1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the visitAnd, visitOr, and visitXor functions.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "InstCombineInternal.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/Utils/CmpInstAnalysis.h"
20 #include "llvm/Transforms/Utils/Local.h"
21 using namespace llvm;
22 using namespace PatternMatch;
23
24 #define DEBUG_TYPE "instcombine"
25
dyn_castNotVal(Value * V)26 static inline Value *dyn_castNotVal(Value *V) {
27 // If this is not(not(x)) don't return that this is a not: we want the two
28 // not's to be folded first.
29 if (BinaryOperator::isNot(V)) {
30 Value *Operand = BinaryOperator::getNotArgument(V);
31 if (!IsFreeToInvert(Operand, Operand->hasOneUse()))
32 return Operand;
33 }
34
35 // Constants can be considered to be not'ed values...
36 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
37 return ConstantInt::get(C->getType(), ~C->getValue());
38 return nullptr;
39 }
40
41 /// Similar to getICmpCode but for FCmpInst. This encodes a fcmp predicate into
42 /// a four bit mask.
getFCmpCode(FCmpInst::Predicate CC)43 static unsigned getFCmpCode(FCmpInst::Predicate CC) {
44 assert(FCmpInst::FCMP_FALSE <= CC && CC <= FCmpInst::FCMP_TRUE &&
45 "Unexpected FCmp predicate!");
46 // Take advantage of the bit pattern of FCmpInst::Predicate here.
47 // U L G E
48 static_assert(FCmpInst::FCMP_FALSE == 0, ""); // 0 0 0 0
49 static_assert(FCmpInst::FCMP_OEQ == 1, ""); // 0 0 0 1
50 static_assert(FCmpInst::FCMP_OGT == 2, ""); // 0 0 1 0
51 static_assert(FCmpInst::FCMP_OGE == 3, ""); // 0 0 1 1
52 static_assert(FCmpInst::FCMP_OLT == 4, ""); // 0 1 0 0
53 static_assert(FCmpInst::FCMP_OLE == 5, ""); // 0 1 0 1
54 static_assert(FCmpInst::FCMP_ONE == 6, ""); // 0 1 1 0
55 static_assert(FCmpInst::FCMP_ORD == 7, ""); // 0 1 1 1
56 static_assert(FCmpInst::FCMP_UNO == 8, ""); // 1 0 0 0
57 static_assert(FCmpInst::FCMP_UEQ == 9, ""); // 1 0 0 1
58 static_assert(FCmpInst::FCMP_UGT == 10, ""); // 1 0 1 0
59 static_assert(FCmpInst::FCMP_UGE == 11, ""); // 1 0 1 1
60 static_assert(FCmpInst::FCMP_ULT == 12, ""); // 1 1 0 0
61 static_assert(FCmpInst::FCMP_ULE == 13, ""); // 1 1 0 1
62 static_assert(FCmpInst::FCMP_UNE == 14, ""); // 1 1 1 0
63 static_assert(FCmpInst::FCMP_TRUE == 15, ""); // 1 1 1 1
64 return CC;
65 }
66
67 /// This is the complement of getICmpCode, which turns an opcode and two
68 /// operands into either a constant true or false, or a brand new ICmp
69 /// instruction. The sign is passed in to determine which kind of predicate to
70 /// use in the new icmp instruction.
getNewICmpValue(bool Sign,unsigned Code,Value * LHS,Value * RHS,InstCombiner::BuilderTy * Builder)71 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
72 InstCombiner::BuilderTy *Builder) {
73 ICmpInst::Predicate NewPred;
74 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
75 return NewConstant;
76 return Builder->CreateICmp(NewPred, LHS, RHS);
77 }
78
79 /// This is the complement of getFCmpCode, which turns an opcode and two
80 /// operands into either a FCmp instruction, or a true/false constant.
getFCmpValue(unsigned Code,Value * LHS,Value * RHS,InstCombiner::BuilderTy * Builder)81 static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS,
82 InstCombiner::BuilderTy *Builder) {
83 const auto Pred = static_cast<FCmpInst::Predicate>(Code);
84 assert(FCmpInst::FCMP_FALSE <= Pred && Pred <= FCmpInst::FCMP_TRUE &&
85 "Unexpected FCmp predicate!");
86 if (Pred == FCmpInst::FCMP_FALSE)
87 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
88 if (Pred == FCmpInst::FCMP_TRUE)
89 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
90 return Builder->CreateFCmp(Pred, LHS, RHS);
91 }
92
93 /// \brief Transform BITWISE_OP(BSWAP(A),BSWAP(B)) to BSWAP(BITWISE_OP(A, B))
94 /// \param I Binary operator to transform.
95 /// \return Pointer to node that must replace the original binary operator, or
96 /// null pointer if no transformation was made.
SimplifyBSwap(BinaryOperator & I)97 Value *InstCombiner::SimplifyBSwap(BinaryOperator &I) {
98 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
99
100 // Can't do vectors.
101 if (I.getType()->isVectorTy()) return nullptr;
102
103 // Can only do bitwise ops.
104 unsigned Op = I.getOpcode();
105 if (Op != Instruction::And && Op != Instruction::Or &&
106 Op != Instruction::Xor)
107 return nullptr;
108
109 Value *OldLHS = I.getOperand(0);
110 Value *OldRHS = I.getOperand(1);
111 ConstantInt *ConstLHS = dyn_cast<ConstantInt>(OldLHS);
112 ConstantInt *ConstRHS = dyn_cast<ConstantInt>(OldRHS);
113 IntrinsicInst *IntrLHS = dyn_cast<IntrinsicInst>(OldLHS);
114 IntrinsicInst *IntrRHS = dyn_cast<IntrinsicInst>(OldRHS);
115 bool IsBswapLHS = (IntrLHS && IntrLHS->getIntrinsicID() == Intrinsic::bswap);
116 bool IsBswapRHS = (IntrRHS && IntrRHS->getIntrinsicID() == Intrinsic::bswap);
117
118 if (!IsBswapLHS && !IsBswapRHS)
119 return nullptr;
120
121 if (!IsBswapLHS && !ConstLHS)
122 return nullptr;
123
124 if (!IsBswapRHS && !ConstRHS)
125 return nullptr;
126
127 /// OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
128 /// OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
129 Value *NewLHS = IsBswapLHS ? IntrLHS->getOperand(0) :
130 Builder->getInt(ConstLHS->getValue().byteSwap());
131
132 Value *NewRHS = IsBswapRHS ? IntrRHS->getOperand(0) :
133 Builder->getInt(ConstRHS->getValue().byteSwap());
134
135 Value *BinOp = nullptr;
136 if (Op == Instruction::And)
137 BinOp = Builder->CreateAnd(NewLHS, NewRHS);
138 else if (Op == Instruction::Or)
139 BinOp = Builder->CreateOr(NewLHS, NewRHS);
140 else //if (Op == Instruction::Xor)
141 BinOp = Builder->CreateXor(NewLHS, NewRHS);
142
143 Function *F = Intrinsic::getDeclaration(I.getModule(), Intrinsic::bswap, ITy);
144 return Builder->CreateCall(F, BinOp);
145 }
146
147 /// This handles expressions of the form ((val OP C1) & C2). Where
148 /// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
149 /// guaranteed to be a binary operator.
OptAndOp(Instruction * Op,ConstantInt * OpRHS,ConstantInt * AndRHS,BinaryOperator & TheAnd)150 Instruction *InstCombiner::OptAndOp(Instruction *Op,
151 ConstantInt *OpRHS,
152 ConstantInt *AndRHS,
153 BinaryOperator &TheAnd) {
154 Value *X = Op->getOperand(0);
155 Constant *Together = nullptr;
156 if (!Op->isShift())
157 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
158
159 switch (Op->getOpcode()) {
160 case Instruction::Xor:
161 if (Op->hasOneUse()) {
162 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
163 Value *And = Builder->CreateAnd(X, AndRHS);
164 And->takeName(Op);
165 return BinaryOperator::CreateXor(And, Together);
166 }
167 break;
168 case Instruction::Or:
169 if (Op->hasOneUse()){
170 if (Together != OpRHS) {
171 // (X | C1) & C2 --> (X | (C1&C2)) & C2
172 Value *Or = Builder->CreateOr(X, Together);
173 Or->takeName(Op);
174 return BinaryOperator::CreateAnd(Or, AndRHS);
175 }
176
177 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
178 if (TogetherCI && !TogetherCI->isZero()){
179 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
180 // NOTE: This reduces the number of bits set in the & mask, which
181 // can expose opportunities for store narrowing.
182 Together = ConstantExpr::getXor(AndRHS, Together);
183 Value *And = Builder->CreateAnd(X, Together);
184 And->takeName(Op);
185 return BinaryOperator::CreateOr(And, OpRHS);
186 }
187 }
188
189 break;
190 case Instruction::Add:
191 if (Op->hasOneUse()) {
192 // Adding a one to a single bit bit-field should be turned into an XOR
193 // of the bit. First thing to check is to see if this AND is with a
194 // single bit constant.
195 const APInt &AndRHSV = AndRHS->getValue();
196
197 // If there is only one bit set.
198 if (AndRHSV.isPowerOf2()) {
199 // Ok, at this point, we know that we are masking the result of the
200 // ADD down to exactly one bit. If the constant we are adding has
201 // no bits set below this bit, then we can eliminate the ADD.
202 const APInt& AddRHS = OpRHS->getValue();
203
204 // Check to see if any bits below the one bit set in AndRHSV are set.
205 if ((AddRHS & (AndRHSV-1)) == 0) {
206 // If not, the only thing that can effect the output of the AND is
207 // the bit specified by AndRHSV. If that bit is set, the effect of
208 // the XOR is to toggle the bit. If it is clear, then the ADD has
209 // no effect.
210 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
211 TheAnd.setOperand(0, X);
212 return &TheAnd;
213 } else {
214 // Pull the XOR out of the AND.
215 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
216 NewAnd->takeName(Op);
217 return BinaryOperator::CreateXor(NewAnd, AndRHS);
218 }
219 }
220 }
221 }
222 break;
223
224 case Instruction::Shl: {
225 // We know that the AND will not produce any of the bits shifted in, so if
226 // the anded constant includes them, clear them now!
227 //
228 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
229 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
230 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
231 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
232
233 if (CI->getValue() == ShlMask)
234 // Masking out bits that the shift already masks.
235 return replaceInstUsesWith(TheAnd, Op); // No need for the and.
236
237 if (CI != AndRHS) { // Reducing bits set in and.
238 TheAnd.setOperand(1, CI);
239 return &TheAnd;
240 }
241 break;
242 }
243 case Instruction::LShr: {
244 // We know that the AND will not produce any of the bits shifted in, so if
245 // the anded constant includes them, clear them now! This only applies to
246 // unsigned shifts, because a signed shr may bring in set bits!
247 //
248 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
249 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
250 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
251 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
252
253 if (CI->getValue() == ShrMask)
254 // Masking out bits that the shift already masks.
255 return replaceInstUsesWith(TheAnd, Op);
256
257 if (CI != AndRHS) {
258 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
259 return &TheAnd;
260 }
261 break;
262 }
263 case Instruction::AShr:
264 // Signed shr.
265 // See if this is shifting in some sign extension, then masking it out
266 // with an and.
267 if (Op->hasOneUse()) {
268 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
269 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
270 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
271 Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
272 if (C == AndRHS) { // Masking out bits shifted in.
273 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
274 // Make the argument unsigned.
275 Value *ShVal = Op->getOperand(0);
276 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
277 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
278 }
279 }
280 break;
281 }
282 return nullptr;
283 }
284
285 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
286 /// (V < Lo || V >= Hi). In practice, we emit the more efficient
287 /// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
288 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
289 /// insert new instructions.
InsertRangeTest(Value * V,Constant * Lo,Constant * Hi,bool isSigned,bool Inside)290 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
291 bool isSigned, bool Inside) {
292 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
293 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
294 "Lo is not <= Hi in range emission code!");
295
296 if (Inside) {
297 if (Lo == Hi) // Trivially false.
298 return Builder->getFalse();
299
300 // V >= Min && V < Hi --> V < Hi
301 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
302 ICmpInst::Predicate pred = (isSigned ?
303 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
304 return Builder->CreateICmp(pred, V, Hi);
305 }
306
307 // Emit V-Lo <u Hi-Lo
308 Constant *NegLo = ConstantExpr::getNeg(Lo);
309 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
310 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
311 return Builder->CreateICmpULT(Add, UpperBound);
312 }
313
314 if (Lo == Hi) // Trivially true.
315 return Builder->getTrue();
316
317 // V < Min || V >= Hi -> V > Hi-1
318 Hi = SubOne(cast<ConstantInt>(Hi));
319 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
320 ICmpInst::Predicate pred = (isSigned ?
321 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
322 return Builder->CreateICmp(pred, V, Hi);
323 }
324
325 // Emit V-Lo >u Hi-1-Lo
326 // Note that Hi has already had one subtracted from it, above.
327 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
328 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
329 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
330 return Builder->CreateICmpUGT(Add, LowerBound);
331 }
332
333 /// Returns true iff Val consists of one contiguous run of 1s with any number
334 /// of 0s on either side. The 1s are allowed to wrap from LSB to MSB,
335 /// so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
336 /// not, since all 1s are not contiguous.
isRunOfOnes(ConstantInt * Val,uint32_t & MB,uint32_t & ME)337 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
338 const APInt& V = Val->getValue();
339 uint32_t BitWidth = Val->getType()->getBitWidth();
340 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
341
342 // look for the first zero bit after the run of ones
343 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
344 // look for the first non-zero bit
345 ME = V.getActiveBits();
346 return true;
347 }
348
349 /// This is part of an expression (LHS +/- RHS) & Mask, where isSub determines
350 /// whether the operator is a sub. If we can fold one of the following xforms:
351 ///
352 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
353 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
354 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
355 ///
356 /// return (A +/- B).
357 ///
FoldLogicalPlusAnd(Value * LHS,Value * RHS,ConstantInt * Mask,bool isSub,Instruction & I)358 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
359 ConstantInt *Mask, bool isSub,
360 Instruction &I) {
361 Instruction *LHSI = dyn_cast<Instruction>(LHS);
362 if (!LHSI || LHSI->getNumOperands() != 2 ||
363 !isa<ConstantInt>(LHSI->getOperand(1))) return nullptr;
364
365 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
366
367 switch (LHSI->getOpcode()) {
368 default: return nullptr;
369 case Instruction::And:
370 if (ConstantExpr::getAnd(N, Mask) == Mask) {
371 // If the AndRHS is a power of two minus one (0+1+), this is simple.
372 if ((Mask->getValue().countLeadingZeros() +
373 Mask->getValue().countPopulation()) ==
374 Mask->getValue().getBitWidth())
375 break;
376
377 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
378 // part, we don't need any explicit masks to take them out of A. If that
379 // is all N is, ignore it.
380 uint32_t MB = 0, ME = 0;
381 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
382 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
383 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
384 if (MaskedValueIsZero(RHS, Mask, 0, &I))
385 break;
386 }
387 }
388 return nullptr;
389 case Instruction::Or:
390 case Instruction::Xor:
391 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
392 if ((Mask->getValue().countLeadingZeros() +
393 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
394 && ConstantExpr::getAnd(N, Mask)->isNullValue())
395 break;
396 return nullptr;
397 }
398
399 if (isSub)
400 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
401 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
402 }
403
404 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
405 /// One of A and B is considered the mask, the other the value. This is
406 /// described as the "AMask" or "BMask" part of the enum. If the enum
407 /// contains only "Mask", then both A and B can be considered masks.
408 /// If A is the mask, then it was proven, that (A & C) == C. This
409 /// is trivial if C == A, or C == 0. If both A and C are constants, this
410 /// proof is also easy.
411 /// For the following explanations we assume that A is the mask.
412 /// The part "AllOnes" declares, that the comparison is true only
413 /// if (A & B) == A, or all bits of A are set in B.
414 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
415 /// The part "AllZeroes" declares, that the comparison is true only
416 /// if (A & B) == 0, or all bits of A are cleared in B.
417 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
418 /// The part "Mixed" declares, that (A & B) == C and C might or might not
419 /// contain any number of one bits and zero bits.
420 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
421 /// The Part "Not" means, that in above descriptions "==" should be replaced
422 /// by "!=".
423 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
424 /// If the mask A contains a single bit, then the following is equivalent:
425 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
426 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
427 enum MaskedICmpType {
428 FoldMskICmp_AMask_AllOnes = 1,
429 FoldMskICmp_AMask_NotAllOnes = 2,
430 FoldMskICmp_BMask_AllOnes = 4,
431 FoldMskICmp_BMask_NotAllOnes = 8,
432 FoldMskICmp_Mask_AllZeroes = 16,
433 FoldMskICmp_Mask_NotAllZeroes = 32,
434 FoldMskICmp_AMask_Mixed = 64,
435 FoldMskICmp_AMask_NotMixed = 128,
436 FoldMskICmp_BMask_Mixed = 256,
437 FoldMskICmp_BMask_NotMixed = 512
438 };
439
440 /// Return the set of pattern classes (from MaskedICmpType)
441 /// that (icmp SCC (A & B), C) satisfies.
getTypeOfMaskedICmp(Value * A,Value * B,Value * C,ICmpInst::Predicate SCC)442 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
443 ICmpInst::Predicate SCC)
444 {
445 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
446 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
447 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
448 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
449 bool icmp_abit = (ACst && !ACst->isZero() &&
450 ACst->getValue().isPowerOf2());
451 bool icmp_bbit = (BCst && !BCst->isZero() &&
452 BCst->getValue().isPowerOf2());
453 unsigned result = 0;
454 if (CCst && CCst->isZero()) {
455 // if C is zero, then both A and B qualify as mask
456 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
457 FoldMskICmp_AMask_Mixed |
458 FoldMskICmp_BMask_Mixed)
459 : (FoldMskICmp_Mask_NotAllZeroes |
460 FoldMskICmp_AMask_NotMixed |
461 FoldMskICmp_BMask_NotMixed));
462 if (icmp_abit)
463 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
464 FoldMskICmp_AMask_NotMixed)
465 : (FoldMskICmp_AMask_AllOnes |
466 FoldMskICmp_AMask_Mixed));
467 if (icmp_bbit)
468 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
469 FoldMskICmp_BMask_NotMixed)
470 : (FoldMskICmp_BMask_AllOnes |
471 FoldMskICmp_BMask_Mixed));
472 return result;
473 }
474 if (A == C) {
475 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
476 FoldMskICmp_AMask_Mixed)
477 : (FoldMskICmp_AMask_NotAllOnes |
478 FoldMskICmp_AMask_NotMixed));
479 if (icmp_abit)
480 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
481 FoldMskICmp_AMask_NotMixed)
482 : (FoldMskICmp_Mask_AllZeroes |
483 FoldMskICmp_AMask_Mixed));
484 } else if (ACst && CCst &&
485 ConstantExpr::getAnd(ACst, CCst) == CCst) {
486 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
487 : FoldMskICmp_AMask_NotMixed);
488 }
489 if (B == C) {
490 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
491 FoldMskICmp_BMask_Mixed)
492 : (FoldMskICmp_BMask_NotAllOnes |
493 FoldMskICmp_BMask_NotMixed));
494 if (icmp_bbit)
495 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
496 FoldMskICmp_BMask_NotMixed)
497 : (FoldMskICmp_Mask_AllZeroes |
498 FoldMskICmp_BMask_Mixed));
499 } else if (BCst && CCst &&
500 ConstantExpr::getAnd(BCst, CCst) == CCst) {
501 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
502 : FoldMskICmp_BMask_NotMixed);
503 }
504 return result;
505 }
506
507 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
508 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
509 /// is adjacent to the corresponding normal flag (recording ==), this just
510 /// involves swapping those bits over.
conjugateICmpMask(unsigned Mask)511 static unsigned conjugateICmpMask(unsigned Mask) {
512 unsigned NewMask;
513 NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes |
514 FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed |
515 FoldMskICmp_BMask_Mixed))
516 << 1;
517
518 NewMask |=
519 (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes |
520 FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed |
521 FoldMskICmp_BMask_NotMixed))
522 >> 1;
523
524 return NewMask;
525 }
526
527 /// Decompose an icmp into the form ((X & Y) pred Z) if possible.
528 /// The returned predicate is either == or !=. Returns false if
529 /// decomposition fails.
decomposeBitTestICmp(const ICmpInst * I,ICmpInst::Predicate & Pred,Value * & X,Value * & Y,Value * & Z)530 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
531 Value *&X, Value *&Y, Value *&Z) {
532 ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1));
533 if (!C)
534 return false;
535
536 switch (I->getPredicate()) {
537 default:
538 return false;
539 case ICmpInst::ICMP_SLT:
540 // X < 0 is equivalent to (X & SignBit) != 0.
541 if (!C->isZero())
542 return false;
543 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
544 Pred = ICmpInst::ICMP_NE;
545 break;
546 case ICmpInst::ICMP_SGT:
547 // X > -1 is equivalent to (X & SignBit) == 0.
548 if (!C->isAllOnesValue())
549 return false;
550 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
551 Pred = ICmpInst::ICMP_EQ;
552 break;
553 case ICmpInst::ICMP_ULT:
554 // X <u 2^n is equivalent to (X & ~(2^n-1)) == 0.
555 if (!C->getValue().isPowerOf2())
556 return false;
557 Y = ConstantInt::get(I->getContext(), -C->getValue());
558 Pred = ICmpInst::ICMP_EQ;
559 break;
560 case ICmpInst::ICMP_UGT:
561 // X >u 2^n-1 is equivalent to (X & ~(2^n-1)) != 0.
562 if (!(C->getValue() + 1).isPowerOf2())
563 return false;
564 Y = ConstantInt::get(I->getContext(), ~C->getValue());
565 Pred = ICmpInst::ICMP_NE;
566 break;
567 }
568
569 X = I->getOperand(0);
570 Z = ConstantInt::getNullValue(C->getType());
571 return true;
572 }
573
574 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
575 /// Return the set of pattern classes (from MaskedICmpType)
576 /// that both LHS and RHS satisfy.
foldLogOpOfMaskedICmpsHelper(Value * & A,Value * & B,Value * & C,Value * & D,Value * & E,ICmpInst * LHS,ICmpInst * RHS,ICmpInst::Predicate & LHSCC,ICmpInst::Predicate & RHSCC)577 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
578 Value*& B, Value*& C,
579 Value*& D, Value*& E,
580 ICmpInst *LHS, ICmpInst *RHS,
581 ICmpInst::Predicate &LHSCC,
582 ICmpInst::Predicate &RHSCC) {
583 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
584 // vectors are not (yet?) supported
585 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
586
587 // Here comes the tricky part:
588 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
589 // and L11 & L12 == L21 & L22. The same goes for RHS.
590 // Now we must find those components L** and R**, that are equal, so
591 // that we can extract the parameters A, B, C, D, and E for the canonical
592 // above.
593 Value *L1 = LHS->getOperand(0);
594 Value *L2 = LHS->getOperand(1);
595 Value *L11,*L12,*L21,*L22;
596 // Check whether the icmp can be decomposed into a bit test.
597 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
598 L21 = L22 = L1 = nullptr;
599 } else {
600 // Look for ANDs in the LHS icmp.
601 if (!L1->getType()->isIntegerTy()) {
602 // You can icmp pointers, for example. They really aren't masks.
603 L11 = L12 = nullptr;
604 } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
605 // Any icmp can be viewed as being trivially masked; if it allows us to
606 // remove one, it's worth it.
607 L11 = L1;
608 L12 = Constant::getAllOnesValue(L1->getType());
609 }
610
611 if (!L2->getType()->isIntegerTy()) {
612 // You can icmp pointers, for example. They really aren't masks.
613 L21 = L22 = nullptr;
614 } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
615 L21 = L2;
616 L22 = Constant::getAllOnesValue(L2->getType());
617 }
618 }
619
620 // Bail if LHS was a icmp that can't be decomposed into an equality.
621 if (!ICmpInst::isEquality(LHSCC))
622 return 0;
623
624 Value *R1 = RHS->getOperand(0);
625 Value *R2 = RHS->getOperand(1);
626 Value *R11,*R12;
627 bool ok = false;
628 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
629 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
630 A = R11; D = R12;
631 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
632 A = R12; D = R11;
633 } else {
634 return 0;
635 }
636 E = R2; R1 = nullptr; ok = true;
637 } else if (R1->getType()->isIntegerTy()) {
638 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
639 // As before, model no mask as a trivial mask if it'll let us do an
640 // optimization.
641 R11 = R1;
642 R12 = Constant::getAllOnesValue(R1->getType());
643 }
644
645 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
646 A = R11; D = R12; E = R2; ok = true;
647 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
648 A = R12; D = R11; E = R2; ok = true;
649 }
650 }
651
652 // Bail if RHS was a icmp that can't be decomposed into an equality.
653 if (!ICmpInst::isEquality(RHSCC))
654 return 0;
655
656 // Look for ANDs on the right side of the RHS icmp.
657 if (!ok && R2->getType()->isIntegerTy()) {
658 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
659 R11 = R2;
660 R12 = Constant::getAllOnesValue(R2->getType());
661 }
662
663 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
664 A = R11; D = R12; E = R1; ok = true;
665 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
666 A = R12; D = R11; E = R1; ok = true;
667 } else {
668 return 0;
669 }
670 }
671 if (!ok)
672 return 0;
673
674 if (L11 == A) {
675 B = L12; C = L2;
676 } else if (L12 == A) {
677 B = L11; C = L2;
678 } else if (L21 == A) {
679 B = L22; C = L1;
680 } else if (L22 == A) {
681 B = L21; C = L1;
682 }
683
684 unsigned LeftType = getTypeOfMaskedICmp(A, B, C, LHSCC);
685 unsigned RightType = getTypeOfMaskedICmp(A, D, E, RHSCC);
686 return LeftType & RightType;
687 }
688
689 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
690 /// into a single (icmp(A & X) ==/!= Y).
foldLogOpOfMaskedICmps(ICmpInst * LHS,ICmpInst * RHS,bool IsAnd,llvm::InstCombiner::BuilderTy * Builder)691 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
692 llvm::InstCombiner::BuilderTy *Builder) {
693 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
694 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
695 unsigned Mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
696 LHSCC, RHSCC);
697 if (Mask == 0) return nullptr;
698 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
699 "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
700
701 // In full generality:
702 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
703 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
704 //
705 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
706 // equivalent to (icmp (A & X) !Op Y).
707 //
708 // Therefore, we can pretend for the rest of this function that we're dealing
709 // with the conjunction, provided we flip the sense of any comparisons (both
710 // input and output).
711
712 // In most cases we're going to produce an EQ for the "&&" case.
713 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
714 if (!IsAnd) {
715 // Convert the masking analysis into its equivalent with negated
716 // comparisons.
717 Mask = conjugateICmpMask(Mask);
718 }
719
720 if (Mask & FoldMskICmp_Mask_AllZeroes) {
721 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
722 // -> (icmp eq (A & (B|D)), 0)
723 Value *NewOr = Builder->CreateOr(B, D);
724 Value *NewAnd = Builder->CreateAnd(A, NewOr);
725 // We can't use C as zero because we might actually handle
726 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
727 // with B and D, having a single bit set.
728 Value *Zero = Constant::getNullValue(A->getType());
729 return Builder->CreateICmp(NewCC, NewAnd, Zero);
730 }
731 if (Mask & FoldMskICmp_BMask_AllOnes) {
732 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
733 // -> (icmp eq (A & (B|D)), (B|D))
734 Value *NewOr = Builder->CreateOr(B, D);
735 Value *NewAnd = Builder->CreateAnd(A, NewOr);
736 return Builder->CreateICmp(NewCC, NewAnd, NewOr);
737 }
738 if (Mask & FoldMskICmp_AMask_AllOnes) {
739 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
740 // -> (icmp eq (A & (B&D)), A)
741 Value *NewAnd1 = Builder->CreateAnd(B, D);
742 Value *NewAnd2 = Builder->CreateAnd(A, NewAnd1);
743 return Builder->CreateICmp(NewCC, NewAnd2, A);
744 }
745
746 // Remaining cases assume at least that B and D are constant, and depend on
747 // their actual values. This isn't strictly necessary, just a "handle the
748 // easy cases for now" decision.
749 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
750 if (!BCst) return nullptr;
751 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
752 if (!DCst) return nullptr;
753
754 if (Mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) {
755 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
756 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
757 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
758 // Only valid if one of the masks is a superset of the other (check "B&D" is
759 // the same as either B or D).
760 APInt NewMask = BCst->getValue() & DCst->getValue();
761
762 if (NewMask == BCst->getValue())
763 return LHS;
764 else if (NewMask == DCst->getValue())
765 return RHS;
766 }
767 if (Mask & FoldMskICmp_AMask_NotAllOnes) {
768 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
769 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
770 // Only valid if one of the masks is a superset of the other (check "B|D" is
771 // the same as either B or D).
772 APInt NewMask = BCst->getValue() | DCst->getValue();
773
774 if (NewMask == BCst->getValue())
775 return LHS;
776 else if (NewMask == DCst->getValue())
777 return RHS;
778 }
779 if (Mask & FoldMskICmp_BMask_Mixed) {
780 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
781 // We already know that B & C == C && D & E == E.
782 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
783 // C and E, which are shared by both the mask B and the mask D, don't
784 // contradict, then we can transform to
785 // -> (icmp eq (A & (B|D)), (C|E))
786 // Currently, we only handle the case of B, C, D, and E being constant.
787 // We can't simply use C and E because we might actually handle
788 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
789 // with B and D, having a single bit set.
790 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
791 if (!CCst) return nullptr;
792 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
793 if (!ECst) return nullptr;
794 if (LHSCC != NewCC)
795 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
796 if (RHSCC != NewCC)
797 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
798 // If there is a conflict, we should actually return a false for the
799 // whole construct.
800 if (((BCst->getValue() & DCst->getValue()) &
801 (CCst->getValue() ^ ECst->getValue())) != 0)
802 return ConstantInt::get(LHS->getType(), !IsAnd);
803 Value *NewOr1 = Builder->CreateOr(B, D);
804 Value *NewOr2 = ConstantExpr::getOr(CCst, ECst);
805 Value *NewAnd = Builder->CreateAnd(A, NewOr1);
806 return Builder->CreateICmp(NewCC, NewAnd, NewOr2);
807 }
808 return nullptr;
809 }
810
811 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
812 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
813 /// If \p Inverted is true then the check is for the inverted range, e.g.
814 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
simplifyRangeCheck(ICmpInst * Cmp0,ICmpInst * Cmp1,bool Inverted)815 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
816 bool Inverted) {
817 // Check the lower range comparison, e.g. x >= 0
818 // InstCombine already ensured that if there is a constant it's on the RHS.
819 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
820 if (!RangeStart)
821 return nullptr;
822
823 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
824 Cmp0->getPredicate());
825
826 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
827 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
828 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
829 return nullptr;
830
831 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
832 Cmp1->getPredicate());
833
834 Value *Input = Cmp0->getOperand(0);
835 Value *RangeEnd;
836 if (Cmp1->getOperand(0) == Input) {
837 // For the upper range compare we have: icmp x, n
838 RangeEnd = Cmp1->getOperand(1);
839 } else if (Cmp1->getOperand(1) == Input) {
840 // For the upper range compare we have: icmp n, x
841 RangeEnd = Cmp1->getOperand(0);
842 Pred1 = ICmpInst::getSwappedPredicate(Pred1);
843 } else {
844 return nullptr;
845 }
846
847 // Check the upper range comparison, e.g. x < n
848 ICmpInst::Predicate NewPred;
849 switch (Pred1) {
850 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
851 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
852 default: return nullptr;
853 }
854
855 // This simplification is only valid if the upper range is not negative.
856 bool IsNegative, IsNotNegative;
857 ComputeSignBit(RangeEnd, IsNotNegative, IsNegative, /*Depth=*/0, Cmp1);
858 if (!IsNotNegative)
859 return nullptr;
860
861 if (Inverted)
862 NewPred = ICmpInst::getInversePredicate(NewPred);
863
864 return Builder->CreateICmp(NewPred, Input, RangeEnd);
865 }
866
867 /// Fold (icmp)&(icmp) if possible.
FoldAndOfICmps(ICmpInst * LHS,ICmpInst * RHS)868 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
869 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
870
871 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
872 if (PredicatesFoldable(LHSCC, RHSCC)) {
873 if (LHS->getOperand(0) == RHS->getOperand(1) &&
874 LHS->getOperand(1) == RHS->getOperand(0))
875 LHS->swapOperands();
876 if (LHS->getOperand(0) == RHS->getOperand(0) &&
877 LHS->getOperand(1) == RHS->getOperand(1)) {
878 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
879 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
880 bool isSigned = LHS->isSigned() || RHS->isSigned();
881 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
882 }
883 }
884
885 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
886 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
887 return V;
888
889 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
890 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
891 return V;
892
893 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
894 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
895 return V;
896
897 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
898 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
899 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
900 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
901 if (!LHSCst || !RHSCst) return nullptr;
902
903 if (LHSCst == RHSCst && LHSCC == RHSCC) {
904 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
905 // where C is a power of 2 or
906 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
907 if ((LHSCC == ICmpInst::ICMP_ULT && LHSCst->getValue().isPowerOf2()) ||
908 (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero())) {
909 Value *NewOr = Builder->CreateOr(Val, Val2);
910 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
911 }
912 }
913
914 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
915 // where CMAX is the all ones value for the truncated type,
916 // iff the lower bits of C2 and CA are zero.
917 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
918 LHS->hasOneUse() && RHS->hasOneUse()) {
919 Value *V;
920 ConstantInt *AndCst, *SmallCst = nullptr, *BigCst = nullptr;
921
922 // (trunc x) == C1 & (and x, CA) == C2
923 // (and x, CA) == C2 & (trunc x) == C1
924 if (match(Val2, m_Trunc(m_Value(V))) &&
925 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
926 SmallCst = RHSCst;
927 BigCst = LHSCst;
928 } else if (match(Val, m_Trunc(m_Value(V))) &&
929 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
930 SmallCst = LHSCst;
931 BigCst = RHSCst;
932 }
933
934 if (SmallCst && BigCst) {
935 unsigned BigBitSize = BigCst->getType()->getBitWidth();
936 unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
937
938 // Check that the low bits are zero.
939 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
940 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
941 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
942 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
943 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
944 return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
945 }
946 }
947 }
948
949 // From here on, we only handle:
950 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
951 if (Val != Val2) return nullptr;
952
953 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
954 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
955 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
956 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
957 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
958 return nullptr;
959
960 // We can't fold (ugt x, C) & (sgt x, C2).
961 if (!PredicatesFoldable(LHSCC, RHSCC))
962 return nullptr;
963
964 // Ensure that the larger constant is on the RHS.
965 bool ShouldSwap;
966 if (CmpInst::isSigned(LHSCC) ||
967 (ICmpInst::isEquality(LHSCC) &&
968 CmpInst::isSigned(RHSCC)))
969 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
970 else
971 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
972
973 if (ShouldSwap) {
974 std::swap(LHS, RHS);
975 std::swap(LHSCst, RHSCst);
976 std::swap(LHSCC, RHSCC);
977 }
978
979 // At this point, we know we have two icmp instructions
980 // comparing a value against two constants and and'ing the result
981 // together. Because of the above check, we know that we only have
982 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
983 // (from the icmp folding check above), that the two constants
984 // are not equal and that the larger constant is on the RHS
985 assert(LHSCst != RHSCst && "Compares not folded above?");
986
987 switch (LHSCC) {
988 default: llvm_unreachable("Unknown integer condition code!");
989 case ICmpInst::ICMP_EQ:
990 switch (RHSCC) {
991 default: llvm_unreachable("Unknown integer condition code!");
992 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
993 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
994 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
995 return LHS;
996 }
997 case ICmpInst::ICMP_NE:
998 switch (RHSCC) {
999 default: llvm_unreachable("Unknown integer condition code!");
1000 case ICmpInst::ICMP_ULT:
1001 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
1002 return Builder->CreateICmpULT(Val, LHSCst);
1003 if (LHSCst->isNullValue()) // (X != 0 & X u< 14) -> X-1 u< 13
1004 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
1005 break; // (X != 13 & X u< 15) -> no change
1006 case ICmpInst::ICMP_SLT:
1007 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
1008 return Builder->CreateICmpSLT(Val, LHSCst);
1009 break; // (X != 13 & X s< 15) -> no change
1010 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
1011 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
1012 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
1013 return RHS;
1014 case ICmpInst::ICMP_NE:
1015 // Special case to get the ordering right when the values wrap around
1016 // zero.
1017 if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue())
1018 std::swap(LHSCst, RHSCst);
1019 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
1020 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1021 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1022 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1),
1023 Val->getName()+".cmp");
1024 }
1025 break; // (X != 13 & X != 15) -> no change
1026 }
1027 break;
1028 case ICmpInst::ICMP_ULT:
1029 switch (RHSCC) {
1030 default: llvm_unreachable("Unknown integer condition code!");
1031 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
1032 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
1033 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1034 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
1035 break;
1036 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
1037 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
1038 return LHS;
1039 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
1040 break;
1041 }
1042 break;
1043 case ICmpInst::ICMP_SLT:
1044 switch (RHSCC) {
1045 default: llvm_unreachable("Unknown integer condition code!");
1046 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
1047 break;
1048 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
1049 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
1050 return LHS;
1051 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
1052 break;
1053 }
1054 break;
1055 case ICmpInst::ICMP_UGT:
1056 switch (RHSCC) {
1057 default: llvm_unreachable("Unknown integer condition code!");
1058 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
1059 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
1060 return RHS;
1061 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
1062 break;
1063 case ICmpInst::ICMP_NE:
1064 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
1065 return Builder->CreateICmp(LHSCC, Val, RHSCst);
1066 break; // (X u> 13 & X != 15) -> no change
1067 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
1068 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
1069 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
1070 break;
1071 }
1072 break;
1073 case ICmpInst::ICMP_SGT:
1074 switch (RHSCC) {
1075 default: llvm_unreachable("Unknown integer condition code!");
1076 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
1077 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
1078 return RHS;
1079 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
1080 break;
1081 case ICmpInst::ICMP_NE:
1082 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
1083 return Builder->CreateICmp(LHSCC, Val, RHSCst);
1084 break; // (X s> 13 & X != 15) -> no change
1085 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
1086 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
1087 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
1088 break;
1089 }
1090 break;
1091 }
1092
1093 return nullptr;
1094 }
1095
1096 /// Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of instcombine, this returns
1097 /// a Value which should already be inserted into the function.
FoldAndOfFCmps(FCmpInst * LHS,FCmpInst * RHS)1098 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1099 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1100 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1101 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1102
1103 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1104 // Swap RHS operands to match LHS.
1105 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1106 std::swap(Op1LHS, Op1RHS);
1107 }
1108
1109 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1110 // Suppose the relation between x and y is R, where R is one of
1111 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for
1112 // testing the desired relations.
1113 //
1114 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1115 // bool(R & CC0) && bool(R & CC1)
1116 // = bool((R & CC0) & (R & CC1))
1117 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency
1118 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS)
1119 return getFCmpValue(getFCmpCode(Op0CC) & getFCmpCode(Op1CC), Op0LHS, Op0RHS,
1120 Builder);
1121
1122 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
1123 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
1124 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
1125 return nullptr;
1126
1127 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
1128 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1129 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1130 // If either of the constants are nans, then the whole thing returns
1131 // false.
1132 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1133 return Builder->getFalse();
1134 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1135 }
1136
1137 // Handle vector zeros. This occurs because the canonical form of
1138 // "fcmp ord x,x" is "fcmp ord x, 0".
1139 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1140 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1141 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1142 return nullptr;
1143 }
1144
1145 return nullptr;
1146 }
1147
1148 /// Match De Morgan's Laws:
1149 /// (~A & ~B) == (~(A | B))
1150 /// (~A | ~B) == (~(A & B))
matchDeMorgansLaws(BinaryOperator & I,InstCombiner::BuilderTy * Builder)1151 static Instruction *matchDeMorgansLaws(BinaryOperator &I,
1152 InstCombiner::BuilderTy *Builder) {
1153 auto Opcode = I.getOpcode();
1154 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1155 "Trying to match De Morgan's Laws with something other than and/or");
1156 // Flip the logic operation.
1157 if (Opcode == Instruction::And)
1158 Opcode = Instruction::Or;
1159 else
1160 Opcode = Instruction::And;
1161
1162 Value *Op0 = I.getOperand(0);
1163 Value *Op1 = I.getOperand(1);
1164 // TODO: Use pattern matchers instead of dyn_cast.
1165 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1166 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1167 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1168 Value *LogicOp = Builder->CreateBinOp(Opcode, Op0NotVal, Op1NotVal,
1169 I.getName() + ".demorgan");
1170 return BinaryOperator::CreateNot(LogicOp);
1171 }
1172
1173 // De Morgan's Law in disguise:
1174 // (zext(bool A) ^ 1) & (zext(bool B) ^ 1) -> zext(~(A | B))
1175 // (zext(bool A) ^ 1) | (zext(bool B) ^ 1) -> zext(~(A & B))
1176 Value *A = nullptr;
1177 Value *B = nullptr;
1178 ConstantInt *C1 = nullptr;
1179 if (match(Op0, m_OneUse(m_Xor(m_ZExt(m_Value(A)), m_ConstantInt(C1)))) &&
1180 match(Op1, m_OneUse(m_Xor(m_ZExt(m_Value(B)), m_Specific(C1))))) {
1181 // TODO: This check could be loosened to handle different type sizes.
1182 // Alternatively, we could fix the definition of m_Not to recognize a not
1183 // operation hidden by a zext?
1184 if (A->getType()->isIntegerTy(1) && B->getType()->isIntegerTy(1) &&
1185 C1->isOne()) {
1186 Value *LogicOp = Builder->CreateBinOp(Opcode, A, B,
1187 I.getName() + ".demorgan");
1188 Value *Not = Builder->CreateNot(LogicOp);
1189 return CastInst::CreateZExtOrBitCast(Not, I.getType());
1190 }
1191 }
1192
1193 return nullptr;
1194 }
1195
foldCastedBitwiseLogic(BinaryOperator & I)1196 Instruction *InstCombiner::foldCastedBitwiseLogic(BinaryOperator &I) {
1197 auto LogicOpc = I.getOpcode();
1198 assert((LogicOpc == Instruction::And || LogicOpc == Instruction::Or ||
1199 LogicOpc == Instruction::Xor) &&
1200 "Unexpected opcode for bitwise logic folding");
1201
1202 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1203 CastInst *Cast0 = dyn_cast<CastInst>(Op0);
1204 if (!Cast0)
1205 return nullptr;
1206
1207 // This must be a cast from an integer or integer vector source type to allow
1208 // transformation of the logic operation to the source type.
1209 Type *DestTy = I.getType();
1210 Type *SrcTy = Cast0->getSrcTy();
1211 if (!SrcTy->isIntOrIntVectorTy())
1212 return nullptr;
1213
1214 // If one operand is a bitcast and the other is a constant, move the logic
1215 // operation ahead of the bitcast. That is, do the logic operation in the
1216 // original type. This can eliminate useless bitcasts and allow normal
1217 // combines that would otherwise be impeded by the bitcast. Canonicalization
1218 // ensures that if there is a constant operand, it will be the second operand.
1219 Value *BC = nullptr;
1220 Constant *C = nullptr;
1221 if ((match(Op0, m_BitCast(m_Value(BC))) && match(Op1, m_Constant(C)))) {
1222 Value *NewConstant = ConstantExpr::getBitCast(C, SrcTy);
1223 Value *NewOp = Builder->CreateBinOp(LogicOpc, BC, NewConstant, I.getName());
1224 return CastInst::CreateBitOrPointerCast(NewOp, DestTy);
1225 }
1226
1227 CastInst *Cast1 = dyn_cast<CastInst>(Op1);
1228 if (!Cast1)
1229 return nullptr;
1230
1231 // Both operands of the logic operation are casts. The casts must be of the
1232 // same type for reduction.
1233 auto CastOpcode = Cast0->getOpcode();
1234 if (CastOpcode != Cast1->getOpcode() || SrcTy != Cast1->getSrcTy())
1235 return nullptr;
1236
1237 Value *Cast0Src = Cast0->getOperand(0);
1238 Value *Cast1Src = Cast1->getOperand(0);
1239
1240 // fold (logic (cast A), (cast B)) -> (cast (logic A, B))
1241
1242 // Only do this if the casts both really cause code to be generated.
1243 if ((!isa<ICmpInst>(Cast0Src) || !isa<ICmpInst>(Cast1Src)) &&
1244 ShouldOptimizeCast(CastOpcode, Cast0Src, DestTy) &&
1245 ShouldOptimizeCast(CastOpcode, Cast1Src, DestTy)) {
1246 Value *NewOp = Builder->CreateBinOp(LogicOpc, Cast0Src, Cast1Src,
1247 I.getName());
1248 return CastInst::Create(CastOpcode, NewOp, DestTy);
1249 }
1250
1251 // For now, only 'and'/'or' have optimizations after this.
1252 if (LogicOpc == Instruction::Xor)
1253 return nullptr;
1254
1255 // If this is logic(cast(icmp), cast(icmp)), try to fold this even if the
1256 // cast is otherwise not optimizable. This happens for vector sexts.
1257 ICmpInst *ICmp0 = dyn_cast<ICmpInst>(Cast0Src);
1258 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(Cast1Src);
1259 if (ICmp0 && ICmp1) {
1260 Value *Res = LogicOpc == Instruction::And ? FoldAndOfICmps(ICmp0, ICmp1)
1261 : FoldOrOfICmps(ICmp0, ICmp1, &I);
1262 if (Res)
1263 return CastInst::Create(CastOpcode, Res, DestTy);
1264 return nullptr;
1265 }
1266
1267 // If this is logic(cast(fcmp), cast(fcmp)), try to fold this even if the
1268 // cast is otherwise not optimizable. This happens for vector sexts.
1269 FCmpInst *FCmp0 = dyn_cast<FCmpInst>(Cast0Src);
1270 FCmpInst *FCmp1 = dyn_cast<FCmpInst>(Cast1Src);
1271 if (FCmp0 && FCmp1) {
1272 Value *Res = LogicOpc == Instruction::And ? FoldAndOfFCmps(FCmp0, FCmp1)
1273 : FoldOrOfFCmps(FCmp0, FCmp1);
1274 if (Res)
1275 return CastInst::Create(CastOpcode, Res, DestTy);
1276 return nullptr;
1277 }
1278
1279 return nullptr;
1280 }
1281
foldBoolSextMaskToSelect(BinaryOperator & I)1282 static Instruction *foldBoolSextMaskToSelect(BinaryOperator &I) {
1283 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1284
1285 // Canonicalize SExt or Not to the LHS
1286 if (match(Op1, m_SExt(m_Value())) || match(Op1, m_Not(m_Value()))) {
1287 std::swap(Op0, Op1);
1288 }
1289
1290 // Fold (and (sext bool to A), B) --> (select bool, B, 0)
1291 Value *X = nullptr;
1292 if (match(Op0, m_SExt(m_Value(X))) &&
1293 X->getType()->getScalarType()->isIntegerTy(1)) {
1294 Value *Zero = Constant::getNullValue(Op1->getType());
1295 return SelectInst::Create(X, Op1, Zero);
1296 }
1297
1298 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
1299 if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
1300 X->getType()->getScalarType()->isIntegerTy(1)) {
1301 Value *Zero = Constant::getNullValue(Op0->getType());
1302 return SelectInst::Create(X, Zero, Op1);
1303 }
1304
1305 return nullptr;
1306 }
1307
visitAnd(BinaryOperator & I)1308 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1309 bool Changed = SimplifyAssociativeOrCommutative(I);
1310 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1311
1312 if (Value *V = SimplifyVectorOp(I))
1313 return replaceInstUsesWith(I, V);
1314
1315 if (Value *V = SimplifyAndInst(Op0, Op1, DL, TLI, DT, AC))
1316 return replaceInstUsesWith(I, V);
1317
1318 // (A|B)&(A|C) -> A|(B&C) etc
1319 if (Value *V = SimplifyUsingDistributiveLaws(I))
1320 return replaceInstUsesWith(I, V);
1321
1322 // See if we can simplify any instructions used by the instruction whose sole
1323 // purpose is to compute bits we don't care about.
1324 if (SimplifyDemandedInstructionBits(I))
1325 return &I;
1326
1327 if (Value *V = SimplifyBSwap(I))
1328 return replaceInstUsesWith(I, V);
1329
1330 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1331 const APInt &AndRHSMask = AndRHS->getValue();
1332
1333 // Optimize a variety of ((val OP C1) & C2) combinations...
1334 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1335 Value *Op0LHS = Op0I->getOperand(0);
1336 Value *Op0RHS = Op0I->getOperand(1);
1337 switch (Op0I->getOpcode()) {
1338 default: break;
1339 case Instruction::Xor:
1340 case Instruction::Or: {
1341 // If the mask is only needed on one incoming arm, push it up.
1342 if (!Op0I->hasOneUse()) break;
1343
1344 APInt NotAndRHS(~AndRHSMask);
1345 if (MaskedValueIsZero(Op0LHS, NotAndRHS, 0, &I)) {
1346 // Not masking anything out for the LHS, move to RHS.
1347 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1348 Op0RHS->getName()+".masked");
1349 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1350 }
1351 if (!isa<Constant>(Op0RHS) &&
1352 MaskedValueIsZero(Op0RHS, NotAndRHS, 0, &I)) {
1353 // Not masking anything out for the RHS, move to LHS.
1354 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1355 Op0LHS->getName()+".masked");
1356 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1357 }
1358
1359 break;
1360 }
1361 case Instruction::Add:
1362 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1363 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1364 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1365 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1366 return BinaryOperator::CreateAnd(V, AndRHS);
1367 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1368 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1369 break;
1370
1371 case Instruction::Sub:
1372 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1373 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1374 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1375 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1376 return BinaryOperator::CreateAnd(V, AndRHS);
1377
1378 // -x & 1 -> x & 1
1379 if (AndRHSMask == 1 && match(Op0LHS, m_Zero()))
1380 return BinaryOperator::CreateAnd(Op0RHS, AndRHS);
1381
1382 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1383 // has 1's for all bits that the subtraction with A might affect.
1384 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1385 uint32_t BitWidth = AndRHSMask.getBitWidth();
1386 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1387 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1388
1389 if (MaskedValueIsZero(Op0LHS, Mask, 0, &I)) {
1390 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1391 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1392 }
1393 }
1394 break;
1395
1396 case Instruction::Shl:
1397 case Instruction::LShr:
1398 // (1 << x) & 1 --> zext(x == 0)
1399 // (1 >> x) & 1 --> zext(x == 0)
1400 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1401 Value *NewICmp =
1402 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1403 return new ZExtInst(NewICmp, I.getType());
1404 }
1405 break;
1406 }
1407
1408 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1409 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1410 return Res;
1411 }
1412
1413 // If this is an integer truncation, and if the source is an 'and' with
1414 // immediate, transform it. This frequently occurs for bitfield accesses.
1415 {
1416 Value *X = nullptr; ConstantInt *YC = nullptr;
1417 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1418 // Change: and (trunc (and X, YC) to T), C2
1419 // into : and (trunc X to T), trunc(YC) & C2
1420 // This will fold the two constants together, which may allow
1421 // other simplifications.
1422 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1423 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1424 C3 = ConstantExpr::getAnd(C3, AndRHS);
1425 return BinaryOperator::CreateAnd(NewCast, C3);
1426 }
1427 }
1428
1429 // Try to fold constant and into select arguments.
1430 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1431 if (Instruction *R = FoldOpIntoSelect(I, SI))
1432 return R;
1433 if (isa<PHINode>(Op0))
1434 if (Instruction *NV = FoldOpIntoPhi(I))
1435 return NV;
1436 }
1437
1438 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
1439 return DeMorgan;
1440
1441 {
1442 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
1443 // (A|B) & ~(A&B) -> A^B
1444 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1445 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1446 ((A == C && B == D) || (A == D && B == C)))
1447 return BinaryOperator::CreateXor(A, B);
1448
1449 // ~(A&B) & (A|B) -> A^B
1450 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1451 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1452 ((A == C && B == D) || (A == D && B == C)))
1453 return BinaryOperator::CreateXor(A, B);
1454
1455 // A&(A^B) => A & ~B
1456 {
1457 Value *tmpOp0 = Op0;
1458 Value *tmpOp1 = Op1;
1459 if (match(Op0, m_OneUse(m_Xor(m_Value(A), m_Value(B))))) {
1460 if (A == Op1 || B == Op1 ) {
1461 tmpOp1 = Op0;
1462 tmpOp0 = Op1;
1463 // Simplify below
1464 }
1465 }
1466
1467 if (match(tmpOp1, m_OneUse(m_Xor(m_Value(A), m_Value(B))))) {
1468 if (B == tmpOp0) {
1469 std::swap(A, B);
1470 }
1471 // Notice that the pattern (A&(~B)) is actually (A&(-1^B)), so if
1472 // A is originally -1 (or a vector of -1 and undefs), then we enter
1473 // an endless loop. By checking that A is non-constant we ensure that
1474 // we will never get to the loop.
1475 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1476 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1477 }
1478 }
1479
1480 // (A&((~A)|B)) -> A&B
1481 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1482 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1483 return BinaryOperator::CreateAnd(A, Op1);
1484 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1485 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1486 return BinaryOperator::CreateAnd(A, Op0);
1487
1488 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1489 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1490 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1491 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
1492 return BinaryOperator::CreateAnd(Op0, Builder->CreateNot(C));
1493
1494 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1495 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1496 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1497 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
1498 return BinaryOperator::CreateAnd(Op1, Builder->CreateNot(C));
1499
1500 // (A | B) & ((~A) ^ B) -> (A & B)
1501 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1502 match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
1503 return BinaryOperator::CreateAnd(A, B);
1504
1505 // ((~A) ^ B) & (A | B) -> (A & B)
1506 if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1507 match(Op1, m_Or(m_Specific(A), m_Specific(B))))
1508 return BinaryOperator::CreateAnd(A, B);
1509 }
1510
1511 {
1512 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1513 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1514 if (LHS && RHS)
1515 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1516 return replaceInstUsesWith(I, Res);
1517
1518 // TODO: Make this recursive; it's a little tricky because an arbitrary
1519 // number of 'and' instructions might have to be created.
1520 Value *X, *Y;
1521 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1522 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1523 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1524 return replaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1525 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1526 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1527 return replaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1528 }
1529 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1530 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1531 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1532 return replaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1533 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1534 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1535 return replaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1536 }
1537 }
1538
1539 // If and'ing two fcmp, try combine them into one.
1540 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1541 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1542 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1543 return replaceInstUsesWith(I, Res);
1544
1545 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))
1546 return CastedAnd;
1547
1548 if (Instruction *Select = foldBoolSextMaskToSelect(I))
1549 return Select;
1550
1551 return Changed ? &I : nullptr;
1552 }
1553
1554 /// Given an OR instruction, check to see if this is a bswap idiom. If so,
1555 /// insert the new intrinsic and return it.
MatchBSwap(BinaryOperator & I)1556 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1557 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1558
1559 // Look through zero extends.
1560 if (Instruction *Ext = dyn_cast<ZExtInst>(Op0))
1561 Op0 = Ext->getOperand(0);
1562
1563 if (Instruction *Ext = dyn_cast<ZExtInst>(Op1))
1564 Op1 = Ext->getOperand(0);
1565
1566 // (A | B) | C and A | (B | C) -> bswap if possible.
1567 bool OrOfOrs = match(Op0, m_Or(m_Value(), m_Value())) ||
1568 match(Op1, m_Or(m_Value(), m_Value()));
1569
1570 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1571 bool OrOfShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1572 match(Op1, m_LogicalShift(m_Value(), m_Value()));
1573
1574 // (A & B) | (C & D) -> bswap if possible.
1575 bool OrOfAnds = match(Op0, m_And(m_Value(), m_Value())) &&
1576 match(Op1, m_And(m_Value(), m_Value()));
1577
1578 if (!OrOfOrs && !OrOfShifts && !OrOfAnds)
1579 return nullptr;
1580
1581 SmallVector<Instruction*, 4> Insts;
1582 if (!recognizeBSwapOrBitReverseIdiom(&I, true, false, Insts))
1583 return nullptr;
1584 Instruction *LastInst = Insts.pop_back_val();
1585 LastInst->removeFromParent();
1586
1587 for (auto *Inst : Insts)
1588 Worklist.Add(Inst);
1589 return LastInst;
1590 }
1591
1592 /// If all elements of two constant vectors are 0/-1 and inverses, return true.
areInverseVectorBitmasks(Constant * C1,Constant * C2)1593 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {
1594 unsigned NumElts = C1->getType()->getVectorNumElements();
1595 for (unsigned i = 0; i != NumElts; ++i) {
1596 Constant *EltC1 = C1->getAggregateElement(i);
1597 Constant *EltC2 = C2->getAggregateElement(i);
1598 if (!EltC1 || !EltC2)
1599 return false;
1600
1601 // One element must be all ones, and the other must be all zeros.
1602 // FIXME: Allow undef elements.
1603 if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) ||
1604 (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes()))))
1605 return false;
1606 }
1607 return true;
1608 }
1609
1610 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or
1611 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of
1612 /// B, it can be used as the condition operand of a select instruction.
getSelectCondition(Value * A,Value * B,InstCombiner::BuilderTy & Builder)1613 static Value *getSelectCondition(Value *A, Value *B,
1614 InstCombiner::BuilderTy &Builder) {
1615 // If these are scalars or vectors of i1, A can be used directly.
1616 Type *Ty = A->getType();
1617 if (match(A, m_Not(m_Specific(B))) && Ty->getScalarType()->isIntegerTy(1))
1618 return A;
1619
1620 // If A and B are sign-extended, look through the sexts to find the booleans.
1621 Value *Cond;
1622 if (match(A, m_SExt(m_Value(Cond))) &&
1623 Cond->getType()->getScalarType()->isIntegerTy(1) &&
1624 match(B, m_CombineOr(m_Not(m_SExt(m_Specific(Cond))),
1625 m_SExt(m_Not(m_Specific(Cond))))))
1626 return Cond;
1627
1628 // All scalar (and most vector) possibilities should be handled now.
1629 // Try more matches that only apply to non-splat constant vectors.
1630 if (!Ty->isVectorTy())
1631 return nullptr;
1632
1633 // If both operands are constants, see if the constants are inverse bitmasks.
1634 Constant *AC, *BC;
1635 if (match(A, m_Constant(AC)) && match(B, m_Constant(BC)) &&
1636 areInverseVectorBitmasks(AC, BC))
1637 return ConstantExpr::getTrunc(AC, CmpInst::makeCmpResultType(Ty));
1638
1639 // If both operands are xor'd with constants using the same sexted boolean
1640 // operand, see if the constants are inverse bitmasks.
1641 if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AC)))) &&
1642 match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BC)))) &&
1643 Cond->getType()->getScalarType()->isIntegerTy(1) &&
1644 areInverseVectorBitmasks(AC, BC)) {
1645 AC = ConstantExpr::getTrunc(AC, CmpInst::makeCmpResultType(Ty));
1646 return Builder.CreateXor(Cond, AC);
1647 }
1648 return nullptr;
1649 }
1650
1651 /// We have an expression of the form (A & C) | (B & D). Try to simplify this
1652 /// to "A' ? C : D", where A' is a boolean or vector of booleans.
matchSelectFromAndOr(Value * A,Value * C,Value * B,Value * D,InstCombiner::BuilderTy & Builder)1653 static Value *matchSelectFromAndOr(Value *A, Value *C, Value *B, Value *D,
1654 InstCombiner::BuilderTy &Builder) {
1655 // The potential condition of the select may be bitcasted. In that case, look
1656 // through its bitcast and the corresponding bitcast of the 'not' condition.
1657 Type *OrigType = A->getType();
1658 Value *SrcA, *SrcB;
1659 if (match(A, m_OneUse(m_BitCast(m_Value(SrcA)))) &&
1660 match(B, m_OneUse(m_BitCast(m_Value(SrcB))))) {
1661 A = SrcA;
1662 B = SrcB;
1663 }
1664
1665 if (Value *Cond = getSelectCondition(A, B, Builder)) {
1666 // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D))
1667 // The bitcasts will either all exist or all not exist. The builder will
1668 // not create unnecessary casts if the types already match.
1669 Value *BitcastC = Builder.CreateBitCast(C, A->getType());
1670 Value *BitcastD = Builder.CreateBitCast(D, A->getType());
1671 Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD);
1672 return Builder.CreateBitCast(Select, OrigType);
1673 }
1674
1675 return nullptr;
1676 }
1677
1678 /// Fold (icmp)|(icmp) if possible.
FoldOrOfICmps(ICmpInst * LHS,ICmpInst * RHS,Instruction * CxtI)1679 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1680 Instruction *CxtI) {
1681 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1682
1683 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
1684 // if K1 and K2 are a one-bit mask.
1685 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1686 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1687
1688 if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() &&
1689 RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1690
1691 BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0));
1692 BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0));
1693 if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() &&
1694 LAnd->getOpcode() == Instruction::And &&
1695 RAnd->getOpcode() == Instruction::And) {
1696
1697 Value *Mask = nullptr;
1698 Value *Masked = nullptr;
1699 if (LAnd->getOperand(0) == RAnd->getOperand(0) &&
1700 isKnownToBeAPowerOfTwo(LAnd->getOperand(1), DL, false, 0, AC, CxtI,
1701 DT) &&
1702 isKnownToBeAPowerOfTwo(RAnd->getOperand(1), DL, false, 0, AC, CxtI,
1703 DT)) {
1704 Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1));
1705 Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask);
1706 } else if (LAnd->getOperand(1) == RAnd->getOperand(1) &&
1707 isKnownToBeAPowerOfTwo(LAnd->getOperand(0), DL, false, 0, AC,
1708 CxtI, DT) &&
1709 isKnownToBeAPowerOfTwo(RAnd->getOperand(0), DL, false, 0, AC,
1710 CxtI, DT)) {
1711 Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0));
1712 Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask);
1713 }
1714
1715 if (Masked)
1716 return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask);
1717 }
1718 }
1719
1720 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
1721 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
1722 // The original condition actually refers to the following two ranges:
1723 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
1724 // We can fold these two ranges if:
1725 // 1) C1 and C2 is unsigned greater than C3.
1726 // 2) The two ranges are separated.
1727 // 3) C1 ^ C2 is one-bit mask.
1728 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
1729 // This implies all values in the two ranges differ by exactly one bit.
1730
1731 if ((LHSCC == ICmpInst::ICMP_ULT || LHSCC == ICmpInst::ICMP_ULE) &&
1732 LHSCC == RHSCC && LHSCst && RHSCst && LHS->hasOneUse() &&
1733 RHS->hasOneUse() && LHSCst->getType() == RHSCst->getType() &&
1734 LHSCst->getValue() == (RHSCst->getValue())) {
1735
1736 Value *LAdd = LHS->getOperand(0);
1737 Value *RAdd = RHS->getOperand(0);
1738
1739 Value *LAddOpnd, *RAddOpnd;
1740 ConstantInt *LAddCst, *RAddCst;
1741 if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddCst))) &&
1742 match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddCst))) &&
1743 LAddCst->getValue().ugt(LHSCst->getValue()) &&
1744 RAddCst->getValue().ugt(LHSCst->getValue())) {
1745
1746 APInt DiffCst = LAddCst->getValue() ^ RAddCst->getValue();
1747 if (LAddOpnd == RAddOpnd && DiffCst.isPowerOf2()) {
1748 ConstantInt *MaxAddCst = nullptr;
1749 if (LAddCst->getValue().ult(RAddCst->getValue()))
1750 MaxAddCst = RAddCst;
1751 else
1752 MaxAddCst = LAddCst;
1753
1754 APInt RRangeLow = -RAddCst->getValue();
1755 APInt RRangeHigh = RRangeLow + LHSCst->getValue();
1756 APInt LRangeLow = -LAddCst->getValue();
1757 APInt LRangeHigh = LRangeLow + LHSCst->getValue();
1758 APInt LowRangeDiff = RRangeLow ^ LRangeLow;
1759 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
1760 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
1761 : RRangeLow - LRangeLow;
1762
1763 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
1764 RangeDiff.ugt(LHSCst->getValue())) {
1765 Value *MaskCst = ConstantInt::get(LAddCst->getType(), ~DiffCst);
1766
1767 Value *NewAnd = Builder->CreateAnd(LAddOpnd, MaskCst);
1768 Value *NewAdd = Builder->CreateAdd(NewAnd, MaxAddCst);
1769 return (Builder->CreateICmp(LHS->getPredicate(), NewAdd, LHSCst));
1770 }
1771 }
1772 }
1773 }
1774
1775 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1776 if (PredicatesFoldable(LHSCC, RHSCC)) {
1777 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1778 LHS->getOperand(1) == RHS->getOperand(0))
1779 LHS->swapOperands();
1780 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1781 LHS->getOperand(1) == RHS->getOperand(1)) {
1782 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1783 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1784 bool isSigned = LHS->isSigned() || RHS->isSigned();
1785 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1786 }
1787 }
1788
1789 // handle (roughly):
1790 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1791 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
1792 return V;
1793
1794 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1795 if (LHS->hasOneUse() || RHS->hasOneUse()) {
1796 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
1797 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
1798 Value *A = nullptr, *B = nullptr;
1799 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
1800 B = Val;
1801 if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
1802 A = Val2;
1803 else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
1804 A = RHS->getOperand(1);
1805 }
1806 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
1807 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
1808 else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1809 B = Val2;
1810 if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
1811 A = Val;
1812 else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
1813 A = LHS->getOperand(1);
1814 }
1815 if (A && B)
1816 return Builder->CreateICmp(
1817 ICmpInst::ICMP_UGE,
1818 Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
1819 }
1820
1821 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
1822 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
1823 return V;
1824
1825 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
1826 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
1827 return V;
1828
1829 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1830 if (!LHSCst || !RHSCst) return nullptr;
1831
1832 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1833 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1834 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1835 Value *NewOr = Builder->CreateOr(Val, Val2);
1836 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1837 }
1838 }
1839
1840 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1841 // iff C2 + CA == C1.
1842 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1843 ConstantInt *AddCst;
1844 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1845 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1846 return Builder->CreateICmpULE(Val, LHSCst);
1847 }
1848
1849 // From here on, we only handle:
1850 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1851 if (Val != Val2) return nullptr;
1852
1853 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1854 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1855 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1856 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1857 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1858 return nullptr;
1859
1860 // We can't fold (ugt x, C) | (sgt x, C2).
1861 if (!PredicatesFoldable(LHSCC, RHSCC))
1862 return nullptr;
1863
1864 // Ensure that the larger constant is on the RHS.
1865 bool ShouldSwap;
1866 if (CmpInst::isSigned(LHSCC) ||
1867 (ICmpInst::isEquality(LHSCC) &&
1868 CmpInst::isSigned(RHSCC)))
1869 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1870 else
1871 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1872
1873 if (ShouldSwap) {
1874 std::swap(LHS, RHS);
1875 std::swap(LHSCst, RHSCst);
1876 std::swap(LHSCC, RHSCC);
1877 }
1878
1879 // At this point, we know we have two icmp instructions
1880 // comparing a value against two constants and or'ing the result
1881 // together. Because of the above check, we know that we only have
1882 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1883 // icmp folding check above), that the two constants are not
1884 // equal.
1885 assert(LHSCst != RHSCst && "Compares not folded above?");
1886
1887 switch (LHSCC) {
1888 default: llvm_unreachable("Unknown integer condition code!");
1889 case ICmpInst::ICMP_EQ:
1890 switch (RHSCC) {
1891 default: llvm_unreachable("Unknown integer condition code!");
1892 case ICmpInst::ICMP_EQ:
1893 if (LHS->getOperand(0) == RHS->getOperand(0)) {
1894 // if LHSCst and RHSCst differ only by one bit:
1895 // (A == C1 || A == C2) -> (A | (C1 ^ C2)) == C2
1896 assert(LHSCst->getValue().ule(LHSCst->getValue()));
1897
1898 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
1899 if (Xor.isPowerOf2()) {
1900 Value *Cst = Builder->getInt(Xor);
1901 Value *Or = Builder->CreateOr(LHS->getOperand(0), Cst);
1902 return Builder->CreateICmp(ICmpInst::ICMP_EQ, Or, RHSCst);
1903 }
1904 }
1905
1906 if (LHSCst == SubOne(RHSCst)) {
1907 // (X == 13 | X == 14) -> X-13 <u 2
1908 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1909 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1910 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1911 return Builder->CreateICmpULT(Add, AddCST);
1912 }
1913
1914 break; // (X == 13 | X == 15) -> no change
1915 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1916 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1917 break;
1918 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1919 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1920 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1921 return RHS;
1922 }
1923 break;
1924 case ICmpInst::ICMP_NE:
1925 switch (RHSCC) {
1926 default: llvm_unreachable("Unknown integer condition code!");
1927 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1928 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1929 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1930 return LHS;
1931 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1932 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1933 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1934 return Builder->getTrue();
1935 }
1936 case ICmpInst::ICMP_ULT:
1937 switch (RHSCC) {
1938 default: llvm_unreachable("Unknown integer condition code!");
1939 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1940 break;
1941 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1942 // If RHSCst is [us]MAXINT, it is always false. Not handling
1943 // this can cause overflow.
1944 if (RHSCst->isMaxValue(false))
1945 return LHS;
1946 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1947 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1948 break;
1949 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1950 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1951 return RHS;
1952 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1953 break;
1954 }
1955 break;
1956 case ICmpInst::ICMP_SLT:
1957 switch (RHSCC) {
1958 default: llvm_unreachable("Unknown integer condition code!");
1959 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1960 break;
1961 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1962 // If RHSCst is [us]MAXINT, it is always false. Not handling
1963 // this can cause overflow.
1964 if (RHSCst->isMaxValue(true))
1965 return LHS;
1966 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1967 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1968 break;
1969 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1970 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1971 return RHS;
1972 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1973 break;
1974 }
1975 break;
1976 case ICmpInst::ICMP_UGT:
1977 switch (RHSCC) {
1978 default: llvm_unreachable("Unknown integer condition code!");
1979 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1980 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1981 return LHS;
1982 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1983 break;
1984 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1985 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1986 return Builder->getTrue();
1987 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1988 break;
1989 }
1990 break;
1991 case ICmpInst::ICMP_SGT:
1992 switch (RHSCC) {
1993 default: llvm_unreachable("Unknown integer condition code!");
1994 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1995 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1996 return LHS;
1997 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1998 break;
1999 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
2000 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
2001 return Builder->getTrue();
2002 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
2003 break;
2004 }
2005 break;
2006 }
2007 return nullptr;
2008 }
2009
2010 /// Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of instcombine, this returns
2011 /// a Value which should already be inserted into the function.
FoldOrOfFCmps(FCmpInst * LHS,FCmpInst * RHS)2012 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
2013 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
2014 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
2015 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
2016
2017 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
2018 // Swap RHS operands to match LHS.
2019 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
2020 std::swap(Op1LHS, Op1RHS);
2021 }
2022
2023 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
2024 // This is a similar transformation to the one in FoldAndOfFCmps.
2025 //
2026 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
2027 // bool(R & CC0) || bool(R & CC1)
2028 // = bool((R & CC0) | (R & CC1))
2029 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)
2030 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS)
2031 return getFCmpValue(getFCmpCode(Op0CC) | getFCmpCode(Op1CC), Op0LHS, Op0RHS,
2032 Builder);
2033
2034 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
2035 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
2036 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
2037 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
2038 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
2039 // If either of the constants are nans, then the whole thing returns
2040 // true.
2041 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
2042 return Builder->getTrue();
2043
2044 // Otherwise, no need to compare the two constants, compare the
2045 // rest.
2046 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2047 }
2048
2049 // Handle vector zeros. This occurs because the canonical form of
2050 // "fcmp uno x,x" is "fcmp uno x, 0".
2051 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
2052 isa<ConstantAggregateZero>(RHS->getOperand(1)))
2053 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2054
2055 return nullptr;
2056 }
2057
2058 return nullptr;
2059 }
2060
2061 /// This helper function folds:
2062 ///
2063 /// ((A | B) & C1) | (B & C2)
2064 ///
2065 /// into:
2066 ///
2067 /// (A & C1) | B
2068 ///
2069 /// when the XOR of the two constants is "all ones" (-1).
FoldOrWithConstants(BinaryOperator & I,Value * Op,Value * A,Value * B,Value * C)2070 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
2071 Value *A, Value *B, Value *C) {
2072 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2073 if (!CI1) return nullptr;
2074
2075 Value *V1 = nullptr;
2076 ConstantInt *CI2 = nullptr;
2077 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr;
2078
2079 APInt Xor = CI1->getValue() ^ CI2->getValue();
2080 if (!Xor.isAllOnesValue()) return nullptr;
2081
2082 if (V1 == A || V1 == B) {
2083 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
2084 return BinaryOperator::CreateOr(NewOp, V1);
2085 }
2086
2087 return nullptr;
2088 }
2089
2090 /// \brief This helper function folds:
2091 ///
2092 /// ((A | B) & C1) ^ (B & C2)
2093 ///
2094 /// into:
2095 ///
2096 /// (A & C1) ^ B
2097 ///
2098 /// when the XOR of the two constants is "all ones" (-1).
FoldXorWithConstants(BinaryOperator & I,Value * Op,Value * A,Value * B,Value * C)2099 Instruction *InstCombiner::FoldXorWithConstants(BinaryOperator &I, Value *Op,
2100 Value *A, Value *B, Value *C) {
2101 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2102 if (!CI1)
2103 return nullptr;
2104
2105 Value *V1 = nullptr;
2106 ConstantInt *CI2 = nullptr;
2107 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2))))
2108 return nullptr;
2109
2110 APInt Xor = CI1->getValue() ^ CI2->getValue();
2111 if (!Xor.isAllOnesValue())
2112 return nullptr;
2113
2114 if (V1 == A || V1 == B) {
2115 Value *NewOp = Builder->CreateAnd(V1 == A ? B : A, CI1);
2116 return BinaryOperator::CreateXor(NewOp, V1);
2117 }
2118
2119 return nullptr;
2120 }
2121
visitOr(BinaryOperator & I)2122 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2123 bool Changed = SimplifyAssociativeOrCommutative(I);
2124 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2125
2126 if (Value *V = SimplifyVectorOp(I))
2127 return replaceInstUsesWith(I, V);
2128
2129 if (Value *V = SimplifyOrInst(Op0, Op1, DL, TLI, DT, AC))
2130 return replaceInstUsesWith(I, V);
2131
2132 // (A&B)|(A&C) -> A&(B|C) etc
2133 if (Value *V = SimplifyUsingDistributiveLaws(I))
2134 return replaceInstUsesWith(I, V);
2135
2136 // See if we can simplify any instructions used by the instruction whose sole
2137 // purpose is to compute bits we don't care about.
2138 if (SimplifyDemandedInstructionBits(I))
2139 return &I;
2140
2141 if (Value *V = SimplifyBSwap(I))
2142 return replaceInstUsesWith(I, V);
2143
2144 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2145 ConstantInt *C1 = nullptr; Value *X = nullptr;
2146 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2147 // iff (C1 & C2) == 0.
2148 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
2149 (RHS->getValue() & C1->getValue()) != 0 &&
2150 Op0->hasOneUse()) {
2151 Value *Or = Builder->CreateOr(X, RHS);
2152 Or->takeName(Op0);
2153 return BinaryOperator::CreateAnd(Or,
2154 Builder->getInt(RHS->getValue() | C1->getValue()));
2155 }
2156
2157 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2158 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
2159 Op0->hasOneUse()) {
2160 Value *Or = Builder->CreateOr(X, RHS);
2161 Or->takeName(Op0);
2162 return BinaryOperator::CreateXor(Or,
2163 Builder->getInt(C1->getValue() & ~RHS->getValue()));
2164 }
2165
2166 // Try to fold constant and into select arguments.
2167 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2168 if (Instruction *R = FoldOpIntoSelect(I, SI))
2169 return R;
2170
2171 if (isa<PHINode>(Op0))
2172 if (Instruction *NV = FoldOpIntoPhi(I))
2173 return NV;
2174 }
2175
2176 // Given an OR instruction, check to see if this is a bswap.
2177 if (Instruction *BSwap = MatchBSwap(I))
2178 return BSwap;
2179
2180 Value *A = nullptr, *B = nullptr;
2181 ConstantInt *C1 = nullptr, *C2 = nullptr;
2182
2183 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2184 if (Op0->hasOneUse() &&
2185 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2186 MaskedValueIsZero(Op1, C1->getValue(), 0, &I)) {
2187 Value *NOr = Builder->CreateOr(A, Op1);
2188 NOr->takeName(Op0);
2189 return BinaryOperator::CreateXor(NOr, C1);
2190 }
2191
2192 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2193 if (Op1->hasOneUse() &&
2194 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2195 MaskedValueIsZero(Op0, C1->getValue(), 0, &I)) {
2196 Value *NOr = Builder->CreateOr(A, Op0);
2197 NOr->takeName(Op0);
2198 return BinaryOperator::CreateXor(NOr, C1);
2199 }
2200
2201 // ((~A & B) | A) -> (A | B)
2202 if (match(Op0, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2203 match(Op1, m_Specific(A)))
2204 return BinaryOperator::CreateOr(A, B);
2205
2206 // ((A & B) | ~A) -> (~A | B)
2207 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2208 match(Op1, m_Not(m_Specific(A))))
2209 return BinaryOperator::CreateOr(Builder->CreateNot(A), B);
2210
2211 // (A & (~B)) | (A ^ B) -> (A ^ B)
2212 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2213 match(Op1, m_Xor(m_Specific(A), m_Specific(B))))
2214 return BinaryOperator::CreateXor(A, B);
2215
2216 // (A ^ B) | ( A & (~B)) -> (A ^ B)
2217 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2218 match(Op1, m_And(m_Specific(A), m_Not(m_Specific(B)))))
2219 return BinaryOperator::CreateXor(A, B);
2220
2221 // (A & C)|(B & D)
2222 Value *C = nullptr, *D = nullptr;
2223 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2224 match(Op1, m_And(m_Value(B), m_Value(D)))) {
2225 Value *V1 = nullptr, *V2 = nullptr;
2226 C1 = dyn_cast<ConstantInt>(C);
2227 C2 = dyn_cast<ConstantInt>(D);
2228 if (C1 && C2) { // (A & C1)|(B & C2)
2229 if ((C1->getValue() & C2->getValue()) == 0) {
2230 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2231 // iff (C1&C2) == 0 and (N&~C1) == 0
2232 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2233 ((V1 == B &&
2234 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2235 (V2 == B &&
2236 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V)
2237 return BinaryOperator::CreateAnd(A,
2238 Builder->getInt(C1->getValue()|C2->getValue()));
2239 // Or commutes, try both ways.
2240 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2241 ((V1 == A &&
2242 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2243 (V2 == A &&
2244 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V)
2245 return BinaryOperator::CreateAnd(B,
2246 Builder->getInt(C1->getValue()|C2->getValue()));
2247
2248 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2249 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2250 ConstantInt *C3 = nullptr, *C4 = nullptr;
2251 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2252 (C3->getValue() & ~C1->getValue()) == 0 &&
2253 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2254 (C4->getValue() & ~C2->getValue()) == 0) {
2255 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2256 return BinaryOperator::CreateAnd(V2,
2257 Builder->getInt(C1->getValue()|C2->getValue()));
2258 }
2259 }
2260 }
2261
2262 // Don't try to form a select if it's unlikely that we'll get rid of at
2263 // least one of the operands. A select is generally more expensive than the
2264 // 'or' that it is replacing.
2265 if (Op0->hasOneUse() || Op1->hasOneUse()) {
2266 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.
2267 if (Value *V = matchSelectFromAndOr(A, C, B, D, *Builder))
2268 return replaceInstUsesWith(I, V);
2269 if (Value *V = matchSelectFromAndOr(A, C, D, B, *Builder))
2270 return replaceInstUsesWith(I, V);
2271 if (Value *V = matchSelectFromAndOr(C, A, B, D, *Builder))
2272 return replaceInstUsesWith(I, V);
2273 if (Value *V = matchSelectFromAndOr(C, A, D, B, *Builder))
2274 return replaceInstUsesWith(I, V);
2275 if (Value *V = matchSelectFromAndOr(B, D, A, C, *Builder))
2276 return replaceInstUsesWith(I, V);
2277 if (Value *V = matchSelectFromAndOr(B, D, C, A, *Builder))
2278 return replaceInstUsesWith(I, V);
2279 if (Value *V = matchSelectFromAndOr(D, B, A, C, *Builder))
2280 return replaceInstUsesWith(I, V);
2281 if (Value *V = matchSelectFromAndOr(D, B, C, A, *Builder))
2282 return replaceInstUsesWith(I, V);
2283 }
2284
2285 // ((A&~B)|(~A&B)) -> A^B
2286 if ((match(C, m_Not(m_Specific(D))) &&
2287 match(B, m_Not(m_Specific(A)))))
2288 return BinaryOperator::CreateXor(A, D);
2289 // ((~B&A)|(~A&B)) -> A^B
2290 if ((match(A, m_Not(m_Specific(D))) &&
2291 match(B, m_Not(m_Specific(C)))))
2292 return BinaryOperator::CreateXor(C, D);
2293 // ((A&~B)|(B&~A)) -> A^B
2294 if ((match(C, m_Not(m_Specific(B))) &&
2295 match(D, m_Not(m_Specific(A)))))
2296 return BinaryOperator::CreateXor(A, B);
2297 // ((~B&A)|(B&~A)) -> A^B
2298 if ((match(A, m_Not(m_Specific(B))) &&
2299 match(D, m_Not(m_Specific(C)))))
2300 return BinaryOperator::CreateXor(C, B);
2301
2302 // ((A|B)&1)|(B&-2) -> (A&1) | B
2303 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
2304 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
2305 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
2306 if (Ret) return Ret;
2307 }
2308 // (B&-2)|((A|B)&1) -> (A&1) | B
2309 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
2310 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
2311 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
2312 if (Ret) return Ret;
2313 }
2314 // ((A^B)&1)|(B&-2) -> (A&1) ^ B
2315 if (match(A, m_Xor(m_Value(V1), m_Specific(B))) ||
2316 match(A, m_Xor(m_Specific(B), m_Value(V1)))) {
2317 Instruction *Ret = FoldXorWithConstants(I, Op1, V1, B, C);
2318 if (Ret) return Ret;
2319 }
2320 // (B&-2)|((A^B)&1) -> (A&1) ^ B
2321 if (match(B, m_Xor(m_Specific(A), m_Value(V1))) ||
2322 match(B, m_Xor(m_Value(V1), m_Specific(A)))) {
2323 Instruction *Ret = FoldXorWithConstants(I, Op0, A, V1, D);
2324 if (Ret) return Ret;
2325 }
2326 }
2327
2328 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2329 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2330 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2331 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
2332 return BinaryOperator::CreateOr(Op0, C);
2333
2334 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2335 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2336 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2337 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
2338 return BinaryOperator::CreateOr(Op1, C);
2339
2340 // ((B | C) & A) | B -> B | (A & C)
2341 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2342 return BinaryOperator::CreateOr(Op1, Builder->CreateAnd(A, C));
2343
2344 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
2345 return DeMorgan;
2346
2347 // Canonicalize xor to the RHS.
2348 bool SwappedForXor = false;
2349 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2350 std::swap(Op0, Op1);
2351 SwappedForXor = true;
2352 }
2353
2354 // A | ( A ^ B) -> A | B
2355 // A | (~A ^ B) -> A | ~B
2356 // (A & B) | (A ^ B)
2357 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2358 if (Op0 == A || Op0 == B)
2359 return BinaryOperator::CreateOr(A, B);
2360
2361 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2362 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2363 return BinaryOperator::CreateOr(A, B);
2364
2365 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2366 Value *Not = Builder->CreateNot(B, B->getName()+".not");
2367 return BinaryOperator::CreateOr(Not, Op0);
2368 }
2369 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2370 Value *Not = Builder->CreateNot(A, A->getName()+".not");
2371 return BinaryOperator::CreateOr(Not, Op0);
2372 }
2373 }
2374
2375 // A | ~(A | B) -> A | ~B
2376 // A | ~(A ^ B) -> A | ~B
2377 if (match(Op1, m_Not(m_Value(A))))
2378 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2379 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2380 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2381 B->getOpcode() == Instruction::Xor)) {
2382 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2383 B->getOperand(0);
2384 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
2385 return BinaryOperator::CreateOr(Not, Op0);
2386 }
2387
2388 // (A & B) | ((~A) ^ B) -> (~A ^ B)
2389 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2390 match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
2391 return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2392
2393 // ((~A) ^ B) | (A & B) -> (~A ^ B)
2394 if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
2395 match(Op1, m_And(m_Specific(A), m_Specific(B))))
2396 return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2397
2398 if (SwappedForXor)
2399 std::swap(Op0, Op1);
2400
2401 {
2402 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2403 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2404 if (LHS && RHS)
2405 if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2406 return replaceInstUsesWith(I, Res);
2407
2408 // TODO: Make this recursive; it's a little tricky because an arbitrary
2409 // number of 'or' instructions might have to be created.
2410 Value *X, *Y;
2411 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2412 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2413 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2414 return replaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2415 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2416 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2417 return replaceInstUsesWith(I, Builder->CreateOr(Res, X));
2418 }
2419 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2420 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2421 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2422 return replaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2423 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2424 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2425 return replaceInstUsesWith(I, Builder->CreateOr(Res, X));
2426 }
2427 }
2428
2429 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
2430 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2431 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2432 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2433 return replaceInstUsesWith(I, Res);
2434
2435 if (Instruction *CastedOr = foldCastedBitwiseLogic(I))
2436 return CastedOr;
2437
2438 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.
2439 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
2440 A->getType()->getScalarType()->isIntegerTy(1))
2441 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2442 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
2443 A->getType()->getScalarType()->isIntegerTy(1))
2444 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2445
2446 // Note: If we've gotten to the point of visiting the outer OR, then the
2447 // inner one couldn't be simplified. If it was a constant, then it won't
2448 // be simplified by a later pass either, so we try swapping the inner/outer
2449 // ORs in the hopes that we'll be able to simplify it this way.
2450 // (X|C) | V --> (X|V) | C
2451 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2452 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2453 Value *Inner = Builder->CreateOr(A, Op1);
2454 Inner->takeName(Op0);
2455 return BinaryOperator::CreateOr(Inner, C1);
2456 }
2457
2458 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2459 // Since this OR statement hasn't been optimized further yet, we hope
2460 // that this transformation will allow the new ORs to be optimized.
2461 {
2462 Value *X = nullptr, *Y = nullptr;
2463 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2464 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2465 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2466 Value *orTrue = Builder->CreateOr(A, C);
2467 Value *orFalse = Builder->CreateOr(B, D);
2468 return SelectInst::Create(X, orTrue, orFalse);
2469 }
2470 }
2471
2472 return Changed ? &I : nullptr;
2473 }
2474
visitXor(BinaryOperator & I)2475 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2476 bool Changed = SimplifyAssociativeOrCommutative(I);
2477 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2478
2479 if (Value *V = SimplifyVectorOp(I))
2480 return replaceInstUsesWith(I, V);
2481
2482 if (Value *V = SimplifyXorInst(Op0, Op1, DL, TLI, DT, AC))
2483 return replaceInstUsesWith(I, V);
2484
2485 // (A&B)^(A&C) -> A&(B^C) etc
2486 if (Value *V = SimplifyUsingDistributiveLaws(I))
2487 return replaceInstUsesWith(I, V);
2488
2489 // See if we can simplify any instructions used by the instruction whose sole
2490 // purpose is to compute bits we don't care about.
2491 if (SimplifyDemandedInstructionBits(I))
2492 return &I;
2493
2494 if (Value *V = SimplifyBSwap(I))
2495 return replaceInstUsesWith(I, V);
2496
2497 // Is this a ~ operation?
2498 if (Value *NotOp = dyn_castNotVal(&I)) {
2499 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2500 if (Op0I->getOpcode() == Instruction::And ||
2501 Op0I->getOpcode() == Instruction::Or) {
2502 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2503 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2504 if (dyn_castNotVal(Op0I->getOperand(1)))
2505 Op0I->swapOperands();
2506 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2507 Value *NotY =
2508 Builder->CreateNot(Op0I->getOperand(1),
2509 Op0I->getOperand(1)->getName()+".not");
2510 if (Op0I->getOpcode() == Instruction::And)
2511 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2512 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2513 }
2514
2515 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2516 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2517 if (IsFreeToInvert(Op0I->getOperand(0),
2518 Op0I->getOperand(0)->hasOneUse()) &&
2519 IsFreeToInvert(Op0I->getOperand(1),
2520 Op0I->getOperand(1)->hasOneUse())) {
2521 Value *NotX =
2522 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2523 Value *NotY =
2524 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2525 if (Op0I->getOpcode() == Instruction::And)
2526 return BinaryOperator::CreateOr(NotX, NotY);
2527 return BinaryOperator::CreateAnd(NotX, NotY);
2528 }
2529
2530 } else if (Op0I->getOpcode() == Instruction::AShr) {
2531 // ~(~X >>s Y) --> (X >>s Y)
2532 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2533 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2534 }
2535 }
2536 }
2537
2538 if (Constant *RHS = dyn_cast<Constant>(Op1)) {
2539 if (RHS->isAllOnesValue() && Op0->hasOneUse())
2540 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2541 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2542 return CmpInst::Create(CI->getOpcode(),
2543 CI->getInversePredicate(),
2544 CI->getOperand(0), CI->getOperand(1));
2545 }
2546
2547 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2548 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2549 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2550 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2551 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2552 Instruction::CastOps Opcode = Op0C->getOpcode();
2553 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2554 (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
2555 Op0C->getDestTy()))) {
2556 CI->setPredicate(CI->getInversePredicate());
2557 return CastInst::Create(Opcode, CI, Op0C->getType());
2558 }
2559 }
2560 }
2561 }
2562
2563 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2564 // ~(c-X) == X-c-1 == X+(-c-1)
2565 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2566 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2567 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2568 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2569 ConstantInt::get(I.getType(), 1));
2570 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2571 }
2572
2573 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2574 if (Op0I->getOpcode() == Instruction::Add) {
2575 // ~(X-c) --> (-c-1)-X
2576 if (RHS->isAllOnesValue()) {
2577 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2578 return BinaryOperator::CreateSub(
2579 ConstantExpr::getSub(NegOp0CI,
2580 ConstantInt::get(I.getType(), 1)),
2581 Op0I->getOperand(0));
2582 } else if (RHS->getValue().isSignBit()) {
2583 // (X + C) ^ signbit -> (X + C + signbit)
2584 Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
2585 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2586
2587 }
2588 } else if (Op0I->getOpcode() == Instruction::Or) {
2589 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2590 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue(),
2591 0, &I)) {
2592 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2593 // Anything in both C1 and C2 is known to be zero, remove it from
2594 // NewRHS.
2595 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2596 NewRHS = ConstantExpr::getAnd(NewRHS,
2597 ConstantExpr::getNot(CommonBits));
2598 Worklist.Add(Op0I);
2599 I.setOperand(0, Op0I->getOperand(0));
2600 I.setOperand(1, NewRHS);
2601 return &I;
2602 }
2603 } else if (Op0I->getOpcode() == Instruction::LShr) {
2604 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2605 // E1 = "X ^ C1"
2606 BinaryOperator *E1;
2607 ConstantInt *C1;
2608 if (Op0I->hasOneUse() &&
2609 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2610 E1->getOpcode() == Instruction::Xor &&
2611 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2612 // fold (C1 >> C2) ^ C3
2613 ConstantInt *C2 = Op0CI, *C3 = RHS;
2614 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2615 FoldConst ^= C3->getValue();
2616 // Prepare the two operands.
2617 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
2618 Opnd0->takeName(Op0I);
2619 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2620 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2621
2622 return BinaryOperator::CreateXor(Opnd0, FoldVal);
2623 }
2624 }
2625 }
2626 }
2627
2628 // Try to fold constant and into select arguments.
2629 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2630 if (Instruction *R = FoldOpIntoSelect(I, SI))
2631 return R;
2632 if (isa<PHINode>(Op0))
2633 if (Instruction *NV = FoldOpIntoPhi(I))
2634 return NV;
2635 }
2636
2637 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2638 if (Op1I) {
2639 Value *A, *B;
2640 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2641 if (A == Op0) { // B^(B|A) == (A|B)^B
2642 Op1I->swapOperands();
2643 I.swapOperands();
2644 std::swap(Op0, Op1);
2645 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2646 I.swapOperands(); // Simplified below.
2647 std::swap(Op0, Op1);
2648 }
2649 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2650 Op1I->hasOneUse()){
2651 if (A == Op0) { // A^(A&B) -> A^(B&A)
2652 Op1I->swapOperands();
2653 std::swap(A, B);
2654 }
2655 if (B == Op0) { // A^(B&A) -> (B&A)^A
2656 I.swapOperands(); // Simplified below.
2657 std::swap(Op0, Op1);
2658 }
2659 }
2660 }
2661
2662 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2663 if (Op0I) {
2664 Value *A, *B;
2665 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2666 Op0I->hasOneUse()) {
2667 if (A == Op1) // (B|A)^B == (A|B)^B
2668 std::swap(A, B);
2669 if (B == Op1) // (A|B)^B == A & ~B
2670 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2671 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2672 Op0I->hasOneUse()){
2673 if (A == Op1) // (A&B)^A -> (B&A)^A
2674 std::swap(A, B);
2675 if (B == Op1 && // (B&A)^A == ~B & A
2676 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2677 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2678 }
2679 }
2680 }
2681
2682 if (Op0I && Op1I) {
2683 Value *A, *B, *C, *D;
2684 // (A & B)^(A | B) -> A ^ B
2685 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2686 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2687 if ((A == C && B == D) || (A == D && B == C))
2688 return BinaryOperator::CreateXor(A, B);
2689 }
2690 // (A | B)^(A & B) -> A ^ B
2691 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2692 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2693 if ((A == C && B == D) || (A == D && B == C))
2694 return BinaryOperator::CreateXor(A, B);
2695 }
2696 // (A | ~B) ^ (~A | B) -> A ^ B
2697 if (match(Op0I, m_Or(m_Value(A), m_Not(m_Value(B)))) &&
2698 match(Op1I, m_Or(m_Not(m_Specific(A)), m_Specific(B)))) {
2699 return BinaryOperator::CreateXor(A, B);
2700 }
2701 // (~A | B) ^ (A | ~B) -> A ^ B
2702 if (match(Op0I, m_Or(m_Not(m_Value(A)), m_Value(B))) &&
2703 match(Op1I, m_Or(m_Specific(A), m_Not(m_Specific(B))))) {
2704 return BinaryOperator::CreateXor(A, B);
2705 }
2706 // (A & ~B) ^ (~A & B) -> A ^ B
2707 if (match(Op0I, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2708 match(Op1I, m_And(m_Not(m_Specific(A)), m_Specific(B)))) {
2709 return BinaryOperator::CreateXor(A, B);
2710 }
2711 // (~A & B) ^ (A & ~B) -> A ^ B
2712 if (match(Op0I, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2713 match(Op1I, m_And(m_Specific(A), m_Not(m_Specific(B))))) {
2714 return BinaryOperator::CreateXor(A, B);
2715 }
2716 // (A ^ C)^(A | B) -> ((~A) & B) ^ C
2717 if (match(Op0I, m_Xor(m_Value(D), m_Value(C))) &&
2718 match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2719 if (D == A)
2720 return BinaryOperator::CreateXor(
2721 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2722 if (D == B)
2723 return BinaryOperator::CreateXor(
2724 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2725 }
2726 // (A | B)^(A ^ C) -> ((~A) & B) ^ C
2727 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2728 match(Op1I, m_Xor(m_Value(D), m_Value(C)))) {
2729 if (D == A)
2730 return BinaryOperator::CreateXor(
2731 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2732 if (D == B)
2733 return BinaryOperator::CreateXor(
2734 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2735 }
2736 // (A & B) ^ (A ^ B) -> (A | B)
2737 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2738 match(Op1I, m_Xor(m_Specific(A), m_Specific(B))))
2739 return BinaryOperator::CreateOr(A, B);
2740 // (A ^ B) ^ (A & B) -> (A | B)
2741 if (match(Op0I, m_Xor(m_Value(A), m_Value(B))) &&
2742 match(Op1I, m_And(m_Specific(A), m_Specific(B))))
2743 return BinaryOperator::CreateOr(A, B);
2744 }
2745
2746 Value *A = nullptr, *B = nullptr;
2747 // (A & ~B) ^ (~A) -> ~(A & B)
2748 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2749 match(Op1, m_Not(m_Specific(A))))
2750 return BinaryOperator::CreateNot(Builder->CreateAnd(A, B));
2751
2752 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2753 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2754 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2755 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2756 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2757 LHS->getOperand(1) == RHS->getOperand(0))
2758 LHS->swapOperands();
2759 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2760 LHS->getOperand(1) == RHS->getOperand(1)) {
2761 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2762 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2763 bool isSigned = LHS->isSigned() || RHS->isSigned();
2764 return replaceInstUsesWith(I,
2765 getNewICmpValue(isSigned, Code, Op0, Op1,
2766 Builder));
2767 }
2768 }
2769
2770 if (Instruction *CastedXor = foldCastedBitwiseLogic(I))
2771 return CastedXor;
2772
2773 return Changed ? &I : nullptr;
2774 }
2775