.
1994 // Add/mul commute the constant to RHS; sub with constant RHS becomes add.
1995 Value *Op0 = And.getOperand(i_nocapture: 0), *Op1 = And.getOperand(i_nocapture: 1);
1996 Constant *C;
1997 if (!match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Specific(V: Op1), R: m_Constant(C)))) &&
1998 !match(V: Op0, P: m_OneUse(SubPattern: m_Mul(L: m_Specific(V: Op1), R: m_Constant(C)))) &&
1999 !match(V: Op0, P: m_OneUse(SubPattern: m_LShr(L: m_Specific(V: Op1), R: m_Constant(C)))) &&
2000 !match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Specific(V: Op1), R: m_Constant(C)))) &&
2001 !match(V: Op0, P: m_OneUse(SubPattern: m_Sub(L: m_Constant(C), R: m_Specific(V: Op1)))))
2002 return nullptr;
2003
2004 Value *X;
2005 if (!match(V: Op1, P: m_ZExt(Op: m_Value(V&: X))) || Op1->hasNUsesOrMore(N: 3))
2006 return nullptr;
2007
2008 Type *Ty = And.getType();
2009 if (!isa<VectorType>(Val: Ty) && !shouldChangeType(From: Ty, To: X->getType()))
2010 return nullptr;
2011
2012 // If we're narrowing a shift, the shift amount must be safe (less than the
2013 // width) in the narrower type. If the shift amount is greater, instsimplify
2014 // usually handles that case, but we can't guarantee/assert it.
2015 Instruction::BinaryOps Opc = cast<BinaryOperator>(Val: Op0)->getOpcode();
2016 if (Opc == Instruction::LShr || Opc == Instruction::Shl)
2017 if (!canNarrowShiftAmt(C, BitWidth: X->getType()->getScalarSizeInBits()))
2018 return nullptr;
2019
2020 // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X)
2021 // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X)
2022 Value *NewC = ConstantExpr::getTrunc(C, Ty: X->getType());
2023 Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, LHS: NewC, RHS: X)
2024 : Builder.CreateBinOp(Opc, LHS: X, RHS: NewC);
2025 return new ZExtInst(Builder.CreateAnd(LHS: NewBO, RHS: X), Ty);
2026}
2027
2028/// Try folding relatively complex patterns for both And and Or operations
2029/// with all And and Or swapped.
2030static Instruction *foldComplexAndOrPatterns(BinaryOperator &I,
2031 InstCombiner::BuilderTy &Builder) {
2032 const Instruction::BinaryOps Opcode = I.getOpcode();
2033 assert(Opcode == Instruction::And || Opcode == Instruction::Or);
2034
2035 // Flip the logic operation.
2036 const Instruction::BinaryOps FlippedOpcode =
2037 (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;
2038
2039 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
2040 Value *A, *B, *C, *X, *Y, *Dummy;
2041
2042 // Match following expressions:
2043 // (~(A | B) & C)
2044 // (~(A & B) | C)
2045 // Captures X = ~(A | B) or ~(A & B)
2046 const auto matchNotOrAnd =
2047 [Opcode, FlippedOpcode](Value *Op, auto m_A, auto m_B, auto m_C,
2048 Value *&X, bool CountUses = false) -> bool {
2049 if (CountUses && !Op->hasOneUse())
2050 return false;
2051
2052 if (match(Op,
2053 m_c_BinOp(FlippedOpcode,
2054 m_Value(X, m_Not(m_c_BinOp(Opcode, m_A, m_B))), m_C)))
2055 return !CountUses || X->hasOneUse();
2056
2057 return false;
2058 };
2059
2060 // (~(A | B) & C) | ... --> ...
2061 // (~(A & B) | C) & ... --> ...
2062 // TODO: One use checks are conservative. We just need to check that a total
2063 // number of multiple used values does not exceed reduction
2064 // in operations.
2065 if (matchNotOrAnd(Op0, m_Value(V&: A), m_Value(V&: B), m_Value(V&: C), X)) {
2066 // (~(A | B) & C) | (~(A | C) & B) --> (B ^ C) & ~A
2067 // (~(A & B) | C) & (~(A & C) | B) --> ~((B ^ C) & A)
2068 if (matchNotOrAnd(Op1, m_Specific(V: A), m_Specific(V: C), m_Specific(V: B), Dummy,
2069 true)) {
2070 Value *Xor = Builder.CreateXor(LHS: B, RHS: C);
2071 return (Opcode == Instruction::Or)
2072 ? BinaryOperator::CreateAnd(V1: Xor, V2: Builder.CreateNot(V: A))
2073 : BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: Xor, RHS: A));
2074 }
2075
2076 // (~(A | B) & C) | (~(B | C) & A) --> (A ^ C) & ~B
2077 // (~(A & B) | C) & (~(B & C) | A) --> ~((A ^ C) & B)
2078 if (matchNotOrAnd(Op1, m_Specific(V: B), m_Specific(V: C), m_Specific(V: A), Dummy,
2079 true)) {
2080 Value *Xor = Builder.CreateXor(LHS: A, RHS: C);
2081 return (Opcode == Instruction::Or)
2082 ? BinaryOperator::CreateAnd(V1: Xor, V2: Builder.CreateNot(V: B))
2083 : BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: Xor, RHS: B));
2084 }
2085
2086 // (~(A | B) & C) | ~(A | C) --> ~((B & C) | A)
2087 // (~(A & B) | C) & ~(A & C) --> ~((B | C) & A)
2088 if (match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_OneUse(
2089 SubPattern: m_c_BinOp(Opcode, L: m_Specific(V: A), R: m_Specific(V: C)))))))
2090 return BinaryOperator::CreateNot(Op: Builder.CreateBinOp(
2091 Opc: Opcode, LHS: Builder.CreateBinOp(Opc: FlippedOpcode, LHS: B, RHS: C), RHS: A));
2092
2093 // (~(A | B) & C) | ~(B | C) --> ~((A & C) | B)
2094 // (~(A & B) | C) & ~(B & C) --> ~((A | C) & B)
2095 if (match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_OneUse(
2096 SubPattern: m_c_BinOp(Opcode, L: m_Specific(V: B), R: m_Specific(V: C)))))))
2097 return BinaryOperator::CreateNot(Op: Builder.CreateBinOp(
2098 Opc: Opcode, LHS: Builder.CreateBinOp(Opc: FlippedOpcode, LHS: A, RHS: C), RHS: B));
2099
2100 // (~(A | B) & C) | ~(C | (A ^ B)) --> ~((A | B) & (C | (A ^ B)))
2101 // Note, the pattern with swapped and/or is not handled because the
2102 // result is more undefined than a source:
2103 // (~(A & B) | C) & ~(C & (A ^ B)) --> (A ^ B ^ C) | ~(A | C) is invalid.
2104 if (Opcode == Instruction::Or && Op0->hasOneUse() &&
2105 match(V: Op1,
2106 P: m_OneUse(SubPattern: m_Not(V: m_Value(
2107 V&: Y, P: m_c_BinOp(Opcode, L: m_Specific(V: C),
2108 R: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B)))))))) {
2109 // X = ~(A | B)
2110 // Y = (C | (A ^ B)
2111 Value *Or = cast<BinaryOperator>(Val: X)->getOperand(i_nocapture: 0);
2112 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: Or, RHS: Y));
2113 }
2114 }
2115
2116 // (~A & B & C) | ... --> ...
2117 // (~A | B | C) | ... --> ...
2118 // TODO: One use checks are conservative. We just need to check that a total
2119 // number of multiple used values does not exceed reduction
2120 // in operations.
2121 if (match(V: Op0,
2122 P: m_OneUse(SubPattern: m_c_BinOp(Opcode: FlippedOpcode,
2123 L: m_BinOp(Opcode: FlippedOpcode, L: m_Value(V&: B), R: m_Value(V&: C)),
2124 R: m_Value(V&: X, P: m_Not(V: m_Value(V&: A)))))) ||
2125 match(V: Op0, P: m_OneUse(SubPattern: m_c_BinOp(Opcode: FlippedOpcode,
2126 L: m_c_BinOp(Opcode: FlippedOpcode, L: m_Value(V&: C),
2127 R: m_Value(V&: X, P: m_Not(V: m_Value(V&: A)))),
2128 R: m_Value(V&: B))))) {
2129 // X = ~A
2130 // (~A & B & C) | ~(A | B | C) --> ~(A | (B ^ C))
2131 // (~A | B | C) & ~(A & B & C) --> (~A | (B ^ C))
2132 if (match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_c_BinOp(
2133 Opcode, L: m_c_BinOp(Opcode, L: m_Specific(V: A), R: m_Specific(V: B)),
2134 R: m_Specific(V: C))))) ||
2135 match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_c_BinOp(
2136 Opcode, L: m_c_BinOp(Opcode, L: m_Specific(V: B), R: m_Specific(V: C)),
2137 R: m_Specific(V: A))))) ||
2138 match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_c_BinOp(
2139 Opcode, L: m_c_BinOp(Opcode, L: m_Specific(V: A), R: m_Specific(V: C)),
2140 R: m_Specific(V: B)))))) {
2141 Value *Xor = Builder.CreateXor(LHS: B, RHS: C);
2142 return (Opcode == Instruction::Or)
2143 ? BinaryOperator::CreateNot(Op: Builder.CreateOr(LHS: Xor, RHS: A))
2144 : BinaryOperator::CreateOr(V1: Xor, V2: X);
2145 }
2146
2147 // (~A & B & C) | ~(A | B) --> (C | ~B) & ~A
2148 // (~A | B | C) & ~(A & B) --> (C & ~B) | ~A
2149 if (match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_OneUse(
2150 SubPattern: m_c_BinOp(Opcode, L: m_Specific(V: A), R: m_Specific(V: B)))))))
2151 return BinaryOperator::Create(
2152 Op: FlippedOpcode, S1: Builder.CreateBinOp(Opc: Opcode, LHS: C, RHS: Builder.CreateNot(V: B)),
2153 S2: X);
2154
2155 // (~A & B & C) | ~(A | C) --> (B | ~C) & ~A
2156 // (~A | B | C) & ~(A & C) --> (B & ~C) | ~A
2157 if (match(V: Op1, P: m_OneUse(SubPattern: m_Not(V: m_OneUse(
2158 SubPattern: m_c_BinOp(Opcode, L: m_Specific(V: A), R: m_Specific(V: C)))))))
2159 return BinaryOperator::Create(
2160 Op: FlippedOpcode, S1: Builder.CreateBinOp(Opc: Opcode, LHS: B, RHS: Builder.CreateNot(V: C)),
2161 S2: X);
2162 }
2163
2164 return nullptr;
2165}
2166
2167/// Try to reassociate a pair of binops so that values with one use only are
2168/// part of the same instruction. This may enable folds that are limited with
2169/// multi-use restrictions and makes it more likely to match other patterns that
2170/// are looking for a common operand.
2171static Instruction *reassociateForUses(BinaryOperator &BO,
2172 InstCombinerImpl::BuilderTy &Builder) {
2173 Instruction::BinaryOps Opcode = BO.getOpcode();
2174 Value *X, *Y, *Z;
2175 if (match(V: &BO,
2176 P: m_c_BinOp(Opcode, L: m_OneUse(SubPattern: m_BinOp(Opcode, L: m_Value(V&: X), R: m_Value(V&: Y))),
2177 R: m_OneUse(SubPattern: m_Value(V&: Z))))) {
2178 if (!isa<Constant>(Val: X) && !isa<Constant>(Val: Y) && !isa<Constant>(Val: Z)) {
2179 // (X op Y) op Z --> (Y op Z) op X
2180 if (!X->hasOneUse()) {
2181 Value *YZ = Builder.CreateBinOp(Opc: Opcode, LHS: Y, RHS: Z);
2182 return BinaryOperator::Create(Op: Opcode, S1: YZ, S2: X);
2183 }
2184 // (X op Y) op Z --> (X op Z) op Y
2185 if (!Y->hasOneUse()) {
2186 Value *XZ = Builder.CreateBinOp(Opc: Opcode, LHS: X, RHS: Z);
2187 return BinaryOperator::Create(Op: Opcode, S1: XZ, S2: Y);
2188 }
2189 }
2190 }
2191
2192 return nullptr;
2193}
2194
2195// Match
2196// (X + C2) | C
2197// (X + C2) ^ C
2198// (X + C2) & C
2199// and convert to do the bitwise logic first:
2200// (X | C) + C2
2201// (X ^ C) + C2
2202// (X & C) + C2
2203// iff bits affected by logic op are lower than last bit affected by math op
2204static Instruction *canonicalizeLogicFirst(BinaryOperator &I,
2205 InstCombiner::BuilderTy &Builder) {
2206 Type *Ty = I.getType();
2207 Instruction::BinaryOps OpC = I.getOpcode();
2208 Value *Op0 = I.getOperand(i_nocapture: 0);
2209 Value *Op1 = I.getOperand(i_nocapture: 1);
2210 Value *X;
2211 const APInt *C, *C2;
2212
2213 if (!(match(V: Op0, P: m_OneUse(SubPattern: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C2)))) &&
2214 match(V: Op1, P: m_APInt(Res&: C))))
2215 return nullptr;
2216
2217 unsigned Width = Ty->getScalarSizeInBits();
2218 unsigned LastOneMath = Width - C2->countr_zero();
2219
2220 switch (OpC) {
2221 case Instruction::And:
2222 if (C->countl_one() < LastOneMath)
2223 return nullptr;
2224 break;
2225 case Instruction::Xor:
2226 case Instruction::Or:
2227 if (C->countl_zero() < LastOneMath)
2228 return nullptr;
2229 break;
2230 default:
2231 llvm_unreachable("Unexpected BinaryOp!");
2232 }
2233
2234 Value *NewBinOp = Builder.CreateBinOp(Opc: OpC, LHS: X, RHS: ConstantInt::get(Ty, V: *C));
2235 return BinaryOperator::CreateWithCopiedFlags(Opc: Instruction::Add, V1: NewBinOp,
2236 V2: ConstantInt::get(Ty, V: *C2), CopyO: Op0);
2237}
2238
2239// binop(shift(ShiftedC1, ShAmt), shift(ShiftedC2, add(ShAmt, AddC))) ->
2240// shift(binop(ShiftedC1, shift(ShiftedC2, AddC)), ShAmt)
2241// where both shifts are the same and AddC is a valid shift amount.
2242Instruction *InstCombinerImpl::foldBinOpOfDisplacedShifts(BinaryOperator &I) {
2243 assert((I.isBitwiseLogicOp() || I.getOpcode() == Instruction::Add) &&
2244 "Unexpected opcode");
2245
2246 Value *ShAmt;
2247 Constant *ShiftedC1, *ShiftedC2, *AddC;
2248 Type *Ty = I.getType();
2249 unsigned BitWidth = Ty->getScalarSizeInBits();
2250 if (!match(V: &I, P: m_c_BinOp(L: m_Shift(L: m_ImmConstant(C&: ShiftedC1), R: m_Value(V&: ShAmt)),
2251 R: m_Shift(L: m_ImmConstant(C&: ShiftedC2),
2252 R: m_AddLike(L: m_Deferred(V: ShAmt),
2253 R: m_ImmConstant(C&: AddC))))))
2254 return nullptr;
2255
2256 // Make sure the add constant is a valid shift amount.
2257 if (!match(V: AddC,
2258 P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_ULT, Threshold: APInt(BitWidth, BitWidth))))
2259 return nullptr;
2260
2261 // Avoid constant expressions.
2262 auto *Op0Inst = dyn_cast<Instruction>(Val: I.getOperand(i_nocapture: 0));
2263 auto *Op1Inst = dyn_cast<Instruction>(Val: I.getOperand(i_nocapture: 1));
2264 if (!Op0Inst || !Op1Inst)
2265 return nullptr;
2266
2267 // Both shifts must be the same.
2268 Instruction::BinaryOps ShiftOp =
2269 static_cast<Instruction::BinaryOps>(Op0Inst->getOpcode());
2270 if (ShiftOp != Op1Inst->getOpcode())
2271 return nullptr;
2272
2273 // For adds, only left shifts are supported.
2274 if (I.getOpcode() == Instruction::Add && ShiftOp != Instruction::Shl)
2275 return nullptr;
2276
2277 Value *NewC = Builder.CreateBinOp(
2278 Opc: I.getOpcode(), LHS: ShiftedC1, RHS: Builder.CreateBinOp(Opc: ShiftOp, LHS: ShiftedC2, RHS: AddC));
2279 return BinaryOperator::Create(Op: ShiftOp, S1: NewC, S2: ShAmt);
2280}
2281
2282// Fold and/or/xor with two equal intrinsic IDs:
2283// bitwise(fshl (A, B, ShAmt), fshl(C, D, ShAmt))
2284// -> fshl(bitwise(A, C), bitwise(B, D), ShAmt)
2285// bitwise(fshr (A, B, ShAmt), fshr(C, D, ShAmt))
2286// -> fshr(bitwise(A, C), bitwise(B, D), ShAmt)
2287// bitwise(bswap(A), bswap(B)) -> bswap(bitwise(A, B))
2288// bitwise(bswap(A), C) -> bswap(bitwise(A, bswap(C)))
2289// bitwise(bitreverse(A), bitreverse(B)) -> bitreverse(bitwise(A, B))
2290// bitwise(bitreverse(A), C) -> bitreverse(bitwise(A, bitreverse(C)))
2291static Instruction *
2292foldBitwiseLogicWithIntrinsics(BinaryOperator &I,
2293 InstCombiner::BuilderTy &Builder) {
2294 assert(I.isBitwiseLogicOp() && "Should and/or/xor");
2295 if (!I.getOperand(i_nocapture: 0)->hasOneUse())
2296 return nullptr;
2297 IntrinsicInst *X = dyn_cast<IntrinsicInst>(Val: I.getOperand(i_nocapture: 0));
2298 if (!X)
2299 return nullptr;
2300
2301 IntrinsicInst *Y = dyn_cast<IntrinsicInst>(Val: I.getOperand(i_nocapture: 1));
2302 if (Y && (!Y->hasOneUse() || X->getIntrinsicID() != Y->getIntrinsicID()))
2303 return nullptr;
2304
2305 Intrinsic::ID IID = X->getIntrinsicID();
2306 const APInt *RHSC;
2307 // Try to match constant RHS.
2308 if (!Y && (!(IID == Intrinsic::bswap || IID == Intrinsic::bitreverse) ||
2309 !match(V: I.getOperand(i_nocapture: 1), P: m_APInt(Res&: RHSC))))
2310 return nullptr;
2311
2312 switch (IID) {
2313 case Intrinsic::fshl:
2314 case Intrinsic::fshr: {
2315 if (X->getOperand(i_nocapture: 2) != Y->getOperand(i_nocapture: 2))
2316 return nullptr;
2317 Value *NewOp0 =
2318 Builder.CreateBinOp(Opc: I.getOpcode(), LHS: X->getOperand(i_nocapture: 0), RHS: Y->getOperand(i_nocapture: 0));
2319 Value *NewOp1 =
2320 Builder.CreateBinOp(Opc: I.getOpcode(), LHS: X->getOperand(i_nocapture: 1), RHS: Y->getOperand(i_nocapture: 1));
2321 Function *F =
2322 Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: IID, OverloadTys: I.getType());
2323 return CallInst::Create(Func: F, Args: {NewOp0, NewOp1, X->getOperand(i_nocapture: 2)});
2324 }
2325 case Intrinsic::bswap:
2326 case Intrinsic::bitreverse: {
2327 Value *NewOp0 = Builder.CreateBinOp(
2328 Opc: I.getOpcode(), LHS: X->getOperand(i_nocapture: 0),
2329 RHS: Y ? Y->getOperand(i_nocapture: 0)
2330 : ConstantInt::get(Ty: I.getType(), V: IID == Intrinsic::bswap
2331 ? RHSC->byteSwap()
2332 : RHSC->reverseBits()));
2333 Function *F =
2334 Intrinsic::getOrInsertDeclaration(M: I.getModule(), id: IID, OverloadTys: I.getType());
2335 return CallInst::Create(Func: F, Args: {NewOp0});
2336 }
2337 default:
2338 return nullptr;
2339 }
2340}
2341
2342// Try to simplify V by replacing occurrences of Op with RepOp, but only look
2343// through bitwise operations. In particular, for X | Y we try to replace Y with
2344// 0 inside X and for X & Y we try to replace Y with -1 inside X.
2345// Return the simplified result of X if successful, and nullptr otherwise.
2346// If SimplifyOnly is true, no new instructions will be created.
2347static Value *simplifyAndOrWithOpReplaced(Value *V, Value *Op, Value *RepOp,
2348 bool SimplifyOnly,
2349 InstCombinerImpl &IC,
2350 unsigned Depth = 0) {
2351 if (Op == RepOp)
2352 return nullptr;
2353
2354 if (V == Op)
2355 return RepOp;
2356
2357 auto *I = dyn_cast<BinaryOperator>(Val: V);
2358 if (!I || !I->isBitwiseLogicOp() || Depth >= 3)
2359 return nullptr;
2360
2361 if (!I->hasOneUse())
2362 SimplifyOnly = true;
2363
2364 Value *NewOp0 = simplifyAndOrWithOpReplaced(V: I->getOperand(i_nocapture: 0), Op, RepOp,
2365 SimplifyOnly, IC, Depth: Depth + 1);
2366 Value *NewOp1 = simplifyAndOrWithOpReplaced(V: I->getOperand(i_nocapture: 1), Op, RepOp,
2367 SimplifyOnly, IC, Depth: Depth + 1);
2368 if (!NewOp0 && !NewOp1)
2369 return nullptr;
2370
2371 if (!NewOp0)
2372 NewOp0 = I->getOperand(i_nocapture: 0);
2373 if (!NewOp1)
2374 NewOp1 = I->getOperand(i_nocapture: 1);
2375
2376 if (Value *Res = simplifyBinOp(Opcode: I->getOpcode(), LHS: NewOp0, RHS: NewOp1,
2377 Q: IC.getSimplifyQuery().getWithInstruction(I)))
2378 return Res;
2379
2380 if (SimplifyOnly)
2381 return nullptr;
2382 return IC.Builder.CreateBinOp(Opc: I->getOpcode(), LHS: NewOp0, RHS: NewOp1);
2383}
2384
2385/// Reassociate and/or expressions to see if we can fold the inner and/or ops.
2386/// TODO: Make this recursive; it's a little tricky because an arbitrary
2387/// number of and/or instructions might have to be created.
2388Value *InstCombinerImpl::reassociateBooleanAndOr(Value *LHS, Value *X, Value *Y,
2389 Instruction &I, bool IsAnd,
2390 bool RHSIsLogical) {
2391 Instruction::BinaryOps Opcode = IsAnd ? Instruction::And : Instruction::Or;
2392 Value *Folded = nullptr;
2393 // LHS bop (X lop Y) --> (LHS bop X) lop Y
2394 // LHS bop (X bop Y) --> (LHS bop X) bop Y
2395 if (Value *Res = foldBooleanAndOr(LHS, RHS: X, I, IsAnd, /*IsLogical=*/false))
2396 Folded = RHSIsLogical ? Builder.CreateLogicalOp(Opc: Opcode, Cond1: Res, Cond2: Y)
2397 : Builder.CreateBinOp(Opc: Opcode, LHS: Res, RHS: Y);
2398 // LHS bop (X bop Y) --> X bop (LHS bop Y)
2399 // LHS bop (X lop Y) --> X lop (LHS bop Y)
2400 else if (Value *Res = foldBooleanAndOr(LHS, RHS: Y, I, IsAnd, /*IsLogical=*/false))
2401 Folded = RHSIsLogical ? Builder.CreateLogicalOp(Opc: Opcode, Cond1: X, Cond2: Res)
2402 : Builder.CreateBinOp(Opc: Opcode, LHS: X, RHS: Res);
2403 if (SelectInst *SI = dyn_cast_or_null<SelectInst>(Val: Folded);
2404 SI != nullptr && !ProfcheckDisableMetadataFixes)
2405 // If the bop I was originally a lop, we could recover branch weight
2406 // information using that lop's weights. However, InstCombine usually
2407 // replaces the lop with a bop by the time we get here, deleting the branch
2408 // weight information. Therefore, we can only assume unknown branch weights.
2409 // TODO: see if it's possible to recover branch weight information from the
2410 // original lop (https://github.com/llvm/llvm-project/issues/183864).
2411 setExplicitlyUnknownBranchWeightsIfProfiled(I&: *SI, DEBUG_TYPE,
2412 F: I.getFunction());
2413 return Folded;
2414}
2415
2416// FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2417// here. We should standardize that construct where it is needed or choose some
2418// other way to ensure that commutated variants of patterns are not missed.
2419Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
2420 Type *Ty = I.getType();
2421
2422 if (Value *V = simplifyAndInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
2423 Q: SQ.getWithInstruction(I: &I)))
2424 return replaceInstUsesWith(I, V);
2425
2426 if (SimplifyAssociativeOrCommutative(I))
2427 return &I;
2428
2429 if (Instruction *X = foldVectorBinop(Inst&: I))
2430 return X;
2431
2432 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
2433 return Phi;
2434
2435 // See if we can simplify any instructions used by the instruction whose sole
2436 // purpose is to compute bits we don't care about.
2437 if (SimplifyDemandedInstructionBits(Inst&: I))
2438 return &I;
2439
2440 // Do this before using distributive laws to catch simple and/or/not patterns.
2441 if (Instruction *Xor = foldAndToXor(I, Builder))
2442 return Xor;
2443
2444 if (Instruction *X = foldComplexAndOrPatterns(I, Builder))
2445 return X;
2446
2447 // (A|B)&(A|C) -> A|(B&C) etc
2448 if (Value *V = foldUsingDistributiveLaws(I))
2449 return replaceInstUsesWith(I, V);
2450
2451 if (Instruction *R = foldBinOpShiftWithShift(I))
2452 return R;
2453
2454 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
2455
2456 Value *X, *Y;
2457 const APInt *C;
2458 if ((match(V: Op0, P: m_OneUse(SubPattern: m_LogicalShift(L: m_One(), R: m_Value(V&: X)))) ||
2459 (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_APInt(Res&: C), R: m_Value(V&: X)))) && (*C)[0])) &&
2460 match(V: Op1, P: m_One())) {
2461 // (1 >> X) & 1 --> zext(X == 0)
2462 // (C << X) & 1 --> zext(X == 0), when C is odd
2463 Value *IsZero = Builder.CreateICmpEQ(LHS: X, RHS: ConstantInt::get(Ty, V: 0));
2464 return new ZExtInst(IsZero, Ty);
2465 }
2466
2467 // (-(X & 1)) & Y --> (X & 1) == 0 ? 0 : Y
2468 Value *Neg;
2469 if (match(V: &I,
2470 P: m_c_And(L: m_Value(V&: Neg, P: m_OneUse(SubPattern: m_Neg(V: m_And(L: m_Value(), R: m_One())))),
2471 R: m_Value(V&: Y)))) {
2472 Value *Cmp = Builder.CreateIsNull(Arg: Neg);
2473 return createSelectInstWithUnknownProfile(C: Cmp,
2474 S1: ConstantInt::getNullValue(Ty), S2: Y);
2475 }
2476
2477 // Canonicalize:
2478 // (X +/- Y) & Y --> ~X & Y when Y is a power of 2.
2479 if (match(V: &I, P: m_c_And(L: m_Value(V&: Y), R: m_OneUse(SubPattern: m_CombineOr(
2480 Ps: m_c_Add(L: m_Value(V&: X), R: m_Deferred(V: Y)),
2481 Ps: m_Sub(L: m_Value(V&: X), R: m_Deferred(V: Y)))))) &&
2482 isKnownToBeAPowerOfTwo(V: Y, /*OrZero*/ true, CxtI: &I))
2483 return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: X), V2: Y);
2484
2485 if (match(V: Op1, P: m_APInt(Res&: C))) {
2486 const APInt *XorC;
2487 if (match(V: Op0, P: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: X), R: m_APInt(Res&: XorC))))) {
2488 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2489 Constant *NewC = ConstantInt::get(Ty, V: *C & *XorC);
2490 Value *And = Builder.CreateAnd(LHS: X, RHS: Op1);
2491 And->takeName(V: Op0);
2492 return BinaryOperator::CreateXor(V1: And, V2: NewC);
2493 }
2494
2495 const APInt *OrC;
2496 if (match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_APInt(Res&: OrC))))) {
2497 // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2)
2498 // NOTE: This reduces the number of bits set in the & mask, which
2499 // can expose opportunities for store narrowing for scalars.
2500 // NOTE: SimplifyDemandedBits should have already removed bits from C1
2501 // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in
2502 // above, but this feels safer.
2503 APInt Together = *C & *OrC;
2504 Value *And = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty, V: Together ^ *C));
2505 And->takeName(V: Op0);
2506 return BinaryOperator::CreateOr(V1: And, V2: ConstantInt::get(Ty, V: Together));
2507 }
2508
2509 unsigned Width = Ty->getScalarSizeInBits();
2510 const APInt *ShiftC;
2511 if (match(V: Op0, P: m_OneUse(SubPattern: m_SExt(Op: m_AShr(L: m_Value(V&: X), R: m_APInt(Res&: ShiftC))))) &&
2512 ShiftC->ult(RHS: Width)) {
2513 if (*C == APInt::getLowBitsSet(numBits: Width, loBitsSet: Width - ShiftC->getZExtValue())) {
2514 // We are clearing high bits that were potentially set by sext+ashr:
2515 // and (sext (ashr X, ShiftC)), C --> lshr (sext X), ShiftC
2516 Value *Sext = Builder.CreateSExt(V: X, DestTy: Ty);
2517 Constant *ShAmtC = ConstantInt::get(Ty, V: ShiftC->zext(width: Width));
2518 return BinaryOperator::CreateLShr(V1: Sext, V2: ShAmtC);
2519 }
2520 }
2521
2522 // If this 'and' clears the sign-bits added by ashr, replace with lshr:
2523 // and (ashr X, ShiftC), C --> lshr X, ShiftC
2524 if (match(V: Op0, P: m_AShr(L: m_Value(V&: X), R: m_APInt(Res&: ShiftC))) && ShiftC->ult(RHS: Width) &&
2525 C->isMask(numBits: Width - ShiftC->getZExtValue()))
2526 return BinaryOperator::CreateLShr(V1: X, V2: ConstantInt::get(Ty, V: *ShiftC));
2527
2528 const APInt *AddC;
2529 if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: AddC)))) {
2530 // If we are masking the result of the add down to exactly one bit and
2531 // the constant we are adding has no bits set below that bit, then the
2532 // add is flipping a single bit. Example:
2533 // (X + 4) & 4 --> (X & 4) ^ 4
2534 if (Op0->hasOneUse() && C->isPowerOf2() && (*AddC & (*C - 1)) == 0) {
2535 assert((*C & *AddC) != 0 && "Expected common bit");
2536 Value *NewAnd = Builder.CreateAnd(LHS: X, RHS: Op1);
2537 return BinaryOperator::CreateXor(V1: NewAnd, V2: Op1);
2538 }
2539 }
2540
2541 // ((C1 OP zext(X)) & C2) -> zext((C1 OP X) & C2) if C2 fits in the
2542 // bitwidth of X and OP behaves well when given trunc(C1) and X.
2543 auto isNarrowableBinOpcode = [](BinaryOperator *B) {
2544 switch (B->getOpcode()) {
2545 case Instruction::Xor:
2546 case Instruction::Or:
2547 case Instruction::Mul:
2548 case Instruction::Add:
2549 case Instruction::Sub:
2550 return true;
2551 default:
2552 return false;
2553 }
2554 };
2555 BinaryOperator *BO;
2556 if (match(V: Op0, P: m_OneUse(SubPattern: m_BinOp(I&: BO))) && isNarrowableBinOpcode(BO)) {
2557 Instruction::BinaryOps BOpcode = BO->getOpcode();
2558 Value *X;
2559 const APInt *C1;
2560 // TODO: The one-use restrictions could be relaxed a little if the AND
2561 // is going to be removed.
2562 // Try to narrow the 'and' and a binop with constant operand:
2563 // and (bo (zext X), C1), C --> zext (and (bo X, TruncC1), TruncC)
2564 if (match(V: BO, P: m_c_BinOp(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))), R: m_APInt(Res&: C1))) &&
2565 C->isIntN(N: X->getType()->getScalarSizeInBits())) {
2566 unsigned XWidth = X->getType()->getScalarSizeInBits();
2567 Constant *TruncC1 = ConstantInt::get(Ty: X->getType(), V: C1->trunc(width: XWidth));
2568 Value *BinOp = isa<ZExtInst>(Val: BO->getOperand(i_nocapture: 0))
2569 ? Builder.CreateBinOp(Opc: BOpcode, LHS: X, RHS: TruncC1)
2570 : Builder.CreateBinOp(Opc: BOpcode, LHS: TruncC1, RHS: X);
2571 Constant *TruncC = ConstantInt::get(Ty: X->getType(), V: C->trunc(width: XWidth));
2572 Value *And = Builder.CreateAnd(LHS: BinOp, RHS: TruncC);
2573 return new ZExtInst(And, Ty);
2574 }
2575
2576 // Similar to above: if the mask matches the zext input width, then the
2577 // 'and' can be eliminated, so we can truncate the other variable op:
2578 // and (bo (zext X), Y), C --> zext (bo X, (trunc Y))
2579 if (isa<Instruction>(Val: BO->getOperand(i_nocapture: 0)) &&
2580 match(V: BO->getOperand(i_nocapture: 0), P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) &&
2581 C->isMask(numBits: X->getType()->getScalarSizeInBits())) {
2582 Y = BO->getOperand(i_nocapture: 1);
2583 Value *TrY = Builder.CreateTrunc(V: Y, DestTy: X->getType(), Name: Y->getName() + ".tr");
2584 Value *NewBO =
2585 Builder.CreateBinOp(Opc: BOpcode, LHS: X, RHS: TrY, Name: BO->getName() + ".narrow");
2586 return new ZExtInst(NewBO, Ty);
2587 }
2588 // and (bo Y, (zext X)), C --> zext (bo (trunc Y), X)
2589 if (isa<Instruction>(Val: BO->getOperand(i_nocapture: 1)) &&
2590 match(V: BO->getOperand(i_nocapture: 1), P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) &&
2591 C->isMask(numBits: X->getType()->getScalarSizeInBits())) {
2592 Y = BO->getOperand(i_nocapture: 0);
2593 Value *TrY = Builder.CreateTrunc(V: Y, DestTy: X->getType(), Name: Y->getName() + ".tr");
2594 Value *NewBO =
2595 Builder.CreateBinOp(Opc: BOpcode, LHS: TrY, RHS: X, Name: BO->getName() + ".narrow");
2596 return new ZExtInst(NewBO, Ty);
2597 }
2598 }
2599
2600 // This is intentionally placed after the narrowing transforms for
2601 // efficiency (transform directly to the narrow logic op if possible).
2602 // If the mask is only needed on one incoming arm, push the 'and' op up.
2603 if (match(V: Op0, P: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: X), R: m_Value(V&: Y)))) ||
2604 match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
2605 APInt NotAndMask(~(*C));
2606 BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Val: Op0)->getOpcode();
2607 if (MaskedValueIsZero(V: X, Mask: NotAndMask, CxtI: &I)) {
2608 // Not masking anything out for the LHS, move mask to RHS.
2609 // and ({x}or X, Y), C --> {x}or X, (and Y, C)
2610 Value *NewRHS = Builder.CreateAnd(LHS: Y, RHS: Op1, Name: Y->getName() + ".masked");
2611 return BinaryOperator::Create(Op: BinOp, S1: X, S2: NewRHS);
2612 }
2613 if (!isa<Constant>(Val: Y) && MaskedValueIsZero(V: Y, Mask: NotAndMask, CxtI: &I)) {
2614 // Not masking anything out for the RHS, move mask to LHS.
2615 // and ({x}or X, Y), C --> {x}or (and X, C), Y
2616 Value *NewLHS = Builder.CreateAnd(LHS: X, RHS: Op1, Name: X->getName() + ".masked");
2617 return BinaryOperator::Create(Op: BinOp, S1: NewLHS, S2: Y);
2618 }
2619 }
2620
2621 // When the mask is a power-of-2 constant and op0 is a shifted-power-of-2
2622 // constant, test if the shift amount equals the offset bit index:
2623 // (ShiftC << X) & C --> X == (log2(C) - log2(ShiftC)) ? C : 0
2624 // (ShiftC >> X) & C --> X == (log2(ShiftC) - log2(C)) ? C : 0
2625 if (C->isPowerOf2() &&
2626 match(V: Op0, P: m_OneUse(SubPattern: m_LogicalShift(L: m_Power2(V&: ShiftC), R: m_Value(V&: X))))) {
2627 int Log2ShiftC = ShiftC->exactLogBase2();
2628 int Log2C = C->exactLogBase2();
2629 bool IsShiftLeft =
2630 cast<BinaryOperator>(Val: Op0)->getOpcode() == Instruction::Shl;
2631 int BitNum = IsShiftLeft ? Log2C - Log2ShiftC : Log2ShiftC - Log2C;
2632 assert(BitNum >= 0 && "Expected demanded bits to handle impossible mask");
2633 Value *Cmp = Builder.CreateICmpEQ(LHS: X, RHS: ConstantInt::get(Ty, V: BitNum));
2634 return createSelectInstWithUnknownProfile(C: Cmp, S1: ConstantInt::get(Ty, V: *C),
2635 S2: ConstantInt::getNullValue(Ty));
2636 }
2637
2638 Constant *C1, *C2;
2639 const APInt *C3 = C;
2640 Value *X;
2641 if (C3->isPowerOf2()) {
2642 Constant *Log2C3 = ConstantInt::get(Ty, V: C3->countr_zero());
2643 if (match(V: Op0, P: m_OneUse(SubPattern: m_LShr(L: m_Shl(L: m_ImmConstant(C&: C1), R: m_Value(V&: X)),
2644 R: m_ImmConstant(C&: C2)))) &&
2645 match(V: C1, P: m_Power2())) {
2646 Constant *Log2C1 = ConstantExpr::getExactLogBase2(C: C1);
2647 Constant *LshrC = ConstantExpr::getAdd(C1: C2, C2: Log2C3);
2648 KnownBits KnownLShrc = computeKnownBits(V: LshrC, CxtI: nullptr);
2649 if (KnownLShrc.getMaxValue().ult(RHS: Width)) {
2650 // iff C1,C3 is pow2 and C2 + cttz(C3) < BitWidth:
2651 // ((C1 << X) >> C2) & C3 -> X == (cttz(C3)+C2-cttz(C1)) ? C3 : 0
2652 Constant *CmpC = ConstantExpr::getSub(C1: LshrC, C2: Log2C1);
2653 Value *Cmp = Builder.CreateICmpEQ(LHS: X, RHS: CmpC);
2654 return createSelectInstWithUnknownProfile(
2655 C: Cmp, S1: ConstantInt::get(Ty, V: *C3), S2: ConstantInt::getNullValue(Ty));
2656 }
2657 }
2658
2659 if (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_LShr(L: m_ImmConstant(C&: C1), R: m_Value(V&: X)),
2660 R: m_ImmConstant(C&: C2)))) &&
2661 match(V: C1, P: m_Power2())) {
2662 Constant *Log2C1 = ConstantExpr::getExactLogBase2(C: C1);
2663 Constant *Cmp =
2664 ConstantFoldCompareInstOperands(Predicate: ICmpInst::ICMP_ULT, LHS: Log2C3, RHS: C2, DL);
2665 if (Cmp && Cmp->isNullValue()) {
2666 // iff C1,C3 is pow2 and Log2(C3) >= C2:
2667 // ((C1 >> X) << C2) & C3 -> X == (cttz(C1)+C2-cttz(C3)) ? C3 : 0
2668 Constant *ShlC = ConstantExpr::getAdd(C1: C2, C2: Log2C1);
2669 Constant *CmpC = ConstantExpr::getSub(C1: ShlC, C2: Log2C3);
2670 Value *Cmp = Builder.CreateICmpEQ(LHS: X, RHS: CmpC);
2671 return createSelectInstWithUnknownProfile(
2672 C: Cmp, S1: ConstantInt::get(Ty, V: *C3), S2: ConstantInt::getNullValue(Ty));
2673 }
2674 }
2675 }
2676 }
2677
2678 // If we are clearing the sign bit of a floating-point value, convert this to
2679 // fabs, then cast back to integer.
2680 //
2681 // This is a generous interpretation for noimplicitfloat, this is not a true
2682 // floating-point operation.
2683 //
2684 // Assumes any IEEE-represented type has the sign bit in the high bit.
2685 // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt
2686 Value *CastOp;
2687 if (match(V: Op0, P: m_ElementWiseBitCast(Op: m_Value(V&: CastOp))) &&
2688 match(V: Op1, P: m_MaxSignedValue()) &&
2689 !Builder.GetInsertBlock()->getParent()->hasFnAttribute(
2690 Kind: Attribute::NoImplicitFloat)) {
2691 Type *EltTy = CastOp->getType()->getScalarType();
2692 if (EltTy->isFloatingPointTy() &&
2693 APFloat::hasSignBitInMSB(EltTy->getFltSemantics())) {
2694 Value *FAbs = Builder.CreateFAbs(V: CastOp);
2695 return new BitCastInst(FAbs, I.getType());
2696 }
2697 }
2698
2699 // and(shl(zext(X), Y), SignMask) -> and(sext(X), SignMask)
2700 // where Y is a valid shift amount.
2701 if (match(V: &I, P: m_And(L: m_OneUse(SubPattern: m_Shl(L: m_ZExt(Op: m_Value(V&: X)), R: m_Value(V&: Y))),
2702 R: m_SignMask())) &&
2703 match(V: Y, P: m_SpecificInt_ICMP(
2704 Predicate: ICmpInst::Predicate::ICMP_EQ,
2705 Threshold: APInt(Ty->getScalarSizeInBits(),
2706 Ty->getScalarSizeInBits() -
2707 X->getType()->getScalarSizeInBits())))) {
2708 auto *SExt = Builder.CreateSExt(V: X, DestTy: Ty, Name: X->getName() + ".signext");
2709 return BinaryOperator::CreateAnd(V1: SExt, V2: Op1);
2710 }
2711
2712 if (Instruction *Z = narrowMaskedBinOp(And&: I))
2713 return Z;
2714
2715 if (I.getType()->isIntOrIntVectorTy(BitWidth: 1)) {
2716 if (auto *SI0 = dyn_cast<SelectInst>(Val: Op0)) {
2717 if (auto *R =
2718 foldAndOrOfSelectUsingImpliedCond(Op: Op1, SI&: *SI0, /* IsAnd */ true))
2719 return R;
2720 }
2721 if (auto *SI1 = dyn_cast<SelectInst>(Val: Op1)) {
2722 if (auto *R =
2723 foldAndOrOfSelectUsingImpliedCond(Op: Op0, SI&: *SI1, /* IsAnd */ true))
2724 return R;
2725 }
2726 }
2727
2728 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
2729 return FoldedLogic;
2730
2731 if (Instruction *DeMorgan = matchDeMorgansLaws(I, IC&: *this))
2732 return DeMorgan;
2733
2734 {
2735 Value *A, *B, *C;
2736 // A & ~(A ^ B) --> A & B
2737 if (match(V: Op1, P: m_Not(V: m_c_Xor(L: m_Specific(V: Op0), R: m_Value(V&: B)))))
2738 return BinaryOperator::CreateAnd(V1: Op0, V2: B);
2739 // ~(A ^ B) & A --> A & B
2740 if (match(V: Op0, P: m_Not(V: m_c_Xor(L: m_Specific(V: Op1), R: m_Value(V&: B)))))
2741 return BinaryOperator::CreateAnd(V1: Op1, V2: B);
2742
2743 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
2744 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
2745 match(V: Op1, P: m_Xor(L: m_Xor(L: m_Specific(V: B), R: m_Value(V&: C)), R: m_Specific(V: A)))) {
2746 Value *NotC = Op1->hasOneUse()
2747 ? Builder.CreateNot(V: C)
2748 : getFreelyInverted(V: C, WillInvertAllUses: C->hasOneUse(), Builder: &Builder);
2749 if (NotC != nullptr)
2750 return BinaryOperator::CreateAnd(V1: Op0, V2: NotC);
2751 }
2752
2753 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
2754 if (match(V: Op0, P: m_Xor(L: m_Xor(L: m_Value(V&: A), R: m_Value(V&: C)), R: m_Value(V&: B))) &&
2755 match(V: Op1, P: m_Xor(L: m_Specific(V: B), R: m_Specific(V: A)))) {
2756 Value *NotC = Op0->hasOneUse()
2757 ? Builder.CreateNot(V: C)
2758 : getFreelyInverted(V: C, WillInvertAllUses: C->hasOneUse(), Builder: &Builder);
2759 if (NotC != nullptr)
2760 return BinaryOperator::CreateAnd(V1: Op1, V2: NotC);
2761 }
2762
2763 // (A | B) & (~A ^ B) -> A & B
2764 // (A | B) & (B ^ ~A) -> A & B
2765 // (B | A) & (~A ^ B) -> A & B
2766 // (B | A) & (B ^ ~A) -> A & B
2767 if (match(V: Op1, P: m_c_Xor(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2768 match(V: Op0, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2769 return BinaryOperator::CreateAnd(V1: A, V2: B);
2770
2771 // (~A ^ B) & (A | B) -> A & B
2772 // (~A ^ B) & (B | A) -> A & B
2773 // (B ^ ~A) & (A | B) -> A & B
2774 // (B ^ ~A) & (B | A) -> A & B
2775 if (match(V: Op0, P: m_c_Xor(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2776 match(V: Op1, P: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))
2777 return BinaryOperator::CreateAnd(V1: A, V2: B);
2778
2779 // (~A | B) & (A ^ B) -> ~A & B
2780 // (~A | B) & (B ^ A) -> ~A & B
2781 // (B | ~A) & (A ^ B) -> ~A & B
2782 // (B | ~A) & (B ^ A) -> ~A & B
2783 if (match(V: Op0, P: m_c_Or(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2784 match(V: Op1, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2785 return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: A), V2: B);
2786
2787 // (A ^ B) & (~A | B) -> ~A & B
2788 // (B ^ A) & (~A | B) -> ~A & B
2789 // (A ^ B) & (B | ~A) -> ~A & B
2790 // (B ^ A) & (B | ~A) -> ~A & B
2791 if (match(V: Op1, P: m_c_Or(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B))) &&
2792 match(V: Op0, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
2793 return BinaryOperator::CreateAnd(V1: Builder.CreateNot(V: A), V2: B);
2794 }
2795
2796 if (Value *Res =
2797 foldBooleanAndOr(LHS: Op0, RHS: Op1, I, /*IsAnd=*/true, /*IsLogical=*/false))
2798 return replaceInstUsesWith(I, V: Res);
2799
2800 if (match(V: Op1, P: m_OneUse(SubPattern: m_LogicalAnd(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
2801 bool IsLogical = isa<SelectInst>(Val: Op1);
2802 if (auto *V = reassociateBooleanAndOr(LHS: Op0, X, Y, I, /*IsAnd=*/true,
2803 /*RHSIsLogical=*/IsLogical))
2804 return replaceInstUsesWith(I, V);
2805 }
2806 if (match(V: Op0, P: m_OneUse(SubPattern: m_LogicalAnd(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
2807 bool IsLogical = isa<SelectInst>(Val: Op0);
2808 if (auto *V = reassociateBooleanAndOr(LHS: Op1, X, Y, I, /*IsAnd=*/true,
2809 /*RHSIsLogical=*/IsLogical))
2810 return replaceInstUsesWith(I, V);
2811 }
2812
2813 if (Instruction *FoldedFCmps = reassociateFCmps(BO&: I, Builder))
2814 return FoldedFCmps;
2815
2816 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))
2817 return CastedAnd;
2818
2819 if (Instruction *Sel = foldBinopOfSextBoolToSelect(I))
2820 return Sel;
2821
2822 // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>.
2823 // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold
2824 // with binop identity constant. But creating a select with non-constant
2825 // arm may not be reversible due to poison semantics. Is that a good
2826 // canonicalization?
2827 Value *A, *B;
2828 if (match(V: &I, P: m_c_And(L: m_SExt(Op: m_Value(V&: A)), R: m_Value(V&: B))) &&
2829 A->getType()->isIntOrIntVectorTy(BitWidth: 1))
2830 return createSelectInstWithUnknownProfile(C: A, S1: B, S2: Constant::getNullValue(Ty));
2831
2832 // Similarly, a 'not' of the bool translates to a swap of the select arms:
2833 // ~sext(A) & B / B & ~sext(A) --> A ? 0 : B
2834 if (match(V: &I, P: m_c_And(L: m_Not(V: m_SExt(Op: m_Value(V&: A))), R: m_Value(V&: B))) &&
2835 A->getType()->isIntOrIntVectorTy(BitWidth: 1))
2836 return createSelectInstWithUnknownProfile(C: A, S1: Constant::getNullValue(Ty), S2: B);
2837
2838 // and(zext(A), B) -> A ? (B & 1) : 0
2839 if (match(V: &I, P: m_c_And(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: A))), R: m_Value(V&: B))) &&
2840 A->getType()->isIntOrIntVectorTy(BitWidth: 1))
2841 return createSelectInstWithUnknownProfile(
2842 C: A, S1: Builder.CreateAnd(LHS: B, RHS: ConstantInt::get(Ty, V: 1)),
2843 S2: Constant::getNullValue(Ty));
2844
2845 // (-1 + A) & B --> A ? 0 : B where A is 0/1.
2846 if (match(V: &I, P: m_c_And(L: m_OneUse(SubPattern: m_Add(L: m_ZExtOrSelf(Op: m_Value(V&: A)), R: m_AllOnes())),
2847 R: m_Value(V&: B)))) {
2848 if (A->getType()->isIntOrIntVectorTy(BitWidth: 1))
2849 return createSelectInstWithUnknownProfile(C: A, S1: Constant::getNullValue(Ty),
2850 S2: B);
2851 if (computeKnownBits(V: A, CxtI: &I).countMaxActiveBits() <= 1) {
2852 return createSelectInstWithUnknownProfile(
2853 C: Builder.CreateICmpEQ(LHS: A, RHS: Constant::getNullValue(Ty: A->getType())), S1: B,
2854 S2: Constant::getNullValue(Ty));
2855 }
2856 }
2857
2858 // (iN X s>> (N-1)) & Y --> (X s< 0) ? Y : 0 -- with optional sext
2859 if (match(V: &I, P: m_c_And(L: m_OneUse(SubPattern: m_SExtOrSelf(
2860 Op: m_AShr(L: m_Value(V&: X), R: m_APIntAllowPoison(Res&: C)))),
2861 R: m_Value(V&: Y))) &&
2862 *C == X->getType()->getScalarSizeInBits() - 1) {
2863 Value *IsNeg = Builder.CreateIsNeg(Arg: X, Name: "isneg");
2864 return createSelectInstWithUnknownProfile(C: IsNeg, S1: Y,
2865 S2: ConstantInt::getNullValue(Ty));
2866 }
2867 // If there's a 'not' of the shifted value, swap the select operands:
2868 // ~(iN X s>> (N-1)) & Y --> (X s< 0) ? 0 : Y -- with optional sext
2869 if (match(V: &I, P: m_c_And(L: m_OneUse(SubPattern: m_SExtOrSelf(
2870 Op: m_Not(V: m_AShr(L: m_Value(V&: X), R: m_APIntAllowPoison(Res&: C))))),
2871 R: m_Value(V&: Y))) &&
2872 *C == X->getType()->getScalarSizeInBits() - 1) {
2873 Value *IsNeg = Builder.CreateIsNeg(Arg: X, Name: "isneg");
2874 return createSelectInstWithUnknownProfile(C: IsNeg,
2875 S1: ConstantInt::getNullValue(Ty), S2: Y);
2876 }
2877
2878 // (~x) & y --> ~(x | (~y)) iff that gets rid of inversions
2879 if (sinkNotIntoOtherHandOfLogicalOp(I))
2880 return &I;
2881
2882 // An and recurrence w/loop invariant step is equivelent to (and start, step)
2883 PHINode *PN = nullptr;
2884 Value *Start = nullptr, *Step = nullptr;
2885 if (matchSimpleRecurrence(I: &I, P&: PN, Start, Step) && DT.dominates(Def: Step, User: PN))
2886 return replaceInstUsesWith(I, V: Builder.CreateAnd(LHS: Start, RHS: Step));
2887
2888 if (Instruction *R = reassociateForUses(BO&: I, Builder))
2889 return R;
2890
2891 if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))
2892 return Canonicalized;
2893
2894 if (Instruction *Folded = foldLogicOfIsFPClass(BO&: I, Op0, Op1))
2895 return Folded;
2896
2897 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
2898 return Res;
2899
2900 if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))
2901 return Res;
2902
2903 if (Value *V =
2904 simplifyAndOrWithOpReplaced(V: Op0, Op: Op1, RepOp: Constant::getAllOnesValue(Ty),
2905 /*SimplifyOnly*/ false, IC&: *this))
2906 return BinaryOperator::CreateAnd(V1: V, V2: Op1);
2907 if (Value *V =
2908 simplifyAndOrWithOpReplaced(V: Op1, Op: Op0, RepOp: Constant::getAllOnesValue(Ty),
2909 /*SimplifyOnly*/ false, IC&: *this))
2910 return BinaryOperator::CreateAnd(V1: Op0, V2: V);
2911
2912 return nullptr;
2913}
2914
2915Instruction *InstCombinerImpl::matchBSwapOrBitReverse(Instruction &I,
2916 bool MatchBSwaps,
2917 bool MatchBitReversals) {
2918 SmallVector<Instruction *, 4> Insts;
2919 if (!recognizeBSwapOrBitReverseIdiom(I: &I, MatchBSwaps, MatchBitReversals,
2920 InsertedInsts&: Insts))
2921 return nullptr;
2922 Instruction *LastInst = Insts.pop_back_val();
2923 LastInst->removeFromParent();
2924
2925 for (auto *Inst : Insts) {
2926 Inst->setDebugLoc(I.getDebugLoc());
2927 Worklist.push(I: Inst);
2928 }
2929 return LastInst;
2930}
2931
2932std::optional<std::pair<Intrinsic::ID, SmallVector<Value *, 3>>>
2933InstCombinerImpl::convertOrOfShiftsToFunnelShift(Instruction &Or) {
2934 // TODO: Can we reduce the code duplication between this and the related
2935 // rotate matching code under visitSelect and visitTrunc?
2936 assert(Or.getOpcode() == BinaryOperator::Or && "Expecting or instruction");
2937
2938 unsigned Width = Or.getType()->getScalarSizeInBits();
2939
2940 Instruction *Or0, *Or1;
2941 if (!match(V: Or.getOperand(i: 0), P: m_Instruction(I&: Or0)) ||
2942 !match(V: Or.getOperand(i: 1), P: m_Instruction(I&: Or1)))
2943 return std::nullopt;
2944
2945 bool IsFshl = true; // Sub on LSHR.
2946 SmallVector<Value *, 3> FShiftArgs;
2947
2948 // First, find an or'd pair of opposite shifts:
2949 // or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1)
2950 if (isa<BinaryOperator>(Val: Or0) && isa<BinaryOperator>(Val: Or1)) {
2951 Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1;
2952 if (!match(V: Or0,
2953 P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: ShVal0), R: m_Value(V&: ShAmt0)))) ||
2954 !match(V: Or1,
2955 P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: ShVal1), R: m_Value(V&: ShAmt1)))) ||
2956 Or0->getOpcode() == Or1->getOpcode())
2957 return std::nullopt;
2958
2959 // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)).
2960 if (Or0->getOpcode() == BinaryOperator::LShr) {
2961 std::swap(a&: Or0, b&: Or1);
2962 std::swap(a&: ShVal0, b&: ShVal1);
2963 std::swap(a&: ShAmt0, b&: ShAmt1);
2964 }
2965 assert(Or0->getOpcode() == BinaryOperator::Shl &&
2966 Or1->getOpcode() == BinaryOperator::LShr &&
2967 "Illegal or(shift,shift) pair");
2968
2969 // Match the shift amount operands for a funnel shift pattern. This always
2970 // matches a subtraction on the R operand.
2971 auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * {
2972 // Check for constant shift amounts that sum to the bitwidth.
2973 const APInt *LI, *RI;
2974 if (match(V: L, P: m_APIntAllowPoison(Res&: LI)) && match(V: R, P: m_APIntAllowPoison(Res&: RI)))
2975 if (LI->ult(RHS: Width) && RI->ult(RHS: Width) && (*LI + *RI) == Width)
2976 return ConstantInt::get(Ty: L->getType(), V: *LI);
2977
2978 Constant *LC, *RC;
2979 if (match(V: L, P: m_Constant(C&: LC)) && match(V: R, P: m_Constant(C&: RC)) &&
2980 match(V: L,
2981 P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_ULT, Threshold: APInt(Width, Width))) &&
2982 match(V: R,
2983 P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_ULT, Threshold: APInt(Width, Width))) &&
2984 match(V: ConstantExpr::getAdd(C1: LC, C2: RC), P: m_SpecificIntAllowPoison(V: Width)))
2985 return ConstantExpr::mergeUndefsWith(C: LC, Other: RC);
2986
2987 // (shl ShVal, X) | (lshr ShVal, (Width - x)) iff X < Width.
2988 // We limit this to X < Width in case the backend re-expands the
2989 // intrinsic, and has to reintroduce a shift modulo operation (InstCombine
2990 // might remove it after this fold). This still doesn't guarantee that the
2991 // final codegen will match this original pattern.
2992 if (match(V: R, P: m_OneUse(SubPattern: m_Sub(L: m_SpecificInt(V: Width), R: m_Specific(V: L))))) {
2993 KnownBits KnownL = computeKnownBits(V: L, CxtI: &Or);
2994 return KnownL.getMaxValue().ult(RHS: Width) ? L : nullptr;
2995 }
2996
2997 // For non-constant cases, the following patterns currently only work for
2998 // rotation patterns.
2999 // TODO: Add general funnel-shift compatible patterns.
3000 if (ShVal0 != ShVal1)
3001 return nullptr;
3002
3003 // For non-constant cases we don't support non-pow2 shift masks.
3004 // TODO: Is it worth matching urem as well?
3005 if (!isPowerOf2_32(Value: Width))
3006 return nullptr;
3007
3008 // The shift amount may be masked with negation:
3009 // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1)))
3010 Value *X;
3011 unsigned Mask = Width - 1;
3012 if (match(V: L, P: m_And(L: m_Value(V&: X), R: m_SpecificInt(V: Mask))) &&
3013 match(V: R, P: m_And(L: m_Neg(V: m_Specific(V: X)), R: m_SpecificInt(V: Mask))))
3014 return X;
3015
3016 // (shl ShVal, X) | (lshr ShVal, ((-X) & (Width - 1)))
3017 if (match(V: R, P: m_And(L: m_Neg(V: m_Specific(V: L)), R: m_SpecificInt(V: Mask))))
3018 return L;
3019
3020 // Similar to above, but the shift amount may be extended after masking,
3021 // so return the extended value as the parameter for the intrinsic.
3022 if (match(V: L, P: m_ZExt(Op: m_And(L: m_Value(V&: X), R: m_SpecificInt(V: Mask)))) &&
3023 match(V: R,
3024 P: m_And(L: m_Neg(V: m_ZExt(Op: m_And(L: m_Specific(V: X), R: m_SpecificInt(V: Mask)))),
3025 R: m_SpecificInt(V: Mask))))
3026 return L;
3027
3028 if (match(V: L, P: m_ZExt(Op: m_And(L: m_Value(V&: X), R: m_SpecificInt(V: Mask)))) &&
3029 match(V: R, P: m_ZExt(Op: m_And(L: m_Neg(V: m_Specific(V: X)), R: m_SpecificInt(V: Mask)))))
3030 return L;
3031
3032 return nullptr;
3033 };
3034
3035 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width);
3036 if (!ShAmt) {
3037 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width);
3038 IsFshl = false; // Sub on SHL.
3039 }
3040 if (!ShAmt)
3041 return std::nullopt;
3042
3043 FShiftArgs = {ShVal0, ShVal1, ShAmt};
3044 } else if (isa<ZExtInst>(Val: Or0) || isa<ZExtInst>(Val: Or1)) {
3045 // If there are two 'or' instructions concat variables in opposite order:
3046 //
3047 // Slot1 and Slot2 are all zero bits.
3048 // | Slot1 | Low | Slot2 | High |
3049 // LowHigh = or (shl (zext Low), ZextLowShlAmt), (zext High)
3050 // | Slot2 | High | Slot1 | Low |
3051 // HighLow = or (shl (zext High), ZextHighShlAmt), (zext Low)
3052 //
3053 // the latter 'or' can be safely convert to
3054 // -> HighLow = fshl LowHigh, LowHigh, ZextHighShlAmt
3055 // if ZextLowShlAmt + ZextHighShlAmt == Width.
3056 if (!isa<ZExtInst>(Val: Or1))
3057 std::swap(a&: Or0, b&: Or1);
3058
3059 Value *High, *ZextHigh, *Low;
3060 const APInt *ZextHighShlAmt;
3061 if (!match(V: Or0,
3062 P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: ZextHigh), R: m_APInt(Res&: ZextHighShlAmt)))))
3063 return std::nullopt;
3064
3065 if (!match(V: Or1, P: m_ZExt(Op: m_Value(V&: Low))) ||
3066 !match(V: ZextHigh, P: m_ZExt(Op: m_Value(V&: High))))
3067 return std::nullopt;
3068
3069 unsigned HighSize = High->getType()->getScalarSizeInBits();
3070 unsigned LowSize = Low->getType()->getScalarSizeInBits();
3071 // Make sure High does not overlap with Low and most significant bits of
3072 // High aren't shifted out.
3073 if (ZextHighShlAmt->ult(RHS: LowSize) || ZextHighShlAmt->ugt(RHS: Width - HighSize))
3074 return std::nullopt;
3075
3076 for (User *U : ZextHigh->users()) {
3077 Value *X, *Y;
3078 if (!match(V: U, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))))
3079 continue;
3080
3081 if (!isa<ZExtInst>(Val: Y))
3082 std::swap(a&: X, b&: Y);
3083
3084 const APInt *ZextLowShlAmt;
3085 if (!match(V: X, P: m_Shl(L: m_Specific(V: Or1), R: m_APInt(Res&: ZextLowShlAmt))) ||
3086 !match(V: Y, P: m_Specific(V: ZextHigh)) || !DT.dominates(Def: U, User: &Or))
3087 continue;
3088
3089 // HighLow is good concat. If sum of two shifts amount equals to Width,
3090 // LowHigh must also be a good concat.
3091 if (*ZextLowShlAmt + *ZextHighShlAmt != Width)
3092 continue;
3093
3094 // Low must not overlap with High and most significant bits of Low must
3095 // not be shifted out.
3096 assert(ZextLowShlAmt->uge(HighSize) &&
3097 ZextLowShlAmt->ule(Width - LowSize) && "Invalid concat");
3098
3099 // We cannot reuse the result if it may produce poison.
3100 // Drop poison generating flags in the expression tree.
3101 // Or
3102 cast<Instruction>(Val: U)->dropPoisonGeneratingFlags();
3103 // Shl
3104 cast<Instruction>(Val: X)->dropPoisonGeneratingFlags();
3105
3106 FShiftArgs = {U, U, ConstantInt::get(Ty: Or0->getType(), V: *ZextHighShlAmt)};
3107 break;
3108 }
3109 }
3110
3111 if (FShiftArgs.empty())
3112 return std::nullopt;
3113
3114 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
3115 return std::make_pair(x&: IID, y&: FShiftArgs);
3116}
3117
3118/// Match UB-safe variants of the funnel shift intrinsic.
3119static Instruction *matchFunnelShift(Instruction &Or, InstCombinerImpl &IC) {
3120 if (auto Opt = IC.convertOrOfShiftsToFunnelShift(Or)) {
3121 auto [IID, FShiftArgs] = *Opt;
3122 Function *F =
3123 Intrinsic::getOrInsertDeclaration(M: Or.getModule(), id: IID, OverloadTys: Or.getType());
3124 return CallInst::Create(Func: F, Args: FShiftArgs);
3125 }
3126
3127 return nullptr;
3128}
3129
3130/// Attempt to combine or(zext(x),shl(zext(y),bw/2) concat packing patterns.
3131static Value *matchOrConcat(Instruction &Or, InstCombiner::BuilderTy &Builder) {
3132 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'");
3133 Value *Op0 = Or.getOperand(i: 0), *Op1 = Or.getOperand(i: 1);
3134 Type *Ty = Or.getType();
3135
3136 unsigned Width = Ty->getScalarSizeInBits();
3137 if ((Width & 1) != 0)
3138 return nullptr;
3139 unsigned HalfWidth = Width / 2;
3140
3141 // Canonicalize zext (lower half) to LHS.
3142 if (!isa<ZExtInst>(Val: Op0))
3143 std::swap(a&: Op0, b&: Op1);
3144
3145 // Find lower/upper half.
3146 Value *LowerSrc, *ShlVal, *UpperSrc;
3147 const APInt *C;
3148 if (!match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: LowerSrc)))) ||
3149 !match(V: Op1, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: ShlVal), R: m_APInt(Res&: C)))) ||
3150 !match(V: ShlVal, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: UpperSrc)))))
3151 return nullptr;
3152 if (*C != HalfWidth || LowerSrc->getType() != UpperSrc->getType() ||
3153 LowerSrc->getType()->getScalarSizeInBits() != HalfWidth)
3154 return nullptr;
3155
3156 auto ConcatIntrinsicCalls = [&](Intrinsic::ID id, Value *Lo, Value *Hi) {
3157 Value *NewLower = Builder.CreateZExt(V: Lo, DestTy: Ty);
3158 Value *NewUpper = Builder.CreateZExt(V: Hi, DestTy: Ty);
3159 NewUpper = Builder.CreateShl(LHS: NewUpper, RHS: HalfWidth);
3160 Value *BinOp = Builder.CreateDisjointOr(LHS: NewLower, RHS: NewUpper);
3161 return Builder.CreateIntrinsic(ID: id, OverloadTypes: Ty, Args: BinOp);
3162 };
3163
3164 // BSWAP: Push the concat down, swapping the lower/upper sources.
3165 // concat(bswap(x),bswap(y)) -> bswap(concat(x,y))
3166 Value *LowerBSwap, *UpperBSwap;
3167 if (match(V: LowerSrc, P: m_BSwap(Op0: m_Value(V&: LowerBSwap))) &&
3168 match(V: UpperSrc, P: m_BSwap(Op0: m_Value(V&: UpperBSwap))))
3169 return ConcatIntrinsicCalls(Intrinsic::bswap, UpperBSwap, LowerBSwap);
3170
3171 // BITREVERSE: Push the concat down, swapping the lower/upper sources.
3172 // concat(bitreverse(x),bitreverse(y)) -> bitreverse(concat(x,y))
3173 Value *LowerBRev, *UpperBRev;
3174 if (match(V: LowerSrc, P: m_BitReverse(Op0: m_Value(V&: LowerBRev))) &&
3175 match(V: UpperSrc, P: m_BitReverse(Op0: m_Value(V&: UpperBRev))))
3176 return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev);
3177
3178 // iX ext split: extending or(zext(x),shl(zext(y),bw/2) pattern
3179 // to consume sext/ashr:
3180 // or(zext(sext(x)),shl(zext(sext(ashr(x,xbw-1))),bw/2)
3181 // or(zext(x),shl(zext(ashr(x,xbw-1)),bw/2)
3182 Value *X;
3183 if (match(V: LowerSrc, P: m_SExtOrSelf(Op: m_Value(V&: X))) &&
3184 match(V: UpperSrc,
3185 P: m_SExtOrSelf(Op: m_AShr(
3186 L: m_Specific(V: X),
3187 R: m_SpecificInt(V: X->getType()->getScalarSizeInBits() - 1)))))
3188 return Builder.CreateSExt(V: X, DestTy: Ty);
3189
3190 return nullptr;
3191}
3192
3193/// If all elements of two constant vectors are 0/-1 and inverses, return true.
3194static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {
3195 unsigned NumElts = cast<FixedVectorType>(Val: C1->getType())->getNumElements();
3196 for (unsigned i = 0; i != NumElts; ++i) {
3197 Constant *EltC1 = C1->getAggregateElement(Elt: i);
3198 Constant *EltC2 = C2->getAggregateElement(Elt: i);
3199 if (!EltC1 || !EltC2)
3200 return false;
3201
3202 // One element must be all ones, and the other must be all zeros.
3203 if (!((match(V: EltC1, P: m_Zero()) && match(V: EltC2, P: m_AllOnes())) ||
3204 (match(V: EltC2, P: m_Zero()) && match(V: EltC1, P: m_AllOnes()))))
3205 return false;
3206 }
3207 return true;
3208}
3209
3210/// We have an expression of the form (A & C) | (B & D). If A is a scalar or
3211/// vector composed of all-zeros or all-ones values and is the bitwise 'not' of
3212/// B, it can be used as the condition operand of a select instruction.
3213/// We will detect (A & C) | ~(B | D) when the flag ABIsTheSame enabled.
3214Value *InstCombinerImpl::getSelectCondition(Value *A, Value *B,
3215 bool ABIsTheSame) {
3216 // We may have peeked through bitcasts in the caller.
3217 // Exit immediately if we don't have (vector) integer types.
3218 Type *Ty = A->getType();
3219 if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy())
3220 return nullptr;
3221
3222 // If A is the 'not' operand of B and has enough signbits, we have our answer.
3223 if (ABIsTheSame ? (A == B) : match(V: B, P: m_Not(V: m_Specific(V: A)))) {
3224 // If these are scalars or vectors of i1, A can be used directly.
3225 if (Ty->isIntOrIntVectorTy(BitWidth: 1))
3226 return A;
3227
3228 // If we look through a vector bitcast, the caller will bitcast the operands
3229 // to match the condition's number of bits (N x i1).
3230 // To make this poison-safe, disallow bitcast from wide element to narrow
3231 // element. That could allow poison in lanes where it was not present in the
3232 // original code.
3233 A = peekThroughBitcast(V: A);
3234 if (A->getType()->isIntOrIntVectorTy()) {
3235 unsigned NumSignBits = ComputeNumSignBits(Op: A);
3236 if (NumSignBits == A->getType()->getScalarSizeInBits() &&
3237 NumSignBits <= Ty->getScalarSizeInBits())
3238 return Builder.CreateTrunc(V: A, DestTy: CmpInst::makeCmpResultType(opnd_type: A->getType()));
3239 }
3240 return nullptr;
3241 }
3242
3243 // TODO: add support for sext and constant case
3244 if (ABIsTheSame)
3245 return nullptr;
3246
3247 // If both operands are constants, see if the constants are inverse bitmasks.
3248 Constant *AConst, *BConst;
3249 if (match(V: A, P: m_Constant(C&: AConst)) && match(V: B, P: m_Constant(C&: BConst)))
3250 if (AConst == ConstantExpr::getNot(C: BConst) &&
3251 ComputeNumSignBits(Op: A) == Ty->getScalarSizeInBits())
3252 return Builder.CreateZExtOrTrunc(V: A, DestTy: CmpInst::makeCmpResultType(opnd_type: Ty));
3253
3254 // Look for more complex patterns. The 'not' op may be hidden behind various
3255 // casts. Look through sexts and bitcasts to find the booleans.
3256 Value *Cond;
3257 Value *NotB;
3258 if (match(V: A, P: m_SExt(Op: m_Value(V&: Cond))) &&
3259 Cond->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
3260 // A = sext i1 Cond; B = sext (not (i1 Cond))
3261 if (match(V: B, P: m_SExt(Op: m_Not(V: m_Specific(V: Cond)))))
3262 return Cond;
3263
3264 // A = sext i1 Cond; B = not ({bitcast} (sext (i1 Cond)))
3265 // TODO: The one-use checks are unnecessary or misplaced. If the caller
3266 // checked for uses on logic ops/casts, that should be enough to
3267 // make this transform worthwhile.
3268 if (match(V: B, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: NotB))))) {
3269 NotB = peekThroughBitcast(V: NotB, OneUseOnly: true);
3270 if (match(V: NotB, P: m_SExt(Op: m_Specific(V: Cond))))
3271 return Cond;
3272 }
3273 }
3274
3275 // All scalar (and most vector) possibilities should be handled now.
3276 // Try more matches that only apply to non-splat constant vectors.
3277 if (!Ty->isVectorTy())
3278 return nullptr;
3279
3280 // If both operands are xor'd with constants using the same sexted boolean
3281 // operand, see if the constants are inverse bitmasks.
3282 // TODO: Use ConstantExpr::getNot()?
3283 if (match(V: A, P: (m_Xor(L: m_SExt(Op: m_Value(V&: Cond)), R: m_Constant(C&: AConst)))) &&
3284 match(V: B, P: (m_Xor(L: m_SExt(Op: m_Specific(V: Cond)), R: m_Constant(C&: BConst)))) &&
3285 Cond->getType()->isIntOrIntVectorTy(BitWidth: 1) &&
3286 areInverseVectorBitmasks(C1: AConst, C2: BConst)) {
3287 AConst = ConstantExpr::getTrunc(C: AConst, Ty: CmpInst::makeCmpResultType(opnd_type: Ty));
3288 return Builder.CreateXor(LHS: Cond, RHS: AConst);
3289 }
3290 return nullptr;
3291}
3292
3293/// We have an expression of the form (A & B) | (C & D). Try to simplify this
3294/// to "A' ? B : D", where A' is a boolean or vector of booleans.
3295/// When InvertFalseVal is set to true, we try to match the pattern
3296/// where we have peeked through a 'not' op and A and C are the same:
3297/// (A & B) | ~(A | D) --> (A & B) | (~A & ~D) --> A' ? B : ~D
3298Value *InstCombinerImpl::matchSelectFromAndOr(Value *A, Value *B, Value *C,
3299 Value *D, bool InvertFalseVal) {
3300 // The potential condition of the select may be bitcasted. In that case, look
3301 // through its bitcast and the corresponding bitcast of the 'not' condition.
3302 Type *OrigType = A->getType();
3303 A = peekThroughBitcast(V: A, OneUseOnly: true);
3304 C = peekThroughBitcast(V: C, OneUseOnly: true);
3305 if (Value *Cond = getSelectCondition(A, B: C, ABIsTheSame: InvertFalseVal)) {
3306 // ((bc Cond) & B) | ((bc ~Cond) & D) --> bc (select Cond, (bc B), (bc D))
3307 // If this is a vector, we may need to cast to match the condition's length.
3308 // The bitcasts will either all exist or all not exist. The builder will
3309 // not create unnecessary casts if the types already match.
3310 Type *SelTy = A->getType();
3311 if (auto *VecTy = dyn_cast<VectorType>(Val: Cond->getType())) {
3312 // For a fixed or scalable vector get N from <{vscale x} N x iM>
3313 unsigned Elts = VecTy->getElementCount().getKnownMinValue();
3314 // For a fixed or scalable vector, get the size in bits of N x iM; for a
3315 // scalar this is just M.
3316 unsigned SelEltSize = SelTy->getPrimitiveSizeInBits().getKnownMinValue();
3317 Type *EltTy = Builder.getIntNTy(N: SelEltSize / Elts);
3318 SelTy = VectorType::get(ElementType: EltTy, EC: VecTy->getElementCount());
3319 }
3320 Value *BitcastB = Builder.CreateBitCast(V: B, DestTy: SelTy);
3321 if (InvertFalseVal)
3322 D = Builder.CreateNot(V: D);
3323 Value *BitcastD = Builder.CreateBitCast(V: D, DestTy: SelTy);
3324 Value *Select = Builder.CreateSelect(C: Cond, True: BitcastB, False: BitcastD);
3325 return Builder.CreateBitCast(V: Select, DestTy: OrigType);
3326 }
3327
3328 return nullptr;
3329}
3330
3331// (icmp eq X, C) | (icmp ult Other, (X - C)) -> (icmp ule Other, (X - (C + 1)))
3332// (icmp ne X, C) & (icmp uge Other, (X - C)) -> (icmp ugt Other, (X - (C + 1)))
3333static Value *foldAndOrOfICmpEqConstantAndICmp(ICmpInst *LHS, ICmpInst *RHS,
3334 bool IsAnd, bool IsLogical,
3335 IRBuilderBase &Builder) {
3336 Value *LHS0 = LHS->getOperand(i_nocapture: 0);
3337 Value *RHS0 = RHS->getOperand(i_nocapture: 0);
3338 Value *RHS1 = RHS->getOperand(i_nocapture: 1);
3339
3340 ICmpInst::Predicate LPred =
3341 IsAnd ? LHS->getInversePredicate() : LHS->getPredicate();
3342 ICmpInst::Predicate RPred =
3343 IsAnd ? RHS->getInversePredicate() : RHS->getPredicate();
3344
3345 const APInt *CInt;
3346 if (LPred != ICmpInst::ICMP_EQ ||
3347 !match(V: LHS->getOperand(i_nocapture: 1), P: m_APIntAllowPoison(Res&: CInt)) ||
3348 !LHS0->getType()->isIntOrIntVectorTy() ||
3349 !(LHS->hasOneUse() || RHS->hasOneUse()))
3350 return nullptr;
3351
3352 auto MatchRHSOp = [LHS0, CInt](const Value *RHSOp) {
3353 return match(V: RHSOp,
3354 P: m_Add(L: m_Specific(V: LHS0), R: m_SpecificIntAllowPoison(V: -*CInt))) ||
3355 (CInt->isZero() && RHSOp == LHS0);
3356 };
3357
3358 Value *Other;
3359 if (RPred == ICmpInst::ICMP_ULT && MatchRHSOp(RHS1))
3360 Other = RHS0;
3361 else if (RPred == ICmpInst::ICMP_UGT && MatchRHSOp(RHS0))
3362 Other = RHS1;
3363 else
3364 return nullptr;
3365
3366 if (IsLogical)
3367 Other = Builder.CreateFreeze(V: Other);
3368
3369 return Builder.CreateICmp(
3370 P: IsAnd ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE,
3371 LHS: Builder.CreateSub(LHS: LHS0, RHS: ConstantInt::get(Ty: LHS0->getType(), V: *CInt + 1)),
3372 RHS: Other);
3373}
3374
3375/// Fold (icmp)&(icmp) or (icmp)|(icmp) if possible.
3376/// If IsLogical is true, then the and/or is in select form and the transform
3377/// must be poison-safe.
3378Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
3379 Instruction &I, bool IsAnd,
3380 bool IsLogical) {
3381 const SimplifyQuery Q = SQ.getWithInstruction(I: &I);
3382
3383 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
3384 Value *LHS0 = LHS->getOperand(i_nocapture: 0), *RHS0 = RHS->getOperand(i_nocapture: 0);
3385 Value *LHS1 = LHS->getOperand(i_nocapture: 1), *RHS1 = RHS->getOperand(i_nocapture: 1);
3386
3387 const APInt *LHSC = nullptr, *RHSC = nullptr;
3388 match(V: LHS1, P: m_APInt(Res&: LHSC));
3389 match(V: RHS1, P: m_APInt(Res&: RHSC));
3390
3391 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3392 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3393 if (predicatesFoldable(P1: PredL, P2: PredR)) {
3394 if (LHS0 == RHS1 && LHS1 == RHS0) {
3395 PredL = ICmpInst::getSwappedPredicate(pred: PredL);
3396 std::swap(a&: LHS0, b&: LHS1);
3397 }
3398 if (LHS0 == RHS0 && LHS1 == RHS1) {
3399 unsigned Code = IsAnd ? getICmpCode(Pred: PredL) & getICmpCode(Pred: PredR)
3400 : getICmpCode(Pred: PredL) | getICmpCode(Pred: PredR);
3401 bool IsSigned = LHS->isSigned() || RHS->isSigned();
3402 return getNewICmpValue(Code, Sign: IsSigned, LHS: LHS0, RHS: LHS1, Builder);
3403 }
3404 }
3405
3406 if (Value *V =
3407 foldAndOrOfICmpEqConstantAndICmp(LHS, RHS, IsAnd, IsLogical, Builder))
3408 return V;
3409 // We can treat logical like bitwise here, because both operands are used on
3410 // the LHS, and as such poison from both will propagate.
3411 if (Value *V = foldAndOrOfICmpEqConstantAndICmp(LHS: RHS, RHS: LHS, IsAnd,
3412 /*IsLogical*/ false, Builder))
3413 return V;
3414
3415 if (Value *V = foldAndOrOfICmpsWithConstEq(Cmp0: LHS, Cmp1: RHS, IsAnd, IsLogical,
3416 Builder, Q, I))
3417 return V;
3418 // We can convert this case to bitwise and, because both operands are used
3419 // on the LHS, and as such poison from both will propagate.
3420 if (Value *V = foldAndOrOfICmpsWithConstEq(
3421 Cmp0: RHS, Cmp1: LHS, IsAnd, /*IsLogical=*/false, Builder, Q, I)) {
3422 // If RHS is still used, we should drop samesign flag.
3423 if (IsLogical && RHS->hasSameSign() && !RHS->use_empty()) {
3424 RHS->setSameSign(false);
3425 addToWorklist(I: RHS);
3426 }
3427 return V;
3428 }
3429
3430 if (Value *V = foldIsPowerOf2OrZero(Cmp0: LHS, Cmp1: RHS, IsAnd, Builder, IC&: *this))
3431 return V;
3432 if (Value *V = foldIsPowerOf2OrZero(Cmp0: RHS, Cmp1: LHS, IsAnd, Builder, IC&: *this))
3433 return V;
3434
3435 // TODO: One of these directions is fine with logical and/or, the other could
3436 // be supported by inserting freeze.
3437 if (!IsLogical) {
3438 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
3439 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
3440 if (Value *V = simplifyRangeCheck(Cmp0: LHS, Cmp1: RHS, /*Inverted=*/!IsAnd))
3441 return V;
3442
3443 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
3444 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
3445 if (Value *V = simplifyRangeCheck(Cmp0: RHS, Cmp1: LHS, /*Inverted=*/!IsAnd))
3446 return V;
3447 }
3448
3449 // TODO: Add conjugated or fold, check whether it is safe for logical and/or.
3450 if (IsAnd && !IsLogical)
3451 if (Value *V = foldSignedTruncationCheck(ICmp0: LHS, ICmp1: RHS, CxtI&: I, Builder))
3452 return V;
3453
3454 if (Value *V = foldIsPowerOf2(Cmp0: LHS, Cmp1: RHS, JoinedByAnd: IsAnd, Builder, IC&: *this))
3455 return V;
3456
3457 if (Value *V = foldPowerOf2AndShiftedMask(Cmp0: LHS, Cmp1: RHS, JoinedByAnd: IsAnd, Builder))
3458 return V;
3459
3460 // TODO: Verify whether this is safe for logical and/or.
3461 if (!IsLogical) {
3462 if (Value *X = foldUnsignedUnderflowCheck(ZeroICmp: LHS, UnsignedICmp: RHS, IsAnd, Q, Builder))
3463 return X;
3464 if (Value *X = foldUnsignedUnderflowCheck(ZeroICmp: RHS, UnsignedICmp: LHS, IsAnd, Q, Builder))
3465 return X;
3466 }
3467
3468 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
3469 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
3470 // TODO: Remove this and below when foldLogOpOfMaskedICmps can handle undefs.
3471 if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&
3472 PredL == PredR && match(V: LHS1, P: m_ZeroInt()) && match(V: RHS1, P: m_ZeroInt()) &&
3473 LHS0->getType() == RHS0->getType() &&
3474 (!IsLogical || isGuaranteedNotToBePoison(V: RHS0))) {
3475 Value *NewOr = Builder.CreateOr(LHS: LHS0, RHS: RHS0);
3476 return Builder.CreateICmp(P: PredL, LHS: NewOr,
3477 RHS: Constant::getNullValue(Ty: NewOr->getType()));
3478 }
3479
3480 // (icmp ne A, -1) | (icmp ne B, -1) --> (icmp ne (A&B), -1)
3481 // (icmp eq A, -1) & (icmp eq B, -1) --> (icmp eq (A&B), -1)
3482 if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&
3483 PredL == PredR && match(V: LHS1, P: m_AllOnes()) && match(V: RHS1, P: m_AllOnes()) &&
3484 LHS0->getType() == RHS0->getType() &&
3485 (!IsLogical || isGuaranteedNotToBePoison(V: RHS0))) {
3486 Value *NewAnd = Builder.CreateAnd(LHS: LHS0, RHS: RHS0);
3487 return Builder.CreateICmp(P: PredL, LHS: NewAnd,
3488 RHS: Constant::getAllOnesValue(Ty: LHS0->getType()));
3489 }
3490
3491 if (!IsLogical)
3492 if (Value *V =
3493 foldAndOrOfICmpsWithPow2AndWithZero(Builder, LHS, RHS, IsAnd, Q))
3494 return V;
3495
3496 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
3497 if (!LHSC || !RHSC)
3498 return nullptr;
3499
3500 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
3501 // (trunc x) != C1 | (and x, CA) != C2 -> (and x, CA|CMAX) != C1|C2
3502 // where CMAX is the all ones value for the truncated type,
3503 // iff the lower bits of C2 and CA are zero.
3504 if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&
3505 PredL == PredR && LHS->hasOneUse() && RHS->hasOneUse()) {
3506 Value *V;
3507 const APInt *AndC, *SmallC = nullptr, *BigC = nullptr;
3508
3509 // (trunc x) == C1 & (and x, CA) == C2
3510 // (and x, CA) == C2 & (trunc x) == C1
3511 if (match(V: RHS0, P: m_Trunc(Op: m_Value(V))) &&
3512 match(V: LHS0, P: m_And(L: m_Specific(V), R: m_APInt(Res&: AndC)))) {
3513 SmallC = RHSC;
3514 BigC = LHSC;
3515 } else if (match(V: LHS0, P: m_Trunc(Op: m_Value(V))) &&
3516 match(V: RHS0, P: m_And(L: m_Specific(V), R: m_APInt(Res&: AndC)))) {
3517 SmallC = LHSC;
3518 BigC = RHSC;
3519 }
3520
3521 if (SmallC && BigC) {
3522 unsigned BigBitSize = BigC->getBitWidth();
3523 unsigned SmallBitSize = SmallC->getBitWidth();
3524
3525 // Check that the low bits are zero.
3526 APInt Low = APInt::getLowBitsSet(numBits: BigBitSize, loBitsSet: SmallBitSize);
3527 if ((Low & *AndC).isZero() && (Low & *BigC).isZero()) {
3528 Value *NewAnd = Builder.CreateAnd(LHS: V, RHS: Low | *AndC);
3529 APInt N = SmallC->zext(width: BigBitSize) | *BigC;
3530 Value *NewVal = ConstantInt::get(Ty: NewAnd->getType(), V: N);
3531 return Builder.CreateICmp(P: PredL, LHS: NewAnd, RHS: NewVal);
3532 }
3533 }
3534 }
3535
3536 // Match naive pattern (and its inverted form) for checking if two values
3537 // share same sign. An example of the pattern:
3538 // (icmp slt (X & Y), 0) | (icmp sgt (X | Y), -1) -> (icmp sgt (X ^ Y), -1)
3539 // Inverted form (example):
3540 // (icmp slt (X | Y), 0) & (icmp sgt (X & Y), -1) -> (icmp slt (X ^ Y), 0)
3541 bool TrueIfSignedL, TrueIfSignedR;
3542 if (isSignBitCheck(Pred: PredL, RHS: *LHSC, TrueIfSigned&: TrueIfSignedL) &&
3543 isSignBitCheck(Pred: PredR, RHS: *RHSC, TrueIfSigned&: TrueIfSignedR) &&
3544 (RHS->hasOneUse() || LHS->hasOneUse())) {
3545 Value *X, *Y;
3546 if (IsAnd) {
3547 if ((TrueIfSignedL && !TrueIfSignedR &&
3548 match(V: LHS0, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
3549 match(V: RHS0, P: m_c_And(L: m_Specific(V: X), R: m_Specific(V: Y)))) ||
3550 (!TrueIfSignedL && TrueIfSignedR &&
3551 match(V: LHS0, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
3552 match(V: RHS0, P: m_c_Or(L: m_Specific(V: X), R: m_Specific(V: Y))))) {
3553 Value *NewXor = Builder.CreateXor(LHS: X, RHS: Y);
3554 return Builder.CreateIsNeg(Arg: NewXor);
3555 }
3556 } else {
3557 if ((TrueIfSignedL && !TrueIfSignedR &&
3558 match(V: LHS0, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
3559 match(V: RHS0, P: m_c_Or(L: m_Specific(V: X), R: m_Specific(V: Y)))) ||
3560 (!TrueIfSignedL && TrueIfSignedR &&
3561 match(V: LHS0, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))) &&
3562 match(V: RHS0, P: m_c_And(L: m_Specific(V: X), R: m_Specific(V: Y))))) {
3563 Value *NewXor = Builder.CreateXor(LHS: X, RHS: Y);
3564 return Builder.CreateIsNotNeg(Arg: NewXor);
3565 }
3566 }
3567 }
3568
3569 // (X & ExpMask) != 0 && (X & ExpMask) != ExpMask -> isnormal(X)
3570 // (X & ExpMask) == 0 || (X & ExpMask) == ExpMask -> !isnormal(X)
3571 Value *X;
3572 const APInt *MaskC;
3573 if (LHS0 == RHS0 && PredL == PredR &&
3574 PredL == (IsAnd ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ) &&
3575 !I.getFunction()->hasFnAttribute(Kind: Attribute::NoImplicitFloat) &&
3576 LHS->hasOneUse() && RHS->hasOneUse() &&
3577 match(V: LHS0, P: m_And(L: m_ElementWiseBitCast(Op: m_Value(V&: X)), R: m_APInt(Res&: MaskC))) &&
3578 X->getType()->getScalarType()->isIEEELikeFPTy() &&
3579 APFloat(X->getType()->getScalarType()->getFltSemantics(), *MaskC)
3580 .isPosInfinity() &&
3581 ((LHSC->isZero() && *RHSC == *MaskC) ||
3582 (RHSC->isZero() && *LHSC == *MaskC)))
3583 return Builder.createIsFPClass(FPNum: X, Test: IsAnd ? FPClassTest::fcNormal
3584 : ~FPClassTest::fcNormal);
3585
3586 return foldAndOrOfICmpsUsingRanges(ICmp1: LHS, ICmp2: RHS, IsAnd);
3587}
3588
3589/// If IsLogical is true, then the and/or is in select form and the transform
3590/// must be poison-safe.
3591Value *InstCombinerImpl::foldBooleanAndOr(Value *LHS, Value *RHS,
3592 Instruction &I, bool IsAnd,
3593 bool IsLogical) {
3594 if (!LHS->getType()->isIntOrIntVectorTy(BitWidth: 1))
3595 return nullptr;
3596
3597 // handle (roughly):
3598 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
3599 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
3600 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, IsAnd, IsLogical, Builder,
3601 Q: SQ.getWithInstruction(I: &I)))
3602 return V;
3603
3604 if (auto *LHSCmp = dyn_cast<ICmpInst>(Val: LHS))
3605 if (auto *RHSCmp = dyn_cast<ICmpInst>(Val: RHS))
3606 if (Value *Res = foldAndOrOfICmps(LHS: LHSCmp, RHS: RHSCmp, I, IsAnd, IsLogical))
3607 return Res;
3608
3609 if (auto *LHSCmp = dyn_cast<FCmpInst>(Val: LHS))
3610 if (auto *RHSCmp = dyn_cast<FCmpInst>(Val: RHS))
3611 if (Value *Res = foldLogicOfFCmps(LHS: LHSCmp, RHS: RHSCmp, IsAnd, IsLogicalSelect: IsLogical))
3612 return Res;
3613
3614 if (Value *Res = foldEqOfParts(Cmp0: LHS, Cmp1: RHS, IsAnd))
3615 return Res;
3616
3617 return nullptr;
3618}
3619
3620static Value *foldOrOfInversions(BinaryOperator &I,
3621 InstCombiner::BuilderTy &Builder) {
3622 assert(I.getOpcode() == Instruction::Or &&
3623 "Simplification only supports or at the moment.");
3624
3625 Value *Cmp1, *Cmp2, *Cmp3, *Cmp4;
3626 if (!match(V: I.getOperand(i_nocapture: 0), P: m_And(L: m_Value(V&: Cmp1), R: m_Value(V&: Cmp2))) ||
3627 !match(V: I.getOperand(i_nocapture: 1), P: m_And(L: m_Value(V&: Cmp3), R: m_Value(V&: Cmp4))))
3628 return nullptr;
3629
3630 // Check if any two pairs of the and operations are inversions of each other.
3631 if (isKnownInversion(X: Cmp1, Y: Cmp3) && isKnownInversion(X: Cmp2, Y: Cmp4))
3632 return Builder.CreateXor(LHS: Cmp1, RHS: Cmp4);
3633 if (isKnownInversion(X: Cmp1, Y: Cmp4) && isKnownInversion(X: Cmp2, Y: Cmp3))
3634 return Builder.CreateXor(LHS: Cmp1, RHS: Cmp3);
3635
3636 return nullptr;
3637}
3638
3639/// Match \p V as "shufflevector -> bitcast" or "extractelement -> zext -> shl"
3640/// patterns, which extract vector elements and pack them in the same relative
3641/// positions.
3642///
3643/// \p Vec is the underlying vector being extracted from.
3644/// \p Mask is a bitmask identifying which packed elements are obtained from the
3645/// vector.
3646/// \p VecOffset is the vector element corresponding to index 0 of the
3647/// mask.
3648static bool matchSubIntegerPackFromVector(Value *V, Value *&Vec,
3649 int64_t &VecOffset,
3650 SmallBitVector &Mask,
3651 const DataLayout &DL) {
3652 // First try to match extractelement -> zext -> shl
3653 uint64_t VecIdx, ShlAmt;
3654 if (match(V, P: m_ShlOrSelf(L: m_ZExtOrSelf(Op: m_ExtractElt(Val: m_Value(V&: Vec),
3655 Idx: m_ConstantInt(V&: VecIdx))),
3656 R&: ShlAmt))) {
3657 auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType());
3658 if (!VecTy)
3659 return false;
3660 auto *EltTy = dyn_cast<IntegerType>(Val: VecTy->getElementType());
3661 if (!EltTy)
3662 return false;
3663
3664 const unsigned EltBitWidth = EltTy->getBitWidth();
3665 const unsigned TargetBitWidth = V->getType()->getIntegerBitWidth();
3666 if (TargetBitWidth % EltBitWidth != 0 || ShlAmt % EltBitWidth != 0)
3667 return false;
3668 const unsigned TargetEltWidth = TargetBitWidth / EltBitWidth;
3669 const unsigned ShlEltAmt = ShlAmt / EltBitWidth;
3670
3671 const unsigned MaskIdx =
3672 DL.isLittleEndian() ? ShlEltAmt : TargetEltWidth - ShlEltAmt - 1;
3673
3674 VecOffset = static_cast<int64_t>(VecIdx) - static_cast<int64_t>(MaskIdx);
3675 Mask.resize(N: TargetEltWidth);
3676 Mask.set(MaskIdx);
3677 return true;
3678 }
3679
3680 // Now try to match a bitcasted subvector.
3681 Instruction *SrcVecI;
3682 if (!match(V, P: m_BitCast(Op: m_Instruction(I&: SrcVecI))))
3683 return false;
3684
3685 auto *SrcTy = dyn_cast<FixedVectorType>(Val: SrcVecI->getType());
3686 if (!SrcTy)
3687 return false;
3688
3689 Mask.resize(N: SrcTy->getNumElements());
3690
3691 // First check for a subvector obtained from a shufflevector.
3692 if (isa<ShuffleVectorInst>(Val: SrcVecI)) {
3693 Constant *ConstVec;
3694 ArrayRef<int> ShuffleMask;
3695 if (!match(V: SrcVecI, P: m_Shuffle(v1: m_Value(V&: Vec), v2: m_Constant(C&: ConstVec),
3696 mask: m_Mask(ShuffleMask))))
3697 return false;
3698
3699 auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType());
3700 if (!VecTy)
3701 return false;
3702
3703 const unsigned NumVecElts = VecTy->getNumElements();
3704 bool FoundVecOffset = false;
3705 for (unsigned Idx = 0; Idx < ShuffleMask.size(); ++Idx) {
3706 if (ShuffleMask[Idx] == PoisonMaskElem)
3707 return false;
3708 const unsigned ShuffleIdx = ShuffleMask[Idx];
3709 if (ShuffleIdx >= NumVecElts) {
3710 const unsigned ConstIdx = ShuffleIdx - NumVecElts;
3711 auto *ConstElt =
3712 dyn_cast<ConstantInt>(Val: ConstVec->getAggregateElement(Elt: ConstIdx));
3713 if (!ConstElt || !ConstElt->isNullValue())
3714 return false;
3715 continue;
3716 }
3717
3718 if (FoundVecOffset) {
3719 if (VecOffset + Idx != ShuffleIdx)
3720 return false;
3721 } else {
3722 if (ShuffleIdx < Idx)
3723 return false;
3724 VecOffset = ShuffleIdx - Idx;
3725 FoundVecOffset = true;
3726 }
3727 Mask.set(Idx);
3728 }
3729 return FoundVecOffset;
3730 }
3731
3732 // Check for a subvector obtained as an (insertelement V, 0, idx)
3733 uint64_t InsertIdx;
3734 if (!match(V: SrcVecI,
3735 P: m_InsertElt(Val: m_Value(V&: Vec), Elt: m_Zero(), Idx: m_ConstantInt(V&: InsertIdx))))
3736 return false;
3737
3738 auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType());
3739 if (!VecTy)
3740 return false;
3741 VecOffset = 0;
3742 bool AlreadyInsertedMaskedElt = Mask.test(Idx: InsertIdx);
3743 Mask.set();
3744 if (!AlreadyInsertedMaskedElt)
3745 Mask.reset(Idx: InsertIdx);
3746 return true;
3747}
3748
3749/// Try to fold the join of two scalar integers whose contents are packed
3750/// elements of the same vector.
3751static Instruction *foldIntegerPackFromVector(Instruction &I,
3752 InstCombiner::BuilderTy &Builder,
3753 const DataLayout &DL) {
3754 assert(I.getOpcode() == Instruction::Or);
3755 Value *LhsVec, *RhsVec;
3756 int64_t LhsVecOffset, RhsVecOffset;
3757 SmallBitVector Mask;
3758 if (!matchSubIntegerPackFromVector(V: I.getOperand(i: 0), Vec&: LhsVec, VecOffset&: LhsVecOffset,
3759 Mask, DL))
3760 return nullptr;
3761 if (!matchSubIntegerPackFromVector(V: I.getOperand(i: 1), Vec&: RhsVec, VecOffset&: RhsVecOffset,
3762 Mask, DL))
3763 return nullptr;
3764 if (LhsVec != RhsVec || LhsVecOffset != RhsVecOffset)
3765 return nullptr;
3766
3767 // Convert into shufflevector -> bitcast;
3768 const unsigned ZeroVecIdx =
3769 cast<FixedVectorType>(Val: LhsVec->getType())->getNumElements();
3770 SmallVector<int> ShuffleMask(Mask.size(), ZeroVecIdx);
3771 for (unsigned Idx : Mask.set_bits()) {
3772 assert(LhsVecOffset + Idx >= 0);
3773 ShuffleMask[Idx] = LhsVecOffset + Idx;
3774 }
3775
3776 Value *MaskedVec = Builder.CreateShuffleVector(
3777 V1: LhsVec, V2: Constant::getNullValue(Ty: LhsVec->getType()), Mask: ShuffleMask,
3778 Name: I.getName() + ".v");
3779 return CastInst::Create(Instruction::BitCast, S: MaskedVec, Ty: I.getType());
3780}
3781
3782/// Match \p V as "lshr -> mask -> zext -> shl".
3783///
3784/// \p Int is the underlying integer being extracted from.
3785/// \p Mask is a bitmask identifying which bits of the integer are being
3786/// extracted. \p Offset identifies which bit of the result \p V corresponds to
3787/// the least significant bit of \p Int
3788static bool matchZExtedSubInteger(Value *V, Value *&Int, APInt &Mask,
3789 uint64_t &Offset, bool &IsShlNUW,
3790 bool &IsShlNSW) {
3791 Value *ShlOp0;
3792 uint64_t ShlAmt = 0;
3793 if (!match(V, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: ShlOp0), R: m_ConstantInt(V&: ShlAmt)))))
3794 return false;
3795
3796 IsShlNUW = cast<BinaryOperator>(Val: V)->hasNoUnsignedWrap();
3797 IsShlNSW = cast<BinaryOperator>(Val: V)->hasNoSignedWrap();
3798
3799 Value *ZExtOp0;
3800 if (!match(V: ShlOp0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: ZExtOp0)))))
3801 return false;
3802
3803 Value *MaskedOp0;
3804 const APInt *ShiftedMaskConst = nullptr;
3805 if (!match(V: ZExtOp0, P: m_CombineOr(Ps: m_OneUse(SubPattern: m_And(L: m_Value(V&: MaskedOp0),
3806 R: m_APInt(Res&: ShiftedMaskConst))),
3807 Ps: m_Value(V&: MaskedOp0))))
3808 return false;
3809
3810 uint64_t LShrAmt = 0;
3811 if (!match(V: MaskedOp0,
3812 P: m_CombineOr(Ps: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: Int), R: m_ConstantInt(V&: LShrAmt))),
3813 Ps: m_Value(V&: Int))))
3814 return false;
3815
3816 if (LShrAmt > ShlAmt)
3817 return false;
3818 Offset = ShlAmt - LShrAmt;
3819
3820 Mask = ShiftedMaskConst ? ShiftedMaskConst->shl(shiftAmt: LShrAmt)
3821 : APInt::getBitsSetFrom(
3822 numBits: Int->getType()->getScalarSizeInBits(), loBit: LShrAmt);
3823
3824 return true;
3825}
3826
3827/// Try to fold the join of two scalar integers whose bits are unpacked and
3828/// zexted from the same source integer.
3829static Value *foldIntegerRepackThroughZExt(Value *Lhs, Value *Rhs,
3830 InstCombiner::BuilderTy &Builder) {
3831
3832 Value *LhsInt, *RhsInt;
3833 APInt LhsMask, RhsMask;
3834 uint64_t LhsOffset, RhsOffset;
3835 bool IsLhsShlNUW, IsLhsShlNSW, IsRhsShlNUW, IsRhsShlNSW;
3836 if (!matchZExtedSubInteger(V: Lhs, Int&: LhsInt, Mask&: LhsMask, Offset&: LhsOffset, IsShlNUW&: IsLhsShlNUW,
3837 IsShlNSW&: IsLhsShlNSW))
3838 return nullptr;
3839 if (!matchZExtedSubInteger(V: Rhs, Int&: RhsInt, Mask&: RhsMask, Offset&: RhsOffset, IsShlNUW&: IsRhsShlNUW,
3840 IsShlNSW&: IsRhsShlNSW))
3841 return nullptr;
3842 if (LhsInt != RhsInt || LhsOffset != RhsOffset)
3843 return nullptr;
3844
3845 APInt Mask = LhsMask | RhsMask;
3846
3847 Type *DestTy = Lhs->getType();
3848 Value *Res = Builder.CreateShl(
3849 LHS: Builder.CreateZExt(
3850 V: Builder.CreateAnd(LHS: LhsInt, RHS: Mask, Name: LhsInt->getName() + ".mask"), DestTy,
3851 Name: LhsInt->getName() + ".zext"),
3852 RHS: ConstantInt::get(Ty: DestTy, V: LhsOffset), Name: "", HasNUW: IsLhsShlNUW && IsRhsShlNUW,
3853 HasNSW: IsLhsShlNSW && IsRhsShlNSW);
3854 Res->takeName(V: Lhs);
3855 return Res;
3856}
3857
3858// A decomposition of ((X & Mask) * Factor). The NUW / NSW bools
3859// track these properities for preservation. Note that we can decompose
3860// equivalent select form of this expression (e.g. (!(X & Mask) ? 0 : Mask *
3861// Factor))
3862struct DecomposedBitMaskMul {
3863 Value *X;
3864 APInt Factor;
3865 APInt Mask;
3866 bool NUW;
3867 bool NSW;
3868
3869 bool isCombineableWith(const DecomposedBitMaskMul Other) {
3870 return X == Other.X && !Mask.intersects(RHS: Other.Mask) &&
3871 Factor == Other.Factor;
3872 }
3873};
3874
3875static std::optional<DecomposedBitMaskMul> matchBitmaskMul(Value *V) {
3876 Instruction *Op = dyn_cast<Instruction>(Val: V);
3877 if (!Op)
3878 return std::nullopt;
3879
3880 // Decompose (A & N) * C) into BitMaskMul
3881 Value *Original = nullptr;
3882 const APInt *Mask = nullptr;
3883 const APInt *MulConst = nullptr;
3884 if (match(V: Op, P: m_Mul(L: m_And(L: m_Value(V&: Original), R: m_APInt(Res&: Mask)),
3885 R: m_APInt(Res&: MulConst)))) {
3886 if (MulConst->isZero() || Mask->isZero())
3887 return std::nullopt;
3888
3889 return std::optional<DecomposedBitMaskMul>(
3890 {.X: Original, .Factor: *MulConst, .Mask: *Mask,
3891 .NUW: cast<BinaryOperator>(Val: Op)->hasNoUnsignedWrap(),
3892 .NSW: cast<BinaryOperator>(Val: Op)->hasNoSignedWrap()});
3893 }
3894
3895 Value *Cond = nullptr;
3896 const APInt *EqZero = nullptr, *NeZero = nullptr;
3897
3898 // Decompose ((A & N) ? 0 : N * C) into BitMaskMul
3899 if (match(V: Op, P: m_Select(C: m_Value(V&: Cond), L: m_APInt(Res&: EqZero), R: m_APInt(Res&: NeZero)))) {
3900 auto ICmpDecompose =
3901 decomposeBitTest(Cond, /*LookThroughTrunc=*/true,
3902 /*AllowNonZeroC=*/false, /*DecomposeBitMask=*/DecomposeAnd: true);
3903 if (!ICmpDecompose.has_value())
3904 return std::nullopt;
3905
3906 // decomposeBitTest may provide a scalar bit test for a vector select.
3907 // Ensure the types match.
3908 if (ICmpDecompose->X->getType() != V->getType())
3909 return std::nullopt;
3910
3911 assert(ICmpInst::isEquality(ICmpDecompose->Pred) &&
3912 ICmpDecompose->C.isZero());
3913
3914 if (ICmpDecompose->Pred == ICmpInst::ICMP_NE)
3915 std::swap(a&: EqZero, b&: NeZero);
3916
3917 if (!EqZero->isZero() || NeZero->isZero())
3918 return std::nullopt;
3919
3920 if (!ICmpDecompose->Mask.isPowerOf2() || ICmpDecompose->Mask.isZero())
3921 return std::nullopt;
3922
3923 if (!NeZero->urem(RHS: ICmpDecompose->Mask).isZero())
3924 return std::nullopt;
3925
3926 return std::optional<DecomposedBitMaskMul>(
3927 {.X: ICmpDecompose->X, .Factor: NeZero->udiv(RHS: ICmpDecompose->Mask),
3928 .Mask: ICmpDecompose->Mask, /*NUW=*/false, /*NSW=*/false});
3929 }
3930
3931 return std::nullopt;
3932}
3933
3934/// (A & N) * C + (A & M) * C -> (A & (N + M)) & C
3935/// This also accepts the equivalent select form of (A & N) * C
3936/// expressions i.e. !(A & N) ? 0 : N * C)
3937static Value *foldBitmaskMul(Value *Op0, Value *Op1,
3938 InstCombiner::BuilderTy &Builder) {
3939 auto Decomp1 = matchBitmaskMul(V: Op1);
3940 if (!Decomp1)
3941 return nullptr;
3942
3943 auto Decomp0 = matchBitmaskMul(V: Op0);
3944 if (!Decomp0)
3945 return nullptr;
3946
3947 if (Decomp0->isCombineableWith(Other: *Decomp1)) {
3948 Value *NewAnd = Builder.CreateAnd(
3949 LHS: Decomp0->X,
3950 RHS: ConstantInt::get(Ty: Decomp0->X->getType(), V: Decomp0->Mask + Decomp1->Mask));
3951
3952 return Builder.CreateMul(
3953 LHS: NewAnd, RHS: ConstantInt::get(Ty: NewAnd->getType(), V: Decomp1->Factor), Name: "",
3954 HasNUW: Decomp0->NUW && Decomp1->NUW, HasNSW: Decomp0->NSW && Decomp1->NSW);
3955 }
3956
3957 return nullptr;
3958}
3959
3960Value *InstCombinerImpl::foldDisjointOr(Value *LHS, Value *RHS) {
3961 if (Value *Res = foldBitmaskMul(Op0: LHS, Op1: RHS, Builder))
3962 return Res;
3963 if (Value *Res = foldIntegerRepackThroughZExt(Lhs: LHS, Rhs: RHS, Builder))
3964 return Res;
3965
3966 return nullptr;
3967}
3968
3969Value *InstCombinerImpl::reassociateDisjointOr(Value *LHS, Value *RHS) {
3970
3971 Value *X, *Y;
3972 if (match(V: RHS, P: m_OneUse(SubPattern: m_DisjointOr(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
3973 if (Value *Res = foldDisjointOr(LHS, RHS: X))
3974 return Builder.CreateDisjointOr(LHS: Res, RHS: Y);
3975 if (Value *Res = foldDisjointOr(LHS, RHS: Y))
3976 return Builder.CreateDisjointOr(LHS: Res, RHS: X);
3977 }
3978
3979 if (match(V: LHS, P: m_OneUse(SubPattern: m_DisjointOr(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
3980 if (Value *Res = foldDisjointOr(LHS: X, RHS))
3981 return Builder.CreateDisjointOr(LHS: Res, RHS: Y);
3982 if (Value *Res = foldDisjointOr(LHS: Y, RHS))
3983 return Builder.CreateDisjointOr(LHS: Res, RHS: X);
3984 }
3985
3986 return nullptr;
3987}
3988
3989/// Fold Res, Overflow = (umul.with.overflow x c1); (or Overflow (ugt Res c2))
3990/// --> (ugt x (c2/c1)). This code checks whether a multiplication of two
3991/// unsigned numbers (one is a constant) is mathematically greater than a
3992/// second constant.
3993static Value *foldOrUnsignedUMulOverflowICmp(BinaryOperator &I,
3994 InstCombiner::BuilderTy &Builder,
3995 const DataLayout &DL) {
3996 Value *WOV, *X;
3997 const APInt *C1, *C2;
3998 if (match(V: &I,
3999 P: m_c_Or(L: m_ExtractValue<1>(
4000 V: m_Value(V&: WOV, P: m_Intrinsic<Intrinsic::umul_with_overflow>(
4001 Op0: m_Value(V&: X), Op1: m_APInt(Res&: C1)))),
4002 R: m_OneUse(SubPattern: m_SpecificCmp(MatchPred: ICmpInst::ICMP_UGT,
4003 L: m_ExtractValue<0>(V: m_Deferred(V: WOV)),
4004 R: m_APInt(Res&: C2))))) &&
4005 !C1->isZero()) {
4006 Constant *NewC = ConstantInt::get(Ty: X->getType(), V: C2->udiv(RHS: *C1));
4007 return Builder.CreateICmp(P: ICmpInst::ICMP_UGT, LHS: X, RHS: NewC);
4008 }
4009 return nullptr;
4010}
4011
4012/// Fold select(X >s 0, 0, -X) | smax(X, 0) --> abs(X)
4013/// select(X <s 0, -X, 0) | smax(X, 0) --> abs(X)
4014static Value *FoldOrOfSelectSmaxToAbs(BinaryOperator &I,
4015 InstCombiner::BuilderTy &Builder) {
4016 Value *X;
4017 Value *Sel;
4018 if (match(V: &I,
4019 P: m_c_Or(L: m_Value(V&: Sel), R: m_OneUse(SubPattern: m_SMax(Op0: m_Value(V&: X), Op1: m_ZeroInt()))))) {
4020 auto NegX = m_Neg(V: m_Specific(V: X));
4021 if (match(V: Sel, P: m_Select(C: m_SpecificICmp(MatchPred: ICmpInst::ICMP_SGT, L: m_Specific(V: X),
4022 R: m_ZeroInt()),
4023 L: m_ZeroInt(), R: NegX)) ||
4024 match(V: Sel, P: m_Select(C: m_SpecificICmp(MatchPred: ICmpInst::ICMP_SLT, L: m_Specific(V: X),
4025 R: m_ZeroInt()),
4026 L: NegX, R: m_ZeroInt())))
4027 return Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: X,
4028 RHS: Builder.getFalse());
4029 }
4030 return nullptr;
4031}
4032
4033Instruction *InstCombinerImpl::FoldOrOfLogicalAnds(Value *Op0, Value *Op1) {
4034 Value *C, *A, *B;
4035 // (C && A) || (!C && B)
4036 // (C && A) || (B && !C)
4037 // (A && C) || (!C && B)
4038 // (A && C) || (B && !C) (may require freeze)
4039 //
4040 // => select C, A, B
4041 if (match(V: Op1, P: m_c_LogicalAnd(L: m_Not(V: m_Value(V&: C)), R: m_Value(V&: B))) &&
4042 match(V: Op0, P: m_c_LogicalAnd(L: m_Specific(V: C), R: m_Value(V&: A)))) {
4043 auto *SelOp0 = dyn_cast<SelectInst>(Val: Op0);
4044 auto *SelOp1 = dyn_cast<SelectInst>(Val: Op1);
4045
4046 bool MayNeedFreeze = SelOp0 && SelOp1 &&
4047 match(V: SelOp1->getTrueValue(),
4048 P: m_Not(V: m_Specific(V: SelOp0->getTrueValue())));
4049 if (MayNeedFreeze)
4050 C = Builder.CreateFreeze(V: C);
4051 if (!ProfcheckDisableMetadataFixes) {
4052 Value *C2 = nullptr, *A2 = nullptr, *B2 = nullptr;
4053 if (match(V: Op0, P: m_LogicalAnd(L: m_Specific(V: C), R: m_Value(V&: A2))) && SelOp0) {
4054 return SelectInst::Create(C, S1: A, S2: B, NameStr: "", InsertBefore: nullptr, MDFrom: SelOp0);
4055 } else if (match(V: Op1, P: m_LogicalAnd(L: m_Not(V: m_Value(V&: C2)), R: m_Value(V&: B2))) &&
4056 SelOp1) {
4057 SelectInst *NewSI = SelectInst::Create(C, S1: A, S2: B, NameStr: "", InsertBefore: nullptr, MDFrom: SelOp1);
4058 NewSI->swapProfMetadata();
4059 return NewSI;
4060 } else {
4061 return createSelectInstWithUnknownProfile(C, S1: A, S2: B);
4062 }
4063 }
4064 return SelectInst::Create(C, S1: A, S2: B);
4065 }
4066
4067 // (!C && A) || (C && B)
4068 // (A && !C) || (C && B)
4069 // (!C && A) || (B && C)
4070 // (A && !C) || (B && C) (may require freeze)
4071 //
4072 // => select C, B, A
4073 if (match(V: Op0, P: m_c_LogicalAnd(L: m_Not(V: m_Value(V&: C)), R: m_Value(V&: A))) &&
4074 match(V: Op1, P: m_c_LogicalAnd(L: m_Specific(V: C), R: m_Value(V&: B)))) {
4075 auto *SelOp0 = dyn_cast<SelectInst>(Val: Op0);
4076 auto *SelOp1 = dyn_cast<SelectInst>(Val: Op1);
4077 bool MayNeedFreeze = SelOp0 && SelOp1 &&
4078 match(V: SelOp0->getTrueValue(),
4079 P: m_Not(V: m_Specific(V: SelOp1->getTrueValue())));
4080 if (MayNeedFreeze)
4081 C = Builder.CreateFreeze(V: C);
4082 if (!ProfcheckDisableMetadataFixes) {
4083 Value *C2 = nullptr, *A2 = nullptr, *B2 = nullptr;
4084 if (match(V: Op0, P: m_LogicalAnd(L: m_Not(V: m_Value(V&: C2)), R: m_Value(V&: A2))) && SelOp0) {
4085 SelectInst *NewSI = SelectInst::Create(C, S1: B, S2: A, NameStr: "", InsertBefore: nullptr, MDFrom: SelOp0);
4086 NewSI->swapProfMetadata();
4087 return NewSI;
4088 } else if (match(V: Op1, P: m_LogicalAnd(L: m_Specific(V: C), R: m_Value(V&: B2))) &&
4089 SelOp1) {
4090 return SelectInst::Create(C, S1: B, S2: A, NameStr: "", InsertBefore: nullptr, MDFrom: SelOp1);
4091 } else {
4092 return createSelectInstWithUnknownProfile(C, S1: B, S2: A);
4093 }
4094 }
4095 return SelectInst::Create(C, S1: B, S2: A);
4096 }
4097
4098 return nullptr;
4099}
4100
4101// FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
4102// here. We should standardize that construct where it is needed or choose some
4103// other way to ensure that commutated variants of patterns are not missed.
4104Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
4105 if (Value *V = simplifyOrInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
4106 Q: SQ.getWithInstruction(I: &I)))
4107 return replaceInstUsesWith(I, V);
4108
4109 if (SimplifyAssociativeOrCommutative(I))
4110 return &I;
4111
4112 if (Instruction *X = foldVectorBinop(Inst&: I))
4113 return X;
4114
4115 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
4116 return Phi;
4117
4118 // See if we can simplify any instructions used by the instruction whose sole
4119 // purpose is to compute bits we don't care about.
4120 if (SimplifyDemandedInstructionBits(Inst&: I))
4121 return &I;
4122
4123 // Do this before using distributive laws to catch simple and/or/not patterns.
4124 if (Instruction *Xor = foldOrToXor(I, Builder))
4125 return Xor;
4126
4127 if (Instruction *X = foldComplexAndOrPatterns(I, Builder))
4128 return X;
4129
4130 if (Instruction *X = foldIntegerPackFromVector(I, Builder, DL))
4131 return X;
4132
4133 // (A & B) | (C & D) -> A ^ D where A == ~C && B == ~D
4134 // (A & B) | (C & D) -> A ^ C where A == ~D && B == ~C
4135 if (Value *V = foldOrOfInversions(I, Builder))
4136 return replaceInstUsesWith(I, V);
4137
4138 // (A&B)|(A&C) -> A&(B|C) etc
4139 if (Value *V = foldUsingDistributiveLaws(I))
4140 return replaceInstUsesWith(I, V);
4141
4142 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
4143 Type *Ty = I.getType();
4144 if (Ty->isIntOrIntVectorTy(BitWidth: 1)) {
4145 if (auto *SI0 = dyn_cast<SelectInst>(Val: Op0)) {
4146 if (auto *R =
4147 foldAndOrOfSelectUsingImpliedCond(Op: Op1, SI&: *SI0, /* IsAnd */ false))
4148 return R;
4149 }
4150 if (auto *SI1 = dyn_cast<SelectInst>(Val: Op1)) {
4151 if (auto *R =
4152 foldAndOrOfSelectUsingImpliedCond(Op: Op0, SI&: *SI1, /* IsAnd */ false))
4153 return R;
4154 }
4155 }
4156
4157 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
4158 return FoldedLogic;
4159
4160 if (Instruction *FoldedLogic = foldBinOpSelectBinOp(Op&: I))
4161 return FoldedLogic;
4162
4163 if (Instruction *BitOp = matchBSwapOrBitReverse(I, /*MatchBSwaps*/ true,
4164 /*MatchBitReversals*/ true))
4165 return BitOp;
4166
4167 if (Instruction *Funnel = matchFunnelShift(Or&: I, IC&: *this))
4168 return Funnel;
4169
4170 if (Value *Concat = matchOrConcat(Or&: I, Builder))
4171 return replaceInstUsesWith(I, V: Concat);
4172
4173 if (Instruction *R = foldBinOpShiftWithShift(I))
4174 return R;
4175
4176 if (Instruction *R = tryFoldInstWithCtpopWithNot(I: &I))
4177 return R;
4178
4179 if (cast<PossiblyDisjointInst>(Val&: I).isDisjoint()) {
4180 if (Instruction *R =
4181 foldAddLikeCommutative(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
4182 /*NSW=*/true, /*NUW=*/true))
4183 return R;
4184 if (Instruction *R =
4185 foldAddLikeCommutative(LHS: I.getOperand(i_nocapture: 1), RHS: I.getOperand(i_nocapture: 0),
4186 /*NSW=*/true, /*NUW=*/true))
4187 return R;
4188
4189 if (Value *Res = foldDisjointOr(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1)))
4190 return replaceInstUsesWith(I, V: Res);
4191
4192 if (Value *Res = reassociateDisjointOr(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1)))
4193 return replaceInstUsesWith(I, V: Res);
4194 }
4195
4196 Value *X, *Y;
4197 const APInt *CV;
4198 if (match(V: &I, P: m_c_Or(L: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: X), R: m_APInt(Res&: CV))), R: m_Value(V&: Y))) &&
4199 !CV->isAllOnes() && MaskedValueIsZero(V: Y, Mask: *CV, CxtI: &I)) {
4200 // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 0
4201 // The check for a 'not' op is for efficiency (if Y is known zero --> ~X).
4202 Value *Or = Builder.CreateOr(LHS: X, RHS: Y);
4203 return BinaryOperator::CreateXor(V1: Or, V2: ConstantInt::get(Ty, V: *CV));
4204 }
4205
4206 // If the operands have no common bits set:
4207 // or (mul X, Y), X --> add (mul X, Y), X --> mul X, (Y + 1)
4208 if (match(V: &I, P: m_c_DisjointOr(L: m_OneUse(SubPattern: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Y))),
4209 R: m_Deferred(V: X)))) {
4210 Value *IncrementY = Builder.CreateAdd(LHS: Y, RHS: ConstantInt::get(Ty, V: 1));
4211 return BinaryOperator::CreateMul(V1: X, V2: IncrementY);
4212 }
4213
4214 // Canonicalization to achieve lowering to Bit Manipulation Instructions (BMI)
4215 // ~X | (X-1) => ~(X & -X)
4216 Value *Op;
4217 if (match(V: &I, P: m_c_Or(L: m_OneUse(SubPattern: m_Not(V: m_Value(V&: Op))),
4218 R: m_OneUse(SubPattern: m_Add(L: m_Deferred(V: Op), R: m_AllOnes()))))) {
4219 Value *NegX = Builder.CreateNeg(V: Op);
4220 Value *And = Builder.CreateAnd(LHS: Op, RHS: NegX);
4221 return BinaryOperator::CreateNot(Op: And);
4222 }
4223
4224 // (C && A) || (C && B) => select C, A, B (and similar cases)
4225 //
4226 // Note: This is the same transformation used in `foldSelectOfBools`,
4227 // except that it's an `or` instead of `select`.
4228 if (I.getType()->isIntOrIntVectorTy(BitWidth: 1) &&
4229 (Op0->hasOneUse() || Op1->hasOneUse())) {
4230 if (Instruction *V = FoldOrOfLogicalAnds(Op0, Op1)) {
4231 return V;
4232 }
4233 }
4234
4235 // (A & C) | (B & D)
4236 Value *A, *B, *C, *D;
4237 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: C))) &&
4238 match(V: Op1, P: m_And(L: m_Value(V&: B), R: m_Value(V&: D)))) {
4239
4240 // (A & C0) | (B & C1)
4241 const APInt *C0, *C1;
4242 if (match(V: C, P: m_APInt(Res&: C0)) && match(V: D, P: m_APInt(Res&: C1))) {
4243 Value *X;
4244 if (*C0 == ~*C1) {
4245 // ((X | B) & MaskC) | (B & ~MaskC) -> (X & MaskC) | B
4246 if (match(V: A, P: m_c_Or(L: m_Value(V&: X), R: m_Specific(V: B))))
4247 return BinaryOperator::CreateOr(V1: Builder.CreateAnd(LHS: X, RHS: *C0), V2: B);
4248 // (A & MaskC) | ((X | A) & ~MaskC) -> (X & ~MaskC) | A
4249 if (match(V: B, P: m_c_Or(L: m_Specific(V: A), R: m_Value(V&: X))))
4250 return BinaryOperator::CreateOr(V1: Builder.CreateAnd(LHS: X, RHS: *C1), V2: A);
4251
4252 // ((X ^ B) & MaskC) | (B & ~MaskC) -> (X & MaskC) ^ B
4253 if (match(V: A, P: m_c_Xor(L: m_Value(V&: X), R: m_Specific(V: B))))
4254 return BinaryOperator::CreateXor(V1: Builder.CreateAnd(LHS: X, RHS: *C0), V2: B);
4255 // (A & MaskC) | ((X ^ A) & ~MaskC) -> (X & ~MaskC) ^ A
4256 if (match(V: B, P: m_c_Xor(L: m_Specific(V: A), R: m_Value(V&: X))))
4257 return BinaryOperator::CreateXor(V1: Builder.CreateAnd(LHS: X, RHS: *C1), V2: A);
4258 }
4259
4260 if ((*C0 & *C1).isZero()) {
4261 // ((X | B) & C0) | (B & C1) --> (X | B) & (C0 | C1)
4262 // iff (C0 & C1) == 0 and (X & ~C0) == 0
4263 if (match(V: A, P: m_c_Or(L: m_Value(V&: X), R: m_Specific(V: B))) &&
4264 MaskedValueIsZero(V: X, Mask: ~*C0, CxtI: &I)) {
4265 Constant *C01 = ConstantInt::get(Ty, V: *C0 | *C1);
4266 return BinaryOperator::CreateAnd(V1: A, V2: C01);
4267 }
4268 // (A & C0) | ((X | A) & C1) --> (X | A) & (C0 | C1)
4269 // iff (C0 & C1) == 0 and (X & ~C1) == 0
4270 if (match(V: B, P: m_c_Or(L: m_Value(V&: X), R: m_Specific(V: A))) &&
4271 MaskedValueIsZero(V: X, Mask: ~*C1, CxtI: &I)) {
4272 Constant *C01 = ConstantInt::get(Ty, V: *C0 | *C1);
4273 return BinaryOperator::CreateAnd(V1: B, V2: C01);
4274 }
4275 // ((X | C2) & C0) | ((X | C3) & C1) --> (X | C2 | C3) & (C0 | C1)
4276 // iff (C0 & C1) == 0 and (C2 & ~C0) == 0 and (C3 & ~C1) == 0.
4277 const APInt *C2, *C3;
4278 if (match(V: A, P: m_Or(L: m_Value(V&: X), R: m_APInt(Res&: C2))) &&
4279 match(V: B, P: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C3))) &&
4280 (*C2 & ~*C0).isZero() && (*C3 & ~*C1).isZero()) {
4281 Value *Or = Builder.CreateOr(LHS: X, RHS: *C2 | *C3, Name: "bitfield");
4282 Constant *C01 = ConstantInt::get(Ty, V: *C0 | *C1);
4283 return BinaryOperator::CreateAnd(V1: Or, V2: C01);
4284 }
4285 }
4286 }
4287
4288 // Don't try to form a select if it's unlikely that we'll get rid of at
4289 // least one of the operands. A select is generally more expensive than the
4290 // 'or' that it is replacing.
4291 if (Op0->hasOneUse() || Op1->hasOneUse()) {
4292 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.
4293 if (Value *V = matchSelectFromAndOr(A, B: C, C: B, D))
4294 return replaceInstUsesWith(I, V);
4295 if (Value *V = matchSelectFromAndOr(A, B: C, C: D, D: B))
4296 return replaceInstUsesWith(I, V);
4297 if (Value *V = matchSelectFromAndOr(A: C, B: A, C: B, D))
4298 return replaceInstUsesWith(I, V);
4299 if (Value *V = matchSelectFromAndOr(A: C, B: A, C: D, D: B))
4300 return replaceInstUsesWith(I, V);
4301 if (Value *V = matchSelectFromAndOr(A: B, B: D, C: A, D: C))
4302 return replaceInstUsesWith(I, V);
4303 if (Value *V = matchSelectFromAndOr(A: B, B: D, C, D: A))
4304 return replaceInstUsesWith(I, V);
4305 if (Value *V = matchSelectFromAndOr(A: D, B, C: A, D: C))
4306 return replaceInstUsesWith(I, V);
4307 if (Value *V = matchSelectFromAndOr(A: D, B, C, D: A))
4308 return replaceInstUsesWith(I, V);
4309 }
4310 }
4311
4312 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: C))) &&
4313 match(V: Op1, P: m_Not(V: m_Or(L: m_Value(V&: B), R: m_Value(V&: D)))) &&
4314 (Op0->hasOneUse() || Op1->hasOneUse())) {
4315 // (Cond & C) | ~(Cond | D) -> Cond ? C : ~D
4316 if (Value *V = matchSelectFromAndOr(A, B: C, C: B, D, InvertFalseVal: true))
4317 return replaceInstUsesWith(I, V);
4318 if (Value *V = matchSelectFromAndOr(A, B: C, C: D, D: B, InvertFalseVal: true))
4319 return replaceInstUsesWith(I, V);
4320 if (Value *V = matchSelectFromAndOr(A: C, B: A, C: B, D, InvertFalseVal: true))
4321 return replaceInstUsesWith(I, V);
4322 if (Value *V = matchSelectFromAndOr(A: C, B: A, C: D, D: B, InvertFalseVal: true))
4323 return replaceInstUsesWith(I, V);
4324 }
4325
4326 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
4327 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))))
4328 if (match(V: Op1,
4329 P: m_c_Xor(L: m_c_Xor(L: m_Specific(V: B), R: m_Value(V&: C)), R: m_Specific(V: A))) ||
4330 match(V: Op1, P: m_c_Xor(L: m_c_Xor(L: m_Specific(V: A), R: m_Value(V&: C)), R: m_Specific(V: B))))
4331 return BinaryOperator::CreateOr(V1: Op0, V2: C);
4332
4333 // ((B ^ C) ^ A) | (A ^ B) -> (A ^ B) | C
4334 if (match(V: Op1, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))))
4335 if (match(V: Op0,
4336 P: m_c_Xor(L: m_c_Xor(L: m_Specific(V: B), R: m_Value(V&: C)), R: m_Specific(V: A))) ||
4337 match(V: Op0, P: m_c_Xor(L: m_c_Xor(L: m_Specific(V: A), R: m_Value(V&: C)), R: m_Specific(V: B))))
4338 return BinaryOperator::CreateOr(V1: Op1, V2: C);
4339
4340 if (Instruction *DeMorgan = matchDeMorgansLaws(I, IC&: *this))
4341 return DeMorgan;
4342
4343 // Canonicalize xor to the RHS.
4344 bool SwappedForXor = false;
4345 if (match(V: Op0, P: m_Xor(L: m_Value(), R: m_Value()))) {
4346 std::swap(a&: Op0, b&: Op1);
4347 SwappedForXor = true;
4348 }
4349
4350 if (match(V: Op1, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B)))) {
4351 // (A | ?) | (A ^ B) --> (A | ?) | B
4352 // (B | ?) | (A ^ B) --> (B | ?) | A
4353 if (match(V: Op0, P: m_c_Or(L: m_Specific(V: A), R: m_Value())))
4354 return BinaryOperator::CreateOr(V1: Op0, V2: B);
4355 if (match(V: Op0, P: m_c_Or(L: m_Specific(V: B), R: m_Value())))
4356 return BinaryOperator::CreateOr(V1: Op0, V2: A);
4357
4358 // (A & B) | (A ^ B) --> A | B
4359 // (B & A) | (A ^ B) --> A | B
4360 if (match(V: Op0, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
4361 return BinaryOperator::CreateOr(V1: A, V2: B);
4362
4363 // ~A | (A ^ B) --> ~(A & B)
4364 // ~B | (A ^ B) --> ~(A & B)
4365 // The swap above should always make Op0 the 'not'.
4366 if ((Op0->hasOneUse() || Op1->hasOneUse()) &&
4367 (match(V: Op0, P: m_Not(V: m_Specific(V: A))) || match(V: Op0, P: m_Not(V: m_Specific(V: B)))))
4368 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: A, RHS: B));
4369
4370 // Same as above, but peek through an 'and' to the common operand:
4371 // ~(A & ?) | (A ^ B) --> ~((A & ?) & B)
4372 // ~(B & ?) | (A ^ B) --> ~((B & ?) & A)
4373 Instruction *And;
4374 if ((Op0->hasOneUse() || Op1->hasOneUse()) &&
4375 match(V: Op0,
4376 P: m_Not(V: m_Instruction(I&: And, P: m_c_And(L: m_Specific(V: A), R: m_Value())))))
4377 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: And, RHS: B));
4378 if ((Op0->hasOneUse() || Op1->hasOneUse()) &&
4379 match(V: Op0,
4380 P: m_Not(V: m_Instruction(I&: And, P: m_c_And(L: m_Specific(V: B), R: m_Value())))))
4381 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: And, RHS: A));
4382
4383 // (~A | C) | (A ^ B) --> ~(A & B) | C
4384 // (~B | C) | (A ^ B) --> ~(A & B) | C
4385 if (Op0->hasOneUse() && Op1->hasOneUse() &&
4386 (match(V: Op0, P: m_c_Or(L: m_Not(V: m_Specific(V: A)), R: m_Value(V&: C))) ||
4387 match(V: Op0, P: m_c_Or(L: m_Not(V: m_Specific(V: B)), R: m_Value(V&: C))))) {
4388 Value *Nand = Builder.CreateNot(V: Builder.CreateAnd(LHS: A, RHS: B), Name: "nand");
4389 return BinaryOperator::CreateOr(V1: Nand, V2: C);
4390 }
4391 }
4392
4393 if (SwappedForXor)
4394 std::swap(a&: Op0, b&: Op1);
4395
4396 if (Value *Res =
4397 foldBooleanAndOr(LHS: Op0, RHS: Op1, I, /*IsAnd=*/false, /*IsLogical=*/false))
4398 return replaceInstUsesWith(I, V: Res);
4399
4400 if (match(V: Op1, P: m_OneUse(SubPattern: m_LogicalOr(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
4401 bool IsLogical = isa<SelectInst>(Val: Op1);
4402 if (auto *V = reassociateBooleanAndOr(LHS: Op0, X, Y, I, /*IsAnd=*/false,
4403 /*RHSIsLogical=*/IsLogical))
4404 return replaceInstUsesWith(I, V);
4405 }
4406 if (match(V: Op0, P: m_OneUse(SubPattern: m_LogicalOr(L: m_Value(V&: X), R: m_Value(V&: Y))))) {
4407 bool IsLogical = isa<SelectInst>(Val: Op0);
4408 if (auto *V = reassociateBooleanAndOr(LHS: Op1, X, Y, I, /*IsAnd=*/false,
4409 /*RHSIsLogical=*/IsLogical))
4410 return replaceInstUsesWith(I, V);
4411 }
4412
4413 if (Instruction *FoldedFCmps = reassociateFCmps(BO&: I, Builder))
4414 return FoldedFCmps;
4415
4416 if (Instruction *CastedOr = foldCastedBitwiseLogic(I))
4417 return CastedOr;
4418
4419 if (Instruction *Sel = foldBinopOfSextBoolToSelect(I))
4420 return Sel;
4421
4422 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.
4423 // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold
4424 // with binop identity constant. But creating a select with non-constant
4425 // arm may not be reversible due to poison semantics. Is that a good
4426 // canonicalization?
4427 if (match(V: &I, P: m_c_Or(L: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: A))), R: m_Value(V&: B))) &&
4428 A->getType()->isIntOrIntVectorTy(BitWidth: 1))
4429 return createSelectInstWithUnknownProfile(
4430 C: A, S1: ConstantInt::getAllOnesValue(Ty), S2: B);
4431
4432 // Note: If we've gotten to the point of visiting the outer OR, then the
4433 // inner one couldn't be simplified. If it was a constant, then it won't
4434 // be simplified by a later pass either, so we try swapping the inner/outer
4435 // ORs in the hopes that we'll be able to simplify it this way.
4436 // (X|C) | V --> (X|V) | C
4437 // Pass the disjoint flag in the following two patterns:
4438 // 1. or-disjoint (or-disjoint X, C), V -->
4439 // or-disjoint (or-disjoint X, V), C
4440 //
4441 // 2. or-disjoint (or X, C), V -->
4442 // or (or-disjoint X, V), C
4443 ConstantInt *CI;
4444 if (Op0->hasOneUse() && !match(V: Op1, P: m_ConstantInt()) &&
4445 match(V: Op0, P: m_Or(L: m_Value(V&: A), R: m_ConstantInt(CI)))) {
4446 bool IsDisjointOuter = cast<PossiblyDisjointInst>(Val&: I).isDisjoint();
4447 bool IsDisjointInner = cast<PossiblyDisjointInst>(Val: Op0)->isDisjoint();
4448 Value *Inner = Builder.CreateOr(LHS: A, RHS: Op1, Name: "", /*IsDisjoint=*/IsDisjointOuter);
4449 Inner->takeName(V: Op0);
4450 return IsDisjointOuter && IsDisjointInner
4451 ? BinaryOperator::CreateDisjointOr(V1: Inner, V2: CI)
4452 : BinaryOperator::CreateOr(V1: Inner, V2: CI);
4453 }
4454
4455 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
4456 // Since this OR statement hasn't been optimized further yet, we hope
4457 // that this transformation will allow the new ORs to be optimized.
4458 {
4459 Value *X = nullptr, *Y = nullptr;
4460 if (Op0->hasOneUse() && Op1->hasOneUse() &&
4461 match(V: Op0, P: m_Select(C: m_Value(V&: X), L: m_Value(V&: A), R: m_Value(V&: B))) &&
4462 match(V: Op1, P: m_Select(C: m_Value(V&: Y), L: m_Value(V&: C), R: m_Value(V&: D))) && X == Y) {
4463 Value *orTrue = Builder.CreateOr(LHS: A, RHS: C);
4464 Value *orFalse = Builder.CreateOr(LHS: B, RHS: D);
4465 return SelectInst::Create(C: X, S1: orTrue, S2: orFalse);
4466 }
4467 }
4468
4469 // or(ashr(subNSW(Y, X), ScalarSizeInBits(Y) - 1), X) --> X s> Y ? -1 : X.
4470 {
4471 Value *X, *Y;
4472 if (match(V: &I, P: m_c_Or(L: m_OneUse(SubPattern: m_AShr(
4473 L: m_NSWSub(L: m_Value(V&: Y), R: m_Value(V&: X)),
4474 R: m_SpecificInt(V: Ty->getScalarSizeInBits() - 1))),
4475 R: m_Deferred(V: X)))) {
4476 Value *NewICmpInst = Builder.CreateICmpSGT(LHS: X, RHS: Y);
4477 Value *AllOnes = ConstantInt::getAllOnesValue(Ty);
4478 return createSelectInstWithUnknownProfile(C: NewICmpInst, S1: AllOnes, S2: X);
4479 }
4480 }
4481
4482 {
4483 // ((A & B) ^ A) | ((A & B) ^ B) -> A ^ B
4484 // (A ^ (A & B)) | (B ^ (A & B)) -> A ^ B
4485 // ((A & B) ^ B) | ((A & B) ^ A) -> A ^ B
4486 // (B ^ (A & B)) | (A ^ (A & B)) -> A ^ B
4487 const auto TryXorOpt = [&](Value *Lhs, Value *Rhs) -> Instruction * {
4488 if (match(V: Lhs, P: m_c_Xor(L: m_And(L: m_Value(V&: A), R: m_Value(V&: B)), R: m_Deferred(V: A))) &&
4489 match(V: Rhs,
4490 P: m_c_Xor(L: m_And(L: m_Specific(V: A), R: m_Specific(V: B)), R: m_Specific(V: B)))) {
4491 return BinaryOperator::CreateXor(V1: A, V2: B);
4492 }
4493 return nullptr;
4494 };
4495
4496 if (Instruction *Result = TryXorOpt(Op0, Op1))
4497 return Result;
4498 if (Instruction *Result = TryXorOpt(Op1, Op0))
4499 return Result;
4500 }
4501
4502 if (Instruction *V =
4503 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
4504 return V;
4505
4506 CmpPredicate Pred;
4507 Value *Mul, *Ov, *MulIsNotZero, *UMulWithOv;
4508 // Check if the OR weakens the overflow condition for umul.with.overflow by
4509 // treating any non-zero result as overflow. In that case, we overflow if both
4510 // umul.with.overflow operands are != 0, as in that case the result can only
4511 // be 0, iff the multiplication overflows.
4512 if (match(V: &I, P: m_c_Or(L: m_Value(V&: Ov, P: m_ExtractValue<1>(V: m_Value(V&: UMulWithOv))),
4513 R: m_Value(V&: MulIsNotZero,
4514 P: m_SpecificICmp(
4515 MatchPred: ICmpInst::ICMP_NE,
4516 L: m_Value(V&: Mul, P: m_ExtractValue<0>(
4517 V: m_Deferred(V: UMulWithOv))),
4518 R: m_ZeroInt())))) &&
4519 (Ov->hasOneUse() || (MulIsNotZero->hasOneUse() && Mul->hasOneUse()))) {
4520 Value *A, *B;
4521 if (match(V: UMulWithOv, P: m_Intrinsic<Intrinsic::umul_with_overflow>(
4522 Op0: m_Value(V&: A), Op1: m_Value(V&: B)))) {
4523 Value *NotNullA = Builder.CreateIsNotNull(Arg: A);
4524 Value *NotNullB = Builder.CreateIsNotNull(Arg: B);
4525 return BinaryOperator::CreateAnd(V1: NotNullA, V2: NotNullB);
4526 }
4527 }
4528
4529 /// Res, Overflow = xxx_with_overflow X, C1
4530 /// Try to canonicalize the pattern "Overflow | icmp pred Res, C2" into
4531 /// "Overflow | icmp pred X, C2 +/- C1".
4532 const WithOverflowInst *WO;
4533 const Value *WOV;
4534 const APInt *C1, *C2;
4535 if (match(V: &I, P: m_c_Or(L: m_Value(V&: Ov, P: m_ExtractValue<1>(
4536 V: m_Value(V&: WOV, P: m_WithOverflowInst(I&: WO)))),
4537 R: m_OneUse(SubPattern: m_ICmp(Pred, L: m_ExtractValue<0>(V: m_Deferred(V: WOV)),
4538 R: m_APInt(Res&: C2))))) &&
4539 (WO->getBinaryOp() == Instruction::Add ||
4540 WO->getBinaryOp() == Instruction::Sub) &&
4541 (ICmpInst::isEquality(P: Pred) ||
4542 WO->isSigned() == ICmpInst::isSigned(Pred)) &&
4543 match(V: WO->getRHS(), P: m_APInt(Res&: C1))) {
4544 bool Overflow;
4545 APInt NewC = WO->getBinaryOp() == Instruction::Add
4546 ? (ICmpInst::isSigned(Pred) ? C2->ssub_ov(RHS: *C1, Overflow)
4547 : C2->usub_ov(RHS: *C1, Overflow))
4548 : (ICmpInst::isSigned(Pred) ? C2->sadd_ov(RHS: *C1, Overflow)
4549 : C2->uadd_ov(RHS: *C1, Overflow));
4550 if (!Overflow || ICmpInst::isEquality(P: Pred)) {
4551 Value *NewCmp = Builder.CreateICmp(
4552 P: Pred, LHS: WO->getLHS(), RHS: ConstantInt::get(Ty: WO->getLHS()->getType(), V: NewC));
4553 return BinaryOperator::CreateOr(V1: Ov, V2: NewCmp);
4554 }
4555 }
4556
4557 // Try to fold the pattern "Overflow | icmp pred Res, C2" into a single
4558 // comparison instruction for umul.with.overflow.
4559 if (Value *R = foldOrUnsignedUMulOverflowICmp(I, Builder, DL))
4560 return replaceInstUsesWith(I, V: R);
4561
4562 // (~x) | y --> ~(x & (~y)) iff that gets rid of inversions
4563 if (sinkNotIntoOtherHandOfLogicalOp(I))
4564 return &I;
4565
4566 // Improve "get low bit mask up to and including bit X" pattern:
4567 // (1 << X) | ((1 << X) + -1) --> -1 l>> (bitwidth(x) - 1 - X)
4568 if (match(V: &I, P: m_c_Or(L: m_Add(L: m_Shl(L: m_One(), R: m_Value(V&: X)), R: m_AllOnes()),
4569 R: m_Shl(L: m_One(), R: m_Deferred(V: X)))) &&
4570 match(V: &I, P: m_c_Or(L: m_OneUse(SubPattern: m_Value()), R: m_Value()))) {
4571 Value *Sub = Builder.CreateSub(
4572 LHS: ConstantInt::get(Ty, V: Ty->getScalarSizeInBits() - 1), RHS: X);
4573 return BinaryOperator::CreateLShr(V1: Constant::getAllOnesValue(Ty), V2: Sub);
4574 }
4575
4576 // An or recurrence w/loop invariant step is equivelent to (or start, step)
4577 PHINode *PN = nullptr;
4578 Value *Start = nullptr, *Step = nullptr;
4579 if (matchSimpleRecurrence(I: &I, P&: PN, Start, Step) && DT.dominates(Def: Step, User: PN))
4580 return replaceInstUsesWith(I, V: Builder.CreateOr(LHS: Start, RHS: Step));
4581
4582 // (A & B) | (C | D) or (C | D) | (A & B)
4583 // Can be combined if C or D is of type (A/B & X)
4584 if (match(V: &I, P: m_c_Or(L: m_OneUse(SubPattern: m_And(L: m_Value(V&: A), R: m_Value(V&: B))),
4585 R: m_OneUse(SubPattern: m_Or(L: m_Value(V&: C), R: m_Value(V&: D)))))) {
4586 // (A & B) | (C | ?) -> C | (? | (A & B))
4587 // (A & B) | (C | ?) -> C | (? | (A & B))
4588 // (A & B) | (C | ?) -> C | (? | (A & B))
4589 // (A & B) | (C | ?) -> C | (? | (A & B))
4590 // (C | ?) | (A & B) -> C | (? | (A & B))
4591 // (C | ?) | (A & B) -> C | (? | (A & B))
4592 // (C | ?) | (A & B) -> C | (? | (A & B))
4593 // (C | ?) | (A & B) -> C | (? | (A & B))
4594 if (match(V: D, P: m_OneUse(SubPattern: m_c_And(L: m_Specific(V: A), R: m_Value()))) ||
4595 match(V: D, P: m_OneUse(SubPattern: m_c_And(L: m_Specific(V: B), R: m_Value()))))
4596 return BinaryOperator::CreateOr(
4597 V1: C, V2: Builder.CreateOr(LHS: D, RHS: Builder.CreateAnd(LHS: A, RHS: B)));
4598 // (A & B) | (? | D) -> (? | (A & B)) | D
4599 // (A & B) | (? | D) -> (? | (A & B)) | D
4600 // (A & B) | (? | D) -> (? | (A & B)) | D
4601 // (A & B) | (? | D) -> (? | (A & B)) | D
4602 // (? | D) | (A & B) -> (? | (A & B)) | D
4603 // (? | D) | (A & B) -> (? | (A & B)) | D
4604 // (? | D) | (A & B) -> (? | (A & B)) | D
4605 // (? | D) | (A & B) -> (? | (A & B)) | D
4606 if (match(V: C, P: m_OneUse(SubPattern: m_c_And(L: m_Specific(V: A), R: m_Value()))) ||
4607 match(V: C, P: m_OneUse(SubPattern: m_c_And(L: m_Specific(V: B), R: m_Value()))))
4608 return BinaryOperator::CreateOr(
4609 V1: Builder.CreateOr(LHS: C, RHS: Builder.CreateAnd(LHS: A, RHS: B)), V2: D);
4610 }
4611
4612 if (Instruction *R = reassociateForUses(BO&: I, Builder))
4613 return R;
4614
4615 if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))
4616 return Canonicalized;
4617
4618 if (Instruction *Folded = foldLogicOfIsFPClass(BO&: I, Op0, Op1))
4619 return Folded;
4620
4621 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
4622 return Res;
4623
4624 // If we are setting the sign bit of a floating-point value, convert
4625 // this to fneg(fabs), then cast back to integer.
4626 //
4627 // If the result isn't immediately cast back to a float, this will increase
4628 // the number of instructions. This is still probably a better canonical form
4629 // as it enables FP value tracking.
4630 //
4631 // Assumes any IEEE-represented type has the sign bit in the high bit.
4632 //
4633 // This is generous interpretation of noimplicitfloat, this is not a true
4634 // floating-point operation.
4635 Value *CastOp;
4636 if (match(V: Op0, P: m_ElementWiseBitCast(Op: m_Value(V&: CastOp))) &&
4637 match(V: Op1, P: m_SignMask()) &&
4638 !Builder.GetInsertBlock()->getParent()->hasFnAttribute(
4639 Kind: Attribute::NoImplicitFloat)) {
4640 Type *EltTy = CastOp->getType()->getScalarType();
4641 if (EltTy->isFloatingPointTy() &&
4642 APFloat::hasSignBitInMSB(EltTy->getFltSemantics())) {
4643 Value *FAbs = Builder.CreateFAbs(V: CastOp);
4644 Value *FNegFAbs = Builder.CreateFNeg(V: FAbs);
4645 return new BitCastInst(FNegFAbs, I.getType());
4646 }
4647 }
4648
4649 // (X & C1) | C2 -> X & (C1 | C2) iff (X & C2) == C2
4650 if (match(V: Op0, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) &&
4651 match(V: Op1, P: m_APInt(Res&: C2))) {
4652 KnownBits KnownX = computeKnownBits(V: X, CxtI: &I);
4653 if ((KnownX.One & *C2) == *C2)
4654 return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty, V: *C1 | *C2));
4655 }
4656
4657 if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))
4658 return Res;
4659
4660 if (Value *V =
4661 simplifyAndOrWithOpReplaced(V: Op0, Op: Op1, RepOp: Constant::getNullValue(Ty),
4662 /*SimplifyOnly*/ false, IC&: *this))
4663 return BinaryOperator::CreateOr(V1: V, V2: Op1);
4664 if (Value *V =
4665 simplifyAndOrWithOpReplaced(V: Op1, Op: Op0, RepOp: Constant::getNullValue(Ty),
4666 /*SimplifyOnly*/ false, IC&: *this))
4667 return BinaryOperator::CreateOr(V1: Op0, V2: V);
4668
4669 if (cast<PossiblyDisjointInst>(Val&: I).isDisjoint())
4670 if (Value *V = SimplifyAddWithRemainder(I))
4671 return replaceInstUsesWith(I, V);
4672
4673 if (Value *Res = FoldOrOfSelectSmaxToAbs(I, Builder))
4674 return replaceInstUsesWith(I, V: Res);
4675
4676 // signum: or (ashr X, BW-1), zext (icmp ne|sgt X, 0) --> scmp(X, 0)
4677 // The ashr already supplies -1 for negative X, so any predicate that
4678 // produces 1 for positive X and 0 for X == 0 yields the same result here.
4679 {
4680 Value *X;
4681 CmpPredicate SignPred;
4682 unsigned BitWidth = Ty->getScalarSizeInBits();
4683 if (match(V: &I,
4684 P: m_c_Or(L: m_AShr(L: m_Value(V&: X), R: m_SpecificIntAllowPoison(V: BitWidth - 1)),
4685 R: m_ZExt(Op: m_ICmp(Pred&: SignPred, L: m_Deferred(V: X), R: m_ZeroInt())))) &&
4686 (SignPred == ICmpInst::ICMP_NE || SignPred == ICmpInst::ICMP_SGT) &&
4687 (Op0->hasOneUse() || Op1->hasOneUse()))
4688 return replaceInstUsesWith(
4689 I, V: Builder.CreateIntrinsic(RetTy: Ty, ID: Intrinsic::scmp,
4690 Args: {X, Constant::getNullValue(Ty)}));
4691 }
4692
4693 return nullptr;
4694}
4695
4696/// A ^ B can be specified using other logic ops in a variety of patterns. We
4697/// can fold these early and efficiently by morphing an existing instruction.
4698static Instruction *foldXorToXor(BinaryOperator &I,
4699 InstCombiner::BuilderTy &Builder) {
4700 assert(I.getOpcode() == Instruction::Xor);
4701 Value *Op0 = I.getOperand(i_nocapture: 0);
4702 Value *Op1 = I.getOperand(i_nocapture: 1);
4703 Value *A, *B;
4704
4705 // There are 4 commuted variants for each of the basic patterns.
4706
4707 // (A & B) ^ (A | B) -> A ^ B
4708 // (A & B) ^ (B | A) -> A ^ B
4709 // (A | B) ^ (A & B) -> A ^ B
4710 // (A | B) ^ (B & A) -> A ^ B
4711 if (match(V: &I, P: m_c_Xor(L: m_And(L: m_Value(V&: A), R: m_Value(V&: B)),
4712 R: m_c_Or(L: m_Deferred(V: A), R: m_Deferred(V: B)))))
4713 return BinaryOperator::CreateXor(V1: A, V2: B);
4714
4715 // (A | ~B) ^ (~A | B) -> A ^ B
4716 // (~B | A) ^ (~A | B) -> A ^ B
4717 // (~A | B) ^ (A | ~B) -> A ^ B
4718 // (B | ~A) ^ (A | ~B) -> A ^ B
4719 if (match(V: &I, P: m_Xor(L: m_c_Or(L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B))),
4720 R: m_c_Or(L: m_Not(V: m_Deferred(V: A)), R: m_Deferred(V: B)))))
4721 return BinaryOperator::CreateXor(V1: A, V2: B);
4722
4723 // (A & ~B) ^ (~A & B) -> A ^ B
4724 // (~B & A) ^ (~A & B) -> A ^ B
4725 // (~A & B) ^ (A & ~B) -> A ^ B
4726 // (B & ~A) ^ (A & ~B) -> A ^ B
4727 if (match(V: &I, P: m_Xor(L: m_c_And(L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B))),
4728 R: m_c_And(L: m_Not(V: m_Deferred(V: A)), R: m_Deferred(V: B)))))
4729 return BinaryOperator::CreateXor(V1: A, V2: B);
4730
4731 // For the remaining cases we need to get rid of one of the operands.
4732 if (!Op0->hasOneUse() && !Op1->hasOneUse())
4733 return nullptr;
4734
4735 // (A | B) ^ ~(A & B) -> ~(A ^ B)
4736 // (A | B) ^ ~(B & A) -> ~(A ^ B)
4737 // (A & B) ^ ~(A | B) -> ~(A ^ B)
4738 // (A & B) ^ ~(B | A) -> ~(A ^ B)
4739 // Complexity sorting ensures the not will be on the right side.
4740 if ((match(V: Op0, P: m_Or(L: m_Value(V&: A), R: m_Value(V&: B))) &&
4741 match(V: Op1, P: m_Not(V: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))) ||
4742 (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
4743 match(V: Op1, P: m_Not(V: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))))
4744 return BinaryOperator::CreateNot(Op: Builder.CreateXor(LHS: A, RHS: B));
4745
4746 return nullptr;
4747}
4748
4749Value *InstCombinerImpl::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS,
4750 BinaryOperator &I) {
4751 assert(I.getOpcode() == Instruction::Xor && I.getOperand(0) == LHS &&
4752 I.getOperand(1) == RHS && "Should be 'xor' with these operands");
4753
4754 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
4755 Value *LHS0 = LHS->getOperand(i_nocapture: 0), *LHS1 = LHS->getOperand(i_nocapture: 1);
4756 Value *RHS0 = RHS->getOperand(i_nocapture: 0), *RHS1 = RHS->getOperand(i_nocapture: 1);
4757
4758 if (predicatesFoldable(P1: PredL, P2: PredR)) {
4759 if (LHS0 == RHS1 && LHS1 == RHS0) {
4760 std::swap(a&: LHS0, b&: LHS1);
4761 PredL = ICmpInst::getSwappedPredicate(pred: PredL);
4762 }
4763 if (LHS0 == RHS0 && LHS1 == RHS1) {
4764 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4765 unsigned Code = getICmpCode(Pred: PredL) ^ getICmpCode(Pred: PredR);
4766 bool IsSigned = LHS->isSigned() || RHS->isSigned();
4767 return getNewICmpValue(Code, Sign: IsSigned, LHS: LHS0, RHS: LHS1, Builder);
4768 }
4769 }
4770
4771 const APInt *LC, *RC;
4772 if (match(V: LHS1, P: m_APInt(Res&: LC)) && match(V: RHS1, P: m_APInt(Res&: RC)) &&
4773 LHS0->getType() == RHS0->getType() &&
4774 LHS0->getType()->isIntOrIntVectorTy()) {
4775 // Convert xor of signbit tests to signbit test of xor'd values:
4776 // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0
4777 // (X < 0) ^ (Y < 0) --> (X ^ Y) < 0
4778 // (X > -1) ^ (Y < 0) --> (X ^ Y) > -1
4779 // (X < 0) ^ (Y > -1) --> (X ^ Y) > -1
4780 bool TrueIfSignedL, TrueIfSignedR;
4781 if ((LHS->hasOneUse() || RHS->hasOneUse()) &&
4782 isSignBitCheck(Pred: PredL, RHS: *LC, TrueIfSigned&: TrueIfSignedL) &&
4783 isSignBitCheck(Pred: PredR, RHS: *RC, TrueIfSigned&: TrueIfSignedR)) {
4784 Value *XorLR = Builder.CreateXor(LHS: LHS0, RHS: RHS0);
4785 return TrueIfSignedL == TrueIfSignedR ? Builder.CreateIsNeg(Arg: XorLR) :
4786 Builder.CreateIsNotNeg(Arg: XorLR);
4787 }
4788
4789 // Fold (icmp pred1 X, C1) ^ (icmp pred2 X, C2)
4790 // into a single comparison using range-based reasoning.
4791 if (LHS0 == RHS0) {
4792 ConstantRange CR1 = ConstantRange::makeExactICmpRegion(Pred: PredL, Other: *LC);
4793 ConstantRange CR2 = ConstantRange::makeExactICmpRegion(Pred: PredR, Other: *RC);
4794 auto CRUnion = CR1.exactUnionWith(CR: CR2);
4795 auto CRIntersect = CR1.exactIntersectWith(CR: CR2);
4796 if (CRUnion && CRIntersect)
4797 if (auto CR = CRUnion->exactIntersectWith(CR: CRIntersect->inverse())) {
4798 if (CR->isFullSet())
4799 return ConstantInt::getTrue(Ty: I.getType());
4800 if (CR->isEmptySet())
4801 return ConstantInt::getFalse(Ty: I.getType());
4802
4803 CmpInst::Predicate NewPred;
4804 APInt NewC, Offset;
4805 CR->getEquivalentICmp(Pred&: NewPred, RHS&: NewC, Offset);
4806
4807 if ((Offset.isZero() && (LHS->hasOneUse() || RHS->hasOneUse())) ||
4808 (LHS->hasOneUse() && RHS->hasOneUse())) {
4809 Value *NewV = LHS0;
4810 Type *Ty = LHS0->getType();
4811 if (!Offset.isZero())
4812 NewV = Builder.CreateAdd(LHS: NewV, RHS: ConstantInt::get(Ty, V: Offset));
4813 return Builder.CreateICmp(P: NewPred, LHS: NewV,
4814 RHS: ConstantInt::get(Ty, V: NewC));
4815 }
4816 }
4817 }
4818
4819 // Fold (icmp eq/ne (X & Pow2), 0) ^ (icmp eq/ne (Y & Pow2), 0) into
4820 // (icmp eq/ne ((X ^ Y) & Pow2), 0)
4821 Value *X, *Y, *Pow2;
4822 if (ICmpInst::isEquality(P: PredL) && ICmpInst::isEquality(P: PredR) &&
4823 LC->isZero() && RC->isZero() && LHS->hasOneUse() && RHS->hasOneUse() &&
4824 match(V: LHS0, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Pow2))) &&
4825 match(V: RHS0, P: m_And(L: m_Value(V&: Y), R: m_Specific(V: Pow2))) &&
4826 isKnownToBeAPowerOfTwo(V: Pow2, /*OrZero=*/true, CxtI: &I)) {
4827 Value *Xor = Builder.CreateXor(LHS: X, RHS: Y);
4828 Value *And = Builder.CreateAnd(LHS: Xor, RHS: Pow2);
4829 return Builder.CreateICmp(P: PredL == PredR ? ICmpInst::ICMP_NE
4830 : ICmpInst::ICMP_EQ,
4831 LHS: And, RHS: ConstantInt::getNullValue(Ty: Xor->getType()));
4832 }
4833 }
4834
4835 // Instead of trying to imitate the folds for and/or, decompose this 'xor'
4836 // into those logic ops. That is, try to turn this into an and-of-icmps
4837 // because we have many folds for that pattern.
4838 //
4839 // This is based on a truth table definition of xor:
4840 // X ^ Y --> (X | Y) & !(X & Y)
4841 if (Value *OrICmp = simplifyBinOp(Opcode: Instruction::Or, LHS, RHS, Q: SQ)) {
4842 // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y).
4843 // TODO: If OrICmp is false, the whole thing is false (InstSimplify?).
4844 if (Value *AndICmp = simplifyBinOp(Opcode: Instruction::And, LHS, RHS, Q: SQ)) {
4845 // TODO: Independently handle cases where the 'and' side is a constant.
4846 ICmpInst *X = nullptr, *Y = nullptr;
4847 if (OrICmp == LHS && AndICmp == RHS) {
4848 // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS --> X & !Y
4849 X = LHS;
4850 Y = RHS;
4851 }
4852 if (OrICmp == RHS && AndICmp == LHS) {
4853 // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS --> !Y & X
4854 X = RHS;
4855 Y = LHS;
4856 }
4857 if (X && Y && (Y->hasOneUse() || canFreelyInvertAllUsersOf(V: Y, IgnoredUser: &I))) {
4858 // Invert the predicate of 'Y', thus inverting its output.
4859 Y->setPredicate(Y->getInversePredicate());
4860 // So, are there other uses of Y?
4861 if (!Y->hasOneUse()) {
4862 // We need to adapt other uses of Y though. Get a value that matches
4863 // the original value of Y before inversion. While this increases
4864 // immediate instruction count, we have just ensured that all the
4865 // users are freely-invertible, so that 'not' *will* get folded away.
4866 BuilderTy::InsertPointGuard Guard(Builder);
4867 // Set insertion point to right after the Y.
4868 Builder.SetInsertPoint(TheBB: Y->getParent(), IP: ++(Y->getIterator()));
4869 Value *NotY = Builder.CreateNot(V: Y, Name: Y->getName() + ".not");
4870 // Replace all uses of Y (excluding the one in NotY!) with NotY.
4871 Worklist.pushUsersToWorkList(I&: *Y);
4872 Y->replaceUsesWithIf(New: NotY,
4873 ShouldReplace: [NotY](Use &U) { return U.getUser() != NotY; });
4874 }
4875 // All done.
4876 return Builder.CreateAnd(LHS, RHS);
4877 }
4878 }
4879 }
4880
4881 return nullptr;
4882}
4883
4884/// If we have a masked merge, in the canonical form of:
4885/// (assuming that A only has one use.)
4886/// | A | |B|
4887/// ((x ^ y) & M) ^ y
4888/// | D |
4889/// * If M is inverted:
4890/// | D |
4891/// ((x ^ y) & ~M) ^ y
4892/// We can canonicalize by swapping the final xor operand
4893/// to eliminate the 'not' of the mask.
4894/// ((x ^ y) & M) ^ x
4895/// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops
4896/// because that shortens the dependency chain and improves analysis:
4897/// (x & M) | (y & ~M)
4898static Instruction *visitMaskedMerge(BinaryOperator &I,
4899 InstCombiner::BuilderTy &Builder) {
4900 Value *B, *X, *D;
4901 Value *M;
4902 if (!match(V: &I, P: m_c_Xor(L: m_Value(V&: B),
4903 R: m_OneUse(SubPattern: m_c_And(
4904 L: m_Value(V&: D, P: m_c_Xor(L: m_Deferred(V: B), R: m_Value(V&: X))),
4905 R: m_Value(V&: M))))))
4906 return nullptr;
4907
4908 Value *NotM;
4909 if (match(V: M, P: m_Not(V: m_Value(V&: NotM)))) {
4910 // De-invert the mask and swap the value in B part.
4911 Value *NewA = Builder.CreateAnd(LHS: D, RHS: NotM);
4912 return BinaryOperator::CreateXor(V1: NewA, V2: X);
4913 }
4914
4915 Constant *C;
4916 if (D->hasOneUse() && match(V: M, P: m_Constant(C))) {
4917 // Propagating undef is unsafe. Clamp undef elements to -1.
4918 Type *EltTy = C->getType()->getScalarType();
4919 C = Constant::replaceUndefsWith(C, Replacement: ConstantInt::getAllOnesValue(Ty: EltTy));
4920 // Unfold.
4921 Value *LHS = Builder.CreateAnd(LHS: X, RHS: C);
4922 Value *NotC = Builder.CreateNot(V: C);
4923 Value *RHS = Builder.CreateAnd(LHS: B, RHS: NotC);
4924 return BinaryOperator::CreateOr(V1: LHS, V2: RHS);
4925 }
4926
4927 return nullptr;
4928}
4929
4930static Instruction *foldNotXor(BinaryOperator &I,
4931 InstCombiner::BuilderTy &Builder) {
4932 Value *X, *Y;
4933 // FIXME: one-use check is not needed in general, but currently we are unable
4934 // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182)
4935 if (!match(V: &I, P: m_Not(V: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: X), R: m_Value(V&: Y))))))
4936 return nullptr;
4937
4938 auto hasCommonOperand = [](Value *A, Value *B, Value *C, Value *D) {
4939 return A == C || A == D || B == C || B == D;
4940 };
4941
4942 Value *A, *B, *C, *D;
4943 // Canonicalize ~((A & B) ^ (A | ?)) -> (A & B) | ~(A | ?)
4944 // 4 commuted variants
4945 if (match(V: X, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
4946 match(V: Y, P: m_Or(L: m_Value(V&: C), R: m_Value(V&: D))) && hasCommonOperand(A, B, C, D)) {
4947 Value *NotY = Builder.CreateNot(V: Y);
4948 return BinaryOperator::CreateOr(V1: X, V2: NotY);
4949 };
4950
4951 // Canonicalize ~((A | ?) ^ (A & B)) -> (A & B) | ~(A | ?)
4952 // 4 commuted variants
4953 if (match(V: Y, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
4954 match(V: X, P: m_Or(L: m_Value(V&: C), R: m_Value(V&: D))) && hasCommonOperand(A, B, C, D)) {
4955 Value *NotX = Builder.CreateNot(V: X);
4956 return BinaryOperator::CreateOr(V1: Y, V2: NotX);
4957 };
4958
4959 return nullptr;
4960}
4961
4962/// Canonicalize a shifty way to code absolute value to the more common pattern
4963/// that uses negation and select.
4964static Instruction *canonicalizeAbs(BinaryOperator &Xor,
4965 InstCombiner::BuilderTy &Builder) {
4966 assert(Xor.getOpcode() == Instruction::Xor && "Expected an xor instruction.");
4967
4968 // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1.
4969 // We're relying on the fact that we only do this transform when the shift has
4970 // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase
4971 // instructions).
4972 Value *Op0 = Xor.getOperand(i_nocapture: 0), *Op1 = Xor.getOperand(i_nocapture: 1);
4973 if (Op0->hasNUses(N: 2))
4974 std::swap(a&: Op0, b&: Op1);
4975
4976 Type *Ty = Xor.getType();
4977 Value *A;
4978 const APInt *ShAmt;
4979 if (match(V: Op1, P: m_AShr(L: m_Value(V&: A), R: m_APInt(Res&: ShAmt))) &&
4980 Op1->hasNUses(N: 2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
4981 match(V: Op0, P: m_OneUse(SubPattern: m_c_Add(L: m_Specific(V: A), R: m_Specific(V: Op1))))) {
4982 // Op1 = ashr i32 A, 31 ; smear the sign bit
4983 // xor (add A, Op1), Op1 ; add -1 and flip bits if negative
4984 // --> (A < 0) ? -A : A
4985 Value *IsNeg = Builder.CreateIsNeg(Arg: A);
4986 // Copy the nsw flags from the add to the negate.
4987 auto *Add = cast<BinaryOperator>(Val: Op0);
4988 Value *NegA = Add->hasNoUnsignedWrap()
4989 ? Constant::getNullValue(Ty: A->getType())
4990 : Builder.CreateNeg(V: A, Name: "", HasNSW: Add->hasNoSignedWrap());
4991 return SelectInst::Create(C: IsNeg, S1: NegA, S2: A);
4992 }
4993 return nullptr;
4994}
4995
4996static bool canFreelyInvert(InstCombiner &IC, Value *Op,
4997 Instruction *IgnoredUser) {
4998 auto *I = dyn_cast<Instruction>(Val: Op);
4999 return I && IC.isFreeToInvert(V: I, /*WillInvertAllUses=*/true) &&
5000 IC.canFreelyInvertAllUsersOf(V: I, IgnoredUser);
5001}
5002
5003static Value *freelyInvert(InstCombinerImpl &IC, Value *Op,
5004 Instruction *IgnoredUser) {
5005 auto *I = cast<Instruction>(Val: Op);
5006 IC.Builder.SetInsertPoint(*I->getInsertionPointAfterDef());
5007 Value *NotOp = IC.Builder.CreateNot(V: Op, Name: Op->getName() + ".not");
5008 Op->replaceUsesWithIf(New: NotOp,
5009 ShouldReplace: [NotOp](Use &U) { return U.getUser() != NotOp; });
5010 IC.freelyInvertAllUsersOf(V: NotOp, IgnoredUser);
5011 return NotOp;
5012}
5013
5014// Transform
5015// z = ~(x &/| y)
5016// into:
5017// z = ((~x) |/& (~y))
5018// iff both x and y are free to invert and all uses of z can be freely updated.
5019bool InstCombinerImpl::sinkNotIntoLogicalOp(Instruction &I) {
5020 Value *Op0, *Op1;
5021 if (!match(V: &I, P: m_LogicalOp(L: m_Value(V&: Op0), R: m_Value(V&: Op1))))
5022 return false;
5023
5024 // If this logic op has not been simplified yet, just bail out and let that
5025 // happen first. Otherwise, the code below may wrongly invert.
5026 if (Op0 == Op1)
5027 return false;
5028
5029 // If one of the operands is a user of the other,
5030 // freelyInvert->freelyInvertAllUsersOf will change the operands of I, which
5031 // may cause miscompilation.
5032 if (match(V: Op0, P: m_Not(V: m_Specific(V: Op1))) || match(V: Op1, P: m_Not(V: m_Specific(V: Op0))))
5033 return false;
5034
5035 Instruction::BinaryOps NewOpc =
5036 match(V: &I, P: m_LogicalAnd()) ? Instruction::Or : Instruction::And;
5037 bool IsBinaryOp = isa<BinaryOperator>(Val: I);
5038
5039 // Can our users be adapted?
5040 if (!InstCombiner::canFreelyInvertAllUsersOf(V: &I, /*IgnoredUser=*/nullptr))
5041 return false;
5042
5043 // And can the operands be adapted?
5044 if (!canFreelyInvert(IC&: *this, Op: Op0, IgnoredUser: &I) || !canFreelyInvert(IC&: *this, Op: Op1, IgnoredUser: &I))
5045 return false;
5046
5047 Op0 = freelyInvert(IC&: *this, Op: Op0, IgnoredUser: &I);
5048 Op1 = freelyInvert(IC&: *this, Op: Op1, IgnoredUser: &I);
5049
5050 Builder.SetInsertPoint(*I.getInsertionPointAfterDef());
5051 Value *NewLogicOp;
5052 if (IsBinaryOp) {
5053 NewLogicOp = Builder.CreateBinOp(Opc: NewOpc, LHS: Op0, RHS: Op1, Name: I.getName() + ".not");
5054 } else {
5055 NewLogicOp =
5056 Builder.CreateLogicalOp(Opc: NewOpc, Cond1: Op0, Cond2: Op1, Name: I.getName() + ".not",
5057 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : &I);
5058 if (SelectInst *SI = dyn_cast<SelectInst>(Val: NewLogicOp))
5059 SI->swapProfMetadata();
5060 }
5061
5062 replaceInstUsesWith(I, V: NewLogicOp);
5063 // We can not just create an outer `not`, it will most likely be immediately
5064 // folded back, reconstructing our initial pattern, and causing an
5065 // infinite combine loop, so immediately manually fold it away.
5066 freelyInvertAllUsersOf(V: NewLogicOp);
5067 return true;
5068}
5069
5070// Transform
5071// z = (~x) &/| y
5072// into:
5073// z = ~(x |/& (~y))
5074// iff y is free to invert and all uses of z can be freely updated.
5075bool InstCombinerImpl::sinkNotIntoOtherHandOfLogicalOp(Instruction &I) {
5076 Value *Op0, *Op1;
5077 if (!match(V: &I, P: m_LogicalOp(L: m_Value(V&: Op0), R: m_Value(V&: Op1))))
5078 return false;
5079 Instruction::BinaryOps NewOpc =
5080 match(V: &I, P: m_LogicalAnd()) ? Instruction::Or : Instruction::And;
5081 bool IsBinaryOp = isa<BinaryOperator>(Val: I);
5082
5083 Value *NotOp0 = nullptr;
5084 Value *NotOp1 = nullptr;
5085 Value **OpToInvert = nullptr;
5086 if (match(V: Op0, P: m_Not(V: m_Value(V&: NotOp0))) && canFreelyInvert(IC&: *this, Op: Op1, IgnoredUser: &I)) {
5087 Op0 = NotOp0;
5088 OpToInvert = &Op1;
5089 } else if (match(V: Op1, P: m_Not(V: m_Value(V&: NotOp1))) &&
5090 canFreelyInvert(IC&: *this, Op: Op0, IgnoredUser: &I)) {
5091 Op1 = NotOp1;
5092 OpToInvert = &Op0;
5093 } else
5094 return false;
5095
5096 // And can our users be adapted?
5097 if (!InstCombiner::canFreelyInvertAllUsersOf(V: &I, /*IgnoredUser=*/nullptr))
5098 return false;
5099
5100 *OpToInvert = freelyInvert(IC&: *this, Op: *OpToInvert, IgnoredUser: &I);
5101
5102 Builder.SetInsertPoint(*I.getInsertionPointAfterDef());
5103 Value *NewBinOp;
5104 if (IsBinaryOp)
5105 NewBinOp = Builder.CreateBinOp(Opc: NewOpc, LHS: Op0, RHS: Op1, Name: I.getName() + ".not");
5106 else
5107 NewBinOp = Builder.CreateLogicalOp(Opc: NewOpc, Cond1: Op0, Cond2: Op1, Name: I.getName() + ".not");
5108 replaceInstUsesWith(I, V: NewBinOp);
5109 // We can not just create an outer `not`, it will most likely be immediately
5110 // folded back, reconstructing our initial pattern, and causing an
5111 // infinite combine loop, so immediately manually fold it away.
5112 freelyInvertAllUsersOf(V: NewBinOp);
5113 return true;
5114}
5115
5116Instruction *InstCombinerImpl::foldNot(BinaryOperator &I) {
5117 Value *NotOp;
5118 if (!match(V: &I, P: m_Not(V: m_Value(V&: NotOp))))
5119 return nullptr;
5120
5121 // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand.
5122 // We must eliminate the and/or (one-use) for these transforms to not increase
5123 // the instruction count.
5124 //
5125 // ~(~X & Y) --> (X | ~Y)
5126 // ~(Y & ~X) --> (X | ~Y)
5127 //
5128 // Note: The logical matches do not check for the commuted patterns because
5129 // those are handled via SimplifySelectsFeedingBinaryOp().
5130 Type *Ty = I.getType();
5131 Value *X, *Y;
5132 if (match(V: NotOp, P: m_OneUse(SubPattern: m_c_And(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y))))) {
5133 Value *NotY = Builder.CreateNot(V: Y, Name: Y->getName() + ".not");
5134 return BinaryOperator::CreateOr(V1: X, V2: NotY);
5135 }
5136 if (match(V: NotOp, P: m_OneUse(SubPattern: m_LogicalAnd(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y))))) {
5137 Value *NotY = Builder.CreateNot(V: Y, Name: Y->getName() + ".not");
5138 SelectInst *SI = SelectInst::Create(
5139 C: X, S1: ConstantInt::getTrue(Ty), S2: NotY, NameStr: "", InsertBefore: nullptr,
5140 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : cast<Instruction>(Val: NotOp));
5141 SI->swapProfMetadata();
5142 return SI;
5143 }
5144
5145 // ~(~X | Y) --> (X & ~Y)
5146 // ~(Y | ~X) --> (X & ~Y)
5147 if (match(V: NotOp, P: m_OneUse(SubPattern: m_c_Or(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y))))) {
5148 Value *NotY = Builder.CreateNot(V: Y, Name: Y->getName() + ".not");
5149 return BinaryOperator::CreateAnd(V1: X, V2: NotY);
5150 }
5151 if (match(V: NotOp, P: m_OneUse(SubPattern: m_LogicalOr(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y))))) {
5152 Value *NotY = Builder.CreateNot(V: Y, Name: Y->getName() + ".not");
5153 SelectInst *SI = SelectInst::Create(
5154 C: X, S1: NotY, S2: ConstantInt::getFalse(Ty), NameStr: "", InsertBefore: nullptr,
5155 MDFrom: ProfcheckDisableMetadataFixes ? nullptr : cast<Instruction>(Val: NotOp));
5156 SI->swapProfMetadata();
5157 return SI;
5158 }
5159
5160 // Is this a 'not' (~) fed by a binary operator?
5161 BinaryOperator *NotVal;
5162 if (match(V: NotOp, P: m_BinOp(I&: NotVal))) {
5163 // ~((-X) | Y) --> (X - 1) & (~Y)
5164 if (match(V: NotVal,
5165 P: m_OneUse(SubPattern: m_c_Or(L: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X))), R: m_Value(V&: Y))))) {
5166 Value *DecX = Builder.CreateAdd(LHS: X, RHS: ConstantInt::getAllOnesValue(Ty));
5167 Value *NotY = Builder.CreateNot(V: Y);
5168 return BinaryOperator::CreateAnd(V1: DecX, V2: NotY);
5169 }
5170
5171 // ~(~X >>s Y) --> (X >>s Y)
5172 if (match(V: NotVal, P: m_AShr(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y))))
5173 return BinaryOperator::CreateAShr(V1: X, V2: Y);
5174
5175 // Treat lshr with non-negative operand as ashr.
5176 // ~(~X >>u Y) --> (X >>s Y) iff X is known negative
5177 if (match(V: NotVal, P: m_LShr(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y))) &&
5178 isKnownNegative(V: X, SQ: SQ.getWithInstruction(I: NotVal)))
5179 return BinaryOperator::CreateAShr(V1: X, V2: Y);
5180
5181 // Bit-hack form of a signbit test for iN type:
5182 // ~(X >>s (N - 1)) --> sext i1 (X > -1) to iN
5183 unsigned FullShift = Ty->getScalarSizeInBits() - 1;
5184 if (match(V: NotVal, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: X), R: m_SpecificInt(V: FullShift))))) {
5185 Value *IsNotNeg = Builder.CreateIsNotNeg(Arg: X, Name: "isnotneg");
5186 return new SExtInst(IsNotNeg, Ty);
5187 }
5188
5189 // If we are inverting a right-shifted constant, we may be able to eliminate
5190 // the 'not' by inverting the constant and using the opposite shift type.
5191 // Canonicalization rules ensure that only a negative constant uses 'ashr',
5192 // but we must check that in case that transform has not fired yet.
5193
5194 // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits)
5195 Constant *C;
5196 if (match(V: NotVal, P: m_AShr(L: m_Constant(C), R: m_Value(V&: Y))) &&
5197 match(V: C, P: m_Negative()))
5198 return BinaryOperator::CreateLShr(V1: ConstantExpr::getNot(C), V2: Y);
5199
5200 // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits)
5201 if (match(V: NotVal, P: m_LShr(L: m_Constant(C), R: m_Value(V&: Y))) &&
5202 match(V: C, P: m_NonNegative()))
5203 return BinaryOperator::CreateAShr(V1: ConstantExpr::getNot(C), V2: Y);
5204
5205 // ~(X + C) --> ~C - X
5206 if (match(V: NotVal, P: m_Add(L: m_Value(V&: X), R: m_ImmConstant(C))))
5207 return BinaryOperator::CreateSub(V1: ConstantExpr::getNot(C), V2: X);
5208
5209 // ~(X - Y) --> ~X + Y
5210 // FIXME: is it really beneficial to sink the `not` here?
5211 if (match(V: NotVal, P: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y))))
5212 if (isa<Constant>(Val: X) || NotVal->hasOneUse())
5213 return BinaryOperator::CreateAdd(V1: Builder.CreateNot(V: X), V2: Y);
5214
5215 // ~(~X + Y) --> X - Y
5216 if (match(V: NotVal, P: m_c_Add(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y))))
5217 return BinaryOperator::CreateWithCopiedFlags(Opc: Instruction::Sub, V1: X, V2: Y,
5218 CopyO: NotVal);
5219 }
5220
5221 // not (cmp A, B) = !cmp A, B
5222 CmpPredicate Pred;
5223 if (match(V: NotOp, P: m_Cmp(Pred, L: m_Value(), R: m_Value())) &&
5224 (NotOp->hasOneUse() ||
5225 InstCombiner::canFreelyInvertAllUsersOf(V: cast<Instruction>(Val: NotOp),
5226 /*IgnoredUser=*/nullptr))) {
5227 cast<CmpInst>(Val: NotOp)->setPredicate(CmpInst::getInversePredicate(pred: Pred));
5228 freelyInvertAllUsersOf(V: NotOp);
5229 return &I;
5230 }
5231
5232 // not (bitcast (cmp A, B) --> bitcast (!cmp A, B)
5233 if (match(V: NotOp, P: m_OneUse(SubPattern: m_BitCast(Op: m_Value(V&: X)))) &&
5234 match(V: X, P: m_OneUse(SubPattern: m_Cmp(Pred, L: m_Value(), R: m_Value())))) {
5235 cast<CmpInst>(Val: X)->setPredicate(CmpInst::getInversePredicate(pred: Pred));
5236 return new BitCastInst(X, Ty);
5237 }
5238
5239 // Move a 'not' ahead of casts of a bool to enable logic reduction:
5240 // not (bitcast (sext i1 X)) --> bitcast (sext (not i1 X))
5241 if (match(V: NotOp, P: m_OneUse(SubPattern: m_BitCast(Op: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X)))))) &&
5242 X->getType()->isIntOrIntVectorTy(BitWidth: 1)) {
5243 Type *SextTy = cast<BitCastOperator>(Val: NotOp)->getSrcTy();
5244 Value *NotX = Builder.CreateNot(V: X);
5245 Value *Sext = Builder.CreateSExt(V: NotX, DestTy: SextTy);
5246 return new BitCastInst(Sext, Ty);
5247 }
5248
5249 if (auto *NotOpI = dyn_cast<Instruction>(Val: NotOp))
5250 if (sinkNotIntoLogicalOp(I&: *NotOpI))
5251 return &I;
5252
5253 // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max:
5254 // ~min(~X, ~Y) --> max(X, Y)
5255 // ~max(~X, Y) --> min(X, ~Y)
5256 auto *II = dyn_cast<IntrinsicInst>(Val: NotOp);
5257 if (II && II->hasOneUse()) {
5258 if (match(V: NotOp, P: m_c_MaxOrMin(L: m_Not(V: m_Value(V&: X)), R: m_Value(V&: Y)))) {
5259 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(MinMaxID: II->getIntrinsicID());
5260 Value *NotY = Builder.CreateNot(V: Y);
5261 Value *InvMaxMin = Builder.CreateBinaryIntrinsic(ID: InvID, LHS: X, RHS: NotY);
5262 return replaceInstUsesWith(I, V: InvMaxMin);
5263 }
5264
5265 if (II->getIntrinsicID() == Intrinsic::is_fpclass) {
5266 ConstantInt *ClassMask = cast<ConstantInt>(Val: II->getArgOperand(i: 1));
5267 II->setArgOperand(
5268 i: 1, v: ConstantInt::get(Ty: ClassMask->getType(),
5269 V: ~ClassMask->getZExtValue() & fcAllFlags));
5270 return replaceInstUsesWith(I, V: II);
5271 }
5272 }
5273
5274 if (NotOp->hasOneUse()) {
5275 // Pull 'not' into operands of select if both operands are one-use compares
5276 // or one is one-use compare and the other one is a constant.
5277 // Inverting the predicates eliminates the 'not' operation.
5278 // Example:
5279 // not (select ?, (cmp TPred, ?, ?), (cmp FPred, ?, ?) -->
5280 // select ?, (cmp InvTPred, ?, ?), (cmp InvFPred, ?, ?)
5281 // not (select ?, (cmp TPred, ?, ?), true -->
5282 // select ?, (cmp InvTPred, ?, ?), false
5283 if (auto *Sel = dyn_cast<SelectInst>(Val: NotOp)) {
5284 Value *TV = Sel->getTrueValue();
5285 Value *FV = Sel->getFalseValue();
5286 auto *CmpT = dyn_cast<CmpInst>(Val: TV);
5287 auto *CmpF = dyn_cast<CmpInst>(Val: FV);
5288 bool InvertibleT = (CmpT && CmpT->hasOneUse()) || isa<Constant>(Val: TV);
5289 bool InvertibleF = (CmpF && CmpF->hasOneUse()) || isa<Constant>(Val: FV);
5290 if (InvertibleT && InvertibleF) {
5291 if (CmpT)
5292 CmpT->setPredicate(CmpT->getInversePredicate());
5293 else
5294 Sel->setTrueValue(ConstantExpr::getNot(C: cast<Constant>(Val: TV)));
5295 if (CmpF)
5296 CmpF->setPredicate(CmpF->getInversePredicate());
5297 else
5298 Sel->setFalseValue(ConstantExpr::getNot(C: cast<Constant>(Val: FV)));
5299 return replaceInstUsesWith(I, V: Sel);
5300 }
5301 }
5302 }
5303
5304 if (Instruction *NewXor = foldNotXor(I, Builder))
5305 return NewXor;
5306
5307 // TODO: Could handle multi-use better by checking if all uses of NotOp (other
5308 // than I) can be inverted.
5309 if (Value *R = getFreelyInverted(V: NotOp, WillInvertAllUses: NotOp->hasOneUse(), Builder: &Builder))
5310 return replaceInstUsesWith(I, V: R);
5311
5312 return nullptr;
5313}
5314
5315// ((X + C) & M) ^ M --> (~C − X) & M
5316static Instruction *foldMaskedAddXorPattern(BinaryOperator &I,
5317 InstCombiner::BuilderTy &Builder) {
5318 Value *X, *Mask;
5319 Constant *AddC;
5320 BinaryOperator *AddInst;
5321 if (match(V: &I,
5322 P: m_Xor(L: m_OneUse(SubPattern: m_And(L: m_OneUse(SubPattern: m_CombineAnd(
5323 Ps: m_BinOp(I&: AddInst),
5324 Ps: m_Add(L: m_Value(V&: X), R: m_ImmConstant(C&: AddC)))),
5325 R: m_Value(V&: Mask))),
5326 R: m_Deferred(V: Mask)))) {
5327 Value *NotC = Builder.CreateNot(V: AddC);
5328 Value *NewSub = Builder.CreateSub(LHS: NotC, RHS: X, Name: "", HasNUW: AddInst->hasNoUnsignedWrap(),
5329 HasNSW: AddInst->hasNoSignedWrap());
5330 return BinaryOperator::CreateAnd(V1: NewSub, V2: Mask);
5331 }
5332
5333 return nullptr;
5334}
5335
5336// FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
5337// here. We should standardize that construct where it is needed or choose some
5338// other way to ensure that commutated variants of patterns are not missed.
5339Instruction *InstCombinerImpl::visitXor(BinaryOperator &I) {
5340 if (Value *V = simplifyXorInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1),
5341 Q: SQ.getWithInstruction(I: &I)))
5342 return replaceInstUsesWith(I, V);
5343
5344 if (SimplifyAssociativeOrCommutative(I))
5345 return &I;
5346
5347 if (Instruction *X = foldVectorBinop(Inst&: I))
5348 return X;
5349
5350 if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I))
5351 return Phi;
5352
5353 if (Instruction *NewXor = foldXorToXor(I, Builder))
5354 return NewXor;
5355
5356 // (A&B)^(A&C) -> A&(B^C) etc
5357 if (Value *V = foldUsingDistributiveLaws(I))
5358 return replaceInstUsesWith(I, V);
5359
5360 // See if we can simplify any instructions used by the instruction whose sole
5361 // purpose is to compute bits we don't care about.
5362 if (SimplifyDemandedInstructionBits(Inst&: I))
5363 return &I;
5364
5365 if (Instruction *R = foldNot(I))
5366 return R;
5367
5368 if (Instruction *R = foldBinOpShiftWithShift(I))
5369 return R;
5370
5371 Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1);
5372 Value *X, *Y, *M;
5373
5374 // (X | Y) ^ M -> (X ^ M) ^ Y
5375 // (X | Y) ^ M -> (Y ^ M) ^ X
5376 if (match(V: &I, P: m_c_Xor(L: m_OneUse(SubPattern: m_DisjointOr(L: m_Value(V&: X), R: m_Value(V&: Y))),
5377 R: m_Value(V&: M)))) {
5378 if (Value *XorAC = simplifyXorInst(LHS: X, RHS: M, Q: SQ.getWithInstruction(I: &I)))
5379 return BinaryOperator::CreateXor(V1: XorAC, V2: Y);
5380
5381 if (Value *XorBC = simplifyXorInst(LHS: Y, RHS: M, Q: SQ.getWithInstruction(I: &I)))
5382 return BinaryOperator::CreateXor(V1: XorBC, V2: X);
5383 }
5384
5385 // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M)
5386 // This it a special case in haveNoCommonBitsSet, but the computeKnownBits
5387 // calls in there are unnecessary as SimplifyDemandedInstructionBits should
5388 // have already taken care of those cases.
5389 if (match(V: &I, P: m_c_Xor(L: m_c_And(L: m_Not(V: m_Value(V&: M)), R: m_Value()),
5390 R: m_c_And(L: m_Deferred(V: M), R: m_Value())))) {
5391 if (isGuaranteedNotToBeUndef(V: M))
5392 return BinaryOperator::CreateDisjointOr(V1: Op0, V2: Op1);
5393 else
5394 return BinaryOperator::CreateOr(V1: Op0, V2: Op1);
5395 }
5396
5397 if (Instruction *Xor = visitMaskedMerge(I, Builder))
5398 return Xor;
5399
5400 Constant *C1;
5401 if (match(V: Op1, P: m_Constant(C&: C1))) {
5402 Constant *C2;
5403
5404 if (match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: X), R: m_ImmConstant(C&: C2)))) &&
5405 match(V: C1, P: m_ImmConstant())) {
5406 // (X | C2) ^ C1 --> (X & ~C2) ^ (C1^C2)
5407 C2 = Constant::replaceUndefsWith(
5408 C: C2, Replacement: Constant::getAllOnesValue(Ty: C2->getType()->getScalarType()));
5409 Value *And = Builder.CreateAnd(
5410 LHS: X, RHS: Constant::mergeUndefsWith(C: ConstantExpr::getNot(C: C2), Other: C1));
5411 return BinaryOperator::CreateXor(
5412 V1: And, V2: Constant::mergeUndefsWith(C: ConstantExpr::getXor(C1, C2), Other: C1));
5413 }
5414
5415 // Use DeMorgan and reassociation to eliminate a 'not' op.
5416 if (match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Not(V: m_Value(V&: X)), R: m_Constant(C&: C2))))) {
5417 // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C1
5418 Value *And = Builder.CreateAnd(LHS: X, RHS: ConstantExpr::getNot(C: C2));
5419 return BinaryOperator::CreateXor(V1: And, V2: ConstantExpr::getNot(C: C1));
5420 }
5421 if (match(V: Op0, P: m_OneUse(SubPattern: m_And(L: m_Not(V: m_Value(V&: X)), R: m_Constant(C&: C2))))) {
5422 // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C1
5423 Value *Or = Builder.CreateOr(LHS: X, RHS: ConstantExpr::getNot(C: C2));
5424 return BinaryOperator::CreateXor(V1: Or, V2: ConstantExpr::getNot(C: C1));
5425 }
5426
5427 // Convert xor ([trunc] (ashr X, BW-1)), C =>
5428 // select(X >s -1, C, ~C)
5429 // The ashr creates "AllZeroOrAllOne's", which then optionally inverses the
5430 // constant depending on whether this input is less than 0.
5431 const APInt *CA;
5432 if (match(V: Op0, P: m_OneUse(SubPattern: m_TruncOrSelf(
5433 Op: m_AShr(L: m_Value(V&: X), R: m_APIntAllowPoison(Res&: CA))))) &&
5434 *CA == X->getType()->getScalarSizeInBits() - 1 &&
5435 !match(V: C1, P: m_AllOnes())) {
5436 assert(!C1->isNullValue() && "Unexpected xor with 0");
5437 Value *IsNotNeg = Builder.CreateIsNotNeg(Arg: X);
5438 return createSelectInstWithUnknownProfile(C: IsNotNeg, S1: Op1,
5439 S2: Builder.CreateNot(V: Op1));
5440 }
5441 }
5442
5443 Type *Ty = I.getType();
5444 {
5445 const APInt *RHSC;
5446 if (match(V: Op1, P: m_APInt(Res&: RHSC))) {
5447 Value *X;
5448 const APInt *C;
5449 // (C - X) ^ signmaskC --> (C + signmaskC) - X
5450 if (RHSC->isSignMask() && match(V: Op0, P: m_Sub(L: m_APInt(Res&: C), R: m_Value(V&: X))))
5451 return BinaryOperator::CreateSub(V1: ConstantInt::get(Ty, V: *C + *RHSC), V2: X);
5452
5453 // (X + C) ^ signmaskC --> X + (C + signmaskC)
5454 if (RHSC->isSignMask() && match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_APInt(Res&: C))))
5455 return BinaryOperator::CreateAdd(V1: X, V2: ConstantInt::get(Ty, V: *C + *RHSC));
5456
5457 // (X | C) ^ RHSC --> X ^ (C ^ RHSC) iff X & C == 0
5458 if (match(V: Op0, P: m_Or(L: m_Value(V&: X), R: m_APInt(Res&: C))) &&
5459 MaskedValueIsZero(V: X, Mask: *C, CxtI: &I))
5460 return BinaryOperator::CreateXor(V1: X, V2: ConstantInt::get(Ty, V: *C ^ *RHSC));
5461
5462 // When X is a power-of-two or zero and zero input is poison:
5463 // ctlz(i32 X) ^ 31 --> cttz(X)
5464 // cttz(i32 X) ^ 31 --> ctlz(X)
5465 auto *II = dyn_cast<IntrinsicInst>(Val: Op0);
5466 if (II && II->hasOneUse() && *RHSC == Ty->getScalarSizeInBits() - 1) {
5467 Intrinsic::ID IID = II->getIntrinsicID();
5468 if ((IID == Intrinsic::ctlz || IID == Intrinsic::cttz) &&
5469 match(V: II->getArgOperand(i: 1), P: m_One()) &&
5470 isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 0), /*OrZero */ true)) {
5471 IID = (IID == Intrinsic::ctlz) ? Intrinsic::cttz : Intrinsic::ctlz;
5472 Function *F =
5473 Intrinsic::getOrInsertDeclaration(M: II->getModule(), id: IID, OverloadTys: Ty);
5474 return CallInst::Create(Func: F, Args: {II->getArgOperand(i: 0), Builder.getTrue()});
5475 }
5476 }
5477
5478 // If RHSC is inverting the remaining bits of shifted X,
5479 // canonicalize to a 'not' before the shift to help SCEV and codegen:
5480 // (X << C) ^ RHSC --> ~X << C
5481 if (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C)))) &&
5482 *RHSC == APInt::getAllOnes(numBits: Ty->getScalarSizeInBits()).shl(ShiftAmt: *C)) {
5483 Value *NotX = Builder.CreateNot(V: X);
5484 return BinaryOperator::CreateShl(V1: NotX, V2: ConstantInt::get(Ty, V: *C));
5485 }
5486 // (X >>u C) ^ RHSC --> ~X >>u C
5487 if (match(V: Op0, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C)))) &&
5488 *RHSC == APInt::getAllOnes(numBits: Ty->getScalarSizeInBits()).lshr(ShiftAmt: *C)) {
5489 Value *NotX = Builder.CreateNot(V: X);
5490 return BinaryOperator::CreateLShr(V1: NotX, V2: ConstantInt::get(Ty, V: *C));
5491 }
5492 // TODO: We could handle 'ashr' here as well. That would be matching
5493 // a 'not' op and moving it before the shift. Doing that requires
5494 // preventing the inverse fold in canShiftBinOpWithConstantRHS().
5495 }
5496
5497 // If we are XORing the sign bit of a floating-point value, convert
5498 // this to fneg, then cast back to integer.
5499 //
5500 // This is generous interpretation of noimplicitfloat, this is not a true
5501 // floating-point operation.
5502 //
5503 // Assumes any IEEE-represented type has the sign bit in the high bit.
5504 // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt
5505 Value *CastOp;
5506 if (match(V: Op0, P: m_ElementWiseBitCast(Op: m_Value(V&: CastOp))) &&
5507 match(V: Op1, P: m_SignMask()) &&
5508 !Builder.GetInsertBlock()->getParent()->hasFnAttribute(
5509 Kind: Attribute::NoImplicitFloat)) {
5510 Type *EltTy = CastOp->getType()->getScalarType();
5511 if (EltTy->isFloatingPointTy() &&
5512 APFloat::hasSignBitInMSB(EltTy->getFltSemantics())) {
5513 Value *FNeg = Builder.CreateFNeg(V: CastOp);
5514 return new BitCastInst(FNeg, I.getType());
5515 }
5516 }
5517 }
5518
5519 // FIXME: This should not be limited to scalar (pull into APInt match above).
5520 {
5521 Value *X;
5522 ConstantInt *C1, *C2, *C3;
5523 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
5524 if (match(V: Op1, P: m_ConstantInt(CI&: C3)) &&
5525 match(V: Op0, P: m_LShr(L: m_Xor(L: m_Value(V&: X), R: m_ConstantInt(CI&: C1)),
5526 R: m_ConstantInt(CI&: C2))) &&
5527 Op0->hasOneUse()) {
5528 // fold (C1 >> C2) ^ C3
5529 APInt FoldConst = C1->getValue().lshr(ShiftAmt: C2->getValue());
5530 FoldConst ^= C3->getValue();
5531 // Prepare the two operands.
5532 auto *Opnd0 = Builder.CreateLShr(LHS: X, RHS: C2);
5533 Opnd0->takeName(V: Op0);
5534 return BinaryOperator::CreateXor(V1: Opnd0, V2: ConstantInt::get(Ty, V: FoldConst));
5535 }
5536 }
5537
5538 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
5539 return FoldedLogic;
5540
5541 if (Instruction *FoldedLogic = foldBinOpSelectBinOp(Op&: I))
5542 return FoldedLogic;
5543
5544 // Y ^ (X | Y) --> X & ~Y
5545 // Y ^ (Y | X) --> X & ~Y
5546 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_Or(L: m_Value(V&: X), R: m_Specific(V: Op0)))))
5547 return BinaryOperator::CreateAnd(V1: X, V2: Builder.CreateNot(V: Op0));
5548 // (X | Y) ^ Y --> X & ~Y
5549 // (Y | X) ^ Y --> X & ~Y
5550 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Value(V&: X), R: m_Specific(V: Op1)))))
5551 return BinaryOperator::CreateAnd(V1: X, V2: Builder.CreateNot(V: Op1));
5552
5553 // Y ^ (X & Y) --> ~X & Y
5554 // Y ^ (Y & X) --> ~X & Y
5555 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: X), R: m_Specific(V: Op0)))))
5556 return BinaryOperator::CreateAnd(V1: Op0, V2: Builder.CreateNot(V: X));
5557 // (X & Y) ^ Y --> ~X & Y
5558 // (Y & X) ^ Y --> ~X & Y
5559 // Canonical form is (X & C) ^ C; don't touch that.
5560 // TODO: A 'not' op is better for analysis and codegen, but demanded bits must
5561 // be fixed to prefer that (otherwise we get infinite looping).
5562 if (!match(V: Op1, P: m_Constant()) &&
5563 match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: X), R: m_Specific(V: Op1)))))
5564 return BinaryOperator::CreateAnd(V1: Op1, V2: Builder.CreateNot(V: X));
5565
5566 Value *A, *B, *C;
5567 // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants.
5568 if (match(V: &I, P: m_c_Xor(L: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))),
5569 R: m_OneUse(SubPattern: m_c_Or(L: m_Deferred(V: A), R: m_Value(V&: C))))))
5570 return BinaryOperator::CreateXor(
5571 V1: Builder.CreateAnd(LHS: Builder.CreateNot(V: A), RHS: C), V2: B);
5572
5573 // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants.
5574 if (match(V: &I, P: m_c_Xor(L: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))),
5575 R: m_OneUse(SubPattern: m_c_Or(L: m_Deferred(V: B), R: m_Value(V&: C))))))
5576 return BinaryOperator::CreateXor(
5577 V1: Builder.CreateAnd(LHS: Builder.CreateNot(V: B), RHS: C), V2: A);
5578
5579 // (A & B) ^ (A ^ B) -> (A | B)
5580 if (match(V: Op0, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) &&
5581 match(V: Op1, P: m_c_Xor(L: m_Specific(V: A), R: m_Specific(V: B))))
5582 return BinaryOperator::CreateOr(V1: A, V2: B);
5583 // (A ^ B) ^ (A & B) -> (A | B)
5584 if (match(V: Op0, P: m_Xor(L: m_Value(V&: A), R: m_Value(V&: B))) &&
5585 match(V: Op1, P: m_c_And(L: m_Specific(V: A), R: m_Specific(V: B))))
5586 return BinaryOperator::CreateOr(V1: A, V2: B);
5587
5588 // (A & ~B) ^ ~A -> ~(A & B)
5589 // (~B & A) ^ ~A -> ~(A & B)
5590 if (match(V: Op0, P: m_c_And(L: m_Value(V&: A), R: m_Not(V: m_Value(V&: B)))) &&
5591 match(V: Op1, P: m_Not(V: m_Specific(V: A))))
5592 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: A, RHS: B));
5593
5594 // (~A & B) ^ A --> A | B -- There are 4 commuted variants.
5595 if (match(V: &I, P: m_c_Xor(L: m_c_And(L: m_Not(V: m_Value(V&: A)), R: m_Value(V&: B)), R: m_Deferred(V: A))))
5596 return BinaryOperator::CreateOr(V1: A, V2: B);
5597
5598 // (~A | B) ^ A --> ~(A & B)
5599 if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Not(V: m_Specific(V: Op1)), R: m_Value(V&: B)))))
5600 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: Op1, RHS: B));
5601
5602 // A ^ (~A | B) --> ~(A & B)
5603 if (match(V: Op1, P: m_OneUse(SubPattern: m_c_Or(L: m_Not(V: m_Specific(V: Op0)), R: m_Value(V&: B)))))
5604 return BinaryOperator::CreateNot(Op: Builder.CreateAnd(LHS: Op0, RHS: B));
5605
5606 // (A | B) ^ (A | C) --> (B ^ C) & ~A -- There are 4 commuted variants.
5607 // TODO: Loosen one-use restriction if common operand is a constant.
5608 Value *D;
5609 if (match(V: Op0, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: A), R: m_Value(V&: B)))) &&
5610 match(V: Op1, P: m_OneUse(SubPattern: m_Or(L: m_Value(V&: C), R: m_Value(V&: D))))) {
5611 if (B == C || B == D)
5612 std::swap(a&: A, b&: B);
5613 if (A == C)
5614 std::swap(a&: C, b&: D);
5615 if (A == D) {
5616 Value *NotA = Builder.CreateNot(V: A);
5617 return BinaryOperator::CreateAnd(V1: Builder.CreateXor(LHS: B, RHS: C), V2: NotA);
5618 }
5619 }
5620
5621 // (A & B) ^ (A | C) --> A ? ~B : C -- There are 4 commuted variants.
5622 if (I.getType()->isIntOrIntVectorTy(BitWidth: 1) &&
5623 match(V: &I, P: m_c_Xor(L: m_OneUse(SubPattern: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))),
5624 R: m_OneUse(SubPattern: m_LogicalOr(L: m_Value(V&: C), R: m_Value(V&: D)))))) {
5625 bool NeedFreeze = isa<SelectInst>(Val: Op0) && isa<SelectInst>(Val: Op1) && B == D;
5626 Instruction *MDFrom = cast<Instruction>(Val: Op0);
5627 if (B == C || B == D) {
5628 std::swap(a&: A, b&: B);
5629 MDFrom = B == C ? cast<Instruction>(Val: Op1) : nullptr;
5630 }
5631 if (A == C)
5632 std::swap(a&: C, b&: D);
5633 if (A == D) {
5634 if (NeedFreeze)
5635 A = Builder.CreateFreeze(V: A);
5636 Value *NotB = Builder.CreateNot(V: B);
5637 return MDFrom == nullptr || ProfcheckDisableMetadataFixes
5638 ? createSelectInstWithUnknownProfile(C: A, S1: NotB, S2: C)
5639 : SelectInst::Create(C: A, S1: NotB, S2: C, NameStr: "", InsertBefore: nullptr, MDFrom);
5640 }
5641 }
5642
5643 if (auto *LHS = dyn_cast<ICmpInst>(Val: I.getOperand(i_nocapture: 0)))
5644 if (auto *RHS = dyn_cast<ICmpInst>(Val: I.getOperand(i_nocapture: 1)))
5645 if (Value *V = foldXorOfICmps(LHS, RHS, I))
5646 return replaceInstUsesWith(I, V);
5647
5648 if (Instruction *CastedXor = foldCastedBitwiseLogic(I))
5649 return CastedXor;
5650
5651 if (Instruction *Abs = canonicalizeAbs(Xor&: I, Builder))
5652 return Abs;
5653
5654 // Otherwise, if all else failed, try to hoist the xor-by-constant:
5655 // (X ^ C) ^ Y --> (X ^ Y) ^ C
5656 // Just like we do in other places, we completely avoid the fold
5657 // for constantexprs, at least to avoid endless combine loop.
5658 if (match(V: &I, P: m_c_Xor(L: m_OneUse(SubPattern: m_Xor(L: m_Value(V&: X, P: m_Unless(M: m_ConstantExpr())),
5659 R: m_ImmConstant(C&: C1))),
5660 R: m_Value(V&: Y))))
5661 return BinaryOperator::CreateXor(V1: Builder.CreateXor(LHS: X, RHS: Y), V2: C1);
5662
5663 if (Instruction *R = reassociateForUses(BO&: I, Builder))
5664 return R;
5665
5666 if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))
5667 return Canonicalized;
5668
5669 if (Instruction *Folded = foldLogicOfIsFPClass(BO&: I, Op0, Op1))
5670 return Folded;
5671
5672 if (Instruction *Folded = canonicalizeConditionalNegationViaMathToSelect(I))
5673 return Folded;
5674
5675 if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
5676 return Res;
5677
5678 if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))
5679 return Res;
5680
5681 if (Instruction *Res = foldMaskedAddXorPattern(I, Builder))
5682 return Res;
5683
5684 return nullptr;
5685}
5686