1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements routines for folding instructions into simpler forms
11 // that do not require creating new instructions. This does constant folding
12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15 // simplified: This is usually true and assuming it simplifies the logic (if
16 // they have not been simplified then results are correct but maybe suboptimal).
18 //===----------------------------------------------------------------------===//
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/CaptureTracking.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/MemoryBuiltins.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/Analysis/VectorUtils.h"
29 #include "llvm/IR/ConstantRange.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/GetElementPtrTypeIterator.h"
33 #include "llvm/IR/GlobalAlias.h"
34 #include "llvm/IR/Operator.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/ValueHandle.h"
39 using namespace llvm::PatternMatch;
41 #define DEBUG_TYPE "instsimplify"
43 enum { RecursionLimit = 3 };
45 STATISTIC(NumExpand, "Number of expansions");
46 STATISTIC(NumReassoc, "Number of reassociations");
51 const TargetLibraryInfo *TLI;
52 const DominatorTree *DT;
54 const Instruction *CxtI;
56 Query(const DataLayout &DL, const TargetLibraryInfo *tli,
57 const DominatorTree *dt, AssumptionCache *ac = nullptr,
58 const Instruction *cxti = nullptr)
59 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {}
61 } // end anonymous namespace
63 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
64 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
66 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
67 const Query &, unsigned);
68 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
70 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
71 const Query &Q, unsigned MaxRecurse);
72 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
73 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
74 static Value *SimplifyCastInst(unsigned, Value *, Type *,
75 const Query &, unsigned);
77 /// For a boolean type, or a vector of boolean type, return false, or
78 /// a vector with every element false, as appropriate for the type.
79 static Constant *getFalse(Type *Ty) {
80 assert(Ty->getScalarType()->isIntegerTy(1) &&
81 "Expected i1 type or a vector of i1!");
82 return Constant::getNullValue(Ty);
85 /// For a boolean type, or a vector of boolean type, return true, or
86 /// a vector with every element true, as appropriate for the type.
87 static Constant *getTrue(Type *Ty) {
88 assert(Ty->getScalarType()->isIntegerTy(1) &&
89 "Expected i1 type or a vector of i1!");
90 return Constant::getAllOnesValue(Ty);
93 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
94 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
96 CmpInst *Cmp = dyn_cast<CmpInst>(V);
99 CmpInst::Predicate CPred = Cmp->getPredicate();
100 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
101 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
103 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
107 /// Does the given value dominate the specified phi node?
108 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
109 Instruction *I = dyn_cast<Instruction>(V);
111 // Arguments and constants dominate all instructions.
114 // If we are processing instructions (and/or basic blocks) that have not been
115 // fully added to a function, the parent nodes may still be null. Simply
116 // return the conservative answer in these cases.
117 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
120 // If we have a DominatorTree then do a precise test.
122 if (!DT->isReachableFromEntry(P->getParent()))
124 if (!DT->isReachableFromEntry(I->getParent()))
126 return DT->dominates(I, P);
129 // Otherwise, if the instruction is in the entry block and is not an invoke,
130 // then it obviously dominates all phi nodes.
131 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
138 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
139 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
140 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
141 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
142 /// Returns the simplified value, or null if no simplification was performed.
143 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
144 unsigned OpcToExpand, const Query &Q,
145 unsigned MaxRecurse) {
146 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
147 // Recursion is always used, so bail out at once if we already hit the limit.
151 // Check whether the expression has the form "(A op' B) op C".
152 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
153 if (Op0->getOpcode() == OpcodeToExpand) {
154 // It does! Try turning it into "(A op C) op' (B op C)".
155 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
156 // Do "A op C" and "B op C" both simplify?
157 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
158 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
159 // They do! Return "L op' R" if it simplifies or is already available.
160 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
161 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
162 && L == B && R == A)) {
166 // Otherwise return "L op' R" if it simplifies.
167 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
174 // Check whether the expression has the form "A op (B op' C)".
175 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
176 if (Op1->getOpcode() == OpcodeToExpand) {
177 // It does! Try turning it into "(A op B) op' (A op C)".
178 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
179 // Do "A op B" and "A op C" both simplify?
180 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
181 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
182 // They do! Return "L op' R" if it simplifies or is already available.
183 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
184 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
185 && L == C && R == B)) {
189 // Otherwise return "L op' R" if it simplifies.
190 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
200 /// Generic simplifications for associative binary operations.
201 /// Returns the simpler value, or null if none was found.
202 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
203 const Query &Q, unsigned MaxRecurse) {
204 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
205 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
207 // Recursion is always used, so bail out at once if we already hit the limit.
211 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
212 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
214 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
215 if (Op0 && Op0->getOpcode() == Opcode) {
216 Value *A = Op0->getOperand(0);
217 Value *B = Op0->getOperand(1);
220 // Does "B op C" simplify?
221 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
222 // It does! Return "A op V" if it simplifies or is already available.
223 // If V equals B then "A op V" is just the LHS.
224 if (V == B) return LHS;
225 // Otherwise return "A op V" if it simplifies.
226 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
233 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
234 if (Op1 && Op1->getOpcode() == Opcode) {
236 Value *B = Op1->getOperand(0);
237 Value *C = Op1->getOperand(1);
239 // Does "A op B" simplify?
240 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
241 // It does! Return "V op C" if it simplifies or is already available.
242 // If V equals B then "V op C" is just the RHS.
243 if (V == B) return RHS;
244 // Otherwise return "V op C" if it simplifies.
245 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
252 // The remaining transforms require commutativity as well as associativity.
253 if (!Instruction::isCommutative(Opcode))
256 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
257 if (Op0 && Op0->getOpcode() == Opcode) {
258 Value *A = Op0->getOperand(0);
259 Value *B = Op0->getOperand(1);
262 // Does "C op A" simplify?
263 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
264 // It does! Return "V op B" if it simplifies or is already available.
265 // If V equals A then "V op B" is just the LHS.
266 if (V == A) return LHS;
267 // Otherwise return "V op B" if it simplifies.
268 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
275 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
276 if (Op1 && Op1->getOpcode() == Opcode) {
278 Value *B = Op1->getOperand(0);
279 Value *C = Op1->getOperand(1);
281 // Does "C op A" simplify?
282 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
283 // It does! Return "B op V" if it simplifies or is already available.
284 // If V equals C then "B op V" is just the RHS.
285 if (V == C) return RHS;
286 // Otherwise return "B op V" if it simplifies.
287 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
297 /// In the case of a binary operation with a select instruction as an operand,
298 /// try to simplify the binop by seeing whether evaluating it on both branches
299 /// of the select results in the same value. Returns the common value if so,
300 /// otherwise returns null.
301 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
302 const Query &Q, unsigned MaxRecurse) {
303 // Recursion is always used, so bail out at once if we already hit the limit.
308 if (isa<SelectInst>(LHS)) {
309 SI = cast<SelectInst>(LHS);
311 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
312 SI = cast<SelectInst>(RHS);
315 // Evaluate the BinOp on the true and false branches of the select.
319 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
320 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
322 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
323 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
326 // If they simplified to the same value, then return the common value.
327 // If they both failed to simplify then return null.
331 // If one branch simplified to undef, return the other one.
332 if (TV && isa<UndefValue>(TV))
334 if (FV && isa<UndefValue>(FV))
337 // If applying the operation did not change the true and false select values,
338 // then the result of the binop is the select itself.
339 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
342 // If one branch simplified and the other did not, and the simplified
343 // value is equal to the unsimplified one, return the simplified value.
344 // For example, select (cond, X, X & Z) & Z -> X & Z.
345 if ((FV && !TV) || (TV && !FV)) {
346 // Check that the simplified value has the form "X op Y" where "op" is the
347 // same as the original operation.
348 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
349 if (Simplified && Simplified->getOpcode() == Opcode) {
350 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
351 // We already know that "op" is the same as for the simplified value. See
352 // if the operands match too. If so, return the simplified value.
353 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
354 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
355 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
356 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
357 Simplified->getOperand(1) == UnsimplifiedRHS)
359 if (Simplified->isCommutative() &&
360 Simplified->getOperand(1) == UnsimplifiedLHS &&
361 Simplified->getOperand(0) == UnsimplifiedRHS)
369 /// In the case of a comparison with a select instruction, try to simplify the
370 /// comparison by seeing whether both branches of the select result in the same
371 /// value. Returns the common value if so, otherwise returns null.
372 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
373 Value *RHS, const Query &Q,
374 unsigned MaxRecurse) {
375 // Recursion is always used, so bail out at once if we already hit the limit.
379 // Make sure the select is on the LHS.
380 if (!isa<SelectInst>(LHS)) {
382 Pred = CmpInst::getSwappedPredicate(Pred);
384 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
385 SelectInst *SI = cast<SelectInst>(LHS);
386 Value *Cond = SI->getCondition();
387 Value *TV = SI->getTrueValue();
388 Value *FV = SI->getFalseValue();
390 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
391 // Does "cmp TV, RHS" simplify?
392 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
394 // It not only simplified, it simplified to the select condition. Replace
396 TCmp = getTrue(Cond->getType());
398 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
399 // condition then we can replace it with 'true'. Otherwise give up.
400 if (!isSameCompare(Cond, Pred, TV, RHS))
402 TCmp = getTrue(Cond->getType());
405 // Does "cmp FV, RHS" simplify?
406 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
408 // It not only simplified, it simplified to the select condition. Replace
410 FCmp = getFalse(Cond->getType());
412 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
413 // condition then we can replace it with 'false'. Otherwise give up.
414 if (!isSameCompare(Cond, Pred, FV, RHS))
416 FCmp = getFalse(Cond->getType());
419 // If both sides simplified to the same value, then use it as the result of
420 // the original comparison.
424 // The remaining cases only make sense if the select condition has the same
425 // type as the result of the comparison, so bail out if this is not so.
426 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
428 // If the false value simplified to false, then the result of the compare
429 // is equal to "Cond && TCmp". This also catches the case when the false
430 // value simplified to false and the true value to true, returning "Cond".
431 if (match(FCmp, m_Zero()))
432 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
434 // If the true value simplified to true, then the result of the compare
435 // is equal to "Cond || FCmp".
436 if (match(TCmp, m_One()))
437 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
439 // Finally, if the false value simplified to true and the true value to
440 // false, then the result of the compare is equal to "!Cond".
441 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
443 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
450 /// In the case of a binary operation with an operand that is a PHI instruction,
451 /// try to simplify the binop by seeing whether evaluating it on the incoming
452 /// phi values yields the same result for every value. If so returns the common
453 /// value, otherwise returns null.
454 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
455 const Query &Q, unsigned MaxRecurse) {
456 // Recursion is always used, so bail out at once if we already hit the limit.
461 if (isa<PHINode>(LHS)) {
462 PI = cast<PHINode>(LHS);
463 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
464 if (!ValueDominatesPHI(RHS, PI, Q.DT))
467 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
468 PI = cast<PHINode>(RHS);
469 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
470 if (!ValueDominatesPHI(LHS, PI, Q.DT))
474 // Evaluate the BinOp on the incoming phi values.
475 Value *CommonValue = nullptr;
476 for (Value *Incoming : PI->incoming_values()) {
477 // If the incoming value is the phi node itself, it can safely be skipped.
478 if (Incoming == PI) continue;
479 Value *V = PI == LHS ?
480 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
481 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
482 // If the operation failed to simplify, or simplified to a different value
483 // to previously, then give up.
484 if (!V || (CommonValue && V != CommonValue))
492 /// In the case of a comparison with a PHI instruction, try to simplify the
493 /// comparison by seeing whether comparing with all of the incoming phi values
494 /// yields the same result every time. If so returns the common result,
495 /// otherwise returns null.
496 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
497 const Query &Q, unsigned MaxRecurse) {
498 // Recursion is always used, so bail out at once if we already hit the limit.
502 // Make sure the phi is on the LHS.
503 if (!isa<PHINode>(LHS)) {
505 Pred = CmpInst::getSwappedPredicate(Pred);
507 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
508 PHINode *PI = cast<PHINode>(LHS);
510 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
511 if (!ValueDominatesPHI(RHS, PI, Q.DT))
514 // Evaluate the BinOp on the incoming phi values.
515 Value *CommonValue = nullptr;
516 for (Value *Incoming : PI->incoming_values()) {
517 // If the incoming value is the phi node itself, it can safely be skipped.
518 if (Incoming == PI) continue;
519 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
520 // If the operation failed to simplify, or simplified to a different value
521 // to previously, then give up.
522 if (!V || (CommonValue && V != CommonValue))
530 /// Given operands for an Add, see if we can fold the result.
531 /// If not, this returns null.
532 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
533 const Query &Q, unsigned MaxRecurse) {
534 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
535 if (Constant *CRHS = dyn_cast<Constant>(Op1))
536 return ConstantFoldBinaryOpOperands(Instruction::Add, CLHS, CRHS, Q.DL);
538 // Canonicalize the constant to the RHS.
542 // X + undef -> undef
543 if (match(Op1, m_Undef()))
547 if (match(Op1, m_Zero()))
554 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
555 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
558 // X + ~X -> -1 since ~X = -X-1
559 if (match(Op0, m_Not(m_Specific(Op1))) ||
560 match(Op1, m_Not(m_Specific(Op0))))
561 return Constant::getAllOnesValue(Op0->getType());
564 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
565 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
568 // Try some generic simplifications for associative operations.
569 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
573 // Threading Add over selects and phi nodes is pointless, so don't bother.
574 // Threading over the select in "A + select(cond, B, C)" means evaluating
575 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
576 // only if B and C are equal. If B and C are equal then (since we assume
577 // that operands have already been simplified) "select(cond, B, C)" should
578 // have been simplified to the common value of B and C already. Analysing
579 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
580 // for threading over phi nodes.
585 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
586 const DataLayout &DL, const TargetLibraryInfo *TLI,
587 const DominatorTree *DT, AssumptionCache *AC,
588 const Instruction *CxtI) {
589 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
593 /// \brief Compute the base pointer and cumulative constant offsets for V.
595 /// This strips all constant offsets off of V, leaving it the base pointer, and
596 /// accumulates the total constant offset applied in the returned constant. It
597 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
598 /// no constant offsets applied.
600 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
601 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
603 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
604 bool AllowNonInbounds = false) {
605 assert(V->getType()->getScalarType()->isPointerTy());
607 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
608 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
610 // Even though we don't look through PHI nodes, we could be called on an
611 // instruction in an unreachable block, which may be on a cycle.
612 SmallPtrSet<Value *, 4> Visited;
615 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
616 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
617 !GEP->accumulateConstantOffset(DL, Offset))
619 V = GEP->getPointerOperand();
620 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
621 V = cast<Operator>(V)->getOperand(0);
622 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
623 if (GA->isInterposable())
625 V = GA->getAliasee();
627 if (auto CS = CallSite(V))
628 if (Value *RV = CS.getReturnedArgOperand()) {
634 assert(V->getType()->getScalarType()->isPointerTy() &&
635 "Unexpected operand type!");
636 } while (Visited.insert(V).second);
638 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
639 if (V->getType()->isVectorTy())
640 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
645 /// \brief Compute the constant difference between two pointer values.
646 /// If the difference is not a constant, returns zero.
647 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
649 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
650 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
652 // If LHS and RHS are not related via constant offsets to the same base
653 // value, there is nothing we can do here.
657 // Otherwise, the difference of LHS - RHS can be computed as:
659 // = (LHSOffset + Base) - (RHSOffset + Base)
660 // = LHSOffset - RHSOffset
661 return ConstantExpr::getSub(LHSOffset, RHSOffset);
664 /// Given operands for a Sub, see if we can fold the result.
665 /// If not, this returns null.
666 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
667 const Query &Q, unsigned MaxRecurse) {
668 if (Constant *CLHS = dyn_cast<Constant>(Op0))
669 if (Constant *CRHS = dyn_cast<Constant>(Op1))
670 return ConstantFoldBinaryOpOperands(Instruction::Sub, CLHS, CRHS, Q.DL);
672 // X - undef -> undef
673 // undef - X -> undef
674 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
675 return UndefValue::get(Op0->getType());
678 if (match(Op1, m_Zero()))
683 return Constant::getNullValue(Op0->getType());
685 // Is this a negation?
686 if (match(Op0, m_Zero())) {
687 // 0 - X -> 0 if the sub is NUW.
691 unsigned BitWidth = Op1->getType()->getScalarSizeInBits();
692 APInt KnownZero(BitWidth, 0);
693 APInt KnownOne(BitWidth, 0);
694 computeKnownBits(Op1, KnownZero, KnownOne, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
695 if (KnownZero == ~APInt::getSignBit(BitWidth)) {
696 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
697 // Op1 must be 0 because negating the minimum signed value is undefined.
701 // 0 - X -> X if X is 0 or the minimum signed value.
706 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
707 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
708 Value *X = nullptr, *Y = nullptr, *Z = Op1;
709 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
710 // See if "V === Y - Z" simplifies.
711 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
712 // It does! Now see if "X + V" simplifies.
713 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
714 // It does, we successfully reassociated!
718 // See if "V === X - Z" simplifies.
719 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
720 // It does! Now see if "Y + V" simplifies.
721 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
722 // It does, we successfully reassociated!
728 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
729 // For example, X - (X + 1) -> -1
731 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
732 // See if "V === X - Y" simplifies.
733 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
734 // It does! Now see if "V - Z" simplifies.
735 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
736 // It does, we successfully reassociated!
740 // See if "V === X - Z" simplifies.
741 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
742 // It does! Now see if "V - Y" simplifies.
743 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
744 // It does, we successfully reassociated!
750 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
751 // For example, X - (X - Y) -> Y.
753 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
754 // See if "V === Z - X" simplifies.
755 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
756 // It does! Now see if "V + Y" simplifies.
757 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
758 // It does, we successfully reassociated!
763 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
764 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
765 match(Op1, m_Trunc(m_Value(Y))))
766 if (X->getType() == Y->getType())
767 // See if "V === X - Y" simplifies.
768 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
769 // It does! Now see if "trunc V" simplifies.
770 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
772 // It does, return the simplified "trunc V".
775 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
776 if (match(Op0, m_PtrToInt(m_Value(X))) &&
777 match(Op1, m_PtrToInt(m_Value(Y))))
778 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
779 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
782 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
783 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
786 // Threading Sub over selects and phi nodes is pointless, so don't bother.
787 // Threading over the select in "A - select(cond, B, C)" means evaluating
788 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
789 // only if B and C are equal. If B and C are equal then (since we assume
790 // that operands have already been simplified) "select(cond, B, C)" should
791 // have been simplified to the common value of B and C already. Analysing
792 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
793 // for threading over phi nodes.
798 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
799 const DataLayout &DL, const TargetLibraryInfo *TLI,
800 const DominatorTree *DT, AssumptionCache *AC,
801 const Instruction *CxtI) {
802 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
806 /// Given operands for an FAdd, see if we can fold the result. If not, this
808 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
809 const Query &Q, unsigned MaxRecurse) {
810 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
811 if (Constant *CRHS = dyn_cast<Constant>(Op1))
812 return ConstantFoldBinaryOpOperands(Instruction::FAdd, CLHS, CRHS, Q.DL);
814 // Canonicalize the constant to the RHS.
819 if (match(Op1, m_NegZero()))
822 // fadd X, 0 ==> X, when we know X is not -0
823 if (match(Op1, m_Zero()) &&
824 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
827 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
828 // where nnan and ninf have to occur at least once somewhere in this
830 Value *SubOp = nullptr;
831 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
833 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
836 Instruction *FSub = cast<Instruction>(SubOp);
837 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
838 (FMF.noInfs() || FSub->hasNoInfs()))
839 return Constant::getNullValue(Op0->getType());
845 /// Given operands for an FSub, see if we can fold the result. If not, this
847 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
848 const Query &Q, unsigned MaxRecurse) {
849 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
850 if (Constant *CRHS = dyn_cast<Constant>(Op1))
851 return ConstantFoldBinaryOpOperands(Instruction::FSub, CLHS, CRHS, Q.DL);
855 if (match(Op1, m_Zero()))
858 // fsub X, -0 ==> X, when we know X is not -0
859 if (match(Op1, m_NegZero()) &&
860 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
863 // fsub -0.0, (fsub -0.0, X) ==> X
865 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
868 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
869 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
870 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
873 // fsub nnan x, x ==> 0.0
874 if (FMF.noNaNs() && Op0 == Op1)
875 return Constant::getNullValue(Op0->getType());
880 /// Given the operands for an FMul, see if we can fold the result
881 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
884 unsigned MaxRecurse) {
885 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
886 if (Constant *CRHS = dyn_cast<Constant>(Op1))
887 return ConstantFoldBinaryOpOperands(Instruction::FMul, CLHS, CRHS, Q.DL);
889 // Canonicalize the constant to the RHS.
894 if (match(Op1, m_FPOne()))
897 // fmul nnan nsz X, 0 ==> 0
898 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
904 /// Given operands for a Mul, see if we can fold the result.
905 /// If not, this returns null.
906 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
907 unsigned MaxRecurse) {
908 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
909 if (Constant *CRHS = dyn_cast<Constant>(Op1))
910 return ConstantFoldBinaryOpOperands(Instruction::Mul, CLHS, CRHS, Q.DL);
912 // Canonicalize the constant to the RHS.
917 if (match(Op1, m_Undef()))
918 return Constant::getNullValue(Op0->getType());
921 if (match(Op1, m_Zero()))
925 if (match(Op1, m_One()))
928 // (X / Y) * Y -> X if the division is exact.
930 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
931 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
935 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
936 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
939 // Try some generic simplifications for associative operations.
940 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
944 // Mul distributes over Add. Try some generic simplifications based on this.
945 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
949 // If the operation is with the result of a select instruction, check whether
950 // operating on either branch of the select always yields the same value.
951 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
952 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
956 // If the operation is with the result of a phi instruction, check whether
957 // operating on all incoming values of the phi always yields the same value.
958 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
959 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
966 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
967 const DataLayout &DL,
968 const TargetLibraryInfo *TLI,
969 const DominatorTree *DT, AssumptionCache *AC,
970 const Instruction *CxtI) {
971 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
975 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
976 const DataLayout &DL,
977 const TargetLibraryInfo *TLI,
978 const DominatorTree *DT, AssumptionCache *AC,
979 const Instruction *CxtI) {
980 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
984 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
985 const DataLayout &DL,
986 const TargetLibraryInfo *TLI,
987 const DominatorTree *DT, AssumptionCache *AC,
988 const Instruction *CxtI) {
989 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
993 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout &DL,
994 const TargetLibraryInfo *TLI,
995 const DominatorTree *DT, AssumptionCache *AC,
996 const Instruction *CxtI) {
997 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1001 /// Given operands for an SDiv or UDiv, see if we can fold the result.
1002 /// If not, this returns null.
1003 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1004 const Query &Q, unsigned MaxRecurse) {
1005 if (Constant *C0 = dyn_cast<Constant>(Op0))
1006 if (Constant *C1 = dyn_cast<Constant>(Op1))
1007 return ConstantFoldBinaryOpOperands(Opcode, C0, C1, Q.DL);
1009 bool isSigned = Opcode == Instruction::SDiv;
1011 // X / undef -> undef
1012 if (match(Op1, m_Undef()))
1015 // X / 0 -> undef, we don't need to preserve faults!
1016 if (match(Op1, m_Zero()))
1017 return UndefValue::get(Op1->getType());
1020 if (match(Op0, m_Undef()))
1021 return Constant::getNullValue(Op0->getType());
1023 // 0 / X -> 0, we don't need to preserve faults!
1024 if (match(Op0, m_Zero()))
1028 if (match(Op1, m_One()))
1031 if (Op0->getType()->isIntegerTy(1))
1032 // It can't be division by zero, hence it must be division by one.
1037 return ConstantInt::get(Op0->getType(), 1);
1039 // (X * Y) / Y -> X if the multiplication does not overflow.
1040 Value *X = nullptr, *Y = nullptr;
1041 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1042 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1043 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1044 // If the Mul knows it does not overflow, then we are good to go.
1045 if ((isSigned && Mul->hasNoSignedWrap()) ||
1046 (!isSigned && Mul->hasNoUnsignedWrap()))
1048 // If X has the form X = A / Y then X * Y cannot overflow.
1049 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1050 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1054 // (X rem Y) / Y -> 0
1055 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1056 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1057 return Constant::getNullValue(Op0->getType());
1059 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1060 ConstantInt *C1, *C2;
1061 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1062 match(Op1, m_ConstantInt(C2))) {
1064 C1->getValue().umul_ov(C2->getValue(), Overflow);
1066 return Constant::getNullValue(Op0->getType());
1069 // If the operation is with the result of a select instruction, check whether
1070 // operating on either branch of the select always yields the same value.
1071 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1072 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1075 // If the operation is with the result of a phi instruction, check whether
1076 // operating on all incoming values of the phi always yields the same value.
1077 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1078 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1084 /// Given operands for an SDiv, see if we can fold the result.
1085 /// If not, this returns null.
1086 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1087 unsigned MaxRecurse) {
1088 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1094 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1095 const TargetLibraryInfo *TLI,
1096 const DominatorTree *DT, AssumptionCache *AC,
1097 const Instruction *CxtI) {
1098 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1102 /// Given operands for a UDiv, see if we can fold the result.
1103 /// If not, this returns null.
1104 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1105 unsigned MaxRecurse) {
1106 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1109 // udiv %V, C -> 0 if %V < C
1111 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1112 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1113 if (C->isAllOnesValue()) {
1114 return Constant::getNullValue(Op0->getType());
1122 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1123 const TargetLibraryInfo *TLI,
1124 const DominatorTree *DT, AssumptionCache *AC,
1125 const Instruction *CxtI) {
1126 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1130 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1131 const Query &Q, unsigned) {
1132 // undef / X -> undef (the undef could be a snan).
1133 if (match(Op0, m_Undef()))
1136 // X / undef -> undef
1137 if (match(Op1, m_Undef()))
1141 if (match(Op1, m_FPOne()))
1145 // Requires that NaNs are off (X could be zero) and signed zeroes are
1146 // ignored (X could be positive or negative, so the output sign is unknown).
1147 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1151 // X / X -> 1.0 is legal when NaNs are ignored.
1153 return ConstantFP::get(Op0->getType(), 1.0);
1155 // -X / X -> -1.0 and
1156 // X / -X -> -1.0 are legal when NaNs are ignored.
1157 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1158 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1159 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1160 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1161 BinaryOperator::getFNegArgument(Op1) == Op0))
1162 return ConstantFP::get(Op0->getType(), -1.0);
1168 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1169 const DataLayout &DL,
1170 const TargetLibraryInfo *TLI,
1171 const DominatorTree *DT, AssumptionCache *AC,
1172 const Instruction *CxtI) {
1173 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1177 /// Given operands for an SRem or URem, see if we can fold the result.
1178 /// If not, this returns null.
1179 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1180 const Query &Q, unsigned MaxRecurse) {
1181 if (Constant *C0 = dyn_cast<Constant>(Op0))
1182 if (Constant *C1 = dyn_cast<Constant>(Op1))
1183 return ConstantFoldBinaryOpOperands(Opcode, C0, C1, Q.DL);
1185 // X % undef -> undef
1186 if (match(Op1, m_Undef()))
1190 if (match(Op0, m_Undef()))
1191 return Constant::getNullValue(Op0->getType());
1193 // 0 % X -> 0, we don't need to preserve faults!
1194 if (match(Op0, m_Zero()))
1197 // X % 0 -> undef, we don't need to preserve faults!
1198 if (match(Op1, m_Zero()))
1199 return UndefValue::get(Op0->getType());
1202 if (match(Op1, m_One()))
1203 return Constant::getNullValue(Op0->getType());
1205 if (Op0->getType()->isIntegerTy(1))
1206 // It can't be remainder by zero, hence it must be remainder by one.
1207 return Constant::getNullValue(Op0->getType());
1211 return Constant::getNullValue(Op0->getType());
1213 // (X % Y) % Y -> X % Y
1214 if ((Opcode == Instruction::SRem &&
1215 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1216 (Opcode == Instruction::URem &&
1217 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1220 // If the operation is with the result of a select instruction, check whether
1221 // operating on either branch of the select always yields the same value.
1222 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1223 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1226 // If the operation is with the result of a phi instruction, check whether
1227 // operating on all incoming values of the phi always yields the same value.
1228 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1229 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1235 /// Given operands for an SRem, see if we can fold the result.
1236 /// If not, this returns null.
1237 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1238 unsigned MaxRecurse) {
1239 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1245 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1246 const TargetLibraryInfo *TLI,
1247 const DominatorTree *DT, AssumptionCache *AC,
1248 const Instruction *CxtI) {
1249 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1253 /// Given operands for a URem, see if we can fold the result.
1254 /// If not, this returns null.
1255 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1256 unsigned MaxRecurse) {
1257 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1260 // urem %V, C -> %V if %V < C
1262 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1263 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1264 if (C->isAllOnesValue()) {
1273 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1274 const TargetLibraryInfo *TLI,
1275 const DominatorTree *DT, AssumptionCache *AC,
1276 const Instruction *CxtI) {
1277 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1281 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1282 const Query &, unsigned) {
1283 // undef % X -> undef (the undef could be a snan).
1284 if (match(Op0, m_Undef()))
1287 // X % undef -> undef
1288 if (match(Op1, m_Undef()))
1292 // Requires that NaNs are off (X could be zero) and signed zeroes are
1293 // ignored (X could be positive or negative, so the output sign is unknown).
1294 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1300 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1301 const DataLayout &DL,
1302 const TargetLibraryInfo *TLI,
1303 const DominatorTree *DT, AssumptionCache *AC,
1304 const Instruction *CxtI) {
1305 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1309 /// Returns true if a shift by \c Amount always yields undef.
1310 static bool isUndefShift(Value *Amount) {
1311 Constant *C = dyn_cast<Constant>(Amount);
1315 // X shift by undef -> undef because it may shift by the bitwidth.
1316 if (isa<UndefValue>(C))
1319 // Shifting by the bitwidth or more is undefined.
1320 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1321 if (CI->getValue().getLimitedValue() >=
1322 CI->getType()->getScalarSizeInBits())
1325 // If all lanes of a vector shift are undefined the whole shift is.
1326 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1327 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1328 if (!isUndefShift(C->getAggregateElement(I)))
1336 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1337 /// If not, this returns null.
1338 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1339 const Query &Q, unsigned MaxRecurse) {
1340 if (Constant *C0 = dyn_cast<Constant>(Op0))
1341 if (Constant *C1 = dyn_cast<Constant>(Op1))
1342 return ConstantFoldBinaryOpOperands(Opcode, C0, C1, Q.DL);
1344 // 0 shift by X -> 0
1345 if (match(Op0, m_Zero()))
1348 // X shift by 0 -> X
1349 if (match(Op1, m_Zero()))
1352 // Fold undefined shifts.
1353 if (isUndefShift(Op1))
1354 return UndefValue::get(Op0->getType());
1356 // If the operation is with the result of a select instruction, check whether
1357 // operating on either branch of the select always yields the same value.
1358 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1359 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1362 // If the operation is with the result of a phi instruction, check whether
1363 // operating on all incoming values of the phi always yields the same value.
1364 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1365 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1368 // If any bits in the shift amount make that value greater than or equal to
1369 // the number of bits in the type, the shift is undefined.
1370 unsigned BitWidth = Op1->getType()->getScalarSizeInBits();
1371 APInt KnownZero(BitWidth, 0);
1372 APInt KnownOne(BitWidth, 0);
1373 computeKnownBits(Op1, KnownZero, KnownOne, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1374 if (KnownOne.getLimitedValue() >= BitWidth)
1375 return UndefValue::get(Op0->getType());
1377 // If all valid bits in the shift amount are known zero, the first operand is
1379 unsigned NumValidShiftBits = Log2_32_Ceil(BitWidth);
1380 APInt ShiftAmountMask = APInt::getLowBitsSet(BitWidth, NumValidShiftBits);
1381 if ((KnownZero & ShiftAmountMask) == ShiftAmountMask)
1387 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1388 /// fold the result. If not, this returns null.
1389 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1,
1390 bool isExact, const Query &Q,
1391 unsigned MaxRecurse) {
1392 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1397 return Constant::getNullValue(Op0->getType());
1400 // undef >> X -> undef (if it's exact)
1401 if (match(Op0, m_Undef()))
1402 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1404 // The low bit cannot be shifted out of an exact shift if it is set.
1406 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
1407 APInt Op0KnownZero(BitWidth, 0);
1408 APInt Op0KnownOne(BitWidth, 0);
1409 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC,
1418 /// Given operands for an Shl, see if we can fold the result.
1419 /// If not, this returns null.
1420 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1421 const Query &Q, unsigned MaxRecurse) {
1422 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1426 // undef << X -> undef if (if it's NSW/NUW)
1427 if (match(Op0, m_Undef()))
1428 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1430 // (X >> A) << A -> X
1432 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1437 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1438 const DataLayout &DL, const TargetLibraryInfo *TLI,
1439 const DominatorTree *DT, AssumptionCache *AC,
1440 const Instruction *CxtI) {
1441 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
1445 /// Given operands for an LShr, see if we can fold the result.
1446 /// If not, this returns null.
1447 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1448 const Query &Q, unsigned MaxRecurse) {
1449 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1453 // (X << A) >> A -> X
1455 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1461 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1462 const DataLayout &DL,
1463 const TargetLibraryInfo *TLI,
1464 const DominatorTree *DT, AssumptionCache *AC,
1465 const Instruction *CxtI) {
1466 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1470 /// Given operands for an AShr, see if we can fold the result.
1471 /// If not, this returns null.
1472 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1473 const Query &Q, unsigned MaxRecurse) {
1474 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1478 // all ones >>a X -> all ones
1479 if (match(Op0, m_AllOnes()))
1482 // (X << A) >> A -> X
1484 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1487 // Arithmetic shifting an all-sign-bit value is a no-op.
1488 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1489 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1495 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1496 const DataLayout &DL,
1497 const TargetLibraryInfo *TLI,
1498 const DominatorTree *DT, AssumptionCache *AC,
1499 const Instruction *CxtI) {
1500 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1504 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1505 ICmpInst *UnsignedICmp, bool IsAnd) {
1508 ICmpInst::Predicate EqPred;
1509 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1510 !ICmpInst::isEquality(EqPred))
1513 ICmpInst::Predicate UnsignedPred;
1514 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1515 ICmpInst::isUnsigned(UnsignedPred))
1517 else if (match(UnsignedICmp,
1518 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1519 ICmpInst::isUnsigned(UnsignedPred))
1520 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1524 // X < Y && Y != 0 --> X < Y
1525 // X < Y || Y != 0 --> Y != 0
1526 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1527 return IsAnd ? UnsignedICmp : ZeroICmp;
1529 // X >= Y || Y != 0 --> true
1530 // X >= Y || Y == 0 --> X >= Y
1531 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1532 if (EqPred == ICmpInst::ICMP_NE)
1533 return getTrue(UnsignedICmp->getType());
1534 return UnsignedICmp;
1537 // X < Y && Y == 0 --> false
1538 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1540 return getFalse(UnsignedICmp->getType());
1545 /// Commuted variants are assumed to be handled by calling this function again
1546 /// with the parameters swapped.
1547 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1548 ICmpInst::Predicate Pred0, Pred1;
1550 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1551 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1554 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1555 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1556 // can eliminate Op1 from this 'and'.
1557 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1560 // Check for any combination of predicates that are guaranteed to be disjoint.
1561 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1562 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1563 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1564 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1565 return getFalse(Op0->getType());
1570 /// Commuted variants are assumed to be handled by calling this function again
1571 /// with the parameters swapped.
1572 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1573 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1576 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1579 // Look for this pattern: (icmp V, C0) & (icmp V, C1)).
1580 Type *ITy = Op0->getType();
1581 ICmpInst::Predicate Pred0, Pred1;
1582 const APInt *C0, *C1;
1584 if (match(Op0, m_ICmp(Pred0, m_Value(V), m_APInt(C0))) &&
1585 match(Op1, m_ICmp(Pred1, m_Specific(V), m_APInt(C1)))) {
1586 // Make a constant range that's the intersection of the two icmp ranges.
1587 // If the intersection is empty, we know that the result is false.
1588 auto Range0 = ConstantRange::makeAllowedICmpRegion(Pred0, *C0);
1589 auto Range1 = ConstantRange::makeAllowedICmpRegion(Pred1, *C1);
1590 if (Range0.intersectWith(Range1).isEmptySet())
1591 return getFalse(ITy);
1594 // (icmp (add V, C0), C1) & (icmp V, C0)
1595 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1598 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1601 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1602 if (AddInst->getOperand(1) != Op1->getOperand(1))
1605 bool isNSW = AddInst->hasNoSignedWrap();
1606 bool isNUW = AddInst->hasNoUnsignedWrap();
1608 const APInt Delta = *C1 - *C0;
1609 if (C0->isStrictlyPositive()) {
1611 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1612 return getFalse(ITy);
1613 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1614 return getFalse(ITy);
1617 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1618 return getFalse(ITy);
1619 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1620 return getFalse(ITy);
1623 if (C0->getBoolValue() && isNUW) {
1625 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1626 return getFalse(ITy);
1628 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1629 return getFalse(ITy);
1635 /// Given operands for an And, see if we can fold the result.
1636 /// If not, this returns null.
1637 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1638 unsigned MaxRecurse) {
1639 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1640 if (Constant *CRHS = dyn_cast<Constant>(Op1))
1641 return ConstantFoldBinaryOpOperands(Instruction::And, CLHS, CRHS, Q.DL);
1643 // Canonicalize the constant to the RHS.
1644 std::swap(Op0, Op1);
1648 if (match(Op1, m_Undef()))
1649 return Constant::getNullValue(Op0->getType());
1656 if (match(Op1, m_Zero()))
1660 if (match(Op1, m_AllOnes()))
1663 // A & ~A = ~A & A = 0
1664 if (match(Op0, m_Not(m_Specific(Op1))) ||
1665 match(Op1, m_Not(m_Specific(Op0))))
1666 return Constant::getNullValue(Op0->getType());
1669 Value *A = nullptr, *B = nullptr;
1670 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1671 (A == Op1 || B == Op1))
1675 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1676 (A == Op0 || B == Op0))
1679 // A & (-A) = A if A is a power of two or zero.
1680 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1681 match(Op1, m_Neg(m_Specific(Op0)))) {
1682 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1685 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1690 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1691 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1692 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS))
1694 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS))
1699 // The compares may be hidden behind casts. Look through those and try the
1700 // same folds as above.
1701 auto *Cast0 = dyn_cast<CastInst>(Op0);
1702 auto *Cast1 = dyn_cast<CastInst>(Op1);
1703 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1704 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1705 auto *Cmp0 = dyn_cast<ICmpInst>(Cast0->getOperand(0));
1706 auto *Cmp1 = dyn_cast<ICmpInst>(Cast1->getOperand(0));
1708 Instruction::CastOps CastOpc = Cast0->getOpcode();
1709 Type *ResultType = Cast0->getType();
1710 if (auto *V = dyn_cast_or_null<Constant>(SimplifyAndOfICmps(Cmp0, Cmp1)))
1711 return ConstantExpr::getCast(CastOpc, V, ResultType);
1712 if (auto *V = dyn_cast_or_null<Constant>(SimplifyAndOfICmps(Cmp1, Cmp0)))
1713 return ConstantExpr::getCast(CastOpc, V, ResultType);
1717 // Try some generic simplifications for associative operations.
1718 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1722 // And distributes over Or. Try some generic simplifications based on this.
1723 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1727 // And distributes over Xor. Try some generic simplifications based on this.
1728 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1732 // If the operation is with the result of a select instruction, check whether
1733 // operating on either branch of the select always yields the same value.
1734 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1735 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1739 // If the operation is with the result of a phi instruction, check whether
1740 // operating on all incoming values of the phi always yields the same value.
1741 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1742 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1749 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout &DL,
1750 const TargetLibraryInfo *TLI,
1751 const DominatorTree *DT, AssumptionCache *AC,
1752 const Instruction *CxtI) {
1753 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1757 /// Commuted variants are assumed to be handled by calling this function again
1758 /// with the parameters swapped.
1759 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1760 ICmpInst::Predicate Pred0, Pred1;
1762 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1763 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1766 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1767 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1768 // can eliminate Op0 from this 'or'.
1769 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1772 // Check for any combination of predicates that cover the entire range of
1774 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1775 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1776 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1777 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1778 return getTrue(Op0->getType());
1783 /// Commuted variants are assumed to be handled by calling this function again
1784 /// with the parameters swapped.
1785 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1786 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1789 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1792 // (icmp (add V, C0), C1) | (icmp V, C0)
1793 ICmpInst::Predicate Pred0, Pred1;
1794 const APInt *C0, *C1;
1796 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1799 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1802 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1803 if (AddInst->getOperand(1) != Op1->getOperand(1))
1806 Type *ITy = Op0->getType();
1807 bool isNSW = AddInst->hasNoSignedWrap();
1808 bool isNUW = AddInst->hasNoUnsignedWrap();
1810 const APInt Delta = *C1 - *C0;
1811 if (C0->isStrictlyPositive()) {
1813 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1814 return getTrue(ITy);
1815 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1816 return getTrue(ITy);
1819 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1820 return getTrue(ITy);
1821 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1822 return getTrue(ITy);
1825 if (C0->getBoolValue() && isNUW) {
1827 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1828 return getTrue(ITy);
1830 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1831 return getTrue(ITy);
1837 /// Given operands for an Or, see if we can fold the result.
1838 /// If not, this returns null.
1839 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1840 unsigned MaxRecurse) {
1841 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1842 if (Constant *CRHS = dyn_cast<Constant>(Op1))
1843 return ConstantFoldBinaryOpOperands(Instruction::Or, CLHS, CRHS, Q.DL);
1845 // Canonicalize the constant to the RHS.
1846 std::swap(Op0, Op1);
1850 if (match(Op1, m_Undef()))
1851 return Constant::getAllOnesValue(Op0->getType());
1858 if (match(Op1, m_Zero()))
1862 if (match(Op1, m_AllOnes()))
1865 // A | ~A = ~A | A = -1
1866 if (match(Op0, m_Not(m_Specific(Op1))) ||
1867 match(Op1, m_Not(m_Specific(Op0))))
1868 return Constant::getAllOnesValue(Op0->getType());
1871 Value *A = nullptr, *B = nullptr;
1872 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1873 (A == Op1 || B == Op1))
1877 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1878 (A == Op0 || B == Op0))
1881 // ~(A & ?) | A = -1
1882 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1883 (A == Op1 || B == Op1))
1884 return Constant::getAllOnesValue(Op1->getType());
1886 // A | ~(A & ?) = -1
1887 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1888 (A == Op0 || B == Op0))
1889 return Constant::getAllOnesValue(Op0->getType());
1891 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1892 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1893 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS))
1895 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS))
1900 // Try some generic simplifications for associative operations.
1901 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1905 // Or distributes over And. Try some generic simplifications based on this.
1906 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1910 // If the operation is with the result of a select instruction, check whether
1911 // operating on either branch of the select always yields the same value.
1912 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1913 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1918 Value *C = nullptr, *D = nullptr;
1919 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1920 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1921 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1922 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1923 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1924 // (A & C1)|(B & C2)
1925 // If we have: ((V + N) & C1) | (V & C2)
1926 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1927 // replace with V+N.
1929 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1930 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1931 // Add commutes, try both ways.
1933 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1936 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1939 // Or commutes, try both ways.
1940 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1941 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1942 // Add commutes, try both ways.
1944 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1947 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1953 // If the operation is with the result of a phi instruction, check whether
1954 // operating on all incoming values of the phi always yields the same value.
1955 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1956 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1962 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout &DL,
1963 const TargetLibraryInfo *TLI,
1964 const DominatorTree *DT, AssumptionCache *AC,
1965 const Instruction *CxtI) {
1966 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1970 /// Given operands for a Xor, see if we can fold the result.
1971 /// If not, this returns null.
1972 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1973 unsigned MaxRecurse) {
1974 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1975 if (Constant *CRHS = dyn_cast<Constant>(Op1))
1976 return ConstantFoldBinaryOpOperands(Instruction::Xor, CLHS, CRHS, Q.DL);
1978 // Canonicalize the constant to the RHS.
1979 std::swap(Op0, Op1);
1982 // A ^ undef -> undef
1983 if (match(Op1, m_Undef()))
1987 if (match(Op1, m_Zero()))
1992 return Constant::getNullValue(Op0->getType());
1994 // A ^ ~A = ~A ^ A = -1
1995 if (match(Op0, m_Not(m_Specific(Op1))) ||
1996 match(Op1, m_Not(m_Specific(Op0))))
1997 return Constant::getAllOnesValue(Op0->getType());
1999 // Try some generic simplifications for associative operations.
2000 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2004 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2005 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2006 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2007 // only if B and C are equal. If B and C are equal then (since we assume
2008 // that operands have already been simplified) "select(cond, B, C)" should
2009 // have been simplified to the common value of B and C already. Analysing
2010 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2011 // for threading over phi nodes.
2016 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout &DL,
2017 const TargetLibraryInfo *TLI,
2018 const DominatorTree *DT, AssumptionCache *AC,
2019 const Instruction *CxtI) {
2020 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
2024 static Type *GetCompareTy(Value *Op) {
2025 return CmpInst::makeCmpResultType(Op->getType());
2028 /// Rummage around inside V looking for something equivalent to the comparison
2029 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2030 /// Helper function for analyzing max/min idioms.
2031 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2032 Value *LHS, Value *RHS) {
2033 SelectInst *SI = dyn_cast<SelectInst>(V);
2036 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2039 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2040 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2042 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2043 LHS == CmpRHS && RHS == CmpLHS)
2048 // A significant optimization not implemented here is assuming that alloca
2049 // addresses are not equal to incoming argument values. They don't *alias*,
2050 // as we say, but that doesn't mean they aren't equal, so we take a
2051 // conservative approach.
2053 // This is inspired in part by C++11 5.10p1:
2054 // "Two pointers of the same type compare equal if and only if they are both
2055 // null, both point to the same function, or both represent the same
2058 // This is pretty permissive.
2060 // It's also partly due to C11 6.5.9p6:
2061 // "Two pointers compare equal if and only if both are null pointers, both are
2062 // pointers to the same object (including a pointer to an object and a
2063 // subobject at its beginning) or function, both are pointers to one past the
2064 // last element of the same array object, or one is a pointer to one past the
2065 // end of one array object and the other is a pointer to the start of a
2066 // different array object that happens to immediately follow the first array
2067 // object in the address space.)
2069 // C11's version is more restrictive, however there's no reason why an argument
2070 // couldn't be a one-past-the-end value for a stack object in the caller and be
2071 // equal to the beginning of a stack object in the callee.
2073 // If the C and C++ standards are ever made sufficiently restrictive in this
2074 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2075 // this optimization.
2077 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2078 const DominatorTree *DT, CmpInst::Predicate Pred,
2079 const Instruction *CxtI, Value *LHS, Value *RHS) {
2080 // First, skip past any trivial no-ops.
2081 LHS = LHS->stripPointerCasts();
2082 RHS = RHS->stripPointerCasts();
2084 // A non-null pointer is not equal to a null pointer.
2085 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) &&
2086 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2087 return ConstantInt::get(GetCompareTy(LHS),
2088 !CmpInst::isTrueWhenEqual(Pred));
2090 // We can only fold certain predicates on pointer comparisons.
2095 // Equality comaprisons are easy to fold.
2096 case CmpInst::ICMP_EQ:
2097 case CmpInst::ICMP_NE:
2100 // We can only handle unsigned relational comparisons because 'inbounds' on
2101 // a GEP only protects against unsigned wrapping.
2102 case CmpInst::ICMP_UGT:
2103 case CmpInst::ICMP_UGE:
2104 case CmpInst::ICMP_ULT:
2105 case CmpInst::ICMP_ULE:
2106 // However, we have to switch them to their signed variants to handle
2107 // negative indices from the base pointer.
2108 Pred = ICmpInst::getSignedPredicate(Pred);
2112 // Strip off any constant offsets so that we can reason about them.
2113 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2114 // here and compare base addresses like AliasAnalysis does, however there are
2115 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2116 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2117 // doesn't need to guarantee pointer inequality when it says NoAlias.
2118 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2119 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2121 // If LHS and RHS are related via constant offsets to the same base
2122 // value, we can replace it with an icmp which just compares the offsets.
2124 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2126 // Various optimizations for (in)equality comparisons.
2127 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2128 // Different non-empty allocations that exist at the same time have
2129 // different addresses (if the program can tell). Global variables always
2130 // exist, so they always exist during the lifetime of each other and all
2131 // allocas. Two different allocas usually have different addresses...
2133 // However, if there's an @llvm.stackrestore dynamically in between two
2134 // allocas, they may have the same address. It's tempting to reduce the
2135 // scope of the problem by only looking at *static* allocas here. That would
2136 // cover the majority of allocas while significantly reducing the likelihood
2137 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2138 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2139 // an entry block. Also, if we have a block that's not attached to a
2140 // function, we can't tell if it's "static" under the current definition.
2141 // Theoretically, this problem could be fixed by creating a new kind of
2142 // instruction kind specifically for static allocas. Such a new instruction
2143 // could be required to be at the top of the entry block, thus preventing it
2144 // from being subject to a @llvm.stackrestore. Instcombine could even
2145 // convert regular allocas into these special allocas. It'd be nifty.
2146 // However, until then, this problem remains open.
2148 // So, we'll assume that two non-empty allocas have different addresses
2151 // With all that, if the offsets are within the bounds of their allocations
2152 // (and not one-past-the-end! so we can't use inbounds!), and their
2153 // allocations aren't the same, the pointers are not equal.
2155 // Note that it's not necessary to check for LHS being a global variable
2156 // address, due to canonicalization and constant folding.
2157 if (isa<AllocaInst>(LHS) &&
2158 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2159 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2160 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2161 uint64_t LHSSize, RHSSize;
2162 if (LHSOffsetCI && RHSOffsetCI &&
2163 getObjectSize(LHS, LHSSize, DL, TLI) &&
2164 getObjectSize(RHS, RHSSize, DL, TLI)) {
2165 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2166 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2167 if (!LHSOffsetValue.isNegative() &&
2168 !RHSOffsetValue.isNegative() &&
2169 LHSOffsetValue.ult(LHSSize) &&
2170 RHSOffsetValue.ult(RHSSize)) {
2171 return ConstantInt::get(GetCompareTy(LHS),
2172 !CmpInst::isTrueWhenEqual(Pred));
2176 // Repeat the above check but this time without depending on DataLayout
2177 // or being able to compute a precise size.
2178 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2179 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2180 LHSOffset->isNullValue() &&
2181 RHSOffset->isNullValue())
2182 return ConstantInt::get(GetCompareTy(LHS),
2183 !CmpInst::isTrueWhenEqual(Pred));
2186 // Even if an non-inbounds GEP occurs along the path we can still optimize
2187 // equality comparisons concerning the result. We avoid walking the whole
2188 // chain again by starting where the last calls to
2189 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2190 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2191 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2193 return ConstantExpr::getICmp(Pred,
2194 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2195 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2197 // If one side of the equality comparison must come from a noalias call
2198 // (meaning a system memory allocation function), and the other side must
2199 // come from a pointer that cannot overlap with dynamically-allocated
2200 // memory within the lifetime of the current function (allocas, byval
2201 // arguments, globals), then determine the comparison result here.
2202 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2203 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2204 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2206 // Is the set of underlying objects all noalias calls?
2207 auto IsNAC = [](ArrayRef<Value *> Objects) {
2208 return all_of(Objects, isNoAliasCall);
2211 // Is the set of underlying objects all things which must be disjoint from
2212 // noalias calls. For allocas, we consider only static ones (dynamic
2213 // allocas might be transformed into calls to malloc not simultaneously
2214 // live with the compared-to allocation). For globals, we exclude symbols
2215 // that might be resolve lazily to symbols in another dynamically-loaded
2216 // library (and, thus, could be malloc'ed by the implementation).
2217 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2218 return all_of(Objects, [](Value *V) {
2219 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2220 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2221 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2222 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2223 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2224 !GV->isThreadLocal();
2225 if (const Argument *A = dyn_cast<Argument>(V))
2226 return A->hasByValAttr();
2231 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2232 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2233 return ConstantInt::get(GetCompareTy(LHS),
2234 !CmpInst::isTrueWhenEqual(Pred));
2236 // Fold comparisons for non-escaping pointer even if the allocation call
2237 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2238 // dynamic allocation call could be either of the operands.
2239 Value *MI = nullptr;
2240 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT))
2242 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT))
2244 // FIXME: We should also fold the compare when the pointer escapes, but the
2245 // compare dominates the pointer escape
2246 if (MI && !PointerMayBeCaptured(MI, true, true))
2247 return ConstantInt::get(GetCompareTy(LHS),
2248 CmpInst::isFalseWhenEqual(Pred));
2255 /// Fold an icmp when its operands have i1 scalar type.
2256 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2257 Value *RHS, const Query &Q) {
2258 Type *ITy = GetCompareTy(LHS); // The return type.
2259 Type *OpTy = LHS->getType(); // The operand type.
2260 if (!OpTy->getScalarType()->isIntegerTy(1))
2266 case ICmpInst::ICMP_EQ:
2268 if (match(RHS, m_One()))
2271 case ICmpInst::ICMP_NE:
2273 if (match(RHS, m_Zero()))
2276 case ICmpInst::ICMP_UGT:
2278 if (match(RHS, m_Zero()))
2281 case ICmpInst::ICMP_UGE:
2283 if (match(RHS, m_One()))
2285 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2286 return getTrue(ITy);
2288 case ICmpInst::ICMP_SGE:
2289 /// For signed comparison, the values for an i1 are 0 and -1
2290 /// respectively. This maps into a truth table of:
2291 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2292 /// 0 | 0 | 1 (0 >= 0) | 1
2293 /// 0 | 1 | 1 (0 >= -1) | 1
2294 /// 1 | 0 | 0 (-1 >= 0) | 0
2295 /// 1 | 1 | 1 (-1 >= -1) | 1
2296 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2297 return getTrue(ITy);
2299 case ICmpInst::ICMP_SLT:
2301 if (match(RHS, m_Zero()))
2304 case ICmpInst::ICMP_SLE:
2306 if (match(RHS, m_One()))
2309 case ICmpInst::ICMP_ULE:
2310 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2311 return getTrue(ITy);
2318 /// Try hard to fold icmp with zero RHS because this is a common case.
2319 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2320 Value *RHS, const Query &Q) {
2321 if (!match(RHS, m_Zero()))
2324 Type *ITy = GetCompareTy(LHS); // The return type.
2325 bool LHSKnownNonNegative, LHSKnownNegative;
2328 llvm_unreachable("Unknown ICmp predicate!");
2329 case ICmpInst::ICMP_ULT:
2330 return getFalse(ITy);
2331 case ICmpInst::ICMP_UGE:
2332 return getTrue(ITy);
2333 case ICmpInst::ICMP_EQ:
2334 case ICmpInst::ICMP_ULE:
2335 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2336 return getFalse(ITy);
2338 case ICmpInst::ICMP_NE:
2339 case ICmpInst::ICMP_UGT:
2340 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2341 return getTrue(ITy);
2343 case ICmpInst::ICMP_SLT:
2344 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2346 if (LHSKnownNegative)
2347 return getTrue(ITy);
2348 if (LHSKnownNonNegative)
2349 return getFalse(ITy);
2351 case ICmpInst::ICMP_SLE:
2352 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2354 if (LHSKnownNegative)
2355 return getTrue(ITy);
2356 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2357 return getFalse(ITy);
2359 case ICmpInst::ICMP_SGE:
2360 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2362 if (LHSKnownNegative)
2363 return getFalse(ITy);
2364 if (LHSKnownNonNegative)
2365 return getTrue(ITy);
2367 case ICmpInst::ICMP_SGT:
2368 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2370 if (LHSKnownNegative)
2371 return getFalse(ITy);
2372 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2373 return getTrue(ITy);
2380 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2383 if (!match(RHS, m_APInt(C)))
2386 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2387 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2388 if (RHS_CR.isEmptySet())
2389 return ConstantInt::getFalse(GetCompareTy(RHS));
2390 if (RHS_CR.isFullSet())
2391 return ConstantInt::getTrue(GetCompareTy(RHS));
2393 // Many binary operators with constant RHS have easy to compute constant
2394 // range. Use them to check whether the comparison is a tautology.
2395 unsigned Width = C->getBitWidth();
2396 APInt Lower = APInt(Width, 0);
2397 APInt Upper = APInt(Width, 0);
2399 if (match(LHS, m_URem(m_Value(), m_APInt(C2)))) {
2400 // 'urem x, C2' produces [0, C2).
2402 } else if (match(LHS, m_SRem(m_Value(), m_APInt(C2)))) {
2403 // 'srem x, C2' produces (-|C2|, |C2|).
2405 Lower = (-Upper) + 1;
2406 } else if (match(LHS, m_UDiv(m_APInt(C2), m_Value()))) {
2407 // 'udiv C2, x' produces [0, C2].
2409 } else if (match(LHS, m_UDiv(m_Value(), m_APInt(C2)))) {
2410 // 'udiv x, C2' produces [0, UINT_MAX / C2].
2411 APInt NegOne = APInt::getAllOnesValue(Width);
2413 Upper = NegOne.udiv(*C2) + 1;
2414 } else if (match(LHS, m_SDiv(m_APInt(C2), m_Value()))) {
2415 if (C2->isMinSignedValue()) {
2416 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2418 Upper = Lower.lshr(1) + 1;
2420 // 'sdiv C2, x' produces [-|C2|, |C2|].
2421 Upper = C2->abs() + 1;
2422 Lower = (-Upper) + 1;
2424 } else if (match(LHS, m_SDiv(m_Value(), m_APInt(C2)))) {
2425 APInt IntMin = APInt::getSignedMinValue(Width);
2426 APInt IntMax = APInt::getSignedMaxValue(Width);
2427 if (C2->isAllOnesValue()) {
2428 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2429 // where C2 != -1 and C2 != 0 and C2 != 1
2432 } else if (C2->countLeadingZeros() < Width - 1) {
2433 // 'sdiv x, C2' produces [INT_MIN / C2, INT_MAX / C2]
2434 // where C2 != -1 and C2 != 0 and C2 != 1
2435 Lower = IntMin.sdiv(*C2);
2436 Upper = IntMax.sdiv(*C2);
2437 if (Lower.sgt(Upper))
2438 std::swap(Lower, Upper);
2440 assert(Upper != Lower && "Upper part of range has wrapped!");
2442 } else if (match(LHS, m_NUWShl(m_APInt(C2), m_Value()))) {
2443 // 'shl nuw C2, x' produces [C2, C2 << CLZ(C2)]
2445 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2446 } else if (match(LHS, m_NSWShl(m_APInt(C2), m_Value()))) {
2447 if (C2->isNegative()) {
2448 // 'shl nsw C2, x' produces [C2 << CLO(C2)-1, C2]
2449 unsigned ShiftAmount = C2->countLeadingOnes() - 1;
2450 Lower = C2->shl(ShiftAmount);
2453 // 'shl nsw C2, x' produces [C2, C2 << CLZ(C2)-1]
2454 unsigned ShiftAmount = C2->countLeadingZeros() - 1;
2456 Upper = C2->shl(ShiftAmount) + 1;
2458 } else if (match(LHS, m_LShr(m_Value(), m_APInt(C2)))) {
2459 // 'lshr x, C2' produces [0, UINT_MAX >> C2].
2460 APInt NegOne = APInt::getAllOnesValue(Width);
2462 Upper = NegOne.lshr(*C2) + 1;
2463 } else if (match(LHS, m_LShr(m_APInt(C2), m_Value()))) {
2464 // 'lshr C2, x' produces [C2 >> (Width-1), C2].
2465 unsigned ShiftAmount = Width - 1;
2466 if (*C2 != 0 && cast<BinaryOperator>(LHS)->isExact())
2467 ShiftAmount = C2->countTrailingZeros();
2468 Lower = C2->lshr(ShiftAmount);
2470 } else if (match(LHS, m_AShr(m_Value(), m_APInt(C2)))) {
2471 // 'ashr x, C2' produces [INT_MIN >> C2, INT_MAX >> C2].
2472 APInt IntMin = APInt::getSignedMinValue(Width);
2473 APInt IntMax = APInt::getSignedMaxValue(Width);
2474 if (C2->ult(Width)) {
2475 Lower = IntMin.ashr(*C2);
2476 Upper = IntMax.ashr(*C2) + 1;
2478 } else if (match(LHS, m_AShr(m_APInt(C2), m_Value()))) {
2479 unsigned ShiftAmount = Width - 1;
2480 if (*C2 != 0 && cast<BinaryOperator>(LHS)->isExact())
2481 ShiftAmount = C2->countTrailingZeros();
2482 if (C2->isNegative()) {
2483 // 'ashr C2, x' produces [C2, C2 >> (Width-1)]
2485 Upper = C2->ashr(ShiftAmount) + 1;
2487 // 'ashr C2, x' produces [C2 >> (Width-1), C2]
2488 Lower = C2->ashr(ShiftAmount);
2491 } else if (match(LHS, m_Or(m_Value(), m_APInt(C2)))) {
2492 // 'or x, C2' produces [C2, UINT_MAX].
2494 } else if (match(LHS, m_And(m_Value(), m_APInt(C2)))) {
2495 // 'and x, C2' produces [0, C2].
2497 } else if (match(LHS, m_NUWAdd(m_Value(), m_APInt(C2)))) {
2498 // 'add nuw x, C2' produces [C2, UINT_MAX].
2502 ConstantRange LHS_CR =
2503 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2505 if (auto *I = dyn_cast<Instruction>(LHS))
2506 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2507 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2509 if (!LHS_CR.isFullSet()) {
2510 if (RHS_CR.contains(LHS_CR))
2511 return ConstantInt::getTrue(GetCompareTy(RHS));
2512 if (RHS_CR.inverse().contains(LHS_CR))
2513 return ConstantInt::getFalse(GetCompareTy(RHS));
2519 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2520 Value *RHS, const Query &Q,
2521 unsigned MaxRecurse) {
2522 Type *ITy = GetCompareTy(LHS); // The return type.
2524 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2525 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2526 if (MaxRecurse && (LBO || RBO)) {
2527 // Analyze the case when either LHS or RHS is an add instruction.
2528 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2529 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2530 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2531 if (LBO && LBO->getOpcode() == Instruction::Add) {
2532 A = LBO->getOperand(0);
2533 B = LBO->getOperand(1);
2535 ICmpInst::isEquality(Pred) ||
2536 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2537 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2539 if (RBO && RBO->getOpcode() == Instruction::Add) {
2540 C = RBO->getOperand(0);
2541 D = RBO->getOperand(1);
2543 ICmpInst::isEquality(Pred) ||
2544 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2545 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2548 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2549 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2550 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2551 Constant::getNullValue(RHS->getType()), Q,
2555 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2556 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2558 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2559 C == LHS ? D : C, Q, MaxRecurse - 1))
2562 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2563 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2565 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2568 // C + B == C + D -> B == D
2571 } else if (A == D) {
2572 // D + B == C + D -> B == C
2575 } else if (B == C) {
2576 // A + C == C + D -> A == D
2581 // A + D == C + D -> A == C
2585 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2592 // icmp pred (or X, Y), X
2593 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2594 if (Pred == ICmpInst::ICMP_ULT)
2595 return getFalse(ITy);
2596 if (Pred == ICmpInst::ICMP_UGE)
2597 return getTrue(ITy);
2599 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2600 bool RHSKnownNonNegative, RHSKnownNegative;
2601 bool YKnownNonNegative, YKnownNegative;
2602 ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, Q.DL, 0,
2603 Q.AC, Q.CxtI, Q.DT);
2604 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Q.DL, 0, Q.AC,
2606 if (RHSKnownNonNegative && YKnownNegative)
2607 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2608 if (RHSKnownNegative || YKnownNonNegative)
2609 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2612 // icmp pred X, (or X, Y)
2613 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2614 if (Pred == ICmpInst::ICMP_ULE)
2615 return getTrue(ITy);
2616 if (Pred == ICmpInst::ICMP_UGT)
2617 return getFalse(ITy);
2619 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2620 bool LHSKnownNonNegative, LHSKnownNegative;
2621 bool YKnownNonNegative, YKnownNegative;
2622 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0,
2623 Q.AC, Q.CxtI, Q.DT);
2624 ComputeSignBit(Y, YKnownNonNegative, YKnownNegative, Q.DL, 0, Q.AC,
2626 if (LHSKnownNonNegative && YKnownNegative)
2627 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2628 if (LHSKnownNegative || YKnownNonNegative)
2629 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2634 // icmp pred (and X, Y), X
2635 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)),
2636 m_And(m_Specific(RHS), m_Value())))) {
2637 if (Pred == ICmpInst::ICMP_UGT)
2638 return getFalse(ITy);
2639 if (Pred == ICmpInst::ICMP_ULE)
2640 return getTrue(ITy);
2642 // icmp pred X, (and X, Y)
2643 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)),
2644 m_And(m_Specific(LHS), m_Value())))) {
2645 if (Pred == ICmpInst::ICMP_UGE)
2646 return getTrue(ITy);
2647 if (Pred == ICmpInst::ICMP_ULT)
2648 return getFalse(ITy);
2651 // 0 - (zext X) pred C
2652 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2653 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2654 if (RHSC->getValue().isStrictlyPositive()) {
2655 if (Pred == ICmpInst::ICMP_SLT)
2656 return ConstantInt::getTrue(RHSC->getContext());
2657 if (Pred == ICmpInst::ICMP_SGE)
2658 return ConstantInt::getFalse(RHSC->getContext());
2659 if (Pred == ICmpInst::ICMP_EQ)
2660 return ConstantInt::getFalse(RHSC->getContext());
2661 if (Pred == ICmpInst::ICMP_NE)
2662 return ConstantInt::getTrue(RHSC->getContext());
2664 if (RHSC->getValue().isNonNegative()) {
2665 if (Pred == ICmpInst::ICMP_SLE)
2666 return ConstantInt::getTrue(RHSC->getContext());
2667 if (Pred == ICmpInst::ICMP_SGT)
2668 return ConstantInt::getFalse(RHSC->getContext());
2673 // icmp pred (urem X, Y), Y
2674 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2675 bool KnownNonNegative, KnownNegative;
2679 case ICmpInst::ICMP_SGT:
2680 case ICmpInst::ICMP_SGE:
2681 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2683 if (!KnownNonNegative)
2686 case ICmpInst::ICMP_EQ:
2687 case ICmpInst::ICMP_UGT:
2688 case ICmpInst::ICMP_UGE:
2689 return getFalse(ITy);
2690 case ICmpInst::ICMP_SLT:
2691 case ICmpInst::ICMP_SLE:
2692 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2694 if (!KnownNonNegative)
2697 case ICmpInst::ICMP_NE:
2698 case ICmpInst::ICMP_ULT:
2699 case ICmpInst::ICMP_ULE:
2700 return getTrue(ITy);
2704 // icmp pred X, (urem Y, X)
2705 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2706 bool KnownNonNegative, KnownNegative;
2710 case ICmpInst::ICMP_SGT:
2711 case ICmpInst::ICMP_SGE:
2712 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2714 if (!KnownNonNegative)
2717 case ICmpInst::ICMP_NE:
2718 case ICmpInst::ICMP_UGT:
2719 case ICmpInst::ICMP_UGE:
2720 return getTrue(ITy);
2721 case ICmpInst::ICMP_SLT:
2722 case ICmpInst::ICMP_SLE:
2723 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2725 if (!KnownNonNegative)
2728 case ICmpInst::ICMP_EQ:
2729 case ICmpInst::ICMP_ULT:
2730 case ICmpInst::ICMP_ULE:
2731 return getFalse(ITy);
2737 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2738 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2739 // icmp pred (X op Y), X
2740 if (Pred == ICmpInst::ICMP_UGT)
2741 return getFalse(ITy);
2742 if (Pred == ICmpInst::ICMP_ULE)
2743 return getTrue(ITy);
2748 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2749 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2750 // icmp pred X, (X op Y)
2751 if (Pred == ICmpInst::ICMP_ULT)
2752 return getFalse(ITy);
2753 if (Pred == ICmpInst::ICMP_UGE)
2754 return getTrue(ITy);
2761 // where CI2 is a power of 2 and CI isn't
2762 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2763 const APInt *CI2Val, *CIVal = &CI->getValue();
2764 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2765 CI2Val->isPowerOf2()) {
2766 if (!CIVal->isPowerOf2()) {
2767 // CI2 << X can equal zero in some circumstances,
2768 // this simplification is unsafe if CI is zero.
2770 // We know it is safe if:
2771 // - The shift is nsw, we can't shift out the one bit.
2772 // - The shift is nuw, we can't shift out the one bit.
2775 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2776 *CI2Val == 1 || !CI->isZero()) {
2777 if (Pred == ICmpInst::ICMP_EQ)
2778 return ConstantInt::getFalse(RHS->getContext());
2779 if (Pred == ICmpInst::ICMP_NE)
2780 return ConstantInt::getTrue(RHS->getContext());
2783 if (CIVal->isSignBit() && *CI2Val == 1) {
2784 if (Pred == ICmpInst::ICMP_UGT)
2785 return ConstantInt::getFalse(RHS->getContext());
2786 if (Pred == ICmpInst::ICMP_ULE)
2787 return ConstantInt::getTrue(RHS->getContext());
2792 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2793 LBO->getOperand(1) == RBO->getOperand(1)) {
2794 switch (LBO->getOpcode()) {
2797 case Instruction::UDiv:
2798 case Instruction::LShr:
2799 if (ICmpInst::isSigned(Pred))
2802 case Instruction::SDiv:
2803 case Instruction::AShr:
2804 if (!LBO->isExact() || !RBO->isExact())
2806 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2807 RBO->getOperand(0), Q, MaxRecurse - 1))
2810 case Instruction::Shl: {
2811 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2812 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2815 if (!NSW && ICmpInst::isSigned(Pred))
2817 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2818 RBO->getOperand(0), Q, MaxRecurse - 1))
2827 /// Simplify integer comparisons where at least one operand of the compare
2828 /// matches an integer min/max idiom.
2829 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2830 Value *RHS, const Query &Q,
2831 unsigned MaxRecurse) {
2832 Type *ITy = GetCompareTy(LHS); // The return type.
2834 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2835 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2837 // Signed variants on "max(a,b)>=a -> true".
2838 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2840 std::swap(A, B); // smax(A, B) pred A.
2841 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2842 // We analyze this as smax(A, B) pred A.
2844 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2845 (A == LHS || B == LHS)) {
2847 std::swap(A, B); // A pred smax(A, B).
2848 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2849 // We analyze this as smax(A, B) swapped-pred A.
2850 P = CmpInst::getSwappedPredicate(Pred);
2851 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2852 (A == RHS || B == RHS)) {
2854 std::swap(A, B); // smin(A, B) pred A.
2855 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2856 // We analyze this as smax(-A, -B) swapped-pred -A.
2857 // Note that we do not need to actually form -A or -B thanks to EqP.
2858 P = CmpInst::getSwappedPredicate(Pred);
2859 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2860 (A == LHS || B == LHS)) {
2862 std::swap(A, B); // A pred smin(A, B).
2863 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2864 // We analyze this as smax(-A, -B) pred -A.
2865 // Note that we do not need to actually form -A or -B thanks to EqP.
2868 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2869 // Cases correspond to "max(A, B) p A".
2873 case CmpInst::ICMP_EQ:
2874 case CmpInst::ICMP_SLE:
2875 // Equivalent to "A EqP B". This may be the same as the condition tested
2876 // in the max/min; if so, we can just return that.
2877 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2879 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2881 // Otherwise, see if "A EqP B" simplifies.
2883 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2886 case CmpInst::ICMP_NE:
2887 case CmpInst::ICMP_SGT: {
2888 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2889 // Equivalent to "A InvEqP B". This may be the same as the condition
2890 // tested in the max/min; if so, we can just return that.
2891 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2893 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2895 // Otherwise, see if "A InvEqP B" simplifies.
2897 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2901 case CmpInst::ICMP_SGE:
2903 return getTrue(ITy);
2904 case CmpInst::ICMP_SLT:
2906 return getFalse(ITy);
2910 // Unsigned variants on "max(a,b)>=a -> true".
2911 P = CmpInst::BAD_ICMP_PREDICATE;
2912 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2914 std::swap(A, B); // umax(A, B) pred A.
2915 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2916 // We analyze this as umax(A, B) pred A.
2918 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2919 (A == LHS || B == LHS)) {
2921 std::swap(A, B); // A pred umax(A, B).
2922 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2923 // We analyze this as umax(A, B) swapped-pred A.
2924 P = CmpInst::getSwappedPredicate(Pred);
2925 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2926 (A == RHS || B == RHS)) {
2928 std::swap(A, B); // umin(A, B) pred A.
2929 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2930 // We analyze this as umax(-A, -B) swapped-pred -A.
2931 // Note that we do not need to actually form -A or -B thanks to EqP.
2932 P = CmpInst::getSwappedPredicate(Pred);
2933 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2934 (A == LHS || B == LHS)) {
2936 std::swap(A, B); // A pred umin(A, B).
2937 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2938 // We analyze this as umax(-A, -B) pred -A.
2939 // Note that we do not need to actually form -A or -B thanks to EqP.
2942 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2943 // Cases correspond to "max(A, B) p A".
2947 case CmpInst::ICMP_EQ:
2948 case CmpInst::ICMP_ULE:
2949 // Equivalent to "A EqP B". This may be the same as the condition tested
2950 // in the max/min; if so, we can just return that.
2951 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2953 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2955 // Otherwise, see if "A EqP B" simplifies.
2957 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2960 case CmpInst::ICMP_NE:
2961 case CmpInst::ICMP_UGT: {
2962 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2963 // Equivalent to "A InvEqP B". This may be the same as the condition
2964 // tested in the max/min; if so, we can just return that.
2965 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2967 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2969 // Otherwise, see if "A InvEqP B" simplifies.
2971 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2975 case CmpInst::ICMP_UGE:
2977 return getTrue(ITy);
2978 case CmpInst::ICMP_ULT:
2980 return getFalse(ITy);
2984 // Variants on "max(x,y) >= min(x,z)".
2986 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2987 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2988 (A == C || A == D || B == C || B == D)) {
2989 // max(x, ?) pred min(x, ?).
2990 if (Pred == CmpInst::ICMP_SGE)
2992 return getTrue(ITy);
2993 if (Pred == CmpInst::ICMP_SLT)
2995 return getFalse(ITy);
2996 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2997 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2998 (A == C || A == D || B == C || B == D)) {
2999 // min(x, ?) pred max(x, ?).
3000 if (Pred == CmpInst::ICMP_SLE)
3002 return getTrue(ITy);
3003 if (Pred == CmpInst::ICMP_SGT)
3005 return getFalse(ITy);
3006 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3007 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3008 (A == C || A == D || B == C || B == D)) {
3009 // max(x, ?) pred min(x, ?).
3010 if (Pred == CmpInst::ICMP_UGE)
3012 return getTrue(ITy);
3013 if (Pred == CmpInst::ICMP_ULT)
3015 return getFalse(ITy);
3016 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3017 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3018 (A == C || A == D || B == C || B == D)) {
3019 // min(x, ?) pred max(x, ?).
3020 if (Pred == CmpInst::ICMP_ULE)
3022 return getTrue(ITy);
3023 if (Pred == CmpInst::ICMP_UGT)
3025 return getFalse(ITy);
3031 /// Given operands for an ICmpInst, see if we can fold the result.
3032 /// If not, this returns null.
3033 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3034 const Query &Q, unsigned MaxRecurse) {
3035 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3036 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3038 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3039 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3040 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3042 // If we have a constant, make sure it is on the RHS.
3043 std::swap(LHS, RHS);
3044 Pred = CmpInst::getSwappedPredicate(Pred);
3047 Type *ITy = GetCompareTy(LHS); // The return type.
3049 // icmp X, X -> true/false
3050 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3051 // because X could be 0.
3052 if (LHS == RHS || isa<UndefValue>(RHS))
3053 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3055 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3058 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3061 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3064 // If both operands have range metadata, use the metadata
3065 // to simplify the comparison.
3066 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3067 auto RHS_Instr = dyn_cast<Instruction>(RHS);
3068 auto LHS_Instr = dyn_cast<Instruction>(LHS);
3070 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3071 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3072 auto RHS_CR = getConstantRangeFromMetadata(
3073 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3074 auto LHS_CR = getConstantRangeFromMetadata(
3075 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3077 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3078 if (Satisfied_CR.contains(LHS_CR))
3079 return ConstantInt::getTrue(RHS->getContext());
3081 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3082 CmpInst::getInversePredicate(Pred), RHS_CR);
3083 if (InversedSatisfied_CR.contains(LHS_CR))
3084 return ConstantInt::getFalse(RHS->getContext());
3088 // Compare of cast, for example (zext X) != 0 -> X != 0
3089 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3090 Instruction *LI = cast<CastInst>(LHS);
3091 Value *SrcOp = LI->getOperand(0);
3092 Type *SrcTy = SrcOp->getType();
3093 Type *DstTy = LI->getType();
3095 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3096 // if the integer type is the same size as the pointer type.
3097 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3098 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3099 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3100 // Transfer the cast to the constant.
3101 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3102 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3105 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3106 if (RI->getOperand(0)->getType() == SrcTy)
3107 // Compare without the cast.
3108 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3114 if (isa<ZExtInst>(LHS)) {
3115 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3117 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3118 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3119 // Compare X and Y. Note that signed predicates become unsigned.
3120 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3121 SrcOp, RI->getOperand(0), Q,
3125 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3126 // too. If not, then try to deduce the result of the comparison.
3127 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3128 // Compute the constant that would happen if we truncated to SrcTy then
3129 // reextended to DstTy.
3130 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3131 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3133 // If the re-extended constant didn't change then this is effectively
3134 // also a case of comparing two zero-extended values.
3135 if (RExt == CI && MaxRecurse)
3136 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3137 SrcOp, Trunc, Q, MaxRecurse-1))
3140 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3141 // there. Use this to work out the result of the comparison.
3144 default: llvm_unreachable("Unknown ICmp predicate!");
3146 case ICmpInst::ICMP_EQ:
3147 case ICmpInst::ICMP_UGT:
3148 case ICmpInst::ICMP_UGE:
3149 return ConstantInt::getFalse(CI->getContext());
3151 case ICmpInst::ICMP_NE:
3152 case ICmpInst::ICMP_ULT:
3153 case ICmpInst::ICMP_ULE:
3154 return ConstantInt::getTrue(CI->getContext());
3156 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3157 // is non-negative then LHS <s RHS.
3158 case ICmpInst::ICMP_SGT:
3159 case ICmpInst::ICMP_SGE:
3160 return CI->getValue().isNegative() ?
3161 ConstantInt::getTrue(CI->getContext()) :
3162 ConstantInt::getFalse(CI->getContext());
3164 case ICmpInst::ICMP_SLT:
3165 case ICmpInst::ICMP_SLE:
3166 return CI->getValue().isNegative() ?
3167 ConstantInt::getFalse(CI->getContext()) :
3168 ConstantInt::getTrue(CI->getContext());
3174 if (isa<SExtInst>(LHS)) {
3175 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3177 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3178 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3179 // Compare X and Y. Note that the predicate does not change.
3180 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3184 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3185 // too. If not, then try to deduce the result of the comparison.
3186 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3187 // Compute the constant that would happen if we truncated to SrcTy then
3188 // reextended to DstTy.
3189 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3190 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3192 // If the re-extended constant didn't change then this is effectively
3193 // also a case of comparing two sign-extended values.
3194 if (RExt == CI && MaxRecurse)
3195 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3198 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3199 // bits there. Use this to work out the result of the comparison.
3202 default: llvm_unreachable("Unknown ICmp predicate!");
3203 case ICmpInst::ICMP_EQ:
3204 return ConstantInt::getFalse(CI->getContext());
3205 case ICmpInst::ICMP_NE:
3206 return ConstantInt::getTrue(CI->getContext());
3208 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3210 case ICmpInst::ICMP_SGT:
3211 case ICmpInst::ICMP_SGE:
3212 return CI->getValue().isNegative() ?
3213 ConstantInt::getTrue(CI->getContext()) :
3214 ConstantInt::getFalse(CI->getContext());
3215 case ICmpInst::ICMP_SLT:
3216 case ICmpInst::ICMP_SLE:
3217 return CI->getValue().isNegative() ?
3218 ConstantInt::getFalse(CI->getContext()) :
3219 ConstantInt::getTrue(CI->getContext());
3221 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3223 case ICmpInst::ICMP_UGT:
3224 case ICmpInst::ICMP_UGE:
3225 // Comparison is true iff the LHS <s 0.
3227 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3228 Constant::getNullValue(SrcTy),
3232 case ICmpInst::ICMP_ULT:
3233 case ICmpInst::ICMP_ULE:
3234 // Comparison is true iff the LHS >=s 0.
3236 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3237 Constant::getNullValue(SrcTy),
3247 // icmp eq|ne X, Y -> false|true if X != Y
3248 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
3249 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3250 LLVMContext &Ctx = LHS->getType()->getContext();
3251 return Pred == ICmpInst::ICMP_NE ?
3252 ConstantInt::getTrue(Ctx) : ConstantInt::getFalse(Ctx);
3255 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3258 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3261 // Simplify comparisons of related pointers using a powerful, recursive
3262 // GEP-walk when we have target data available..
3263 if (LHS->getType()->isPointerTy())
3264 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS))
3266 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3267 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3268 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3269 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3270 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3271 Q.DL.getTypeSizeInBits(CRHS->getType()))
3272 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI,
3273 CLHS->getPointerOperand(),
3274 CRHS->getPointerOperand()))
3277 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3278 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3279 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3280 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3281 (ICmpInst::isEquality(Pred) ||
3282 (GLHS->isInBounds() && GRHS->isInBounds() &&
3283 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3284 // The bases are equal and the indices are constant. Build a constant
3285 // expression GEP with the same indices and a null base pointer to see
3286 // what constant folding can make out of it.
3287 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3288 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3289 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3290 GLHS->getSourceElementType(), Null, IndicesLHS);
3292 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3293 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3294 GLHS->getSourceElementType(), Null, IndicesRHS);
3295 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3300 // If a bit is known to be zero for A and known to be one for B,
3301 // then A and B cannot be equal.
3302 if (ICmpInst::isEquality(Pred)) {
3303 const APInt *RHSVal;
3304 if (match(RHS, m_APInt(RHSVal))) {
3305 unsigned BitWidth = RHSVal->getBitWidth();
3306 APInt LHSKnownZero(BitWidth, 0);
3307 APInt LHSKnownOne(BitWidth, 0);
3308 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC,
3310 if (((LHSKnownZero & *RHSVal) != 0) || ((LHSKnownOne & ~(*RHSVal)) != 0))
3311 return Pred == ICmpInst::ICMP_EQ ? ConstantInt::getFalse(ITy)
3312 : ConstantInt::getTrue(ITy);
3316 // If the comparison is with the result of a select instruction, check whether
3317 // comparing with either branch of the select always yields the same value.
3318 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3319 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3322 // If the comparison is with the result of a phi instruction, check whether
3323 // doing the compare with each incoming phi value yields a common result.
3324 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3325 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3331 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3332 const DataLayout &DL,
3333 const TargetLibraryInfo *TLI,
3334 const DominatorTree *DT, AssumptionCache *AC,
3335 const Instruction *CxtI) {
3336 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3340 /// Given operands for an FCmpInst, see if we can fold the result.
3341 /// If not, this returns null.
3342 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3343 FastMathFlags FMF, const Query &Q,
3344 unsigned MaxRecurse) {
3345 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3346 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3348 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3349 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3350 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3352 // If we have a constant, make sure it is on the RHS.
3353 std::swap(LHS, RHS);
3354 Pred = CmpInst::getSwappedPredicate(Pred);
3357 // Fold trivial predicates.
3358 Type *RetTy = GetCompareTy(LHS);
3359 if (Pred == FCmpInst::FCMP_FALSE)
3360 return getFalse(RetTy);
3361 if (Pred == FCmpInst::FCMP_TRUE)
3362 return getTrue(RetTy);
3364 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3366 if (Pred == FCmpInst::FCMP_UNO)
3367 return getFalse(RetTy);
3368 if (Pred == FCmpInst::FCMP_ORD)
3369 return getTrue(RetTy);
3372 // fcmp pred x, undef and fcmp pred undef, x
3373 // fold to true if unordered, false if ordered
3374 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3375 // Choosing NaN for the undef will always make unordered comparison succeed
3376 // and ordered comparison fail.
3377 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3380 // fcmp x,x -> true/false. Not all compares are foldable.
3382 if (CmpInst::isTrueWhenEqual(Pred))
3383 return getTrue(RetTy);
3384 if (CmpInst::isFalseWhenEqual(Pred))
3385 return getFalse(RetTy);
3388 // Handle fcmp with constant RHS
3389 const ConstantFP *CFP = nullptr;
3390 if (const auto *RHSC = dyn_cast<Constant>(RHS)) {
3391 if (RHS->getType()->isVectorTy())
3392 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue());
3394 CFP = dyn_cast<ConstantFP>(RHSC);
3397 // If the constant is a nan, see if we can fold the comparison based on it.
3398 if (CFP->getValueAPF().isNaN()) {
3399 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3400 return getFalse(RetTy);
3401 assert(FCmpInst::isUnordered(Pred) &&
3402 "Comparison must be either ordered or unordered!");
3403 // True if unordered.
3404 return getTrue(RetTy);
3406 // Check whether the constant is an infinity.
3407 if (CFP->getValueAPF().isInfinity()) {
3408 if (CFP->getValueAPF().isNegative()) {
3410 case FCmpInst::FCMP_OLT:
3411 // No value is ordered and less than negative infinity.
3412 return getFalse(RetTy);
3413 case FCmpInst::FCMP_UGE:
3414 // All values are unordered with or at least negative infinity.
3415 return getTrue(RetTy);
3421 case FCmpInst::FCMP_OGT:
3422 // No value is ordered and greater than infinity.
3423 return getFalse(RetTy);
3424 case FCmpInst::FCMP_ULE:
3425 // All values are unordered with and at most infinity.
3426 return getTrue(RetTy);
3432 if (CFP->getValueAPF().isZero()) {
3434 case FCmpInst::FCMP_UGE:
3435 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3436 return getTrue(RetTy);
3438 case FCmpInst::FCMP_OLT:
3440 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3441 return getFalse(RetTy);
3449 // If the comparison is with the result of a select instruction, check whether
3450 // comparing with either branch of the select always yields the same value.
3451 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3452 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3455 // If the comparison is with the result of a phi instruction, check whether
3456 // doing the compare with each incoming phi value yields a common result.
3457 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3458 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3464 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3465 FastMathFlags FMF, const DataLayout &DL,
3466 const TargetLibraryInfo *TLI,
3467 const DominatorTree *DT, AssumptionCache *AC,
3468 const Instruction *CxtI) {
3469 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF,
3470 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3473 /// See if V simplifies when its operand Op is replaced with RepOp.
3474 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3476 unsigned MaxRecurse) {
3477 // Trivial replacement.
3481 auto *I = dyn_cast<Instruction>(V);
3485 // If this is a binary operator, try to simplify it with the replaced op.
3486 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3488 // %cmp = icmp eq i32 %x, 2147483647
3489 // %add = add nsw i32 %x, 1
3490 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3492 // We can't replace %sel with %add unless we strip away the flags.
3493 if (isa<OverflowingBinaryOperator>(B))
3494 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3496 if (isa<PossiblyExactOperator>(B))
3501 if (B->getOperand(0) == Op)
3502 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3504 if (B->getOperand(1) == Op)
3505 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3510 // Same for CmpInsts.
3511 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3513 if (C->getOperand(0) == Op)
3514 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3516 if (C->getOperand(1) == Op)
3517 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3522 // TODO: We could hand off more cases to instsimplify here.
3524 // If all operands are constant after substituting Op for RepOp then we can
3525 // constant fold the instruction.
3526 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3527 // Build a list of all constant operands.
3528 SmallVector<Constant *, 8> ConstOps;
3529 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3530 if (I->getOperand(i) == Op)
3531 ConstOps.push_back(CRepOp);
3532 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3533 ConstOps.push_back(COp);
3538 // All operands were constants, fold it.
3539 if (ConstOps.size() == I->getNumOperands()) {
3540 if (CmpInst *C = dyn_cast<CmpInst>(I))
3541 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3542 ConstOps[1], Q.DL, Q.TLI);
3544 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3545 if (!LI->isVolatile())
3546 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3548 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3555 /// Try to simplify a select instruction when its condition operand is an
3556 /// integer comparison where one operand of the compare is a constant.
3557 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3558 const APInt *Y, bool TrueWhenUnset) {
3561 // (X & Y) == 0 ? X & ~Y : X --> X
3562 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3563 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3565 return TrueWhenUnset ? FalseVal : TrueVal;
3567 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3568 // (X & Y) != 0 ? X : X & ~Y --> X
3569 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3571 return TrueWhenUnset ? FalseVal : TrueVal;
3573 if (Y->isPowerOf2()) {
3574 // (X & Y) == 0 ? X | Y : X --> X | Y
3575 // (X & Y) != 0 ? X | Y : X --> X
3576 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3578 return TrueWhenUnset ? TrueVal : FalseVal;
3580 // (X & Y) == 0 ? X : X | Y --> X
3581 // (X & Y) != 0 ? X : X | Y --> X | Y
3582 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3584 return TrueWhenUnset ? TrueVal : FalseVal;
3590 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3592 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal,
3594 bool TrueWhenUnset) {
3595 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
3599 APInt MinSignedValue;
3601 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) {
3602 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0
3603 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0
3604 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits();
3605 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth);
3607 // icmp slt X, 0 <--> icmp ne (and X, C), 0
3608 // icmp sgt X, -1 <--> icmp eq (and X, C), 0
3610 MinSignedValue = APInt::getSignedMinValue(BitWidth);
3613 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue,
3620 /// Try to simplify a select instruction when its condition operand is an
3621 /// integer comparison.
3622 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3623 Value *FalseVal, const Query &Q,
3624 unsigned MaxRecurse) {
3625 ICmpInst::Predicate Pred;
3626 Value *CmpLHS, *CmpRHS;
3627 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3630 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring
3631 // decomposeBitTestICmp() might help.
3632 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3635 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3636 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3637 Pred == ICmpInst::ICMP_EQ))
3639 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3640 // Comparing signed-less-than 0 checks if the sign bit is set.
3641 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3644 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3645 // Comparing signed-greater-than -1 checks if the sign bit is not set.
3646 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3651 if (CondVal->hasOneUse()) {
3653 if (match(CmpRHS, m_APInt(C))) {
3654 // X < MIN ? T : F --> F
3655 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3657 // X < MIN ? T : F --> F
3658 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3660 // X > MAX ? T : F --> F
3661 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3663 // X > MAX ? T : F --> F
3664 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3669 // If we have an equality comparison, then we know the value in one of the
3670 // arms of the select. See if substituting this value into the arm and
3671 // simplifying the result yields the same value as the other arm.
3672 if (Pred == ICmpInst::ICMP_EQ) {
3673 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3675 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3678 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3680 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3683 } else if (Pred == ICmpInst::ICMP_NE) {
3684 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3686 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3689 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3691 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3699 /// Given operands for a SelectInst, see if we can fold the result.
3700 /// If not, this returns null.
3701 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3702 Value *FalseVal, const Query &Q,
3703 unsigned MaxRecurse) {
3704 // select true, X, Y -> X
3705 // select false, X, Y -> Y
3706 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3707 if (CB->isAllOnesValue())
3709 if (CB->isNullValue())
3713 // select C, X, X -> X
3714 if (TrueVal == FalseVal)
3717 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3718 if (isa<Constant>(TrueVal))
3722 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3724 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3728 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3734 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3735 const DataLayout &DL,
3736 const TargetLibraryInfo *TLI,
3737 const DominatorTree *DT, AssumptionCache *AC,
3738 const Instruction *CxtI) {
3739 return ::SimplifySelectInst(Cond, TrueVal, FalseVal,
3740 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3743 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3744 /// If not, this returns null.
3745 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3746 const Query &Q, unsigned) {
3747 // The type of the GEP pointer operand.
3749 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3751 // getelementptr P -> P.
3752 if (Ops.size() == 1)
3755 // Compute the (pointer) type returned by the GEP instruction.
3756 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3757 Type *GEPTy = PointerType::get(LastType, AS);
3758 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3759 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3761 if (isa<UndefValue>(Ops[0]))
3762 return UndefValue::get(GEPTy);
3764 if (Ops.size() == 2) {
3765 // getelementptr P, 0 -> P.
3766 if (match(Ops[1], m_Zero()))
3770 if (Ty->isSized()) {
3773 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3774 // getelementptr P, N -> P if P points to a type of zero size.
3775 if (TyAllocSize == 0)
3778 // The following transforms are only safe if the ptrtoint cast
3779 // doesn't truncate the pointers.
3780 if (Ops[1]->getType()->getScalarSizeInBits() ==
3781 Q.DL.getPointerSizeInBits(AS)) {
3782 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3783 if (match(P, m_Zero()))
3784 return Constant::getNullValue(GEPTy);
3786 if (match(P, m_PtrToInt(m_Value(Temp))))
3787 if (Temp->getType() == GEPTy)
3792 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3793 if (TyAllocSize == 1 &&
3794 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3795 if (Value *R = PtrToIntOrZero(P))
3798 // getelementptr V, (ashr (sub P, V), C) -> Q
3799 // if P points to a type of size 1 << C.
3801 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3802 m_ConstantInt(C))) &&
3803 TyAllocSize == 1ULL << C)
3804 if (Value *R = PtrToIntOrZero(P))
3807 // getelementptr V, (sdiv (sub P, V), C) -> Q
3808 // if P points to a type of size C.
3810 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3811 m_SpecificInt(TyAllocSize))))
3812 if (Value *R = PtrToIntOrZero(P))
3818 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3819 all_of(Ops.slice(1).drop_back(1),
3820 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3822 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3823 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3824 APInt BasePtrOffset(PtrWidth, 0);
3825 Value *StrippedBasePtr =
3826 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3829 // gep (gep V, C), (sub 0, V) -> C
3830 if (match(Ops.back(),
3831 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3832 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3833 return ConstantExpr::getIntToPtr(CI, GEPTy);
3835 // gep (gep V, C), (xor V, -1) -> C-1
3836 if (match(Ops.back(),
3837 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3838 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3839 return ConstantExpr::getIntToPtr(CI, GEPTy);
3844 // Check to see if this is constant foldable.
3845 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
3846 if (!isa<Constant>(Ops[i]))
3849 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3853 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3854 const DataLayout &DL,
3855 const TargetLibraryInfo *TLI,
3856 const DominatorTree *DT, AssumptionCache *AC,
3857 const Instruction *CxtI) {
3858 return ::SimplifyGEPInst(SrcTy, Ops,
3859 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3862 /// Given operands for an InsertValueInst, see if we can fold the result.
3863 /// If not, this returns null.
3864 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3865 ArrayRef<unsigned> Idxs, const Query &Q,
3867 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3868 if (Constant *CVal = dyn_cast<Constant>(Val))
3869 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3871 // insertvalue x, undef, n -> x
3872 if (match(Val, m_Undef()))
3875 // insertvalue x, (extractvalue y, n), n
3876 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3877 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3878 EV->getIndices() == Idxs) {
3879 // insertvalue undef, (extractvalue y, n), n -> y
3880 if (match(Agg, m_Undef()))
3881 return EV->getAggregateOperand();
3883 // insertvalue y, (extractvalue y, n), n -> y
3884 if (Agg == EV->getAggregateOperand())
3891 Value *llvm::SimplifyInsertValueInst(
3892 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL,
3893 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC,
3894 const Instruction *CxtI) {
3895 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI),
3899 /// Given operands for an ExtractValueInst, see if we can fold the result.
3900 /// If not, this returns null.
3901 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3902 const Query &, unsigned) {
3903 if (auto *CAgg = dyn_cast<Constant>(Agg))
3904 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3906 // extractvalue x, (insertvalue y, elt, n), n -> elt
3907 unsigned NumIdxs = Idxs.size();
3908 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3909 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3910 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3911 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3912 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3913 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3914 Idxs.slice(0, NumCommonIdxs)) {
3915 if (NumIdxs == NumInsertValueIdxs)
3916 return IVI->getInsertedValueOperand();
3924 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3925 const DataLayout &DL,
3926 const TargetLibraryInfo *TLI,
3927 const DominatorTree *DT,
3928 AssumptionCache *AC,
3929 const Instruction *CxtI) {
3930 return ::SimplifyExtractValueInst(Agg, Idxs, Query(DL, TLI, DT, AC, CxtI),
3934 /// Given operands for an ExtractElementInst, see if we can fold the result.
3935 /// If not, this returns null.
3936 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const Query &,
3938 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3939 if (auto *CIdx = dyn_cast<Constant>(Idx))
3940 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3942 // The index is not relevant if our vector is a splat.
3943 if (auto *Splat = CVec->getSplatValue())
3946 if (isa<UndefValue>(Vec))
3947 return UndefValue::get(Vec->getType()->getVectorElementType());
3950 // If extracting a specified index from the vector, see if we can recursively
3951 // find a previously computed scalar that was inserted into the vector.
3952 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3953 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3959 Value *llvm::SimplifyExtractElementInst(
3960 Value *Vec, Value *Idx, const DataLayout &DL, const TargetLibraryInfo *TLI,
3961 const DominatorTree *DT, AssumptionCache *AC, const Instruction *CxtI) {
3962 return ::SimplifyExtractElementInst(Vec, Idx, Query(DL, TLI, DT, AC, CxtI),
3966 /// See if we can fold the given phi. If not, returns null.
3967 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
3968 // If all of the PHI's incoming values are the same then replace the PHI node
3969 // with the common value.
3970 Value *CommonValue = nullptr;
3971 bool HasUndefInput = false;
3972 for (Value *Incoming : PN->incoming_values()) {
3973 // If the incoming value is the phi node itself, it can safely be skipped.
3974 if (Incoming == PN) continue;
3975 if (isa<UndefValue>(Incoming)) {
3976 // Remember that we saw an undef value, but otherwise ignore them.
3977 HasUndefInput = true;
3980 if (CommonValue && Incoming != CommonValue)
3981 return nullptr; // Not the same, bail out.
3982 CommonValue = Incoming;
3985 // If CommonValue is null then all of the incoming values were either undef or
3986 // equal to the phi node itself.
3988 return UndefValue::get(PN->getType());
3990 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3991 // instruction, we cannot return X as the result of the PHI node unless it
3992 // dominates the PHI block.
3994 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3999 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4000 Type *Ty, const Query &Q, unsigned MaxRecurse) {
4001 if (auto *C = dyn_cast<Constant>(Op))
4002 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4004 if (auto *CI = dyn_cast<CastInst>(Op)) {
4005 auto *Src = CI->getOperand(0);
4006 Type *SrcTy = Src->getType();
4007 Type *MidTy = CI->getType();
4009 if (Src->getType() == Ty) {
4010 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4011 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4013 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4015 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4017 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4018 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4019 SrcIntPtrTy, MidIntPtrTy,
4020 DstIntPtrTy) == Instruction::BitCast)
4026 if (CastOpc == Instruction::BitCast)
4027 if (Op->getType() == Ty)
4033 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4034 const DataLayout &DL,
4035 const TargetLibraryInfo *TLI,
4036 const DominatorTree *DT, AssumptionCache *AC,
4037 const Instruction *CxtI) {
4038 return ::SimplifyCastInst(CastOpc, Op, Ty, Query(DL, TLI, DT, AC, CxtI),
4042 //=== Helper functions for higher up the class hierarchy.
4044 /// Given operands for a BinaryOperator, see if we can fold the result.
4045 /// If not, this returns null.
4046 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4047 const Query &Q, unsigned MaxRecurse) {
4049 case Instruction::Add:
4050 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
4052 case Instruction::FAdd:
4053 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4055 case Instruction::Sub:
4056 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
4058 case Instruction::FSub:
4059 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4061 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
4062 case Instruction::FMul:
4063 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4064 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4065 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4066 case Instruction::FDiv:
4067 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4068 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4069 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4070 case Instruction::FRem:
4071 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4072 case Instruction::Shl:
4073 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
4075 case Instruction::LShr:
4076 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
4077 case Instruction::AShr:
4078 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
4079 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4080 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
4081 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4083 if (Constant *CLHS = dyn_cast<Constant>(LHS))
4084 if (Constant *CRHS = dyn_cast<Constant>(RHS))
4085 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
4087 // If the operation is associative, try some generic simplifications.
4088 if (Instruction::isAssociative(Opcode))
4089 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
4092 // If the operation is with the result of a select instruction check whether
4093 // operating on either branch of the select always yields the same value.
4094 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
4095 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
4098 // If the operation is with the result of a phi instruction, check whether
4099 // operating on all incoming values of the phi always yields the same value.
4100 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
4101 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
4108 /// Given operands for a BinaryOperator, see if we can fold the result.
4109 /// If not, this returns null.
4110 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4111 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4112 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4113 const FastMathFlags &FMF, const Query &Q,
4114 unsigned MaxRecurse) {
4116 case Instruction::FAdd:
4117 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4118 case Instruction::FSub:
4119 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4120 case Instruction::FMul:
4121 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4122 case Instruction::FDiv:
4123 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4125 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4129 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4130 const DataLayout &DL, const TargetLibraryInfo *TLI,
4131 const DominatorTree *DT, AssumptionCache *AC,
4132 const Instruction *CxtI) {
4133 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
4137 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4138 const FastMathFlags &FMF, const DataLayout &DL,
4139 const TargetLibraryInfo *TLI,
4140 const DominatorTree *DT, AssumptionCache *AC,
4141 const Instruction *CxtI) {
4142 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI),
4146 /// Given operands for a CmpInst, see if we can fold the result.
4147 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4148 const Query &Q, unsigned MaxRecurse) {
4149 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4150 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4151 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4154 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4155 const DataLayout &DL, const TargetLibraryInfo *TLI,
4156 const DominatorTree *DT, AssumptionCache *AC,
4157 const Instruction *CxtI) {
4158 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
4162 static bool IsIdempotent(Intrinsic::ID ID) {
4164 default: return false;
4166 // Unary idempotent: f(f(x)) = f(x)
4167 case Intrinsic::fabs:
4168 case Intrinsic::floor:
4169 case Intrinsic::ceil:
4170 case Intrinsic::trunc:
4171 case Intrinsic::rint:
4172 case Intrinsic::nearbyint:
4173 case Intrinsic::round:
4178 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4179 const DataLayout &DL) {
4180 GlobalValue *PtrSym;
4182 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4185 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4186 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4187 Type *Int32PtrTy = Int32Ty->getPointerTo();
4188 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4190 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4191 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4194 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4195 if (OffsetInt % 4 != 0)
4198 Constant *C = ConstantExpr::getGetElementPtr(
4199 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4200 ConstantInt::get(Int64Ty, OffsetInt / 4));
4201 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4205 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4209 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4210 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4215 if (LoadedCE->getOpcode() != Instruction::Sub)
4218 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4219 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4221 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4223 Constant *LoadedRHS = LoadedCE->getOperand(1);
4224 GlobalValue *LoadedRHSSym;
4225 APInt LoadedRHSOffset;
4226 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4228 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4231 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4234 static bool maskIsAllZeroOrUndef(Value *Mask) {
4235 auto *ConstMask = dyn_cast<Constant>(Mask);
4238 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4240 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4242 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4243 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4250 template <typename IterTy>
4251 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4252 const Query &Q, unsigned MaxRecurse) {
4253 Intrinsic::ID IID = F->getIntrinsicID();
4254 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4257 if (NumOperands == 1) {
4258 // Perform idempotent optimizations
4259 if (IsIdempotent(IID)) {
4260 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4261 if (II->getIntrinsicID() == IID)
4267 case Intrinsic::fabs: {
4268 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4277 if (NumOperands == 2) {
4278 Value *LHS = *ArgBegin;
4279 Value *RHS = *(ArgBegin + 1);
4280 Type *ReturnType = F->getReturnType();
4283 case Intrinsic::usub_with_overflow:
4284 case Intrinsic::ssub_with_overflow: {
4285 // X - X -> { 0, false }
4287 return Constant::getNullValue(ReturnType);
4289 // X - undef -> undef
4290 // undef - X -> undef
4291 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4292 return UndefValue::get(ReturnType);
4296 case Intrinsic::uadd_with_overflow:
4297 case Intrinsic::sadd_with_overflow: {
4298 // X + undef -> undef
4299 if (isa<UndefValue>(RHS))
4300 return UndefValue::get(ReturnType);
4304 case Intrinsic::umul_with_overflow:
4305 case Intrinsic::smul_with_overflow: {
4306 // X * 0 -> { 0, false }
4307 if (match(RHS, m_Zero()))
4308 return Constant::getNullValue(ReturnType);
4310 // X * undef -> { 0, false }
4311 if (match(RHS, m_Undef()))
4312 return Constant::getNullValue(ReturnType);
4316 case Intrinsic::load_relative: {
4317 Constant *C0 = dyn_cast<Constant>(LHS);
4318 Constant *C1 = dyn_cast<Constant>(RHS);
4320 return SimplifyRelativeLoad(C0, C1, Q.DL);
4328 // Simplify calls to llvm.masked.load.*
4330 case Intrinsic::masked_load: {
4331 Value *MaskArg = ArgBegin[2];
4332 Value *PassthruArg = ArgBegin[3];
4333 // If the mask is all zeros or undef, the "passthru" argument is the result.
4334 if (maskIsAllZeroOrUndef(MaskArg))
4343 template <typename IterTy>
4344 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
4345 const Query &Q, unsigned MaxRecurse) {
4346 Type *Ty = V->getType();
4347 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4348 Ty = PTy->getElementType();
4349 FunctionType *FTy = cast<FunctionType>(Ty);
4351 // call undef -> undef
4352 // call null -> undef
4353 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4354 return UndefValue::get(FTy->getReturnType());
4356 Function *F = dyn_cast<Function>(V);
4360 if (F->isIntrinsic())
4361 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4364 if (!canConstantFoldCallTo(F))
4367 SmallVector<Constant *, 4> ConstantArgs;
4368 ConstantArgs.reserve(ArgEnd - ArgBegin);
4369 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4370 Constant *C = dyn_cast<Constant>(*I);
4373 ConstantArgs.push_back(C);
4376 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
4379 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
4380 User::op_iterator ArgEnd, const DataLayout &DL,
4381 const TargetLibraryInfo *TLI, const DominatorTree *DT,
4382 AssumptionCache *AC, const Instruction *CxtI) {
4383 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI),
4387 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
4388 const DataLayout &DL, const TargetLibraryInfo *TLI,
4389 const DominatorTree *DT, AssumptionCache *AC,
4390 const Instruction *CxtI) {
4391 return ::SimplifyCall(V, Args.begin(), Args.end(),
4392 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
4395 /// See if we can compute a simplified version of this instruction.
4396 /// If not, this returns null.
4397 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL,
4398 const TargetLibraryInfo *TLI,
4399 const DominatorTree *DT, AssumptionCache *AC) {
4402 switch (I->getOpcode()) {
4404 Result = ConstantFoldInstruction(I, DL, TLI);
4406 case Instruction::FAdd:
4407 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4408 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4410 case Instruction::Add:
4411 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4412 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4413 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
4416 case Instruction::FSub:
4417 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4418 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4420 case Instruction::Sub:
4421 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4422 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4423 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
4426 case Instruction::FMul:
4427 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4428 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4430 case Instruction::Mul:
4432 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
4434 case Instruction::SDiv:
4435 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
4438 case Instruction::UDiv:
4439 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
4442 case Instruction::FDiv:
4443 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4444 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4446 case Instruction::SRem:
4447 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
4450 case Instruction::URem:
4451 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
4454 case Instruction::FRem:
4455 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4456 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4458 case Instruction::Shl:
4459 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4460 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4461 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
4464 case Instruction::LShr:
4465 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4466 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
4469 case Instruction::AShr:
4470 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4471 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
4474 case Instruction::And:
4476 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
4478 case Instruction::Or:
4480 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
4482 case Instruction::Xor:
4484 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
4486 case Instruction::ICmp:
4488 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0),
4489 I->getOperand(1), DL, TLI, DT, AC, I);
4491 case Instruction::FCmp:
4492 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
4493 I->getOperand(0), I->getOperand(1),
4494 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4496 case Instruction::Select:
4497 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4498 I->getOperand(2), DL, TLI, DT, AC, I);
4500 case Instruction::GetElementPtr: {
4501 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
4502 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4503 Ops, DL, TLI, DT, AC, I);
4506 case Instruction::InsertValue: {
4507 InsertValueInst *IV = cast<InsertValueInst>(I);
4508 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4509 IV->getInsertedValueOperand(),
4510 IV->getIndices(), DL, TLI, DT, AC, I);
4513 case Instruction::ExtractValue: {
4514 auto *EVI = cast<ExtractValueInst>(I);
4515 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4516 EVI->getIndices(), DL, TLI, DT, AC, I);
4519 case Instruction::ExtractElement: {
4520 auto *EEI = cast<ExtractElementInst>(I);
4521 Result = SimplifyExtractElementInst(
4522 EEI->getVectorOperand(), EEI->getIndexOperand(), DL, TLI, DT, AC, I);
4525 case Instruction::PHI:
4526 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I));
4528 case Instruction::Call: {
4529 CallSite CS(cast<CallInst>(I));
4530 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL,
4534 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4535 #include "llvm/IR/Instruction.def"
4536 #undef HANDLE_CAST_INST
4537 Result = SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(),
4538 DL, TLI, DT, AC, I);
4542 // In general, it is possible for computeKnownBits to determine all bits in a
4543 // value even when the operands are not all constants.
4544 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4545 unsigned BitWidth = I->getType()->getScalarSizeInBits();
4546 APInt KnownZero(BitWidth, 0);
4547 APInt KnownOne(BitWidth, 0);
4548 computeKnownBits(I, KnownZero, KnownOne, DL, /*Depth*/0, AC, I, DT);
4549 if ((KnownZero | KnownOne).isAllOnesValue())
4550 Result = ConstantInt::get(I->getType(), KnownOne);
4553 /// If called on unreachable code, the above logic may report that the
4554 /// instruction simplified to itself. Make life easier for users by
4555 /// detecting that case here, returning a safe value instead.
4556 return Result == I ? UndefValue::get(I->getType()) : Result;
4559 /// \brief Implementation of recursive simplification through an instruction's
4562 /// This is the common implementation of the recursive simplification routines.
4563 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4564 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4565 /// instructions to process and attempt to simplify it using
4566 /// InstructionSimplify.
4568 /// This routine returns 'true' only when *it* simplifies something. The passed
4569 /// in simplified value does not count toward this.
4570 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4571 const TargetLibraryInfo *TLI,
4572 const DominatorTree *DT,
4573 AssumptionCache *AC) {
4574 bool Simplified = false;
4575 SmallSetVector<Instruction *, 8> Worklist;
4576 const DataLayout &DL = I->getModule()->getDataLayout();
4578 // If we have an explicit value to collapse to, do that round of the
4579 // simplification loop by hand initially.
4581 for (User *U : I->users())
4583 Worklist.insert(cast<Instruction>(U));
4585 // Replace the instruction with its simplified value.
4586 I->replaceAllUsesWith(SimpleV);
4588 // Gracefully handle edge cases where the instruction is not wired into any
4590 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4591 !I->mayHaveSideEffects())
4592 I->eraseFromParent();
4597 // Note that we must test the size on each iteration, the worklist can grow.
4598 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4601 // See if this instruction simplifies.
4602 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC);
4608 // Stash away all the uses of the old instruction so we can check them for
4609 // recursive simplifications after a RAUW. This is cheaper than checking all
4610 // uses of To on the recursive step in most cases.
4611 for (User *U : I->users())
4612 Worklist.insert(cast<Instruction>(U));
4614 // Replace the instruction with its simplified value.
4615 I->replaceAllUsesWith(SimpleV);
4617 // Gracefully handle edge cases where the instruction is not wired into any
4619 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4620 !I->mayHaveSideEffects())
4621 I->eraseFromParent();
4626 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4627 const TargetLibraryInfo *TLI,
4628 const DominatorTree *DT,
4629 AssumptionCache *AC) {
4630 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4633 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4634 const TargetLibraryInfo *TLI,
4635 const DominatorTree *DT,
4636 AssumptionCache *AC) {
4637 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4638 assert(SimpleV && "Must provide a simplified value.");
4639 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);