xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (revision 8c2f6c3be0125142d3c1782e4b0ee0634c584b9e)
1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visitAnd, visitOr, and visitXor functions.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstCombineInternal.h"
14 #include "llvm/Analysis/CmpInstAnalysis.h"
15 #include "llvm/Analysis/InstructionSimplify.h"
16 #include "llvm/IR/ConstantRange.h"
17 #include "llvm/IR/Intrinsics.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Transforms/InstCombine/InstCombiner.h"
20 #include "llvm/Transforms/Utils/Local.h"
21 
22 using namespace llvm;
23 using namespace PatternMatch;
24 
25 #define DEBUG_TYPE "instcombine"
26 
27 /// This is the complement of getICmpCode, which turns an opcode and two
28 /// operands into either a constant true or false, or a brand new ICmp
29 /// instruction. The sign is passed in to determine which kind of predicate to
30 /// use in the new icmp instruction.
31 static Value *getNewICmpValue(unsigned Code, bool Sign, Value *LHS, Value *RHS,
32                               InstCombiner::BuilderTy &Builder) {
33   ICmpInst::Predicate NewPred;
34   if (Constant *TorF = getPredForICmpCode(Code, Sign, LHS->getType(), NewPred))
35     return TorF;
36   return Builder.CreateICmp(NewPred, LHS, RHS);
37 }
38 
39 /// This is the complement of getFCmpCode, which turns an opcode and two
40 /// operands into either a FCmp instruction, or a true/false constant.
41 static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS,
42                            InstCombiner::BuilderTy &Builder) {
43   FCmpInst::Predicate NewPred;
44   if (Constant *TorF = getPredForFCmpCode(Code, LHS->getType(), NewPred))
45     return TorF;
46   return Builder.CreateFCmp(NewPred, LHS, RHS);
47 }
48 
49 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
50 /// (V < Lo || V >= Hi). This method expects that Lo < Hi. IsSigned indicates
51 /// whether to treat V, Lo, and Hi as signed or not.
52 Value *InstCombinerImpl::insertRangeTest(Value *V, const APInt &Lo,
53                                          const APInt &Hi, bool isSigned,
54                                          bool Inside) {
55   assert((isSigned ? Lo.slt(Hi) : Lo.ult(Hi)) &&
56          "Lo is not < Hi in range emission code!");
57 
58   Type *Ty = V->getType();
59 
60   // V >= Min && V <  Hi --> V <  Hi
61   // V <  Min || V >= Hi --> V >= Hi
62   ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
63   if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) {
64     Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred;
65     return Builder.CreateICmp(Pred, V, ConstantInt::get(Ty, Hi));
66   }
67 
68   // V >= Lo && V <  Hi --> V - Lo u<  Hi - Lo
69   // V <  Lo || V >= Hi --> V - Lo u>= Hi - Lo
70   Value *VMinusLo =
71       Builder.CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off");
72   Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo);
73   return Builder.CreateICmp(Pred, VMinusLo, HiMinusLo);
74 }
75 
76 /// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns
77 /// that can be simplified.
78 /// One of A and B is considered the mask. The other is the value. This is
79 /// described as the "AMask" or "BMask" part of the enum. If the enum contains
80 /// only "Mask", then both A and B can be considered masks. If A is the mask,
81 /// then it was proven that (A & C) == C. This is trivial if C == A or C == 0.
82 /// If both A and C are constants, this proof is also easy.
83 /// For the following explanations, we assume that A is the mask.
84 ///
85 /// "AllOnes" declares that the comparison is true only if (A & B) == A or all
86 /// bits of A are set in B.
87 ///   Example: (icmp eq (A & 3), 3) -> AMask_AllOnes
88 ///
89 /// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all
90 /// bits of A are cleared in B.
91 ///   Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes
92 ///
93 /// "Mixed" declares that (A & B) == C and C might or might not contain any
94 /// number of one bits and zero bits.
95 ///   Example: (icmp eq (A & 3), 1) -> AMask_Mixed
96 ///
97 /// "Not" means that in above descriptions "==" should be replaced by "!=".
98 ///   Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes
99 ///
100 /// If the mask A contains a single bit, then the following is equivalent:
101 ///    (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
102 ///    (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
103 enum MaskedICmpType {
104   AMask_AllOnes           =     1,
105   AMask_NotAllOnes        =     2,
106   BMask_AllOnes           =     4,
107   BMask_NotAllOnes        =     8,
108   Mask_AllZeros           =    16,
109   Mask_NotAllZeros        =    32,
110   AMask_Mixed             =    64,
111   AMask_NotMixed          =   128,
112   BMask_Mixed             =   256,
113   BMask_NotMixed          =   512
114 };
115 
116 /// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C)
117 /// satisfies.
118 static unsigned getMaskedICmpType(Value *A, Value *B, Value *C,
119                                   ICmpInst::Predicate Pred) {
120   const APInt *ConstA = nullptr, *ConstB = nullptr, *ConstC = nullptr;
121   match(A, m_APInt(ConstA));
122   match(B, m_APInt(ConstB));
123   match(C, m_APInt(ConstC));
124   bool IsEq = (Pred == ICmpInst::ICMP_EQ);
125   bool IsAPow2 = ConstA && ConstA->isPowerOf2();
126   bool IsBPow2 = ConstB && ConstB->isPowerOf2();
127   unsigned MaskVal = 0;
128   if (ConstC && ConstC->isZero()) {
129     // if C is zero, then both A and B qualify as mask
130     MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed)
131                      : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed));
132     if (IsAPow2)
133       MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed)
134                        : (AMask_AllOnes | AMask_Mixed));
135     if (IsBPow2)
136       MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed)
137                        : (BMask_AllOnes | BMask_Mixed));
138     return MaskVal;
139   }
140 
141   if (A == C) {
142     MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed)
143                      : (AMask_NotAllOnes | AMask_NotMixed));
144     if (IsAPow2)
145       MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed)
146                        : (Mask_AllZeros | AMask_Mixed));
147   } else if (ConstA && ConstC && ConstC->isSubsetOf(*ConstA)) {
148     MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed);
149   }
150 
151   if (B == C) {
152     MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed)
153                      : (BMask_NotAllOnes | BMask_NotMixed));
154     if (IsBPow2)
155       MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed)
156                        : (Mask_AllZeros | BMask_Mixed));
157   } else if (ConstB && ConstC && ConstC->isSubsetOf(*ConstB)) {
158     MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed);
159   }
160 
161   return MaskVal;
162 }
163 
164 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
165 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
166 /// is adjacent to the corresponding normal flag (recording ==), this just
167 /// involves swapping those bits over.
168 static unsigned conjugateICmpMask(unsigned Mask) {
169   unsigned NewMask;
170   NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros |
171                      AMask_Mixed | BMask_Mixed))
172             << 1;
173 
174   NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros |
175                       AMask_NotMixed | BMask_NotMixed))
176              >> 1;
177 
178   return NewMask;
179 }
180 
181 // Adapts the external decomposeBitTestICmp for local use.
182 static bool decomposeBitTestICmp(Value *LHS, Value *RHS, CmpInst::Predicate &Pred,
183                                  Value *&X, Value *&Y, Value *&Z) {
184   APInt Mask;
185   if (!llvm::decomposeBitTestICmp(LHS, RHS, Pred, X, Mask))
186     return false;
187 
188   Y = ConstantInt::get(X->getType(), Mask);
189   Z = ConstantInt::get(X->getType(), 0);
190   return true;
191 }
192 
193 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E).
194 /// Return the pattern classes (from MaskedICmpType) for the left hand side and
195 /// the right hand side as a pair.
196 /// LHS and RHS are the left hand side and the right hand side ICmps and PredL
197 /// and PredR are their predicates, respectively.
198 static std::optional<std::pair<unsigned, unsigned>> getMaskedTypeForICmpPair(
199     Value *&A, Value *&B, Value *&C, Value *&D, Value *&E, ICmpInst *LHS,
200     ICmpInst *RHS, ICmpInst::Predicate &PredL, ICmpInst::Predicate &PredR) {
201   // Don't allow pointers. Splat vectors are fine.
202   if (!LHS->getOperand(0)->getType()->isIntOrIntVectorTy() ||
203       !RHS->getOperand(0)->getType()->isIntOrIntVectorTy())
204     return std::nullopt;
205 
206   // Here comes the tricky part:
207   // LHS might be of the form L11 & L12 == X, X == L21 & L22,
208   // and L11 & L12 == L21 & L22. The same goes for RHS.
209   // Now we must find those components L** and R**, that are equal, so
210   // that we can extract the parameters A, B, C, D, and E for the canonical
211   // above.
212   Value *L1 = LHS->getOperand(0);
213   Value *L2 = LHS->getOperand(1);
214   Value *L11, *L12, *L21, *L22;
215   // Check whether the icmp can be decomposed into a bit test.
216   if (decomposeBitTestICmp(L1, L2, PredL, L11, L12, L2)) {
217     L21 = L22 = L1 = nullptr;
218   } else {
219     // Look for ANDs in the LHS icmp.
220     if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
221       // Any icmp can be viewed as being trivially masked; if it allows us to
222       // remove one, it's worth it.
223       L11 = L1;
224       L12 = Constant::getAllOnesValue(L1->getType());
225     }
226 
227     if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
228       L21 = L2;
229       L22 = Constant::getAllOnesValue(L2->getType());
230     }
231   }
232 
233   // Bail if LHS was a icmp that can't be decomposed into an equality.
234   if (!ICmpInst::isEquality(PredL))
235     return std::nullopt;
236 
237   Value *R1 = RHS->getOperand(0);
238   Value *R2 = RHS->getOperand(1);
239   Value *R11, *R12;
240   bool Ok = false;
241   if (decomposeBitTestICmp(R1, R2, PredR, R11, R12, R2)) {
242     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
243       A = R11;
244       D = R12;
245     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
246       A = R12;
247       D = R11;
248     } else {
249       return std::nullopt;
250     }
251     E = R2;
252     R1 = nullptr;
253     Ok = true;
254   } else {
255     if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
256       // As before, model no mask as a trivial mask if it'll let us do an
257       // optimization.
258       R11 = R1;
259       R12 = Constant::getAllOnesValue(R1->getType());
260     }
261 
262     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
263       A = R11;
264       D = R12;
265       E = R2;
266       Ok = true;
267     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
268       A = R12;
269       D = R11;
270       E = R2;
271       Ok = true;
272     }
273   }
274 
275   // Bail if RHS was a icmp that can't be decomposed into an equality.
276   if (!ICmpInst::isEquality(PredR))
277     return std::nullopt;
278 
279   // Look for ANDs on the right side of the RHS icmp.
280   if (!Ok) {
281     if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
282       R11 = R2;
283       R12 = Constant::getAllOnesValue(R2->getType());
284     }
285 
286     if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
287       A = R11;
288       D = R12;
289       E = R1;
290       Ok = true;
291     } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
292       A = R12;
293       D = R11;
294       E = R1;
295       Ok = true;
296     } else {
297       return std::nullopt;
298     }
299 
300     assert(Ok && "Failed to find AND on the right side of the RHS icmp.");
301   }
302 
303   if (L11 == A) {
304     B = L12;
305     C = L2;
306   } else if (L12 == A) {
307     B = L11;
308     C = L2;
309   } else if (L21 == A) {
310     B = L22;
311     C = L1;
312   } else if (L22 == A) {
313     B = L21;
314     C = L1;
315   }
316 
317   unsigned LeftType = getMaskedICmpType(A, B, C, PredL);
318   unsigned RightType = getMaskedICmpType(A, D, E, PredR);
319   return std::optional<std::pair<unsigned, unsigned>>(
320       std::make_pair(LeftType, RightType));
321 }
322 
323 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single
324 /// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros
325 /// and the right hand side is of type BMask_Mixed. For example,
326 /// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8).
327 /// Also used for logical and/or, must be poison safe.
328 static Value *foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
329     ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, Value *A, Value *B, Value *C,
330     Value *D, Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
331     InstCombiner::BuilderTy &Builder) {
332   // We are given the canonical form:
333   //   (icmp ne (A & B), 0) & (icmp eq (A & D), E).
334   // where D & E == E.
335   //
336   // If IsAnd is false, we get it in negated form:
337   //   (icmp eq (A & B), 0) | (icmp ne (A & D), E) ->
338   //      !((icmp ne (A & B), 0) & (icmp eq (A & D), E)).
339   //
340   // We currently handle the case of B, C, D, E are constant.
341   //
342   const APInt *BCst, *CCst, *DCst, *OrigECst;
343   if (!match(B, m_APInt(BCst)) || !match(C, m_APInt(CCst)) ||
344       !match(D, m_APInt(DCst)) || !match(E, m_APInt(OrigECst)))
345     return nullptr;
346 
347   ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
348 
349   // Update E to the canonical form when D is a power of two and RHS is
350   // canonicalized as,
351   // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or
352   // (icmp ne (A & D), D) -> (icmp eq (A & D), 0).
353   APInt ECst = *OrigECst;
354   if (PredR != NewCC)
355     ECst ^= *DCst;
356 
357   // If B or D is zero, skip because if LHS or RHS can be trivially folded by
358   // other folding rules and this pattern won't apply any more.
359   if (*BCst == 0 || *DCst == 0)
360     return nullptr;
361 
362   // If B and D don't intersect, ie. (B & D) == 0, no folding because we can't
363   // deduce anything from it.
364   // For example,
365   // (icmp ne (A & 12), 0) & (icmp eq (A & 3), 1) -> no folding.
366   if ((*BCst & *DCst) == 0)
367     return nullptr;
368 
369   // If the following two conditions are met:
370   //
371   // 1. mask B covers only a single bit that's not covered by mask D, that is,
372   // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of
373   // B and D has only one bit set) and,
374   //
375   // 2. RHS (and E) indicates that the rest of B's bits are zero (in other
376   // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0
377   //
378   // then that single bit in B must be one and thus the whole expression can be
379   // folded to
380   //   (A & (B | D)) == (B & (B ^ D)) | E.
381   //
382   // For example,
383   // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9)
384   // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8)
385   if ((((*BCst & *DCst) & ECst) == 0) &&
386       (*BCst & (*BCst ^ *DCst)).isPowerOf2()) {
387     APInt BorD = *BCst | *DCst;
388     APInt BandBxorDorE = (*BCst & (*BCst ^ *DCst)) | ECst;
389     Value *NewMask = ConstantInt::get(A->getType(), BorD);
390     Value *NewMaskedValue = ConstantInt::get(A->getType(), BandBxorDorE);
391     Value *NewAnd = Builder.CreateAnd(A, NewMask);
392     return Builder.CreateICmp(NewCC, NewAnd, NewMaskedValue);
393   }
394 
395   auto IsSubSetOrEqual = [](const APInt *C1, const APInt *C2) {
396     return (*C1 & *C2) == *C1;
397   };
398   auto IsSuperSetOrEqual = [](const APInt *C1, const APInt *C2) {
399     return (*C1 & *C2) == *C2;
400   };
401 
402   // In the following, we consider only the cases where B is a superset of D, B
403   // is a subset of D, or B == D because otherwise there's at least one bit
404   // covered by B but not D, in which case we can't deduce much from it, so
405   // no folding (aside from the single must-be-one bit case right above.)
406   // For example,
407   // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding.
408   if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst))
409     return nullptr;
410 
411   // At this point, either B is a superset of D, B is a subset of D or B == D.
412 
413   // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict
414   // and the whole expression becomes false (or true if negated), otherwise, no
415   // folding.
416   // For example,
417   // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false.
418   // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding.
419   if (ECst.isZero()) {
420     if (IsSubSetOrEqual(BCst, DCst))
421       return ConstantInt::get(LHS->getType(), !IsAnd);
422     return nullptr;
423   }
424 
425   // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B ==
426   // D. If B is a superset of (or equal to) D, since E is not zero, LHS is
427   // subsumed by RHS (RHS implies LHS.) So the whole expression becomes
428   // RHS. For example,
429   // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
430   // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
431   if (IsSuperSetOrEqual(BCst, DCst))
432     return RHS;
433   // Otherwise, B is a subset of D. If B and E have a common bit set,
434   // ie. (B & E) != 0, then LHS is subsumed by RHS. For example.
435   // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
436   assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code");
437   if ((*BCst & ECst) != 0)
438     return RHS;
439   // Otherwise, LHS and RHS contradict and the whole expression becomes false
440   // (or true if negated.) For example,
441   // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false.
442   // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false.
443   return ConstantInt::get(LHS->getType(), !IsAnd);
444 }
445 
446 /// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single
447 /// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side
448 /// aren't of the common mask pattern type.
449 /// Also used for logical and/or, must be poison safe.
450 static Value *foldLogOpOfMaskedICmpsAsymmetric(
451     ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, Value *A, Value *B, Value *C,
452     Value *D, Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
453     unsigned LHSMask, unsigned RHSMask, InstCombiner::BuilderTy &Builder) {
454   assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
455          "Expected equality predicates for masked type of icmps.");
456   // Handle Mask_NotAllZeros-BMask_Mixed cases.
457   // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or
458   // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E)
459   //    which gets swapped to
460   //    (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C).
461   if (!IsAnd) {
462     LHSMask = conjugateICmpMask(LHSMask);
463     RHSMask = conjugateICmpMask(RHSMask);
464   }
465   if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) {
466     if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
467             LHS, RHS, IsAnd, A, B, C, D, E,
468             PredL, PredR, Builder)) {
469       return V;
470     }
471   } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) {
472     if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
473             RHS, LHS, IsAnd, A, D, E, B, C,
474             PredR, PredL, Builder)) {
475       return V;
476     }
477   }
478   return nullptr;
479 }
480 
481 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
482 /// into a single (icmp(A & X) ==/!= Y).
483 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
484                                      bool IsLogical,
485                                      InstCombiner::BuilderTy &Builder) {
486   Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
487   ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
488   std::optional<std::pair<unsigned, unsigned>> MaskPair =
489       getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR);
490   if (!MaskPair)
491     return nullptr;
492   assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
493          "Expected equality predicates for masked type of icmps.");
494   unsigned LHSMask = MaskPair->first;
495   unsigned RHSMask = MaskPair->second;
496   unsigned Mask = LHSMask & RHSMask;
497   if (Mask == 0) {
498     // Even if the two sides don't share a common pattern, check if folding can
499     // still happen.
500     if (Value *V = foldLogOpOfMaskedICmpsAsymmetric(
501             LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask,
502             Builder))
503       return V;
504     return nullptr;
505   }
506 
507   // In full generality:
508   //     (icmp (A & B) Op C) | (icmp (A & D) Op E)
509   // ==  ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
510   //
511   // If the latter can be converted into (icmp (A & X) Op Y) then the former is
512   // equivalent to (icmp (A & X) !Op Y).
513   //
514   // Therefore, we can pretend for the rest of this function that we're dealing
515   // with the conjunction, provided we flip the sense of any comparisons (both
516   // input and output).
517 
518   // In most cases we're going to produce an EQ for the "&&" case.
519   ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
520   if (!IsAnd) {
521     // Convert the masking analysis into its equivalent with negated
522     // comparisons.
523     Mask = conjugateICmpMask(Mask);
524   }
525 
526   if (Mask & Mask_AllZeros) {
527     // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
528     // -> (icmp eq (A & (B|D)), 0)
529     if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D))
530       return nullptr; // TODO: Use freeze?
531     Value *NewOr = Builder.CreateOr(B, D);
532     Value *NewAnd = Builder.CreateAnd(A, NewOr);
533     // We can't use C as zero because we might actually handle
534     //   (icmp ne (A & B), B) & (icmp ne (A & D), D)
535     // with B and D, having a single bit set.
536     Value *Zero = Constant::getNullValue(A->getType());
537     return Builder.CreateICmp(NewCC, NewAnd, Zero);
538   }
539   if (Mask & BMask_AllOnes) {
540     // (icmp eq (A & B), B) & (icmp eq (A & D), D)
541     // -> (icmp eq (A & (B|D)), (B|D))
542     if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D))
543       return nullptr; // TODO: Use freeze?
544     Value *NewOr = Builder.CreateOr(B, D);
545     Value *NewAnd = Builder.CreateAnd(A, NewOr);
546     return Builder.CreateICmp(NewCC, NewAnd, NewOr);
547   }
548   if (Mask & AMask_AllOnes) {
549     // (icmp eq (A & B), A) & (icmp eq (A & D), A)
550     // -> (icmp eq (A & (B&D)), A)
551     if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D))
552       return nullptr; // TODO: Use freeze?
553     Value *NewAnd1 = Builder.CreateAnd(B, D);
554     Value *NewAnd2 = Builder.CreateAnd(A, NewAnd1);
555     return Builder.CreateICmp(NewCC, NewAnd2, A);
556   }
557 
558   // Remaining cases assume at least that B and D are constant, and depend on
559   // their actual values. This isn't strictly necessary, just a "handle the
560   // easy cases for now" decision.
561   const APInt *ConstB, *ConstD;
562   if (!match(B, m_APInt(ConstB)) || !match(D, m_APInt(ConstD)))
563     return nullptr;
564 
565   if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) {
566     // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
567     // (icmp ne (A & B), B) & (icmp ne (A & D), D)
568     //     -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
569     // Only valid if one of the masks is a superset of the other (check "B&D" is
570     // the same as either B or D).
571     APInt NewMask = *ConstB & *ConstD;
572     if (NewMask == *ConstB)
573       return LHS;
574     else if (NewMask == *ConstD)
575       return RHS;
576   }
577 
578   if (Mask & AMask_NotAllOnes) {
579     // (icmp ne (A & B), B) & (icmp ne (A & D), D)
580     //     -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
581     // Only valid if one of the masks is a superset of the other (check "B|D" is
582     // the same as either B or D).
583     APInt NewMask = *ConstB | *ConstD;
584     if (NewMask == *ConstB)
585       return LHS;
586     else if (NewMask == *ConstD)
587       return RHS;
588   }
589 
590   if (Mask & (BMask_Mixed | BMask_NotMixed)) {
591     // Mixed:
592     // (icmp eq (A & B), C) & (icmp eq (A & D), E)
593     // We already know that B & C == C && D & E == E.
594     // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
595     // C and E, which are shared by both the mask B and the mask D, don't
596     // contradict, then we can transform to
597     // -> (icmp eq (A & (B|D)), (C|E))
598     // Currently, we only handle the case of B, C, D, and E being constant.
599     // We can't simply use C and E because we might actually handle
600     //   (icmp ne (A & B), B) & (icmp eq (A & D), D)
601     // with B and D, having a single bit set.
602 
603     // NotMixed:
604     // (icmp ne (A & B), C) & (icmp ne (A & D), E)
605     // -> (icmp ne (A & (B & D)), (C & E))
606     // Check the intersection (B & D) for inequality.
607     // Assume that (B & D) == B || (B & D) == D, i.e B/D is a subset of D/B
608     // and (B & D) & (C ^ E) == 0, bits of C and E, which are shared by both the
609     // B and the D, don't contradict.
610     // Note that we can assume (~B & C) == 0 && (~D & E) == 0, previous
611     // operation should delete these icmps if it hadn't been met.
612 
613     const APInt *OldConstC, *OldConstE;
614     if (!match(C, m_APInt(OldConstC)) || !match(E, m_APInt(OldConstE)))
615       return nullptr;
616 
617     auto FoldBMixed = [&](ICmpInst::Predicate CC, bool IsNot) -> Value * {
618       CC = IsNot ? CmpInst::getInversePredicate(CC) : CC;
619       const APInt ConstC = PredL != CC ? *ConstB ^ *OldConstC : *OldConstC;
620       const APInt ConstE = PredR != CC ? *ConstD ^ *OldConstE : *OldConstE;
621 
622       if (((*ConstB & *ConstD) & (ConstC ^ ConstE)).getBoolValue())
623         return IsNot ? nullptr : ConstantInt::get(LHS->getType(), !IsAnd);
624 
625       if (IsNot && !ConstB->isSubsetOf(*ConstD) && !ConstD->isSubsetOf(*ConstB))
626         return nullptr;
627 
628       APInt BD, CE;
629       if (IsNot) {
630         BD = *ConstB & *ConstD;
631         CE = ConstC & ConstE;
632       } else {
633         BD = *ConstB | *ConstD;
634         CE = ConstC | ConstE;
635       }
636       Value *NewAnd = Builder.CreateAnd(A, BD);
637       Value *CEVal = ConstantInt::get(A->getType(), CE);
638       return Builder.CreateICmp(CC, CEVal, NewAnd);
639     };
640 
641     if (Mask & BMask_Mixed)
642       return FoldBMixed(NewCC, false);
643     if (Mask & BMask_NotMixed) // can be else also
644       return FoldBMixed(NewCC, true);
645   }
646   return nullptr;
647 }
648 
649 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
650 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
651 /// If \p Inverted is true then the check is for the inverted range, e.g.
652 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
653 Value *InstCombinerImpl::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
654                                             bool Inverted) {
655   // Check the lower range comparison, e.g. x >= 0
656   // InstCombine already ensured that if there is a constant it's on the RHS.
657   ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
658   if (!RangeStart)
659     return nullptr;
660 
661   ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
662                                Cmp0->getPredicate());
663 
664   // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
665   if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
666         (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
667     return nullptr;
668 
669   ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
670                                Cmp1->getPredicate());
671 
672   Value *Input = Cmp0->getOperand(0);
673   Value *RangeEnd;
674   if (Cmp1->getOperand(0) == Input) {
675     // For the upper range compare we have: icmp x, n
676     RangeEnd = Cmp1->getOperand(1);
677   } else if (Cmp1->getOperand(1) == Input) {
678     // For the upper range compare we have: icmp n, x
679     RangeEnd = Cmp1->getOperand(0);
680     Pred1 = ICmpInst::getSwappedPredicate(Pred1);
681   } else {
682     return nullptr;
683   }
684 
685   // Check the upper range comparison, e.g. x < n
686   ICmpInst::Predicate NewPred;
687   switch (Pred1) {
688     case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
689     case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
690     default: return nullptr;
691   }
692 
693   // This simplification is only valid if the upper range is not negative.
694   KnownBits Known = computeKnownBits(RangeEnd, /*Depth=*/0, Cmp1);
695   if (!Known.isNonNegative())
696     return nullptr;
697 
698   if (Inverted)
699     NewPred = ICmpInst::getInversePredicate(NewPred);
700 
701   return Builder.CreateICmp(NewPred, Input, RangeEnd);
702 }
703 
704 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
705 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2)
706 Value *InstCombinerImpl::foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS,
707                                                        ICmpInst *RHS,
708                                                        Instruction *CxtI,
709                                                        bool IsAnd,
710                                                        bool IsLogical) {
711   CmpInst::Predicate Pred = IsAnd ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
712   if (LHS->getPredicate() != Pred || RHS->getPredicate() != Pred)
713     return nullptr;
714 
715   if (!match(LHS->getOperand(1), m_Zero()) ||
716       !match(RHS->getOperand(1), m_Zero()))
717     return nullptr;
718 
719   Value *L1, *L2, *R1, *R2;
720   if (match(LHS->getOperand(0), m_And(m_Value(L1), m_Value(L2))) &&
721       match(RHS->getOperand(0), m_And(m_Value(R1), m_Value(R2)))) {
722     if (L1 == R2 || L2 == R2)
723       std::swap(R1, R2);
724     if (L2 == R1)
725       std::swap(L1, L2);
726 
727     if (L1 == R1 &&
728         isKnownToBeAPowerOfTwo(L2, false, 0, CxtI) &&
729         isKnownToBeAPowerOfTwo(R2, false, 0, CxtI)) {
730       // If this is a logical and/or, then we must prevent propagation of a
731       // poison value from the RHS by inserting freeze.
732       if (IsLogical)
733         R2 = Builder.CreateFreeze(R2);
734       Value *Mask = Builder.CreateOr(L2, R2);
735       Value *Masked = Builder.CreateAnd(L1, Mask);
736       auto NewPred = IsAnd ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
737       return Builder.CreateICmp(NewPred, Masked, Mask);
738     }
739   }
740 
741   return nullptr;
742 }
743 
744 /// General pattern:
745 ///   X & Y
746 ///
747 /// Where Y is checking that all the high bits (covered by a mask 4294967168)
748 /// are uniform, i.e.  %arg & 4294967168  can be either  4294967168  or  0
749 /// Pattern can be one of:
750 ///   %t = add        i32 %arg,    128
751 ///   %r = icmp   ult i32 %t,      256
752 /// Or
753 ///   %t0 = shl       i32 %arg,    24
754 ///   %t1 = ashr      i32 %t0,     24
755 ///   %r  = icmp  eq  i32 %t1,     %arg
756 /// Or
757 ///   %t0 = trunc     i32 %arg  to i8
758 ///   %t1 = sext      i8  %t0   to i32
759 ///   %r  = icmp  eq  i32 %t1,     %arg
760 /// This pattern is a signed truncation check.
761 ///
762 /// And X is checking that some bit in that same mask is zero.
763 /// I.e. can be one of:
764 ///   %r = icmp sgt i32   %arg,    -1
765 /// Or
766 ///   %t = and      i32   %arg,    2147483648
767 ///   %r = icmp eq  i32   %t,      0
768 ///
769 /// Since we are checking that all the bits in that mask are the same,
770 /// and a particular bit is zero, what we are really checking is that all the
771 /// masked bits are zero.
772 /// So this should be transformed to:
773 ///   %r = icmp ult i32 %arg, 128
774 static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1,
775                                         Instruction &CxtI,
776                                         InstCombiner::BuilderTy &Builder) {
777   assert(CxtI.getOpcode() == Instruction::And);
778 
779   // Match  icmp ult (add %arg, C01), C1   (C1 == C01 << 1; powers of two)
780   auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X,
781                                             APInt &SignBitMask) -> bool {
782     CmpInst::Predicate Pred;
783     const APInt *I01, *I1; // powers of two; I1 == I01 << 1
784     if (!(match(ICmp,
785                 m_ICmp(Pred, m_Add(m_Value(X), m_Power2(I01)), m_Power2(I1))) &&
786           Pred == ICmpInst::ICMP_ULT && I1->ugt(*I01) && I01->shl(1) == *I1))
787       return false;
788     // Which bit is the new sign bit as per the 'signed truncation' pattern?
789     SignBitMask = *I01;
790     return true;
791   };
792 
793   // One icmp needs to be 'signed truncation check'.
794   // We need to match this first, else we will mismatch commutative cases.
795   Value *X1;
796   APInt HighestBit;
797   ICmpInst *OtherICmp;
798   if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit))
799     OtherICmp = ICmp0;
800   else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit))
801     OtherICmp = ICmp1;
802   else
803     return nullptr;
804 
805   assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)");
806 
807   // Try to match/decompose into:  icmp eq (X & Mask), 0
808   auto tryToDecompose = [](ICmpInst *ICmp, Value *&X,
809                            APInt &UnsetBitsMask) -> bool {
810     CmpInst::Predicate Pred = ICmp->getPredicate();
811     // Can it be decomposed into  icmp eq (X & Mask), 0  ?
812     if (llvm::decomposeBitTestICmp(ICmp->getOperand(0), ICmp->getOperand(1),
813                                    Pred, X, UnsetBitsMask,
814                                    /*LookThroughTrunc=*/false) &&
815         Pred == ICmpInst::ICMP_EQ)
816       return true;
817     // Is it  icmp eq (X & Mask), 0  already?
818     const APInt *Mask;
819     if (match(ICmp, m_ICmp(Pred, m_And(m_Value(X), m_APInt(Mask)), m_Zero())) &&
820         Pred == ICmpInst::ICMP_EQ) {
821       UnsetBitsMask = *Mask;
822       return true;
823     }
824     return false;
825   };
826 
827   // And the other icmp needs to be decomposable into a bit test.
828   Value *X0;
829   APInt UnsetBitsMask;
830   if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask))
831     return nullptr;
832 
833   assert(!UnsetBitsMask.isZero() && "empty mask makes no sense.");
834 
835   // Are they working on the same value?
836   Value *X;
837   if (X1 == X0) {
838     // Ok as is.
839     X = X1;
840   } else if (match(X0, m_Trunc(m_Specific(X1)))) {
841     UnsetBitsMask = UnsetBitsMask.zext(X1->getType()->getScalarSizeInBits());
842     X = X1;
843   } else
844     return nullptr;
845 
846   // So which bits should be uniform as per the 'signed truncation check'?
847   // (all the bits starting with (i.e. including) HighestBit)
848   APInt SignBitsMask = ~(HighestBit - 1U);
849 
850   // UnsetBitsMask must have some common bits with SignBitsMask,
851   if (!UnsetBitsMask.intersects(SignBitsMask))
852     return nullptr;
853 
854   // Does UnsetBitsMask contain any bits outside of SignBitsMask?
855   if (!UnsetBitsMask.isSubsetOf(SignBitsMask)) {
856     APInt OtherHighestBit = (~UnsetBitsMask) + 1U;
857     if (!OtherHighestBit.isPowerOf2())
858       return nullptr;
859     HighestBit = APIntOps::umin(HighestBit, OtherHighestBit);
860   }
861   // Else, if it does not, then all is ok as-is.
862 
863   // %r = icmp ult %X, SignBit
864   return Builder.CreateICmpULT(X, ConstantInt::get(X->getType(), HighestBit),
865                                CxtI.getName() + ".simplified");
866 }
867 
868 /// Fold (icmp eq ctpop(X) 1) | (icmp eq X 0) into (icmp ult ctpop(X) 2) and
869 /// fold (icmp ne ctpop(X) 1) & (icmp ne X 0) into (icmp ugt ctpop(X) 1).
870 /// Also used for logical and/or, must be poison safe.
871 static Value *foldIsPowerOf2OrZero(ICmpInst *Cmp0, ICmpInst *Cmp1, bool IsAnd,
872                                    InstCombiner::BuilderTy &Builder) {
873   CmpInst::Predicate Pred0, Pred1;
874   Value *X;
875   if (!match(Cmp0, m_ICmp(Pred0, m_Intrinsic<Intrinsic::ctpop>(m_Value(X)),
876                           m_SpecificInt(1))) ||
877       !match(Cmp1, m_ICmp(Pred1, m_Specific(X), m_ZeroInt())))
878     return nullptr;
879 
880   Value *CtPop = Cmp0->getOperand(0);
881   if (IsAnd && Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_NE)
882     return Builder.CreateICmpUGT(CtPop, ConstantInt::get(CtPop->getType(), 1));
883   if (!IsAnd && Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_EQ)
884     return Builder.CreateICmpULT(CtPop, ConstantInt::get(CtPop->getType(), 2));
885 
886   return nullptr;
887 }
888 
889 /// Reduce a pair of compares that check if a value has exactly 1 bit set.
890 /// Also used for logical and/or, must be poison safe.
891 static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd,
892                              InstCombiner::BuilderTy &Builder) {
893   // Handle 'and' / 'or' commutation: make the equality check the first operand.
894   if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE)
895     std::swap(Cmp0, Cmp1);
896   else if (!JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_EQ)
897     std::swap(Cmp0, Cmp1);
898 
899   // (X != 0) && (ctpop(X) u< 2) --> ctpop(X) == 1
900   CmpInst::Predicate Pred0, Pred1;
901   Value *X;
902   if (JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) &&
903       match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),
904                          m_SpecificInt(2))) &&
905       Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_ULT) {
906     Value *CtPop = Cmp1->getOperand(0);
907     return Builder.CreateICmpEQ(CtPop, ConstantInt::get(CtPop->getType(), 1));
908   }
909   // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1
910   if (!JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) &&
911       match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),
912                          m_SpecificInt(1))) &&
913       Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_UGT) {
914     Value *CtPop = Cmp1->getOperand(0);
915     return Builder.CreateICmpNE(CtPop, ConstantInt::get(CtPop->getType(), 1));
916   }
917   return nullptr;
918 }
919 
920 /// Try to fold (icmp(A & B) == 0) & (icmp(A & D) != E) into (icmp A u< D) iff
921 /// B is a contiguous set of ones starting from the most significant bit
922 /// (negative power of 2), D and E are equal, and D is a contiguous set of ones
923 /// starting at the most significant zero bit in B. Parameter B supports masking
924 /// using undef/poison in either scalar or vector values.
925 static Value *foldNegativePower2AndShiftedMask(
926     Value *A, Value *B, Value *D, Value *E, ICmpInst::Predicate PredL,
927     ICmpInst::Predicate PredR, InstCombiner::BuilderTy &Builder) {
928   assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
929          "Expected equality predicates for masked type of icmps.");
930   if (PredL != ICmpInst::ICMP_EQ || PredR != ICmpInst::ICMP_NE)
931     return nullptr;
932 
933   if (!match(B, m_NegatedPower2()) || !match(D, m_ShiftedMask()) ||
934       !match(E, m_ShiftedMask()))
935     return nullptr;
936 
937   // Test scalar arguments for conversion. B has been validated earlier to be a
938   // negative power of two and thus is guaranteed to have one or more contiguous
939   // ones starting from the MSB followed by zero or more contiguous zeros. D has
940   // been validated earlier to be a shifted set of one or more contiguous ones.
941   // In order to match, B leading ones and D leading zeros should be equal. The
942   // predicate that B be a negative power of 2 prevents the condition of there
943   // ever being zero leading ones. Thus 0 == 0 cannot occur. The predicate that
944   // D always be a shifted mask prevents the condition of D equaling 0. This
945   // prevents matching the condition where B contains the maximum number of
946   // leading one bits (-1) and D contains the maximum number of leading zero
947   // bits (0).
948   auto isReducible = [](const Value *B, const Value *D, const Value *E) {
949     const APInt *BCst, *DCst, *ECst;
950     return match(B, m_APIntAllowUndef(BCst)) && match(D, m_APInt(DCst)) &&
951            match(E, m_APInt(ECst)) && *DCst == *ECst &&
952            (isa<UndefValue>(B) ||
953             (BCst->countLeadingOnes() == DCst->countLeadingZeros()));
954   };
955 
956   // Test vector type arguments for conversion.
957   if (const auto *BVTy = dyn_cast<VectorType>(B->getType())) {
958     const auto *BFVTy = dyn_cast<FixedVectorType>(BVTy);
959     const auto *BConst = dyn_cast<Constant>(B);
960     const auto *DConst = dyn_cast<Constant>(D);
961     const auto *EConst = dyn_cast<Constant>(E);
962 
963     if (!BFVTy || !BConst || !DConst || !EConst)
964       return nullptr;
965 
966     for (unsigned I = 0; I != BFVTy->getNumElements(); ++I) {
967       const auto *BElt = BConst->getAggregateElement(I);
968       const auto *DElt = DConst->getAggregateElement(I);
969       const auto *EElt = EConst->getAggregateElement(I);
970 
971       if (!BElt || !DElt || !EElt)
972         return nullptr;
973       if (!isReducible(BElt, DElt, EElt))
974         return nullptr;
975     }
976   } else {
977     // Test scalar type arguments for conversion.
978     if (!isReducible(B, D, E))
979       return nullptr;
980   }
981   return Builder.CreateICmp(ICmpInst::ICMP_ULT, A, D);
982 }
983 
984 /// Try to fold ((icmp X u< P) & (icmp(X & M) != M)) or ((icmp X s> -1) &
985 /// (icmp(X & M) != M)) into (icmp X u< M). Where P is a power of 2, M < P, and
986 /// M is a contiguous shifted mask starting at the right most significant zero
987 /// bit in P. SGT is supported as when P is the largest representable power of
988 /// 2, an earlier optimization converts the expression into (icmp X s> -1).
989 /// Parameter P supports masking using undef/poison in either scalar or vector
990 /// values.
991 static Value *foldPowerOf2AndShiftedMask(ICmpInst *Cmp0, ICmpInst *Cmp1,
992                                          bool JoinedByAnd,
993                                          InstCombiner::BuilderTy &Builder) {
994   if (!JoinedByAnd)
995     return nullptr;
996   Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
997   ICmpInst::Predicate CmpPred0 = Cmp0->getPredicate(),
998                       CmpPred1 = Cmp1->getPredicate();
999   // Assuming P is a 2^n, getMaskedTypeForICmpPair will normalize (icmp X u<
1000   // 2^n) into (icmp (X & ~(2^n-1)) == 0) and (icmp X s> -1) into (icmp (X &
1001   // SignMask) == 0).
1002   std::optional<std::pair<unsigned, unsigned>> MaskPair =
1003       getMaskedTypeForICmpPair(A, B, C, D, E, Cmp0, Cmp1, CmpPred0, CmpPred1);
1004   if (!MaskPair)
1005     return nullptr;
1006 
1007   const auto compareBMask = BMask_NotMixed | BMask_NotAllOnes;
1008   unsigned CmpMask0 = MaskPair->first;
1009   unsigned CmpMask1 = MaskPair->second;
1010   if ((CmpMask0 & Mask_AllZeros) && (CmpMask1 == compareBMask)) {
1011     if (Value *V = foldNegativePower2AndShiftedMask(A, B, D, E, CmpPred0,
1012                                                     CmpPred1, Builder))
1013       return V;
1014   } else if ((CmpMask0 == compareBMask) && (CmpMask1 & Mask_AllZeros)) {
1015     if (Value *V = foldNegativePower2AndShiftedMask(A, D, B, C, CmpPred1,
1016                                                     CmpPred0, Builder))
1017       return V;
1018   }
1019   return nullptr;
1020 }
1021 
1022 /// Commuted variants are assumed to be handled by calling this function again
1023 /// with the parameters swapped.
1024 static Value *foldUnsignedUnderflowCheck(ICmpInst *ZeroICmp,
1025                                          ICmpInst *UnsignedICmp, bool IsAnd,
1026                                          const SimplifyQuery &Q,
1027                                          InstCombiner::BuilderTy &Builder) {
1028   Value *ZeroCmpOp;
1029   ICmpInst::Predicate EqPred;
1030   if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(ZeroCmpOp), m_Zero())) ||
1031       !ICmpInst::isEquality(EqPred))
1032     return nullptr;
1033 
1034   auto IsKnownNonZero = [&](Value *V) {
1035     return isKnownNonZero(V, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1036   };
1037 
1038   ICmpInst::Predicate UnsignedPred;
1039 
1040   Value *A, *B;
1041   if (match(UnsignedICmp,
1042             m_c_ICmp(UnsignedPred, m_Specific(ZeroCmpOp), m_Value(A))) &&
1043       match(ZeroCmpOp, m_c_Add(m_Specific(A), m_Value(B))) &&
1044       (ZeroICmp->hasOneUse() || UnsignedICmp->hasOneUse())) {
1045     auto GetKnownNonZeroAndOther = [&](Value *&NonZero, Value *&Other) {
1046       if (!IsKnownNonZero(NonZero))
1047         std::swap(NonZero, Other);
1048       return IsKnownNonZero(NonZero);
1049     };
1050 
1051     // Given  ZeroCmpOp = (A + B)
1052     //   ZeroCmpOp <  A && ZeroCmpOp != 0  -->  (0-X) <  Y  iff
1053     //   ZeroCmpOp >= A || ZeroCmpOp == 0  -->  (0-X) >= Y  iff
1054     //     with X being the value (A/B) that is known to be non-zero,
1055     //     and Y being remaining value.
1056     if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE &&
1057         IsAnd && GetKnownNonZeroAndOther(B, A))
1058       return Builder.CreateICmpULT(Builder.CreateNeg(B), A);
1059     if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ &&
1060         !IsAnd && GetKnownNonZeroAndOther(B, A))
1061       return Builder.CreateICmpUGE(Builder.CreateNeg(B), A);
1062   }
1063 
1064   return nullptr;
1065 }
1066 
1067 struct IntPart {
1068   Value *From;
1069   unsigned StartBit;
1070   unsigned NumBits;
1071 };
1072 
1073 /// Match an extraction of bits from an integer.
1074 static std::optional<IntPart> matchIntPart(Value *V) {
1075   Value *X;
1076   if (!match(V, m_OneUse(m_Trunc(m_Value(X)))))
1077     return std::nullopt;
1078 
1079   unsigned NumOriginalBits = X->getType()->getScalarSizeInBits();
1080   unsigned NumExtractedBits = V->getType()->getScalarSizeInBits();
1081   Value *Y;
1082   const APInt *Shift;
1083   // For a trunc(lshr Y, Shift) pattern, make sure we're only extracting bits
1084   // from Y, not any shifted-in zeroes.
1085   if (match(X, m_OneUse(m_LShr(m_Value(Y), m_APInt(Shift)))) &&
1086       Shift->ule(NumOriginalBits - NumExtractedBits))
1087     return {{Y, (unsigned)Shift->getZExtValue(), NumExtractedBits}};
1088   return {{X, 0, NumExtractedBits}};
1089 }
1090 
1091 /// Materialize an extraction of bits from an integer in IR.
1092 static Value *extractIntPart(const IntPart &P, IRBuilderBase &Builder) {
1093   Value *V = P.From;
1094   if (P.StartBit)
1095     V = Builder.CreateLShr(V, P.StartBit);
1096   Type *TruncTy = V->getType()->getWithNewBitWidth(P.NumBits);
1097   if (TruncTy != V->getType())
1098     V = Builder.CreateTrunc(V, TruncTy);
1099   return V;
1100 }
1101 
1102 /// (icmp eq X0, Y0) & (icmp eq X1, Y1) -> icmp eq X01, Y01
1103 /// (icmp ne X0, Y0) | (icmp ne X1, Y1) -> icmp ne X01, Y01
1104 /// where X0, X1 and Y0, Y1 are adjacent parts extracted from an integer.
1105 Value *InstCombinerImpl::foldEqOfParts(ICmpInst *Cmp0, ICmpInst *Cmp1,
1106                                        bool IsAnd) {
1107   if (!Cmp0->hasOneUse() || !Cmp1->hasOneUse())
1108     return nullptr;
1109 
1110   CmpInst::Predicate Pred = IsAnd ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE;
1111   auto GetMatchPart = [&](ICmpInst *Cmp,
1112                           unsigned OpNo) -> std::optional<IntPart> {
1113     if (Pred == Cmp->getPredicate())
1114       return matchIntPart(Cmp->getOperand(OpNo));
1115 
1116     const APInt *C;
1117     // (icmp eq (lshr x, C), (lshr y, C)) gets optimized to:
1118     // (icmp ult (xor x, y), 1 << C) so also look for that.
1119     if (Pred == CmpInst::ICMP_EQ && Cmp->getPredicate() == CmpInst::ICMP_ULT) {
1120       if (!match(Cmp->getOperand(1), m_Power2(C)) ||
1121           !match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value())))
1122         return std::nullopt;
1123     }
1124 
1125     // (icmp ne (lshr x, C), (lshr y, C)) gets optimized to:
1126     // (icmp ugt (xor x, y), (1 << C) - 1) so also look for that.
1127     else if (Pred == CmpInst::ICMP_NE &&
1128              Cmp->getPredicate() == CmpInst::ICMP_UGT) {
1129       if (!match(Cmp->getOperand(1), m_LowBitMask(C)) ||
1130           !match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value())))
1131         return std::nullopt;
1132     } else {
1133       return std::nullopt;
1134     }
1135 
1136     unsigned From = Pred == CmpInst::ICMP_NE ? C->popcount() : C->countr_zero();
1137     Instruction *I = cast<Instruction>(Cmp->getOperand(0));
1138     return {{I->getOperand(OpNo), From, C->getBitWidth() - From}};
1139   };
1140 
1141   std::optional<IntPart> L0 = GetMatchPart(Cmp0, 0);
1142   std::optional<IntPart> R0 = GetMatchPart(Cmp0, 1);
1143   std::optional<IntPart> L1 = GetMatchPart(Cmp1, 0);
1144   std::optional<IntPart> R1 = GetMatchPart(Cmp1, 1);
1145   if (!L0 || !R0 || !L1 || !R1)
1146     return nullptr;
1147 
1148   // Make sure the LHS/RHS compare a part of the same value, possibly after
1149   // an operand swap.
1150   if (L0->From != L1->From || R0->From != R1->From) {
1151     if (L0->From != R1->From || R0->From != L1->From)
1152       return nullptr;
1153     std::swap(L1, R1);
1154   }
1155 
1156   // Make sure the extracted parts are adjacent, canonicalizing to L0/R0 being
1157   // the low part and L1/R1 being the high part.
1158   if (L0->StartBit + L0->NumBits != L1->StartBit ||
1159       R0->StartBit + R0->NumBits != R1->StartBit) {
1160     if (L1->StartBit + L1->NumBits != L0->StartBit ||
1161         R1->StartBit + R1->NumBits != R0->StartBit)
1162       return nullptr;
1163     std::swap(L0, L1);
1164     std::swap(R0, R1);
1165   }
1166 
1167   // We can simplify to a comparison of these larger parts of the integers.
1168   IntPart L = {L0->From, L0->StartBit, L0->NumBits + L1->NumBits};
1169   IntPart R = {R0->From, R0->StartBit, R0->NumBits + R1->NumBits};
1170   Value *LValue = extractIntPart(L, Builder);
1171   Value *RValue = extractIntPart(R, Builder);
1172   return Builder.CreateICmp(Pred, LValue, RValue);
1173 }
1174 
1175 /// Reduce logic-of-compares with equality to a constant by substituting a
1176 /// common operand with the constant. Callers are expected to call this with
1177 /// Cmp0/Cmp1 switched to handle logic op commutativity.
1178 static Value *foldAndOrOfICmpsWithConstEq(ICmpInst *Cmp0, ICmpInst *Cmp1,
1179                                           bool IsAnd, bool IsLogical,
1180                                           InstCombiner::BuilderTy &Builder,
1181                                           const SimplifyQuery &Q) {
1182   // Match an equality compare with a non-poison constant as Cmp0.
1183   // Also, give up if the compare can be constant-folded to avoid looping.
1184   ICmpInst::Predicate Pred0;
1185   Value *X;
1186   Constant *C;
1187   if (!match(Cmp0, m_ICmp(Pred0, m_Value(X), m_Constant(C))) ||
1188       !isGuaranteedNotToBeUndefOrPoison(C) || isa<Constant>(X))
1189     return nullptr;
1190   if ((IsAnd && Pred0 != ICmpInst::ICMP_EQ) ||
1191       (!IsAnd && Pred0 != ICmpInst::ICMP_NE))
1192     return nullptr;
1193 
1194   // The other compare must include a common operand (X). Canonicalize the
1195   // common operand as operand 1 (Pred1 is swapped if the common operand was
1196   // operand 0).
1197   Value *Y;
1198   ICmpInst::Predicate Pred1;
1199   if (!match(Cmp1, m_c_ICmp(Pred1, m_Value(Y), m_Deferred(X))))
1200     return nullptr;
1201 
1202   // Replace variable with constant value equivalence to remove a variable use:
1203   // (X == C) && (Y Pred1 X) --> (X == C) && (Y Pred1 C)
1204   // (X != C) || (Y Pred1 X) --> (X != C) || (Y Pred1 C)
1205   // Can think of the 'or' substitution with the 'and' bool equivalent:
1206   // A || B --> A || (!A && B)
1207   Value *SubstituteCmp = simplifyICmpInst(Pred1, Y, C, Q);
1208   if (!SubstituteCmp) {
1209     // If we need to create a new instruction, require that the old compare can
1210     // be removed.
1211     if (!Cmp1->hasOneUse())
1212       return nullptr;
1213     SubstituteCmp = Builder.CreateICmp(Pred1, Y, C);
1214   }
1215   if (IsLogical)
1216     return IsAnd ? Builder.CreateLogicalAnd(Cmp0, SubstituteCmp)
1217                  : Builder.CreateLogicalOr(Cmp0, SubstituteCmp);
1218   return Builder.CreateBinOp(IsAnd ? Instruction::And : Instruction::Or, Cmp0,
1219                              SubstituteCmp);
1220 }
1221 
1222 /// Fold (icmp Pred1 V1, C1) & (icmp Pred2 V2, C2)
1223 /// or   (icmp Pred1 V1, C1) | (icmp Pred2 V2, C2)
1224 /// into a single comparison using range-based reasoning.
1225 /// NOTE: This is also used for logical and/or, must be poison-safe!
1226 Value *InstCombinerImpl::foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1,
1227                                                      ICmpInst *ICmp2,
1228                                                      bool IsAnd) {
1229   ICmpInst::Predicate Pred1, Pred2;
1230   Value *V1, *V2;
1231   const APInt *C1, *C2;
1232   if (!match(ICmp1, m_ICmp(Pred1, m_Value(V1), m_APInt(C1))) ||
1233       !match(ICmp2, m_ICmp(Pred2, m_Value(V2), m_APInt(C2))))
1234     return nullptr;
1235 
1236   // Look through add of a constant offset on V1, V2, or both operands. This
1237   // allows us to interpret the V + C' < C'' range idiom into a proper range.
1238   const APInt *Offset1 = nullptr, *Offset2 = nullptr;
1239   if (V1 != V2) {
1240     Value *X;
1241     if (match(V1, m_Add(m_Value(X), m_APInt(Offset1))))
1242       V1 = X;
1243     if (match(V2, m_Add(m_Value(X), m_APInt(Offset2))))
1244       V2 = X;
1245   }
1246 
1247   if (V1 != V2)
1248     return nullptr;
1249 
1250   ConstantRange CR1 = ConstantRange::makeExactICmpRegion(
1251       IsAnd ? ICmpInst::getInversePredicate(Pred1) : Pred1, *C1);
1252   if (Offset1)
1253     CR1 = CR1.subtract(*Offset1);
1254 
1255   ConstantRange CR2 = ConstantRange::makeExactICmpRegion(
1256       IsAnd ? ICmpInst::getInversePredicate(Pred2) : Pred2, *C2);
1257   if (Offset2)
1258     CR2 = CR2.subtract(*Offset2);
1259 
1260   Type *Ty = V1->getType();
1261   Value *NewV = V1;
1262   std::optional<ConstantRange> CR = CR1.exactUnionWith(CR2);
1263   if (!CR) {
1264     if (!(ICmp1->hasOneUse() && ICmp2->hasOneUse()) || CR1.isWrappedSet() ||
1265         CR2.isWrappedSet())
1266       return nullptr;
1267 
1268     // Check whether we have equal-size ranges that only differ by one bit.
1269     // In that case we can apply a mask to map one range onto the other.
1270     APInt LowerDiff = CR1.getLower() ^ CR2.getLower();
1271     APInt UpperDiff = (CR1.getUpper() - 1) ^ (CR2.getUpper() - 1);
1272     APInt CR1Size = CR1.getUpper() - CR1.getLower();
1273     if (!LowerDiff.isPowerOf2() || LowerDiff != UpperDiff ||
1274         CR1Size != CR2.getUpper() - CR2.getLower())
1275       return nullptr;
1276 
1277     CR = CR1.getLower().ult(CR2.getLower()) ? CR1 : CR2;
1278     NewV = Builder.CreateAnd(NewV, ConstantInt::get(Ty, ~LowerDiff));
1279   }
1280 
1281   if (IsAnd)
1282     CR = CR->inverse();
1283 
1284   CmpInst::Predicate NewPred;
1285   APInt NewC, Offset;
1286   CR->getEquivalentICmp(NewPred, NewC, Offset);
1287 
1288   if (Offset != 0)
1289     NewV = Builder.CreateAdd(NewV, ConstantInt::get(Ty, Offset));
1290   return Builder.CreateICmp(NewPred, NewV, ConstantInt::get(Ty, NewC));
1291 }
1292 
1293 /// Ignore all operations which only change the sign of a value, returning the
1294 /// underlying magnitude value.
1295 static Value *stripSignOnlyFPOps(Value *Val) {
1296   match(Val, m_FNeg(m_Value(Val)));
1297   match(Val, m_FAbs(m_Value(Val)));
1298   match(Val, m_CopySign(m_Value(Val), m_Value()));
1299   return Val;
1300 }
1301 
1302 /// Matches canonical form of isnan, fcmp ord x, 0
1303 static bool matchIsNotNaN(FCmpInst::Predicate P, Value *LHS, Value *RHS) {
1304   return P == FCmpInst::FCMP_ORD && match(RHS, m_AnyZeroFP());
1305 }
1306 
1307 /// Matches fcmp u__ x, +/-inf
1308 static bool matchUnorderedInfCompare(FCmpInst::Predicate P, Value *LHS,
1309                                      Value *RHS) {
1310   return FCmpInst::isUnordered(P) && match(RHS, m_Inf());
1311 }
1312 
1313 /// and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf
1314 ///
1315 /// Clang emits this pattern for doing an isfinite check in __builtin_isnormal.
1316 static Value *matchIsFiniteTest(InstCombiner::BuilderTy &Builder, FCmpInst *LHS,
1317                                 FCmpInst *RHS) {
1318   Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1319   Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1320   FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1321 
1322   if (!matchIsNotNaN(PredL, LHS0, LHS1) ||
1323       !matchUnorderedInfCompare(PredR, RHS0, RHS1))
1324     return nullptr;
1325 
1326   IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1327   FastMathFlags FMF = LHS->getFastMathFlags();
1328   FMF &= RHS->getFastMathFlags();
1329   Builder.setFastMathFlags(FMF);
1330 
1331   return Builder.CreateFCmp(FCmpInst::getOrderedPredicate(PredR), RHS0, RHS1);
1332 }
1333 
1334 Value *InstCombinerImpl::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS,
1335                                           bool IsAnd, bool IsLogicalSelect) {
1336   Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1337   Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1338   FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1339 
1340   if (LHS0 == RHS1 && RHS0 == LHS1) {
1341     // Swap RHS operands to match LHS.
1342     PredR = FCmpInst::getSwappedPredicate(PredR);
1343     std::swap(RHS0, RHS1);
1344   }
1345 
1346   // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1347   // Suppose the relation between x and y is R, where R is one of
1348   // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for
1349   // testing the desired relations.
1350   //
1351   // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1352   //    bool(R & CC0) && bool(R & CC1)
1353   //  = bool((R & CC0) & (R & CC1))
1354   //  = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency
1355   //
1356   // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1357   //    bool(R & CC0) || bool(R & CC1)
1358   //  = bool((R & CC0) | (R & CC1))
1359   //  = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)
1360   if (LHS0 == RHS0 && LHS1 == RHS1) {
1361     unsigned FCmpCodeL = getFCmpCode(PredL);
1362     unsigned FCmpCodeR = getFCmpCode(PredR);
1363     unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR;
1364 
1365     // Intersect the fast math flags.
1366     // TODO: We can union the fast math flags unless this is a logical select.
1367     IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1368     FastMathFlags FMF = LHS->getFastMathFlags();
1369     FMF &= RHS->getFastMathFlags();
1370     Builder.setFastMathFlags(FMF);
1371 
1372     return getFCmpValue(NewPred, LHS0, LHS1, Builder);
1373   }
1374 
1375   // This transform is not valid for a logical select.
1376   if (!IsLogicalSelect &&
1377       ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1378        (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO &&
1379         !IsAnd))) {
1380     if (LHS0->getType() != RHS0->getType())
1381       return nullptr;
1382 
1383     // FCmp canonicalization ensures that (fcmp ord/uno X, X) and
1384     // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0).
1385     if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP()))
1386       // Ignore the constants because they are obviously not NANs:
1387       // (fcmp ord x, 0.0) & (fcmp ord y, 0.0)  -> (fcmp ord x, y)
1388       // (fcmp uno x, 0.0) | (fcmp uno y, 0.0)  -> (fcmp uno x, y)
1389       return Builder.CreateFCmp(PredL, LHS0, RHS0);
1390   }
1391 
1392   if (IsAnd && stripSignOnlyFPOps(LHS0) == stripSignOnlyFPOps(RHS0)) {
1393     // and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf
1394     // and (fcmp ord x, 0), (fcmp u* fabs(x), inf) -> fcmp o* x, inf
1395     if (Value *Left = matchIsFiniteTest(Builder, LHS, RHS))
1396       return Left;
1397     if (Value *Right = matchIsFiniteTest(Builder, RHS, LHS))
1398       return Right;
1399   }
1400 
1401   // Turn at least two fcmps with constants into llvm.is.fpclass.
1402   //
1403   // If we can represent a combined value test with one class call, we can
1404   // potentially eliminate 4-6 instructions. If we can represent a test with a
1405   // single fcmp with fneg and fabs, that's likely a better canonical form.
1406   if (LHS->hasOneUse() && RHS->hasOneUse()) {
1407     auto [ClassValRHS, ClassMaskRHS] =
1408         fcmpToClassTest(PredR, *RHS->getFunction(), RHS0, RHS1);
1409     if (ClassValRHS) {
1410       auto [ClassValLHS, ClassMaskLHS] =
1411           fcmpToClassTest(PredL, *LHS->getFunction(), LHS0, LHS1);
1412       if (ClassValLHS == ClassValRHS) {
1413         unsigned CombinedMask = IsAnd ? (ClassMaskLHS & ClassMaskRHS)
1414                                       : (ClassMaskLHS | ClassMaskRHS);
1415         return Builder.CreateIntrinsic(
1416             Intrinsic::is_fpclass, {ClassValLHS->getType()},
1417             {ClassValLHS, Builder.getInt32(CombinedMask)});
1418       }
1419     }
1420   }
1421 
1422   return nullptr;
1423 }
1424 
1425 /// Match an fcmp against a special value that performs a test possible by
1426 /// llvm.is.fpclass.
1427 static bool matchIsFPClassLikeFCmp(Value *Op, Value *&ClassVal,
1428                                    uint64_t &ClassMask) {
1429   auto *FCmp = dyn_cast<FCmpInst>(Op);
1430   if (!FCmp || !FCmp->hasOneUse())
1431     return false;
1432 
1433   std::tie(ClassVal, ClassMask) =
1434       fcmpToClassTest(FCmp->getPredicate(), *FCmp->getParent()->getParent(),
1435                       FCmp->getOperand(0), FCmp->getOperand(1));
1436   return ClassVal != nullptr;
1437 }
1438 
1439 /// or (is_fpclass x, mask0), (is_fpclass x, mask1)
1440 ///     -> is_fpclass x, (mask0 | mask1)
1441 /// and (is_fpclass x, mask0), (is_fpclass x, mask1)
1442 ///     -> is_fpclass x, (mask0 & mask1)
1443 /// xor (is_fpclass x, mask0), (is_fpclass x, mask1)
1444 ///     -> is_fpclass x, (mask0 ^ mask1)
1445 Instruction *InstCombinerImpl::foldLogicOfIsFPClass(BinaryOperator &BO,
1446                                                     Value *Op0, Value *Op1) {
1447   Value *ClassVal0 = nullptr;
1448   Value *ClassVal1 = nullptr;
1449   uint64_t ClassMask0, ClassMask1;
1450 
1451   // Restrict to folding one fcmp into one is.fpclass for now, don't introduce a
1452   // new class.
1453   //
1454   // TODO: Support forming is.fpclass out of 2 separate fcmps when codegen is
1455   // better.
1456 
1457   bool IsLHSClass =
1458       match(Op0, m_OneUse(m_Intrinsic<Intrinsic::is_fpclass>(
1459                      m_Value(ClassVal0), m_ConstantInt(ClassMask0))));
1460   bool IsRHSClass =
1461       match(Op1, m_OneUse(m_Intrinsic<Intrinsic::is_fpclass>(
1462                      m_Value(ClassVal1), m_ConstantInt(ClassMask1))));
1463   if ((((IsLHSClass || matchIsFPClassLikeFCmp(Op0, ClassVal0, ClassMask0)) &&
1464         (IsRHSClass || matchIsFPClassLikeFCmp(Op1, ClassVal1, ClassMask1)))) &&
1465       ClassVal0 == ClassVal1) {
1466     unsigned NewClassMask;
1467     switch (BO.getOpcode()) {
1468     case Instruction::And:
1469       NewClassMask = ClassMask0 & ClassMask1;
1470       break;
1471     case Instruction::Or:
1472       NewClassMask = ClassMask0 | ClassMask1;
1473       break;
1474     case Instruction::Xor:
1475       NewClassMask = ClassMask0 ^ ClassMask1;
1476       break;
1477     default:
1478       llvm_unreachable("not a binary logic operator");
1479     }
1480 
1481     if (IsLHSClass) {
1482       auto *II = cast<IntrinsicInst>(Op0);
1483       II->setArgOperand(
1484           1, ConstantInt::get(II->getArgOperand(1)->getType(), NewClassMask));
1485       return replaceInstUsesWith(BO, II);
1486     }
1487 
1488     if (IsRHSClass) {
1489       auto *II = cast<IntrinsicInst>(Op1);
1490       II->setArgOperand(
1491           1, ConstantInt::get(II->getArgOperand(1)->getType(), NewClassMask));
1492       return replaceInstUsesWith(BO, II);
1493     }
1494 
1495     CallInst *NewClass =
1496         Builder.CreateIntrinsic(Intrinsic::is_fpclass, {ClassVal0->getType()},
1497                                 {ClassVal0, Builder.getInt32(NewClassMask)});
1498     return replaceInstUsesWith(BO, NewClass);
1499   }
1500 
1501   return nullptr;
1502 }
1503 
1504 /// Look for the pattern that conditionally negates a value via math operations:
1505 ///   cond.splat = sext i1 cond
1506 ///   sub = add cond.splat, x
1507 ///   xor = xor sub, cond.splat
1508 /// and rewrite it to do the same, but via logical operations:
1509 ///   value.neg = sub 0, value
1510 ///   cond = select i1 neg, value.neg, value
1511 Instruction *InstCombinerImpl::canonicalizeConditionalNegationViaMathToSelect(
1512     BinaryOperator &I) {
1513   assert(I.getOpcode() == BinaryOperator::Xor && "Only for xor!");
1514   Value *Cond, *X;
1515   // As per complexity ordering, `xor` is not commutative here.
1516   if (!match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())) ||
1517       !match(I.getOperand(1), m_SExt(m_Value(Cond))) ||
1518       !Cond->getType()->isIntOrIntVectorTy(1) ||
1519       !match(I.getOperand(0), m_c_Add(m_SExt(m_Deferred(Cond)), m_Value(X))))
1520     return nullptr;
1521   return SelectInst::Create(Cond, Builder.CreateNeg(X, X->getName() + ".neg"),
1522                             X);
1523 }
1524 
1525 /// This a limited reassociation for a special case (see above) where we are
1526 /// checking if two values are either both NAN (unordered) or not-NAN (ordered).
1527 /// This could be handled more generally in '-reassociation', but it seems like
1528 /// an unlikely pattern for a large number of logic ops and fcmps.
1529 static Instruction *reassociateFCmps(BinaryOperator &BO,
1530                                      InstCombiner::BuilderTy &Builder) {
1531   Instruction::BinaryOps Opcode = BO.getOpcode();
1532   assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1533          "Expecting and/or op for fcmp transform");
1534 
1535   // There are 4 commuted variants of the pattern. Canonicalize operands of this
1536   // logic op so an fcmp is operand 0 and a matching logic op is operand 1.
1537   Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1), *X;
1538   FCmpInst::Predicate Pred;
1539   if (match(Op1, m_FCmp(Pred, m_Value(), m_AnyZeroFP())))
1540     std::swap(Op0, Op1);
1541 
1542   // Match inner binop and the predicate for combining 2 NAN checks into 1.
1543   Value *BO10, *BO11;
1544   FCmpInst::Predicate NanPred = Opcode == Instruction::And ? FCmpInst::FCMP_ORD
1545                                                            : FCmpInst::FCMP_UNO;
1546   if (!match(Op0, m_FCmp(Pred, m_Value(X), m_AnyZeroFP())) || Pred != NanPred ||
1547       !match(Op1, m_BinOp(Opcode, m_Value(BO10), m_Value(BO11))))
1548     return nullptr;
1549 
1550   // The inner logic op must have a matching fcmp operand.
1551   Value *Y;
1552   if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) ||
1553       Pred != NanPred || X->getType() != Y->getType())
1554     std::swap(BO10, BO11);
1555 
1556   if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) ||
1557       Pred != NanPred || X->getType() != Y->getType())
1558     return nullptr;
1559 
1560   // and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z
1561   // or  (fcmp uno X, 0), (or  (fcmp uno Y, 0), Z) --> or  (fcmp uno X, Y), Z
1562   Value *NewFCmp = Builder.CreateFCmp(Pred, X, Y);
1563   if (auto *NewFCmpInst = dyn_cast<FCmpInst>(NewFCmp)) {
1564     // Intersect FMF from the 2 source fcmps.
1565     NewFCmpInst->copyIRFlags(Op0);
1566     NewFCmpInst->andIRFlags(BO10);
1567   }
1568   return BinaryOperator::Create(Opcode, NewFCmp, BO11);
1569 }
1570 
1571 /// Match variations of De Morgan's Laws:
1572 /// (~A & ~B) == (~(A | B))
1573 /// (~A | ~B) == (~(A & B))
1574 static Instruction *matchDeMorgansLaws(BinaryOperator &I,
1575                                        InstCombiner &IC) {
1576   const Instruction::BinaryOps Opcode = I.getOpcode();
1577   assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1578          "Trying to match De Morgan's Laws with something other than and/or");
1579 
1580   // Flip the logic operation.
1581   const Instruction::BinaryOps FlippedOpcode =
1582       (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;
1583 
1584   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1585   Value *A, *B;
1586   if (match(Op0, m_OneUse(m_Not(m_Value(A)))) &&
1587       match(Op1, m_OneUse(m_Not(m_Value(B)))) &&
1588       !IC.isFreeToInvert(A, A->hasOneUse()) &&
1589       !IC.isFreeToInvert(B, B->hasOneUse())) {
1590     Value *AndOr =
1591         IC.Builder.CreateBinOp(FlippedOpcode, A, B, I.getName() + ".demorgan");
1592     return BinaryOperator::CreateNot(AndOr);
1593   }
1594 
1595   // The 'not' ops may require reassociation.
1596   // (A & ~B) & ~C --> A & ~(B | C)
1597   // (~B & A) & ~C --> A & ~(B | C)
1598   // (A | ~B) | ~C --> A | ~(B & C)
1599   // (~B | A) | ~C --> A | ~(B & C)
1600   Value *C;
1601   if (match(Op0, m_OneUse(m_c_BinOp(Opcode, m_Value(A), m_Not(m_Value(B))))) &&
1602       match(Op1, m_Not(m_Value(C)))) {
1603     Value *FlippedBO = IC.Builder.CreateBinOp(FlippedOpcode, B, C);
1604     return BinaryOperator::Create(Opcode, A, IC.Builder.CreateNot(FlippedBO));
1605   }
1606 
1607   return nullptr;
1608 }
1609 
1610 bool InstCombinerImpl::shouldOptimizeCast(CastInst *CI) {
1611   Value *CastSrc = CI->getOperand(0);
1612 
1613   // Noop casts and casts of constants should be eliminated trivially.
1614   if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc))
1615     return false;
1616 
1617   // If this cast is paired with another cast that can be eliminated, we prefer
1618   // to have it eliminated.
1619   if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc))
1620     if (isEliminableCastPair(PrecedingCI, CI))
1621       return false;
1622 
1623   return true;
1624 }
1625 
1626 /// Fold {and,or,xor} (cast X), C.
1627 static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast,
1628                                           InstCombinerImpl &IC) {
1629   Constant *C = dyn_cast<Constant>(Logic.getOperand(1));
1630   if (!C)
1631     return nullptr;
1632 
1633   auto LogicOpc = Logic.getOpcode();
1634   Type *DestTy = Logic.getType();
1635   Type *SrcTy = Cast->getSrcTy();
1636 
1637   // Move the logic operation ahead of a zext or sext if the constant is
1638   // unchanged in the smaller source type. Performing the logic in a smaller
1639   // type may provide more information to later folds, and the smaller logic
1640   // instruction may be cheaper (particularly in the case of vectors).
1641   Value *X;
1642   if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) {
1643     if (Constant *TruncC = IC.getLosslessUnsignedTrunc(C, SrcTy)) {
1644       // LogicOpc (zext X), C --> zext (LogicOpc X, C)
1645       Value *NewOp = IC.Builder.CreateBinOp(LogicOpc, X, TruncC);
1646       return new ZExtInst(NewOp, DestTy);
1647     }
1648   }
1649 
1650   if (match(Cast, m_OneUse(m_SExt(m_Value(X))))) {
1651     if (Constant *TruncC = IC.getLosslessSignedTrunc(C, SrcTy)) {
1652       // LogicOpc (sext X), C --> sext (LogicOpc X, C)
1653       Value *NewOp = IC.Builder.CreateBinOp(LogicOpc, X, TruncC);
1654       return new SExtInst(NewOp, DestTy);
1655     }
1656   }
1657 
1658   return nullptr;
1659 }
1660 
1661 /// Fold {and,or,xor} (cast X), Y.
1662 Instruction *InstCombinerImpl::foldCastedBitwiseLogic(BinaryOperator &I) {
1663   auto LogicOpc = I.getOpcode();
1664   assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding");
1665 
1666   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1667 
1668   // fold bitwise(A >> BW - 1, zext(icmp))     (BW is the scalar bits of the
1669   // type of A)
1670   //   -> bitwise(zext(A < 0), zext(icmp))
1671   //   -> zext(bitwise(A < 0, icmp))
1672   auto FoldBitwiseICmpZeroWithICmp = [&](Value *Op0,
1673                                          Value *Op1) -> Instruction * {
1674     ICmpInst::Predicate Pred;
1675     Value *A;
1676     bool IsMatched =
1677         match(Op0,
1678               m_OneUse(m_LShr(
1679                   m_Value(A),
1680                   m_SpecificInt(Op0->getType()->getScalarSizeInBits() - 1)))) &&
1681         match(Op1, m_OneUse(m_ZExt(m_ICmp(Pred, m_Value(), m_Value()))));
1682 
1683     if (!IsMatched)
1684       return nullptr;
1685 
1686     auto *ICmpL =
1687         Builder.CreateICmpSLT(A, Constant::getNullValue(A->getType()));
1688     auto *ICmpR = cast<ZExtInst>(Op1)->getOperand(0);
1689     auto *BitwiseOp = Builder.CreateBinOp(LogicOpc, ICmpL, ICmpR);
1690 
1691     return new ZExtInst(BitwiseOp, Op0->getType());
1692   };
1693 
1694   if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op0, Op1))
1695     return Ret;
1696 
1697   if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op1, Op0))
1698     return Ret;
1699 
1700   CastInst *Cast0 = dyn_cast<CastInst>(Op0);
1701   if (!Cast0)
1702     return nullptr;
1703 
1704   // This must be a cast from an integer or integer vector source type to allow
1705   // transformation of the logic operation to the source type.
1706   Type *DestTy = I.getType();
1707   Type *SrcTy = Cast0->getSrcTy();
1708   if (!SrcTy->isIntOrIntVectorTy())
1709     return nullptr;
1710 
1711   if (Instruction *Ret = foldLogicCastConstant(I, Cast0, *this))
1712     return Ret;
1713 
1714   CastInst *Cast1 = dyn_cast<CastInst>(Op1);
1715   if (!Cast1)
1716     return nullptr;
1717 
1718   // Both operands of the logic operation are casts. The casts must be the
1719   // same kind for reduction.
1720   Instruction::CastOps CastOpcode = Cast0->getOpcode();
1721   if (CastOpcode != Cast1->getOpcode())
1722     return nullptr;
1723 
1724   // If the source types do not match, but the casts are matching extends, we
1725   // can still narrow the logic op.
1726   if (SrcTy != Cast1->getSrcTy()) {
1727     Value *X, *Y;
1728     if (match(Cast0, m_OneUse(m_ZExtOrSExt(m_Value(X)))) &&
1729         match(Cast1, m_OneUse(m_ZExtOrSExt(m_Value(Y))))) {
1730       // Cast the narrower source to the wider source type.
1731       unsigned XNumBits = X->getType()->getScalarSizeInBits();
1732       unsigned YNumBits = Y->getType()->getScalarSizeInBits();
1733       if (XNumBits < YNumBits)
1734         X = Builder.CreateCast(CastOpcode, X, Y->getType());
1735       else
1736         Y = Builder.CreateCast(CastOpcode, Y, X->getType());
1737       // Do the logic op in the intermediate width, then widen more.
1738       Value *NarrowLogic = Builder.CreateBinOp(LogicOpc, X, Y);
1739       return CastInst::Create(CastOpcode, NarrowLogic, DestTy);
1740     }
1741 
1742     // Give up for other cast opcodes.
1743     return nullptr;
1744   }
1745 
1746   Value *Cast0Src = Cast0->getOperand(0);
1747   Value *Cast1Src = Cast1->getOperand(0);
1748 
1749   // fold logic(cast(A), cast(B)) -> cast(logic(A, B))
1750   if ((Cast0->hasOneUse() || Cast1->hasOneUse()) &&
1751       shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {
1752     Value *NewOp = Builder.CreateBinOp(LogicOpc, Cast0Src, Cast1Src,
1753                                        I.getName());
1754     return CastInst::Create(CastOpcode, NewOp, DestTy);
1755   }
1756 
1757   return nullptr;
1758 }
1759 
1760 static Instruction *foldAndToXor(BinaryOperator &I,
1761                                  InstCombiner::BuilderTy &Builder) {
1762   assert(I.getOpcode() == Instruction::And);
1763   Value *Op0 = I.getOperand(0);
1764   Value *Op1 = I.getOperand(1);
1765   Value *A, *B;
1766 
1767   // Operand complexity canonicalization guarantees that the 'or' is Op0.
1768   // (A | B) & ~(A & B) --> A ^ B
1769   // (A | B) & ~(B & A) --> A ^ B
1770   if (match(&I, m_BinOp(m_Or(m_Value(A), m_Value(B)),
1771                         m_Not(m_c_And(m_Deferred(A), m_Deferred(B))))))
1772     return BinaryOperator::CreateXor(A, B);
1773 
1774   // (A | ~B) & (~A | B) --> ~(A ^ B)
1775   // (A | ~B) & (B | ~A) --> ~(A ^ B)
1776   // (~B | A) & (~A | B) --> ~(A ^ B)
1777   // (~B | A) & (B | ~A) --> ~(A ^ B)
1778   if (Op0->hasOneUse() || Op1->hasOneUse())
1779     if (match(&I, m_BinOp(m_c_Or(m_Value(A), m_Not(m_Value(B))),
1780                           m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))
1781       return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1782 
1783   return nullptr;
1784 }
1785 
1786 static Instruction *foldOrToXor(BinaryOperator &I,
1787                                 InstCombiner::BuilderTy &Builder) {
1788   assert(I.getOpcode() == Instruction::Or);
1789   Value *Op0 = I.getOperand(0);
1790   Value *Op1 = I.getOperand(1);
1791   Value *A, *B;
1792 
1793   // Operand complexity canonicalization guarantees that the 'and' is Op0.
1794   // (A & B) | ~(A | B) --> ~(A ^ B)
1795   // (A & B) | ~(B | A) --> ~(A ^ B)
1796   if (Op0->hasOneUse() || Op1->hasOneUse())
1797     if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1798         match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
1799       return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1800 
1801   // Operand complexity canonicalization guarantees that the 'xor' is Op0.
1802   // (A ^ B) | ~(A | B) --> ~(A & B)
1803   // (A ^ B) | ~(B | A) --> ~(A & B)
1804   if (Op0->hasOneUse() || Op1->hasOneUse())
1805     if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1806         match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
1807       return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));
1808 
1809   // (A & ~B) | (~A & B) --> A ^ B
1810   // (A & ~B) | (B & ~A) --> A ^ B
1811   // (~B & A) | (~A & B) --> A ^ B
1812   // (~B & A) | (B & ~A) --> A ^ B
1813   if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
1814       match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B))))
1815     return BinaryOperator::CreateXor(A, B);
1816 
1817   return nullptr;
1818 }
1819 
1820 /// Return true if a constant shift amount is always less than the specified
1821 /// bit-width. If not, the shift could create poison in the narrower type.
1822 static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) {
1823   APInt Threshold(C->getType()->getScalarSizeInBits(), BitWidth);
1824   return match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold));
1825 }
1826 
1827 /// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and
1828 /// a common zext operand: and (binop (zext X), C), (zext X).
1829 Instruction *InstCombinerImpl::narrowMaskedBinOp(BinaryOperator &And) {
1830   // This transform could also apply to {or, and, xor}, but there are better
1831   // folds for those cases, so we don't expect those patterns here. AShr is not
1832   // handled because it should always be transformed to LShr in this sequence.
1833   // The subtract transform is different because it has a constant on the left.
1834   // Add/mul commute the constant to RHS; sub with constant RHS becomes add.
1835   Value *Op0 = And.getOperand(0), *Op1 = And.getOperand(1);
1836   Constant *C;
1837   if (!match(Op0, m_OneUse(m_Add(m_Specific(Op1), m_Constant(C)))) &&
1838       !match(Op0, m_OneUse(m_Mul(m_Specific(Op1), m_Constant(C)))) &&
1839       !match(Op0, m_OneUse(m_LShr(m_Specific(Op1), m_Constant(C)))) &&
1840       !match(Op0, m_OneUse(m_Shl(m_Specific(Op1), m_Constant(C)))) &&
1841       !match(Op0, m_OneUse(m_Sub(m_Constant(C), m_Specific(Op1)))))
1842     return nullptr;
1843 
1844   Value *X;
1845   if (!match(Op1, m_ZExt(m_Value(X))) || Op1->hasNUsesOrMore(3))
1846     return nullptr;
1847 
1848   Type *Ty = And.getType();
1849   if (!isa<VectorType>(Ty) && !shouldChangeType(Ty, X->getType()))
1850     return nullptr;
1851 
1852   // If we're narrowing a shift, the shift amount must be safe (less than the
1853   // width) in the narrower type. If the shift amount is greater, instsimplify
1854   // usually handles that case, but we can't guarantee/assert it.
1855   Instruction::BinaryOps Opc = cast<BinaryOperator>(Op0)->getOpcode();
1856   if (Opc == Instruction::LShr || Opc == Instruction::Shl)
1857     if (!canNarrowShiftAmt(C, X->getType()->getScalarSizeInBits()))
1858       return nullptr;
1859 
1860   // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X)
1861   // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X)
1862   Value *NewC = ConstantExpr::getTrunc(C, X->getType());
1863   Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, NewC, X)
1864                                          : Builder.CreateBinOp(Opc, X, NewC);
1865   return new ZExtInst(Builder.CreateAnd(NewBO, X), Ty);
1866 }
1867 
1868 /// Try folding relatively complex patterns for both And and Or operations
1869 /// with all And and Or swapped.
1870 static Instruction *foldComplexAndOrPatterns(BinaryOperator &I,
1871                                              InstCombiner::BuilderTy &Builder) {
1872   const Instruction::BinaryOps Opcode = I.getOpcode();
1873   assert(Opcode == Instruction::And || Opcode == Instruction::Or);
1874 
1875   // Flip the logic operation.
1876   const Instruction::BinaryOps FlippedOpcode =
1877       (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;
1878 
1879   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1880   Value *A, *B, *C, *X, *Y, *Dummy;
1881 
1882   // Match following expressions:
1883   // (~(A | B) & C)
1884   // (~(A & B) | C)
1885   // Captures X = ~(A | B) or ~(A & B)
1886   const auto matchNotOrAnd =
1887       [Opcode, FlippedOpcode](Value *Op, auto m_A, auto m_B, auto m_C,
1888                               Value *&X, bool CountUses = false) -> bool {
1889     if (CountUses && !Op->hasOneUse())
1890       return false;
1891 
1892     if (match(Op, m_c_BinOp(FlippedOpcode,
1893                             m_CombineAnd(m_Value(X),
1894                                          m_Not(m_c_BinOp(Opcode, m_A, m_B))),
1895                             m_C)))
1896       return !CountUses || X->hasOneUse();
1897 
1898     return false;
1899   };
1900 
1901   // (~(A | B) & C) | ... --> ...
1902   // (~(A & B) | C) & ... --> ...
1903   // TODO: One use checks are conservative. We just need to check that a total
1904   //       number of multiple used values does not exceed reduction
1905   //       in operations.
1906   if (matchNotOrAnd(Op0, m_Value(A), m_Value(B), m_Value(C), X)) {
1907     // (~(A | B) & C) | (~(A | C) & B) --> (B ^ C) & ~A
1908     // (~(A & B) | C) & (~(A & C) | B) --> ~((B ^ C) & A)
1909     if (matchNotOrAnd(Op1, m_Specific(A), m_Specific(C), m_Specific(B), Dummy,
1910                       true)) {
1911       Value *Xor = Builder.CreateXor(B, C);
1912       return (Opcode == Instruction::Or)
1913                  ? BinaryOperator::CreateAnd(Xor, Builder.CreateNot(A))
1914                  : BinaryOperator::CreateNot(Builder.CreateAnd(Xor, A));
1915     }
1916 
1917     // (~(A | B) & C) | (~(B | C) & A) --> (A ^ C) & ~B
1918     // (~(A & B) | C) & (~(B & C) | A) --> ~((A ^ C) & B)
1919     if (matchNotOrAnd(Op1, m_Specific(B), m_Specific(C), m_Specific(A), Dummy,
1920                       true)) {
1921       Value *Xor = Builder.CreateXor(A, C);
1922       return (Opcode == Instruction::Or)
1923                  ? BinaryOperator::CreateAnd(Xor, Builder.CreateNot(B))
1924                  : BinaryOperator::CreateNot(Builder.CreateAnd(Xor, B));
1925     }
1926 
1927     // (~(A | B) & C) | ~(A | C) --> ~((B & C) | A)
1928     // (~(A & B) | C) & ~(A & C) --> ~((B | C) & A)
1929     if (match(Op1, m_OneUse(m_Not(m_OneUse(
1930                        m_c_BinOp(Opcode, m_Specific(A), m_Specific(C)))))))
1931       return BinaryOperator::CreateNot(Builder.CreateBinOp(
1932           Opcode, Builder.CreateBinOp(FlippedOpcode, B, C), A));
1933 
1934     // (~(A | B) & C) | ~(B | C) --> ~((A & C) | B)
1935     // (~(A & B) | C) & ~(B & C) --> ~((A | C) & B)
1936     if (match(Op1, m_OneUse(m_Not(m_OneUse(
1937                        m_c_BinOp(Opcode, m_Specific(B), m_Specific(C)))))))
1938       return BinaryOperator::CreateNot(Builder.CreateBinOp(
1939           Opcode, Builder.CreateBinOp(FlippedOpcode, A, C), B));
1940 
1941     // (~(A | B) & C) | ~(C | (A ^ B)) --> ~((A | B) & (C | (A ^ B)))
1942     // Note, the pattern with swapped and/or is not handled because the
1943     // result is more undefined than a source:
1944     // (~(A & B) | C) & ~(C & (A ^ B)) --> (A ^ B ^ C) | ~(A | C) is invalid.
1945     if (Opcode == Instruction::Or && Op0->hasOneUse() &&
1946         match(Op1, m_OneUse(m_Not(m_CombineAnd(
1947                        m_Value(Y),
1948                        m_c_BinOp(Opcode, m_Specific(C),
1949                                  m_c_Xor(m_Specific(A), m_Specific(B)))))))) {
1950       // X = ~(A | B)
1951       // Y = (C | (A ^ B)
1952       Value *Or = cast<BinaryOperator>(X)->getOperand(0);
1953       return BinaryOperator::CreateNot(Builder.CreateAnd(Or, Y));
1954     }
1955   }
1956 
1957   // (~A & B & C) | ... --> ...
1958   // (~A | B | C) | ... --> ...
1959   // TODO: One use checks are conservative. We just need to check that a total
1960   //       number of multiple used values does not exceed reduction
1961   //       in operations.
1962   if (match(Op0,
1963             m_OneUse(m_c_BinOp(FlippedOpcode,
1964                                m_BinOp(FlippedOpcode, m_Value(B), m_Value(C)),
1965                                m_CombineAnd(m_Value(X), m_Not(m_Value(A)))))) ||
1966       match(Op0, m_OneUse(m_c_BinOp(
1967                      FlippedOpcode,
1968                      m_c_BinOp(FlippedOpcode, m_Value(C),
1969                                m_CombineAnd(m_Value(X), m_Not(m_Value(A)))),
1970                      m_Value(B))))) {
1971     // X = ~A
1972     // (~A & B & C) | ~(A | B | C) --> ~(A | (B ^ C))
1973     // (~A | B | C) & ~(A & B & C) --> (~A | (B ^ C))
1974     if (match(Op1, m_OneUse(m_Not(m_c_BinOp(
1975                        Opcode, m_c_BinOp(Opcode, m_Specific(A), m_Specific(B)),
1976                        m_Specific(C))))) ||
1977         match(Op1, m_OneUse(m_Not(m_c_BinOp(
1978                        Opcode, m_c_BinOp(Opcode, m_Specific(B), m_Specific(C)),
1979                        m_Specific(A))))) ||
1980         match(Op1, m_OneUse(m_Not(m_c_BinOp(
1981                        Opcode, m_c_BinOp(Opcode, m_Specific(A), m_Specific(C)),
1982                        m_Specific(B)))))) {
1983       Value *Xor = Builder.CreateXor(B, C);
1984       return (Opcode == Instruction::Or)
1985                  ? BinaryOperator::CreateNot(Builder.CreateOr(Xor, A))
1986                  : BinaryOperator::CreateOr(Xor, X);
1987     }
1988 
1989     // (~A & B & C) | ~(A | B) --> (C | ~B) & ~A
1990     // (~A | B | C) & ~(A & B) --> (C & ~B) | ~A
1991     if (match(Op1, m_OneUse(m_Not(m_OneUse(
1992                        m_c_BinOp(Opcode, m_Specific(A), m_Specific(B)))))))
1993       return BinaryOperator::Create(
1994           FlippedOpcode, Builder.CreateBinOp(Opcode, C, Builder.CreateNot(B)),
1995           X);
1996 
1997     // (~A & B & C) | ~(A | C) --> (B | ~C) & ~A
1998     // (~A | B | C) & ~(A & C) --> (B & ~C) | ~A
1999     if (match(Op1, m_OneUse(m_Not(m_OneUse(
2000                        m_c_BinOp(Opcode, m_Specific(A), m_Specific(C)))))))
2001       return BinaryOperator::Create(
2002           FlippedOpcode, Builder.CreateBinOp(Opcode, B, Builder.CreateNot(C)),
2003           X);
2004   }
2005 
2006   return nullptr;
2007 }
2008 
2009 /// Try to reassociate a pair of binops so that values with one use only are
2010 /// part of the same instruction. This may enable folds that are limited with
2011 /// multi-use restrictions and makes it more likely to match other patterns that
2012 /// are looking for a common operand.
2013 static Instruction *reassociateForUses(BinaryOperator &BO,
2014                                        InstCombinerImpl::BuilderTy &Builder) {
2015   Instruction::BinaryOps Opcode = BO.getOpcode();
2016   Value *X, *Y, *Z;
2017   if (match(&BO,
2018             m_c_BinOp(Opcode, m_OneUse(m_BinOp(Opcode, m_Value(X), m_Value(Y))),
2019                       m_OneUse(m_Value(Z))))) {
2020     if (!isa<Constant>(X) && !isa<Constant>(Y) && !isa<Constant>(Z)) {
2021       // (X op Y) op Z --> (Y op Z) op X
2022       if (!X->hasOneUse()) {
2023         Value *YZ = Builder.CreateBinOp(Opcode, Y, Z);
2024         return BinaryOperator::Create(Opcode, YZ, X);
2025       }
2026       // (X op Y) op Z --> (X op Z) op Y
2027       if (!Y->hasOneUse()) {
2028         Value *XZ = Builder.CreateBinOp(Opcode, X, Z);
2029         return BinaryOperator::Create(Opcode, XZ, Y);
2030       }
2031     }
2032   }
2033 
2034   return nullptr;
2035 }
2036 
2037 // Match
2038 // (X + C2) | C
2039 // (X + C2) ^ C
2040 // (X + C2) & C
2041 // and convert to do the bitwise logic first:
2042 // (X | C) + C2
2043 // (X ^ C) + C2
2044 // (X & C) + C2
2045 // iff bits affected by logic op are lower than last bit affected by math op
2046 static Instruction *canonicalizeLogicFirst(BinaryOperator &I,
2047                                            InstCombiner::BuilderTy &Builder) {
2048   Type *Ty = I.getType();
2049   Instruction::BinaryOps OpC = I.getOpcode();
2050   Value *Op0 = I.getOperand(0);
2051   Value *Op1 = I.getOperand(1);
2052   Value *X;
2053   const APInt *C, *C2;
2054 
2055   if (!(match(Op0, m_OneUse(m_Add(m_Value(X), m_APInt(C2)))) &&
2056         match(Op1, m_APInt(C))))
2057     return nullptr;
2058 
2059   unsigned Width = Ty->getScalarSizeInBits();
2060   unsigned LastOneMath = Width - C2->countr_zero();
2061 
2062   switch (OpC) {
2063   case Instruction::And:
2064     if (C->countl_one() < LastOneMath)
2065       return nullptr;
2066     break;
2067   case Instruction::Xor:
2068   case Instruction::Or:
2069     if (C->countl_zero() < LastOneMath)
2070       return nullptr;
2071     break;
2072   default:
2073     llvm_unreachable("Unexpected BinaryOp!");
2074   }
2075 
2076   Value *NewBinOp = Builder.CreateBinOp(OpC, X, ConstantInt::get(Ty, *C));
2077   return BinaryOperator::CreateWithCopiedFlags(Instruction::Add, NewBinOp,
2078                                                ConstantInt::get(Ty, *C2), Op0);
2079 }
2080 
2081 // binop(shift(ShiftedC1, ShAmt), shift(ShiftedC2, add(ShAmt, AddC))) ->
2082 // shift(binop(ShiftedC1, shift(ShiftedC2, AddC)), ShAmt)
2083 // where both shifts are the same and AddC is a valid shift amount.
2084 Instruction *InstCombinerImpl::foldBinOpOfDisplacedShifts(BinaryOperator &I) {
2085   assert((I.isBitwiseLogicOp() || I.getOpcode() == Instruction::Add) &&
2086          "Unexpected opcode");
2087 
2088   Value *ShAmt;
2089   Constant *ShiftedC1, *ShiftedC2, *AddC;
2090   Type *Ty = I.getType();
2091   unsigned BitWidth = Ty->getScalarSizeInBits();
2092   if (!match(&I, m_c_BinOp(m_Shift(m_ImmConstant(ShiftedC1), m_Value(ShAmt)),
2093                            m_Shift(m_ImmConstant(ShiftedC2),
2094                                    m_AddLike(m_Deferred(ShAmt),
2095                                              m_ImmConstant(AddC))))))
2096     return nullptr;
2097 
2098   // Make sure the add constant is a valid shift amount.
2099   if (!match(AddC,
2100              m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(BitWidth, BitWidth))))
2101     return nullptr;
2102 
2103   // Avoid constant expressions.
2104   auto *Op0Inst = dyn_cast<Instruction>(I.getOperand(0));
2105   auto *Op1Inst = dyn_cast<Instruction>(I.getOperand(1));
2106   if (!Op0Inst || !Op1Inst)
2107     return nullptr;
2108 
2109   // Both shifts must be the same.
2110   Instruction::BinaryOps ShiftOp =
2111       static_cast<Instruction::BinaryOps>(Op0Inst->getOpcode());
2112   if (ShiftOp != Op1Inst->getOpcode())
2113     return nullptr;
2114 
2115   // For adds, only left shifts are supported.
2116   if (I.getOpcode() == Instruction::Add && ShiftOp != Instruction::Shl)
2117     return nullptr;
2118 
2119   Value *NewC = Builder.CreateBinOp(
2120       I.getOpcode(), ShiftedC1, Builder.CreateBinOp(ShiftOp, ShiftedC2, AddC));
2121   return BinaryOperator::Create(ShiftOp, NewC, ShAmt);
2122 }
2123 
2124 // Fold and/or/xor with two equal intrinsic IDs:
2125 // bitwise(fshl (A, B, ShAmt), fshl(C, D, ShAmt))
2126 // -> fshl(bitwise(A, C), bitwise(B, D), ShAmt)
2127 // bitwise(fshr (A, B, ShAmt), fshr(C, D, ShAmt))
2128 // -> fshr(bitwise(A, C), bitwise(B, D), ShAmt)
2129 // bitwise(bswap(A), bswap(B)) -> bswap(bitwise(A, B))
2130 // bitwise(bswap(A), C) -> bswap(bitwise(A, bswap(C)))
2131 // bitwise(bitreverse(A), bitreverse(B)) -> bitreverse(bitwise(A, B))
2132 // bitwise(bitreverse(A), C) -> bitreverse(bitwise(A, bitreverse(C)))
2133 static Instruction *
2134 foldBitwiseLogicWithIntrinsics(BinaryOperator &I,
2135                                InstCombiner::BuilderTy &Builder) {
2136   assert(I.isBitwiseLogicOp() && "Should and/or/xor");
2137   if (!I.getOperand(0)->hasOneUse())
2138     return nullptr;
2139   IntrinsicInst *X = dyn_cast<IntrinsicInst>(I.getOperand(0));
2140   if (!X)
2141     return nullptr;
2142 
2143   IntrinsicInst *Y = dyn_cast<IntrinsicInst>(I.getOperand(1));
2144   if (Y && (!Y->hasOneUse() || X->getIntrinsicID() != Y->getIntrinsicID()))
2145     return nullptr;
2146 
2147   Intrinsic::ID IID = X->getIntrinsicID();
2148   const APInt *RHSC;
2149   // Try to match constant RHS.
2150   if (!Y && (!(IID == Intrinsic::bswap || IID == Intrinsic::bitreverse) ||
2151              !match(I.getOperand(1), m_APInt(RHSC))))
2152     return nullptr;
2153 
2154   switch (IID) {
2155   case Intrinsic::fshl:
2156   case Intrinsic::fshr: {
2157     if (X->getOperand(2) != Y->getOperand(2))
2158       return nullptr;
2159     Value *NewOp0 =
2160         Builder.CreateBinOp(I.getOpcode(), X->getOperand(0), Y->getOperand(0));
2161     Value *NewOp1 =
2162         Builder.CreateBinOp(I.getOpcode(), X->getOperand(1), Y->getOperand(1));
2163     Function *F = Intrinsic::getDeclaration(I.getModule(), IID, I.getType());
2164     return CallInst::Create(F, {NewOp0, NewOp1, X->getOperand(2)});
2165   }
2166   case Intrinsic::bswap:
2167   case Intrinsic::bitreverse: {
2168     Value *NewOp0 = Builder.CreateBinOp(
2169         I.getOpcode(), X->getOperand(0),
2170         Y ? Y->getOperand(0)
2171           : ConstantInt::get(I.getType(), IID == Intrinsic::bswap
2172                                               ? RHSC->byteSwap()
2173                                               : RHSC->reverseBits()));
2174     Function *F = Intrinsic::getDeclaration(I.getModule(), IID, I.getType());
2175     return CallInst::Create(F, {NewOp0});
2176   }
2177   default:
2178     return nullptr;
2179   }
2180 }
2181 
2182 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2183 // here. We should standardize that construct where it is needed or choose some
2184 // other way to ensure that commutated variants of patterns are not missed.
2185 Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
2186   Type *Ty = I.getType();
2187 
2188   if (Value *V = simplifyAndInst(I.getOperand(0), I.getOperand(1),
2189                                  SQ.getWithInstruction(&I)))
2190     return replaceInstUsesWith(I, V);
2191 
2192   if (SimplifyAssociativeOrCommutative(I))
2193     return &I;
2194 
2195   if (Instruction *X = foldVectorBinop(I))
2196     return X;
2197 
2198   if (Instruction *Phi = foldBinopWithPhiOperands(I))
2199     return Phi;
2200 
2201   // See if we can simplify any instructions used by the instruction whose sole
2202   // purpose is to compute bits we don't care about.
2203   if (SimplifyDemandedInstructionBits(I))
2204     return &I;
2205 
2206   // Do this before using distributive laws to catch simple and/or/not patterns.
2207   if (Instruction *Xor = foldAndToXor(I, Builder))
2208     return Xor;
2209 
2210   if (Instruction *X = foldComplexAndOrPatterns(I, Builder))
2211     return X;
2212 
2213   // (A|B)&(A|C) -> A|(B&C) etc
2214   if (Value *V = foldUsingDistributiveLaws(I))
2215     return replaceInstUsesWith(I, V);
2216 
2217   if (Instruction *R = foldBinOpShiftWithShift(I))
2218     return R;
2219 
2220   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2221 
2222   Value *X, *Y;
2223   if (match(Op0, m_OneUse(m_LogicalShift(m_One(), m_Value(X)))) &&
2224       match(Op1, m_One())) {
2225     // (1 << X) & 1 --> zext(X == 0)
2226     // (1 >> X) & 1 --> zext(X == 0)
2227     Value *IsZero = Builder.CreateICmpEQ(X, ConstantInt::get(Ty, 0));
2228     return new ZExtInst(IsZero, Ty);
2229   }
2230 
2231   // (-(X & 1)) & Y --> (X & 1) == 0 ? 0 : Y
2232   Value *Neg;
2233   if (match(&I,
2234             m_c_And(m_CombineAnd(m_Value(Neg),
2235                                  m_OneUse(m_Neg(m_And(m_Value(), m_One())))),
2236                     m_Value(Y)))) {
2237     Value *Cmp = Builder.CreateIsNull(Neg);
2238     return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Y);
2239   }
2240 
2241   // Canonicalize:
2242   // (X +/- Y) & Y --> ~X & Y when Y is a power of 2.
2243   if (match(&I, m_c_And(m_Value(Y), m_OneUse(m_CombineOr(
2244                                         m_c_Add(m_Value(X), m_Deferred(Y)),
2245                                         m_Sub(m_Value(X), m_Deferred(Y)))))) &&
2246       isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, /*Depth*/ 0, &I))
2247     return BinaryOperator::CreateAnd(Builder.CreateNot(X), Y);
2248 
2249   const APInt *C;
2250   if (match(Op1, m_APInt(C))) {
2251     const APInt *XorC;
2252     if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_APInt(XorC))))) {
2253       // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2254       Constant *NewC = ConstantInt::get(Ty, *C & *XorC);
2255       Value *And = Builder.CreateAnd(X, Op1);
2256       And->takeName(Op0);
2257       return BinaryOperator::CreateXor(And, NewC);
2258     }
2259 
2260     const APInt *OrC;
2261     if (match(Op0, m_OneUse(m_Or(m_Value(X), m_APInt(OrC))))) {
2262       // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2)
2263       // NOTE: This reduces the number of bits set in the & mask, which
2264       // can expose opportunities for store narrowing for scalars.
2265       // NOTE: SimplifyDemandedBits should have already removed bits from C1
2266       // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in
2267       // above, but this feels safer.
2268       APInt Together = *C & *OrC;
2269       Value *And = Builder.CreateAnd(X, ConstantInt::get(Ty, Together ^ *C));
2270       And->takeName(Op0);
2271       return BinaryOperator::CreateOr(And, ConstantInt::get(Ty, Together));
2272     }
2273 
2274     unsigned Width = Ty->getScalarSizeInBits();
2275     const APInt *ShiftC;
2276     if (match(Op0, m_OneUse(m_SExt(m_AShr(m_Value(X), m_APInt(ShiftC))))) &&
2277         ShiftC->ult(Width)) {
2278       if (*C == APInt::getLowBitsSet(Width, Width - ShiftC->getZExtValue())) {
2279         // We are clearing high bits that were potentially set by sext+ashr:
2280         // and (sext (ashr X, ShiftC)), C --> lshr (sext X), ShiftC
2281         Value *Sext = Builder.CreateSExt(X, Ty);
2282         Constant *ShAmtC = ConstantInt::get(Ty, ShiftC->zext(Width));
2283         return BinaryOperator::CreateLShr(Sext, ShAmtC);
2284       }
2285     }
2286 
2287     // If this 'and' clears the sign-bits added by ashr, replace with lshr:
2288     // and (ashr X, ShiftC), C --> lshr X, ShiftC
2289     if (match(Op0, m_AShr(m_Value(X), m_APInt(ShiftC))) && ShiftC->ult(Width) &&
2290         C->isMask(Width - ShiftC->getZExtValue()))
2291       return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, *ShiftC));
2292 
2293     const APInt *AddC;
2294     if (match(Op0, m_Add(m_Value(X), m_APInt(AddC)))) {
2295       // If we are masking the result of the add down to exactly one bit and
2296       // the constant we are adding has no bits set below that bit, then the
2297       // add is flipping a single bit. Example:
2298       // (X + 4) & 4 --> (X & 4) ^ 4
2299       if (Op0->hasOneUse() && C->isPowerOf2() && (*AddC & (*C - 1)) == 0) {
2300         assert((*C & *AddC) != 0 && "Expected common bit");
2301         Value *NewAnd = Builder.CreateAnd(X, Op1);
2302         return BinaryOperator::CreateXor(NewAnd, Op1);
2303       }
2304     }
2305 
2306     // ((C1 OP zext(X)) & C2) -> zext((C1 OP X) & C2) if C2 fits in the
2307     // bitwidth of X and OP behaves well when given trunc(C1) and X.
2308     auto isNarrowableBinOpcode = [](BinaryOperator *B) {
2309       switch (B->getOpcode()) {
2310       case Instruction::Xor:
2311       case Instruction::Or:
2312       case Instruction::Mul:
2313       case Instruction::Add:
2314       case Instruction::Sub:
2315         return true;
2316       default:
2317         return false;
2318       }
2319     };
2320     BinaryOperator *BO;
2321     if (match(Op0, m_OneUse(m_BinOp(BO))) && isNarrowableBinOpcode(BO)) {
2322       Instruction::BinaryOps BOpcode = BO->getOpcode();
2323       Value *X;
2324       const APInt *C1;
2325       // TODO: The one-use restrictions could be relaxed a little if the AND
2326       // is going to be removed.
2327       // Try to narrow the 'and' and a binop with constant operand:
2328       // and (bo (zext X), C1), C --> zext (and (bo X, TruncC1), TruncC)
2329       if (match(BO, m_c_BinOp(m_OneUse(m_ZExt(m_Value(X))), m_APInt(C1))) &&
2330           C->isIntN(X->getType()->getScalarSizeInBits())) {
2331         unsigned XWidth = X->getType()->getScalarSizeInBits();
2332         Constant *TruncC1 = ConstantInt::get(X->getType(), C1->trunc(XWidth));
2333         Value *BinOp = isa<ZExtInst>(BO->getOperand(0))
2334                            ? Builder.CreateBinOp(BOpcode, X, TruncC1)
2335                            : Builder.CreateBinOp(BOpcode, TruncC1, X);
2336         Constant *TruncC = ConstantInt::get(X->getType(), C->trunc(XWidth));
2337         Value *And = Builder.CreateAnd(BinOp, TruncC);
2338         return new ZExtInst(And, Ty);
2339       }
2340 
2341       // Similar to above: if the mask matches the zext input width, then the
2342       // 'and' can be eliminated, so we can truncate the other variable op:
2343       // and (bo (zext X), Y), C --> zext (bo X, (trunc Y))
2344       if (isa<Instruction>(BO->getOperand(0)) &&
2345           match(BO->getOperand(0), m_OneUse(m_ZExt(m_Value(X)))) &&
2346           C->isMask(X->getType()->getScalarSizeInBits())) {
2347         Y = BO->getOperand(1);
2348         Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr");
2349         Value *NewBO =
2350             Builder.CreateBinOp(BOpcode, X, TrY, BO->getName() + ".narrow");
2351         return new ZExtInst(NewBO, Ty);
2352       }
2353       // and (bo Y, (zext X)), C --> zext (bo (trunc Y), X)
2354       if (isa<Instruction>(BO->getOperand(1)) &&
2355           match(BO->getOperand(1), m_OneUse(m_ZExt(m_Value(X)))) &&
2356           C->isMask(X->getType()->getScalarSizeInBits())) {
2357         Y = BO->getOperand(0);
2358         Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr");
2359         Value *NewBO =
2360             Builder.CreateBinOp(BOpcode, TrY, X, BO->getName() + ".narrow");
2361         return new ZExtInst(NewBO, Ty);
2362       }
2363     }
2364 
2365     // This is intentionally placed after the narrowing transforms for
2366     // efficiency (transform directly to the narrow logic op if possible).
2367     // If the mask is only needed on one incoming arm, push the 'and' op up.
2368     if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_Value(Y)))) ||
2369         match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2370       APInt NotAndMask(~(*C));
2371       BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Op0)->getOpcode();
2372       if (MaskedValueIsZero(X, NotAndMask, 0, &I)) {
2373         // Not masking anything out for the LHS, move mask to RHS.
2374         // and ({x}or X, Y), C --> {x}or X, (and Y, C)
2375         Value *NewRHS = Builder.CreateAnd(Y, Op1, Y->getName() + ".masked");
2376         return BinaryOperator::Create(BinOp, X, NewRHS);
2377       }
2378       if (!isa<Constant>(Y) && MaskedValueIsZero(Y, NotAndMask, 0, &I)) {
2379         // Not masking anything out for the RHS, move mask to LHS.
2380         // and ({x}or X, Y), C --> {x}or (and X, C), Y
2381         Value *NewLHS = Builder.CreateAnd(X, Op1, X->getName() + ".masked");
2382         return BinaryOperator::Create(BinOp, NewLHS, Y);
2383       }
2384     }
2385 
2386     // When the mask is a power-of-2 constant and op0 is a shifted-power-of-2
2387     // constant, test if the shift amount equals the offset bit index:
2388     // (ShiftC << X) & C --> X == (log2(C) - log2(ShiftC)) ? C : 0
2389     // (ShiftC >> X) & C --> X == (log2(ShiftC) - log2(C)) ? C : 0
2390     if (C->isPowerOf2() &&
2391         match(Op0, m_OneUse(m_LogicalShift(m_Power2(ShiftC), m_Value(X))))) {
2392       int Log2ShiftC = ShiftC->exactLogBase2();
2393       int Log2C = C->exactLogBase2();
2394       bool IsShiftLeft =
2395          cast<BinaryOperator>(Op0)->getOpcode() == Instruction::Shl;
2396       int BitNum = IsShiftLeft ? Log2C - Log2ShiftC : Log2ShiftC - Log2C;
2397       assert(BitNum >= 0 && "Expected demanded bits to handle impossible mask");
2398       Value *Cmp = Builder.CreateICmpEQ(X, ConstantInt::get(Ty, BitNum));
2399       return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C),
2400                                 ConstantInt::getNullValue(Ty));
2401     }
2402 
2403     Constant *C1, *C2;
2404     const APInt *C3 = C;
2405     Value *X;
2406     if (C3->isPowerOf2()) {
2407       Constant *Log2C3 = ConstantInt::get(Ty, C3->countr_zero());
2408       if (match(Op0, m_OneUse(m_LShr(m_Shl(m_ImmConstant(C1), m_Value(X)),
2409                                      m_ImmConstant(C2)))) &&
2410           match(C1, m_Power2())) {
2411         Constant *Log2C1 = ConstantExpr::getExactLogBase2(C1);
2412         Constant *LshrC = ConstantExpr::getAdd(C2, Log2C3);
2413         KnownBits KnownLShrc = computeKnownBits(LshrC, 0, nullptr);
2414         if (KnownLShrc.getMaxValue().ult(Width)) {
2415           // iff C1,C3 is pow2 and C2 + cttz(C3) < BitWidth:
2416           // ((C1 << X) >> C2) & C3 -> X == (cttz(C3)+C2-cttz(C1)) ? C3 : 0
2417           Constant *CmpC = ConstantExpr::getSub(LshrC, Log2C1);
2418           Value *Cmp = Builder.CreateICmpEQ(X, CmpC);
2419           return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C3),
2420                                     ConstantInt::getNullValue(Ty));
2421         }
2422       }
2423 
2424       if (match(Op0, m_OneUse(m_Shl(m_LShr(m_ImmConstant(C1), m_Value(X)),
2425                                     m_ImmConstant(C2)))) &&
2426           match(C1, m_Power2())) {
2427         Constant *Log2C1 = ConstantExpr::getExactLogBase2(C1);
2428         Constant *Cmp =
2429             ConstantExpr::getCompare(ICmpInst::ICMP_ULT, Log2C3, C2);
2430         if (Cmp->isZeroValue()) {
2431           // iff C1,C3 is pow2 and Log2(C3) >= C2:
2432           // ((C1 >> X) << C2) & C3 -> X == (cttz(C1)+C2-cttz(C3)) ? C3 : 0
2433           Constant *ShlC = ConstantExpr::getAdd(C2, Log2C1);
2434           Constant *CmpC = ConstantExpr::getSub(ShlC, Log2C3);
2435           Value *Cmp = Builder.CreateICmpEQ(X, CmpC);
2436           return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C3),
2437                                     ConstantInt::getNullValue(Ty));
2438         }
2439       }
2440     }
2441   }
2442 
2443   // If we are clearing the sign bit of a floating-point value, convert this to
2444   // fabs, then cast back to integer.
2445   //
2446   // This is a generous interpretation for noimplicitfloat, this is not a true
2447   // floating-point operation.
2448   //
2449   // Assumes any IEEE-represented type has the sign bit in the high bit.
2450   // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt
2451   Value *CastOp;
2452   if (match(Op0, m_BitCast(m_Value(CastOp))) &&
2453       match(Op1, m_MaxSignedValue()) &&
2454       !Builder.GetInsertBlock()->getParent()->hasFnAttribute(
2455         Attribute::NoImplicitFloat)) {
2456     Type *EltTy = CastOp->getType()->getScalarType();
2457     if (EltTy->isFloatingPointTy() && EltTy->isIEEE() &&
2458         EltTy->getPrimitiveSizeInBits() ==
2459         I.getType()->getScalarType()->getPrimitiveSizeInBits()) {
2460       Value *FAbs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, CastOp);
2461       return new BitCastInst(FAbs, I.getType());
2462     }
2463   }
2464 
2465   if (match(&I, m_And(m_OneUse(m_Shl(m_ZExt(m_Value(X)), m_Value(Y))),
2466                       m_SignMask())) &&
2467       match(Y, m_SpecificInt_ICMP(
2468                    ICmpInst::Predicate::ICMP_EQ,
2469                    APInt(Ty->getScalarSizeInBits(),
2470                          Ty->getScalarSizeInBits() -
2471                              X->getType()->getScalarSizeInBits())))) {
2472     auto *SExt = Builder.CreateSExt(X, Ty, X->getName() + ".signext");
2473     auto *SanitizedSignMask = cast<Constant>(Op1);
2474     // We must be careful with the undef elements of the sign bit mask, however:
2475     // the mask elt can be undef iff the shift amount for that lane was undef,
2476     // otherwise we need to sanitize undef masks to zero.
2477     SanitizedSignMask = Constant::replaceUndefsWith(
2478         SanitizedSignMask, ConstantInt::getNullValue(Ty->getScalarType()));
2479     SanitizedSignMask =
2480         Constant::mergeUndefsWith(SanitizedSignMask, cast<Constant>(Y));
2481     return BinaryOperator::CreateAnd(SExt, SanitizedSignMask);
2482   }
2483 
2484   if (Instruction *Z = narrowMaskedBinOp(I))
2485     return Z;
2486 
2487   if (I.getType()->isIntOrIntVectorTy(1)) {
2488     if (auto *SI0 = dyn_cast<SelectInst>(Op0)) {
2489       if (auto *R =
2490               foldAndOrOfSelectUsingImpliedCond(Op1, *SI0, /* IsAnd */ true))
2491         return R;
2492     }
2493     if (auto *SI1 = dyn_cast<SelectInst>(Op1)) {
2494       if (auto *R =
2495               foldAndOrOfSelectUsingImpliedCond(Op0, *SI1, /* IsAnd */ true))
2496         return R;
2497     }
2498   }
2499 
2500   if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
2501     return FoldedLogic;
2502 
2503   if (Instruction *DeMorgan = matchDeMorgansLaws(I, *this))
2504     return DeMorgan;
2505 
2506   {
2507     Value *A, *B, *C;
2508     // A & (A ^ B) --> A & ~B
2509     if (match(Op1, m_OneUse(m_c_Xor(m_Specific(Op0), m_Value(B)))))
2510       return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(B));
2511     // (A ^ B) & A --> A & ~B
2512     if (match(Op0, m_OneUse(m_c_Xor(m_Specific(Op1), m_Value(B)))))
2513       return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(B));
2514 
2515     // A & ~(A ^ B) --> A & B
2516     if (match(Op1, m_Not(m_c_Xor(m_Specific(Op0), m_Value(B)))))
2517       return BinaryOperator::CreateAnd(Op0, B);
2518     // ~(A ^ B) & A --> A & B
2519     if (match(Op0, m_Not(m_c_Xor(m_Specific(Op1), m_Value(B)))))
2520       return BinaryOperator::CreateAnd(Op1, B);
2521 
2522     // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
2523     if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2524         match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A)))) {
2525       Value *NotC = Op1->hasOneUse()
2526                         ? Builder.CreateNot(C)
2527                         : getFreelyInverted(C, C->hasOneUse(), &Builder);
2528       if (NotC != nullptr)
2529         return BinaryOperator::CreateAnd(Op0, NotC);
2530     }
2531 
2532     // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
2533     if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))) &&
2534         match(Op1, m_Xor(m_Specific(B), m_Specific(A)))) {
2535       Value *NotC = Op0->hasOneUse()
2536                         ? Builder.CreateNot(C)
2537                         : getFreelyInverted(C, C->hasOneUse(), &Builder);
2538       if (NotC != nullptr)
2539         return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(C));
2540     }
2541 
2542     // (A | B) & (~A ^ B) -> A & B
2543     // (A | B) & (B ^ ~A) -> A & B
2544     // (B | A) & (~A ^ B) -> A & B
2545     // (B | A) & (B ^ ~A) -> A & B
2546     if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
2547         match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
2548       return BinaryOperator::CreateAnd(A, B);
2549 
2550     // (~A ^ B) & (A | B) -> A & B
2551     // (~A ^ B) & (B | A) -> A & B
2552     // (B ^ ~A) & (A | B) -> A & B
2553     // (B ^ ~A) & (B | A) -> A & B
2554     if (match(Op0, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
2555         match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
2556       return BinaryOperator::CreateAnd(A, B);
2557 
2558     // (~A | B) & (A ^ B) -> ~A & B
2559     // (~A | B) & (B ^ A) -> ~A & B
2560     // (B | ~A) & (A ^ B) -> ~A & B
2561     // (B | ~A) & (B ^ A) -> ~A & B
2562     if (match(Op0, m_c_Or(m_Not(m_Value(A)), m_Value(B))) &&
2563         match(Op1, m_c_Xor(m_Specific(A), m_Specific(B))))
2564       return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
2565 
2566     // (A ^ B) & (~A | B) -> ~A & B
2567     // (B ^ A) & (~A | B) -> ~A & B
2568     // (A ^ B) & (B | ~A) -> ~A & B
2569     // (B ^ A) & (B | ~A) -> ~A & B
2570     if (match(Op1, m_c_Or(m_Not(m_Value(A)), m_Value(B))) &&
2571         match(Op0, m_c_Xor(m_Specific(A), m_Specific(B))))
2572       return BinaryOperator::CreateAnd(Builder.CreateNot(A), B);
2573   }
2574 
2575   {
2576     ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2577     ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2578     if (LHS && RHS)
2579       if (Value *Res = foldAndOrOfICmps(LHS, RHS, I, /* IsAnd */ true))
2580         return replaceInstUsesWith(I, Res);
2581 
2582     // TODO: Make this recursive; it's a little tricky because an arbitrary
2583     // number of 'and' instructions might have to be created.
2584     if (LHS && match(Op1, m_OneUse(m_LogicalAnd(m_Value(X), m_Value(Y))))) {
2585       bool IsLogical = isa<SelectInst>(Op1);
2586       // LHS & (X && Y) --> (LHS && X) && Y
2587       if (auto *Cmp = dyn_cast<ICmpInst>(X))
2588         if (Value *Res =
2589                 foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ true, IsLogical))
2590           return replaceInstUsesWith(I, IsLogical
2591                                             ? Builder.CreateLogicalAnd(Res, Y)
2592                                             : Builder.CreateAnd(Res, Y));
2593       // LHS & (X && Y) --> X && (LHS & Y)
2594       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2595         if (Value *Res = foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ true,
2596                                           /* IsLogical */ false))
2597           return replaceInstUsesWith(I, IsLogical
2598                                             ? Builder.CreateLogicalAnd(X, Res)
2599                                             : Builder.CreateAnd(X, Res));
2600     }
2601     if (RHS && match(Op0, m_OneUse(m_LogicalAnd(m_Value(X), m_Value(Y))))) {
2602       bool IsLogical = isa<SelectInst>(Op0);
2603       // (X && Y) & RHS --> (X && RHS) && Y
2604       if (auto *Cmp = dyn_cast<ICmpInst>(X))
2605         if (Value *Res =
2606                 foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ true, IsLogical))
2607           return replaceInstUsesWith(I, IsLogical
2608                                             ? Builder.CreateLogicalAnd(Res, Y)
2609                                             : Builder.CreateAnd(Res, Y));
2610       // (X && Y) & RHS --> X && (Y & RHS)
2611       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2612         if (Value *Res = foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ true,
2613                                           /* IsLogical */ false))
2614           return replaceInstUsesWith(I, IsLogical
2615                                             ? Builder.CreateLogicalAnd(X, Res)
2616                                             : Builder.CreateAnd(X, Res));
2617     }
2618   }
2619 
2620   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2621     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2622       if (Value *Res = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ true))
2623         return replaceInstUsesWith(I, Res);
2624 
2625   if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))
2626     return FoldedFCmps;
2627 
2628   if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))
2629     return CastedAnd;
2630 
2631   if (Instruction *Sel = foldBinopOfSextBoolToSelect(I))
2632     return Sel;
2633 
2634   // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>.
2635   // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold
2636   //       with binop identity constant. But creating a select with non-constant
2637   //       arm may not be reversible due to poison semantics. Is that a good
2638   //       canonicalization?
2639   Value *A, *B;
2640   if (match(&I, m_c_And(m_OneUse(m_SExt(m_Value(A))), m_Value(B))) &&
2641       A->getType()->isIntOrIntVectorTy(1))
2642     return SelectInst::Create(A, B, Constant::getNullValue(Ty));
2643 
2644   // Similarly, a 'not' of the bool translates to a swap of the select arms:
2645   // ~sext(A) & B / B & ~sext(A) --> A ? 0 : B
2646   if (match(&I, m_c_And(m_Not(m_SExt(m_Value(A))), m_Value(B))) &&
2647       A->getType()->isIntOrIntVectorTy(1))
2648     return SelectInst::Create(A, Constant::getNullValue(Ty), B);
2649 
2650   // and(zext(A), B) -> A ? (B & 1) : 0
2651   if (match(&I, m_c_And(m_OneUse(m_ZExt(m_Value(A))), m_Value(B))) &&
2652       A->getType()->isIntOrIntVectorTy(1))
2653     return SelectInst::Create(A, Builder.CreateAnd(B, ConstantInt::get(Ty, 1)),
2654                               Constant::getNullValue(Ty));
2655 
2656   // (-1 + A) & B --> A ? 0 : B where A is 0/1.
2657   if (match(&I, m_c_And(m_OneUse(m_Add(m_ZExtOrSelf(m_Value(A)), m_AllOnes())),
2658                         m_Value(B)))) {
2659     if (A->getType()->isIntOrIntVectorTy(1))
2660       return SelectInst::Create(A, Constant::getNullValue(Ty), B);
2661     if (computeKnownBits(A, /* Depth */ 0, &I).countMaxActiveBits() <= 1) {
2662       return SelectInst::Create(
2663           Builder.CreateICmpEQ(A, Constant::getNullValue(A->getType())), B,
2664           Constant::getNullValue(Ty));
2665     }
2666   }
2667 
2668   // (iN X s>> (N-1)) & Y --> (X s< 0) ? Y : 0 -- with optional sext
2669   if (match(&I, m_c_And(m_OneUse(m_SExtOrSelf(
2670                             m_AShr(m_Value(X), m_APIntAllowUndef(C)))),
2671                         m_Value(Y))) &&
2672       *C == X->getType()->getScalarSizeInBits() - 1) {
2673     Value *IsNeg = Builder.CreateIsNeg(X, "isneg");
2674     return SelectInst::Create(IsNeg, Y, ConstantInt::getNullValue(Ty));
2675   }
2676   // If there's a 'not' of the shifted value, swap the select operands:
2677   // ~(iN X s>> (N-1)) & Y --> (X s< 0) ? 0 : Y -- with optional sext
2678   if (match(&I, m_c_And(m_OneUse(m_SExtOrSelf(
2679                             m_Not(m_AShr(m_Value(X), m_APIntAllowUndef(C))))),
2680                         m_Value(Y))) &&
2681       *C == X->getType()->getScalarSizeInBits() - 1) {
2682     Value *IsNeg = Builder.CreateIsNeg(X, "isneg");
2683     return SelectInst::Create(IsNeg, ConstantInt::getNullValue(Ty), Y);
2684   }
2685 
2686   // (~x) & y  -->  ~(x | (~y))  iff that gets rid of inversions
2687   if (sinkNotIntoOtherHandOfLogicalOp(I))
2688     return &I;
2689 
2690   // An and recurrence w/loop invariant step is equivelent to (and start, step)
2691   PHINode *PN = nullptr;
2692   Value *Start = nullptr, *Step = nullptr;
2693   if (matchSimpleRecurrence(&I, PN, Start, Step) && DT.dominates(Step, PN))
2694     return replaceInstUsesWith(I, Builder.CreateAnd(Start, Step));
2695 
2696   if (Instruction *R = reassociateForUses(I, Builder))
2697     return R;
2698 
2699   if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))
2700     return Canonicalized;
2701 
2702   if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1))
2703     return Folded;
2704 
2705   if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
2706     return Res;
2707 
2708   if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))
2709     return Res;
2710 
2711   return nullptr;
2712 }
2713 
2714 Instruction *InstCombinerImpl::matchBSwapOrBitReverse(Instruction &I,
2715                                                       bool MatchBSwaps,
2716                                                       bool MatchBitReversals) {
2717   SmallVector<Instruction *, 4> Insts;
2718   if (!recognizeBSwapOrBitReverseIdiom(&I, MatchBSwaps, MatchBitReversals,
2719                                        Insts))
2720     return nullptr;
2721   Instruction *LastInst = Insts.pop_back_val();
2722   LastInst->removeFromParent();
2723 
2724   for (auto *Inst : Insts)
2725     Worklist.push(Inst);
2726   return LastInst;
2727 }
2728 
2729 /// Match UB-safe variants of the funnel shift intrinsic.
2730 static Instruction *matchFunnelShift(Instruction &Or, InstCombinerImpl &IC,
2731                                      const DominatorTree &DT) {
2732   // TODO: Can we reduce the code duplication between this and the related
2733   // rotate matching code under visitSelect and visitTrunc?
2734   unsigned Width = Or.getType()->getScalarSizeInBits();
2735 
2736   Instruction *Or0, *Or1;
2737   if (!match(Or.getOperand(0), m_Instruction(Or0)) ||
2738       !match(Or.getOperand(1), m_Instruction(Or1)))
2739     return nullptr;
2740 
2741   bool IsFshl = true; // Sub on LSHR.
2742   SmallVector<Value *, 3> FShiftArgs;
2743 
2744   // First, find an or'd pair of opposite shifts:
2745   // or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1)
2746   if (isa<BinaryOperator>(Or0) && isa<BinaryOperator>(Or1)) {
2747     Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1;
2748     if (!match(Or0,
2749                m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) ||
2750         !match(Or1,
2751                m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) ||
2752         Or0->getOpcode() == Or1->getOpcode())
2753       return nullptr;
2754 
2755     // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)).
2756     if (Or0->getOpcode() == BinaryOperator::LShr) {
2757       std::swap(Or0, Or1);
2758       std::swap(ShVal0, ShVal1);
2759       std::swap(ShAmt0, ShAmt1);
2760     }
2761     assert(Or0->getOpcode() == BinaryOperator::Shl &&
2762            Or1->getOpcode() == BinaryOperator::LShr &&
2763            "Illegal or(shift,shift) pair");
2764 
2765     // Match the shift amount operands for a funnel shift pattern. This always
2766     // matches a subtraction on the R operand.
2767     auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * {
2768       // Check for constant shift amounts that sum to the bitwidth.
2769       const APInt *LI, *RI;
2770       if (match(L, m_APIntAllowUndef(LI)) && match(R, m_APIntAllowUndef(RI)))
2771         if (LI->ult(Width) && RI->ult(Width) && (*LI + *RI) == Width)
2772           return ConstantInt::get(L->getType(), *LI);
2773 
2774       Constant *LC, *RC;
2775       if (match(L, m_Constant(LC)) && match(R, m_Constant(RC)) &&
2776           match(L,
2777                 m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) &&
2778           match(R,
2779                 m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) &&
2780           match(ConstantExpr::getAdd(LC, RC), m_SpecificIntAllowUndef(Width)))
2781         return ConstantExpr::mergeUndefsWith(LC, RC);
2782 
2783       // (shl ShVal, X) | (lshr ShVal, (Width - x)) iff X < Width.
2784       // We limit this to X < Width in case the backend re-expands the
2785       // intrinsic, and has to reintroduce a shift modulo operation (InstCombine
2786       // might remove it after this fold). This still doesn't guarantee that the
2787       // final codegen will match this original pattern.
2788       if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L))))) {
2789         KnownBits KnownL = IC.computeKnownBits(L, /*Depth*/ 0, &Or);
2790         return KnownL.getMaxValue().ult(Width) ? L : nullptr;
2791       }
2792 
2793       // For non-constant cases, the following patterns currently only work for
2794       // rotation patterns.
2795       // TODO: Add general funnel-shift compatible patterns.
2796       if (ShVal0 != ShVal1)
2797         return nullptr;
2798 
2799       // For non-constant cases we don't support non-pow2 shift masks.
2800       // TODO: Is it worth matching urem as well?
2801       if (!isPowerOf2_32(Width))
2802         return nullptr;
2803 
2804       // The shift amount may be masked with negation:
2805       // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1)))
2806       Value *X;
2807       unsigned Mask = Width - 1;
2808       if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&
2809           match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))
2810         return X;
2811 
2812       // (shl ShVal, X) | (lshr ShVal, ((-X) & (Width - 1)))
2813       if (match(R, m_And(m_Neg(m_Specific(L)), m_SpecificInt(Mask))))
2814         return L;
2815 
2816       // Similar to above, but the shift amount may be extended after masking,
2817       // so return the extended value as the parameter for the intrinsic.
2818       if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&
2819           match(R,
2820                 m_And(m_Neg(m_ZExt(m_And(m_Specific(X), m_SpecificInt(Mask)))),
2821                       m_SpecificInt(Mask))))
2822         return L;
2823 
2824       if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&
2825           match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))))
2826         return L;
2827 
2828       return nullptr;
2829     };
2830 
2831     Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width);
2832     if (!ShAmt) {
2833       ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width);
2834       IsFshl = false; // Sub on SHL.
2835     }
2836     if (!ShAmt)
2837       return nullptr;
2838 
2839     FShiftArgs = {ShVal0, ShVal1, ShAmt};
2840   } else if (isa<ZExtInst>(Or0) || isa<ZExtInst>(Or1)) {
2841     // If there are two 'or' instructions concat variables in opposite order:
2842     //
2843     // Slot1 and Slot2 are all zero bits.
2844     // | Slot1 | Low | Slot2 | High |
2845     // LowHigh = or (shl (zext Low), ZextLowShlAmt), (zext High)
2846     // | Slot2 | High | Slot1 | Low |
2847     // HighLow = or (shl (zext High), ZextHighShlAmt), (zext Low)
2848     //
2849     // the latter 'or' can be safely convert to
2850     // -> HighLow = fshl LowHigh, LowHigh, ZextHighShlAmt
2851     // if ZextLowShlAmt + ZextHighShlAmt == Width.
2852     if (!isa<ZExtInst>(Or1))
2853       std::swap(Or0, Or1);
2854 
2855     Value *High, *ZextHigh, *Low;
2856     const APInt *ZextHighShlAmt;
2857     if (!match(Or0,
2858                m_OneUse(m_Shl(m_Value(ZextHigh), m_APInt(ZextHighShlAmt)))))
2859       return nullptr;
2860 
2861     if (!match(Or1, m_ZExt(m_Value(Low))) ||
2862         !match(ZextHigh, m_ZExt(m_Value(High))))
2863       return nullptr;
2864 
2865     unsigned HighSize = High->getType()->getScalarSizeInBits();
2866     unsigned LowSize = Low->getType()->getScalarSizeInBits();
2867     // Make sure High does not overlap with Low and most significant bits of
2868     // High aren't shifted out.
2869     if (ZextHighShlAmt->ult(LowSize) || ZextHighShlAmt->ugt(Width - HighSize))
2870       return nullptr;
2871 
2872     for (User *U : ZextHigh->users()) {
2873       Value *X, *Y;
2874       if (!match(U, m_Or(m_Value(X), m_Value(Y))))
2875         continue;
2876 
2877       if (!isa<ZExtInst>(Y))
2878         std::swap(X, Y);
2879 
2880       const APInt *ZextLowShlAmt;
2881       if (!match(X, m_Shl(m_Specific(Or1), m_APInt(ZextLowShlAmt))) ||
2882           !match(Y, m_Specific(ZextHigh)) || !DT.dominates(U, &Or))
2883         continue;
2884 
2885       // HighLow is good concat. If sum of two shifts amount equals to Width,
2886       // LowHigh must also be a good concat.
2887       if (*ZextLowShlAmt + *ZextHighShlAmt != Width)
2888         continue;
2889 
2890       // Low must not overlap with High and most significant bits of Low must
2891       // not be shifted out.
2892       assert(ZextLowShlAmt->uge(HighSize) &&
2893              ZextLowShlAmt->ule(Width - LowSize) && "Invalid concat");
2894 
2895       FShiftArgs = {U, U, ConstantInt::get(Or0->getType(), *ZextHighShlAmt)};
2896       break;
2897     }
2898   }
2899 
2900   if (FShiftArgs.empty())
2901     return nullptr;
2902 
2903   Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
2904   Function *F = Intrinsic::getDeclaration(Or.getModule(), IID, Or.getType());
2905   return CallInst::Create(F, FShiftArgs);
2906 }
2907 
2908 /// Attempt to combine or(zext(x),shl(zext(y),bw/2) concat packing patterns.
2909 static Instruction *matchOrConcat(Instruction &Or,
2910                                   InstCombiner::BuilderTy &Builder) {
2911   assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'");
2912   Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1);
2913   Type *Ty = Or.getType();
2914 
2915   unsigned Width = Ty->getScalarSizeInBits();
2916   if ((Width & 1) != 0)
2917     return nullptr;
2918   unsigned HalfWidth = Width / 2;
2919 
2920   // Canonicalize zext (lower half) to LHS.
2921   if (!isa<ZExtInst>(Op0))
2922     std::swap(Op0, Op1);
2923 
2924   // Find lower/upper half.
2925   Value *LowerSrc, *ShlVal, *UpperSrc;
2926   const APInt *C;
2927   if (!match(Op0, m_OneUse(m_ZExt(m_Value(LowerSrc)))) ||
2928       !match(Op1, m_OneUse(m_Shl(m_Value(ShlVal), m_APInt(C)))) ||
2929       !match(ShlVal, m_OneUse(m_ZExt(m_Value(UpperSrc)))))
2930     return nullptr;
2931   if (*C != HalfWidth || LowerSrc->getType() != UpperSrc->getType() ||
2932       LowerSrc->getType()->getScalarSizeInBits() != HalfWidth)
2933     return nullptr;
2934 
2935   auto ConcatIntrinsicCalls = [&](Intrinsic::ID id, Value *Lo, Value *Hi) {
2936     Value *NewLower = Builder.CreateZExt(Lo, Ty);
2937     Value *NewUpper = Builder.CreateZExt(Hi, Ty);
2938     NewUpper = Builder.CreateShl(NewUpper, HalfWidth);
2939     Value *BinOp = Builder.CreateOr(NewLower, NewUpper);
2940     Function *F = Intrinsic::getDeclaration(Or.getModule(), id, Ty);
2941     return Builder.CreateCall(F, BinOp);
2942   };
2943 
2944   // BSWAP: Push the concat down, swapping the lower/upper sources.
2945   // concat(bswap(x),bswap(y)) -> bswap(concat(x,y))
2946   Value *LowerBSwap, *UpperBSwap;
2947   if (match(LowerSrc, m_BSwap(m_Value(LowerBSwap))) &&
2948       match(UpperSrc, m_BSwap(m_Value(UpperBSwap))))
2949     return ConcatIntrinsicCalls(Intrinsic::bswap, UpperBSwap, LowerBSwap);
2950 
2951   // BITREVERSE: Push the concat down, swapping the lower/upper sources.
2952   // concat(bitreverse(x),bitreverse(y)) -> bitreverse(concat(x,y))
2953   Value *LowerBRev, *UpperBRev;
2954   if (match(LowerSrc, m_BitReverse(m_Value(LowerBRev))) &&
2955       match(UpperSrc, m_BitReverse(m_Value(UpperBRev))))
2956     return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev);
2957 
2958   return nullptr;
2959 }
2960 
2961 /// If all elements of two constant vectors are 0/-1 and inverses, return true.
2962 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {
2963   unsigned NumElts = cast<FixedVectorType>(C1->getType())->getNumElements();
2964   for (unsigned i = 0; i != NumElts; ++i) {
2965     Constant *EltC1 = C1->getAggregateElement(i);
2966     Constant *EltC2 = C2->getAggregateElement(i);
2967     if (!EltC1 || !EltC2)
2968       return false;
2969 
2970     // One element must be all ones, and the other must be all zeros.
2971     if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) ||
2972           (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes()))))
2973       return false;
2974   }
2975   return true;
2976 }
2977 
2978 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or
2979 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of
2980 /// B, it can be used as the condition operand of a select instruction.
2981 /// We will detect (A & C) | ~(B | D) when the flag ABIsTheSame enabled.
2982 Value *InstCombinerImpl::getSelectCondition(Value *A, Value *B,
2983                                             bool ABIsTheSame) {
2984   // We may have peeked through bitcasts in the caller.
2985   // Exit immediately if we don't have (vector) integer types.
2986   Type *Ty = A->getType();
2987   if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy())
2988     return nullptr;
2989 
2990   // If A is the 'not' operand of B and has enough signbits, we have our answer.
2991   if (ABIsTheSame ? (A == B) : match(B, m_Not(m_Specific(A)))) {
2992     // If these are scalars or vectors of i1, A can be used directly.
2993     if (Ty->isIntOrIntVectorTy(1))
2994       return A;
2995 
2996     // If we look through a vector bitcast, the caller will bitcast the operands
2997     // to match the condition's number of bits (N x i1).
2998     // To make this poison-safe, disallow bitcast from wide element to narrow
2999     // element. That could allow poison in lanes where it was not present in the
3000     // original code.
3001     A = peekThroughBitcast(A);
3002     if (A->getType()->isIntOrIntVectorTy()) {
3003       unsigned NumSignBits = ComputeNumSignBits(A);
3004       if (NumSignBits == A->getType()->getScalarSizeInBits() &&
3005           NumSignBits <= Ty->getScalarSizeInBits())
3006         return Builder.CreateTrunc(A, CmpInst::makeCmpResultType(A->getType()));
3007     }
3008     return nullptr;
3009   }
3010 
3011   // TODO: add support for sext and constant case
3012   if (ABIsTheSame)
3013     return nullptr;
3014 
3015   // If both operands are constants, see if the constants are inverse bitmasks.
3016   Constant *AConst, *BConst;
3017   if (match(A, m_Constant(AConst)) && match(B, m_Constant(BConst)))
3018     if (AConst == ConstantExpr::getNot(BConst) &&
3019         ComputeNumSignBits(A) == Ty->getScalarSizeInBits())
3020       return Builder.CreateZExtOrTrunc(A, CmpInst::makeCmpResultType(Ty));
3021 
3022   // Look for more complex patterns. The 'not' op may be hidden behind various
3023   // casts. Look through sexts and bitcasts to find the booleans.
3024   Value *Cond;
3025   Value *NotB;
3026   if (match(A, m_SExt(m_Value(Cond))) &&
3027       Cond->getType()->isIntOrIntVectorTy(1)) {
3028     // A = sext i1 Cond; B = sext (not (i1 Cond))
3029     if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
3030       return Cond;
3031 
3032     // A = sext i1 Cond; B = not ({bitcast} (sext (i1 Cond)))
3033     // TODO: The one-use checks are unnecessary or misplaced. If the caller
3034     //       checked for uses on logic ops/casts, that should be enough to
3035     //       make this transform worthwhile.
3036     if (match(B, m_OneUse(m_Not(m_Value(NotB))))) {
3037       NotB = peekThroughBitcast(NotB, true);
3038       if (match(NotB, m_SExt(m_Specific(Cond))))
3039         return Cond;
3040     }
3041   }
3042 
3043   // All scalar (and most vector) possibilities should be handled now.
3044   // Try more matches that only apply to non-splat constant vectors.
3045   if (!Ty->isVectorTy())
3046     return nullptr;
3047 
3048   // If both operands are xor'd with constants using the same sexted boolean
3049   // operand, see if the constants are inverse bitmasks.
3050   // TODO: Use ConstantExpr::getNot()?
3051   if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AConst)))) &&
3052       match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BConst)))) &&
3053       Cond->getType()->isIntOrIntVectorTy(1) &&
3054       areInverseVectorBitmasks(AConst, BConst)) {
3055     AConst = ConstantExpr::getTrunc(AConst, CmpInst::makeCmpResultType(Ty));
3056     return Builder.CreateXor(Cond, AConst);
3057   }
3058   return nullptr;
3059 }
3060 
3061 /// We have an expression of the form (A & C) | (B & D). Try to simplify this
3062 /// to "A' ? C : D", where A' is a boolean or vector of booleans.
3063 /// When InvertFalseVal is set to true, we try to match the pattern
3064 /// where we have peeked through a 'not' op and A and B are the same:
3065 /// (A & C) | ~(A | D) --> (A & C) | (~A & ~D) --> A' ? C : ~D
3066 Value *InstCombinerImpl::matchSelectFromAndOr(Value *A, Value *C, Value *B,
3067                                               Value *D, bool InvertFalseVal) {
3068   // The potential condition of the select may be bitcasted. In that case, look
3069   // through its bitcast and the corresponding bitcast of the 'not' condition.
3070   Type *OrigType = A->getType();
3071   A = peekThroughBitcast(A, true);
3072   B = peekThroughBitcast(B, true);
3073   if (Value *Cond = getSelectCondition(A, B, InvertFalseVal)) {
3074     // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D))
3075     // If this is a vector, we may need to cast to match the condition's length.
3076     // The bitcasts will either all exist or all not exist. The builder will
3077     // not create unnecessary casts if the types already match.
3078     Type *SelTy = A->getType();
3079     if (auto *VecTy = dyn_cast<VectorType>(Cond->getType())) {
3080       // For a fixed or scalable vector get N from <{vscale x} N x iM>
3081       unsigned Elts = VecTy->getElementCount().getKnownMinValue();
3082       // For a fixed or scalable vector, get the size in bits of N x iM; for a
3083       // scalar this is just M.
3084       unsigned SelEltSize = SelTy->getPrimitiveSizeInBits().getKnownMinValue();
3085       Type *EltTy = Builder.getIntNTy(SelEltSize / Elts);
3086       SelTy = VectorType::get(EltTy, VecTy->getElementCount());
3087     }
3088     Value *BitcastC = Builder.CreateBitCast(C, SelTy);
3089     if (InvertFalseVal)
3090       D = Builder.CreateNot(D);
3091     Value *BitcastD = Builder.CreateBitCast(D, SelTy);
3092     Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD);
3093     return Builder.CreateBitCast(Select, OrigType);
3094   }
3095 
3096   return nullptr;
3097 }
3098 
3099 // (icmp eq X, C) | (icmp ult Other, (X - C)) -> (icmp ule Other, (X - (C + 1)))
3100 // (icmp ne X, C) & (icmp uge Other, (X - C)) -> (icmp ugt Other, (X - (C + 1)))
3101 static Value *foldAndOrOfICmpEqConstantAndICmp(ICmpInst *LHS, ICmpInst *RHS,
3102                                                bool IsAnd, bool IsLogical,
3103                                                IRBuilderBase &Builder) {
3104   Value *LHS0 = LHS->getOperand(0);
3105   Value *RHS0 = RHS->getOperand(0);
3106   Value *RHS1 = RHS->getOperand(1);
3107 
3108   ICmpInst::Predicate LPred =
3109       IsAnd ? LHS->getInversePredicate() : LHS->getPredicate();
3110   ICmpInst::Predicate RPred =
3111       IsAnd ? RHS->getInversePredicate() : RHS->getPredicate();
3112 
3113   const APInt *CInt;
3114   if (LPred != ICmpInst::ICMP_EQ ||
3115       !match(LHS->getOperand(1), m_APIntAllowUndef(CInt)) ||
3116       !LHS0->getType()->isIntOrIntVectorTy() ||
3117       !(LHS->hasOneUse() || RHS->hasOneUse()))
3118     return nullptr;
3119 
3120   auto MatchRHSOp = [LHS0, CInt](const Value *RHSOp) {
3121     return match(RHSOp,
3122                  m_Add(m_Specific(LHS0), m_SpecificIntAllowUndef(-*CInt))) ||
3123            (CInt->isZero() && RHSOp == LHS0);
3124   };
3125 
3126   Value *Other;
3127   if (RPred == ICmpInst::ICMP_ULT && MatchRHSOp(RHS1))
3128     Other = RHS0;
3129   else if (RPred == ICmpInst::ICMP_UGT && MatchRHSOp(RHS0))
3130     Other = RHS1;
3131   else
3132     return nullptr;
3133 
3134   if (IsLogical)
3135     Other = Builder.CreateFreeze(Other);
3136 
3137   return Builder.CreateICmp(
3138       IsAnd ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE,
3139       Builder.CreateSub(LHS0, ConstantInt::get(LHS0->getType(), *CInt + 1)),
3140       Other);
3141 }
3142 
3143 /// Fold (icmp)&(icmp) or (icmp)|(icmp) if possible.
3144 /// If IsLogical is true, then the and/or is in select form and the transform
3145 /// must be poison-safe.
3146 Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
3147                                           Instruction &I, bool IsAnd,
3148                                           bool IsLogical) {
3149   const SimplifyQuery Q = SQ.getWithInstruction(&I);
3150 
3151   // Fold (iszero(A & K1) | iszero(A & K2)) ->  (A & (K1 | K2)) != (K1 | K2)
3152   // Fold (!iszero(A & K1) & !iszero(A & K2)) ->  (A & (K1 | K2)) == (K1 | K2)
3153   // if K1 and K2 are a one-bit mask.
3154   if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, &I, IsAnd, IsLogical))
3155     return V;
3156 
3157   ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
3158   Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);
3159   Value *LHS1 = LHS->getOperand(1), *RHS1 = RHS->getOperand(1);
3160   const APInt *LHSC = nullptr, *RHSC = nullptr;
3161   match(LHS1, m_APInt(LHSC));
3162   match(RHS1, m_APInt(RHSC));
3163 
3164   // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3165   // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3166   if (predicatesFoldable(PredL, PredR)) {
3167     if (LHS0 == RHS1 && LHS1 == RHS0) {
3168       PredL = ICmpInst::getSwappedPredicate(PredL);
3169       std::swap(LHS0, LHS1);
3170     }
3171     if (LHS0 == RHS0 && LHS1 == RHS1) {
3172       unsigned Code = IsAnd ? getICmpCode(PredL) & getICmpCode(PredR)
3173                             : getICmpCode(PredL) | getICmpCode(PredR);
3174       bool IsSigned = LHS->isSigned() || RHS->isSigned();
3175       return getNewICmpValue(Code, IsSigned, LHS0, LHS1, Builder);
3176     }
3177   }
3178 
3179   // handle (roughly):
3180   // (icmp ne (A & B), C) | (icmp ne (A & D), E)
3181   // (icmp eq (A & B), C) & (icmp eq (A & D), E)
3182   if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, IsAnd, IsLogical, Builder))
3183     return V;
3184 
3185   if (Value *V =
3186           foldAndOrOfICmpEqConstantAndICmp(LHS, RHS, IsAnd, IsLogical, Builder))
3187     return V;
3188   // We can treat logical like bitwise here, because both operands are used on
3189   // the LHS, and as such poison from both will propagate.
3190   if (Value *V = foldAndOrOfICmpEqConstantAndICmp(RHS, LHS, IsAnd,
3191                                                   /*IsLogical*/ false, Builder))
3192     return V;
3193 
3194   if (Value *V =
3195           foldAndOrOfICmpsWithConstEq(LHS, RHS, IsAnd, IsLogical, Builder, Q))
3196     return V;
3197   // We can convert this case to bitwise and, because both operands are used
3198   // on the LHS, and as such poison from both will propagate.
3199   if (Value *V = foldAndOrOfICmpsWithConstEq(RHS, LHS, IsAnd,
3200                                              /*IsLogical*/ false, Builder, Q))
3201     return V;
3202 
3203   if (Value *V = foldIsPowerOf2OrZero(LHS, RHS, IsAnd, Builder))
3204     return V;
3205   if (Value *V = foldIsPowerOf2OrZero(RHS, LHS, IsAnd, Builder))
3206     return V;
3207 
3208   // TODO: One of these directions is fine with logical and/or, the other could
3209   // be supported by inserting freeze.
3210   if (!IsLogical) {
3211     // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
3212     // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
3213     if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/!IsAnd))
3214       return V;
3215 
3216     // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
3217     // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
3218     if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/!IsAnd))
3219       return V;
3220   }
3221 
3222   // TODO: Add conjugated or fold, check whether it is safe for logical and/or.
3223   if (IsAnd && !IsLogical)
3224     if (Value *V = foldSignedTruncationCheck(LHS, RHS, I, Builder))
3225       return V;
3226 
3227   if (Value *V = foldIsPowerOf2(LHS, RHS, IsAnd, Builder))
3228     return V;
3229 
3230   if (Value *V = foldPowerOf2AndShiftedMask(LHS, RHS, IsAnd, Builder))
3231     return V;
3232 
3233   // TODO: Verify whether this is safe for logical and/or.
3234   if (!IsLogical) {
3235     if (Value *X = foldUnsignedUnderflowCheck(LHS, RHS, IsAnd, Q, Builder))
3236       return X;
3237     if (Value *X = foldUnsignedUnderflowCheck(RHS, LHS, IsAnd, Q, Builder))
3238       return X;
3239   }
3240 
3241   if (Value *X = foldEqOfParts(LHS, RHS, IsAnd))
3242     return X;
3243 
3244   // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
3245   // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
3246   // TODO: Remove this and below when foldLogOpOfMaskedICmps can handle undefs.
3247   if (!IsLogical && PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&
3248       PredL == PredR && match(LHS1, m_ZeroInt()) && match(RHS1, m_ZeroInt()) &&
3249       LHS0->getType() == RHS0->getType()) {
3250     Value *NewOr = Builder.CreateOr(LHS0, RHS0);
3251     return Builder.CreateICmp(PredL, NewOr,
3252                               Constant::getNullValue(NewOr->getType()));
3253   }
3254 
3255   // (icmp ne A, -1) | (icmp ne B, -1) --> (icmp ne (A&B), -1)
3256   // (icmp eq A, -1) & (icmp eq B, -1) --> (icmp eq (A&B), -1)
3257   if (!IsLogical && PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&
3258       PredL == PredR && match(LHS1, m_AllOnes()) && match(RHS1, m_AllOnes()) &&
3259       LHS0->getType() == RHS0->getType()) {
3260     Value *NewAnd = Builder.CreateAnd(LHS0, RHS0);
3261     return Builder.CreateICmp(PredL, NewAnd,
3262                               Constant::getAllOnesValue(LHS0->getType()));
3263   }
3264 
3265   // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
3266   if (!LHSC || !RHSC)
3267     return nullptr;
3268 
3269   // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
3270   // (trunc x) != C1 | (and x, CA) != C2 -> (and x, CA|CMAX) != C1|C2
3271   // where CMAX is the all ones value for the truncated type,
3272   // iff the lower bits of C2 and CA are zero.
3273   if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) &&
3274       PredL == PredR && LHS->hasOneUse() && RHS->hasOneUse()) {
3275     Value *V;
3276     const APInt *AndC, *SmallC = nullptr, *BigC = nullptr;
3277 
3278     // (trunc x) == C1 & (and x, CA) == C2
3279     // (and x, CA) == C2 & (trunc x) == C1
3280     if (match(RHS0, m_Trunc(m_Value(V))) &&
3281         match(LHS0, m_And(m_Specific(V), m_APInt(AndC)))) {
3282       SmallC = RHSC;
3283       BigC = LHSC;
3284     } else if (match(LHS0, m_Trunc(m_Value(V))) &&
3285                match(RHS0, m_And(m_Specific(V), m_APInt(AndC)))) {
3286       SmallC = LHSC;
3287       BigC = RHSC;
3288     }
3289 
3290     if (SmallC && BigC) {
3291       unsigned BigBitSize = BigC->getBitWidth();
3292       unsigned SmallBitSize = SmallC->getBitWidth();
3293 
3294       // Check that the low bits are zero.
3295       APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
3296       if ((Low & *AndC).isZero() && (Low & *BigC).isZero()) {
3297         Value *NewAnd = Builder.CreateAnd(V, Low | *AndC);
3298         APInt N = SmallC->zext(BigBitSize) | *BigC;
3299         Value *NewVal = ConstantInt::get(NewAnd->getType(), N);
3300         return Builder.CreateICmp(PredL, NewAnd, NewVal);
3301       }
3302     }
3303   }
3304 
3305   // Match naive pattern (and its inverted form) for checking if two values
3306   // share same sign. An example of the pattern:
3307   // (icmp slt (X & Y), 0) | (icmp sgt (X | Y), -1) -> (icmp sgt (X ^ Y), -1)
3308   // Inverted form (example):
3309   // (icmp slt (X | Y), 0) & (icmp sgt (X & Y), -1) -> (icmp slt (X ^ Y), 0)
3310   bool TrueIfSignedL, TrueIfSignedR;
3311   if (isSignBitCheck(PredL, *LHSC, TrueIfSignedL) &&
3312       isSignBitCheck(PredR, *RHSC, TrueIfSignedR) &&
3313       (RHS->hasOneUse() || LHS->hasOneUse())) {
3314     Value *X, *Y;
3315     if (IsAnd) {
3316       if ((TrueIfSignedL && !TrueIfSignedR &&
3317            match(LHS0, m_Or(m_Value(X), m_Value(Y))) &&
3318            match(RHS0, m_c_And(m_Specific(X), m_Specific(Y)))) ||
3319           (!TrueIfSignedL && TrueIfSignedR &&
3320            match(LHS0, m_And(m_Value(X), m_Value(Y))) &&
3321            match(RHS0, m_c_Or(m_Specific(X), m_Specific(Y))))) {
3322         Value *NewXor = Builder.CreateXor(X, Y);
3323         return Builder.CreateIsNeg(NewXor);
3324       }
3325     } else {
3326       if ((TrueIfSignedL && !TrueIfSignedR &&
3327             match(LHS0, m_And(m_Value(X), m_Value(Y))) &&
3328             match(RHS0, m_c_Or(m_Specific(X), m_Specific(Y)))) ||
3329           (!TrueIfSignedL && TrueIfSignedR &&
3330            match(LHS0, m_Or(m_Value(X), m_Value(Y))) &&
3331            match(RHS0, m_c_And(m_Specific(X), m_Specific(Y))))) {
3332         Value *NewXor = Builder.CreateXor(X, Y);
3333         return Builder.CreateIsNotNeg(NewXor);
3334       }
3335     }
3336   }
3337 
3338   return foldAndOrOfICmpsUsingRanges(LHS, RHS, IsAnd);
3339 }
3340 
3341 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
3342 // here. We should standardize that construct where it is needed or choose some
3343 // other way to ensure that commutated variants of patterns are not missed.
3344 Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
3345   if (Value *V = simplifyOrInst(I.getOperand(0), I.getOperand(1),
3346                                 SQ.getWithInstruction(&I)))
3347     return replaceInstUsesWith(I, V);
3348 
3349   if (SimplifyAssociativeOrCommutative(I))
3350     return &I;
3351 
3352   if (Instruction *X = foldVectorBinop(I))
3353     return X;
3354 
3355   if (Instruction *Phi = foldBinopWithPhiOperands(I))
3356     return Phi;
3357 
3358   // See if we can simplify any instructions used by the instruction whose sole
3359   // purpose is to compute bits we don't care about.
3360   if (SimplifyDemandedInstructionBits(I))
3361     return &I;
3362 
3363   // Do this before using distributive laws to catch simple and/or/not patterns.
3364   if (Instruction *Xor = foldOrToXor(I, Builder))
3365     return Xor;
3366 
3367   if (Instruction *X = foldComplexAndOrPatterns(I, Builder))
3368     return X;
3369 
3370   // (A&B)|(A&C) -> A&(B|C) etc
3371   if (Value *V = foldUsingDistributiveLaws(I))
3372     return replaceInstUsesWith(I, V);
3373 
3374   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3375   Type *Ty = I.getType();
3376   if (Ty->isIntOrIntVectorTy(1)) {
3377     if (auto *SI0 = dyn_cast<SelectInst>(Op0)) {
3378       if (auto *R =
3379               foldAndOrOfSelectUsingImpliedCond(Op1, *SI0, /* IsAnd */ false))
3380         return R;
3381     }
3382     if (auto *SI1 = dyn_cast<SelectInst>(Op1)) {
3383       if (auto *R =
3384               foldAndOrOfSelectUsingImpliedCond(Op0, *SI1, /* IsAnd */ false))
3385         return R;
3386     }
3387   }
3388 
3389   if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
3390     return FoldedLogic;
3391 
3392   if (Instruction *BitOp = matchBSwapOrBitReverse(I, /*MatchBSwaps*/ true,
3393                                                   /*MatchBitReversals*/ true))
3394     return BitOp;
3395 
3396   if (Instruction *Funnel = matchFunnelShift(I, *this, DT))
3397     return Funnel;
3398 
3399   if (Instruction *Concat = matchOrConcat(I, Builder))
3400     return replaceInstUsesWith(I, Concat);
3401 
3402   if (Instruction *R = foldBinOpShiftWithShift(I))
3403     return R;
3404 
3405   if (Instruction *R = tryFoldInstWithCtpopWithNot(&I))
3406     return R;
3407 
3408   Value *X, *Y;
3409   const APInt *CV;
3410   if (match(&I, m_c_Or(m_OneUse(m_Xor(m_Value(X), m_APInt(CV))), m_Value(Y))) &&
3411       !CV->isAllOnes() && MaskedValueIsZero(Y, *CV, 0, &I)) {
3412     // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 0
3413     // The check for a 'not' op is for efficiency (if Y is known zero --> ~X).
3414     Value *Or = Builder.CreateOr(X, Y);
3415     return BinaryOperator::CreateXor(Or, ConstantInt::get(Ty, *CV));
3416   }
3417 
3418   // If the operands have no common bits set:
3419   // or (mul X, Y), X --> add (mul X, Y), X --> mul X, (Y + 1)
3420   if (match(&I, m_c_DisjointOr(m_OneUse(m_Mul(m_Value(X), m_Value(Y))),
3421                                m_Deferred(X)))) {
3422     Value *IncrementY = Builder.CreateAdd(Y, ConstantInt::get(Ty, 1));
3423     return BinaryOperator::CreateMul(X, IncrementY);
3424   }
3425 
3426   // X | (X ^ Y) --> X | Y (4 commuted patterns)
3427   if (match(&I, m_c_Or(m_Value(X), m_c_Xor(m_Deferred(X), m_Value(Y)))))
3428     return BinaryOperator::CreateOr(X, Y);
3429 
3430   // (A & C) | (B & D)
3431   Value *A, *B, *C, *D;
3432   if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3433       match(Op1, m_And(m_Value(B), m_Value(D)))) {
3434 
3435     // (A & C0) | (B & C1)
3436     const APInt *C0, *C1;
3437     if (match(C, m_APInt(C0)) && match(D, m_APInt(C1))) {
3438       Value *X;
3439       if (*C0 == ~*C1) {
3440         // ((X | B) & MaskC) | (B & ~MaskC) -> (X & MaskC) | B
3441         if (match(A, m_c_Or(m_Value(X), m_Specific(B))))
3442           return BinaryOperator::CreateOr(Builder.CreateAnd(X, *C0), B);
3443         // (A & MaskC) | ((X | A) & ~MaskC) -> (X & ~MaskC) | A
3444         if (match(B, m_c_Or(m_Specific(A), m_Value(X))))
3445           return BinaryOperator::CreateOr(Builder.CreateAnd(X, *C1), A);
3446 
3447         // ((X ^ B) & MaskC) | (B & ~MaskC) -> (X & MaskC) ^ B
3448         if (match(A, m_c_Xor(m_Value(X), m_Specific(B))))
3449           return BinaryOperator::CreateXor(Builder.CreateAnd(X, *C0), B);
3450         // (A & MaskC) | ((X ^ A) & ~MaskC) -> (X & ~MaskC) ^ A
3451         if (match(B, m_c_Xor(m_Specific(A), m_Value(X))))
3452           return BinaryOperator::CreateXor(Builder.CreateAnd(X, *C1), A);
3453       }
3454 
3455       if ((*C0 & *C1).isZero()) {
3456         // ((X | B) & C0) | (B & C1) --> (X | B) & (C0 | C1)
3457         // iff (C0 & C1) == 0 and (X & ~C0) == 0
3458         if (match(A, m_c_Or(m_Value(X), m_Specific(B))) &&
3459             MaskedValueIsZero(X, ~*C0, 0, &I)) {
3460           Constant *C01 = ConstantInt::get(Ty, *C0 | *C1);
3461           return BinaryOperator::CreateAnd(A, C01);
3462         }
3463         // (A & C0) | ((X | A) & C1) --> (X | A) & (C0 | C1)
3464         // iff (C0 & C1) == 0 and (X & ~C1) == 0
3465         if (match(B, m_c_Or(m_Value(X), m_Specific(A))) &&
3466             MaskedValueIsZero(X, ~*C1, 0, &I)) {
3467           Constant *C01 = ConstantInt::get(Ty, *C0 | *C1);
3468           return BinaryOperator::CreateAnd(B, C01);
3469         }
3470         // ((X | C2) & C0) | ((X | C3) & C1) --> (X | C2 | C3) & (C0 | C1)
3471         // iff (C0 & C1) == 0 and (C2 & ~C0) == 0 and (C3 & ~C1) == 0.
3472         const APInt *C2, *C3;
3473         if (match(A, m_Or(m_Value(X), m_APInt(C2))) &&
3474             match(B, m_Or(m_Specific(X), m_APInt(C3))) &&
3475             (*C2 & ~*C0).isZero() && (*C3 & ~*C1).isZero()) {
3476           Value *Or = Builder.CreateOr(X, *C2 | *C3, "bitfield");
3477           Constant *C01 = ConstantInt::get(Ty, *C0 | *C1);
3478           return BinaryOperator::CreateAnd(Or, C01);
3479         }
3480       }
3481     }
3482 
3483     // Don't try to form a select if it's unlikely that we'll get rid of at
3484     // least one of the operands. A select is generally more expensive than the
3485     // 'or' that it is replacing.
3486     if (Op0->hasOneUse() || Op1->hasOneUse()) {
3487       // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.
3488       if (Value *V = matchSelectFromAndOr(A, C, B, D))
3489         return replaceInstUsesWith(I, V);
3490       if (Value *V = matchSelectFromAndOr(A, C, D, B))
3491         return replaceInstUsesWith(I, V);
3492       if (Value *V = matchSelectFromAndOr(C, A, B, D))
3493         return replaceInstUsesWith(I, V);
3494       if (Value *V = matchSelectFromAndOr(C, A, D, B))
3495         return replaceInstUsesWith(I, V);
3496       if (Value *V = matchSelectFromAndOr(B, D, A, C))
3497         return replaceInstUsesWith(I, V);
3498       if (Value *V = matchSelectFromAndOr(B, D, C, A))
3499         return replaceInstUsesWith(I, V);
3500       if (Value *V = matchSelectFromAndOr(D, B, A, C))
3501         return replaceInstUsesWith(I, V);
3502       if (Value *V = matchSelectFromAndOr(D, B, C, A))
3503         return replaceInstUsesWith(I, V);
3504     }
3505   }
3506 
3507   if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3508       match(Op1, m_Not(m_Or(m_Value(B), m_Value(D)))) &&
3509       (Op0->hasOneUse() || Op1->hasOneUse())) {
3510     // (Cond & C) | ~(Cond | D) -> Cond ? C : ~D
3511     if (Value *V = matchSelectFromAndOr(A, C, B, D, true))
3512       return replaceInstUsesWith(I, V);
3513     if (Value *V = matchSelectFromAndOr(A, C, D, B, true))
3514       return replaceInstUsesWith(I, V);
3515     if (Value *V = matchSelectFromAndOr(C, A, B, D, true))
3516       return replaceInstUsesWith(I, V);
3517     if (Value *V = matchSelectFromAndOr(C, A, D, B, true))
3518       return replaceInstUsesWith(I, V);
3519   }
3520 
3521   // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
3522   if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
3523     if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
3524       return BinaryOperator::CreateOr(Op0, C);
3525 
3526   // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
3527   if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
3528     if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
3529       return BinaryOperator::CreateOr(Op1, C);
3530 
3531   // ((A & B) ^ C) | B -> C | B
3532   if (match(Op0, m_c_Xor(m_c_And(m_Value(A), m_Specific(Op1)), m_Value(C))))
3533     return BinaryOperator::CreateOr(C, Op1);
3534 
3535   // B | ((A & B) ^ C) -> B | C
3536   if (match(Op1, m_c_Xor(m_c_And(m_Value(A), m_Specific(Op0)), m_Value(C))))
3537     return BinaryOperator::CreateOr(Op0, C);
3538 
3539   // ((B | C) & A) | B -> B | (A & C)
3540   if (match(Op0, m_c_And(m_c_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
3541     return BinaryOperator::CreateOr(Op1, Builder.CreateAnd(A, C));
3542 
3543   // B | ((B | C) & A) -> B | (A & C)
3544   if (match(Op1, m_c_And(m_c_Or(m_Specific(Op0), m_Value(C)), m_Value(A))))
3545     return BinaryOperator::CreateOr(Op0, Builder.CreateAnd(A, C));
3546 
3547   if (Instruction *DeMorgan = matchDeMorgansLaws(I, *this))
3548     return DeMorgan;
3549 
3550   // Canonicalize xor to the RHS.
3551   bool SwappedForXor = false;
3552   if (match(Op0, m_Xor(m_Value(), m_Value()))) {
3553     std::swap(Op0, Op1);
3554     SwappedForXor = true;
3555   }
3556 
3557   if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3558     // (A | ?) | (A ^ B) --> (A | ?) | B
3559     // (B | ?) | (A ^ B) --> (B | ?) | A
3560     if (match(Op0, m_c_Or(m_Specific(A), m_Value())))
3561       return BinaryOperator::CreateOr(Op0, B);
3562     if (match(Op0, m_c_Or(m_Specific(B), m_Value())))
3563       return BinaryOperator::CreateOr(Op0, A);
3564 
3565     // (A & B) | (A ^ B) --> A | B
3566     // (B & A) | (A ^ B) --> A | B
3567     if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
3568         match(Op0, m_And(m_Specific(B), m_Specific(A))))
3569       return BinaryOperator::CreateOr(A, B);
3570 
3571     // ~A | (A ^ B) --> ~(A & B)
3572     // ~B | (A ^ B) --> ~(A & B)
3573     // The swap above should always make Op0 the 'not'.
3574     if ((Op0->hasOneUse() || Op1->hasOneUse()) &&
3575         (match(Op0, m_Not(m_Specific(A))) || match(Op0, m_Not(m_Specific(B)))))
3576       return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));
3577 
3578     // Same as above, but peek through an 'and' to the common operand:
3579     // ~(A & ?) | (A ^ B) --> ~((A & ?) & B)
3580     // ~(B & ?) | (A ^ B) --> ~((B & ?) & A)
3581     Instruction *And;
3582     if ((Op0->hasOneUse() || Op1->hasOneUse()) &&
3583         match(Op0, m_Not(m_CombineAnd(m_Instruction(And),
3584                                       m_c_And(m_Specific(A), m_Value())))))
3585       return BinaryOperator::CreateNot(Builder.CreateAnd(And, B));
3586     if ((Op0->hasOneUse() || Op1->hasOneUse()) &&
3587         match(Op0, m_Not(m_CombineAnd(m_Instruction(And),
3588                                       m_c_And(m_Specific(B), m_Value())))))
3589       return BinaryOperator::CreateNot(Builder.CreateAnd(And, A));
3590 
3591     // (~A | C) | (A ^ B) --> ~(A & B) | C
3592     // (~B | C) | (A ^ B) --> ~(A & B) | C
3593     if (Op0->hasOneUse() && Op1->hasOneUse() &&
3594         (match(Op0, m_c_Or(m_Not(m_Specific(A)), m_Value(C))) ||
3595          match(Op0, m_c_Or(m_Not(m_Specific(B)), m_Value(C))))) {
3596       Value *Nand = Builder.CreateNot(Builder.CreateAnd(A, B), "nand");
3597       return BinaryOperator::CreateOr(Nand, C);
3598     }
3599 
3600     // A | (~A ^ B) --> ~B | A
3601     // B | (A ^ ~B) --> ~A | B
3602     if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
3603       Value *NotB = Builder.CreateNot(B, B->getName() + ".not");
3604       return BinaryOperator::CreateOr(NotB, Op0);
3605     }
3606     if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
3607       Value *NotA = Builder.CreateNot(A, A->getName() + ".not");
3608       return BinaryOperator::CreateOr(NotA, Op0);
3609     }
3610   }
3611 
3612   // A | ~(A | B) -> A | ~B
3613   // A | ~(A ^ B) -> A | ~B
3614   if (match(Op1, m_Not(m_Value(A))))
3615     if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
3616       if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
3617           Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
3618                                B->getOpcode() == Instruction::Xor)) {
3619         Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
3620                                                  B->getOperand(0);
3621         Value *Not = Builder.CreateNot(NotOp, NotOp->getName() + ".not");
3622         return BinaryOperator::CreateOr(Not, Op0);
3623       }
3624 
3625   if (SwappedForXor)
3626     std::swap(Op0, Op1);
3627 
3628   {
3629     ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
3630     ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
3631     if (LHS && RHS)
3632       if (Value *Res = foldAndOrOfICmps(LHS, RHS, I, /* IsAnd */ false))
3633         return replaceInstUsesWith(I, Res);
3634 
3635     // TODO: Make this recursive; it's a little tricky because an arbitrary
3636     // number of 'or' instructions might have to be created.
3637     Value *X, *Y;
3638     if (LHS && match(Op1, m_OneUse(m_LogicalOr(m_Value(X), m_Value(Y))))) {
3639       bool IsLogical = isa<SelectInst>(Op1);
3640       // LHS | (X || Y) --> (LHS || X) || Y
3641       if (auto *Cmp = dyn_cast<ICmpInst>(X))
3642         if (Value *Res =
3643                 foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ false, IsLogical))
3644           return replaceInstUsesWith(I, IsLogical
3645                                             ? Builder.CreateLogicalOr(Res, Y)
3646                                             : Builder.CreateOr(Res, Y));
3647       // LHS | (X || Y) --> X || (LHS | Y)
3648       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
3649         if (Value *Res = foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ false,
3650                                           /* IsLogical */ false))
3651           return replaceInstUsesWith(I, IsLogical
3652                                             ? Builder.CreateLogicalOr(X, Res)
3653                                             : Builder.CreateOr(X, Res));
3654     }
3655     if (RHS && match(Op0, m_OneUse(m_LogicalOr(m_Value(X), m_Value(Y))))) {
3656       bool IsLogical = isa<SelectInst>(Op0);
3657       // (X || Y) | RHS --> (X || RHS) || Y
3658       if (auto *Cmp = dyn_cast<ICmpInst>(X))
3659         if (Value *Res =
3660                 foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ false, IsLogical))
3661           return replaceInstUsesWith(I, IsLogical
3662                                             ? Builder.CreateLogicalOr(Res, Y)
3663                                             : Builder.CreateOr(Res, Y));
3664       // (X || Y) | RHS --> X || (Y | RHS)
3665       if (auto *Cmp = dyn_cast<ICmpInst>(Y))
3666         if (Value *Res = foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ false,
3667                                           /* IsLogical */ false))
3668           return replaceInstUsesWith(I, IsLogical
3669                                             ? Builder.CreateLogicalOr(X, Res)
3670                                             : Builder.CreateOr(X, Res));
3671     }
3672   }
3673 
3674   if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
3675     if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
3676       if (Value *Res = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ false))
3677         return replaceInstUsesWith(I, Res);
3678 
3679   if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))
3680     return FoldedFCmps;
3681 
3682   if (Instruction *CastedOr = foldCastedBitwiseLogic(I))
3683     return CastedOr;
3684 
3685   if (Instruction *Sel = foldBinopOfSextBoolToSelect(I))
3686     return Sel;
3687 
3688   // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.
3689   // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold
3690   //       with binop identity constant. But creating a select with non-constant
3691   //       arm may not be reversible due to poison semantics. Is that a good
3692   //       canonicalization?
3693   if (match(&I, m_c_Or(m_OneUse(m_SExt(m_Value(A))), m_Value(B))) &&
3694       A->getType()->isIntOrIntVectorTy(1))
3695     return SelectInst::Create(A, ConstantInt::getAllOnesValue(Ty), B);
3696 
3697   // Note: If we've gotten to the point of visiting the outer OR, then the
3698   // inner one couldn't be simplified.  If it was a constant, then it won't
3699   // be simplified by a later pass either, so we try swapping the inner/outer
3700   // ORs in the hopes that we'll be able to simplify it this way.
3701   // (X|C) | V --> (X|V) | C
3702   ConstantInt *CI;
3703   if (Op0->hasOneUse() && !match(Op1, m_ConstantInt()) &&
3704       match(Op0, m_Or(m_Value(A), m_ConstantInt(CI)))) {
3705     Value *Inner = Builder.CreateOr(A, Op1);
3706     Inner->takeName(Op0);
3707     return BinaryOperator::CreateOr(Inner, CI);
3708   }
3709 
3710   // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
3711   // Since this OR statement hasn't been optimized further yet, we hope
3712   // that this transformation will allow the new ORs to be optimized.
3713   {
3714     Value *X = nullptr, *Y = nullptr;
3715     if (Op0->hasOneUse() && Op1->hasOneUse() &&
3716         match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
3717         match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
3718       Value *orTrue = Builder.CreateOr(A, C);
3719       Value *orFalse = Builder.CreateOr(B, D);
3720       return SelectInst::Create(X, orTrue, orFalse);
3721     }
3722   }
3723 
3724   // or(ashr(subNSW(Y, X), ScalarSizeInBits(Y) - 1), X)  --> X s> Y ? -1 : X.
3725   {
3726     Value *X, *Y;
3727     if (match(&I, m_c_Or(m_OneUse(m_AShr(
3728                              m_NSWSub(m_Value(Y), m_Value(X)),
3729                              m_SpecificInt(Ty->getScalarSizeInBits() - 1))),
3730                          m_Deferred(X)))) {
3731       Value *NewICmpInst = Builder.CreateICmpSGT(X, Y);
3732       Value *AllOnes = ConstantInt::getAllOnesValue(Ty);
3733       return SelectInst::Create(NewICmpInst, AllOnes, X);
3734     }
3735   }
3736 
3737   {
3738     // ((A & B) ^ A) | ((A & B) ^ B) -> A ^ B
3739     // (A ^ (A & B)) | (B ^ (A & B)) -> A ^ B
3740     // ((A & B) ^ B) | ((A & B) ^ A) -> A ^ B
3741     // (B ^ (A & B)) | (A ^ (A & B)) -> A ^ B
3742     const auto TryXorOpt = [&](Value *Lhs, Value *Rhs) -> Instruction * {
3743       if (match(Lhs, m_c_Xor(m_And(m_Value(A), m_Value(B)), m_Deferred(A))) &&
3744           match(Rhs,
3745                 m_c_Xor(m_And(m_Specific(A), m_Specific(B)), m_Deferred(B)))) {
3746         return BinaryOperator::CreateXor(A, B);
3747       }
3748       return nullptr;
3749     };
3750 
3751     if (Instruction *Result = TryXorOpt(Op0, Op1))
3752       return Result;
3753     if (Instruction *Result = TryXorOpt(Op1, Op0))
3754       return Result;
3755   }
3756 
3757   if (Instruction *V =
3758           canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
3759     return V;
3760 
3761   CmpInst::Predicate Pred;
3762   Value *Mul, *Ov, *MulIsNotZero, *UMulWithOv;
3763   // Check if the OR weakens the overflow condition for umul.with.overflow by
3764   // treating any non-zero result as overflow. In that case, we overflow if both
3765   // umul.with.overflow operands are != 0, as in that case the result can only
3766   // be 0, iff the multiplication overflows.
3767   if (match(&I,
3768             m_c_Or(m_CombineAnd(m_ExtractValue<1>(m_Value(UMulWithOv)),
3769                                 m_Value(Ov)),
3770                    m_CombineAnd(m_ICmp(Pred,
3771                                        m_CombineAnd(m_ExtractValue<0>(
3772                                                         m_Deferred(UMulWithOv)),
3773                                                     m_Value(Mul)),
3774                                        m_ZeroInt()),
3775                                 m_Value(MulIsNotZero)))) &&
3776       (Ov->hasOneUse() || (MulIsNotZero->hasOneUse() && Mul->hasOneUse())) &&
3777       Pred == CmpInst::ICMP_NE) {
3778     Value *A, *B;
3779     if (match(UMulWithOv, m_Intrinsic<Intrinsic::umul_with_overflow>(
3780                               m_Value(A), m_Value(B)))) {
3781       Value *NotNullA = Builder.CreateIsNotNull(A);
3782       Value *NotNullB = Builder.CreateIsNotNull(B);
3783       return BinaryOperator::CreateAnd(NotNullA, NotNullB);
3784     }
3785   }
3786 
3787   /// Res, Overflow = xxx_with_overflow X, C1
3788   /// Try to canonicalize the pattern "Overflow | icmp pred Res, C2" into
3789   /// "Overflow | icmp pred X, C2 +/- C1".
3790   const WithOverflowInst *WO;
3791   const Value *WOV;
3792   const APInt *C1, *C2;
3793   if (match(&I, m_c_Or(m_CombineAnd(m_ExtractValue<1>(m_CombineAnd(
3794                                         m_WithOverflowInst(WO), m_Value(WOV))),
3795                                     m_Value(Ov)),
3796                        m_OneUse(m_ICmp(Pred, m_ExtractValue<0>(m_Deferred(WOV)),
3797                                        m_APInt(C2))))) &&
3798       (WO->getBinaryOp() == Instruction::Add ||
3799        WO->getBinaryOp() == Instruction::Sub) &&
3800       (ICmpInst::isEquality(Pred) ||
3801        WO->isSigned() == ICmpInst::isSigned(Pred)) &&
3802       match(WO->getRHS(), m_APInt(C1))) {
3803     bool Overflow;
3804     APInt NewC = WO->getBinaryOp() == Instruction::Add
3805                      ? (ICmpInst::isSigned(Pred) ? C2->ssub_ov(*C1, Overflow)
3806                                                  : C2->usub_ov(*C1, Overflow))
3807                      : (ICmpInst::isSigned(Pred) ? C2->sadd_ov(*C1, Overflow)
3808                                                  : C2->uadd_ov(*C1, Overflow));
3809     if (!Overflow || ICmpInst::isEquality(Pred)) {
3810       Value *NewCmp = Builder.CreateICmp(
3811           Pred, WO->getLHS(), ConstantInt::get(WO->getLHS()->getType(), NewC));
3812       return BinaryOperator::CreateOr(Ov, NewCmp);
3813     }
3814   }
3815 
3816   // (~x) | y  -->  ~(x & (~y))  iff that gets rid of inversions
3817   if (sinkNotIntoOtherHandOfLogicalOp(I))
3818     return &I;
3819 
3820   // Improve "get low bit mask up to and including bit X" pattern:
3821   //   (1 << X) | ((1 << X) + -1)  -->  -1 l>> (bitwidth(x) - 1 - X)
3822   if (match(&I, m_c_Or(m_Add(m_Shl(m_One(), m_Value(X)), m_AllOnes()),
3823                        m_Shl(m_One(), m_Deferred(X)))) &&
3824       match(&I, m_c_Or(m_OneUse(m_Value()), m_Value()))) {
3825     Value *Sub = Builder.CreateSub(
3826         ConstantInt::get(Ty, Ty->getScalarSizeInBits() - 1), X);
3827     return BinaryOperator::CreateLShr(Constant::getAllOnesValue(Ty), Sub);
3828   }
3829 
3830   // An or recurrence w/loop invariant step is equivelent to (or start, step)
3831   PHINode *PN = nullptr;
3832   Value *Start = nullptr, *Step = nullptr;
3833   if (matchSimpleRecurrence(&I, PN, Start, Step) && DT.dominates(Step, PN))
3834     return replaceInstUsesWith(I, Builder.CreateOr(Start, Step));
3835 
3836   // (A & B) | (C | D) or (C | D) | (A & B)
3837   // Can be combined if C or D is of type (A/B & X)
3838   if (match(&I, m_c_Or(m_OneUse(m_And(m_Value(A), m_Value(B))),
3839                        m_OneUse(m_Or(m_Value(C), m_Value(D)))))) {
3840     // (A & B) | (C | ?) -> C | (? | (A & B))
3841     // (A & B) | (C | ?) -> C | (? | (A & B))
3842     // (A & B) | (C | ?) -> C | (? | (A & B))
3843     // (A & B) | (C | ?) -> C | (? | (A & B))
3844     // (C | ?) | (A & B) -> C | (? | (A & B))
3845     // (C | ?) | (A & B) -> C | (? | (A & B))
3846     // (C | ?) | (A & B) -> C | (? | (A & B))
3847     // (C | ?) | (A & B) -> C | (? | (A & B))
3848     if (match(D, m_OneUse(m_c_And(m_Specific(A), m_Value()))) ||
3849         match(D, m_OneUse(m_c_And(m_Specific(B), m_Value()))))
3850       return BinaryOperator::CreateOr(
3851           C, Builder.CreateOr(D, Builder.CreateAnd(A, B)));
3852     // (A & B) | (? | D) -> (? | (A & B)) | D
3853     // (A & B) | (? | D) -> (? | (A & B)) | D
3854     // (A & B) | (? | D) -> (? | (A & B)) | D
3855     // (A & B) | (? | D) -> (? | (A & B)) | D
3856     // (? | D) | (A & B) -> (? | (A & B)) | D
3857     // (? | D) | (A & B) -> (? | (A & B)) | D
3858     // (? | D) | (A & B) -> (? | (A & B)) | D
3859     // (? | D) | (A & B) -> (? | (A & B)) | D
3860     if (match(C, m_OneUse(m_c_And(m_Specific(A), m_Value()))) ||
3861         match(C, m_OneUse(m_c_And(m_Specific(B), m_Value()))))
3862       return BinaryOperator::CreateOr(
3863           Builder.CreateOr(C, Builder.CreateAnd(A, B)), D);
3864   }
3865 
3866   if (Instruction *R = reassociateForUses(I, Builder))
3867     return R;
3868 
3869   if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))
3870     return Canonicalized;
3871 
3872   if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1))
3873     return Folded;
3874 
3875   if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
3876     return Res;
3877 
3878   // If we are setting the sign bit of a floating-point value, convert
3879   // this to fneg(fabs), then cast back to integer.
3880   //
3881   // If the result isn't immediately cast back to a float, this will increase
3882   // the number of instructions. This is still probably a better canonical form
3883   // as it enables FP value tracking.
3884   //
3885   // Assumes any IEEE-represented type has the sign bit in the high bit.
3886   //
3887   // This is generous interpretation of noimplicitfloat, this is not a true
3888   // floating-point operation.
3889   Value *CastOp;
3890   if (match(Op0, m_BitCast(m_Value(CastOp))) && match(Op1, m_SignMask()) &&
3891       !Builder.GetInsertBlock()->getParent()->hasFnAttribute(
3892           Attribute::NoImplicitFloat)) {
3893     Type *EltTy = CastOp->getType()->getScalarType();
3894     if (EltTy->isFloatingPointTy() && EltTy->isIEEE() &&
3895         EltTy->getPrimitiveSizeInBits() ==
3896         I.getType()->getScalarType()->getPrimitiveSizeInBits()) {
3897       Value *FAbs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, CastOp);
3898       Value *FNegFAbs = Builder.CreateFNeg(FAbs);
3899       return new BitCastInst(FNegFAbs, I.getType());
3900     }
3901   }
3902 
3903   // (X & C1) | C2 -> X & (C1 | C2) iff (X & C2) == C2
3904   if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C1)))) &&
3905       match(Op1, m_APInt(C2))) {
3906     KnownBits KnownX = computeKnownBits(X, /*Depth*/ 0, &I);
3907     if ((KnownX.One & *C2) == *C2)
3908       return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *C1 | *C2));
3909   }
3910 
3911   if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))
3912     return Res;
3913 
3914   return nullptr;
3915 }
3916 
3917 /// A ^ B can be specified using other logic ops in a variety of patterns. We
3918 /// can fold these early and efficiently by morphing an existing instruction.
3919 static Instruction *foldXorToXor(BinaryOperator &I,
3920                                  InstCombiner::BuilderTy &Builder) {
3921   assert(I.getOpcode() == Instruction::Xor);
3922   Value *Op0 = I.getOperand(0);
3923   Value *Op1 = I.getOperand(1);
3924   Value *A, *B;
3925 
3926   // There are 4 commuted variants for each of the basic patterns.
3927 
3928   // (A & B) ^ (A | B) -> A ^ B
3929   // (A & B) ^ (B | A) -> A ^ B
3930   // (A | B) ^ (A & B) -> A ^ B
3931   // (A | B) ^ (B & A) -> A ^ B
3932   if (match(&I, m_c_Xor(m_And(m_Value(A), m_Value(B)),
3933                         m_c_Or(m_Deferred(A), m_Deferred(B)))))
3934     return BinaryOperator::CreateXor(A, B);
3935 
3936   // (A | ~B) ^ (~A | B) -> A ^ B
3937   // (~B | A) ^ (~A | B) -> A ^ B
3938   // (~A | B) ^ (A | ~B) -> A ^ B
3939   // (B | ~A) ^ (A | ~B) -> A ^ B
3940   if (match(&I, m_Xor(m_c_Or(m_Value(A), m_Not(m_Value(B))),
3941                       m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))
3942     return BinaryOperator::CreateXor(A, B);
3943 
3944   // (A & ~B) ^ (~A & B) -> A ^ B
3945   // (~B & A) ^ (~A & B) -> A ^ B
3946   // (~A & B) ^ (A & ~B) -> A ^ B
3947   // (B & ~A) ^ (A & ~B) -> A ^ B
3948   if (match(&I, m_Xor(m_c_And(m_Value(A), m_Not(m_Value(B))),
3949                       m_c_And(m_Not(m_Deferred(A)), m_Deferred(B)))))
3950     return BinaryOperator::CreateXor(A, B);
3951 
3952   // For the remaining cases we need to get rid of one of the operands.
3953   if (!Op0->hasOneUse() && !Op1->hasOneUse())
3954     return nullptr;
3955 
3956   // (A | B) ^ ~(A & B) -> ~(A ^ B)
3957   // (A | B) ^ ~(B & A) -> ~(A ^ B)
3958   // (A & B) ^ ~(A | B) -> ~(A ^ B)
3959   // (A & B) ^ ~(B | A) -> ~(A ^ B)
3960   // Complexity sorting ensures the not will be on the right side.
3961   if ((match(Op0, m_Or(m_Value(A), m_Value(B))) &&
3962        match(Op1, m_Not(m_c_And(m_Specific(A), m_Specific(B))))) ||
3963       (match(Op0, m_And(m_Value(A), m_Value(B))) &&
3964        match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))))
3965     return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
3966 
3967   return nullptr;
3968 }
3969 
3970 Value *InstCombinerImpl::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS,
3971                                         BinaryOperator &I) {
3972   assert(I.getOpcode() == Instruction::Xor && I.getOperand(0) == LHS &&
3973          I.getOperand(1) == RHS && "Should be 'xor' with these operands");
3974 
3975   ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
3976   Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
3977   Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
3978 
3979   if (predicatesFoldable(PredL, PredR)) {
3980     if (LHS0 == RHS1 && LHS1 == RHS0) {
3981       std::swap(LHS0, LHS1);
3982       PredL = ICmpInst::getSwappedPredicate(PredL);
3983     }
3984     if (LHS0 == RHS0 && LHS1 == RHS1) {
3985       // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
3986       unsigned Code = getICmpCode(PredL) ^ getICmpCode(PredR);
3987       bool IsSigned = LHS->isSigned() || RHS->isSigned();
3988       return getNewICmpValue(Code, IsSigned, LHS0, LHS1, Builder);
3989     }
3990   }
3991 
3992   // TODO: This can be generalized to compares of non-signbits using
3993   // decomposeBitTestICmp(). It could be enhanced more by using (something like)
3994   // foldLogOpOfMaskedICmps().
3995   const APInt *LC, *RC;
3996   if (match(LHS1, m_APInt(LC)) && match(RHS1, m_APInt(RC)) &&
3997       LHS0->getType() == RHS0->getType() &&
3998       LHS0->getType()->isIntOrIntVectorTy()) {
3999     // Convert xor of signbit tests to signbit test of xor'd values:
4000     // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0
4001     // (X <  0) ^ (Y <  0) --> (X ^ Y) < 0
4002     // (X > -1) ^ (Y <  0) --> (X ^ Y) > -1
4003     // (X <  0) ^ (Y > -1) --> (X ^ Y) > -1
4004     bool TrueIfSignedL, TrueIfSignedR;
4005     if ((LHS->hasOneUse() || RHS->hasOneUse()) &&
4006         isSignBitCheck(PredL, *LC, TrueIfSignedL) &&
4007         isSignBitCheck(PredR, *RC, TrueIfSignedR)) {
4008       Value *XorLR = Builder.CreateXor(LHS0, RHS0);
4009       return TrueIfSignedL == TrueIfSignedR ? Builder.CreateIsNeg(XorLR) :
4010                                               Builder.CreateIsNotNeg(XorLR);
4011     }
4012 
4013     // Fold (icmp pred1 X, C1) ^ (icmp pred2 X, C2)
4014     // into a single comparison using range-based reasoning.
4015     if (LHS0 == RHS0) {
4016       ConstantRange CR1 = ConstantRange::makeExactICmpRegion(PredL, *LC);
4017       ConstantRange CR2 = ConstantRange::makeExactICmpRegion(PredR, *RC);
4018       auto CRUnion = CR1.exactUnionWith(CR2);
4019       auto CRIntersect = CR1.exactIntersectWith(CR2);
4020       if (CRUnion && CRIntersect)
4021         if (auto CR = CRUnion->exactIntersectWith(CRIntersect->inverse())) {
4022           if (CR->isFullSet())
4023             return ConstantInt::getTrue(I.getType());
4024           if (CR->isEmptySet())
4025             return ConstantInt::getFalse(I.getType());
4026 
4027           CmpInst::Predicate NewPred;
4028           APInt NewC, Offset;
4029           CR->getEquivalentICmp(NewPred, NewC, Offset);
4030 
4031           if ((Offset.isZero() && (LHS->hasOneUse() || RHS->hasOneUse())) ||
4032               (LHS->hasOneUse() && RHS->hasOneUse())) {
4033             Value *NewV = LHS0;
4034             Type *Ty = LHS0->getType();
4035             if (!Offset.isZero())
4036               NewV = Builder.CreateAdd(NewV, ConstantInt::get(Ty, Offset));
4037             return Builder.CreateICmp(NewPred, NewV,
4038                                       ConstantInt::get(Ty, NewC));
4039           }
4040         }
4041     }
4042   }
4043 
4044   // Instead of trying to imitate the folds for and/or, decompose this 'xor'
4045   // into those logic ops. That is, try to turn this into an and-of-icmps
4046   // because we have many folds for that pattern.
4047   //
4048   // This is based on a truth table definition of xor:
4049   // X ^ Y --> (X | Y) & !(X & Y)
4050   if (Value *OrICmp = simplifyBinOp(Instruction::Or, LHS, RHS, SQ)) {
4051     // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y).
4052     // TODO: If OrICmp is false, the whole thing is false (InstSimplify?).
4053     if (Value *AndICmp = simplifyBinOp(Instruction::And, LHS, RHS, SQ)) {
4054       // TODO: Independently handle cases where the 'and' side is a constant.
4055       ICmpInst *X = nullptr, *Y = nullptr;
4056       if (OrICmp == LHS && AndICmp == RHS) {
4057         // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS  --> X & !Y
4058         X = LHS;
4059         Y = RHS;
4060       }
4061       if (OrICmp == RHS && AndICmp == LHS) {
4062         // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS  --> !Y & X
4063         X = RHS;
4064         Y = LHS;
4065       }
4066       if (X && Y && (Y->hasOneUse() || canFreelyInvertAllUsersOf(Y, &I))) {
4067         // Invert the predicate of 'Y', thus inverting its output.
4068         Y->setPredicate(Y->getInversePredicate());
4069         // So, are there other uses of Y?
4070         if (!Y->hasOneUse()) {
4071           // We need to adapt other uses of Y though. Get a value that matches
4072           // the original value of Y before inversion. While this increases
4073           // immediate instruction count, we have just ensured that all the
4074           // users are freely-invertible, so that 'not' *will* get folded away.
4075           BuilderTy::InsertPointGuard Guard(Builder);
4076           // Set insertion point to right after the Y.
4077           Builder.SetInsertPoint(Y->getParent(), ++(Y->getIterator()));
4078           Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
4079           // Replace all uses of Y (excluding the one in NotY!) with NotY.
4080           Worklist.pushUsersToWorkList(*Y);
4081           Y->replaceUsesWithIf(NotY,
4082                                [NotY](Use &U) { return U.getUser() != NotY; });
4083         }
4084         // All done.
4085         return Builder.CreateAnd(LHS, RHS);
4086       }
4087     }
4088   }
4089 
4090   return nullptr;
4091 }
4092 
4093 /// If we have a masked merge, in the canonical form of:
4094 /// (assuming that A only has one use.)
4095 ///   |        A  |  |B|
4096 ///   ((x ^ y) & M) ^ y
4097 ///    |  D  |
4098 /// * If M is inverted:
4099 ///      |  D  |
4100 ///     ((x ^ y) & ~M) ^ y
4101 ///   We can canonicalize by swapping the final xor operand
4102 ///   to eliminate the 'not' of the mask.
4103 ///     ((x ^ y) & M) ^ x
4104 /// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops
4105 ///   because that shortens the dependency chain and improves analysis:
4106 ///     (x & M) | (y & ~M)
4107 static Instruction *visitMaskedMerge(BinaryOperator &I,
4108                                      InstCombiner::BuilderTy &Builder) {
4109   Value *B, *X, *D;
4110   Value *M;
4111   if (!match(&I, m_c_Xor(m_Value(B),
4112                          m_OneUse(m_c_And(
4113                              m_CombineAnd(m_c_Xor(m_Deferred(B), m_Value(X)),
4114                                           m_Value(D)),
4115                              m_Value(M))))))
4116     return nullptr;
4117 
4118   Value *NotM;
4119   if (match(M, m_Not(m_Value(NotM)))) {
4120     // De-invert the mask and swap the value in B part.
4121     Value *NewA = Builder.CreateAnd(D, NotM);
4122     return BinaryOperator::CreateXor(NewA, X);
4123   }
4124 
4125   Constant *C;
4126   if (D->hasOneUse() && match(M, m_Constant(C))) {
4127     // Propagating undef is unsafe. Clamp undef elements to -1.
4128     Type *EltTy = C->getType()->getScalarType();
4129     C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy));
4130     // Unfold.
4131     Value *LHS = Builder.CreateAnd(X, C);
4132     Value *NotC = Builder.CreateNot(C);
4133     Value *RHS = Builder.CreateAnd(B, NotC);
4134     return BinaryOperator::CreateOr(LHS, RHS);
4135   }
4136 
4137   return nullptr;
4138 }
4139 
4140 static Instruction *foldNotXor(BinaryOperator &I,
4141                                InstCombiner::BuilderTy &Builder) {
4142   Value *X, *Y;
4143   // FIXME: one-use check is not needed in general, but currently we are unable
4144   // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182)
4145   if (!match(&I, m_Not(m_OneUse(m_Xor(m_Value(X), m_Value(Y))))))
4146     return nullptr;
4147 
4148   auto hasCommonOperand = [](Value *A, Value *B, Value *C, Value *D) {
4149     return A == C || A == D || B == C || B == D;
4150   };
4151 
4152   Value *A, *B, *C, *D;
4153   // Canonicalize ~((A & B) ^ (A | ?)) -> (A & B) | ~(A | ?)
4154   // 4 commuted variants
4155   if (match(X, m_And(m_Value(A), m_Value(B))) &&
4156       match(Y, m_Or(m_Value(C), m_Value(D))) && hasCommonOperand(A, B, C, D)) {
4157     Value *NotY = Builder.CreateNot(Y);
4158     return BinaryOperator::CreateOr(X, NotY);
4159   };
4160 
4161   // Canonicalize ~((A | ?) ^ (A & B)) -> (A & B) | ~(A | ?)
4162   // 4 commuted variants
4163   if (match(Y, m_And(m_Value(A), m_Value(B))) &&
4164       match(X, m_Or(m_Value(C), m_Value(D))) && hasCommonOperand(A, B, C, D)) {
4165     Value *NotX = Builder.CreateNot(X);
4166     return BinaryOperator::CreateOr(Y, NotX);
4167   };
4168 
4169   return nullptr;
4170 }
4171 
4172 /// Canonicalize a shifty way to code absolute value to the more common pattern
4173 /// that uses negation and select.
4174 static Instruction *canonicalizeAbs(BinaryOperator &Xor,
4175                                     InstCombiner::BuilderTy &Builder) {
4176   assert(Xor.getOpcode() == Instruction::Xor && "Expected an xor instruction.");
4177 
4178   // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1.
4179   // We're relying on the fact that we only do this transform when the shift has
4180   // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase
4181   // instructions).
4182   Value *Op0 = Xor.getOperand(0), *Op1 = Xor.getOperand(1);
4183   if (Op0->hasNUses(2))
4184     std::swap(Op0, Op1);
4185 
4186   Type *Ty = Xor.getType();
4187   Value *A;
4188   const APInt *ShAmt;
4189   if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
4190       Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
4191       match(Op0, m_OneUse(m_c_Add(m_Specific(A), m_Specific(Op1))))) {
4192     // Op1 = ashr i32 A, 31   ; smear the sign bit
4193     // xor (add A, Op1), Op1  ; add -1 and flip bits if negative
4194     // --> (A < 0) ? -A : A
4195     Value *IsNeg = Builder.CreateIsNeg(A);
4196     // Copy the nuw/nsw flags from the add to the negate.
4197     auto *Add = cast<BinaryOperator>(Op0);
4198     Value *NegA = Builder.CreateNeg(A, "", Add->hasNoUnsignedWrap(),
4199                                    Add->hasNoSignedWrap());
4200     return SelectInst::Create(IsNeg, NegA, A);
4201   }
4202   return nullptr;
4203 }
4204 
4205 static bool canFreelyInvert(InstCombiner &IC, Value *Op,
4206                             Instruction *IgnoredUser) {
4207   auto *I = dyn_cast<Instruction>(Op);
4208   return I && IC.isFreeToInvert(I, /*WillInvertAllUses=*/true) &&
4209          IC.canFreelyInvertAllUsersOf(I, IgnoredUser);
4210 }
4211 
4212 static Value *freelyInvert(InstCombinerImpl &IC, Value *Op,
4213                            Instruction *IgnoredUser) {
4214   auto *I = cast<Instruction>(Op);
4215   IC.Builder.SetInsertPoint(*I->getInsertionPointAfterDef());
4216   Value *NotOp = IC.Builder.CreateNot(Op, Op->getName() + ".not");
4217   Op->replaceUsesWithIf(NotOp,
4218                         [NotOp](Use &U) { return U.getUser() != NotOp; });
4219   IC.freelyInvertAllUsersOf(NotOp, IgnoredUser);
4220   return NotOp;
4221 }
4222 
4223 // Transform
4224 //   z = ~(x &/| y)
4225 // into:
4226 //   z = ((~x) |/& (~y))
4227 // iff both x and y are free to invert and all uses of z can be freely updated.
4228 bool InstCombinerImpl::sinkNotIntoLogicalOp(Instruction &I) {
4229   Value *Op0, *Op1;
4230   if (!match(&I, m_LogicalOp(m_Value(Op0), m_Value(Op1))))
4231     return false;
4232 
4233   // If this logic op has not been simplified yet, just bail out and let that
4234   // happen first. Otherwise, the code below may wrongly invert.
4235   if (Op0 == Op1)
4236     return false;
4237 
4238   Instruction::BinaryOps NewOpc =
4239       match(&I, m_LogicalAnd()) ? Instruction::Or : Instruction::And;
4240   bool IsBinaryOp = isa<BinaryOperator>(I);
4241 
4242   // Can our users be adapted?
4243   if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
4244     return false;
4245 
4246   // And can the operands be adapted?
4247   if (!canFreelyInvert(*this, Op0, &I) || !canFreelyInvert(*this, Op1, &I))
4248     return false;
4249 
4250   Op0 = freelyInvert(*this, Op0, &I);
4251   Op1 = freelyInvert(*this, Op1, &I);
4252 
4253   Builder.SetInsertPoint(*I.getInsertionPointAfterDef());
4254   Value *NewLogicOp;
4255   if (IsBinaryOp)
4256     NewLogicOp = Builder.CreateBinOp(NewOpc, Op0, Op1, I.getName() + ".not");
4257   else
4258     NewLogicOp =
4259         Builder.CreateLogicalOp(NewOpc, Op0, Op1, I.getName() + ".not");
4260 
4261   replaceInstUsesWith(I, NewLogicOp);
4262   // We can not just create an outer `not`, it will most likely be immediately
4263   // folded back, reconstructing our initial pattern, and causing an
4264   // infinite combine loop, so immediately manually fold it away.
4265   freelyInvertAllUsersOf(NewLogicOp);
4266   return true;
4267 }
4268 
4269 // Transform
4270 //   z = (~x) &/| y
4271 // into:
4272 //   z = ~(x |/& (~y))
4273 // iff y is free to invert and all uses of z can be freely updated.
4274 bool InstCombinerImpl::sinkNotIntoOtherHandOfLogicalOp(Instruction &I) {
4275   Value *Op0, *Op1;
4276   if (!match(&I, m_LogicalOp(m_Value(Op0), m_Value(Op1))))
4277     return false;
4278   Instruction::BinaryOps NewOpc =
4279       match(&I, m_LogicalAnd()) ? Instruction::Or : Instruction::And;
4280   bool IsBinaryOp = isa<BinaryOperator>(I);
4281 
4282   Value *NotOp0 = nullptr;
4283   Value *NotOp1 = nullptr;
4284   Value **OpToInvert = nullptr;
4285   if (match(Op0, m_Not(m_Value(NotOp0))) && canFreelyInvert(*this, Op1, &I)) {
4286     Op0 = NotOp0;
4287     OpToInvert = &Op1;
4288   } else if (match(Op1, m_Not(m_Value(NotOp1))) &&
4289              canFreelyInvert(*this, Op0, &I)) {
4290     Op1 = NotOp1;
4291     OpToInvert = &Op0;
4292   } else
4293     return false;
4294 
4295   // And can our users be adapted?
4296   if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr))
4297     return false;
4298 
4299   *OpToInvert = freelyInvert(*this, *OpToInvert, &I);
4300 
4301   Builder.SetInsertPoint(*I.getInsertionPointAfterDef());
4302   Value *NewBinOp;
4303   if (IsBinaryOp)
4304     NewBinOp = Builder.CreateBinOp(NewOpc, Op0, Op1, I.getName() + ".not");
4305   else
4306     NewBinOp = Builder.CreateLogicalOp(NewOpc, Op0, Op1, I.getName() + ".not");
4307   replaceInstUsesWith(I, NewBinOp);
4308   // We can not just create an outer `not`, it will most likely be immediately
4309   // folded back, reconstructing our initial pattern, and causing an
4310   // infinite combine loop, so immediately manually fold it away.
4311   freelyInvertAllUsersOf(NewBinOp);
4312   return true;
4313 }
4314 
4315 Instruction *InstCombinerImpl::foldNot(BinaryOperator &I) {
4316   Value *NotOp;
4317   if (!match(&I, m_Not(m_Value(NotOp))))
4318     return nullptr;
4319 
4320   // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand.
4321   // We must eliminate the and/or (one-use) for these transforms to not increase
4322   // the instruction count.
4323   //
4324   // ~(~X & Y) --> (X | ~Y)
4325   // ~(Y & ~X) --> (X | ~Y)
4326   //
4327   // Note: The logical matches do not check for the commuted patterns because
4328   //       those are handled via SimplifySelectsFeedingBinaryOp().
4329   Type *Ty = I.getType();
4330   Value *X, *Y;
4331   if (match(NotOp, m_OneUse(m_c_And(m_Not(m_Value(X)), m_Value(Y))))) {
4332     Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
4333     return BinaryOperator::CreateOr(X, NotY);
4334   }
4335   if (match(NotOp, m_OneUse(m_LogicalAnd(m_Not(m_Value(X)), m_Value(Y))))) {
4336     Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
4337     return SelectInst::Create(X, ConstantInt::getTrue(Ty), NotY);
4338   }
4339 
4340   // ~(~X | Y) --> (X & ~Y)
4341   // ~(Y | ~X) --> (X & ~Y)
4342   if (match(NotOp, m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y))))) {
4343     Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
4344     return BinaryOperator::CreateAnd(X, NotY);
4345   }
4346   if (match(NotOp, m_OneUse(m_LogicalOr(m_Not(m_Value(X)), m_Value(Y))))) {
4347     Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
4348     return SelectInst::Create(X, NotY, ConstantInt::getFalse(Ty));
4349   }
4350 
4351   // Is this a 'not' (~) fed by a binary operator?
4352   BinaryOperator *NotVal;
4353   if (match(NotOp, m_BinOp(NotVal))) {
4354     // ~((-X) | Y) --> (X - 1) & (~Y)
4355     if (match(NotVal,
4356               m_OneUse(m_c_Or(m_OneUse(m_Neg(m_Value(X))), m_Value(Y))))) {
4357       Value *DecX = Builder.CreateAdd(X, ConstantInt::getAllOnesValue(Ty));
4358       Value *NotY = Builder.CreateNot(Y);
4359       return BinaryOperator::CreateAnd(DecX, NotY);
4360     }
4361 
4362     // ~(~X >>s Y) --> (X >>s Y)
4363     if (match(NotVal, m_AShr(m_Not(m_Value(X)), m_Value(Y))))
4364       return BinaryOperator::CreateAShr(X, Y);
4365 
4366     // Treat lshr with non-negative operand as ashr.
4367     // ~(~X >>u Y) --> (X >>s Y) iff X is known negative
4368     if (match(NotVal, m_LShr(m_Not(m_Value(X)), m_Value(Y))) &&
4369         isKnownNegative(X, SQ.getWithInstruction(NotVal)))
4370       return BinaryOperator::CreateAShr(X, Y);
4371 
4372     // Bit-hack form of a signbit test for iN type:
4373     // ~(X >>s (N - 1)) --> sext i1 (X > -1) to iN
4374     unsigned FullShift = Ty->getScalarSizeInBits() - 1;
4375     if (match(NotVal, m_OneUse(m_AShr(m_Value(X), m_SpecificInt(FullShift))))) {
4376       Value *IsNotNeg = Builder.CreateIsNotNeg(X, "isnotneg");
4377       return new SExtInst(IsNotNeg, Ty);
4378     }
4379 
4380     // If we are inverting a right-shifted constant, we may be able to eliminate
4381     // the 'not' by inverting the constant and using the opposite shift type.
4382     // Canonicalization rules ensure that only a negative constant uses 'ashr',
4383     // but we must check that in case that transform has not fired yet.
4384 
4385     // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits)
4386     Constant *C;
4387     if (match(NotVal, m_AShr(m_Constant(C), m_Value(Y))) &&
4388         match(C, m_Negative())) {
4389       // We matched a negative constant, so propagating undef is unsafe.
4390       // Clamp undef elements to -1.
4391       Type *EltTy = Ty->getScalarType();
4392       C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy));
4393       return BinaryOperator::CreateLShr(ConstantExpr::getNot(C), Y);
4394     }
4395 
4396     // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits)
4397     if (match(NotVal, m_LShr(m_Constant(C), m_Value(Y))) &&
4398         match(C, m_NonNegative())) {
4399       // We matched a non-negative constant, so propagating undef is unsafe.
4400       // Clamp undef elements to 0.
4401       Type *EltTy = Ty->getScalarType();
4402       C = Constant::replaceUndefsWith(C, ConstantInt::getNullValue(EltTy));
4403       return BinaryOperator::CreateAShr(ConstantExpr::getNot(C), Y);
4404     }
4405 
4406     // ~(X + C) --> ~C - X
4407     if (match(NotVal, m_c_Add(m_Value(X), m_ImmConstant(C))))
4408       return BinaryOperator::CreateSub(ConstantExpr::getNot(C), X);
4409 
4410     // ~(X - Y) --> ~X + Y
4411     // FIXME: is it really beneficial to sink the `not` here?
4412     if (match(NotVal, m_Sub(m_Value(X), m_Value(Y))))
4413       if (isa<Constant>(X) || NotVal->hasOneUse())
4414         return BinaryOperator::CreateAdd(Builder.CreateNot(X), Y);
4415 
4416     // ~(~X + Y) --> X - Y
4417     if (match(NotVal, m_c_Add(m_Not(m_Value(X)), m_Value(Y))))
4418       return BinaryOperator::CreateWithCopiedFlags(Instruction::Sub, X, Y,
4419                                                    NotVal);
4420   }
4421 
4422   // not (cmp A, B) = !cmp A, B
4423   CmpInst::Predicate Pred;
4424   if (match(NotOp, m_Cmp(Pred, m_Value(), m_Value())) &&
4425       (NotOp->hasOneUse() ||
4426        InstCombiner::canFreelyInvertAllUsersOf(cast<Instruction>(NotOp),
4427                                                /*IgnoredUser=*/nullptr))) {
4428     cast<CmpInst>(NotOp)->setPredicate(CmpInst::getInversePredicate(Pred));
4429     freelyInvertAllUsersOf(NotOp);
4430     return &I;
4431   }
4432 
4433   // Move a 'not' ahead of casts of a bool to enable logic reduction:
4434   // not (bitcast (sext i1 X)) --> bitcast (sext (not i1 X))
4435   if (match(NotOp, m_OneUse(m_BitCast(m_OneUse(m_SExt(m_Value(X)))))) && X->getType()->isIntOrIntVectorTy(1)) {
4436     Type *SextTy = cast<BitCastOperator>(NotOp)->getSrcTy();
4437     Value *NotX = Builder.CreateNot(X);
4438     Value *Sext = Builder.CreateSExt(NotX, SextTy);
4439     return CastInst::CreateBitOrPointerCast(Sext, Ty);
4440   }
4441 
4442   if (auto *NotOpI = dyn_cast<Instruction>(NotOp))
4443     if (sinkNotIntoLogicalOp(*NotOpI))
4444       return &I;
4445 
4446   // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max:
4447   // ~min(~X, ~Y) --> max(X, Y)
4448   // ~max(~X, Y) --> min(X, ~Y)
4449   auto *II = dyn_cast<IntrinsicInst>(NotOp);
4450   if (II && II->hasOneUse()) {
4451     if (match(NotOp, m_c_MaxOrMin(m_Not(m_Value(X)), m_Value(Y)))) {
4452       Intrinsic::ID InvID = getInverseMinMaxIntrinsic(II->getIntrinsicID());
4453       Value *NotY = Builder.CreateNot(Y);
4454       Value *InvMaxMin = Builder.CreateBinaryIntrinsic(InvID, X, NotY);
4455       return replaceInstUsesWith(I, InvMaxMin);
4456     }
4457 
4458     if (II->getIntrinsicID() == Intrinsic::is_fpclass) {
4459       ConstantInt *ClassMask = cast<ConstantInt>(II->getArgOperand(1));
4460       II->setArgOperand(
4461           1, ConstantInt::get(ClassMask->getType(),
4462                               ~ClassMask->getZExtValue() & fcAllFlags));
4463       return replaceInstUsesWith(I, II);
4464     }
4465   }
4466 
4467   if (NotOp->hasOneUse()) {
4468     // Pull 'not' into operands of select if both operands are one-use compares
4469     // or one is one-use compare and the other one is a constant.
4470     // Inverting the predicates eliminates the 'not' operation.
4471     // Example:
4472     //   not (select ?, (cmp TPred, ?, ?), (cmp FPred, ?, ?) -->
4473     //     select ?, (cmp InvTPred, ?, ?), (cmp InvFPred, ?, ?)
4474     //   not (select ?, (cmp TPred, ?, ?), true -->
4475     //     select ?, (cmp InvTPred, ?, ?), false
4476     if (auto *Sel = dyn_cast<SelectInst>(NotOp)) {
4477       Value *TV = Sel->getTrueValue();
4478       Value *FV = Sel->getFalseValue();
4479       auto *CmpT = dyn_cast<CmpInst>(TV);
4480       auto *CmpF = dyn_cast<CmpInst>(FV);
4481       bool InvertibleT = (CmpT && CmpT->hasOneUse()) || isa<Constant>(TV);
4482       bool InvertibleF = (CmpF && CmpF->hasOneUse()) || isa<Constant>(FV);
4483       if (InvertibleT && InvertibleF) {
4484         if (CmpT)
4485           CmpT->setPredicate(CmpT->getInversePredicate());
4486         else
4487           Sel->setTrueValue(ConstantExpr::getNot(cast<Constant>(TV)));
4488         if (CmpF)
4489           CmpF->setPredicate(CmpF->getInversePredicate());
4490         else
4491           Sel->setFalseValue(ConstantExpr::getNot(cast<Constant>(FV)));
4492         return replaceInstUsesWith(I, Sel);
4493       }
4494     }
4495   }
4496 
4497   if (Instruction *NewXor = foldNotXor(I, Builder))
4498     return NewXor;
4499 
4500   // TODO: Could handle multi-use better by checking if all uses of NotOp (other
4501   // than I) can be inverted.
4502   if (Value *R = getFreelyInverted(NotOp, NotOp->hasOneUse(), &Builder))
4503     return replaceInstUsesWith(I, R);
4504 
4505   return nullptr;
4506 }
4507 
4508 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
4509 // here. We should standardize that construct where it is needed or choose some
4510 // other way to ensure that commutated variants of patterns are not missed.
4511 Instruction *InstCombinerImpl::visitXor(BinaryOperator &I) {
4512   if (Value *V = simplifyXorInst(I.getOperand(0), I.getOperand(1),
4513                                  SQ.getWithInstruction(&I)))
4514     return replaceInstUsesWith(I, V);
4515 
4516   if (SimplifyAssociativeOrCommutative(I))
4517     return &I;
4518 
4519   if (Instruction *X = foldVectorBinop(I))
4520     return X;
4521 
4522   if (Instruction *Phi = foldBinopWithPhiOperands(I))
4523     return Phi;
4524 
4525   if (Instruction *NewXor = foldXorToXor(I, Builder))
4526     return NewXor;
4527 
4528   // (A&B)^(A&C) -> A&(B^C) etc
4529   if (Value *V = foldUsingDistributiveLaws(I))
4530     return replaceInstUsesWith(I, V);
4531 
4532   // See if we can simplify any instructions used by the instruction whose sole
4533   // purpose is to compute bits we don't care about.
4534   if (SimplifyDemandedInstructionBits(I))
4535     return &I;
4536 
4537   if (Instruction *R = foldNot(I))
4538     return R;
4539 
4540   if (Instruction *R = foldBinOpShiftWithShift(I))
4541     return R;
4542 
4543   // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M)
4544   // This it a special case in haveNoCommonBitsSet, but the computeKnownBits
4545   // calls in there are unnecessary as SimplifyDemandedInstructionBits should
4546   // have already taken care of those cases.
4547   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4548   Value *M;
4549   if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(M)), m_Value()),
4550                         m_c_And(m_Deferred(M), m_Value()))))
4551     return BinaryOperator::CreateDisjointOr(Op0, Op1);
4552 
4553   if (Instruction *Xor = visitMaskedMerge(I, Builder))
4554     return Xor;
4555 
4556   Value *X, *Y;
4557   Constant *C1;
4558   if (match(Op1, m_Constant(C1))) {
4559     Constant *C2;
4560 
4561     if (match(Op0, m_OneUse(m_Or(m_Value(X), m_ImmConstant(C2)))) &&
4562         match(C1, m_ImmConstant())) {
4563       // (X | C2) ^ C1 --> (X & ~C2) ^ (C1^C2)
4564       C2 = Constant::replaceUndefsWith(
4565           C2, Constant::getAllOnesValue(C2->getType()->getScalarType()));
4566       Value *And = Builder.CreateAnd(
4567           X, Constant::mergeUndefsWith(ConstantExpr::getNot(C2), C1));
4568       return BinaryOperator::CreateXor(
4569           And, Constant::mergeUndefsWith(ConstantExpr::getXor(C1, C2), C1));
4570     }
4571 
4572     // Use DeMorgan and reassociation to eliminate a 'not' op.
4573     if (match(Op0, m_OneUse(m_Or(m_Not(m_Value(X)), m_Constant(C2))))) {
4574       // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C1
4575       Value *And = Builder.CreateAnd(X, ConstantExpr::getNot(C2));
4576       return BinaryOperator::CreateXor(And, ConstantExpr::getNot(C1));
4577     }
4578     if (match(Op0, m_OneUse(m_And(m_Not(m_Value(X)), m_Constant(C2))))) {
4579       // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C1
4580       Value *Or = Builder.CreateOr(X, ConstantExpr::getNot(C2));
4581       return BinaryOperator::CreateXor(Or, ConstantExpr::getNot(C1));
4582     }
4583 
4584     // Convert xor ([trunc] (ashr X, BW-1)), C =>
4585     //   select(X >s -1, C, ~C)
4586     // The ashr creates "AllZeroOrAllOne's", which then optionally inverses the
4587     // constant depending on whether this input is less than 0.
4588     const APInt *CA;
4589     if (match(Op0, m_OneUse(m_TruncOrSelf(
4590                        m_AShr(m_Value(X), m_APIntAllowUndef(CA))))) &&
4591         *CA == X->getType()->getScalarSizeInBits() - 1 &&
4592         !match(C1, m_AllOnes())) {
4593       assert(!C1->isZeroValue() && "Unexpected xor with 0");
4594       Value *IsNotNeg = Builder.CreateIsNotNeg(X);
4595       return SelectInst::Create(IsNotNeg, Op1, Builder.CreateNot(Op1));
4596     }
4597   }
4598 
4599   Type *Ty = I.getType();
4600   {
4601     const APInt *RHSC;
4602     if (match(Op1, m_APInt(RHSC))) {
4603       Value *X;
4604       const APInt *C;
4605       // (C - X) ^ signmaskC --> (C + signmaskC) - X
4606       if (RHSC->isSignMask() && match(Op0, m_Sub(m_APInt(C), m_Value(X))))
4607         return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C + *RHSC), X);
4608 
4609       // (X + C) ^ signmaskC --> X + (C + signmaskC)
4610       if (RHSC->isSignMask() && match(Op0, m_Add(m_Value(X), m_APInt(C))))
4611         return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C + *RHSC));
4612 
4613       // (X | C) ^ RHSC --> X ^ (C ^ RHSC) iff X & C == 0
4614       if (match(Op0, m_Or(m_Value(X), m_APInt(C))) &&
4615           MaskedValueIsZero(X, *C, 0, &I))
4616         return BinaryOperator::CreateXor(X, ConstantInt::get(Ty, *C ^ *RHSC));
4617 
4618       // When X is a power-of-two or zero and zero input is poison:
4619       // ctlz(i32 X) ^ 31 --> cttz(X)
4620       // cttz(i32 X) ^ 31 --> ctlz(X)
4621       auto *II = dyn_cast<IntrinsicInst>(Op0);
4622       if (II && II->hasOneUse() && *RHSC == Ty->getScalarSizeInBits() - 1) {
4623         Intrinsic::ID IID = II->getIntrinsicID();
4624         if ((IID == Intrinsic::ctlz || IID == Intrinsic::cttz) &&
4625             match(II->getArgOperand(1), m_One()) &&
4626             isKnownToBeAPowerOfTwo(II->getArgOperand(0), /*OrZero */ true)) {
4627           IID = (IID == Intrinsic::ctlz) ? Intrinsic::cttz : Intrinsic::ctlz;
4628           Function *F = Intrinsic::getDeclaration(II->getModule(), IID, Ty);
4629           return CallInst::Create(F, {II->getArgOperand(0), Builder.getTrue()});
4630         }
4631       }
4632 
4633       // If RHSC is inverting the remaining bits of shifted X,
4634       // canonicalize to a 'not' before the shift to help SCEV and codegen:
4635       // (X << C) ^ RHSC --> ~X << C
4636       if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_APInt(C)))) &&
4637           *RHSC == APInt::getAllOnes(Ty->getScalarSizeInBits()).shl(*C)) {
4638         Value *NotX = Builder.CreateNot(X);
4639         return BinaryOperator::CreateShl(NotX, ConstantInt::get(Ty, *C));
4640       }
4641       // (X >>u C) ^ RHSC --> ~X >>u C
4642       if (match(Op0, m_OneUse(m_LShr(m_Value(X), m_APInt(C)))) &&
4643           *RHSC == APInt::getAllOnes(Ty->getScalarSizeInBits()).lshr(*C)) {
4644         Value *NotX = Builder.CreateNot(X);
4645         return BinaryOperator::CreateLShr(NotX, ConstantInt::get(Ty, *C));
4646       }
4647       // TODO: We could handle 'ashr' here as well. That would be matching
4648       //       a 'not' op and moving it before the shift. Doing that requires
4649       //       preventing the inverse fold in canShiftBinOpWithConstantRHS().
4650     }
4651 
4652     // If we are XORing the sign bit of a floating-point value, convert
4653     // this to fneg, then cast back to integer.
4654     //
4655     // This is generous interpretation of noimplicitfloat, this is not a true
4656     // floating-point operation.
4657     //
4658     // Assumes any IEEE-represented type has the sign bit in the high bit.
4659     // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt
4660     Value *CastOp;
4661     if (match(Op0, m_BitCast(m_Value(CastOp))) && match(Op1, m_SignMask()) &&
4662         !Builder.GetInsertBlock()->getParent()->hasFnAttribute(
4663             Attribute::NoImplicitFloat)) {
4664       Type *EltTy = CastOp->getType()->getScalarType();
4665       if (EltTy->isFloatingPointTy() && EltTy->isIEEE() &&
4666           EltTy->getPrimitiveSizeInBits() ==
4667           I.getType()->getScalarType()->getPrimitiveSizeInBits()) {
4668         Value *FNeg = Builder.CreateFNeg(CastOp);
4669         return new BitCastInst(FNeg, I.getType());
4670       }
4671     }
4672   }
4673 
4674   // FIXME: This should not be limited to scalar (pull into APInt match above).
4675   {
4676     Value *X;
4677     ConstantInt *C1, *C2, *C3;
4678     // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
4679     if (match(Op1, m_ConstantInt(C3)) &&
4680         match(Op0, m_LShr(m_Xor(m_Value(X), m_ConstantInt(C1)),
4681                           m_ConstantInt(C2))) &&
4682         Op0->hasOneUse()) {
4683       // fold (C1 >> C2) ^ C3
4684       APInt FoldConst = C1->getValue().lshr(C2->getValue());
4685       FoldConst ^= C3->getValue();
4686       // Prepare the two operands.
4687       auto *Opnd0 = Builder.CreateLShr(X, C2);
4688       Opnd0->takeName(Op0);
4689       return BinaryOperator::CreateXor(Opnd0, ConstantInt::get(Ty, FoldConst));
4690     }
4691   }
4692 
4693   if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
4694     return FoldedLogic;
4695 
4696   // Y ^ (X | Y) --> X & ~Y
4697   // Y ^ (Y | X) --> X & ~Y
4698   if (match(Op1, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op0)))))
4699     return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op0));
4700   // (X | Y) ^ Y --> X & ~Y
4701   // (Y | X) ^ Y --> X & ~Y
4702   if (match(Op0, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op1)))))
4703     return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op1));
4704 
4705   // Y ^ (X & Y) --> ~X & Y
4706   // Y ^ (Y & X) --> ~X & Y
4707   if (match(Op1, m_OneUse(m_c_And(m_Value(X), m_Specific(Op0)))))
4708     return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(X));
4709   // (X & Y) ^ Y --> ~X & Y
4710   // (Y & X) ^ Y --> ~X & Y
4711   // Canonical form is (X & C) ^ C; don't touch that.
4712   // TODO: A 'not' op is better for analysis and codegen, but demanded bits must
4713   //       be fixed to prefer that (otherwise we get infinite looping).
4714   if (!match(Op1, m_Constant()) &&
4715       match(Op0, m_OneUse(m_c_And(m_Value(X), m_Specific(Op1)))))
4716     return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(X));
4717 
4718   Value *A, *B, *C;
4719   // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants.
4720   if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
4721                         m_OneUse(m_c_Or(m_Deferred(A), m_Value(C))))))
4722       return BinaryOperator::CreateXor(
4723           Builder.CreateAnd(Builder.CreateNot(A), C), B);
4724 
4725   // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants.
4726   if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
4727                         m_OneUse(m_c_Or(m_Deferred(B), m_Value(C))))))
4728       return BinaryOperator::CreateXor(
4729           Builder.CreateAnd(Builder.CreateNot(B), C), A);
4730 
4731   // (A & B) ^ (A ^ B) -> (A | B)
4732   if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
4733       match(Op1, m_c_Xor(m_Specific(A), m_Specific(B))))
4734     return BinaryOperator::CreateOr(A, B);
4735   // (A ^ B) ^ (A & B) -> (A | B)
4736   if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
4737       match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
4738     return BinaryOperator::CreateOr(A, B);
4739 
4740   // (A & ~B) ^ ~A -> ~(A & B)
4741   // (~B & A) ^ ~A -> ~(A & B)
4742   if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
4743       match(Op1, m_Not(m_Specific(A))))
4744     return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));
4745 
4746   // (~A & B) ^ A --> A | B -- There are 4 commuted variants.
4747   if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(A)), m_Value(B)), m_Deferred(A))))
4748     return BinaryOperator::CreateOr(A, B);
4749 
4750   // (~A | B) ^ A --> ~(A & B)
4751   if (match(Op0, m_OneUse(m_c_Or(m_Not(m_Specific(Op1)), m_Value(B)))))
4752     return BinaryOperator::CreateNot(Builder.CreateAnd(Op1, B));
4753 
4754   // A ^ (~A | B) --> ~(A & B)
4755   if (match(Op1, m_OneUse(m_c_Or(m_Not(m_Specific(Op0)), m_Value(B)))))
4756     return BinaryOperator::CreateNot(Builder.CreateAnd(Op0, B));
4757 
4758   // (A | B) ^ (A | C) --> (B ^ C) & ~A -- There are 4 commuted variants.
4759   // TODO: Loosen one-use restriction if common operand is a constant.
4760   Value *D;
4761   if (match(Op0, m_OneUse(m_Or(m_Value(A), m_Value(B)))) &&
4762       match(Op1, m_OneUse(m_Or(m_Value(C), m_Value(D))))) {
4763     if (B == C || B == D)
4764       std::swap(A, B);
4765     if (A == C)
4766       std::swap(C, D);
4767     if (A == D) {
4768       Value *NotA = Builder.CreateNot(A);
4769       return BinaryOperator::CreateAnd(Builder.CreateXor(B, C), NotA);
4770     }
4771   }
4772 
4773   // (A & B) ^ (A | C) --> A ? ~B : C -- There are 4 commuted variants.
4774   if (I.getType()->isIntOrIntVectorTy(1) &&
4775       match(Op0, m_OneUse(m_LogicalAnd(m_Value(A), m_Value(B)))) &&
4776       match(Op1, m_OneUse(m_LogicalOr(m_Value(C), m_Value(D))))) {
4777     bool NeedFreeze = isa<SelectInst>(Op0) && isa<SelectInst>(Op1) && B == D;
4778     if (B == C || B == D)
4779       std::swap(A, B);
4780     if (A == C)
4781       std::swap(C, D);
4782     if (A == D) {
4783       if (NeedFreeze)
4784         A = Builder.CreateFreeze(A);
4785       Value *NotB = Builder.CreateNot(B);
4786       return SelectInst::Create(A, NotB, C);
4787     }
4788   }
4789 
4790   if (auto *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4791     if (auto *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4792       if (Value *V = foldXorOfICmps(LHS, RHS, I))
4793         return replaceInstUsesWith(I, V);
4794 
4795   if (Instruction *CastedXor = foldCastedBitwiseLogic(I))
4796     return CastedXor;
4797 
4798   if (Instruction *Abs = canonicalizeAbs(I, Builder))
4799     return Abs;
4800 
4801   // Otherwise, if all else failed, try to hoist the xor-by-constant:
4802   //   (X ^ C) ^ Y --> (X ^ Y) ^ C
4803   // Just like we do in other places, we completely avoid the fold
4804   // for constantexprs, at least to avoid endless combine loop.
4805   if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_CombineAnd(m_Value(X),
4806                                                     m_Unless(m_ConstantExpr())),
4807                                        m_ImmConstant(C1))),
4808                         m_Value(Y))))
4809     return BinaryOperator::CreateXor(Builder.CreateXor(X, Y), C1);
4810 
4811   if (Instruction *R = reassociateForUses(I, Builder))
4812     return R;
4813 
4814   if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder))
4815     return Canonicalized;
4816 
4817   if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1))
4818     return Folded;
4819 
4820   if (Instruction *Folded = canonicalizeConditionalNegationViaMathToSelect(I))
4821     return Folded;
4822 
4823   if (Instruction *Res = foldBinOpOfDisplacedShifts(I))
4824     return Res;
4825 
4826   if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder))
4827     return Res;
4828 
4829   return nullptr;
4830 }
4831