1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements routines for folding instructions into simpler forms
10 // that do not require creating new instructions. This does constant folding
11 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
12 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
13 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
14 // simplified: This is usually true and assuming it simplifies the logic (if
15 // they have not been simplified then results are correct but maybe suboptimal).
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/AssumptionCache.h"
24 #include "llvm/Analysis/CaptureTracking.h"
25 #include "llvm/Analysis/CmpInstAnalysis.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/LoopAnalysisManager.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/Analysis/VectorUtils.h"
31 #include "llvm/IR/ConstantRange.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/GetElementPtrTypeIterator.h"
35 #include "llvm/IR/GlobalAlias.h"
36 #include "llvm/IR/InstrTypes.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/IR/PatternMatch.h"
40 #include "llvm/IR/ValueHandle.h"
41 #include "llvm/Support/KnownBits.h"
44 using namespace llvm::PatternMatch;
46 #define DEBUG_TYPE "instsimplify"
48 enum { RecursionLimit = 3 };
50 STATISTIC(NumExpand, "Number of expansions");
51 STATISTIC(NumReassoc, "Number of reassociations");
53 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
54 static Value *simplifyUnOp(unsigned, Value *, const SimplifyQuery &, unsigned);
55 static Value *simplifyFPUnOp(unsigned, Value *, const FastMathFlags &,
56 const SimplifyQuery &, unsigned);
57 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
59 static Value *SimplifyBinOp(unsigned, Value *, Value *, const FastMathFlags &,
60 const SimplifyQuery &, unsigned);
61 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
63 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
64 const SimplifyQuery &Q, unsigned MaxRecurse);
65 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
66 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
67 static Value *SimplifyCastInst(unsigned, Value *, Type *,
68 const SimplifyQuery &, unsigned);
69 static Value *SimplifyGEPInst(Type *, ArrayRef<Value *>, const SimplifyQuery &,
72 static Value *foldSelectWithBinaryOp(Value *Cond, Value *TrueVal,
74 BinaryOperator::BinaryOps BinOpCode;
75 if (auto *BO = dyn_cast<BinaryOperator>(Cond))
76 BinOpCode = BO->getOpcode();
80 CmpInst::Predicate ExpectedPred, Pred1, Pred2;
81 if (BinOpCode == BinaryOperator::Or) {
82 ExpectedPred = ICmpInst::ICMP_NE;
83 } else if (BinOpCode == BinaryOperator::And) {
84 ExpectedPred = ICmpInst::ICMP_EQ;
88 // %A = icmp eq %TV, %FV
89 // %B = icmp eq %X, %Y (and one of these is a select operand)
91 // %D = select %C, %TV, %FV
95 // %A = icmp ne %TV, %FV
96 // %B = icmp ne %X, %Y (and one of these is a select operand)
98 // %D = select %C, %TV, %FV
102 if (!match(Cond, m_c_BinOp(m_c_ICmp(Pred1, m_Specific(TrueVal),
103 m_Specific(FalseVal)),
104 m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
105 Pred1 != Pred2 || Pred1 != ExpectedPred)
108 if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
109 return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
114 /// For a boolean type or a vector of boolean type, return false or a vector
115 /// with every element false.
116 static Constant *getFalse(Type *Ty) {
117 return ConstantInt::getFalse(Ty);
120 /// For a boolean type or a vector of boolean type, return true or a vector
121 /// with every element true.
122 static Constant *getTrue(Type *Ty) {
123 return ConstantInt::getTrue(Ty);
126 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
127 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
129 CmpInst *Cmp = dyn_cast<CmpInst>(V);
132 CmpInst::Predicate CPred = Cmp->getPredicate();
133 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
134 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
136 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
140 /// Simplify comparison with true or false branch of select:
141 /// %sel = select i1 %cond, i32 %tv, i32 %fv
142 /// %cmp = icmp sle i32 %sel, %rhs
143 /// Compose new comparison by substituting %sel with either %tv or %fv
144 /// and see if it simplifies.
145 static Value *simplifyCmpSelCase(CmpInst::Predicate Pred, Value *LHS,
146 Value *RHS, Value *Cond,
147 const SimplifyQuery &Q, unsigned MaxRecurse,
148 Constant *TrueOrFalse) {
149 Value *SimplifiedCmp = SimplifyCmpInst(Pred, LHS, RHS, Q, MaxRecurse);
150 if (SimplifiedCmp == Cond) {
151 // %cmp simplified to the select condition (%cond).
153 } else if (!SimplifiedCmp && isSameCompare(Cond, Pred, LHS, RHS)) {
154 // It didn't simplify. However, if composed comparison is equivalent
155 // to the select condition (%cond) then we can replace it.
158 return SimplifiedCmp;
161 /// Simplify comparison with true branch of select
162 static Value *simplifyCmpSelTrueCase(CmpInst::Predicate Pred, Value *LHS,
163 Value *RHS, Value *Cond,
164 const SimplifyQuery &Q,
165 unsigned MaxRecurse) {
166 return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
167 getTrue(Cond->getType()));
170 /// Simplify comparison with false branch of select
171 static Value *simplifyCmpSelFalseCase(CmpInst::Predicate Pred, Value *LHS,
172 Value *RHS, Value *Cond,
173 const SimplifyQuery &Q,
174 unsigned MaxRecurse) {
175 return simplifyCmpSelCase(Pred, LHS, RHS, Cond, Q, MaxRecurse,
176 getFalse(Cond->getType()));
179 /// We know comparison with both branches of select can be simplified, but they
180 /// are not equal. This routine handles some logical simplifications.
181 static Value *handleOtherCmpSelSimplifications(Value *TCmp, Value *FCmp,
183 const SimplifyQuery &Q,
184 unsigned MaxRecurse) {
185 // If the false value simplified to false, then the result of the compare
186 // is equal to "Cond && TCmp". This also catches the case when the false
187 // value simplified to false and the true value to true, returning "Cond".
188 if (match(FCmp, m_Zero()))
189 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
191 // If the true value simplified to true, then the result of the compare
192 // is equal to "Cond || FCmp".
193 if (match(TCmp, m_One()))
194 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
196 // Finally, if the false value simplified to true and the true value to
197 // false, then the result of the compare is equal to "!Cond".
198 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
199 if (Value *V = SimplifyXorInst(
200 Cond, Constant::getAllOnesValue(Cond->getType()), Q, MaxRecurse))
205 /// Does the given value dominate the specified phi node?
206 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
207 Instruction *I = dyn_cast<Instruction>(V);
209 // Arguments and constants dominate all instructions.
212 // If we are processing instructions (and/or basic blocks) that have not been
213 // fully added to a function, the parent nodes may still be null. Simply
214 // return the conservative answer in these cases.
215 if (!I->getParent() || !P->getParent() || !I->getFunction())
218 // If we have a DominatorTree then do a precise test.
220 return DT->dominates(I, P);
222 // Otherwise, if the instruction is in the entry block and is not an invoke,
223 // then it obviously dominates all phi nodes.
224 if (I->getParent() == &I->getFunction()->getEntryBlock() &&
231 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
232 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
233 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
234 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
235 /// Returns the simplified value, or null if no simplification was performed.
236 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
237 Instruction::BinaryOps OpcodeToExpand,
238 const SimplifyQuery &Q, unsigned MaxRecurse) {
239 // Recursion is always used, so bail out at once if we already hit the limit.
243 // Check whether the expression has the form "(A op' B) op C".
244 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
245 if (Op0->getOpcode() == OpcodeToExpand) {
246 // It does! Try turning it into "(A op C) op' (B op C)".
247 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
248 // Do "A op C" and "B op C" both simplify?
249 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
250 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
251 // They do! Return "L op' R" if it simplifies or is already available.
252 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
253 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
254 && L == B && R == A)) {
258 // Otherwise return "L op' R" if it simplifies.
259 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
266 // Check whether the expression has the form "A op (B op' C)".
267 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
268 if (Op1->getOpcode() == OpcodeToExpand) {
269 // It does! Try turning it into "(A op B) op' (A op C)".
270 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
271 // Do "A op B" and "A op C" both simplify?
272 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
273 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
274 // They do! Return "L op' R" if it simplifies or is already available.
275 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
276 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
277 && L == C && R == B)) {
281 // Otherwise return "L op' R" if it simplifies.
282 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
292 /// Generic simplifications for associative binary operations.
293 /// Returns the simpler value, or null if none was found.
294 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
295 Value *LHS, Value *RHS,
296 const SimplifyQuery &Q,
297 unsigned MaxRecurse) {
298 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
300 // Recursion is always used, so bail out at once if we already hit the limit.
304 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
305 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
307 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
308 if (Op0 && Op0->getOpcode() == Opcode) {
309 Value *A = Op0->getOperand(0);
310 Value *B = Op0->getOperand(1);
313 // Does "B op C" simplify?
314 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
315 // It does! Return "A op V" if it simplifies or is already available.
316 // If V equals B then "A op V" is just the LHS.
317 if (V == B) return LHS;
318 // Otherwise return "A op V" if it simplifies.
319 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
326 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
327 if (Op1 && Op1->getOpcode() == Opcode) {
329 Value *B = Op1->getOperand(0);
330 Value *C = Op1->getOperand(1);
332 // Does "A op B" simplify?
333 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
334 // It does! Return "V op C" if it simplifies or is already available.
335 // If V equals B then "V op C" is just the RHS.
336 if (V == B) return RHS;
337 // Otherwise return "V op C" if it simplifies.
338 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
345 // The remaining transforms require commutativity as well as associativity.
346 if (!Instruction::isCommutative(Opcode))
349 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
350 if (Op0 && Op0->getOpcode() == Opcode) {
351 Value *A = Op0->getOperand(0);
352 Value *B = Op0->getOperand(1);
355 // Does "C op A" simplify?
356 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
357 // It does! Return "V op B" if it simplifies or is already available.
358 // If V equals A then "V op B" is just the LHS.
359 if (V == A) return LHS;
360 // Otherwise return "V op B" if it simplifies.
361 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
368 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
369 if (Op1 && Op1->getOpcode() == Opcode) {
371 Value *B = Op1->getOperand(0);
372 Value *C = Op1->getOperand(1);
374 // Does "C op A" simplify?
375 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
376 // It does! Return "B op V" if it simplifies or is already available.
377 // If V equals C then "B op V" is just the RHS.
378 if (V == C) return RHS;
379 // Otherwise return "B op V" if it simplifies.
380 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
390 /// In the case of a binary operation with a select instruction as an operand,
391 /// try to simplify the binop by seeing whether evaluating it on both branches
392 /// of the select results in the same value. Returns the common value if so,
393 /// otherwise returns null.
394 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
395 Value *RHS, const SimplifyQuery &Q,
396 unsigned MaxRecurse) {
397 // Recursion is always used, so bail out at once if we already hit the limit.
402 if (isa<SelectInst>(LHS)) {
403 SI = cast<SelectInst>(LHS);
405 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
406 SI = cast<SelectInst>(RHS);
409 // Evaluate the BinOp on the true and false branches of the select.
413 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
414 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
416 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
417 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
420 // If they simplified to the same value, then return the common value.
421 // If they both failed to simplify then return null.
425 // If one branch simplified to undef, return the other one.
426 if (TV && isa<UndefValue>(TV))
428 if (FV && isa<UndefValue>(FV))
431 // If applying the operation did not change the true and false select values,
432 // then the result of the binop is the select itself.
433 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
436 // If one branch simplified and the other did not, and the simplified
437 // value is equal to the unsimplified one, return the simplified value.
438 // For example, select (cond, X, X & Z) & Z -> X & Z.
439 if ((FV && !TV) || (TV && !FV)) {
440 // Check that the simplified value has the form "X op Y" where "op" is the
441 // same as the original operation.
442 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
443 if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
444 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
445 // We already know that "op" is the same as for the simplified value. See
446 // if the operands match too. If so, return the simplified value.
447 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
448 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
449 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
450 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
451 Simplified->getOperand(1) == UnsimplifiedRHS)
453 if (Simplified->isCommutative() &&
454 Simplified->getOperand(1) == UnsimplifiedLHS &&
455 Simplified->getOperand(0) == UnsimplifiedRHS)
463 /// In the case of a comparison with a select instruction, try to simplify the
464 /// comparison by seeing whether both branches of the select result in the same
465 /// value. Returns the common value if so, otherwise returns null.
466 /// For example, if we have:
467 /// %tmp = select i1 %cmp, i32 1, i32 2
468 /// %cmp1 = icmp sle i32 %tmp, 3
469 /// We can simplify %cmp1 to true, because both branches of select are
470 /// less than 3. We compose new comparison by substituting %tmp with both
471 /// branches of select and see if it can be simplified.
472 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
473 Value *RHS, const SimplifyQuery &Q,
474 unsigned MaxRecurse) {
475 // Recursion is always used, so bail out at once if we already hit the limit.
479 // Make sure the select is on the LHS.
480 if (!isa<SelectInst>(LHS)) {
482 Pred = CmpInst::getSwappedPredicate(Pred);
484 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
485 SelectInst *SI = cast<SelectInst>(LHS);
486 Value *Cond = SI->getCondition();
487 Value *TV = SI->getTrueValue();
488 Value *FV = SI->getFalseValue();
490 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
491 // Does "cmp TV, RHS" simplify?
492 Value *TCmp = simplifyCmpSelTrueCase(Pred, TV, RHS, Cond, Q, MaxRecurse);
496 // Does "cmp FV, RHS" simplify?
497 Value *FCmp = simplifyCmpSelFalseCase(Pred, FV, RHS, Cond, Q, MaxRecurse);
501 // If both sides simplified to the same value, then use it as the result of
502 // the original comparison.
506 // The remaining cases only make sense if the select condition has the same
507 // type as the result of the comparison, so bail out if this is not so.
508 if (Cond->getType()->isVectorTy() == RHS->getType()->isVectorTy())
509 return handleOtherCmpSelSimplifications(TCmp, FCmp, Cond, Q, MaxRecurse);
514 /// In the case of a binary operation with an operand that is a PHI instruction,
515 /// try to simplify the binop by seeing whether evaluating it on the incoming
516 /// phi values yields the same result for every value. If so returns the common
517 /// value, otherwise returns null.
518 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
519 Value *RHS, const SimplifyQuery &Q,
520 unsigned MaxRecurse) {
521 // Recursion is always used, so bail out at once if we already hit the limit.
526 if (isa<PHINode>(LHS)) {
527 PI = cast<PHINode>(LHS);
528 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
529 if (!valueDominatesPHI(RHS, PI, Q.DT))
532 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
533 PI = cast<PHINode>(RHS);
534 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
535 if (!valueDominatesPHI(LHS, PI, Q.DT))
539 // Evaluate the BinOp on the incoming phi values.
540 Value *CommonValue = nullptr;
541 for (Value *Incoming : PI->incoming_values()) {
542 // If the incoming value is the phi node itself, it can safely be skipped.
543 if (Incoming == PI) continue;
544 Value *V = PI == LHS ?
545 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
546 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
547 // If the operation failed to simplify, or simplified to a different value
548 // to previously, then give up.
549 if (!V || (CommonValue && V != CommonValue))
557 /// In the case of a comparison with a PHI instruction, try to simplify the
558 /// comparison by seeing whether comparing with all of the incoming phi values
559 /// yields the same result every time. If so returns the common result,
560 /// otherwise returns null.
561 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
562 const SimplifyQuery &Q, unsigned MaxRecurse) {
563 // Recursion is always used, so bail out at once if we already hit the limit.
567 // Make sure the phi is on the LHS.
568 if (!isa<PHINode>(LHS)) {
570 Pred = CmpInst::getSwappedPredicate(Pred);
572 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
573 PHINode *PI = cast<PHINode>(LHS);
575 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
576 if (!valueDominatesPHI(RHS, PI, Q.DT))
579 // Evaluate the BinOp on the incoming phi values.
580 Value *CommonValue = nullptr;
581 for (unsigned u = 0, e = PI->getNumIncomingValues(); u < e; ++u) {
582 Value *Incoming = PI->getIncomingValue(u);
583 Instruction *InTI = PI->getIncomingBlock(u)->getTerminator();
584 // If the incoming value is the phi node itself, it can safely be skipped.
585 if (Incoming == PI) continue;
586 // Change the context instruction to the "edge" that flows into the phi.
587 // This is important because that is where incoming is actually "evaluated"
588 // even though it is used later somewhere else.
589 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q.getWithInstruction(InTI),
591 // If the operation failed to simplify, or simplified to a different value
592 // to previously, then give up.
593 if (!V || (CommonValue && V != CommonValue))
601 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
602 Value *&Op0, Value *&Op1,
603 const SimplifyQuery &Q) {
604 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
605 if (auto *CRHS = dyn_cast<Constant>(Op1))
606 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
608 // Canonicalize the constant to the RHS if this is a commutative operation.
609 if (Instruction::isCommutative(Opcode))
615 /// Given operands for an Add, see if we can fold the result.
616 /// If not, this returns null.
617 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
618 const SimplifyQuery &Q, unsigned MaxRecurse) {
619 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
622 // X + undef -> undef
623 if (match(Op1, m_Undef()))
627 if (match(Op1, m_Zero()))
630 // If two operands are negative, return 0.
631 if (isKnownNegation(Op0, Op1))
632 return Constant::getNullValue(Op0->getType());
638 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
639 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
642 // X + ~X -> -1 since ~X = -X-1
643 Type *Ty = Op0->getType();
644 if (match(Op0, m_Not(m_Specific(Op1))) ||
645 match(Op1, m_Not(m_Specific(Op0))))
646 return Constant::getAllOnesValue(Ty);
648 // add nsw/nuw (xor Y, signmask), signmask --> Y
649 // The no-wrapping add guarantees that the top bit will be set by the add.
650 // Therefore, the xor must be clearing the already set sign bit of Y.
651 if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
652 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
655 // add nuw %x, -1 -> -1, because %x can only be 0.
656 if (IsNUW && match(Op1, m_AllOnes()))
657 return Op1; // Which is -1.
660 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
661 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
664 // Try some generic simplifications for associative operations.
665 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
669 // Threading Add over selects and phi nodes is pointless, so don't bother.
670 // Threading over the select in "A + select(cond, B, C)" means evaluating
671 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
672 // only if B and C are equal. If B and C are equal then (since we assume
673 // that operands have already been simplified) "select(cond, B, C)" should
674 // have been simplified to the common value of B and C already. Analysing
675 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
676 // for threading over phi nodes.
681 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
682 const SimplifyQuery &Query) {
683 return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
686 /// Compute the base pointer and cumulative constant offsets for V.
688 /// This strips all constant offsets off of V, leaving it the base pointer, and
689 /// accumulates the total constant offset applied in the returned constant. It
690 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
691 /// no constant offsets applied.
693 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
694 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
696 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
697 bool AllowNonInbounds = false) {
698 assert(V->getType()->isPtrOrPtrVectorTy());
700 Type *IntIdxTy = DL.getIndexType(V->getType())->getScalarType();
701 APInt Offset = APInt::getNullValue(IntIdxTy->getIntegerBitWidth());
703 V = V->stripAndAccumulateConstantOffsets(DL, Offset, AllowNonInbounds);
704 // As that strip may trace through `addrspacecast`, need to sext or trunc
705 // the offset calculated.
706 IntIdxTy = DL.getIndexType(V->getType())->getScalarType();
707 Offset = Offset.sextOrTrunc(IntIdxTy->getIntegerBitWidth());
709 Constant *OffsetIntPtr = ConstantInt::get(IntIdxTy, Offset);
710 if (V->getType()->isVectorTy())
711 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
716 /// Compute the constant difference between two pointer values.
717 /// If the difference is not a constant, returns zero.
718 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
720 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
721 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
723 // If LHS and RHS are not related via constant offsets to the same base
724 // value, there is nothing we can do here.
728 // Otherwise, the difference of LHS - RHS can be computed as:
730 // = (LHSOffset + Base) - (RHSOffset + Base)
731 // = LHSOffset - RHSOffset
732 return ConstantExpr::getSub(LHSOffset, RHSOffset);
735 /// Given operands for a Sub, see if we can fold the result.
736 /// If not, this returns null.
737 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
738 const SimplifyQuery &Q, unsigned MaxRecurse) {
739 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
742 // X - undef -> undef
743 // undef - X -> undef
744 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
745 return UndefValue::get(Op0->getType());
748 if (match(Op1, m_Zero()))
753 return Constant::getNullValue(Op0->getType());
755 // Is this a negation?
756 if (match(Op0, m_Zero())) {
757 // 0 - X -> 0 if the sub is NUW.
759 return Constant::getNullValue(Op0->getType());
761 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
762 if (Known.Zero.isMaxSignedValue()) {
763 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
764 // Op1 must be 0 because negating the minimum signed value is undefined.
766 return Constant::getNullValue(Op0->getType());
768 // 0 - X -> X if X is 0 or the minimum signed value.
773 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
774 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
775 Value *X = nullptr, *Y = nullptr, *Z = Op1;
776 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
777 // See if "V === Y - Z" simplifies.
778 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
779 // It does! Now see if "X + V" simplifies.
780 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
781 // It does, we successfully reassociated!
785 // See if "V === X - Z" simplifies.
786 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
787 // It does! Now see if "Y + V" simplifies.
788 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
789 // It does, we successfully reassociated!
795 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
796 // For example, X - (X + 1) -> -1
798 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
799 // See if "V === X - Y" simplifies.
800 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
801 // It does! Now see if "V - Z" simplifies.
802 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
803 // It does, we successfully reassociated!
807 // See if "V === X - Z" simplifies.
808 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
809 // It does! Now see if "V - Y" simplifies.
810 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
811 // It does, we successfully reassociated!
817 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
818 // For example, X - (X - Y) -> Y.
820 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
821 // See if "V === Z - X" simplifies.
822 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
823 // It does! Now see if "V + Y" simplifies.
824 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
825 // It does, we successfully reassociated!
830 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
831 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
832 match(Op1, m_Trunc(m_Value(Y))))
833 if (X->getType() == Y->getType())
834 // See if "V === X - Y" simplifies.
835 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
836 // It does! Now see if "trunc V" simplifies.
837 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
839 // It does, return the simplified "trunc V".
842 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
843 if (match(Op0, m_PtrToInt(m_Value(X))) &&
844 match(Op1, m_PtrToInt(m_Value(Y))))
845 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
846 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
849 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
850 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
853 // Threading Sub over selects and phi nodes is pointless, so don't bother.
854 // Threading over the select in "A - select(cond, B, C)" means evaluating
855 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
856 // only if B and C are equal. If B and C are equal then (since we assume
857 // that operands have already been simplified) "select(cond, B, C)" should
858 // have been simplified to the common value of B and C already. Analysing
859 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
860 // for threading over phi nodes.
865 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
866 const SimplifyQuery &Q) {
867 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
870 /// Given operands for a Mul, see if we can fold the result.
871 /// If not, this returns null.
872 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
873 unsigned MaxRecurse) {
874 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
879 if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
880 return Constant::getNullValue(Op0->getType());
883 if (match(Op1, m_One()))
886 // (X / Y) * Y -> X if the division is exact.
888 if (Q.IIQ.UseInstrInfo &&
890 m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
891 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
895 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
896 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
899 // Try some generic simplifications for associative operations.
900 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
904 // Mul distributes over Add. Try some generic simplifications based on this.
905 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
909 // If the operation is with the result of a select instruction, check whether
910 // operating on either branch of the select always yields the same value.
911 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
912 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
916 // If the operation is with the result of a phi instruction, check whether
917 // operating on all incoming values of the phi always yields the same value.
918 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
919 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
926 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
927 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
930 /// Check for common or similar folds of integer division or integer remainder.
931 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
932 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
933 Type *Ty = Op0->getType();
935 // X / undef -> undef
936 // X % undef -> undef
937 if (match(Op1, m_Undef()))
942 // We don't need to preserve faults!
943 if (match(Op1, m_Zero()))
944 return UndefValue::get(Ty);
946 // If any element of a constant divisor vector is zero or undef, the whole op
948 auto *Op1C = dyn_cast<Constant>(Op1);
949 if (Op1C && Ty->isVectorTy()) {
950 unsigned NumElts = Ty->getVectorNumElements();
951 for (unsigned i = 0; i != NumElts; ++i) {
952 Constant *Elt = Op1C->getAggregateElement(i);
953 if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
954 return UndefValue::get(Ty);
960 if (match(Op0, m_Undef()))
961 return Constant::getNullValue(Ty);
965 if (match(Op0, m_Zero()))
966 return Constant::getNullValue(Op0->getType());
971 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
975 // If this is a boolean op (single-bit element type), we can't have
976 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
977 // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
979 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
980 (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
981 return IsDiv ? Op0 : Constant::getNullValue(Ty);
986 /// Given a predicate and two operands, return true if the comparison is true.
987 /// This is a helper for div/rem simplification where we return some other value
988 /// when we can prove a relationship between the operands.
989 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
990 const SimplifyQuery &Q, unsigned MaxRecurse) {
991 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
992 Constant *C = dyn_cast_or_null<Constant>(V);
993 return (C && C->isAllOnesValue());
996 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
997 /// to simplify X % Y to X.
998 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
999 unsigned MaxRecurse, bool IsSigned) {
1000 // Recursion is always used, so bail out at once if we already hit the limit.
1007 // We require that 1 operand is a simple constant. That could be extended to
1008 // 2 variables if we computed the sign bit for each.
1010 // Make sure that a constant is not the minimum signed value because taking
1011 // the abs() of that is undefined.
1012 Type *Ty = X->getType();
1014 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
1015 // Is the variable divisor magnitude always greater than the constant
1016 // dividend magnitude?
1017 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
1018 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
1019 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
1020 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
1021 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
1024 if (match(Y, m_APInt(C))) {
1025 // Special-case: we can't take the abs() of a minimum signed value. If
1026 // that's the divisor, then all we have to do is prove that the dividend
1027 // is also not the minimum signed value.
1028 if (C->isMinSignedValue())
1029 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1031 // Is the variable dividend magnitude always less than the constant
1032 // divisor magnitude?
1033 // |X| < |C| --> X > -abs(C) and X < abs(C)
1034 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1035 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1036 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1037 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1043 // IsSigned == false.
1044 // Is the dividend unsigned less than the divisor?
1045 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1048 /// These are simplifications common to SDiv and UDiv.
1049 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1050 const SimplifyQuery &Q, unsigned MaxRecurse) {
1051 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1054 if (Value *V = simplifyDivRem(Op0, Op1, true))
1057 bool IsSigned = Opcode == Instruction::SDiv;
1059 // (X * Y) / Y -> X if the multiplication does not overflow.
1061 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1062 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1063 // If the Mul does not overflow, then we are good to go.
1064 if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1065 (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)))
1067 // If X has the form X = A / Y, then X * Y cannot overflow.
1068 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1069 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1073 // (X rem Y) / Y -> 0
1074 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1075 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1076 return Constant::getNullValue(Op0->getType());
1078 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1079 ConstantInt *C1, *C2;
1080 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1081 match(Op1, m_ConstantInt(C2))) {
1083 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1085 return Constant::getNullValue(Op0->getType());
1088 // If the operation is with the result of a select instruction, check whether
1089 // operating on either branch of the select always yields the same value.
1090 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1091 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1094 // If the operation is with the result of a phi instruction, check whether
1095 // operating on all incoming values of the phi always yields the same value.
1096 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1097 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1100 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1101 return Constant::getNullValue(Op0->getType());
1106 /// These are simplifications common to SRem and URem.
1107 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1108 const SimplifyQuery &Q, unsigned MaxRecurse) {
1109 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1112 if (Value *V = simplifyDivRem(Op0, Op1, false))
1115 // (X % Y) % Y -> X % Y
1116 if ((Opcode == Instruction::SRem &&
1117 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1118 (Opcode == Instruction::URem &&
1119 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1122 // (X << Y) % X -> 0
1123 if (Q.IIQ.UseInstrInfo &&
1124 ((Opcode == Instruction::SRem &&
1125 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1126 (Opcode == Instruction::URem &&
1127 match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1128 return Constant::getNullValue(Op0->getType());
1130 // If the operation is with the result of a select instruction, check whether
1131 // operating on either branch of the select always yields the same value.
1132 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1133 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1136 // If the operation is with the result of a phi instruction, check whether
1137 // operating on all incoming values of the phi always yields the same value.
1138 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1139 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1142 // If X / Y == 0, then X % Y == X.
1143 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1149 /// Given operands for an SDiv, see if we can fold the result.
1150 /// If not, this returns null.
1151 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1152 unsigned MaxRecurse) {
1153 // If two operands are negated and no signed overflow, return -1.
1154 if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1155 return Constant::getAllOnesValue(Op0->getType());
1157 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1160 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1161 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1164 /// Given operands for a UDiv, see if we can fold the result.
1165 /// If not, this returns null.
1166 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1167 unsigned MaxRecurse) {
1168 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1171 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1172 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1175 /// Given operands for an SRem, see if we can fold the result.
1176 /// If not, this returns null.
1177 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1178 unsigned MaxRecurse) {
1179 // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1180 // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1182 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1183 return ConstantInt::getNullValue(Op0->getType());
1185 // If the two operands are negated, return 0.
1186 if (isKnownNegation(Op0, Op1))
1187 return ConstantInt::getNullValue(Op0->getType());
1189 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1192 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1193 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1196 /// Given operands for a URem, see if we can fold the result.
1197 /// If not, this returns null.
1198 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1199 unsigned MaxRecurse) {
1200 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1203 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1204 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1207 /// Returns true if a shift by \c Amount always yields undef.
1208 static bool isUndefShift(Value *Amount) {
1209 Constant *C = dyn_cast<Constant>(Amount);
1213 // X shift by undef -> undef because it may shift by the bitwidth.
1214 if (isa<UndefValue>(C))
1217 // Shifting by the bitwidth or more is undefined.
1218 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1219 if (CI->getValue().getLimitedValue() >=
1220 CI->getType()->getScalarSizeInBits())
1223 // If all lanes of a vector shift are undefined the whole shift is.
1224 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1225 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1226 if (!isUndefShift(C->getAggregateElement(I)))
1234 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1235 /// If not, this returns null.
1236 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1237 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1238 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1241 // 0 shift by X -> 0
1242 if (match(Op0, m_Zero()))
1243 return Constant::getNullValue(Op0->getType());
1245 // X shift by 0 -> X
1246 // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1249 if (match(Op1, m_Zero()) ||
1250 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1253 // Fold undefined shifts.
1254 if (isUndefShift(Op1))
1255 return UndefValue::get(Op0->getType());
1257 // If the operation is with the result of a select instruction, check whether
1258 // operating on either branch of the select always yields the same value.
1259 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1260 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1263 // If the operation is with the result of a phi instruction, check whether
1264 // operating on all incoming values of the phi always yields the same value.
1265 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1266 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1269 // If any bits in the shift amount make that value greater than or equal to
1270 // the number of bits in the type, the shift is undefined.
1271 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1272 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1273 return UndefValue::get(Op0->getType());
1275 // If all valid bits in the shift amount are known zero, the first operand is
1277 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1278 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1284 /// Given operands for an Shl, LShr or AShr, see if we can
1285 /// fold the result. If not, this returns null.
1286 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1287 Value *Op1, bool isExact, const SimplifyQuery &Q,
1288 unsigned MaxRecurse) {
1289 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1294 return Constant::getNullValue(Op0->getType());
1297 // undef >> X -> undef (if it's exact)
1298 if (match(Op0, m_Undef()))
1299 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1301 // The low bit cannot be shifted out of an exact shift if it is set.
1303 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1304 if (Op0Known.One[0])
1311 /// Given operands for an Shl, see if we can fold the result.
1312 /// If not, this returns null.
1313 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1314 const SimplifyQuery &Q, unsigned MaxRecurse) {
1315 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1319 // undef << X -> undef if (if it's NSW/NUW)
1320 if (match(Op0, m_Undef()))
1321 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1323 // (X >> A) << A -> X
1325 if (Q.IIQ.UseInstrInfo &&
1326 match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1329 // shl nuw i8 C, %x -> C iff C has sign bit set.
1330 if (isNUW && match(Op0, m_Negative()))
1332 // NOTE: could use computeKnownBits() / LazyValueInfo,
1333 // but the cost-benefit analysis suggests it isn't worth it.
1338 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1339 const SimplifyQuery &Q) {
1340 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1343 /// Given operands for an LShr, see if we can fold the result.
1344 /// If not, this returns null.
1345 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1346 const SimplifyQuery &Q, unsigned MaxRecurse) {
1347 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1351 // (X << A) >> A -> X
1353 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1356 // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1357 // We can return X as we do in the above case since OR alters no bits in X.
1358 // SimplifyDemandedBits in InstCombine can do more general optimization for
1359 // bit manipulation. This pattern aims to provide opportunities for other
1360 // optimizers by supporting a simple but common case in InstSimplify.
1362 const APInt *ShRAmt, *ShLAmt;
1363 if (match(Op1, m_APInt(ShRAmt)) &&
1364 match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1365 *ShRAmt == *ShLAmt) {
1366 const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1367 const unsigned Width = Op0->getType()->getScalarSizeInBits();
1368 const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1369 if (ShRAmt->uge(EffWidthY))
1376 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1377 const SimplifyQuery &Q) {
1378 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1381 /// Given operands for an AShr, see if we can fold the result.
1382 /// If not, this returns null.
1383 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1384 const SimplifyQuery &Q, unsigned MaxRecurse) {
1385 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1389 // all ones >>a X -> -1
1390 // Do not return Op0 because it may contain undef elements if it's a vector.
1391 if (match(Op0, m_AllOnes()))
1392 return Constant::getAllOnesValue(Op0->getType());
1394 // (X << A) >> A -> X
1396 if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1399 // Arithmetic shifting an all-sign-bit value is a no-op.
1400 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1401 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1407 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1408 const SimplifyQuery &Q) {
1409 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1412 /// Commuted variants are assumed to be handled by calling this function again
1413 /// with the parameters swapped.
1414 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1415 ICmpInst *UnsignedICmp, bool IsAnd,
1416 const SimplifyQuery &Q) {
1419 ICmpInst::Predicate EqPred;
1420 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1421 !ICmpInst::isEquality(EqPred))
1424 ICmpInst::Predicate UnsignedPred;
1428 if (match(Y, m_Sub(m_Value(A), m_Value(B)))) {
1429 if (match(UnsignedICmp,
1430 m_c_ICmp(UnsignedPred, m_Specific(A), m_Specific(B))) &&
1431 ICmpInst::isUnsigned(UnsignedPred)) {
1432 if (UnsignedICmp->getOperand(0) != A)
1433 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1435 // A >=/<= B || (A - B) != 0 <--> true
1436 if ((UnsignedPred == ICmpInst::ICMP_UGE ||
1437 UnsignedPred == ICmpInst::ICMP_ULE) &&
1438 EqPred == ICmpInst::ICMP_NE && !IsAnd)
1439 return ConstantInt::getTrue(UnsignedICmp->getType());
1440 // A </> B && (A - B) == 0 <--> false
1441 if ((UnsignedPred == ICmpInst::ICMP_ULT ||
1442 UnsignedPred == ICmpInst::ICMP_UGT) &&
1443 EqPred == ICmpInst::ICMP_EQ && IsAnd)
1444 return ConstantInt::getFalse(UnsignedICmp->getType());
1446 // A </> B && (A - B) != 0 <--> A </> B
1447 // A </> B || (A - B) != 0 <--> (A - B) != 0
1448 if (EqPred == ICmpInst::ICMP_NE && (UnsignedPred == ICmpInst::ICMP_ULT ||
1449 UnsignedPred == ICmpInst::ICMP_UGT))
1450 return IsAnd ? UnsignedICmp : ZeroICmp;
1452 // A <=/>= B && (A - B) == 0 <--> (A - B) == 0
1453 // A <=/>= B || (A - B) == 0 <--> A <=/>= B
1454 if (EqPred == ICmpInst::ICMP_EQ && (UnsignedPred == ICmpInst::ICMP_ULE ||
1455 UnsignedPred == ICmpInst::ICMP_UGE))
1456 return IsAnd ? ZeroICmp : UnsignedICmp;
1459 // Given Y = (A - B)
1460 // Y >= A && Y != 0 --> Y >= A iff B != 0
1461 // Y < A || Y == 0 --> Y < A iff B != 0
1462 if (match(UnsignedICmp,
1463 m_c_ICmp(UnsignedPred, m_Specific(Y), m_Specific(A)))) {
1464 if (UnsignedICmp->getOperand(0) != Y)
1465 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1467 if (UnsignedPred == ICmpInst::ICMP_UGE && IsAnd &&
1468 EqPred == ICmpInst::ICMP_NE &&
1469 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1470 return UnsignedICmp;
1471 if (UnsignedPred == ICmpInst::ICMP_ULT && !IsAnd &&
1472 EqPred == ICmpInst::ICMP_EQ &&
1473 isKnownNonZero(B, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1474 return UnsignedICmp;
1478 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1479 ICmpInst::isUnsigned(UnsignedPred))
1481 else if (match(UnsignedICmp,
1482 m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1483 ICmpInst::isUnsigned(UnsignedPred))
1484 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1488 // X < Y && Y != 0 --> X < Y
1489 // X < Y || Y != 0 --> Y != 0
1490 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1491 return IsAnd ? UnsignedICmp : ZeroICmp;
1493 // X <= Y && Y != 0 --> X <= Y iff X != 0
1494 // X <= Y || Y != 0 --> Y != 0 iff X != 0
1495 if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE &&
1496 isKnownNonZero(X, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1497 return IsAnd ? UnsignedICmp : ZeroICmp;
1499 // X >= Y && Y == 0 --> Y == 0
1500 // X >= Y || Y == 0 --> X >= Y
1501 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ)
1502 return IsAnd ? ZeroICmp : UnsignedICmp;
1504 // X > Y && Y == 0 --> Y == 0 iff X != 0
1505 // X > Y || Y == 0 --> X > Y iff X != 0
1506 if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ &&
1507 isKnownNonZero(X, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT))
1508 return IsAnd ? ZeroICmp : UnsignedICmp;
1510 // X < Y && Y == 0 --> false
1511 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1513 return getFalse(UnsignedICmp->getType());
1515 // X >= Y || Y != 0 --> true
1516 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_NE &&
1518 return getTrue(UnsignedICmp->getType());
1523 /// Commuted variants are assumed to be handled by calling this function again
1524 /// with the parameters swapped.
1525 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1526 ICmpInst::Predicate Pred0, Pred1;
1528 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1529 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1532 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1533 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1534 // can eliminate Op1 from this 'and'.
1535 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1538 // Check for any combination of predicates that are guaranteed to be disjoint.
1539 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1540 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1541 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1542 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1543 return getFalse(Op0->getType());
1548 /// Commuted variants are assumed to be handled by calling this function again
1549 /// with the parameters swapped.
1550 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1551 ICmpInst::Predicate Pred0, Pred1;
1553 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1554 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1557 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1558 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1559 // can eliminate Op0 from this 'or'.
1560 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1563 // Check for any combination of predicates that cover the entire range of
1565 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1566 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1567 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1568 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1569 return getTrue(Op0->getType());
1574 /// Test if a pair of compares with a shared operand and 2 constants has an
1575 /// empty set intersection, full set union, or if one compare is a superset of
1577 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1579 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1580 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1583 const APInt *C0, *C1;
1584 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1585 !match(Cmp1->getOperand(1), m_APInt(C1)))
1588 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1589 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1591 // For and-of-compares, check if the intersection is empty:
1592 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1593 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1594 return getFalse(Cmp0->getType());
1596 // For or-of-compares, check if the union is full:
1597 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1598 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1599 return getTrue(Cmp0->getType());
1601 // Is one range a superset of the other?
1602 // If this is and-of-compares, take the smaller set:
1603 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1604 // If this is or-of-compares, take the larger set:
1605 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1606 if (Range0.contains(Range1))
1607 return IsAnd ? Cmp1 : Cmp0;
1608 if (Range1.contains(Range0))
1609 return IsAnd ? Cmp0 : Cmp1;
1614 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1616 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1617 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1618 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1621 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1624 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1625 Value *X = Cmp0->getOperand(0);
1626 Value *Y = Cmp1->getOperand(0);
1628 // If one of the compares is a masked version of a (not) null check, then
1629 // that compare implies the other, so we eliminate the other. Optionally, look
1630 // through a pointer-to-int cast to match a null check of a pointer type.
1632 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1633 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1634 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1635 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1636 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1637 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1640 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1641 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1642 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1643 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1644 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1645 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1651 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1,
1652 const InstrInfoQuery &IIQ) {
1653 // (icmp (add V, C0), C1) & (icmp V, C0)
1654 ICmpInst::Predicate Pred0, Pred1;
1655 const APInt *C0, *C1;
1657 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1660 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1663 auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1664 if (AddInst->getOperand(1) != Op1->getOperand(1))
1667 Type *ITy = Op0->getType();
1668 bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1669 bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1671 const APInt Delta = *C1 - *C0;
1672 if (C0->isStrictlyPositive()) {
1674 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1675 return getFalse(ITy);
1676 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1677 return getFalse(ITy);
1680 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1681 return getFalse(ITy);
1682 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1683 return getFalse(ITy);
1686 if (C0->getBoolValue() && isNUW) {
1688 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1689 return getFalse(ITy);
1691 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1692 return getFalse(ITy);
1698 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1699 const SimplifyQuery &Q) {
1700 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true, Q))
1702 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true, Q))
1705 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1707 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1710 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1713 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1716 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, Q.IIQ))
1718 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, Q.IIQ))
1724 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1,
1725 const InstrInfoQuery &IIQ) {
1726 // (icmp (add V, C0), C1) | (icmp V, C0)
1727 ICmpInst::Predicate Pred0, Pred1;
1728 const APInt *C0, *C1;
1730 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1733 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1736 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1737 if (AddInst->getOperand(1) != Op1->getOperand(1))
1740 Type *ITy = Op0->getType();
1741 bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1742 bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1744 const APInt Delta = *C1 - *C0;
1745 if (C0->isStrictlyPositive()) {
1747 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1748 return getTrue(ITy);
1749 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1750 return getTrue(ITy);
1753 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1754 return getTrue(ITy);
1755 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1756 return getTrue(ITy);
1759 if (C0->getBoolValue() && isNUW) {
1761 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1762 return getTrue(ITy);
1764 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1765 return getTrue(ITy);
1771 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1772 const SimplifyQuery &Q) {
1773 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false, Q))
1775 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false, Q))
1778 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1780 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1783 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1786 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1789 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, Q.IIQ))
1791 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, Q.IIQ))
1797 static Value *simplifyAndOrOfFCmps(const TargetLibraryInfo *TLI,
1798 FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1799 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1800 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1801 if (LHS0->getType() != RHS0->getType())
1804 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1805 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1806 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1807 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1808 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1809 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1810 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1811 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1812 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1813 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1814 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1815 if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1816 (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1819 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1820 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1821 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1822 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1823 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1824 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1825 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1826 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1827 if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1828 (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1835 static Value *simplifyAndOrOfCmps(const SimplifyQuery &Q,
1836 Value *Op0, Value *Op1, bool IsAnd) {
1837 // Look through casts of the 'and' operands to find compares.
1838 auto *Cast0 = dyn_cast<CastInst>(Op0);
1839 auto *Cast1 = dyn_cast<CastInst>(Op1);
1840 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1841 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1842 Op0 = Cast0->getOperand(0);
1843 Op1 = Cast1->getOperand(0);
1847 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1848 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1850 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q)
1851 : simplifyOrOfICmps(ICmp0, ICmp1, Q);
1853 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1854 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1856 V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1863 // If we looked through casts, we can only handle a constant simplification
1864 // because we are not allowed to create a cast instruction here.
1865 if (auto *C = dyn_cast<Constant>(V))
1866 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1871 /// Check that the Op1 is in expected form, i.e.:
1872 /// %Agg = tail call { i4, i1 } @llvm.[us]mul.with.overflow.i4(i4 %X, i4 %???)
1873 /// %Op1 = extractvalue { i4, i1 } %Agg, 1
1874 static bool omitCheckForZeroBeforeMulWithOverflowInternal(Value *Op1,
1876 auto *Extract = dyn_cast<ExtractValueInst>(Op1);
1877 // We should only be extracting the overflow bit.
1878 if (!Extract || !Extract->getIndices().equals(1))
1880 Value *Agg = Extract->getAggregateOperand();
1881 // This should be a multiplication-with-overflow intrinsic.
1882 if (!match(Agg, m_CombineOr(m_Intrinsic<Intrinsic::umul_with_overflow>(),
1883 m_Intrinsic<Intrinsic::smul_with_overflow>())))
1885 // One of its multipliers should be the value we checked for zero before.
1886 if (!match(Agg, m_CombineOr(m_Argument<0>(m_Specific(X)),
1887 m_Argument<1>(m_Specific(X)))))
1892 /// The @llvm.[us]mul.with.overflow intrinsic could have been folded from some
1893 /// other form of check, e.g. one that was using division; it may have been
1894 /// guarded against division-by-zero. We can drop that check now.
1896 /// %Op0 = icmp ne i4 %X, 0
1897 /// %Agg = tail call { i4, i1 } @llvm.[us]mul.with.overflow.i4(i4 %X, i4 %???)
1898 /// %Op1 = extractvalue { i4, i1 } %Agg, 1
1899 /// %??? = and i1 %Op0, %Op1
1900 /// We can just return %Op1
1901 static Value *omitCheckForZeroBeforeMulWithOverflow(Value *Op0, Value *Op1) {
1902 ICmpInst::Predicate Pred;
1904 if (!match(Op0, m_ICmp(Pred, m_Value(X), m_Zero())) ||
1905 Pred != ICmpInst::Predicate::ICMP_NE)
1907 // Is Op1 in expected form?
1908 if (!omitCheckForZeroBeforeMulWithOverflowInternal(Op1, X))
1910 // Can omit 'and', and just return the overflow bit.
1914 /// The @llvm.[us]mul.with.overflow intrinsic could have been folded from some
1915 /// other form of check, e.g. one that was using division; it may have been
1916 /// guarded against division-by-zero. We can drop that check now.
1918 /// %Op0 = icmp eq i4 %X, 0
1919 /// %Agg = tail call { i4, i1 } @llvm.[us]mul.with.overflow.i4(i4 %X, i4 %???)
1920 /// %Op1 = extractvalue { i4, i1 } %Agg, 1
1921 /// %NotOp1 = xor i1 %Op1, true
1922 /// %or = or i1 %Op0, %NotOp1
1923 /// We can just return %NotOp1
1924 static Value *omitCheckForZeroBeforeInvertedMulWithOverflow(Value *Op0,
1926 ICmpInst::Predicate Pred;
1928 if (!match(Op0, m_ICmp(Pred, m_Value(X), m_Zero())) ||
1929 Pred != ICmpInst::Predicate::ICMP_EQ)
1931 // We expect the other hand of an 'or' to be a 'not'.
1933 if (!match(NotOp1, m_Not(m_Value(Op1))))
1935 // Is Op1 in expected form?
1936 if (!omitCheckForZeroBeforeMulWithOverflowInternal(Op1, X))
1938 // Can omit 'and', and just return the inverted overflow bit.
1942 /// Given operands for an And, see if we can fold the result.
1943 /// If not, this returns null.
1944 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1945 unsigned MaxRecurse) {
1946 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1950 if (match(Op1, m_Undef()))
1951 return Constant::getNullValue(Op0->getType());
1958 if (match(Op1, m_Zero()))
1959 return Constant::getNullValue(Op0->getType());
1962 if (match(Op1, m_AllOnes()))
1965 // A & ~A = ~A & A = 0
1966 if (match(Op0, m_Not(m_Specific(Op1))) ||
1967 match(Op1, m_Not(m_Specific(Op0))))
1968 return Constant::getNullValue(Op0->getType());
1971 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1975 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1978 // A mask that only clears known zeros of a shifted value is a no-op.
1982 if (match(Op1, m_APInt(Mask))) {
1983 // If all bits in the inverted and shifted mask are clear:
1984 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1985 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1986 (~(*Mask)).lshr(*ShAmt).isNullValue())
1989 // If all bits in the inverted and shifted mask are clear:
1990 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1991 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1992 (~(*Mask)).shl(*ShAmt).isNullValue())
1996 // If we have a multiplication overflow check that is being 'and'ed with a
1997 // check that one of the multipliers is not zero, we can omit the 'and', and
1998 // only keep the overflow check.
1999 if (Value *V = omitCheckForZeroBeforeMulWithOverflow(Op0, Op1))
2001 if (Value *V = omitCheckForZeroBeforeMulWithOverflow(Op1, Op0))
2004 // A & (-A) = A if A is a power of two or zero.
2005 if (match(Op0, m_Neg(m_Specific(Op1))) ||
2006 match(Op1, m_Neg(m_Specific(Op0)))) {
2007 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
2010 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
2015 // This is a similar pattern used for checking if a value is a power-of-2:
2016 // (A - 1) & A --> 0 (if A is a power-of-2 or 0)
2017 // A & (A - 1) --> 0 (if A is a power-of-2 or 0)
2018 if (match(Op0, m_Add(m_Specific(Op1), m_AllOnes())) &&
2019 isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
2020 return Constant::getNullValue(Op1->getType());
2021 if (match(Op1, m_Add(m_Specific(Op0), m_AllOnes())) &&
2022 isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
2023 return Constant::getNullValue(Op0->getType());
2025 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
2028 // Try some generic simplifications for associative operations.
2029 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
2033 // And distributes over Or. Try some generic simplifications based on this.
2034 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
2038 // And distributes over Xor. Try some generic simplifications based on this.
2039 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
2043 // If the operation is with the result of a select instruction, check whether
2044 // operating on either branch of the select always yields the same value.
2045 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2046 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
2050 // If the operation is with the result of a phi instruction, check whether
2051 // operating on all incoming values of the phi always yields the same value.
2052 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2053 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
2057 // Assuming the effective width of Y is not larger than A, i.e. all bits
2058 // from X and Y are disjoint in (X << A) | Y,
2059 // if the mask of this AND op covers all bits of X or Y, while it covers
2060 // no bits from the other, we can bypass this AND op. E.g.,
2061 // ((X << A) | Y) & Mask -> Y,
2062 // if Mask = ((1 << effective_width_of(Y)) - 1)
2063 // ((X << A) | Y) & Mask -> X << A,
2064 // if Mask = ((1 << effective_width_of(X)) - 1) << A
2065 // SimplifyDemandedBits in InstCombine can optimize the general case.
2066 // This pattern aims to help other passes for a common case.
2067 Value *Y, *XShifted;
2068 if (match(Op1, m_APInt(Mask)) &&
2069 match(Op0, m_c_Or(m_CombineAnd(m_NUWShl(m_Value(X), m_APInt(ShAmt)),
2072 const unsigned Width = Op0->getType()->getScalarSizeInBits();
2073 const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
2074 const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2075 const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
2076 if (EffWidthY <= ShftCnt) {
2077 const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
2079 const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
2080 const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
2081 const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
2082 // If the mask is extracting all bits from X or Y as is, we can skip
2084 if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
2086 if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
2094 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2095 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
2098 /// Given operands for an Or, see if we can fold the result.
2099 /// If not, this returns null.
2100 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2101 unsigned MaxRecurse) {
2102 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
2107 // Do not return Op1 because it may contain undef elements if it's a vector.
2108 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
2109 return Constant::getAllOnesValue(Op0->getType());
2113 if (Op0 == Op1 || match(Op1, m_Zero()))
2116 // A | ~A = ~A | A = -1
2117 if (match(Op0, m_Not(m_Specific(Op1))) ||
2118 match(Op1, m_Not(m_Specific(Op0))))
2119 return Constant::getAllOnesValue(Op0->getType());
2122 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
2126 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
2129 // ~(A & ?) | A = -1
2130 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
2131 return Constant::getAllOnesValue(Op1->getType());
2133 // A | ~(A & ?) = -1
2134 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
2135 return Constant::getAllOnesValue(Op0->getType());
2138 // (A & ~B) | (A ^ B) -> (A ^ B)
2139 // (~B & A) | (A ^ B) -> (A ^ B)
2140 // (A & ~B) | (B ^ A) -> (B ^ A)
2141 // (~B & A) | (B ^ A) -> (B ^ A)
2142 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2143 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
2144 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
2147 // Commute the 'or' operands.
2148 // (A ^ B) | (A & ~B) -> (A ^ B)
2149 // (A ^ B) | (~B & A) -> (A ^ B)
2150 // (B ^ A) | (A & ~B) -> (B ^ A)
2151 // (B ^ A) | (~B & A) -> (B ^ A)
2152 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2153 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
2154 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
2157 // (A & B) | (~A ^ B) -> (~A ^ B)
2158 // (B & A) | (~A ^ B) -> (~A ^ B)
2159 // (A & B) | (B ^ ~A) -> (B ^ ~A)
2160 // (B & A) | (B ^ ~A) -> (B ^ ~A)
2161 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2162 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
2163 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
2166 // (~A ^ B) | (A & B) -> (~A ^ B)
2167 // (~A ^ B) | (B & A) -> (~A ^ B)
2168 // (B ^ ~A) | (A & B) -> (B ^ ~A)
2169 // (B ^ ~A) | (B & A) -> (B ^ ~A)
2170 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
2171 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
2172 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
2175 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
2178 // If we have a multiplication overflow check that is being 'and'ed with a
2179 // check that one of the multipliers is not zero, we can omit the 'and', and
2180 // only keep the overflow check.
2181 if (Value *V = omitCheckForZeroBeforeInvertedMulWithOverflow(Op0, Op1))
2183 if (Value *V = omitCheckForZeroBeforeInvertedMulWithOverflow(Op1, Op0))
2186 // Try some generic simplifications for associative operations.
2187 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
2191 // Or distributes over And. Try some generic simplifications based on this.
2192 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
2196 // If the operation is with the result of a select instruction, check whether
2197 // operating on either branch of the select always yields the same value.
2198 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2199 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2203 // (A & C1)|(B & C2)
2204 const APInt *C1, *C2;
2205 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2206 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2208 // (A & C1)|(B & C2)
2209 // If we have: ((V + N) & C1) | (V & C2)
2210 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2211 // replace with V+N.
2213 if (C2->isMask() && // C2 == 0+1+
2214 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2215 // Add commutes, try both ways.
2216 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2219 // Or commutes, try both ways.
2221 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2222 // Add commutes, try both ways.
2223 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2229 // If the operation is with the result of a phi instruction, check whether
2230 // operating on all incoming values of the phi always yields the same value.
2231 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2232 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2238 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2239 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
2242 /// Given operands for a Xor, see if we can fold the result.
2243 /// If not, this returns null.
2244 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2245 unsigned MaxRecurse) {
2246 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2249 // A ^ undef -> undef
2250 if (match(Op1, m_Undef()))
2254 if (match(Op1, m_Zero()))
2259 return Constant::getNullValue(Op0->getType());
2261 // A ^ ~A = ~A ^ A = -1
2262 if (match(Op0, m_Not(m_Specific(Op1))) ||
2263 match(Op1, m_Not(m_Specific(Op0))))
2264 return Constant::getAllOnesValue(Op0->getType());
2266 // Try some generic simplifications for associative operations.
2267 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2271 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2272 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2273 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2274 // only if B and C are equal. If B and C are equal then (since we assume
2275 // that operands have already been simplified) "select(cond, B, C)" should
2276 // have been simplified to the common value of B and C already. Analysing
2277 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2278 // for threading over phi nodes.
2283 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2284 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2288 static Type *GetCompareTy(Value *Op) {
2289 return CmpInst::makeCmpResultType(Op->getType());
2292 /// Rummage around inside V looking for something equivalent to the comparison
2293 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2294 /// Helper function for analyzing max/min idioms.
2295 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2296 Value *LHS, Value *RHS) {
2297 SelectInst *SI = dyn_cast<SelectInst>(V);
2300 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2303 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2304 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2306 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2307 LHS == CmpRHS && RHS == CmpLHS)
2312 // A significant optimization not implemented here is assuming that alloca
2313 // addresses are not equal to incoming argument values. They don't *alias*,
2314 // as we say, but that doesn't mean they aren't equal, so we take a
2315 // conservative approach.
2317 // This is inspired in part by C++11 5.10p1:
2318 // "Two pointers of the same type compare equal if and only if they are both
2319 // null, both point to the same function, or both represent the same
2322 // This is pretty permissive.
2324 // It's also partly due to C11 6.5.9p6:
2325 // "Two pointers compare equal if and only if both are null pointers, both are
2326 // pointers to the same object (including a pointer to an object and a
2327 // subobject at its beginning) or function, both are pointers to one past the
2328 // last element of the same array object, or one is a pointer to one past the
2329 // end of one array object and the other is a pointer to the start of a
2330 // different array object that happens to immediately follow the first array
2331 // object in the address space.)
2333 // C11's version is more restrictive, however there's no reason why an argument
2334 // couldn't be a one-past-the-end value for a stack object in the caller and be
2335 // equal to the beginning of a stack object in the callee.
2337 // If the C and C++ standards are ever made sufficiently restrictive in this
2338 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2339 // this optimization.
2341 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2342 const DominatorTree *DT, CmpInst::Predicate Pred,
2343 AssumptionCache *AC, const Instruction *CxtI,
2344 const InstrInfoQuery &IIQ, Value *LHS, Value *RHS) {
2345 // First, skip past any trivial no-ops.
2346 LHS = LHS->stripPointerCasts();
2347 RHS = RHS->stripPointerCasts();
2349 // A non-null pointer is not equal to a null pointer.
2350 if (llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2351 IIQ.UseInstrInfo) &&
2352 isa<ConstantPointerNull>(RHS) &&
2353 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2354 return ConstantInt::get(GetCompareTy(LHS),
2355 !CmpInst::isTrueWhenEqual(Pred));
2357 // We can only fold certain predicates on pointer comparisons.
2362 // Equality comaprisons are easy to fold.
2363 case CmpInst::ICMP_EQ:
2364 case CmpInst::ICMP_NE:
2367 // We can only handle unsigned relational comparisons because 'inbounds' on
2368 // a GEP only protects against unsigned wrapping.
2369 case CmpInst::ICMP_UGT:
2370 case CmpInst::ICMP_UGE:
2371 case CmpInst::ICMP_ULT:
2372 case CmpInst::ICMP_ULE:
2373 // However, we have to switch them to their signed variants to handle
2374 // negative indices from the base pointer.
2375 Pred = ICmpInst::getSignedPredicate(Pred);
2379 // Strip off any constant offsets so that we can reason about them.
2380 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2381 // here and compare base addresses like AliasAnalysis does, however there are
2382 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2383 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2384 // doesn't need to guarantee pointer inequality when it says NoAlias.
2385 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2386 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2388 // If LHS and RHS are related via constant offsets to the same base
2389 // value, we can replace it with an icmp which just compares the offsets.
2391 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2393 // Various optimizations for (in)equality comparisons.
2394 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2395 // Different non-empty allocations that exist at the same time have
2396 // different addresses (if the program can tell). Global variables always
2397 // exist, so they always exist during the lifetime of each other and all
2398 // allocas. Two different allocas usually have different addresses...
2400 // However, if there's an @llvm.stackrestore dynamically in between two
2401 // allocas, they may have the same address. It's tempting to reduce the
2402 // scope of the problem by only looking at *static* allocas here. That would
2403 // cover the majority of allocas while significantly reducing the likelihood
2404 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2405 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2406 // an entry block. Also, if we have a block that's not attached to a
2407 // function, we can't tell if it's "static" under the current definition.
2408 // Theoretically, this problem could be fixed by creating a new kind of
2409 // instruction kind specifically for static allocas. Such a new instruction
2410 // could be required to be at the top of the entry block, thus preventing it
2411 // from being subject to a @llvm.stackrestore. Instcombine could even
2412 // convert regular allocas into these special allocas. It'd be nifty.
2413 // However, until then, this problem remains open.
2415 // So, we'll assume that two non-empty allocas have different addresses
2418 // With all that, if the offsets are within the bounds of their allocations
2419 // (and not one-past-the-end! so we can't use inbounds!), and their
2420 // allocations aren't the same, the pointers are not equal.
2422 // Note that it's not necessary to check for LHS being a global variable
2423 // address, due to canonicalization and constant folding.
2424 if (isa<AllocaInst>(LHS) &&
2425 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2426 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2427 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2428 uint64_t LHSSize, RHSSize;
2429 ObjectSizeOpts Opts;
2430 Opts.NullIsUnknownSize =
2431 NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2432 if (LHSOffsetCI && RHSOffsetCI &&
2433 getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2434 getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2435 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2436 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2437 if (!LHSOffsetValue.isNegative() &&
2438 !RHSOffsetValue.isNegative() &&
2439 LHSOffsetValue.ult(LHSSize) &&
2440 RHSOffsetValue.ult(RHSSize)) {
2441 return ConstantInt::get(GetCompareTy(LHS),
2442 !CmpInst::isTrueWhenEqual(Pred));
2446 // Repeat the above check but this time without depending on DataLayout
2447 // or being able to compute a precise size.
2448 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2449 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2450 LHSOffset->isNullValue() &&
2451 RHSOffset->isNullValue())
2452 return ConstantInt::get(GetCompareTy(LHS),
2453 !CmpInst::isTrueWhenEqual(Pred));
2456 // Even if an non-inbounds GEP occurs along the path we can still optimize
2457 // equality comparisons concerning the result. We avoid walking the whole
2458 // chain again by starting where the last calls to
2459 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2460 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2461 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2463 return ConstantExpr::getICmp(Pred,
2464 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2465 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2467 // If one side of the equality comparison must come from a noalias call
2468 // (meaning a system memory allocation function), and the other side must
2469 // come from a pointer that cannot overlap with dynamically-allocated
2470 // memory within the lifetime of the current function (allocas, byval
2471 // arguments, globals), then determine the comparison result here.
2472 SmallVector<const Value *, 8> LHSUObjs, RHSUObjs;
2473 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2474 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2476 // Is the set of underlying objects all noalias calls?
2477 auto IsNAC = [](ArrayRef<const Value *> Objects) {
2478 return all_of(Objects, isNoAliasCall);
2481 // Is the set of underlying objects all things which must be disjoint from
2482 // noalias calls. For allocas, we consider only static ones (dynamic
2483 // allocas might be transformed into calls to malloc not simultaneously
2484 // live with the compared-to allocation). For globals, we exclude symbols
2485 // that might be resolve lazily to symbols in another dynamically-loaded
2486 // library (and, thus, could be malloc'ed by the implementation).
2487 auto IsAllocDisjoint = [](ArrayRef<const Value *> Objects) {
2488 return all_of(Objects, [](const Value *V) {
2489 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2490 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2491 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2492 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2493 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2494 !GV->isThreadLocal();
2495 if (const Argument *A = dyn_cast<Argument>(V))
2496 return A->hasByValAttr();
2501 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2502 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2503 return ConstantInt::get(GetCompareTy(LHS),
2504 !CmpInst::isTrueWhenEqual(Pred));
2506 // Fold comparisons for non-escaping pointer even if the allocation call
2507 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2508 // dynamic allocation call could be either of the operands.
2509 Value *MI = nullptr;
2510 if (isAllocLikeFn(LHS, TLI) &&
2511 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2513 else if (isAllocLikeFn(RHS, TLI) &&
2514 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2516 // FIXME: We should also fold the compare when the pointer escapes, but the
2517 // compare dominates the pointer escape
2518 if (MI && !PointerMayBeCaptured(MI, true, true))
2519 return ConstantInt::get(GetCompareTy(LHS),
2520 CmpInst::isFalseWhenEqual(Pred));
2527 /// Fold an icmp when its operands have i1 scalar type.
2528 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2529 Value *RHS, const SimplifyQuery &Q) {
2530 Type *ITy = GetCompareTy(LHS); // The return type.
2531 Type *OpTy = LHS->getType(); // The operand type.
2532 if (!OpTy->isIntOrIntVectorTy(1))
2535 // A boolean compared to true/false can be simplified in 14 out of the 20
2536 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2537 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2538 if (match(RHS, m_Zero())) {
2540 case CmpInst::ICMP_NE: // X != 0 -> X
2541 case CmpInst::ICMP_UGT: // X >u 0 -> X
2542 case CmpInst::ICMP_SLT: // X <s 0 -> X
2545 case CmpInst::ICMP_ULT: // X <u 0 -> false
2546 case CmpInst::ICMP_SGT: // X >s 0 -> false
2547 return getFalse(ITy);
2549 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2550 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2551 return getTrue(ITy);
2555 } else if (match(RHS, m_One())) {
2557 case CmpInst::ICMP_EQ: // X == 1 -> X
2558 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2559 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2562 case CmpInst::ICMP_UGT: // X >u 1 -> false
2563 case CmpInst::ICMP_SLT: // X <s -1 -> false
2564 return getFalse(ITy);
2566 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2567 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2568 return getTrue(ITy);
2577 case ICmpInst::ICMP_UGE:
2578 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2579 return getTrue(ITy);
2581 case ICmpInst::ICMP_SGE:
2582 /// For signed comparison, the values for an i1 are 0 and -1
2583 /// respectively. This maps into a truth table of:
2584 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2585 /// 0 | 0 | 1 (0 >= 0) | 1
2586 /// 0 | 1 | 1 (0 >= -1) | 1
2587 /// 1 | 0 | 0 (-1 >= 0) | 0
2588 /// 1 | 1 | 1 (-1 >= -1) | 1
2589 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2590 return getTrue(ITy);
2592 case ICmpInst::ICMP_ULE:
2593 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2594 return getTrue(ITy);
2601 /// Try hard to fold icmp with zero RHS because this is a common case.
2602 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2603 Value *RHS, const SimplifyQuery &Q) {
2604 if (!match(RHS, m_Zero()))
2607 Type *ITy = GetCompareTy(LHS); // The return type.
2610 llvm_unreachable("Unknown ICmp predicate!");
2611 case ICmpInst::ICMP_ULT:
2612 return getFalse(ITy);
2613 case ICmpInst::ICMP_UGE:
2614 return getTrue(ITy);
2615 case ICmpInst::ICMP_EQ:
2616 case ICmpInst::ICMP_ULE:
2617 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2618 return getFalse(ITy);
2620 case ICmpInst::ICMP_NE:
2621 case ICmpInst::ICMP_UGT:
2622 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2623 return getTrue(ITy);
2625 case ICmpInst::ICMP_SLT: {
2626 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2627 if (LHSKnown.isNegative())
2628 return getTrue(ITy);
2629 if (LHSKnown.isNonNegative())
2630 return getFalse(ITy);
2633 case ICmpInst::ICMP_SLE: {
2634 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2635 if (LHSKnown.isNegative())
2636 return getTrue(ITy);
2637 if (LHSKnown.isNonNegative() &&
2638 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2639 return getFalse(ITy);
2642 case ICmpInst::ICMP_SGE: {
2643 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2644 if (LHSKnown.isNegative())
2645 return getFalse(ITy);
2646 if (LHSKnown.isNonNegative())
2647 return getTrue(ITy);
2650 case ICmpInst::ICMP_SGT: {
2651 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2652 if (LHSKnown.isNegative())
2653 return getFalse(ITy);
2654 if (LHSKnown.isNonNegative() &&
2655 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2656 return getTrue(ITy);
2664 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2665 Value *RHS, const InstrInfoQuery &IIQ) {
2666 Type *ITy = GetCompareTy(RHS); // The return type.
2669 // Sign-bit checks can be optimized to true/false after unsigned
2670 // floating-point casts:
2671 // icmp slt (bitcast (uitofp X)), 0 --> false
2672 // icmp sgt (bitcast (uitofp X)), -1 --> true
2673 if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2674 if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2675 return ConstantInt::getFalse(ITy);
2676 if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2677 return ConstantInt::getTrue(ITy);
2681 if (!match(RHS, m_APInt(C)))
2684 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2685 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2686 if (RHS_CR.isEmptySet())
2687 return ConstantInt::getFalse(ITy);
2688 if (RHS_CR.isFullSet())
2689 return ConstantInt::getTrue(ITy);
2691 ConstantRange LHS_CR = computeConstantRange(LHS, IIQ.UseInstrInfo);
2692 if (!LHS_CR.isFullSet()) {
2693 if (RHS_CR.contains(LHS_CR))
2694 return ConstantInt::getTrue(ITy);
2695 if (RHS_CR.inverse().contains(LHS_CR))
2696 return ConstantInt::getFalse(ITy);
2702 /// TODO: A large part of this logic is duplicated in InstCombine's
2703 /// foldICmpBinOp(). We should be able to share that and avoid the code
2705 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2706 Value *RHS, const SimplifyQuery &Q,
2707 unsigned MaxRecurse) {
2708 Type *ITy = GetCompareTy(LHS); // The return type.
2710 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2711 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2712 if (MaxRecurse && (LBO || RBO)) {
2713 // Analyze the case when either LHS or RHS is an add instruction.
2714 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2715 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2716 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2717 if (LBO && LBO->getOpcode() == Instruction::Add) {
2718 A = LBO->getOperand(0);
2719 B = LBO->getOperand(1);
2721 ICmpInst::isEquality(Pred) ||
2722 (CmpInst::isUnsigned(Pred) &&
2723 Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
2724 (CmpInst::isSigned(Pred) &&
2725 Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
2727 if (RBO && RBO->getOpcode() == Instruction::Add) {
2728 C = RBO->getOperand(0);
2729 D = RBO->getOperand(1);
2731 ICmpInst::isEquality(Pred) ||
2732 (CmpInst::isUnsigned(Pred) &&
2733 Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
2734 (CmpInst::isSigned(Pred) &&
2735 Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
2738 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2739 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2740 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2741 Constant::getNullValue(RHS->getType()), Q,
2745 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2746 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2748 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2749 C == LHS ? D : C, Q, MaxRecurse - 1))
2752 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2753 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2755 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2758 // C + B == C + D -> B == D
2761 } else if (A == D) {
2762 // D + B == C + D -> B == C
2765 } else if (B == C) {
2766 // A + C == C + D -> A == D
2771 // A + D == C + D -> A == C
2775 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2782 // icmp pred (or X, Y), X
2783 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2784 if (Pred == ICmpInst::ICMP_ULT)
2785 return getFalse(ITy);
2786 if (Pred == ICmpInst::ICMP_UGE)
2787 return getTrue(ITy);
2789 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2790 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2791 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2792 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2793 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2794 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2795 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2798 // icmp pred X, (or X, Y)
2799 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2800 if (Pred == ICmpInst::ICMP_ULE)
2801 return getTrue(ITy);
2802 if (Pred == ICmpInst::ICMP_UGT)
2803 return getFalse(ITy);
2805 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2806 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2807 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2808 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2809 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2810 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2811 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2816 // icmp pred (and X, Y), X
2817 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2818 if (Pred == ICmpInst::ICMP_UGT)
2819 return getFalse(ITy);
2820 if (Pred == ICmpInst::ICMP_ULE)
2821 return getTrue(ITy);
2823 // icmp pred X, (and X, Y)
2824 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2825 if (Pred == ICmpInst::ICMP_UGE)
2826 return getTrue(ITy);
2827 if (Pred == ICmpInst::ICMP_ULT)
2828 return getFalse(ITy);
2831 // 0 - (zext X) pred C
2832 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2833 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2834 if (RHSC->getValue().isStrictlyPositive()) {
2835 if (Pred == ICmpInst::ICMP_SLT)
2836 return ConstantInt::getTrue(RHSC->getContext());
2837 if (Pred == ICmpInst::ICMP_SGE)
2838 return ConstantInt::getFalse(RHSC->getContext());
2839 if (Pred == ICmpInst::ICMP_EQ)
2840 return ConstantInt::getFalse(RHSC->getContext());
2841 if (Pred == ICmpInst::ICMP_NE)
2842 return ConstantInt::getTrue(RHSC->getContext());
2844 if (RHSC->getValue().isNonNegative()) {
2845 if (Pred == ICmpInst::ICMP_SLE)
2846 return ConstantInt::getTrue(RHSC->getContext());
2847 if (Pred == ICmpInst::ICMP_SGT)
2848 return ConstantInt::getFalse(RHSC->getContext());
2853 // icmp pred (urem X, Y), Y
2854 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2858 case ICmpInst::ICMP_SGT:
2859 case ICmpInst::ICMP_SGE: {
2860 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2861 if (!Known.isNonNegative())
2865 case ICmpInst::ICMP_EQ:
2866 case ICmpInst::ICMP_UGT:
2867 case ICmpInst::ICMP_UGE:
2868 return getFalse(ITy);
2869 case ICmpInst::ICMP_SLT:
2870 case ICmpInst::ICMP_SLE: {
2871 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2872 if (!Known.isNonNegative())
2876 case ICmpInst::ICMP_NE:
2877 case ICmpInst::ICMP_ULT:
2878 case ICmpInst::ICMP_ULE:
2879 return getTrue(ITy);
2883 // icmp pred X, (urem Y, X)
2884 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2888 case ICmpInst::ICMP_SGT:
2889 case ICmpInst::ICMP_SGE: {
2890 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2891 if (!Known.isNonNegative())
2895 case ICmpInst::ICMP_NE:
2896 case ICmpInst::ICMP_UGT:
2897 case ICmpInst::ICMP_UGE:
2898 return getTrue(ITy);
2899 case ICmpInst::ICMP_SLT:
2900 case ICmpInst::ICMP_SLE: {
2901 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2902 if (!Known.isNonNegative())
2906 case ICmpInst::ICMP_EQ:
2907 case ICmpInst::ICMP_ULT:
2908 case ICmpInst::ICMP_ULE:
2909 return getFalse(ITy);
2915 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2916 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2917 // icmp pred (X op Y), X
2918 if (Pred == ICmpInst::ICMP_UGT)
2919 return getFalse(ITy);
2920 if (Pred == ICmpInst::ICMP_ULE)
2921 return getTrue(ITy);
2926 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2927 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2928 // icmp pred X, (X op Y)
2929 if (Pred == ICmpInst::ICMP_ULT)
2930 return getFalse(ITy);
2931 if (Pred == ICmpInst::ICMP_UGE)
2932 return getTrue(ITy);
2939 // where CI2 is a power of 2 and CI isn't
2940 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2941 const APInt *CI2Val, *CIVal = &CI->getValue();
2942 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2943 CI2Val->isPowerOf2()) {
2944 if (!CIVal->isPowerOf2()) {
2945 // CI2 << X can equal zero in some circumstances,
2946 // this simplification is unsafe if CI is zero.
2948 // We know it is safe if:
2949 // - The shift is nsw, we can't shift out the one bit.
2950 // - The shift is nuw, we can't shift out the one bit.
2953 if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2954 Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2955 CI2Val->isOneValue() || !CI->isZero()) {
2956 if (Pred == ICmpInst::ICMP_EQ)
2957 return ConstantInt::getFalse(RHS->getContext());
2958 if (Pred == ICmpInst::ICMP_NE)
2959 return ConstantInt::getTrue(RHS->getContext());
2962 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2963 if (Pred == ICmpInst::ICMP_UGT)
2964 return ConstantInt::getFalse(RHS->getContext());
2965 if (Pred == ICmpInst::ICMP_ULE)
2966 return ConstantInt::getTrue(RHS->getContext());
2971 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2972 LBO->getOperand(1) == RBO->getOperand(1)) {
2973 switch (LBO->getOpcode()) {
2976 case Instruction::UDiv:
2977 case Instruction::LShr:
2978 if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
2979 !Q.IIQ.isExact(RBO))
2981 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2982 RBO->getOperand(0), Q, MaxRecurse - 1))
2985 case Instruction::SDiv:
2986 if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
2987 !Q.IIQ.isExact(RBO))
2989 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2990 RBO->getOperand(0), Q, MaxRecurse - 1))
2993 case Instruction::AShr:
2994 if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
2996 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2997 RBO->getOperand(0), Q, MaxRecurse - 1))
3000 case Instruction::Shl: {
3001 bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
3002 bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
3005 if (!NSW && ICmpInst::isSigned(Pred))
3007 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
3008 RBO->getOperand(0), Q, MaxRecurse - 1))
3017 /// Simplify integer comparisons where at least one operand of the compare
3018 /// matches an integer min/max idiom.
3019 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
3020 Value *RHS, const SimplifyQuery &Q,
3021 unsigned MaxRecurse) {
3022 Type *ITy = GetCompareTy(LHS); // The return type.
3024 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
3025 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
3027 // Signed variants on "max(a,b)>=a -> true".
3028 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3030 std::swap(A, B); // smax(A, B) pred A.
3031 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3032 // We analyze this as smax(A, B) pred A.
3034 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
3035 (A == LHS || B == LHS)) {
3037 std::swap(A, B); // A pred smax(A, B).
3038 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3039 // We analyze this as smax(A, B) swapped-pred A.
3040 P = CmpInst::getSwappedPredicate(Pred);
3041 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3042 (A == RHS || B == RHS)) {
3044 std::swap(A, B); // smin(A, B) pred A.
3045 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3046 // We analyze this as smax(-A, -B) swapped-pred -A.
3047 // Note that we do not need to actually form -A or -B thanks to EqP.
3048 P = CmpInst::getSwappedPredicate(Pred);
3049 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
3050 (A == LHS || B == LHS)) {
3052 std::swap(A, B); // A pred smin(A, B).
3053 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3054 // We analyze this as smax(-A, -B) pred -A.
3055 // Note that we do not need to actually form -A or -B thanks to EqP.
3058 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3059 // Cases correspond to "max(A, B) p A".
3063 case CmpInst::ICMP_EQ:
3064 case CmpInst::ICMP_SLE:
3065 // Equivalent to "A EqP B". This may be the same as the condition tested
3066 // in the max/min; if so, we can just return that.
3067 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3069 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3071 // Otherwise, see if "A EqP B" simplifies.
3073 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3076 case CmpInst::ICMP_NE:
3077 case CmpInst::ICMP_SGT: {
3078 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3079 // Equivalent to "A InvEqP B". This may be the same as the condition
3080 // tested in the max/min; if so, we can just return that.
3081 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3083 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3085 // Otherwise, see if "A InvEqP B" simplifies.
3087 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3091 case CmpInst::ICMP_SGE:
3093 return getTrue(ITy);
3094 case CmpInst::ICMP_SLT:
3096 return getFalse(ITy);
3100 // Unsigned variants on "max(a,b)>=a -> true".
3101 P = CmpInst::BAD_ICMP_PREDICATE;
3102 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3104 std::swap(A, B); // umax(A, B) pred A.
3105 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3106 // We analyze this as umax(A, B) pred A.
3108 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
3109 (A == LHS || B == LHS)) {
3111 std::swap(A, B); // A pred umax(A, B).
3112 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3113 // We analyze this as umax(A, B) swapped-pred A.
3114 P = CmpInst::getSwappedPredicate(Pred);
3115 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3116 (A == RHS || B == RHS)) {
3118 std::swap(A, B); // umin(A, B) pred A.
3119 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3120 // We analyze this as umax(-A, -B) swapped-pred -A.
3121 // Note that we do not need to actually form -A or -B thanks to EqP.
3122 P = CmpInst::getSwappedPredicate(Pred);
3123 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3124 (A == LHS || B == LHS)) {
3126 std::swap(A, B); // A pred umin(A, B).
3127 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3128 // We analyze this as umax(-A, -B) pred -A.
3129 // Note that we do not need to actually form -A or -B thanks to EqP.
3132 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3133 // Cases correspond to "max(A, B) p A".
3137 case CmpInst::ICMP_EQ:
3138 case CmpInst::ICMP_ULE:
3139 // Equivalent to "A EqP B". This may be the same as the condition tested
3140 // in the max/min; if so, we can just return that.
3141 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3143 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3145 // Otherwise, see if "A EqP B" simplifies.
3147 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3150 case CmpInst::ICMP_NE:
3151 case CmpInst::ICMP_UGT: {
3152 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3153 // Equivalent to "A InvEqP B". This may be the same as the condition
3154 // tested in the max/min; if so, we can just return that.
3155 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3157 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3159 // Otherwise, see if "A InvEqP B" simplifies.
3161 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3165 case CmpInst::ICMP_UGE:
3167 return getTrue(ITy);
3168 case CmpInst::ICMP_ULT:
3170 return getFalse(ITy);
3174 // Variants on "max(x,y) >= min(x,z)".
3176 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3177 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3178 (A == C || A == D || B == C || B == D)) {
3179 // max(x, ?) pred min(x, ?).
3180 if (Pred == CmpInst::ICMP_SGE)
3182 return getTrue(ITy);
3183 if (Pred == CmpInst::ICMP_SLT)
3185 return getFalse(ITy);
3186 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3187 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3188 (A == C || A == D || B == C || B == D)) {
3189 // min(x, ?) pred max(x, ?).
3190 if (Pred == CmpInst::ICMP_SLE)
3192 return getTrue(ITy);
3193 if (Pred == CmpInst::ICMP_SGT)
3195 return getFalse(ITy);
3196 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3197 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3198 (A == C || A == D || B == C || B == D)) {
3199 // max(x, ?) pred min(x, ?).
3200 if (Pred == CmpInst::ICMP_UGE)
3202 return getTrue(ITy);
3203 if (Pred == CmpInst::ICMP_ULT)
3205 return getFalse(ITy);
3206 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3207 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3208 (A == C || A == D || B == C || B == D)) {
3209 // min(x, ?) pred max(x, ?).
3210 if (Pred == CmpInst::ICMP_ULE)
3212 return getTrue(ITy);
3213 if (Pred == CmpInst::ICMP_UGT)
3215 return getFalse(ITy);
3221 /// Given operands for an ICmpInst, see if we can fold the result.
3222 /// If not, this returns null.
3223 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3224 const SimplifyQuery &Q, unsigned MaxRecurse) {
3225 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3226 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3228 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3229 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3230 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3232 // If we have a constant, make sure it is on the RHS.
3233 std::swap(LHS, RHS);
3234 Pred = CmpInst::getSwappedPredicate(Pred);
3236 assert(!isa<UndefValue>(LHS) && "Unexpected icmp undef,%X");
3238 Type *ITy = GetCompareTy(LHS); // The return type.
3240 // For EQ and NE, we can always pick a value for the undef to make the
3241 // predicate pass or fail, so we can return undef.
3242 // Matches behavior in llvm::ConstantFoldCompareInstruction.
3243 if (isa<UndefValue>(RHS) && ICmpInst::isEquality(Pred))
3244 return UndefValue::get(ITy);
3246 // icmp X, X -> true/false
3247 // icmp X, undef -> true/false because undef could be X.
3248 if (LHS == RHS || isa<UndefValue>(RHS))
3249 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3251 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3254 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3257 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3260 // If both operands have range metadata, use the metadata
3261 // to simplify the comparison.
3262 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3263 auto RHS_Instr = cast<Instruction>(RHS);
3264 auto LHS_Instr = cast<Instruction>(LHS);
3266 if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3267 Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3268 auto RHS_CR = getConstantRangeFromMetadata(
3269 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3270 auto LHS_CR = getConstantRangeFromMetadata(
3271 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3273 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3274 if (Satisfied_CR.contains(LHS_CR))
3275 return ConstantInt::getTrue(RHS->getContext());
3277 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3278 CmpInst::getInversePredicate(Pred), RHS_CR);
3279 if (InversedSatisfied_CR.contains(LHS_CR))
3280 return ConstantInt::getFalse(RHS->getContext());
3284 // Compare of cast, for example (zext X) != 0 -> X != 0
3285 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3286 Instruction *LI = cast<CastInst>(LHS);
3287 Value *SrcOp = LI->getOperand(0);
3288 Type *SrcTy = SrcOp->getType();
3289 Type *DstTy = LI->getType();
3291 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3292 // if the integer type is the same size as the pointer type.
3293 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3294 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3295 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3296 // Transfer the cast to the constant.
3297 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3298 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3301 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3302 if (RI->getOperand(0)->getType() == SrcTy)
3303 // Compare without the cast.
3304 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3310 if (isa<ZExtInst>(LHS)) {
3311 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3313 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3314 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3315 // Compare X and Y. Note that signed predicates become unsigned.
3316 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3317 SrcOp, RI->getOperand(0), Q,
3321 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3322 // too. If not, then try to deduce the result of the comparison.
3323 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3324 // Compute the constant that would happen if we truncated to SrcTy then
3325 // reextended to DstTy.
3326 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3327 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3329 // If the re-extended constant didn't change then this is effectively
3330 // also a case of comparing two zero-extended values.
3331 if (RExt == CI && MaxRecurse)
3332 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3333 SrcOp, Trunc, Q, MaxRecurse-1))
3336 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3337 // there. Use this to work out the result of the comparison.
3340 default: llvm_unreachable("Unknown ICmp predicate!");
3342 case ICmpInst::ICMP_EQ:
3343 case ICmpInst::ICMP_UGT:
3344 case ICmpInst::ICMP_UGE:
3345 return ConstantInt::getFalse(CI->getContext());
3347 case ICmpInst::ICMP_NE:
3348 case ICmpInst::ICMP_ULT:
3349 case ICmpInst::ICMP_ULE:
3350 return ConstantInt::getTrue(CI->getContext());
3352 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3353 // is non-negative then LHS <s RHS.
3354 case ICmpInst::ICMP_SGT:
3355 case ICmpInst::ICMP_SGE:
3356 return CI->getValue().isNegative() ?
3357 ConstantInt::getTrue(CI->getContext()) :
3358 ConstantInt::getFalse(CI->getContext());
3360 case ICmpInst::ICMP_SLT:
3361 case ICmpInst::ICMP_SLE:
3362 return CI->getValue().isNegative() ?
3363 ConstantInt::getFalse(CI->getContext()) :
3364 ConstantInt::getTrue(CI->getContext());
3370 if (isa<SExtInst>(LHS)) {
3371 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3373 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3374 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3375 // Compare X and Y. Note that the predicate does not change.
3376 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3380 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3381 // too. If not, then try to deduce the result of the comparison.
3382 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3383 // Compute the constant that would happen if we truncated to SrcTy then
3384 // reextended to DstTy.
3385 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3386 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3388 // If the re-extended constant didn't change then this is effectively
3389 // also a case of comparing two sign-extended values.
3390 if (RExt == CI && MaxRecurse)
3391 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3394 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3395 // bits there. Use this to work out the result of the comparison.
3398 default: llvm_unreachable("Unknown ICmp predicate!");
3399 case ICmpInst::ICMP_EQ:
3400 return ConstantInt::getFalse(CI->getContext());
3401 case ICmpInst::ICMP_NE:
3402 return ConstantInt::getTrue(CI->getContext());
3404 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3406 case ICmpInst::ICMP_SGT:
3407 case ICmpInst::ICMP_SGE:
3408 return CI->getValue().isNegative() ?
3409 ConstantInt::getTrue(CI->getContext()) :
3410 ConstantInt::getFalse(CI->getContext());
3411 case ICmpInst::ICMP_SLT:
3412 case ICmpInst::ICMP_SLE:
3413 return CI->getValue().isNegative() ?
3414 ConstantInt::getFalse(CI->getContext()) :
3415 ConstantInt::getTrue(CI->getContext());
3417 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3419 case ICmpInst::ICMP_UGT:
3420 case ICmpInst::ICMP_UGE:
3421 // Comparison is true iff the LHS <s 0.
3423 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3424 Constant::getNullValue(SrcTy),
3428 case ICmpInst::ICMP_ULT:
3429 case ICmpInst::ICMP_ULE:
3430 // Comparison is true iff the LHS >=s 0.
3432 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3433 Constant::getNullValue(SrcTy),
3443 // icmp eq|ne X, Y -> false|true if X != Y
3444 if (ICmpInst::isEquality(Pred) &&
3445 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3446 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3449 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3452 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3455 // Simplify comparisons of related pointers using a powerful, recursive
3456 // GEP-walk when we have target data available..
3457 if (LHS->getType()->isPointerTy())
3458 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3461 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3462 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3463 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3464 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3465 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3466 Q.DL.getTypeSizeInBits(CRHS->getType()))
3467 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3468 Q.IIQ, CLHS->getPointerOperand(),
3469 CRHS->getPointerOperand()))
3472 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3473 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3474 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3475 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3476 (ICmpInst::isEquality(Pred) ||
3477 (GLHS->isInBounds() && GRHS->isInBounds() &&
3478 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3479 // The bases are equal and the indices are constant. Build a constant
3480 // expression GEP with the same indices and a null base pointer to see
3481 // what constant folding can make out of it.
3482 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3483 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3484 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3485 GLHS->getSourceElementType(), Null, IndicesLHS);
3487 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3488 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3489 GLHS->getSourceElementType(), Null, IndicesRHS);
3490 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3495 // If the comparison is with the result of a select instruction, check whether
3496 // comparing with either branch of the select always yields the same value.
3497 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3498 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3501 // If the comparison is with the result of a phi instruction, check whether
3502 // doing the compare with each incoming phi value yields a common result.
3503 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3504 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3510 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3511 const SimplifyQuery &Q) {
3512 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3515 /// Given operands for an FCmpInst, see if we can fold the result.
3516 /// If not, this returns null.
3517 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3518 FastMathFlags FMF, const SimplifyQuery &Q,
3519 unsigned MaxRecurse) {
3520 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3521 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3523 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3524 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3525 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3527 // If we have a constant, make sure it is on the RHS.
3528 std::swap(LHS, RHS);
3529 Pred = CmpInst::getSwappedPredicate(Pred);
3532 // Fold trivial predicates.
3533 Type *RetTy = GetCompareTy(LHS);
3534 if (Pred == FCmpInst::FCMP_FALSE)
3535 return getFalse(RetTy);
3536 if (Pred == FCmpInst::FCMP_TRUE)
3537 return getTrue(RetTy);
3539 // Fold (un)ordered comparison if we can determine there are no NaNs.
3540 if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3542 (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3543 return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3545 // NaN is unordered; NaN is not ordered.
3546 assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&
3547 "Comparison must be either ordered or unordered");
3548 if (match(RHS, m_NaN()))
3549 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3551 // fcmp pred x, undef and fcmp pred undef, x
3552 // fold to true if unordered, false if ordered
3553 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3554 // Choosing NaN for the undef will always make unordered comparison succeed
3555 // and ordered comparison fail.
3556 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3559 // fcmp x,x -> true/false. Not all compares are foldable.
3561 if (CmpInst::isTrueWhenEqual(Pred))
3562 return getTrue(RetTy);
3563 if (CmpInst::isFalseWhenEqual(Pred))
3564 return getFalse(RetTy);
3567 // Handle fcmp with constant RHS.
3568 // TODO: Use match with a specific FP value, so these work with vectors with
3571 if (match(RHS, m_APFloat(C))) {
3572 // Check whether the constant is an infinity.
3573 if (C->isInfinity()) {
3574 if (C->isNegative()) {
3576 case FCmpInst::FCMP_OLT:
3577 // No value is ordered and less than negative infinity.
3578 return getFalse(RetTy);
3579 case FCmpInst::FCMP_UGE:
3580 // All values are unordered with or at least negative infinity.
3581 return getTrue(RetTy);
3587 case FCmpInst::FCMP_OGT:
3588 // No value is ordered and greater than infinity.
3589 return getFalse(RetTy);
3590 case FCmpInst::FCMP_ULE:
3591 // All values are unordered with and at most infinity.
3592 return getTrue(RetTy);
3598 if (C->isNegative() && !C->isNegZero()) {
3599 assert(!C->isNaN() && "Unexpected NaN constant!");
3600 // TODO: We can catch more cases by using a range check rather than
3601 // relying on CannotBeOrderedLessThanZero.
3603 case FCmpInst::FCMP_UGE:
3604 case FCmpInst::FCMP_UGT:
3605 case FCmpInst::FCMP_UNE:
3606 // (X >= 0) implies (X > C) when (C < 0)
3607 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3608 return getTrue(RetTy);
3610 case FCmpInst::FCMP_OEQ:
3611 case FCmpInst::FCMP_OLE:
3612 case FCmpInst::FCMP_OLT:
3613 // (X >= 0) implies !(X < C) when (C < 0)
3614 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3615 return getFalse(RetTy);
3622 // Check comparison of [minnum/maxnum with constant] with other constant.
3624 if ((match(LHS, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_APFloat(C2))) &&
3625 C2->compare(*C) == APFloat::cmpLessThan) ||
3626 (match(LHS, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_APFloat(C2))) &&
3627 C2->compare(*C) == APFloat::cmpGreaterThan)) {
3629 cast<IntrinsicInst>(LHS)->getIntrinsicID() == Intrinsic::maxnum;
3630 // The ordered relationship and minnum/maxnum guarantee that we do not
3631 // have NaN constants, so ordered/unordered preds are handled the same.
3633 case FCmpInst::FCMP_OEQ: case FCmpInst::FCMP_UEQ:
3634 // minnum(X, LesserC) == C --> false
3635 // maxnum(X, GreaterC) == C --> false
3636 return getFalse(RetTy);
3637 case FCmpInst::FCMP_ONE: case FCmpInst::FCMP_UNE:
3638 // minnum(X, LesserC) != C --> true
3639 // maxnum(X, GreaterC) != C --> true
3640 return getTrue(RetTy);
3641 case FCmpInst::FCMP_OGE: case FCmpInst::FCMP_UGE:
3642 case FCmpInst::FCMP_OGT: case FCmpInst::FCMP_UGT:
3643 // minnum(X, LesserC) >= C --> false
3644 // minnum(X, LesserC) > C --> false
3645 // maxnum(X, GreaterC) >= C --> true
3646 // maxnum(X, GreaterC) > C --> true
3647 return ConstantInt::get(RetTy, IsMaxNum);
3648 case FCmpInst::FCMP_OLE: case FCmpInst::FCMP_ULE:
3649 case FCmpInst::FCMP_OLT: case FCmpInst::FCMP_ULT:
3650 // minnum(X, LesserC) <= C --> true
3651 // minnum(X, LesserC) < C --> true
3652 // maxnum(X, GreaterC) <= C --> false
3653 // maxnum(X, GreaterC) < C --> false
3654 return ConstantInt::get(RetTy, !IsMaxNum);
3656 // TRUE/FALSE/ORD/UNO should be handled before this.
3657 llvm_unreachable("Unexpected fcmp predicate");
3662 if (match(RHS, m_AnyZeroFP())) {
3664 case FCmpInst::FCMP_OGE:
3665 case FCmpInst::FCMP_ULT:
3666 // Positive or zero X >= 0.0 --> true
3667 // Positive or zero X < 0.0 --> false
3668 if ((FMF.noNaNs() || isKnownNeverNaN(LHS, Q.TLI)) &&
3669 CannotBeOrderedLessThanZero(LHS, Q.TLI))
3670 return Pred == FCmpInst::FCMP_OGE ? getTrue(RetTy) : getFalse(RetTy);
3672 case FCmpInst::FCMP_UGE:
3673 case FCmpInst::FCMP_OLT:
3674 // Positive or zero or nan X >= 0.0 --> true
3675 // Positive or zero or nan X < 0.0 --> false
3676 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3677 return Pred == FCmpInst::FCMP_UGE ? getTrue(RetTy) : getFalse(RetTy);
3684 // If the comparison is with the result of a select instruction, check whether
3685 // comparing with either branch of the select always yields the same value.
3686 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3687 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3690 // If the comparison is with the result of a phi instruction, check whether
3691 // doing the compare with each incoming phi value yields a common result.
3692 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3693 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3699 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3700 FastMathFlags FMF, const SimplifyQuery &Q) {
3701 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3704 /// See if V simplifies when its operand Op is replaced with RepOp.
3705 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3706 const SimplifyQuery &Q,
3707 unsigned MaxRecurse) {
3708 // Trivial replacement.
3712 // We cannot replace a constant, and shouldn't even try.
3713 if (isa<Constant>(Op))
3716 auto *I = dyn_cast<Instruction>(V);
3720 // If this is a binary operator, try to simplify it with the replaced op.
3721 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3723 // %cmp = icmp eq i32 %x, 2147483647
3724 // %add = add nsw i32 %x, 1
3725 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3727 // We can't replace %sel with %add unless we strip away the flags.
3728 // TODO: This is an unusual limitation because better analysis results in
3729 // worse simplification. InstCombine can do this fold more generally
3730 // by dropping the flags. Remove this fold to save compile-time?
3731 if (isa<OverflowingBinaryOperator>(B))
3732 if (Q.IIQ.hasNoSignedWrap(B) || Q.IIQ.hasNoUnsignedWrap(B))
3734 if (isa<PossiblyExactOperator>(B) && Q.IIQ.isExact(B))
3738 if (B->getOperand(0) == Op)
3739 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3741 if (B->getOperand(1) == Op)
3742 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3747 // Same for CmpInsts.
3748 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3750 if (C->getOperand(0) == Op)
3751 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3753 if (C->getOperand(1) == Op)
3754 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3760 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3762 SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3763 transform(GEP->operands(), NewOps.begin(),
3764 [&](Value *V) { return V == Op ? RepOp : V; });
3765 return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3770 // TODO: We could hand off more cases to instsimplify here.
3772 // If all operands are constant after substituting Op for RepOp then we can
3773 // constant fold the instruction.
3774 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3775 // Build a list of all constant operands.
3776 SmallVector<Constant *, 8> ConstOps;
3777 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3778 if (I->getOperand(i) == Op)
3779 ConstOps.push_back(CRepOp);
3780 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3781 ConstOps.push_back(COp);
3786 // All operands were constants, fold it.
3787 if (ConstOps.size() == I->getNumOperands()) {
3788 if (CmpInst *C = dyn_cast<CmpInst>(I))
3789 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3790 ConstOps[1], Q.DL, Q.TLI);
3792 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3793 if (!LI->isVolatile())
3794 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3796 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3803 /// Try to simplify a select instruction when its condition operand is an
3804 /// integer comparison where one operand of the compare is a constant.
3805 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3806 const APInt *Y, bool TrueWhenUnset) {
3809 // (X & Y) == 0 ? X & ~Y : X --> X
3810 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3811 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3813 return TrueWhenUnset ? FalseVal : TrueVal;
3815 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3816 // (X & Y) != 0 ? X : X & ~Y --> X
3817 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3819 return TrueWhenUnset ? FalseVal : TrueVal;
3821 if (Y->isPowerOf2()) {
3822 // (X & Y) == 0 ? X | Y : X --> X | Y
3823 // (X & Y) != 0 ? X | Y : X --> X
3824 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3826 return TrueWhenUnset ? TrueVal : FalseVal;
3828 // (X & Y) == 0 ? X : X | Y --> X
3829 // (X & Y) != 0 ? X : X | Y --> X | Y
3830 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3832 return TrueWhenUnset ? TrueVal : FalseVal;
3838 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3840 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3841 ICmpInst::Predicate Pred,
3842 Value *TrueVal, Value *FalseVal) {
3845 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3848 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3849 Pred == ICmpInst::ICMP_EQ);
3852 /// Try to simplify a select instruction when its condition operand is an
3853 /// integer comparison.
3854 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3855 Value *FalseVal, const SimplifyQuery &Q,
3856 unsigned MaxRecurse) {
3857 ICmpInst::Predicate Pred;
3858 Value *CmpLHS, *CmpRHS;
3859 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3862 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3865 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3866 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3867 Pred == ICmpInst::ICMP_EQ))
3870 // Test for a bogus zero-shift-guard-op around funnel-shift or rotate.
3872 auto isFsh = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X), m_Value(),
3874 m_Intrinsic<Intrinsic::fshr>(m_Value(), m_Value(X),
3876 // (ShAmt == 0) ? fshl(X, *, ShAmt) : X --> X
3877 // (ShAmt == 0) ? fshr(*, X, ShAmt) : X --> X
3878 if (match(TrueVal, isFsh) && FalseVal == X && CmpLHS == ShAmt &&
3879 Pred == ICmpInst::ICMP_EQ)
3881 // (ShAmt != 0) ? X : fshl(X, *, ShAmt) --> X
3882 // (ShAmt != 0) ? X : fshr(*, X, ShAmt) --> X
3883 if (match(FalseVal, isFsh) && TrueVal == X && CmpLHS == ShAmt &&
3884 Pred == ICmpInst::ICMP_NE)
3887 // Test for a zero-shift-guard-op around rotates. These are used to
3888 // avoid UB from oversized shifts in raw IR rotate patterns, but the
3889 // intrinsics do not have that problem.
3890 // We do not allow this transform for the general funnel shift case because
3891 // that would not preserve the poison safety of the original code.
3892 auto isRotate = m_CombineOr(m_Intrinsic<Intrinsic::fshl>(m_Value(X),
3895 m_Intrinsic<Intrinsic::fshr>(m_Value(X),
3898 // (ShAmt != 0) ? fshl(X, X, ShAmt) : X --> fshl(X, X, ShAmt)
3899 // (ShAmt != 0) ? fshr(X, X, ShAmt) : X --> fshr(X, X, ShAmt)
3900 if (match(TrueVal, isRotate) && FalseVal == X && CmpLHS == ShAmt &&
3901 Pred == ICmpInst::ICMP_NE)
3903 // (ShAmt == 0) ? X : fshl(X, X, ShAmt) --> fshl(X, X, ShAmt)
3904 // (ShAmt == 0) ? X : fshr(X, X, ShAmt) --> fshr(X, X, ShAmt)
3905 if (match(FalseVal, isRotate) && TrueVal == X && CmpLHS == ShAmt &&
3906 Pred == ICmpInst::ICMP_EQ)
3910 // Check for other compares that behave like bit test.
3911 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3915 // If we have an equality comparison, then we know the value in one of the
3916 // arms of the select. See if substituting this value into the arm and
3917 // simplifying the result yields the same value as the other arm.
3918 if (Pred == ICmpInst::ICMP_EQ) {
3919 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3921 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3924 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3926 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3929 } else if (Pred == ICmpInst::ICMP_NE) {
3930 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3932 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3935 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3937 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3945 /// Try to simplify a select instruction when its condition operand is a
3946 /// floating-point comparison.
3947 static Value *simplifySelectWithFCmp(Value *Cond, Value *T, Value *F,
3948 const SimplifyQuery &Q) {
3949 FCmpInst::Predicate Pred;
3950 if (!match(Cond, m_FCmp(Pred, m_Specific(T), m_Specific(F))) &&
3951 !match(Cond, m_FCmp(Pred, m_Specific(F), m_Specific(T))))
3954 // This transform is safe if we do not have (do not care about) -0.0 or if
3955 // at least one operand is known to not be -0.0. Otherwise, the select can
3956 // change the sign of a zero operand.
3957 bool HasNoSignedZeros = Q.CxtI && isa<FPMathOperator>(Q.CxtI) &&
3958 Q.CxtI->hasNoSignedZeros();
3960 if (HasNoSignedZeros || (match(T, m_APFloat(C)) && C->isNonZero()) ||
3961 (match(F, m_APFloat(C)) && C->isNonZero())) {
3962 // (T == F) ? T : F --> F
3963 // (F == T) ? T : F --> F
3964 if (Pred == FCmpInst::FCMP_OEQ)
3967 // (T != F) ? T : F --> T
3968 // (F != T) ? T : F --> T
3969 if (Pred == FCmpInst::FCMP_UNE)
3976 /// Given operands for a SelectInst, see if we can fold the result.
3977 /// If not, this returns null.
3978 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3979 const SimplifyQuery &Q, unsigned MaxRecurse) {
3980 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3981 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3982 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3983 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3985 // select undef, X, Y -> X or Y
3986 if (isa<UndefValue>(CondC))
3987 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3989 // TODO: Vector constants with undef elements don't simplify.
3991 // select true, X, Y -> X
3992 if (CondC->isAllOnesValue())
3994 // select false, X, Y -> Y
3995 if (CondC->isNullValue())
3999 // select i1 Cond, i1 true, i1 false --> i1 Cond
4000 assert(Cond->getType()->isIntOrIntVectorTy(1) &&
4001 "Select must have bool or bool vector condition");
4002 assert(TrueVal->getType() == FalseVal->getType() &&
4003 "Select must have same types for true/false ops");
4004 if (Cond->getType() == TrueVal->getType() &&
4005 match(TrueVal, m_One()) && match(FalseVal, m_ZeroInt()))
4008 // select ?, X, X -> X
4009 if (TrueVal == FalseVal)
4012 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
4014 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
4018 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
4021 if (Value *V = simplifySelectWithFCmp(Cond, TrueVal, FalseVal, Q))
4024 if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
4027 Optional<bool> Imp = isImpliedByDomCondition(Cond, Q.CxtI, Q.DL);
4029 return *Imp ? TrueVal : FalseVal;
4034 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
4035 const SimplifyQuery &Q) {
4036 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
4039 /// Given operands for an GetElementPtrInst, see if we can fold the result.
4040 /// If not, this returns null.
4041 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
4042 const SimplifyQuery &Q, unsigned) {
4043 // The type of the GEP pointer operand.
4045 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
4047 // getelementptr P -> P.
4048 if (Ops.size() == 1)
4051 // Compute the (pointer) type returned by the GEP instruction.
4052 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
4053 Type *GEPTy = PointerType::get(LastType, AS);
4054 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
4055 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
4056 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
4057 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
4059 if (isa<UndefValue>(Ops[0]))
4060 return UndefValue::get(GEPTy);
4062 if (Ops.size() == 2) {
4063 // getelementptr P, 0 -> P.
4064 if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
4068 if (Ty->isSized()) {
4071 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
4072 // getelementptr P, N -> P if P points to a type of zero size.
4073 if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
4076 // The following transforms are only safe if the ptrtoint cast
4077 // doesn't truncate the pointers.
4078 if (Ops[1]->getType()->getScalarSizeInBits() ==
4079 Q.DL.getPointerSizeInBits(AS)) {
4080 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
4081 if (match(P, m_Zero()))
4082 return Constant::getNullValue(GEPTy);
4084 if (match(P, m_PtrToInt(m_Value(Temp))))
4085 if (Temp->getType() == GEPTy)
4090 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
4091 if (TyAllocSize == 1 &&
4092 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
4093 if (Value *R = PtrToIntOrZero(P))
4096 // getelementptr V, (ashr (sub P, V), C) -> Q
4097 // if P points to a type of size 1 << C.
4099 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
4100 m_ConstantInt(C))) &&
4101 TyAllocSize == 1ULL << C)
4102 if (Value *R = PtrToIntOrZero(P))
4105 // getelementptr V, (sdiv (sub P, V), C) -> Q
4106 // if P points to a type of size C.
4108 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
4109 m_SpecificInt(TyAllocSize))))
4110 if (Value *R = PtrToIntOrZero(P))
4116 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
4117 all_of(Ops.slice(1).drop_back(1),
4118 [](Value *Idx) { return match(Idx, m_Zero()); })) {
4120 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
4121 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
4122 APInt BasePtrOffset(IdxWidth, 0);
4123 Value *StrippedBasePtr =
4124 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
4127 // gep (gep V, C), (sub 0, V) -> C
4128 if (match(Ops.back(),
4129 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
4130 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
4131 return ConstantExpr::getIntToPtr(CI, GEPTy);
4133 // gep (gep V, C), (xor V, -1) -> C-1
4134 if (match(Ops.back(),
4135 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
4136 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
4137 return ConstantExpr::getIntToPtr(CI, GEPTy);
4142 // Check to see if this is constant foldable.
4143 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
4146 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
4148 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
4153 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
4154 const SimplifyQuery &Q) {
4155 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
4158 /// Given operands for an InsertValueInst, see if we can fold the result.
4159 /// If not, this returns null.
4160 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
4161 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
4163 if (Constant *CAgg = dyn_cast<Constant>(Agg))
4164 if (Constant *CVal = dyn_cast<Constant>(Val))
4165 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
4167 // insertvalue x, undef, n -> x
4168 if (match(Val, m_Undef()))
4171 // insertvalue x, (extractvalue y, n), n
4172 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
4173 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
4174 EV->getIndices() == Idxs) {
4175 // insertvalue undef, (extractvalue y, n), n -> y
4176 if (match(Agg, m_Undef()))
4177 return EV->getAggregateOperand();
4179 // insertvalue y, (extractvalue y, n), n -> y
4180 if (Agg == EV->getAggregateOperand())
4187 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
4188 ArrayRef<unsigned> Idxs,
4189 const SimplifyQuery &Q) {
4190 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
4193 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
4194 const SimplifyQuery &Q) {
4195 // Try to constant fold.
4196 auto *VecC = dyn_cast<Constant>(Vec);
4197 auto *ValC = dyn_cast<Constant>(Val);
4198 auto *IdxC = dyn_cast<Constant>(Idx);
4199 if (VecC && ValC && IdxC)
4200 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
4202 // Fold into undef if index is out of bounds.
4203 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
4204 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
4205 if (CI->uge(NumElements))
4206 return UndefValue::get(Vec->getType());
4209 // If index is undef, it might be out of bounds (see above case)
4210 if (isa<UndefValue>(Idx))
4211 return UndefValue::get(Vec->getType());
4213 // Inserting an undef scalar? Assume it is the same value as the existing
4215 if (isa<UndefValue>(Val))
4218 // If we are extracting a value from a vector, then inserting it into the same
4219 // place, that's the input vector:
4220 // insertelt Vec, (extractelt Vec, Idx), Idx --> Vec
4221 if (match(Val, m_ExtractElement(m_Specific(Vec), m_Specific(Idx))))
4227 /// Given operands for an ExtractValueInst, see if we can fold the result.
4228 /// If not, this returns null.
4229 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4230 const SimplifyQuery &, unsigned) {
4231 if (auto *CAgg = dyn_cast<Constant>(Agg))
4232 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4234 // extractvalue x, (insertvalue y, elt, n), n -> elt
4235 unsigned NumIdxs = Idxs.size();
4236 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4237 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4238 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4239 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4240 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4241 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4242 Idxs.slice(0, NumCommonIdxs)) {
4243 if (NumIdxs == NumInsertValueIdxs)
4244 return IVI->getInsertedValueOperand();
4252 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4253 const SimplifyQuery &Q) {
4254 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
4257 /// Given operands for an ExtractElementInst, see if we can fold the result.
4258 /// If not, this returns null.
4259 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
4261 if (auto *CVec = dyn_cast<Constant>(Vec)) {
4262 if (auto *CIdx = dyn_cast<Constant>(Idx))
4263 return ConstantFoldExtractElementInstruction(CVec, CIdx);
4265 // The index is not relevant if our vector is a splat.
4266 if (auto *Splat = CVec->getSplatValue())
4269 if (isa<UndefValue>(Vec))
4270 return UndefValue::get(Vec->getType()->getVectorElementType());
4273 // If extracting a specified index from the vector, see if we can recursively
4274 // find a previously computed scalar that was inserted into the vector.
4275 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
4276 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4277 // definitely out of bounds, thus undefined result
4278 return UndefValue::get(Vec->getType()->getVectorElementType());
4279 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4283 // An undef extract index can be arbitrarily chosen to be an out-of-range
4284 // index value, which would result in the instruction being undef.
4285 if (isa<UndefValue>(Idx))
4286 return UndefValue::get(Vec->getType()->getVectorElementType());
4291 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
4292 const SimplifyQuery &Q) {
4293 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
4296 /// See if we can fold the given phi. If not, returns null.
4297 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4298 // If all of the PHI's incoming values are the same then replace the PHI node
4299 // with the common value.
4300 Value *CommonValue = nullptr;
4301 bool HasUndefInput = false;
4302 for (Value *Incoming : PN->incoming_values()) {
4303 // If the incoming value is the phi node itself, it can safely be skipped.
4304 if (Incoming == PN) continue;
4305 if (isa<UndefValue>(Incoming)) {
4306 // Remember that we saw an undef value, but otherwise ignore them.
4307 HasUndefInput = true;
4310 if (CommonValue && Incoming != CommonValue)
4311 return nullptr; // Not the same, bail out.
4312 CommonValue = Incoming;
4315 // If CommonValue is null then all of the incoming values were either undef or
4316 // equal to the phi node itself.
4318 return UndefValue::get(PN->getType());
4320 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4321 // instruction, we cannot return X as the result of the PHI node unless it
4322 // dominates the PHI block.
4324 return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4329 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4330 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4331 if (auto *C = dyn_cast<Constant>(Op))
4332 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4334 if (auto *CI = dyn_cast<CastInst>(Op)) {
4335 auto *Src = CI->getOperand(0);
4336 Type *SrcTy = Src->getType();
4337 Type *MidTy = CI->getType();
4339 if (Src->getType() == Ty) {
4340 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4341 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4343 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4345 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4347 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4348 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4349 SrcIntPtrTy, MidIntPtrTy,
4350 DstIntPtrTy) == Instruction::BitCast)
4356 if (CastOpc == Instruction::BitCast)
4357 if (Op->getType() == Ty)
4363 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4364 const SimplifyQuery &Q) {
4365 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4368 /// For the given destination element of a shuffle, peek through shuffles to
4369 /// match a root vector source operand that contains that element in the same
4370 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4371 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4372 int MaskVal, Value *RootVec,
4373 unsigned MaxRecurse) {
4377 // Bail out if any mask value is undefined. That kind of shuffle may be
4378 // simplified further based on demanded bits or other folds.
4382 // The mask value chooses which source operand we need to look at next.
4383 int InVecNumElts = Op0->getType()->getVectorNumElements();
4384 int RootElt = MaskVal;
4385 Value *SourceOp = Op0;
4386 if (MaskVal >= InVecNumElts) {
4387 RootElt = MaskVal - InVecNumElts;
4391 // If the source operand is a shuffle itself, look through it to find the
4392 // matching root vector.
4393 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4394 return foldIdentityShuffles(
4395 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4396 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4399 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4402 // The source operand is not a shuffle. Initialize the root vector value for
4403 // this shuffle if that has not been done yet.
4407 // Give up as soon as a source operand does not match the existing root value.
4408 if (RootVec != SourceOp)
4411 // The element must be coming from the same lane in the source vector
4412 // (although it may have crossed lanes in intermediate shuffles).
4413 if (RootElt != DestElt)
4419 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4420 Type *RetTy, const SimplifyQuery &Q,
4421 unsigned MaxRecurse) {
4422 if (isa<UndefValue>(Mask))
4423 return UndefValue::get(RetTy);
4425 Type *InVecTy = Op0->getType();
4426 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4427 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4429 SmallVector<int, 32> Indices;
4430 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4431 assert(MaskNumElts == Indices.size() &&
4432 "Size of Indices not same as number of mask elements?");
4434 // Canonicalization: If mask does not select elements from an input vector,
4435 // replace that input vector with undef.
4436 bool MaskSelects0 = false, MaskSelects1 = false;
4437 for (unsigned i = 0; i != MaskNumElts; ++i) {
4438 if (Indices[i] == -1)
4440 if ((unsigned)Indices[i] < InVecNumElts)
4441 MaskSelects0 = true;
4443 MaskSelects1 = true;
4446 Op0 = UndefValue::get(InVecTy);
4448 Op1 = UndefValue::get(InVecTy);
4450 auto *Op0Const = dyn_cast<Constant>(Op0);
4451 auto *Op1Const = dyn_cast<Constant>(Op1);
4453 // If all operands are constant, constant fold the shuffle.
4454 if (Op0Const && Op1Const)
4455 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4457 // Canonicalization: if only one input vector is constant, it shall be the
4459 if (Op0Const && !Op1Const) {
4460 std::swap(Op0, Op1);
4461 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4464 // A splat of an inserted scalar constant becomes a vector constant:
4465 // shuf (inselt ?, C, IndexC), undef, <IndexC, IndexC...> --> <C, C...>
4466 // NOTE: We may have commuted above, so analyze the updated Indices, not the
4467 // original mask constant.
4469 ConstantInt *IndexC;
4470 if (match(Op0, m_InsertElement(m_Value(), m_Constant(C),
4471 m_ConstantInt(IndexC)))) {
4472 // Match a splat shuffle mask of the insert index allowing undef elements.
4473 int InsertIndex = IndexC->getZExtValue();
4474 if (all_of(Indices, [InsertIndex](int MaskElt) {
4475 return MaskElt == InsertIndex || MaskElt == -1;
4477 assert(isa<UndefValue>(Op1) && "Expected undef operand 1 for splat");
4479 // Shuffle mask undefs become undefined constant result elements.
4480 SmallVector<Constant *, 16> VecC(MaskNumElts, C);
4481 for (unsigned i = 0; i != MaskNumElts; ++i)
4482 if (Indices[i] == -1)
4483 VecC[i] = UndefValue::get(C->getType());
4484 return ConstantVector::get(VecC);
4488 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4489 // value type is same as the input vectors' type.
4490 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4491 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4492 OpShuf->getMask()->getSplatValue())
4495 // Don't fold a shuffle with undef mask elements. This may get folded in a
4496 // better way using demanded bits or other analysis.
4497 // TODO: Should we allow this?
4498 if (find(Indices, -1) != Indices.end())
4501 // Check if every element of this shuffle can be mapped back to the
4502 // corresponding element of a single root vector. If so, we don't need this
4503 // shuffle. This handles simple identity shuffles as well as chains of
4504 // shuffles that may widen/narrow and/or move elements across lanes and back.
4505 Value *RootVec = nullptr;
4506 for (unsigned i = 0; i != MaskNumElts; ++i) {
4507 // Note that recursion is limited for each vector element, so if any element
4508 // exceeds the limit, this will fail to simplify.
4510 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4512 // We can't replace a widening/narrowing shuffle with one of its operands.
4513 if (!RootVec || RootVec->getType() != RetTy)
4519 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4520 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4521 Type *RetTy, const SimplifyQuery &Q) {
4522 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4525 static Constant *foldConstant(Instruction::UnaryOps Opcode,
4526 Value *&Op, const SimplifyQuery &Q) {
4527 if (auto *C = dyn_cast<Constant>(Op))
4528 return ConstantFoldUnaryOpOperand(Opcode, C, Q.DL);
4532 /// Given the operand for an FNeg, see if we can fold the result. If not, this
4534 static Value *simplifyFNegInst(Value *Op, FastMathFlags FMF,
4535 const SimplifyQuery &Q, unsigned MaxRecurse) {
4536 if (Constant *C = foldConstant(Instruction::FNeg, Op, Q))
4540 // fneg (fneg X) ==> X
4541 if (match(Op, m_FNeg(m_Value(X))))
4547 Value *llvm::SimplifyFNegInst(Value *Op, FastMathFlags FMF,
4548 const SimplifyQuery &Q) {
4549 return ::simplifyFNegInst(Op, FMF, Q, RecursionLimit);
4552 static Constant *propagateNaN(Constant *In) {
4553 // If the input is a vector with undef elements, just return a default NaN.
4555 return ConstantFP::getNaN(In->getType());
4557 // Propagate the existing NaN constant when possible.
4558 // TODO: Should we quiet a signaling NaN?
4562 /// Perform folds that are common to any floating-point operation. This implies
4563 /// transforms based on undef/NaN because the operation itself makes no
4564 /// difference to the result.
4565 static Constant *simplifyFPOp(ArrayRef<Value *> Ops) {
4566 if (any_of(Ops, [](Value *V) { return isa<UndefValue>(V); }))
4567 return ConstantFP::getNaN(Ops[0]->getType());
4569 for (Value *V : Ops)
4570 if (match(V, m_NaN()))
4571 return propagateNaN(cast<Constant>(V));
4576 /// Given operands for an FAdd, see if we can fold the result. If not, this
4578 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4579 const SimplifyQuery &Q, unsigned MaxRecurse) {
4580 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4583 if (Constant *C = simplifyFPOp({Op0, Op1}))
4587 if (match(Op1, m_NegZeroFP()))
4590 // fadd X, 0 ==> X, when we know X is not -0
4591 if (match(Op1, m_PosZeroFP()) &&
4592 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4595 // With nnan: -X + X --> 0.0 (and commuted variant)
4596 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4597 // Negative zeros are allowed because we always end up with positive zero:
4598 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4599 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4600 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4601 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4603 if (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4604 match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0))))
4605 return ConstantFP::getNullValue(Op0->getType());
4607 if (match(Op0, m_FNeg(m_Specific(Op1))) ||
4608 match(Op1, m_FNeg(m_Specific(Op0))))
4609 return ConstantFP::getNullValue(Op0->getType());
4612 // (X - Y) + Y --> X
4613 // Y + (X - Y) --> X
4615 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4616 (match(Op0, m_FSub(m_Value(X), m_Specific(Op1))) ||
4617 match(Op1, m_FSub(m_Value(X), m_Specific(Op0)))))
4623 /// Given operands for an FSub, see if we can fold the result. If not, this
4625 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4626 const SimplifyQuery &Q, unsigned MaxRecurse) {
4627 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4630 if (Constant *C = simplifyFPOp({Op0, Op1}))
4634 if (match(Op1, m_PosZeroFP()))
4637 // fsub X, -0 ==> X, when we know X is not -0
4638 if (match(Op1, m_NegZeroFP()) &&
4639 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4642 // fsub -0.0, (fsub -0.0, X) ==> X
4643 // fsub -0.0, (fneg X) ==> X
4645 if (match(Op0, m_NegZeroFP()) &&
4646 match(Op1, m_FNeg(m_Value(X))))
4649 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4650 // fsub 0.0, (fneg X) ==> X if signed zeros are ignored.
4651 if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4652 (match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))) ||
4653 match(Op1, m_FNeg(m_Value(X)))))
4656 // fsub nnan x, x ==> 0.0
4657 if (FMF.noNaNs() && Op0 == Op1)
4658 return Constant::getNullValue(Op0->getType());
4660 // Y - (Y - X) --> X
4661 // (X + Y) - Y --> X
4662 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4663 (match(Op1, m_FSub(m_Specific(Op0), m_Value(X))) ||
4664 match(Op0, m_c_FAdd(m_Specific(Op1), m_Value(X)))))
4670 static Value *SimplifyFMAFMul(Value *Op0, Value *Op1, FastMathFlags FMF,
4671 const SimplifyQuery &Q, unsigned MaxRecurse) {
4672 if (Constant *C = simplifyFPOp({Op0, Op1}))
4675 // fmul X, 1.0 ==> X
4676 if (match(Op1, m_FPOne()))
4679 // fmul 1.0, X ==> X
4680 if (match(Op0, m_FPOne()))
4683 // fmul nnan nsz X, 0 ==> 0
4684 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4685 return ConstantFP::getNullValue(Op0->getType());
4687 // fmul nnan nsz 0, X ==> 0
4688 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4689 return ConstantFP::getNullValue(Op1->getType());
4691 // sqrt(X) * sqrt(X) --> X, if we can:
4692 // 1. Remove the intermediate rounding (reassociate).
4693 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4694 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4696 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4697 FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4703 /// Given the operands for an FMul, see if we can fold the result
4704 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4705 const SimplifyQuery &Q, unsigned MaxRecurse) {
4706 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4709 // Now apply simplifications that do not require rounding.
4710 return SimplifyFMAFMul(Op0, Op1, FMF, Q, MaxRecurse);
4713 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4714 const SimplifyQuery &Q) {
4715 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4719 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4720 const SimplifyQuery &Q) {
4721 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4724 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4725 const SimplifyQuery &Q) {
4726 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4729 Value *llvm::SimplifyFMAFMul(Value *Op0, Value *Op1, FastMathFlags FMF,
4730 const SimplifyQuery &Q) {
4731 return ::SimplifyFMAFMul(Op0, Op1, FMF, Q, RecursionLimit);
4734 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4735 const SimplifyQuery &Q, unsigned) {
4736 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4739 if (Constant *C = simplifyFPOp({Op0, Op1}))
4743 if (match(Op1, m_FPOne()))
4747 // Requires that NaNs are off (X could be zero) and signed zeroes are
4748 // ignored (X could be positive or negative, so the output sign is unknown).
4749 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4750 return ConstantFP::getNullValue(Op0->getType());
4753 // X / X -> 1.0 is legal when NaNs are ignored.
4754 // We can ignore infinities because INF/INF is NaN.
4756 return ConstantFP::get(Op0->getType(), 1.0);
4758 // (X * Y) / Y --> X if we can reassociate to the above form.
4760 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4763 // -X / X -> -1.0 and
4764 // X / -X -> -1.0 are legal when NaNs are ignored.
4765 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4766 if (match(Op0, m_FNegNSZ(m_Specific(Op1))) ||
4767 match(Op1, m_FNegNSZ(m_Specific(Op0))))
4768 return ConstantFP::get(Op0->getType(), -1.0);
4774 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4775 const SimplifyQuery &Q) {
4776 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4779 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4780 const SimplifyQuery &Q, unsigned) {
4781 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4784 if (Constant *C = simplifyFPOp({Op0, Op1}))
4787 // Unlike fdiv, the result of frem always matches the sign of the dividend.
4788 // The constant match may include undef elements in a vector, so return a full
4789 // zero constant as the result.
4792 if (match(Op0, m_PosZeroFP()))
4793 return ConstantFP::getNullValue(Op0->getType());
4795 if (match(Op0, m_NegZeroFP()))
4796 return ConstantFP::getNegativeZero(Op0->getType());
4802 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4803 const SimplifyQuery &Q) {
4804 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4807 //=== Helper functions for higher up the class hierarchy.
4809 /// Given the operand for a UnaryOperator, see if we can fold the result.
4810 /// If not, this returns null.
4811 static Value *simplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q,
4812 unsigned MaxRecurse) {
4814 case Instruction::FNeg:
4815 return simplifyFNegInst(Op, FastMathFlags(), Q, MaxRecurse);
4817 llvm_unreachable("Unexpected opcode");
4821 /// Given the operand for a UnaryOperator, see if we can fold the result.
4822 /// If not, this returns null.
4823 /// Try to use FastMathFlags when folding the result.
4824 static Value *simplifyFPUnOp(unsigned Opcode, Value *Op,
4825 const FastMathFlags &FMF,
4826 const SimplifyQuery &Q, unsigned MaxRecurse) {
4828 case Instruction::FNeg:
4829 return simplifyFNegInst(Op, FMF, Q, MaxRecurse);
4831 return simplifyUnOp(Opcode, Op, Q, MaxRecurse);
4835 Value *llvm::SimplifyUnOp(unsigned Opcode, Value *Op, const SimplifyQuery &Q) {
4836 return ::simplifyUnOp(Opcode, Op, Q, RecursionLimit);
4839 Value *llvm::SimplifyUnOp(unsigned Opcode, Value *Op, FastMathFlags FMF,
4840 const SimplifyQuery &Q) {
4841 return ::simplifyFPUnOp(Opcode, Op, FMF, Q, RecursionLimit);
4844 /// Given operands for a BinaryOperator, see if we can fold the result.
4845 /// If not, this returns null.
4846 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4847 const SimplifyQuery &Q, unsigned MaxRecurse) {
4849 case Instruction::Add:
4850 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4851 case Instruction::Sub:
4852 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4853 case Instruction::Mul:
4854 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4855 case Instruction::SDiv:
4856 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4857 case Instruction::UDiv:
4858 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4859 case Instruction::SRem:
4860 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4861 case Instruction::URem:
4862 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4863 case Instruction::Shl:
4864 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4865 case Instruction::LShr:
4866 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4867 case Instruction::AShr:
4868 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4869 case Instruction::And:
4870 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4871 case Instruction::Or:
4872 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4873 case Instruction::Xor:
4874 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4875 case Instruction::FAdd:
4876 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4877 case Instruction::FSub:
4878 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4879 case Instruction::FMul:
4880 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4881 case Instruction::FDiv:
4882 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4883 case Instruction::FRem:
4884 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4886 llvm_unreachable("Unexpected opcode");
4890 /// Given operands for a BinaryOperator, see if we can fold the result.
4891 /// If not, this returns null.
4892 /// Try to use FastMathFlags when folding the result.
4893 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4894 const FastMathFlags &FMF, const SimplifyQuery &Q,
4895 unsigned MaxRecurse) {
4897 case Instruction::FAdd:
4898 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4899 case Instruction::FSub:
4900 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4901 case Instruction::FMul:
4902 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4903 case Instruction::FDiv:
4904 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4906 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4910 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4911 const SimplifyQuery &Q) {
4912 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4915 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4916 FastMathFlags FMF, const SimplifyQuery &Q) {
4917 return ::SimplifyBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4920 /// Given operands for a CmpInst, see if we can fold the result.
4921 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4922 const SimplifyQuery &Q, unsigned MaxRecurse) {
4923 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4924 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4925 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4928 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4929 const SimplifyQuery &Q) {
4930 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4933 static bool IsIdempotent(Intrinsic::ID ID) {
4935 default: return false;
4937 // Unary idempotent: f(f(x)) = f(x)
4938 case Intrinsic::fabs:
4939 case Intrinsic::floor:
4940 case Intrinsic::ceil:
4941 case Intrinsic::trunc:
4942 case Intrinsic::rint:
4943 case Intrinsic::nearbyint:
4944 case Intrinsic::round:
4945 case Intrinsic::canonicalize:
4950 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4951 const DataLayout &DL) {
4952 GlobalValue *PtrSym;
4954 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4957 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4958 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4959 Type *Int32PtrTy = Int32Ty->getPointerTo();
4960 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4962 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4963 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4966 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4967 if (OffsetInt % 4 != 0)
4970 Constant *C = ConstantExpr::getGetElementPtr(
4971 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4972 ConstantInt::get(Int64Ty, OffsetInt / 4));
4973 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4977 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4981 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4982 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4987 if (LoadedCE->getOpcode() != Instruction::Sub)
4990 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4991 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4993 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4995 Constant *LoadedRHS = LoadedCE->getOperand(1);
4996 GlobalValue *LoadedRHSSym;
4997 APInt LoadedRHSOffset;
4998 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
5000 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
5003 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
5006 static Value *simplifyUnaryIntrinsic(Function *F, Value *Op0,
5007 const SimplifyQuery &Q) {
5008 // Idempotent functions return the same result when called repeatedly.
5009 Intrinsic::ID IID = F->getIntrinsicID();
5010 if (IsIdempotent(IID))
5011 if (auto *II = dyn_cast<IntrinsicInst>(Op0))
5012 if (II->getIntrinsicID() == IID)
5017 case Intrinsic::fabs:
5018 if (SignBitMustBeZero(Op0, Q.TLI)) return Op0;
5020 case Intrinsic::bswap:
5021 // bswap(bswap(x)) -> x
5022 if (match(Op0, m_BSwap(m_Value(X)))) return X;
5024 case Intrinsic::bitreverse:
5025 // bitreverse(bitreverse(x)) -> x
5026 if (match(Op0, m_BitReverse(m_Value(X)))) return X;
5028 case Intrinsic::exp:
5030 if (Q.CxtI->hasAllowReassoc() &&
5031 match(Op0, m_Intrinsic<Intrinsic::log>(m_Value(X)))) return X;
5033 case Intrinsic::exp2:
5034 // exp2(log2(x)) -> x
5035 if (Q.CxtI->hasAllowReassoc() &&
5036 match(Op0, m_Intrinsic<Intrinsic::log2>(m_Value(X)))) return X;
5038 case Intrinsic::log:
5040 if (Q.CxtI->hasAllowReassoc() &&
5041 match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X)))) return X;
5043 case Intrinsic::log2:
5044 // log2(exp2(x)) -> x
5045 if (Q.CxtI->hasAllowReassoc() &&
5046 (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) ||
5047 match(Op0, m_Intrinsic<Intrinsic::pow>(m_SpecificFP(2.0),
5048 m_Value(X))))) return X;
5050 case Intrinsic::log10:
5051 // log10(pow(10.0, x)) -> x
5052 if (Q.CxtI->hasAllowReassoc() &&
5053 match(Op0, m_Intrinsic<Intrinsic::pow>(m_SpecificFP(10.0),
5054 m_Value(X)))) return X;
5056 case Intrinsic::floor:
5057 case Intrinsic::trunc:
5058 case Intrinsic::ceil:
5059 case Intrinsic::round:
5060 case Intrinsic::nearbyint:
5061 case Intrinsic::rint: {
5062 // floor (sitofp x) -> sitofp x
5063 // floor (uitofp x) -> uitofp x
5065 // Converting from int always results in a finite integral number or
5066 // infinity. For either of those inputs, these rounding functions always
5067 // return the same value, so the rounding can be eliminated.
5068 if (match(Op0, m_SIToFP(m_Value())) || match(Op0, m_UIToFP(m_Value())))
5079 static Value *simplifyBinaryIntrinsic(Function *F, Value *Op0, Value *Op1,
5080 const SimplifyQuery &Q) {
5081 Intrinsic::ID IID = F->getIntrinsicID();
5082 Type *ReturnType = F->getReturnType();
5084 case Intrinsic::usub_with_overflow:
5085 case Intrinsic::ssub_with_overflow:
5086 // X - X -> { 0, false }
5088 return Constant::getNullValue(ReturnType);
5090 case Intrinsic::uadd_with_overflow:
5091 case Intrinsic::sadd_with_overflow:
5092 // X - undef -> { undef, false }
5093 // undef - X -> { undef, false }
5094 // X + undef -> { undef, false }
5095 // undef + x -> { undef, false }
5096 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1)) {
5097 return ConstantStruct::get(
5098 cast<StructType>(ReturnType),
5099 {UndefValue::get(ReturnType->getStructElementType(0)),
5100 Constant::getNullValue(ReturnType->getStructElementType(1))});
5103 case Intrinsic::umul_with_overflow:
5104 case Intrinsic::smul_with_overflow:
5105 // 0 * X -> { 0, false }
5106 // X * 0 -> { 0, false }
5107 if (match(Op0, m_Zero()) || match(Op1, m_Zero()))
5108 return Constant::getNullValue(ReturnType);
5109 // undef * X -> { 0, false }
5110 // X * undef -> { 0, false }
5111 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
5112 return Constant::getNullValue(ReturnType);
5114 case Intrinsic::uadd_sat:
5115 // sat(MAX + X) -> MAX
5116 // sat(X + MAX) -> MAX
5117 if (match(Op0, m_AllOnes()) || match(Op1, m_AllOnes()))
5118 return Constant::getAllOnesValue(ReturnType);
5120 case Intrinsic::sadd_sat:
5121 // sat(X + undef) -> -1
5122 // sat(undef + X) -> -1
5123 // For unsigned: Assume undef is MAX, thus we saturate to MAX (-1).
5124 // For signed: Assume undef is ~X, in which case X + ~X = -1.
5125 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
5126 return Constant::getAllOnesValue(ReturnType);
5129 if (match(Op1, m_Zero()))
5132 if (match(Op0, m_Zero()))
5135 case Intrinsic::usub_sat:
5136 // sat(0 - X) -> 0, sat(X - MAX) -> 0
5137 if (match(Op0, m_Zero()) || match(Op1, m_AllOnes()))
5138 return Constant::getNullValue(ReturnType);
5140 case Intrinsic::ssub_sat:
5141 // X - X -> 0, X - undef -> 0, undef - X -> 0
5142 if (Op0 == Op1 || match(Op0, m_Undef()) || match(Op1, m_Undef()))
5143 return Constant::getNullValue(ReturnType);
5145 if (match(Op1, m_Zero()))
5148 case Intrinsic::load_relative:
5149 if (auto *C0 = dyn_cast<Constant>(Op0))
5150 if (auto *C1 = dyn_cast<Constant>(Op1))
5151 return SimplifyRelativeLoad(C0, C1, Q.DL);
5153 case Intrinsic::powi:
5154 if (auto *Power = dyn_cast<ConstantInt>(Op1)) {
5155 // powi(x, 0) -> 1.0
5156 if (Power->isZero())
5157 return ConstantFP::get(Op0->getType(), 1.0);
5163 case Intrinsic::copysign:
5164 // copysign X, X --> X
5167 // copysign -X, X --> X
5168 // copysign X, -X --> -X
5169 if (match(Op0, m_FNeg(m_Specific(Op1))) ||
5170 match(Op1, m_FNeg(m_Specific(Op0))))
5173 case Intrinsic::maxnum:
5174 case Intrinsic::minnum:
5175 case Intrinsic::maximum:
5176 case Intrinsic::minimum: {
5177 // If the arguments are the same, this is a no-op.
5178 if (Op0 == Op1) return Op0;
5180 // If one argument is undef, return the other argument.
5181 if (match(Op0, m_Undef()))
5183 if (match(Op1, m_Undef()))
5186 // If one argument is NaN, return other or NaN appropriately.
5187 bool PropagateNaN = IID == Intrinsic::minimum || IID == Intrinsic::maximum;
5188 if (match(Op0, m_NaN()))
5189 return PropagateNaN ? Op0 : Op1;
5190 if (match(Op1, m_NaN()))
5191 return PropagateNaN ? Op1 : Op0;
5193 // Min/max of the same operation with common operand:
5194 // m(m(X, Y)), X --> m(X, Y) (4 commuted variants)
5195 if (auto *M0 = dyn_cast<IntrinsicInst>(Op0))
5196 if (M0->getIntrinsicID() == IID &&
5197 (M0->getOperand(0) == Op1 || M0->getOperand(1) == Op1))
5199 if (auto *M1 = dyn_cast<IntrinsicInst>(Op1))
5200 if (M1->getIntrinsicID() == IID &&
5201 (M1->getOperand(0) == Op0 || M1->getOperand(1) == Op0))
5204 // min(X, -Inf) --> -Inf (and commuted variant)
5205 // max(X, +Inf) --> +Inf (and commuted variant)
5206 bool UseNegInf = IID == Intrinsic::minnum || IID == Intrinsic::minimum;
5208 if ((match(Op0, m_APFloat(C)) && C->isInfinity() &&
5209 C->isNegative() == UseNegInf) ||
5210 (match(Op1, m_APFloat(C)) && C->isInfinity() &&
5211 C->isNegative() == UseNegInf))
5212 return ConstantFP::getInfinity(ReturnType, UseNegInf);
5214 // TODO: minnum(nnan x, inf) -> x
5215 // TODO: minnum(nnan ninf x, flt_max) -> x
5216 // TODO: maxnum(nnan x, -inf) -> x
5217 // TODO: maxnum(nnan ninf x, -flt_max) -> x
5227 static Value *simplifyIntrinsic(CallBase *Call, const SimplifyQuery &Q) {
5229 // Intrinsics with no operands have some kind of side effect. Don't simplify.
5230 unsigned NumOperands = Call->getNumArgOperands();
5234 Function *F = cast<Function>(Call->getCalledFunction());
5235 Intrinsic::ID IID = F->getIntrinsicID();
5236 if (NumOperands == 1)
5237 return simplifyUnaryIntrinsic(F, Call->getArgOperand(0), Q);
5239 if (NumOperands == 2)
5240 return simplifyBinaryIntrinsic(F, Call->getArgOperand(0),
5241 Call->getArgOperand(1), Q);
5243 // Handle intrinsics with 3 or more arguments.
5245 case Intrinsic::masked_load:
5246 case Intrinsic::masked_gather: {
5247 Value *MaskArg = Call->getArgOperand(2);
5248 Value *PassthruArg = Call->getArgOperand(3);
5249 // If the mask is all zeros or undef, the "passthru" argument is the result.
5250 if (maskIsAllZeroOrUndef(MaskArg))
5254 case Intrinsic::fshl:
5255 case Intrinsic::fshr: {
5256 Value *Op0 = Call->getArgOperand(0), *Op1 = Call->getArgOperand(1),
5257 *ShAmtArg = Call->getArgOperand(2);
5259 // If both operands are undef, the result is undef.
5260 if (match(Op0, m_Undef()) && match(Op1, m_Undef()))
5261 return UndefValue::get(F->getReturnType());
5263 // If shift amount is undef, assume it is zero.
5264 if (match(ShAmtArg, m_Undef()))
5265 return Call->getArgOperand(IID == Intrinsic::fshl ? 0 : 1);
5267 const APInt *ShAmtC;
5268 if (match(ShAmtArg, m_APInt(ShAmtC))) {
5269 // If there's effectively no shift, return the 1st arg or 2nd arg.
5270 APInt BitWidth = APInt(ShAmtC->getBitWidth(), ShAmtC->getBitWidth());
5271 if (ShAmtC->urem(BitWidth).isNullValue())
5272 return Call->getArgOperand(IID == Intrinsic::fshl ? 0 : 1);
5276 case Intrinsic::fma:
5277 case Intrinsic::fmuladd: {
5278 Value *Op0 = Call->getArgOperand(0);
5279 Value *Op1 = Call->getArgOperand(1);
5280 Value *Op2 = Call->getArgOperand(2);
5281 if (Value *V = simplifyFPOp({ Op0, Op1, Op2 }))
5290 Value *llvm::SimplifyCall(CallBase *Call, const SimplifyQuery &Q) {
5291 Value *Callee = Call->getCalledValue();
5293 // call undef -> undef
5294 // call null -> undef
5295 if (isa<UndefValue>(Callee) || isa<ConstantPointerNull>(Callee))
5296 return UndefValue::get(Call->getType());
5298 Function *F = dyn_cast<Function>(Callee);
5302 if (F->isIntrinsic())
5303 if (Value *Ret = simplifyIntrinsic(Call, Q))
5306 if (!canConstantFoldCallTo(Call, F))
5309 SmallVector<Constant *, 4> ConstantArgs;
5310 unsigned NumArgs = Call->getNumArgOperands();
5311 ConstantArgs.reserve(NumArgs);
5312 for (auto &Arg : Call->args()) {
5313 Constant *C = dyn_cast<Constant>(&Arg);
5316 ConstantArgs.push_back(C);
5319 return ConstantFoldCall(Call, F, ConstantArgs, Q.TLI);
5322 /// Given operands for a Freeze, see if we can fold the result.
5323 static Value *SimplifyFreezeInst(Value *Op0) {
5324 // Use a utility function defined in ValueTracking.
5325 if (llvm::isGuaranteedNotToBeUndefOrPoison(Op0))
5327 // We have room for improvement.
5331 Value *llvm::SimplifyFreezeInst(Value *Op0, const SimplifyQuery &Q) {
5332 return ::SimplifyFreezeInst(Op0);
5335 /// See if we can compute a simplified version of this instruction.
5336 /// If not, this returns null.
5338 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
5339 OptimizationRemarkEmitter *ORE) {
5340 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
5343 switch (I->getOpcode()) {
5345 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
5347 case Instruction::FNeg:
5348 Result = SimplifyFNegInst(I->getOperand(0), I->getFastMathFlags(), Q);
5350 case Instruction::FAdd:
5351 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
5352 I->getFastMathFlags(), Q);
5354 case Instruction::Add:
5356 SimplifyAddInst(I->getOperand(0), I->getOperand(1),
5357 Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
5358 Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
5360 case Instruction::FSub:
5361 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
5362 I->getFastMathFlags(), Q);
5364 case Instruction::Sub:
5366 SimplifySubInst(I->getOperand(0), I->getOperand(1),
5367 Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
5368 Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
5370 case Instruction::FMul:
5371 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
5372 I->getFastMathFlags(), Q);
5374 case Instruction::Mul:
5375 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
5377 case Instruction::SDiv:
5378 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
5380 case Instruction::UDiv:
5381 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
5383 case Instruction::FDiv:
5384 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
5385 I->getFastMathFlags(), Q);
5387 case Instruction::SRem:
5388 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
5390 case Instruction::URem:
5391 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
5393 case Instruction::FRem:
5394 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
5395 I->getFastMathFlags(), Q);
5397 case Instruction::Shl:
5399 SimplifyShlInst(I->getOperand(0), I->getOperand(1),
5400 Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
5401 Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
5403 case Instruction::LShr:
5404 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
5405 Q.IIQ.isExact(cast<BinaryOperator>(I)), Q);
5407 case Instruction::AShr:
5408 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
5409 Q.IIQ.isExact(cast<BinaryOperator>(I)), Q);
5411 case Instruction::And:
5412 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
5414 case Instruction::Or:
5415 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
5417 case Instruction::Xor:
5418 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
5420 case Instruction::ICmp:
5421 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
5422 I->getOperand(0), I->getOperand(1), Q);
5424 case Instruction::FCmp:
5426 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
5427 I->getOperand(1), I->getFastMathFlags(), Q);
5429 case Instruction::Select:
5430 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
5431 I->getOperand(2), Q);
5433 case Instruction::GetElementPtr: {
5434 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
5435 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
5439 case Instruction::InsertValue: {
5440 InsertValueInst *IV = cast<InsertValueInst>(I);
5441 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
5442 IV->getInsertedValueOperand(),
5443 IV->getIndices(), Q);
5446 case Instruction::InsertElement: {
5447 auto *IE = cast<InsertElementInst>(I);
5448 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
5449 IE->getOperand(2), Q);
5452 case Instruction::ExtractValue: {
5453 auto *EVI = cast<ExtractValueInst>(I);
5454 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
5455 EVI->getIndices(), Q);
5458 case Instruction::ExtractElement: {
5459 auto *EEI = cast<ExtractElementInst>(I);
5460 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
5461 EEI->getIndexOperand(), Q);
5464 case Instruction::ShuffleVector: {
5465 auto *SVI = cast<ShuffleVectorInst>(I);
5466 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5467 SVI->getMask(), SVI->getType(), Q);
5470 case Instruction::PHI:
5471 Result = SimplifyPHINode(cast<PHINode>(I), Q);
5473 case Instruction::Call: {
5474 Result = SimplifyCall(cast<CallInst>(I), Q);
5477 case Instruction::Freeze:
5478 Result = SimplifyFreezeInst(I->getOperand(0), Q);
5480 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
5481 #include "llvm/IR/Instruction.def"
5482 #undef HANDLE_CAST_INST
5484 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
5486 case Instruction::Alloca:
5487 // No simplifications for Alloca and it can't be constant folded.
5492 // In general, it is possible for computeKnownBits to determine all bits in a
5493 // value even when the operands are not all constants.
5494 if (!Result && I->getType()->isIntOrIntVectorTy()) {
5495 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
5496 if (Known.isConstant())
5497 Result = ConstantInt::get(I->getType(), Known.getConstant());
5500 /// If called on unreachable code, the above logic may report that the
5501 /// instruction simplified to itself. Make life easier for users by
5502 /// detecting that case here, returning a safe value instead.
5503 return Result == I ? UndefValue::get(I->getType()) : Result;
5506 /// Implementation of recursive simplification through an instruction's
5509 /// This is the common implementation of the recursive simplification routines.
5510 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
5511 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
5512 /// instructions to process and attempt to simplify it using
5513 /// InstructionSimplify. Recursively visited users which could not be
5514 /// simplified themselves are to the optional UnsimplifiedUsers set for
5515 /// further processing by the caller.
5517 /// This routine returns 'true' only when *it* simplifies something. The passed
5518 /// in simplified value does not count toward this.
5519 static bool replaceAndRecursivelySimplifyImpl(
5520 Instruction *I, Value *SimpleV, const TargetLibraryInfo *TLI,
5521 const DominatorTree *DT, AssumptionCache *AC,
5522 SmallSetVector<Instruction *, 8> *UnsimplifiedUsers = nullptr) {
5523 bool Simplified = false;
5524 SmallSetVector<Instruction *, 8> Worklist;
5525 const DataLayout &DL = I->getModule()->getDataLayout();
5527 // If we have an explicit value to collapse to, do that round of the
5528 // simplification loop by hand initially.
5530 for (User *U : I->users())
5532 Worklist.insert(cast<Instruction>(U));
5534 // Replace the instruction with its simplified value.
5535 I->replaceAllUsesWith(SimpleV);
5537 // Gracefully handle edge cases where the instruction is not wired into any
5539 if (I->getParent() && !I->isEHPad() && !I->isTerminator() &&
5540 !I->mayHaveSideEffects())
5541 I->eraseFromParent();
5546 // Note that we must test the size on each iteration, the worklist can grow.
5547 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
5550 // See if this instruction simplifies.
5551 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
5553 if (UnsimplifiedUsers)
5554 UnsimplifiedUsers->insert(I);
5560 // Stash away all the uses of the old instruction so we can check them for
5561 // recursive simplifications after a RAUW. This is cheaper than checking all
5562 // uses of To on the recursive step in most cases.
5563 for (User *U : I->users())
5564 Worklist.insert(cast<Instruction>(U));
5566 // Replace the instruction with its simplified value.
5567 I->replaceAllUsesWith(SimpleV);
5569 // Gracefully handle edge cases where the instruction is not wired into any
5571 if (I->getParent() && !I->isEHPad() && !I->isTerminator() &&
5572 !I->mayHaveSideEffects())
5573 I->eraseFromParent();
5578 bool llvm::recursivelySimplifyInstruction(Instruction *I,
5579 const TargetLibraryInfo *TLI,
5580 const DominatorTree *DT,
5581 AssumptionCache *AC) {
5582 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC, nullptr);
5585 bool llvm::replaceAndRecursivelySimplify(
5586 Instruction *I, Value *SimpleV, const TargetLibraryInfo *TLI,
5587 const DominatorTree *DT, AssumptionCache *AC,
5588 SmallSetVector<Instruction *, 8> *UnsimplifiedUsers) {
5589 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5590 assert(SimpleV && "Must provide a simplified value.");
5591 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC,
5596 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
5597 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
5598 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
5599 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
5600 auto *TLI = TLIWP ? &TLIWP->getTLI(F) : nullptr;
5601 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5602 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5603 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5606 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5607 const DataLayout &DL) {
5608 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5611 template <class T, class... TArgs>
5612 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5614 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5615 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5616 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5617 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5619 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,