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/ConstantFolding.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/ConstantRange.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/GetElementPtrTypeIterator.h"
31 #include "llvm/IR/GlobalAlias.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/IR/ValueHandle.h"
37 using namespace llvm::PatternMatch;
39 #define DEBUG_TYPE "instsimplify"
41 enum { RecursionLimit = 3 };
43 STATISTIC(NumExpand, "Number of expansions");
44 STATISTIC(NumReassoc, "Number of reassociations");
49 const TargetLibraryInfo *TLI;
50 const DominatorTree *DT;
52 const Instruction *CxtI;
54 Query(const DataLayout &DL, const TargetLibraryInfo *tli,
55 const DominatorTree *dt, AssumptionCache *ac = nullptr,
56 const Instruction *cxti = nullptr)
57 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {}
59 } // end anonymous namespace
61 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
62 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
64 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
65 const Query &, unsigned);
66 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
68 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
69 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
70 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned);
72 /// getFalse - For a boolean type, or a vector of boolean type, return false, or
73 /// a vector with every element false, as appropriate for the type.
74 static Constant *getFalse(Type *Ty) {
75 assert(Ty->getScalarType()->isIntegerTy(1) &&
76 "Expected i1 type or a vector of i1!");
77 return Constant::getNullValue(Ty);
80 /// getTrue - For a boolean type, or a vector of boolean type, return true, or
81 /// a vector with every element true, as appropriate for the type.
82 static Constant *getTrue(Type *Ty) {
83 assert(Ty->getScalarType()->isIntegerTy(1) &&
84 "Expected i1 type or a vector of i1!");
85 return Constant::getAllOnesValue(Ty);
88 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
89 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
91 CmpInst *Cmp = dyn_cast<CmpInst>(V);
94 CmpInst::Predicate CPred = Cmp->getPredicate();
95 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
96 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
98 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
102 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
103 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
104 Instruction *I = dyn_cast<Instruction>(V);
106 // Arguments and constants dominate all instructions.
109 // If we are processing instructions (and/or basic blocks) that have not been
110 // fully added to a function, the parent nodes may still be null. Simply
111 // return the conservative answer in these cases.
112 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
115 // If we have a DominatorTree then do a precise test.
117 if (!DT->isReachableFromEntry(P->getParent()))
119 if (!DT->isReachableFromEntry(I->getParent()))
121 return DT->dominates(I, P);
124 // Otherwise, if the instruction is in the entry block, and is not an invoke,
125 // then it obviously dominates all phi nodes.
126 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
133 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning
134 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
135 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
136 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
137 /// Returns the simplified value, or null if no simplification was performed.
138 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
139 unsigned OpcToExpand, const Query &Q,
140 unsigned MaxRecurse) {
141 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
142 // Recursion is always used, so bail out at once if we already hit the limit.
146 // Check whether the expression has the form "(A op' B) op C".
147 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
148 if (Op0->getOpcode() == OpcodeToExpand) {
149 // It does! Try turning it into "(A op C) op' (B op C)".
150 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
151 // Do "A op C" and "B op C" both simplify?
152 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
153 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
154 // They do! Return "L op' R" if it simplifies or is already available.
155 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
156 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
157 && L == B && R == A)) {
161 // Otherwise return "L op' R" if it simplifies.
162 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
169 // Check whether the expression has the form "A op (B op' C)".
170 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
171 if (Op1->getOpcode() == OpcodeToExpand) {
172 // It does! Try turning it into "(A op B) op' (A op C)".
173 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
174 // Do "A op B" and "A op C" both simplify?
175 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
176 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
177 // They do! Return "L op' R" if it simplifies or is already available.
178 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
179 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
180 && L == C && R == B)) {
184 // Otherwise return "L op' R" if it simplifies.
185 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
195 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
196 /// operations. Returns the simpler value, or null if none was found.
197 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
198 const Query &Q, unsigned MaxRecurse) {
199 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
200 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
202 // Recursion is always used, so bail out at once if we already hit the limit.
206 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
207 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
209 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
210 if (Op0 && Op0->getOpcode() == Opcode) {
211 Value *A = Op0->getOperand(0);
212 Value *B = Op0->getOperand(1);
215 // Does "B op C" simplify?
216 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
217 // It does! Return "A op V" if it simplifies or is already available.
218 // If V equals B then "A op V" is just the LHS.
219 if (V == B) return LHS;
220 // Otherwise return "A op V" if it simplifies.
221 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
228 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
229 if (Op1 && Op1->getOpcode() == Opcode) {
231 Value *B = Op1->getOperand(0);
232 Value *C = Op1->getOperand(1);
234 // Does "A op B" simplify?
235 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
236 // It does! Return "V op C" if it simplifies or is already available.
237 // If V equals B then "V op C" is just the RHS.
238 if (V == B) return RHS;
239 // Otherwise return "V op C" if it simplifies.
240 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
247 // The remaining transforms require commutativity as well as associativity.
248 if (!Instruction::isCommutative(Opcode))
251 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
252 if (Op0 && Op0->getOpcode() == Opcode) {
253 Value *A = Op0->getOperand(0);
254 Value *B = Op0->getOperand(1);
257 // Does "C op A" simplify?
258 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
259 // It does! Return "V op B" if it simplifies or is already available.
260 // If V equals A then "V op B" is just the LHS.
261 if (V == A) return LHS;
262 // Otherwise return "V op B" if it simplifies.
263 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
270 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
271 if (Op1 && Op1->getOpcode() == Opcode) {
273 Value *B = Op1->getOperand(0);
274 Value *C = Op1->getOperand(1);
276 // Does "C op A" simplify?
277 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
278 // It does! Return "B op V" if it simplifies or is already available.
279 // If V equals C then "B op V" is just the RHS.
280 if (V == C) return RHS;
281 // Otherwise return "B op V" if it simplifies.
282 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
292 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
293 /// instruction as an operand, try to simplify the binop by seeing whether
294 /// evaluating it on both branches of the select results in the same value.
295 /// Returns the common value if so, otherwise returns null.
296 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
297 const Query &Q, unsigned MaxRecurse) {
298 // Recursion is always used, so bail out at once if we already hit the limit.
303 if (isa<SelectInst>(LHS)) {
304 SI = cast<SelectInst>(LHS);
306 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
307 SI = cast<SelectInst>(RHS);
310 // Evaluate the BinOp on the true and false branches of the select.
314 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
315 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
317 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
318 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
321 // If they simplified to the same value, then return the common value.
322 // If they both failed to simplify then return null.
326 // If one branch simplified to undef, return the other one.
327 if (TV && isa<UndefValue>(TV))
329 if (FV && isa<UndefValue>(FV))
332 // If applying the operation did not change the true and false select values,
333 // then the result of the binop is the select itself.
334 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
337 // If one branch simplified and the other did not, and the simplified
338 // value is equal to the unsimplified one, return the simplified value.
339 // For example, select (cond, X, X & Z) & Z -> X & Z.
340 if ((FV && !TV) || (TV && !FV)) {
341 // Check that the simplified value has the form "X op Y" where "op" is the
342 // same as the original operation.
343 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
344 if (Simplified && Simplified->getOpcode() == Opcode) {
345 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
346 // We already know that "op" is the same as for the simplified value. See
347 // if the operands match too. If so, return the simplified value.
348 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
349 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
350 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
351 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
352 Simplified->getOperand(1) == UnsimplifiedRHS)
354 if (Simplified->isCommutative() &&
355 Simplified->getOperand(1) == UnsimplifiedLHS &&
356 Simplified->getOperand(0) == UnsimplifiedRHS)
364 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
365 /// try to simplify the comparison by seeing whether both branches of the select
366 /// result in the same value. Returns the common value if so, otherwise returns
368 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
369 Value *RHS, const Query &Q,
370 unsigned MaxRecurse) {
371 // Recursion is always used, so bail out at once if we already hit the limit.
375 // Make sure the select is on the LHS.
376 if (!isa<SelectInst>(LHS)) {
378 Pred = CmpInst::getSwappedPredicate(Pred);
380 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
381 SelectInst *SI = cast<SelectInst>(LHS);
382 Value *Cond = SI->getCondition();
383 Value *TV = SI->getTrueValue();
384 Value *FV = SI->getFalseValue();
386 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
387 // Does "cmp TV, RHS" simplify?
388 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
390 // It not only simplified, it simplified to the select condition. Replace
392 TCmp = getTrue(Cond->getType());
394 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
395 // condition then we can replace it with 'true'. Otherwise give up.
396 if (!isSameCompare(Cond, Pred, TV, RHS))
398 TCmp = getTrue(Cond->getType());
401 // Does "cmp FV, RHS" simplify?
402 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
404 // It not only simplified, it simplified to the select condition. Replace
406 FCmp = getFalse(Cond->getType());
408 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
409 // condition then we can replace it with 'false'. Otherwise give up.
410 if (!isSameCompare(Cond, Pred, FV, RHS))
412 FCmp = getFalse(Cond->getType());
415 // If both sides simplified to the same value, then use it as the result of
416 // the original comparison.
420 // The remaining cases only make sense if the select condition has the same
421 // type as the result of the comparison, so bail out if this is not so.
422 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
424 // If the false value simplified to false, then the result of the compare
425 // is equal to "Cond && TCmp". This also catches the case when the false
426 // value simplified to false and the true value to true, returning "Cond".
427 if (match(FCmp, m_Zero()))
428 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
430 // If the true value simplified to true, then the result of the compare
431 // is equal to "Cond || FCmp".
432 if (match(TCmp, m_One()))
433 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
435 // Finally, if the false value simplified to true and the true value to
436 // false, then the result of the compare is equal to "!Cond".
437 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
439 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
446 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
447 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
448 /// it on the incoming phi values yields the same result for every value. If so
449 /// returns the common value, otherwise returns null.
450 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
451 const Query &Q, unsigned MaxRecurse) {
452 // Recursion is always used, so bail out at once if we already hit the limit.
457 if (isa<PHINode>(LHS)) {
458 PI = cast<PHINode>(LHS);
459 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
460 if (!ValueDominatesPHI(RHS, PI, Q.DT))
463 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
464 PI = cast<PHINode>(RHS);
465 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
466 if (!ValueDominatesPHI(LHS, PI, Q.DT))
470 // Evaluate the BinOp on the incoming phi values.
471 Value *CommonValue = nullptr;
472 for (Value *Incoming : PI->incoming_values()) {
473 // If the incoming value is the phi node itself, it can safely be skipped.
474 if (Incoming == PI) continue;
475 Value *V = PI == LHS ?
476 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
477 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
478 // If the operation failed to simplify, or simplified to a different value
479 // to previously, then give up.
480 if (!V || (CommonValue && V != CommonValue))
488 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
489 /// try to simplify the comparison by seeing whether comparing with all of the
490 /// incoming phi values yields the same result every time. If so returns the
491 /// common result, otherwise returns null.
492 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
493 const Query &Q, unsigned MaxRecurse) {
494 // Recursion is always used, so bail out at once if we already hit the limit.
498 // Make sure the phi is on the LHS.
499 if (!isa<PHINode>(LHS)) {
501 Pred = CmpInst::getSwappedPredicate(Pred);
503 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
504 PHINode *PI = cast<PHINode>(LHS);
506 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
507 if (!ValueDominatesPHI(RHS, PI, Q.DT))
510 // Evaluate the BinOp on the incoming phi values.
511 Value *CommonValue = nullptr;
512 for (Value *Incoming : PI->incoming_values()) {
513 // If the incoming value is the phi node itself, it can safely be skipped.
514 if (Incoming == PI) continue;
515 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
516 // If the operation failed to simplify, or simplified to a different value
517 // to previously, then give up.
518 if (!V || (CommonValue && V != CommonValue))
526 /// SimplifyAddInst - Given operands for an Add, see if we can
527 /// fold the result. If not, this returns null.
528 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
529 const Query &Q, unsigned MaxRecurse) {
530 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
531 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
532 Constant *Ops[] = { CLHS, CRHS };
533 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
537 // Canonicalize the constant to the RHS.
541 // X + undef -> undef
542 if (match(Op1, m_Undef()))
546 if (match(Op1, m_Zero()))
553 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
554 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
557 // X + ~X -> -1 since ~X = -X-1
558 if (match(Op0, m_Not(m_Specific(Op1))) ||
559 match(Op1, m_Not(m_Specific(Op0))))
560 return Constant::getAllOnesValue(Op0->getType());
563 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
564 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
567 // Try some generic simplifications for associative operations.
568 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
572 // Threading Add over selects and phi nodes is pointless, so don't bother.
573 // Threading over the select in "A + select(cond, B, C)" means evaluating
574 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
575 // only if B and C are equal. If B and C are equal then (since we assume
576 // that operands have already been simplified) "select(cond, B, C)" should
577 // have been simplified to the common value of B and C already. Analysing
578 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
579 // for threading over phi nodes.
584 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
585 const DataLayout &DL, const TargetLibraryInfo *TLI,
586 const DominatorTree *DT, AssumptionCache *AC,
587 const Instruction *CxtI) {
588 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
592 /// \brief Compute the base pointer and cumulative constant offsets for V.
594 /// This strips all constant offsets off of V, leaving it the base pointer, and
595 /// accumulates the total constant offset applied in the returned constant. It
596 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
597 /// no constant offsets applied.
599 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
600 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
602 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
603 bool AllowNonInbounds = false) {
604 assert(V->getType()->getScalarType()->isPointerTy());
606 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
607 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
609 // Even though we don't look through PHI nodes, we could be called on an
610 // instruction in an unreachable block, which may be on a cycle.
611 SmallPtrSet<Value *, 4> Visited;
614 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
615 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
616 !GEP->accumulateConstantOffset(DL, Offset))
618 V = GEP->getPointerOperand();
619 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
620 V = cast<Operator>(V)->getOperand(0);
621 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
622 if (GA->mayBeOverridden())
624 V = GA->getAliasee();
628 assert(V->getType()->getScalarType()->isPointerTy() &&
629 "Unexpected operand type!");
630 } while (Visited.insert(V).second);
632 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
633 if (V->getType()->isVectorTy())
634 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
639 /// \brief Compute the constant difference between two pointer values.
640 /// If the difference is not a constant, returns zero.
641 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
643 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
644 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
646 // If LHS and RHS are not related via constant offsets to the same base
647 // value, there is nothing we can do here.
651 // Otherwise, the difference of LHS - RHS can be computed as:
653 // = (LHSOffset + Base) - (RHSOffset + Base)
654 // = LHSOffset - RHSOffset
655 return ConstantExpr::getSub(LHSOffset, RHSOffset);
658 /// SimplifySubInst - Given operands for a Sub, see if we can
659 /// fold the result. If not, this returns null.
660 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
661 const Query &Q, unsigned MaxRecurse) {
662 if (Constant *CLHS = dyn_cast<Constant>(Op0))
663 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
664 Constant *Ops[] = { CLHS, CRHS };
665 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
669 // X - undef -> undef
670 // undef - X -> undef
671 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
672 return UndefValue::get(Op0->getType());
675 if (match(Op1, m_Zero()))
680 return Constant::getNullValue(Op0->getType());
682 // 0 - X -> 0 if the sub is NUW.
683 if (isNUW && match(Op0, m_Zero()))
686 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
687 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
688 Value *X = nullptr, *Y = nullptr, *Z = Op1;
689 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
690 // See if "V === Y - Z" simplifies.
691 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
692 // It does! Now see if "X + V" simplifies.
693 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
694 // It does, we successfully reassociated!
698 // See if "V === X - Z" simplifies.
699 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
700 // It does! Now see if "Y + V" simplifies.
701 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
702 // It does, we successfully reassociated!
708 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
709 // For example, X - (X + 1) -> -1
711 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
712 // See if "V === X - Y" simplifies.
713 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
714 // It does! Now see if "V - Z" simplifies.
715 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
716 // It does, we successfully reassociated!
720 // See if "V === X - Z" simplifies.
721 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
722 // It does! Now see if "V - Y" simplifies.
723 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
724 // It does, we successfully reassociated!
730 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
731 // For example, X - (X - Y) -> Y.
733 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
734 // See if "V === Z - X" simplifies.
735 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
736 // It does! Now see if "V + Y" simplifies.
737 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
738 // It does, we successfully reassociated!
743 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
744 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
745 match(Op1, m_Trunc(m_Value(Y))))
746 if (X->getType() == Y->getType())
747 // See if "V === X - Y" simplifies.
748 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
749 // It does! Now see if "trunc V" simplifies.
750 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
751 // It does, return the simplified "trunc V".
754 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
755 if (match(Op0, m_PtrToInt(m_Value(X))) &&
756 match(Op1, m_PtrToInt(m_Value(Y))))
757 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
758 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
761 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
762 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
765 // Threading Sub over selects and phi nodes is pointless, so don't bother.
766 // Threading over the select in "A - select(cond, B, C)" means evaluating
767 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
768 // only if B and C are equal. If B and C are equal then (since we assume
769 // that operands have already been simplified) "select(cond, B, C)" should
770 // have been simplified to the common value of B and C already. Analysing
771 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
772 // for threading over phi nodes.
777 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
778 const DataLayout &DL, const TargetLibraryInfo *TLI,
779 const DominatorTree *DT, AssumptionCache *AC,
780 const Instruction *CxtI) {
781 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
785 /// Given operands for an FAdd, see if we can fold the result. If not, this
787 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
788 const Query &Q, unsigned MaxRecurse) {
789 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
790 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
791 Constant *Ops[] = { CLHS, CRHS };
792 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(),
796 // Canonicalize the constant to the RHS.
801 if (match(Op1, m_NegZero()))
804 // fadd X, 0 ==> X, when we know X is not -0
805 if (match(Op1, m_Zero()) &&
806 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
809 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
810 // where nnan and ninf have to occur at least once somewhere in this
812 Value *SubOp = nullptr;
813 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
815 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
818 Instruction *FSub = cast<Instruction>(SubOp);
819 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
820 (FMF.noInfs() || FSub->hasNoInfs()))
821 return Constant::getNullValue(Op0->getType());
827 /// Given operands for an FSub, see if we can fold the result. If not, this
829 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
830 const Query &Q, unsigned MaxRecurse) {
831 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
832 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
833 Constant *Ops[] = { CLHS, CRHS };
834 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(),
840 if (match(Op1, m_Zero()))
843 // fsub X, -0 ==> X, when we know X is not -0
844 if (match(Op1, m_NegZero()) &&
845 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
848 // fsub 0, (fsub -0.0, X) ==> X
850 if (match(Op0, m_AnyZero())) {
851 if (match(Op1, m_FSub(m_NegZero(), m_Value(X))))
853 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
857 // fsub nnan x, x ==> 0.0
858 if (FMF.noNaNs() && Op0 == Op1)
859 return Constant::getNullValue(Op0->getType());
864 /// Given the operands for an FMul, see if we can fold the result
865 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
868 unsigned MaxRecurse) {
869 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
870 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
871 Constant *Ops[] = { CLHS, CRHS };
872 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(),
876 // Canonicalize the constant to the RHS.
881 if (match(Op1, m_FPOne()))
884 // fmul nnan nsz X, 0 ==> 0
885 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
891 /// SimplifyMulInst - Given operands for a Mul, see if we can
892 /// fold the result. If not, this returns null.
893 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
894 unsigned MaxRecurse) {
895 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
896 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
897 Constant *Ops[] = { CLHS, CRHS };
898 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(),
902 // Canonicalize the constant to the RHS.
907 if (match(Op1, m_Undef()))
908 return Constant::getNullValue(Op0->getType());
911 if (match(Op1, m_Zero()))
915 if (match(Op1, m_One()))
918 // (X / Y) * Y -> X if the division is exact.
920 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
921 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
925 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
926 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
929 // Try some generic simplifications for associative operations.
930 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
934 // Mul distributes over Add. Try some generic simplifications based on this.
935 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
939 // If the operation is with the result of a select instruction, check whether
940 // operating on either branch of the select always yields the same value.
941 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
942 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
946 // If the operation is with the result of a phi instruction, check whether
947 // operating on all incoming values of the phi always yields the same value.
948 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
949 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
956 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
957 const DataLayout &DL,
958 const TargetLibraryInfo *TLI,
959 const DominatorTree *DT, AssumptionCache *AC,
960 const Instruction *CxtI) {
961 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
965 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
966 const DataLayout &DL,
967 const TargetLibraryInfo *TLI,
968 const DominatorTree *DT, AssumptionCache *AC,
969 const Instruction *CxtI) {
970 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
974 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
975 const DataLayout &DL,
976 const TargetLibraryInfo *TLI,
977 const DominatorTree *DT, AssumptionCache *AC,
978 const Instruction *CxtI) {
979 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
983 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout &DL,
984 const TargetLibraryInfo *TLI,
985 const DominatorTree *DT, AssumptionCache *AC,
986 const Instruction *CxtI) {
987 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
991 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
992 /// fold the result. If not, this returns null.
993 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
994 const Query &Q, unsigned MaxRecurse) {
995 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
996 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
997 Constant *Ops[] = { C0, C1 };
998 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1002 bool isSigned = Opcode == Instruction::SDiv;
1004 // X / undef -> undef
1005 if (match(Op1, m_Undef()))
1008 // X / 0 -> undef, we don't need to preserve faults!
1009 if (match(Op1, m_Zero()))
1010 return UndefValue::get(Op1->getType());
1013 if (match(Op0, m_Undef()))
1014 return Constant::getNullValue(Op0->getType());
1016 // 0 / X -> 0, we don't need to preserve faults!
1017 if (match(Op0, m_Zero()))
1021 if (match(Op1, m_One()))
1024 if (Op0->getType()->isIntegerTy(1))
1025 // It can't be division by zero, hence it must be division by one.
1030 return ConstantInt::get(Op0->getType(), 1);
1032 // (X * Y) / Y -> X if the multiplication does not overflow.
1033 Value *X = nullptr, *Y = nullptr;
1034 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1035 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1036 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1037 // If the Mul knows it does not overflow, then we are good to go.
1038 if ((isSigned && Mul->hasNoSignedWrap()) ||
1039 (!isSigned && Mul->hasNoUnsignedWrap()))
1041 // If X has the form X = A / Y then X * Y cannot overflow.
1042 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1043 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1047 // (X rem Y) / Y -> 0
1048 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1049 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1050 return Constant::getNullValue(Op0->getType());
1052 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1053 ConstantInt *C1, *C2;
1054 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1055 match(Op1, m_ConstantInt(C2))) {
1057 C1->getValue().umul_ov(C2->getValue(), Overflow);
1059 return Constant::getNullValue(Op0->getType());
1062 // If the operation is with the result of a select instruction, check whether
1063 // operating on either branch of the select always yields the same value.
1064 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1065 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1068 // If the operation is with the result of a phi instruction, check whether
1069 // operating on all incoming values of the phi always yields the same value.
1070 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1071 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1077 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1078 /// fold the result. If not, this returns null.
1079 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1080 unsigned MaxRecurse) {
1081 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1087 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1088 const TargetLibraryInfo *TLI,
1089 const DominatorTree *DT, AssumptionCache *AC,
1090 const Instruction *CxtI) {
1091 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1095 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1096 /// fold the result. If not, this returns null.
1097 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1098 unsigned MaxRecurse) {
1099 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1105 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1106 const TargetLibraryInfo *TLI,
1107 const DominatorTree *DT, AssumptionCache *AC,
1108 const Instruction *CxtI) {
1109 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1113 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1114 const Query &Q, unsigned) {
1115 // undef / X -> undef (the undef could be a snan).
1116 if (match(Op0, m_Undef()))
1119 // X / undef -> undef
1120 if (match(Op1, m_Undef()))
1124 // Requires that NaNs are off (X could be zero) and signed zeroes are
1125 // ignored (X could be positive or negative, so the output sign is unknown).
1126 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1130 // X / X -> 1.0 is legal when NaNs are ignored.
1132 return ConstantFP::get(Op0->getType(), 1.0);
1134 // -X / X -> -1.0 and
1135 // X / -X -> -1.0 are legal when NaNs are ignored.
1136 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1137 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1138 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1139 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1140 BinaryOperator::getFNegArgument(Op1) == Op0))
1141 return ConstantFP::get(Op0->getType(), -1.0);
1147 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1148 const DataLayout &DL,
1149 const TargetLibraryInfo *TLI,
1150 const DominatorTree *DT, AssumptionCache *AC,
1151 const Instruction *CxtI) {
1152 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1156 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1157 /// fold the result. If not, this returns null.
1158 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1159 const Query &Q, unsigned MaxRecurse) {
1160 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1161 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1162 Constant *Ops[] = { C0, C1 };
1163 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1167 // X % undef -> undef
1168 if (match(Op1, m_Undef()))
1172 if (match(Op0, m_Undef()))
1173 return Constant::getNullValue(Op0->getType());
1175 // 0 % X -> 0, we don't need to preserve faults!
1176 if (match(Op0, m_Zero()))
1179 // X % 0 -> undef, we don't need to preserve faults!
1180 if (match(Op1, m_Zero()))
1181 return UndefValue::get(Op0->getType());
1184 if (match(Op1, m_One()))
1185 return Constant::getNullValue(Op0->getType());
1187 if (Op0->getType()->isIntegerTy(1))
1188 // It can't be remainder by zero, hence it must be remainder by one.
1189 return Constant::getNullValue(Op0->getType());
1193 return Constant::getNullValue(Op0->getType());
1195 // (X % Y) % Y -> X % Y
1196 if ((Opcode == Instruction::SRem &&
1197 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1198 (Opcode == Instruction::URem &&
1199 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1202 // If the operation is with the result of a select instruction, check whether
1203 // operating on either branch of the select always yields the same value.
1204 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1205 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1208 // If the operation is with the result of a phi instruction, check whether
1209 // operating on all incoming values of the phi always yields the same value.
1210 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1211 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1217 /// SimplifySRemInst - Given operands for an SRem, see if we can
1218 /// fold the result. If not, this returns null.
1219 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1220 unsigned MaxRecurse) {
1221 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1227 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1228 const TargetLibraryInfo *TLI,
1229 const DominatorTree *DT, AssumptionCache *AC,
1230 const Instruction *CxtI) {
1231 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1235 /// SimplifyURemInst - Given operands for a URem, see if we can
1236 /// fold the result. If not, this returns null.
1237 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1238 unsigned MaxRecurse) {
1239 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1245 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1246 const TargetLibraryInfo *TLI,
1247 const DominatorTree *DT, AssumptionCache *AC,
1248 const Instruction *CxtI) {
1249 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1253 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1254 const Query &, unsigned) {
1255 // undef % X -> undef (the undef could be a snan).
1256 if (match(Op0, m_Undef()))
1259 // X % undef -> undef
1260 if (match(Op1, m_Undef()))
1264 // Requires that NaNs are off (X could be zero) and signed zeroes are
1265 // ignored (X could be positive or negative, so the output sign is unknown).
1266 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1272 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1273 const DataLayout &DL,
1274 const TargetLibraryInfo *TLI,
1275 const DominatorTree *DT, AssumptionCache *AC,
1276 const Instruction *CxtI) {
1277 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1281 /// isUndefShift - Returns true if a shift by \c Amount always yields undef.
1282 static bool isUndefShift(Value *Amount) {
1283 Constant *C = dyn_cast<Constant>(Amount);
1287 // X shift by undef -> undef because it may shift by the bitwidth.
1288 if (isa<UndefValue>(C))
1291 // Shifting by the bitwidth or more is undefined.
1292 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1293 if (CI->getValue().getLimitedValue() >=
1294 CI->getType()->getScalarSizeInBits())
1297 // If all lanes of a vector shift are undefined the whole shift is.
1298 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1299 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1300 if (!isUndefShift(C->getAggregateElement(I)))
1308 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1309 /// fold the result. If not, this returns null.
1310 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1311 const Query &Q, unsigned MaxRecurse) {
1312 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1313 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1314 Constant *Ops[] = { C0, C1 };
1315 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1319 // 0 shift by X -> 0
1320 if (match(Op0, m_Zero()))
1323 // X shift by 0 -> X
1324 if (match(Op1, m_Zero()))
1327 // Fold undefined shifts.
1328 if (isUndefShift(Op1))
1329 return UndefValue::get(Op0->getType());
1331 // If the operation is with the result of a select instruction, check whether
1332 // operating on either branch of the select always yields the same value.
1333 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1334 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1337 // If the operation is with the result of a phi instruction, check whether
1338 // operating on all incoming values of the phi always yields the same value.
1339 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1340 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1346 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1347 /// fold the result. If not, this returns null.
1348 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1,
1349 bool isExact, const Query &Q,
1350 unsigned MaxRecurse) {
1351 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1356 return Constant::getNullValue(Op0->getType());
1359 // undef >> X -> undef (if it's exact)
1360 if (match(Op0, m_Undef()))
1361 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1363 // The low bit cannot be shifted out of an exact shift if it is set.
1365 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
1366 APInt Op0KnownZero(BitWidth, 0);
1367 APInt Op0KnownOne(BitWidth, 0);
1368 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC,
1377 /// SimplifyShlInst - Given operands for an Shl, see if we can
1378 /// fold the result. If not, this returns null.
1379 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1380 const Query &Q, unsigned MaxRecurse) {
1381 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1385 // undef << X -> undef if (if it's NSW/NUW)
1386 if (match(Op0, m_Undef()))
1387 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1389 // (X >> A) << A -> X
1391 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1396 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1397 const DataLayout &DL, const TargetLibraryInfo *TLI,
1398 const DominatorTree *DT, AssumptionCache *AC,
1399 const Instruction *CxtI) {
1400 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
1404 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1405 /// fold the result. If not, this returns null.
1406 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1407 const Query &Q, unsigned MaxRecurse) {
1408 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1412 // (X << A) >> A -> X
1414 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1420 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1421 const DataLayout &DL,
1422 const TargetLibraryInfo *TLI,
1423 const DominatorTree *DT, AssumptionCache *AC,
1424 const Instruction *CxtI) {
1425 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1429 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1430 /// fold the result. If not, this returns null.
1431 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1432 const Query &Q, unsigned MaxRecurse) {
1433 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1437 // all ones >>a X -> all ones
1438 if (match(Op0, m_AllOnes()))
1441 // (X << A) >> A -> X
1443 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1446 // Arithmetic shifting an all-sign-bit value is a no-op.
1447 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1448 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1454 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1455 const DataLayout &DL,
1456 const TargetLibraryInfo *TLI,
1457 const DominatorTree *DT, AssumptionCache *AC,
1458 const Instruction *CxtI) {
1459 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1463 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1464 ICmpInst *UnsignedICmp, bool IsAnd) {
1467 ICmpInst::Predicate EqPred;
1468 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1469 !ICmpInst::isEquality(EqPred))
1472 ICmpInst::Predicate UnsignedPred;
1473 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1474 ICmpInst::isUnsigned(UnsignedPred))
1476 else if (match(UnsignedICmp,
1477 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1478 ICmpInst::isUnsigned(UnsignedPred))
1479 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1483 // X < Y && Y != 0 --> X < Y
1484 // X < Y || Y != 0 --> Y != 0
1485 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1486 return IsAnd ? UnsignedICmp : ZeroICmp;
1488 // X >= Y || Y != 0 --> true
1489 // X >= Y || Y == 0 --> X >= Y
1490 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1491 if (EqPred == ICmpInst::ICMP_NE)
1492 return getTrue(UnsignedICmp->getType());
1493 return UnsignedICmp;
1496 // X < Y && Y == 0 --> false
1497 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1499 return getFalse(UnsignedICmp->getType());
1504 // Simplify (and (icmp ...) (icmp ...)) to true when we can tell that the range
1505 // of possible values cannot be satisfied.
1506 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1507 ICmpInst::Predicate Pred0, Pred1;
1508 ConstantInt *CI1, *CI2;
1511 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1514 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)),
1515 m_ConstantInt(CI2))))
1518 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1))))
1521 Type *ITy = Op0->getType();
1523 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1524 bool isNSW = AddInst->hasNoSignedWrap();
1525 bool isNUW = AddInst->hasNoUnsignedWrap();
1527 const APInt &CI1V = CI1->getValue();
1528 const APInt &CI2V = CI2->getValue();
1529 const APInt Delta = CI2V - CI1V;
1530 if (CI1V.isStrictlyPositive()) {
1532 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1533 return getFalse(ITy);
1534 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1535 return getFalse(ITy);
1538 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1539 return getFalse(ITy);
1540 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1541 return getFalse(ITy);
1544 if (CI1V.getBoolValue() && isNUW) {
1546 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1547 return getFalse(ITy);
1549 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1550 return getFalse(ITy);
1556 /// SimplifyAndInst - Given operands for an And, see if we can
1557 /// fold the result. If not, this returns null.
1558 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1559 unsigned MaxRecurse) {
1560 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1561 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1562 Constant *Ops[] = { CLHS, CRHS };
1563 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1567 // Canonicalize the constant to the RHS.
1568 std::swap(Op0, Op1);
1572 if (match(Op1, m_Undef()))
1573 return Constant::getNullValue(Op0->getType());
1580 if (match(Op1, m_Zero()))
1584 if (match(Op1, m_AllOnes()))
1587 // A & ~A = ~A & A = 0
1588 if (match(Op0, m_Not(m_Specific(Op1))) ||
1589 match(Op1, m_Not(m_Specific(Op0))))
1590 return Constant::getNullValue(Op0->getType());
1593 Value *A = nullptr, *B = nullptr;
1594 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1595 (A == Op1 || B == Op1))
1599 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1600 (A == Op0 || B == Op0))
1603 // A & (-A) = A if A is a power of two or zero.
1604 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1605 match(Op1, m_Neg(m_Specific(Op0)))) {
1606 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1609 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1614 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1615 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1616 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS))
1618 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS))
1623 // Try some generic simplifications for associative operations.
1624 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1628 // And distributes over Or. Try some generic simplifications based on this.
1629 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1633 // And distributes over Xor. Try some generic simplifications based on this.
1634 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1638 // If the operation is with the result of a select instruction, check whether
1639 // operating on either branch of the select always yields the same value.
1640 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1641 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1645 // If the operation is with the result of a phi instruction, check whether
1646 // operating on all incoming values of the phi always yields the same value.
1647 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1648 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1655 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout &DL,
1656 const TargetLibraryInfo *TLI,
1657 const DominatorTree *DT, AssumptionCache *AC,
1658 const Instruction *CxtI) {
1659 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1663 // Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union
1664 // contains all possible values.
1665 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1666 ICmpInst::Predicate Pred0, Pred1;
1667 ConstantInt *CI1, *CI2;
1670 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1673 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)),
1674 m_ConstantInt(CI2))))
1677 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1))))
1680 Type *ITy = Op0->getType();
1682 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1683 bool isNSW = AddInst->hasNoSignedWrap();
1684 bool isNUW = AddInst->hasNoUnsignedWrap();
1686 const APInt &CI1V = CI1->getValue();
1687 const APInt &CI2V = CI2->getValue();
1688 const APInt Delta = CI2V - CI1V;
1689 if (CI1V.isStrictlyPositive()) {
1691 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1692 return getTrue(ITy);
1693 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1694 return getTrue(ITy);
1697 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1698 return getTrue(ITy);
1699 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1700 return getTrue(ITy);
1703 if (CI1V.getBoolValue() && isNUW) {
1705 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1706 return getTrue(ITy);
1708 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1709 return getTrue(ITy);
1715 /// SimplifyOrInst - Given operands for an Or, see if we can
1716 /// fold the result. If not, this returns null.
1717 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1718 unsigned MaxRecurse) {
1719 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1720 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1721 Constant *Ops[] = { CLHS, CRHS };
1722 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1726 // Canonicalize the constant to the RHS.
1727 std::swap(Op0, Op1);
1731 if (match(Op1, m_Undef()))
1732 return Constant::getAllOnesValue(Op0->getType());
1739 if (match(Op1, m_Zero()))
1743 if (match(Op1, m_AllOnes()))
1746 // A | ~A = ~A | A = -1
1747 if (match(Op0, m_Not(m_Specific(Op1))) ||
1748 match(Op1, m_Not(m_Specific(Op0))))
1749 return Constant::getAllOnesValue(Op0->getType());
1752 Value *A = nullptr, *B = nullptr;
1753 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1754 (A == Op1 || B == Op1))
1758 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1759 (A == Op0 || B == Op0))
1762 // ~(A & ?) | A = -1
1763 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1764 (A == Op1 || B == Op1))
1765 return Constant::getAllOnesValue(Op1->getType());
1767 // A | ~(A & ?) = -1
1768 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1769 (A == Op0 || B == Op0))
1770 return Constant::getAllOnesValue(Op0->getType());
1772 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1773 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1774 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS))
1776 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS))
1781 // Try some generic simplifications for associative operations.
1782 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1786 // Or distributes over And. Try some generic simplifications based on this.
1787 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1791 // If the operation is with the result of a select instruction, check whether
1792 // operating on either branch of the select always yields the same value.
1793 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1794 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1799 Value *C = nullptr, *D = nullptr;
1800 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1801 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1802 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1803 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1804 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1805 // (A & C1)|(B & C2)
1806 // If we have: ((V + N) & C1) | (V & C2)
1807 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1808 // replace with V+N.
1810 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1811 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1812 // Add commutes, try both ways.
1814 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1817 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1820 // Or commutes, try both ways.
1821 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1822 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1823 // Add commutes, try both ways.
1825 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1828 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1834 // If the operation is with the result of a phi instruction, check whether
1835 // operating on all incoming values of the phi always yields the same value.
1836 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1837 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1843 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout &DL,
1844 const TargetLibraryInfo *TLI,
1845 const DominatorTree *DT, AssumptionCache *AC,
1846 const Instruction *CxtI) {
1847 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1851 /// SimplifyXorInst - Given operands for a Xor, see if we can
1852 /// fold the result. If not, this returns null.
1853 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1854 unsigned MaxRecurse) {
1855 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1856 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1857 Constant *Ops[] = { CLHS, CRHS };
1858 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1862 // Canonicalize the constant to the RHS.
1863 std::swap(Op0, Op1);
1866 // A ^ undef -> undef
1867 if (match(Op1, m_Undef()))
1871 if (match(Op1, m_Zero()))
1876 return Constant::getNullValue(Op0->getType());
1878 // A ^ ~A = ~A ^ A = -1
1879 if (match(Op0, m_Not(m_Specific(Op1))) ||
1880 match(Op1, m_Not(m_Specific(Op0))))
1881 return Constant::getAllOnesValue(Op0->getType());
1883 // Try some generic simplifications for associative operations.
1884 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1888 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1889 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1890 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1891 // only if B and C are equal. If B and C are equal then (since we assume
1892 // that operands have already been simplified) "select(cond, B, C)" should
1893 // have been simplified to the common value of B and C already. Analysing
1894 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1895 // for threading over phi nodes.
1900 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout &DL,
1901 const TargetLibraryInfo *TLI,
1902 const DominatorTree *DT, AssumptionCache *AC,
1903 const Instruction *CxtI) {
1904 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1908 static Type *GetCompareTy(Value *Op) {
1909 return CmpInst::makeCmpResultType(Op->getType());
1912 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1913 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1914 /// otherwise return null. Helper function for analyzing max/min idioms.
1915 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1916 Value *LHS, Value *RHS) {
1917 SelectInst *SI = dyn_cast<SelectInst>(V);
1920 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1923 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1924 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1926 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1927 LHS == CmpRHS && RHS == CmpLHS)
1932 // A significant optimization not implemented here is assuming that alloca
1933 // addresses are not equal to incoming argument values. They don't *alias*,
1934 // as we say, but that doesn't mean they aren't equal, so we take a
1935 // conservative approach.
1937 // This is inspired in part by C++11 5.10p1:
1938 // "Two pointers of the same type compare equal if and only if they are both
1939 // null, both point to the same function, or both represent the same
1942 // This is pretty permissive.
1944 // It's also partly due to C11 6.5.9p6:
1945 // "Two pointers compare equal if and only if both are null pointers, both are
1946 // pointers to the same object (including a pointer to an object and a
1947 // subobject at its beginning) or function, both are pointers to one past the
1948 // last element of the same array object, or one is a pointer to one past the
1949 // end of one array object and the other is a pointer to the start of a
1950 // different array object that happens to immediately follow the first array
1951 // object in the address space.)
1953 // C11's version is more restrictive, however there's no reason why an argument
1954 // couldn't be a one-past-the-end value for a stack object in the caller and be
1955 // equal to the beginning of a stack object in the callee.
1957 // If the C and C++ standards are ever made sufficiently restrictive in this
1958 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1959 // this optimization.
1960 static Constant *computePointerICmp(const DataLayout &DL,
1961 const TargetLibraryInfo *TLI,
1962 CmpInst::Predicate Pred, Value *LHS,
1964 // First, skip past any trivial no-ops.
1965 LHS = LHS->stripPointerCasts();
1966 RHS = RHS->stripPointerCasts();
1968 // A non-null pointer is not equal to a null pointer.
1969 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) &&
1970 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1971 return ConstantInt::get(GetCompareTy(LHS),
1972 !CmpInst::isTrueWhenEqual(Pred));
1974 // We can only fold certain predicates on pointer comparisons.
1979 // Equality comaprisons are easy to fold.
1980 case CmpInst::ICMP_EQ:
1981 case CmpInst::ICMP_NE:
1984 // We can only handle unsigned relational comparisons because 'inbounds' on
1985 // a GEP only protects against unsigned wrapping.
1986 case CmpInst::ICMP_UGT:
1987 case CmpInst::ICMP_UGE:
1988 case CmpInst::ICMP_ULT:
1989 case CmpInst::ICMP_ULE:
1990 // However, we have to switch them to their signed variants to handle
1991 // negative indices from the base pointer.
1992 Pred = ICmpInst::getSignedPredicate(Pred);
1996 // Strip off any constant offsets so that we can reason about them.
1997 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
1998 // here and compare base addresses like AliasAnalysis does, however there are
1999 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2000 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2001 // doesn't need to guarantee pointer inequality when it says NoAlias.
2002 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2003 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2005 // If LHS and RHS are related via constant offsets to the same base
2006 // value, we can replace it with an icmp which just compares the offsets.
2008 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2010 // Various optimizations for (in)equality comparisons.
2011 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2012 // Different non-empty allocations that exist at the same time have
2013 // different addresses (if the program can tell). Global variables always
2014 // exist, so they always exist during the lifetime of each other and all
2015 // allocas. Two different allocas usually have different addresses...
2017 // However, if there's an @llvm.stackrestore dynamically in between two
2018 // allocas, they may have the same address. It's tempting to reduce the
2019 // scope of the problem by only looking at *static* allocas here. That would
2020 // cover the majority of allocas while significantly reducing the likelihood
2021 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2022 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2023 // an entry block. Also, if we have a block that's not attached to a
2024 // function, we can't tell if it's "static" under the current definition.
2025 // Theoretically, this problem could be fixed by creating a new kind of
2026 // instruction kind specifically for static allocas. Such a new instruction
2027 // could be required to be at the top of the entry block, thus preventing it
2028 // from being subject to a @llvm.stackrestore. Instcombine could even
2029 // convert regular allocas into these special allocas. It'd be nifty.
2030 // However, until then, this problem remains open.
2032 // So, we'll assume that two non-empty allocas have different addresses
2035 // With all that, if the offsets are within the bounds of their allocations
2036 // (and not one-past-the-end! so we can't use inbounds!), and their
2037 // allocations aren't the same, the pointers are not equal.
2039 // Note that it's not necessary to check for LHS being a global variable
2040 // address, due to canonicalization and constant folding.
2041 if (isa<AllocaInst>(LHS) &&
2042 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2043 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2044 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2045 uint64_t LHSSize, RHSSize;
2046 if (LHSOffsetCI && RHSOffsetCI &&
2047 getObjectSize(LHS, LHSSize, DL, TLI) &&
2048 getObjectSize(RHS, RHSSize, DL, TLI)) {
2049 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2050 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2051 if (!LHSOffsetValue.isNegative() &&
2052 !RHSOffsetValue.isNegative() &&
2053 LHSOffsetValue.ult(LHSSize) &&
2054 RHSOffsetValue.ult(RHSSize)) {
2055 return ConstantInt::get(GetCompareTy(LHS),
2056 !CmpInst::isTrueWhenEqual(Pred));
2060 // Repeat the above check but this time without depending on DataLayout
2061 // or being able to compute a precise size.
2062 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2063 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2064 LHSOffset->isNullValue() &&
2065 RHSOffset->isNullValue())
2066 return ConstantInt::get(GetCompareTy(LHS),
2067 !CmpInst::isTrueWhenEqual(Pred));
2070 // Even if an non-inbounds GEP occurs along the path we can still optimize
2071 // equality comparisons concerning the result. We avoid walking the whole
2072 // chain again by starting where the last calls to
2073 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2074 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2075 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2077 return ConstantExpr::getICmp(Pred,
2078 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2079 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2081 // If one side of the equality comparison must come from a noalias call
2082 // (meaning a system memory allocation function), and the other side must
2083 // come from a pointer that cannot overlap with dynamically-allocated
2084 // memory within the lifetime of the current function (allocas, byval
2085 // arguments, globals), then determine the comparison result here.
2086 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2087 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2088 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2090 // Is the set of underlying objects all noalias calls?
2091 auto IsNAC = [](SmallVectorImpl<Value *> &Objects) {
2092 return std::all_of(Objects.begin(), Objects.end(),
2093 [](Value *V){ return isNoAliasCall(V); });
2096 // Is the set of underlying objects all things which must be disjoint from
2097 // noalias calls. For allocas, we consider only static ones (dynamic
2098 // allocas might be transformed into calls to malloc not simultaneously
2099 // live with the compared-to allocation). For globals, we exclude symbols
2100 // that might be resolve lazily to symbols in another dynamically-loaded
2101 // library (and, thus, could be malloc'ed by the implementation).
2102 auto IsAllocDisjoint = [](SmallVectorImpl<Value *> &Objects) {
2103 return std::all_of(Objects.begin(), Objects.end(),
2105 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2106 return AI->getParent() && AI->getParent()->getParent() &&
2107 AI->isStaticAlloca();
2108 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2109 return (GV->hasLocalLinkage() ||
2110 GV->hasHiddenVisibility() ||
2111 GV->hasProtectedVisibility() ||
2112 GV->hasUnnamedAddr()) &&
2113 !GV->isThreadLocal();
2114 if (const Argument *A = dyn_cast<Argument>(V))
2115 return A->hasByValAttr();
2120 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2121 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2122 return ConstantInt::get(GetCompareTy(LHS),
2123 !CmpInst::isTrueWhenEqual(Pred));
2130 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
2131 /// fold the result. If not, this returns null.
2132 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2133 const Query &Q, unsigned MaxRecurse) {
2134 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2135 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
2137 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2138 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2139 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
2141 // If we have a constant, make sure it is on the RHS.
2142 std::swap(LHS, RHS);
2143 Pred = CmpInst::getSwappedPredicate(Pred);
2146 Type *ITy = GetCompareTy(LHS); // The return type.
2147 Type *OpTy = LHS->getType(); // The operand type.
2149 // icmp X, X -> true/false
2150 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
2151 // because X could be 0.
2152 if (LHS == RHS || isa<UndefValue>(RHS))
2153 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
2155 // Special case logic when the operands have i1 type.
2156 if (OpTy->getScalarType()->isIntegerTy(1)) {
2159 case ICmpInst::ICMP_EQ:
2161 if (match(RHS, m_One()))
2164 case ICmpInst::ICMP_NE:
2166 if (match(RHS, m_Zero()))
2169 case ICmpInst::ICMP_UGT:
2171 if (match(RHS, m_Zero()))
2174 case ICmpInst::ICMP_UGE:
2176 if (match(RHS, m_One()))
2179 case ICmpInst::ICMP_SLT:
2181 if (match(RHS, m_Zero()))
2184 case ICmpInst::ICMP_SLE:
2186 if (match(RHS, m_One()))
2192 // If we are comparing with zero then try hard since this is a common case.
2193 if (match(RHS, m_Zero())) {
2194 bool LHSKnownNonNegative, LHSKnownNegative;
2196 default: llvm_unreachable("Unknown ICmp predicate!");
2197 case ICmpInst::ICMP_ULT:
2198 return getFalse(ITy);
2199 case ICmpInst::ICMP_UGE:
2200 return getTrue(ITy);
2201 case ICmpInst::ICMP_EQ:
2202 case ICmpInst::ICMP_ULE:
2203 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2204 return getFalse(ITy);
2206 case ICmpInst::ICMP_NE:
2207 case ICmpInst::ICMP_UGT:
2208 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2209 return getTrue(ITy);
2211 case ICmpInst::ICMP_SLT:
2212 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2214 if (LHSKnownNegative)
2215 return getTrue(ITy);
2216 if (LHSKnownNonNegative)
2217 return getFalse(ITy);
2219 case ICmpInst::ICMP_SLE:
2220 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2222 if (LHSKnownNegative)
2223 return getTrue(ITy);
2224 if (LHSKnownNonNegative &&
2225 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2226 return getFalse(ITy);
2228 case ICmpInst::ICMP_SGE:
2229 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2231 if (LHSKnownNegative)
2232 return getFalse(ITy);
2233 if (LHSKnownNonNegative)
2234 return getTrue(ITy);
2236 case ICmpInst::ICMP_SGT:
2237 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2239 if (LHSKnownNegative)
2240 return getFalse(ITy);
2241 if (LHSKnownNonNegative &&
2242 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2243 return getTrue(ITy);
2248 // See if we are doing a comparison with a constant integer.
2249 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2250 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2251 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
2252 if (RHS_CR.isEmptySet())
2253 return ConstantInt::getFalse(CI->getContext());
2254 if (RHS_CR.isFullSet())
2255 return ConstantInt::getTrue(CI->getContext());
2257 // Many binary operators with constant RHS have easy to compute constant
2258 // range. Use them to check whether the comparison is a tautology.
2259 unsigned Width = CI->getBitWidth();
2260 APInt Lower = APInt(Width, 0);
2261 APInt Upper = APInt(Width, 0);
2263 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
2264 // 'urem x, CI2' produces [0, CI2).
2265 Upper = CI2->getValue();
2266 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
2267 // 'srem x, CI2' produces (-|CI2|, |CI2|).
2268 Upper = CI2->getValue().abs();
2269 Lower = (-Upper) + 1;
2270 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
2271 // 'udiv CI2, x' produces [0, CI2].
2272 Upper = CI2->getValue() + 1;
2273 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
2274 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
2275 APInt NegOne = APInt::getAllOnesValue(Width);
2277 Upper = NegOne.udiv(CI2->getValue()) + 1;
2278 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) {
2279 if (CI2->isMinSignedValue()) {
2280 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2281 Lower = CI2->getValue();
2282 Upper = Lower.lshr(1) + 1;
2284 // 'sdiv CI2, x' produces [-|CI2|, |CI2|].
2285 Upper = CI2->getValue().abs() + 1;
2286 Lower = (-Upper) + 1;
2288 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
2289 APInt IntMin = APInt::getSignedMinValue(Width);
2290 APInt IntMax = APInt::getSignedMaxValue(Width);
2291 APInt Val = CI2->getValue();
2292 if (Val.isAllOnesValue()) {
2293 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2294 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2297 } else if (Val.countLeadingZeros() < Width - 1) {
2298 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2]
2299 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2300 Lower = IntMin.sdiv(Val);
2301 Upper = IntMax.sdiv(Val);
2302 if (Lower.sgt(Upper))
2303 std::swap(Lower, Upper);
2305 assert(Upper != Lower && "Upper part of range has wrapped!");
2307 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) {
2308 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)]
2309 Lower = CI2->getValue();
2310 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2311 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) {
2312 if (CI2->isNegative()) {
2313 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2]
2314 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1;
2315 Lower = CI2->getValue().shl(ShiftAmount);
2316 Upper = CI2->getValue() + 1;
2318 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1]
2319 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1;
2320 Lower = CI2->getValue();
2321 Upper = CI2->getValue().shl(ShiftAmount) + 1;
2323 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
2324 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
2325 APInt NegOne = APInt::getAllOnesValue(Width);
2326 if (CI2->getValue().ult(Width))
2327 Upper = NegOne.lshr(CI2->getValue()) + 1;
2328 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) {
2329 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2].
2330 unsigned ShiftAmount = Width - 1;
2331 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2332 ShiftAmount = CI2->getValue().countTrailingZeros();
2333 Lower = CI2->getValue().lshr(ShiftAmount);
2334 Upper = CI2->getValue() + 1;
2335 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
2336 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
2337 APInt IntMin = APInt::getSignedMinValue(Width);
2338 APInt IntMax = APInt::getSignedMaxValue(Width);
2339 if (CI2->getValue().ult(Width)) {
2340 Lower = IntMin.ashr(CI2->getValue());
2341 Upper = IntMax.ashr(CI2->getValue()) + 1;
2343 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) {
2344 unsigned ShiftAmount = Width - 1;
2345 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2346 ShiftAmount = CI2->getValue().countTrailingZeros();
2347 if (CI2->isNegative()) {
2348 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)]
2349 Lower = CI2->getValue();
2350 Upper = CI2->getValue().ashr(ShiftAmount) + 1;
2352 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2]
2353 Lower = CI2->getValue().ashr(ShiftAmount);
2354 Upper = CI2->getValue() + 1;
2356 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
2357 // 'or x, CI2' produces [CI2, UINT_MAX].
2358 Lower = CI2->getValue();
2359 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
2360 // 'and x, CI2' produces [0, CI2].
2361 Upper = CI2->getValue() + 1;
2363 if (Lower != Upper) {
2364 ConstantRange LHS_CR = ConstantRange(Lower, Upper);
2365 if (RHS_CR.contains(LHS_CR))
2366 return ConstantInt::getTrue(RHS->getContext());
2367 if (RHS_CR.inverse().contains(LHS_CR))
2368 return ConstantInt::getFalse(RHS->getContext());
2372 // Compare of cast, for example (zext X) != 0 -> X != 0
2373 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
2374 Instruction *LI = cast<CastInst>(LHS);
2375 Value *SrcOp = LI->getOperand(0);
2376 Type *SrcTy = SrcOp->getType();
2377 Type *DstTy = LI->getType();
2379 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
2380 // if the integer type is the same size as the pointer type.
2381 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
2382 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
2383 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2384 // Transfer the cast to the constant.
2385 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
2386 ConstantExpr::getIntToPtr(RHSC, SrcTy),
2389 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
2390 if (RI->getOperand(0)->getType() == SrcTy)
2391 // Compare without the cast.
2392 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2398 if (isa<ZExtInst>(LHS)) {
2399 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
2401 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
2402 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2403 // Compare X and Y. Note that signed predicates become unsigned.
2404 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2405 SrcOp, RI->getOperand(0), Q,
2409 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
2410 // too. If not, then try to deduce the result of the comparison.
2411 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2412 // Compute the constant that would happen if we truncated to SrcTy then
2413 // reextended to DstTy.
2414 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2415 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
2417 // If the re-extended constant didn't change then this is effectively
2418 // also a case of comparing two zero-extended values.
2419 if (RExt == CI && MaxRecurse)
2420 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2421 SrcOp, Trunc, Q, MaxRecurse-1))
2424 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
2425 // there. Use this to work out the result of the comparison.
2428 default: llvm_unreachable("Unknown ICmp predicate!");
2430 case ICmpInst::ICMP_EQ:
2431 case ICmpInst::ICMP_UGT:
2432 case ICmpInst::ICMP_UGE:
2433 return ConstantInt::getFalse(CI->getContext());
2435 case ICmpInst::ICMP_NE:
2436 case ICmpInst::ICMP_ULT:
2437 case ICmpInst::ICMP_ULE:
2438 return ConstantInt::getTrue(CI->getContext());
2440 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
2441 // is non-negative then LHS <s RHS.
2442 case ICmpInst::ICMP_SGT:
2443 case ICmpInst::ICMP_SGE:
2444 return CI->getValue().isNegative() ?
2445 ConstantInt::getTrue(CI->getContext()) :
2446 ConstantInt::getFalse(CI->getContext());
2448 case ICmpInst::ICMP_SLT:
2449 case ICmpInst::ICMP_SLE:
2450 return CI->getValue().isNegative() ?
2451 ConstantInt::getFalse(CI->getContext()) :
2452 ConstantInt::getTrue(CI->getContext());
2458 if (isa<SExtInst>(LHS)) {
2459 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
2461 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
2462 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2463 // Compare X and Y. Note that the predicate does not change.
2464 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2468 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
2469 // too. If not, then try to deduce the result of the comparison.
2470 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2471 // Compute the constant that would happen if we truncated to SrcTy then
2472 // reextended to DstTy.
2473 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2474 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
2476 // If the re-extended constant didn't change then this is effectively
2477 // also a case of comparing two sign-extended values.
2478 if (RExt == CI && MaxRecurse)
2479 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
2482 // Otherwise the upper bits of LHS are all equal, while RHS has varying
2483 // bits there. Use this to work out the result of the comparison.
2486 default: llvm_unreachable("Unknown ICmp predicate!");
2487 case ICmpInst::ICMP_EQ:
2488 return ConstantInt::getFalse(CI->getContext());
2489 case ICmpInst::ICMP_NE:
2490 return ConstantInt::getTrue(CI->getContext());
2492 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
2494 case ICmpInst::ICMP_SGT:
2495 case ICmpInst::ICMP_SGE:
2496 return CI->getValue().isNegative() ?
2497 ConstantInt::getTrue(CI->getContext()) :
2498 ConstantInt::getFalse(CI->getContext());
2499 case ICmpInst::ICMP_SLT:
2500 case ICmpInst::ICMP_SLE:
2501 return CI->getValue().isNegative() ?
2502 ConstantInt::getFalse(CI->getContext()) :
2503 ConstantInt::getTrue(CI->getContext());
2505 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2507 case ICmpInst::ICMP_UGT:
2508 case ICmpInst::ICMP_UGE:
2509 // Comparison is true iff the LHS <s 0.
2511 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2512 Constant::getNullValue(SrcTy),
2516 case ICmpInst::ICMP_ULT:
2517 case ICmpInst::ICMP_ULE:
2518 // Comparison is true iff the LHS >=s 0.
2520 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2521 Constant::getNullValue(SrcTy),
2531 // Special logic for binary operators.
2532 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2533 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2534 if (MaxRecurse && (LBO || RBO)) {
2535 // Analyze the case when either LHS or RHS is an add instruction.
2536 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2537 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2538 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2539 if (LBO && LBO->getOpcode() == Instruction::Add) {
2540 A = LBO->getOperand(0); B = LBO->getOperand(1);
2541 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2542 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2543 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2545 if (RBO && RBO->getOpcode() == Instruction::Add) {
2546 C = RBO->getOperand(0); D = RBO->getOperand(1);
2547 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2548 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2549 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2552 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2553 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2554 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2555 Constant::getNullValue(RHS->getType()),
2559 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2560 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2561 if (Value *V = SimplifyICmpInst(Pred,
2562 Constant::getNullValue(LHS->getType()),
2563 C == LHS ? D : C, Q, MaxRecurse-1))
2566 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2567 if (A && C && (A == C || A == D || B == C || B == D) &&
2568 NoLHSWrapProblem && NoRHSWrapProblem) {
2569 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2572 // C + B == C + D -> B == D
2575 } else if (A == D) {
2576 // D + B == C + D -> B == C
2579 } else if (B == C) {
2580 // A + C == C + D -> A == D
2585 // A + D == C + D -> A == C
2589 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2594 // icmp pred (or X, Y), X
2595 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)),
2596 m_Or(m_Specific(RHS), m_Value())))) {
2597 if (Pred == ICmpInst::ICMP_ULT)
2598 return getFalse(ITy);
2599 if (Pred == ICmpInst::ICMP_UGE)
2600 return getTrue(ITy);
2602 // icmp pred X, (or X, Y)
2603 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)),
2604 m_Or(m_Specific(LHS), m_Value())))) {
2605 if (Pred == ICmpInst::ICMP_ULE)
2606 return getTrue(ITy);
2607 if (Pred == ICmpInst::ICMP_UGT)
2608 return getFalse(ITy);
2611 // icmp pred (and X, Y), X
2612 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)),
2613 m_And(m_Specific(RHS), m_Value())))) {
2614 if (Pred == ICmpInst::ICMP_UGT)
2615 return getFalse(ITy);
2616 if (Pred == ICmpInst::ICMP_ULE)
2617 return getTrue(ITy);
2619 // icmp pred X, (and X, Y)
2620 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)),
2621 m_And(m_Specific(LHS), m_Value())))) {
2622 if (Pred == ICmpInst::ICMP_UGE)
2623 return getTrue(ITy);
2624 if (Pred == ICmpInst::ICMP_ULT)
2625 return getFalse(ITy);
2628 // 0 - (zext X) pred C
2629 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2630 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2631 if (RHSC->getValue().isStrictlyPositive()) {
2632 if (Pred == ICmpInst::ICMP_SLT)
2633 return ConstantInt::getTrue(RHSC->getContext());
2634 if (Pred == ICmpInst::ICMP_SGE)
2635 return ConstantInt::getFalse(RHSC->getContext());
2636 if (Pred == ICmpInst::ICMP_EQ)
2637 return ConstantInt::getFalse(RHSC->getContext());
2638 if (Pred == ICmpInst::ICMP_NE)
2639 return ConstantInt::getTrue(RHSC->getContext());
2641 if (RHSC->getValue().isNonNegative()) {
2642 if (Pred == ICmpInst::ICMP_SLE)
2643 return ConstantInt::getTrue(RHSC->getContext());
2644 if (Pred == ICmpInst::ICMP_SGT)
2645 return ConstantInt::getFalse(RHSC->getContext());
2650 // icmp pred (urem X, Y), Y
2651 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2652 bool KnownNonNegative, KnownNegative;
2656 case ICmpInst::ICMP_SGT:
2657 case ICmpInst::ICMP_SGE:
2658 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2660 if (!KnownNonNegative)
2663 case ICmpInst::ICMP_EQ:
2664 case ICmpInst::ICMP_UGT:
2665 case ICmpInst::ICMP_UGE:
2666 return getFalse(ITy);
2667 case ICmpInst::ICMP_SLT:
2668 case ICmpInst::ICMP_SLE:
2669 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2671 if (!KnownNonNegative)
2674 case ICmpInst::ICMP_NE:
2675 case ICmpInst::ICMP_ULT:
2676 case ICmpInst::ICMP_ULE:
2677 return getTrue(ITy);
2681 // icmp pred X, (urem Y, X)
2682 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2683 bool KnownNonNegative, KnownNegative;
2687 case ICmpInst::ICMP_SGT:
2688 case ICmpInst::ICMP_SGE:
2689 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2691 if (!KnownNonNegative)
2694 case ICmpInst::ICMP_NE:
2695 case ICmpInst::ICMP_UGT:
2696 case ICmpInst::ICMP_UGE:
2697 return getTrue(ITy);
2698 case ICmpInst::ICMP_SLT:
2699 case ICmpInst::ICMP_SLE:
2700 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2702 if (!KnownNonNegative)
2705 case ICmpInst::ICMP_EQ:
2706 case ICmpInst::ICMP_ULT:
2707 case ICmpInst::ICMP_ULE:
2708 return getFalse(ITy);
2713 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2714 // icmp pred (X /u Y), X
2715 if (Pred == ICmpInst::ICMP_UGT)
2716 return getFalse(ITy);
2717 if (Pred == ICmpInst::ICMP_ULE)
2718 return getTrue(ITy);
2725 // where CI2 is a power of 2 and CI isn't
2726 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2727 const APInt *CI2Val, *CIVal = &CI->getValue();
2728 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2729 CI2Val->isPowerOf2()) {
2730 if (!CIVal->isPowerOf2()) {
2731 // CI2 << X can equal zero in some circumstances,
2732 // this simplification is unsafe if CI is zero.
2734 // We know it is safe if:
2735 // - The shift is nsw, we can't shift out the one bit.
2736 // - The shift is nuw, we can't shift out the one bit.
2739 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2740 *CI2Val == 1 || !CI->isZero()) {
2741 if (Pred == ICmpInst::ICMP_EQ)
2742 return ConstantInt::getFalse(RHS->getContext());
2743 if (Pred == ICmpInst::ICMP_NE)
2744 return ConstantInt::getTrue(RHS->getContext());
2747 if (CIVal->isSignBit() && *CI2Val == 1) {
2748 if (Pred == ICmpInst::ICMP_UGT)
2749 return ConstantInt::getFalse(RHS->getContext());
2750 if (Pred == ICmpInst::ICMP_ULE)
2751 return ConstantInt::getTrue(RHS->getContext());
2756 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2757 LBO->getOperand(1) == RBO->getOperand(1)) {
2758 switch (LBO->getOpcode()) {
2760 case Instruction::UDiv:
2761 case Instruction::LShr:
2762 if (ICmpInst::isSigned(Pred))
2765 case Instruction::SDiv:
2766 case Instruction::AShr:
2767 if (!LBO->isExact() || !RBO->isExact())
2769 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2770 RBO->getOperand(0), Q, MaxRecurse-1))
2773 case Instruction::Shl: {
2774 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2775 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2778 if (!NSW && ICmpInst::isSigned(Pred))
2780 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2781 RBO->getOperand(0), Q, MaxRecurse-1))
2788 // Simplify comparisons involving max/min.
2790 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2791 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2793 // Signed variants on "max(a,b)>=a -> true".
2794 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2795 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2796 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2797 // We analyze this as smax(A, B) pred A.
2799 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2800 (A == LHS || B == LHS)) {
2801 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2802 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2803 // We analyze this as smax(A, B) swapped-pred A.
2804 P = CmpInst::getSwappedPredicate(Pred);
2805 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2806 (A == RHS || B == RHS)) {
2807 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2808 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2809 // We analyze this as smax(-A, -B) swapped-pred -A.
2810 // Note that we do not need to actually form -A or -B thanks to EqP.
2811 P = CmpInst::getSwappedPredicate(Pred);
2812 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2813 (A == LHS || B == LHS)) {
2814 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2815 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2816 // We analyze this as smax(-A, -B) pred -A.
2817 // Note that we do not need to actually form -A or -B thanks to EqP.
2820 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2821 // Cases correspond to "max(A, B) p A".
2825 case CmpInst::ICMP_EQ:
2826 case CmpInst::ICMP_SLE:
2827 // Equivalent to "A EqP B". This may be the same as the condition tested
2828 // in the max/min; if so, we can just return that.
2829 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2831 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2833 // Otherwise, see if "A EqP B" simplifies.
2835 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2838 case CmpInst::ICMP_NE:
2839 case CmpInst::ICMP_SGT: {
2840 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2841 // Equivalent to "A InvEqP B". This may be the same as the condition
2842 // tested in the max/min; if so, we can just return that.
2843 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2845 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2847 // Otherwise, see if "A InvEqP B" simplifies.
2849 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2853 case CmpInst::ICMP_SGE:
2855 return getTrue(ITy);
2856 case CmpInst::ICMP_SLT:
2858 return getFalse(ITy);
2862 // Unsigned variants on "max(a,b)>=a -> true".
2863 P = CmpInst::BAD_ICMP_PREDICATE;
2864 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2865 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2866 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2867 // We analyze this as umax(A, B) pred A.
2869 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2870 (A == LHS || B == LHS)) {
2871 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2872 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2873 // We analyze this as umax(A, B) swapped-pred A.
2874 P = CmpInst::getSwappedPredicate(Pred);
2875 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2876 (A == RHS || B == RHS)) {
2877 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2878 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2879 // We analyze this as umax(-A, -B) swapped-pred -A.
2880 // Note that we do not need to actually form -A or -B thanks to EqP.
2881 P = CmpInst::getSwappedPredicate(Pred);
2882 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2883 (A == LHS || B == LHS)) {
2884 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2885 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2886 // We analyze this as umax(-A, -B) pred -A.
2887 // Note that we do not need to actually form -A or -B thanks to EqP.
2890 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2891 // Cases correspond to "max(A, B) p A".
2895 case CmpInst::ICMP_EQ:
2896 case CmpInst::ICMP_ULE:
2897 // Equivalent to "A EqP B". This may be the same as the condition tested
2898 // in the max/min; if so, we can just return that.
2899 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2901 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2903 // Otherwise, see if "A EqP B" simplifies.
2905 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2908 case CmpInst::ICMP_NE:
2909 case CmpInst::ICMP_UGT: {
2910 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2911 // Equivalent to "A InvEqP B". This may be the same as the condition
2912 // tested in the max/min; if so, we can just return that.
2913 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2915 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2917 // Otherwise, see if "A InvEqP B" simplifies.
2919 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2923 case CmpInst::ICMP_UGE:
2925 return getTrue(ITy);
2926 case CmpInst::ICMP_ULT:
2928 return getFalse(ITy);
2932 // Variants on "max(x,y) >= min(x,z)".
2934 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2935 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2936 (A == C || A == D || B == C || B == D)) {
2937 // max(x, ?) pred min(x, ?).
2938 if (Pred == CmpInst::ICMP_SGE)
2940 return getTrue(ITy);
2941 if (Pred == CmpInst::ICMP_SLT)
2943 return getFalse(ITy);
2944 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2945 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2946 (A == C || A == D || B == C || B == D)) {
2947 // min(x, ?) pred max(x, ?).
2948 if (Pred == CmpInst::ICMP_SLE)
2950 return getTrue(ITy);
2951 if (Pred == CmpInst::ICMP_SGT)
2953 return getFalse(ITy);
2954 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2955 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2956 (A == C || A == D || B == C || B == D)) {
2957 // max(x, ?) pred min(x, ?).
2958 if (Pred == CmpInst::ICMP_UGE)
2960 return getTrue(ITy);
2961 if (Pred == CmpInst::ICMP_ULT)
2963 return getFalse(ITy);
2964 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2965 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2966 (A == C || A == D || B == C || B == D)) {
2967 // min(x, ?) pred max(x, ?).
2968 if (Pred == CmpInst::ICMP_ULE)
2970 return getTrue(ITy);
2971 if (Pred == CmpInst::ICMP_UGT)
2973 return getFalse(ITy);
2976 // Simplify comparisons of related pointers using a powerful, recursive
2977 // GEP-walk when we have target data available..
2978 if (LHS->getType()->isPointerTy())
2979 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS))
2982 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
2983 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
2984 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
2985 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
2986 (ICmpInst::isEquality(Pred) ||
2987 (GLHS->isInBounds() && GRHS->isInBounds() &&
2988 Pred == ICmpInst::getSignedPredicate(Pred)))) {
2989 // The bases are equal and the indices are constant. Build a constant
2990 // expression GEP with the same indices and a null base pointer to see
2991 // what constant folding can make out of it.
2992 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
2993 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
2994 Constant *NewLHS = ConstantExpr::getGetElementPtr(
2995 GLHS->getSourceElementType(), Null, IndicesLHS);
2997 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
2998 Constant *NewRHS = ConstantExpr::getGetElementPtr(
2999 GLHS->getSourceElementType(), Null, IndicesRHS);
3000 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3005 // If a bit is known to be zero for A and known to be one for B,
3006 // then A and B cannot be equal.
3007 if (ICmpInst::isEquality(Pred)) {
3008 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3009 uint32_t BitWidth = CI->getBitWidth();
3010 APInt LHSKnownZero(BitWidth, 0);
3011 APInt LHSKnownOne(BitWidth, 0);
3012 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC,
3014 const APInt &RHSVal = CI->getValue();
3015 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0))
3016 return Pred == ICmpInst::ICMP_EQ
3017 ? ConstantInt::getFalse(CI->getContext())
3018 : ConstantInt::getTrue(CI->getContext());
3022 // If the comparison is with the result of a select instruction, check whether
3023 // comparing with either branch of the select always yields the same value.
3024 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3025 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3028 // If the comparison is with the result of a phi instruction, check whether
3029 // doing the compare with each incoming phi value yields a common result.
3030 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3031 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3037 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3038 const DataLayout &DL,
3039 const TargetLibraryInfo *TLI,
3040 const DominatorTree *DT, AssumptionCache *AC,
3041 Instruction *CxtI) {
3042 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3046 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
3047 /// fold the result. If not, this returns null.
3048 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3049 const Query &Q, unsigned MaxRecurse) {
3050 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3051 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3053 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3054 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3055 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3057 // If we have a constant, make sure it is on the RHS.
3058 std::swap(LHS, RHS);
3059 Pred = CmpInst::getSwappedPredicate(Pred);
3062 // Fold trivial predicates.
3063 if (Pred == FCmpInst::FCMP_FALSE)
3064 return ConstantInt::get(GetCompareTy(LHS), 0);
3065 if (Pred == FCmpInst::FCMP_TRUE)
3066 return ConstantInt::get(GetCompareTy(LHS), 1);
3068 // fcmp pred x, undef and fcmp pred undef, x
3069 // fold to true if unordered, false if ordered
3070 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3071 // Choosing NaN for the undef will always make unordered comparison succeed
3072 // and ordered comparison fail.
3073 return ConstantInt::get(GetCompareTy(LHS), CmpInst::isUnordered(Pred));
3076 // fcmp x,x -> true/false. Not all compares are foldable.
3078 if (CmpInst::isTrueWhenEqual(Pred))
3079 return ConstantInt::get(GetCompareTy(LHS), 1);
3080 if (CmpInst::isFalseWhenEqual(Pred))
3081 return ConstantInt::get(GetCompareTy(LHS), 0);
3084 // Handle fcmp with constant RHS
3085 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
3086 // If the constant is a nan, see if we can fold the comparison based on it.
3087 if (CFP->getValueAPF().isNaN()) {
3088 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3089 return ConstantInt::getFalse(CFP->getContext());
3090 assert(FCmpInst::isUnordered(Pred) &&
3091 "Comparison must be either ordered or unordered!");
3092 // True if unordered.
3093 return ConstantInt::getTrue(CFP->getContext());
3095 // Check whether the constant is an infinity.
3096 if (CFP->getValueAPF().isInfinity()) {
3097 if (CFP->getValueAPF().isNegative()) {
3099 case FCmpInst::FCMP_OLT:
3100 // No value is ordered and less than negative infinity.
3101 return ConstantInt::getFalse(CFP->getContext());
3102 case FCmpInst::FCMP_UGE:
3103 // All values are unordered with or at least negative infinity.
3104 return ConstantInt::getTrue(CFP->getContext());
3110 case FCmpInst::FCMP_OGT:
3111 // No value is ordered and greater than infinity.
3112 return ConstantInt::getFalse(CFP->getContext());
3113 case FCmpInst::FCMP_ULE:
3114 // All values are unordered with and at most infinity.
3115 return ConstantInt::getTrue(CFP->getContext());
3121 if (CFP->getValueAPF().isZero()) {
3123 case FCmpInst::FCMP_UGE:
3124 if (CannotBeOrderedLessThanZero(LHS))
3125 return ConstantInt::getTrue(CFP->getContext());
3127 case FCmpInst::FCMP_OLT:
3129 if (CannotBeOrderedLessThanZero(LHS))
3130 return ConstantInt::getFalse(CFP->getContext());
3138 // If the comparison is with the result of a select instruction, check whether
3139 // comparing with either branch of the select always yields the same value.
3140 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3141 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3144 // If the comparison is with the result of a phi instruction, check whether
3145 // doing the compare with each incoming phi value yields a common result.
3146 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3147 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3153 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3154 const DataLayout &DL,
3155 const TargetLibraryInfo *TLI,
3156 const DominatorTree *DT, AssumptionCache *AC,
3157 const Instruction *CxtI) {
3158 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3162 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is
3163 /// replaced with RepOp.
3164 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3166 unsigned MaxRecurse) {
3167 // Trivial replacement.
3171 auto *I = dyn_cast<Instruction>(V);
3175 // If this is a binary operator, try to simplify it with the replaced op.
3176 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3178 // %cmp = icmp eq i32 %x, 2147483647
3179 // %add = add nsw i32 %x, 1
3180 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3182 // We can't replace %sel with %add unless we strip away the flags.
3183 if (isa<OverflowingBinaryOperator>(B))
3184 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3186 if (isa<PossiblyExactOperator>(B))
3191 if (B->getOperand(0) == Op)
3192 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3194 if (B->getOperand(1) == Op)
3195 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3200 // Same for CmpInsts.
3201 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3203 if (C->getOperand(0) == Op)
3204 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3206 if (C->getOperand(1) == Op)
3207 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3212 // TODO: We could hand off more cases to instsimplify here.
3214 // If all operands are constant after substituting Op for RepOp then we can
3215 // constant fold the instruction.
3216 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3217 // Build a list of all constant operands.
3218 SmallVector<Constant *, 8> ConstOps;
3219 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3220 if (I->getOperand(i) == Op)
3221 ConstOps.push_back(CRepOp);
3222 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3223 ConstOps.push_back(COp);
3228 // All operands were constants, fold it.
3229 if (ConstOps.size() == I->getNumOperands()) {
3230 if (CmpInst *C = dyn_cast<CmpInst>(I))
3231 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3232 ConstOps[1], Q.DL, Q.TLI);
3234 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3235 if (!LI->isVolatile())
3236 return ConstantFoldLoadFromConstPtr(ConstOps[0], Q.DL);
3238 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), ConstOps,
3246 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
3247 /// the result. If not, this returns null.
3248 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3249 Value *FalseVal, const Query &Q,
3250 unsigned MaxRecurse) {
3251 // select true, X, Y -> X
3252 // select false, X, Y -> Y
3253 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3254 if (CB->isAllOnesValue())
3256 if (CB->isNullValue())
3260 // select C, X, X -> X
3261 if (TrueVal == FalseVal)
3264 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3265 if (isa<Constant>(TrueVal))
3269 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3271 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3274 if (const auto *ICI = dyn_cast<ICmpInst>(CondVal)) {
3275 unsigned BitWidth = Q.DL.getTypeSizeInBits(TrueVal->getType());
3276 ICmpInst::Predicate Pred = ICI->getPredicate();
3277 Value *CmpLHS = ICI->getOperand(0);
3278 Value *CmpRHS = ICI->getOperand(1);
3279 APInt MinSignedValue = APInt::getSignBit(BitWidth);
3283 bool IsBitTest = false;
3284 if (ICmpInst::isEquality(Pred) &&
3285 match(CmpLHS, m_And(m_Value(X), m_APInt(Y))) &&
3286 match(CmpRHS, m_Zero())) {
3288 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
3289 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3291 Y = &MinSignedValue;
3293 TrueWhenUnset = false;
3294 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3296 Y = &MinSignedValue;
3298 TrueWhenUnset = true;
3302 // (X & Y) == 0 ? X & ~Y : X --> X
3303 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3304 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3306 return TrueWhenUnset ? FalseVal : TrueVal;
3307 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3308 // (X & Y) != 0 ? X : X & ~Y --> X
3309 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3311 return TrueWhenUnset ? FalseVal : TrueVal;
3313 if (Y->isPowerOf2()) {
3314 // (X & Y) == 0 ? X | Y : X --> X | Y
3315 // (X & Y) != 0 ? X | Y : X --> X
3316 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3318 return TrueWhenUnset ? TrueVal : FalseVal;
3319 // (X & Y) == 0 ? X : X | Y --> X
3320 // (X & Y) != 0 ? X : X | Y --> X | Y
3321 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3323 return TrueWhenUnset ? TrueVal : FalseVal;
3326 if (ICI->hasOneUse()) {
3328 if (match(CmpRHS, m_APInt(C))) {
3329 // X < MIN ? T : F --> F
3330 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3332 // X < MIN ? T : F --> F
3333 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3335 // X > MAX ? T : F --> F
3336 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3338 // X > MAX ? T : F --> F
3339 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3344 // If we have an equality comparison then we know the value in one of the
3345 // arms of the select. See if substituting this value into the arm and
3346 // simplifying the result yields the same value as the other arm.
3347 if (Pred == ICmpInst::ICMP_EQ) {
3348 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3350 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3353 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3355 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3358 } else if (Pred == ICmpInst::ICMP_NE) {
3359 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3361 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3364 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3366 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3375 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3376 const DataLayout &DL,
3377 const TargetLibraryInfo *TLI,
3378 const DominatorTree *DT, AssumptionCache *AC,
3379 const Instruction *CxtI) {
3380 return ::SimplifySelectInst(Cond, TrueVal, FalseVal,
3381 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3384 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
3385 /// fold the result. If not, this returns null.
3386 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3387 const Query &Q, unsigned) {
3388 // The type of the GEP pointer operand.
3390 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3392 // getelementptr P -> P.
3393 if (Ops.size() == 1)
3396 // Compute the (pointer) type returned by the GEP instruction.
3397 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3398 Type *GEPTy = PointerType::get(LastType, AS);
3399 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3400 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3402 if (isa<UndefValue>(Ops[0]))
3403 return UndefValue::get(GEPTy);
3405 if (Ops.size() == 2) {
3406 // getelementptr P, 0 -> P.
3407 if (match(Ops[1], m_Zero()))
3411 if (Ty->isSized()) {
3414 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3415 // getelementptr P, N -> P if P points to a type of zero size.
3416 if (TyAllocSize == 0)
3419 // The following transforms are only safe if the ptrtoint cast
3420 // doesn't truncate the pointers.
3421 if (Ops[1]->getType()->getScalarSizeInBits() ==
3422 Q.DL.getPointerSizeInBits(AS)) {
3423 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3424 if (match(P, m_Zero()))
3425 return Constant::getNullValue(GEPTy);
3427 if (match(P, m_PtrToInt(m_Value(Temp))))
3428 if (Temp->getType() == GEPTy)
3433 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3434 if (TyAllocSize == 1 &&
3435 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3436 if (Value *R = PtrToIntOrZero(P))
3439 // getelementptr V, (ashr (sub P, V), C) -> Q
3440 // if P points to a type of size 1 << C.
3442 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3443 m_ConstantInt(C))) &&
3444 TyAllocSize == 1ULL << C)
3445 if (Value *R = PtrToIntOrZero(P))
3448 // getelementptr V, (sdiv (sub P, V), C) -> Q
3449 // if P points to a type of size C.
3451 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3452 m_SpecificInt(TyAllocSize))))
3453 if (Value *R = PtrToIntOrZero(P))
3459 // Check to see if this is constant foldable.
3460 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
3461 if (!isa<Constant>(Ops[i]))
3464 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3468 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout &DL,
3469 const TargetLibraryInfo *TLI,
3470 const DominatorTree *DT, AssumptionCache *AC,
3471 const Instruction *CxtI) {
3472 return ::SimplifyGEPInst(
3473 cast<PointerType>(Ops[0]->getType()->getScalarType())->getElementType(),
3474 Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3477 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
3478 /// can fold the result. If not, this returns null.
3479 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3480 ArrayRef<unsigned> Idxs, const Query &Q,
3482 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3483 if (Constant *CVal = dyn_cast<Constant>(Val))
3484 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3486 // insertvalue x, undef, n -> x
3487 if (match(Val, m_Undef()))
3490 // insertvalue x, (extractvalue y, n), n
3491 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3492 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3493 EV->getIndices() == Idxs) {
3494 // insertvalue undef, (extractvalue y, n), n -> y
3495 if (match(Agg, m_Undef()))
3496 return EV->getAggregateOperand();
3498 // insertvalue y, (extractvalue y, n), n -> y
3499 if (Agg == EV->getAggregateOperand())
3506 Value *llvm::SimplifyInsertValueInst(
3507 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL,
3508 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC,
3509 const Instruction *CxtI) {
3510 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI),
3514 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
3515 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
3516 // If all of the PHI's incoming values are the same then replace the PHI node
3517 // with the common value.
3518 Value *CommonValue = nullptr;
3519 bool HasUndefInput = false;
3520 for (Value *Incoming : PN->incoming_values()) {
3521 // If the incoming value is the phi node itself, it can safely be skipped.
3522 if (Incoming == PN) continue;
3523 if (isa<UndefValue>(Incoming)) {
3524 // Remember that we saw an undef value, but otherwise ignore them.
3525 HasUndefInput = true;
3528 if (CommonValue && Incoming != CommonValue)
3529 return nullptr; // Not the same, bail out.
3530 CommonValue = Incoming;
3533 // If CommonValue is null then all of the incoming values were either undef or
3534 // equal to the phi node itself.
3536 return UndefValue::get(PN->getType());
3538 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3539 // instruction, we cannot return X as the result of the PHI node unless it
3540 // dominates the PHI block.
3542 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3547 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
3548 if (Constant *C = dyn_cast<Constant>(Op))
3549 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI);
3554 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout &DL,
3555 const TargetLibraryInfo *TLI,
3556 const DominatorTree *DT, AssumptionCache *AC,
3557 const Instruction *CxtI) {
3558 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI),
3562 //=== Helper functions for higher up the class hierarchy.
3564 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
3565 /// fold the result. If not, this returns null.
3566 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3567 const Query &Q, unsigned MaxRecurse) {
3569 case Instruction::Add:
3570 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3572 case Instruction::FAdd:
3573 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3575 case Instruction::Sub:
3576 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3578 case Instruction::FSub:
3579 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3581 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
3582 case Instruction::FMul:
3583 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3584 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
3585 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
3586 case Instruction::FDiv:
3587 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3588 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
3589 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
3590 case Instruction::FRem:
3591 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3592 case Instruction::Shl:
3593 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3595 case Instruction::LShr:
3596 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3597 case Instruction::AShr:
3598 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3599 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
3600 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
3601 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
3603 if (Constant *CLHS = dyn_cast<Constant>(LHS))
3604 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
3605 Constant *COps[] = {CLHS, CRHS};
3606 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL,
3610 // If the operation is associative, try some generic simplifications.
3611 if (Instruction::isAssociative(Opcode))
3612 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
3615 // If the operation is with the result of a select instruction check whether
3616 // operating on either branch of the select always yields the same value.
3617 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3618 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
3621 // If the operation is with the result of a phi instruction, check whether
3622 // operating on all incoming values of the phi always yields the same value.
3623 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3624 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
3631 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can
3632 /// fold the result. If not, this returns null.
3633 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
3634 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
3635 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3636 const FastMathFlags &FMF, const Query &Q,
3637 unsigned MaxRecurse) {
3639 case Instruction::FAdd:
3640 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
3641 case Instruction::FSub:
3642 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
3643 case Instruction::FMul:
3644 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
3646 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
3650 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3651 const DataLayout &DL, const TargetLibraryInfo *TLI,
3652 const DominatorTree *DT, AssumptionCache *AC,
3653 const Instruction *CxtI) {
3654 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3658 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3659 const FastMathFlags &FMF, const DataLayout &DL,
3660 const TargetLibraryInfo *TLI,
3661 const DominatorTree *DT, AssumptionCache *AC,
3662 const Instruction *CxtI) {
3663 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI),
3667 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
3668 /// fold the result.
3669 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3670 const Query &Q, unsigned MaxRecurse) {
3671 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
3672 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3673 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3676 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3677 const DataLayout &DL, const TargetLibraryInfo *TLI,
3678 const DominatorTree *DT, AssumptionCache *AC,
3679 const Instruction *CxtI) {
3680 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3684 static bool IsIdempotent(Intrinsic::ID ID) {
3686 default: return false;
3688 // Unary idempotent: f(f(x)) = f(x)
3689 case Intrinsic::fabs:
3690 case Intrinsic::floor:
3691 case Intrinsic::ceil:
3692 case Intrinsic::trunc:
3693 case Intrinsic::rint:
3694 case Intrinsic::nearbyint:
3695 case Intrinsic::round:
3700 template <typename IterTy>
3701 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
3702 const Query &Q, unsigned MaxRecurse) {
3703 Intrinsic::ID IID = F->getIntrinsicID();
3704 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
3705 Type *ReturnType = F->getReturnType();
3708 if (NumOperands == 2) {
3709 Value *LHS = *ArgBegin;
3710 Value *RHS = *(ArgBegin + 1);
3711 if (IID == Intrinsic::usub_with_overflow ||
3712 IID == Intrinsic::ssub_with_overflow) {
3713 // X - X -> { 0, false }
3715 return Constant::getNullValue(ReturnType);
3717 // X - undef -> undef
3718 // undef - X -> undef
3719 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
3720 return UndefValue::get(ReturnType);
3723 if (IID == Intrinsic::uadd_with_overflow ||
3724 IID == Intrinsic::sadd_with_overflow) {
3725 // X + undef -> undef
3726 if (isa<UndefValue>(RHS))
3727 return UndefValue::get(ReturnType);
3730 if (IID == Intrinsic::umul_with_overflow ||
3731 IID == Intrinsic::smul_with_overflow) {
3732 // X * 0 -> { 0, false }
3733 if (match(RHS, m_Zero()))
3734 return Constant::getNullValue(ReturnType);
3736 // X * undef -> { 0, false }
3737 if (match(RHS, m_Undef()))
3738 return Constant::getNullValue(ReturnType);
3742 // Perform idempotent optimizations
3743 if (!IsIdempotent(IID))
3747 if (NumOperands == 1)
3748 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
3749 if (II->getIntrinsicID() == IID)
3755 template <typename IterTy>
3756 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
3757 const Query &Q, unsigned MaxRecurse) {
3758 Type *Ty = V->getType();
3759 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
3760 Ty = PTy->getElementType();
3761 FunctionType *FTy = cast<FunctionType>(Ty);
3763 // call undef -> undef
3764 if (isa<UndefValue>(V))
3765 return UndefValue::get(FTy->getReturnType());
3767 Function *F = dyn_cast<Function>(V);
3771 if (F->isIntrinsic())
3772 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
3775 if (!canConstantFoldCallTo(F))
3778 SmallVector<Constant *, 4> ConstantArgs;
3779 ConstantArgs.reserve(ArgEnd - ArgBegin);
3780 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
3781 Constant *C = dyn_cast<Constant>(*I);
3784 ConstantArgs.push_back(C);
3787 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
3790 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
3791 User::op_iterator ArgEnd, const DataLayout &DL,
3792 const TargetLibraryInfo *TLI, const DominatorTree *DT,
3793 AssumptionCache *AC, const Instruction *CxtI) {
3794 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI),
3798 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
3799 const DataLayout &DL, const TargetLibraryInfo *TLI,
3800 const DominatorTree *DT, AssumptionCache *AC,
3801 const Instruction *CxtI) {
3802 return ::SimplifyCall(V, Args.begin(), Args.end(),
3803 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3806 /// SimplifyInstruction - See if we can compute a simplified version of this
3807 /// instruction. If not, this returns null.
3808 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL,
3809 const TargetLibraryInfo *TLI,
3810 const DominatorTree *DT, AssumptionCache *AC) {
3813 switch (I->getOpcode()) {
3815 Result = ConstantFoldInstruction(I, DL, TLI);
3817 case Instruction::FAdd:
3818 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
3819 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3821 case Instruction::Add:
3822 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
3823 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3824 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3827 case Instruction::FSub:
3828 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
3829 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3831 case Instruction::Sub:
3832 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
3833 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3834 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3837 case Instruction::FMul:
3838 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
3839 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3841 case Instruction::Mul:
3843 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3845 case Instruction::SDiv:
3846 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3849 case Instruction::UDiv:
3850 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3853 case Instruction::FDiv:
3854 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
3855 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3857 case Instruction::SRem:
3858 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3861 case Instruction::URem:
3862 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3865 case Instruction::FRem:
3866 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
3867 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3869 case Instruction::Shl:
3870 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
3871 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3872 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3875 case Instruction::LShr:
3876 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
3877 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
3880 case Instruction::AShr:
3881 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
3882 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
3885 case Instruction::And:
3887 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3889 case Instruction::Or:
3891 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3893 case Instruction::Xor:
3895 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3897 case Instruction::ICmp:
3899 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0),
3900 I->getOperand(1), DL, TLI, DT, AC, I);
3902 case Instruction::FCmp:
3904 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
3905 I->getOperand(1), DL, TLI, DT, AC, I);
3907 case Instruction::Select:
3908 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
3909 I->getOperand(2), DL, TLI, DT, AC, I);
3911 case Instruction::GetElementPtr: {
3912 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
3913 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I);
3916 case Instruction::InsertValue: {
3917 InsertValueInst *IV = cast<InsertValueInst>(I);
3918 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
3919 IV->getInsertedValueOperand(),
3920 IV->getIndices(), DL, TLI, DT, AC, I);
3923 case Instruction::PHI:
3924 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I));
3926 case Instruction::Call: {
3927 CallSite CS(cast<CallInst>(I));
3928 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL,
3932 case Instruction::Trunc:
3934 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I);
3938 /// If called on unreachable code, the above logic may report that the
3939 /// instruction simplified to itself. Make life easier for users by
3940 /// detecting that case here, returning a safe value instead.
3941 return Result == I ? UndefValue::get(I->getType()) : Result;
3944 /// \brief Implementation of recursive simplification through an instructions
3947 /// This is the common implementation of the recursive simplification routines.
3948 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
3949 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
3950 /// instructions to process and attempt to simplify it using
3951 /// InstructionSimplify.
3953 /// This routine returns 'true' only when *it* simplifies something. The passed
3954 /// in simplified value does not count toward this.
3955 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
3956 const TargetLibraryInfo *TLI,
3957 const DominatorTree *DT,
3958 AssumptionCache *AC) {
3959 bool Simplified = false;
3960 SmallSetVector<Instruction *, 8> Worklist;
3961 const DataLayout &DL = I->getModule()->getDataLayout();
3963 // If we have an explicit value to collapse to, do that round of the
3964 // simplification loop by hand initially.
3966 for (User *U : I->users())
3968 Worklist.insert(cast<Instruction>(U));
3970 // Replace the instruction with its simplified value.
3971 I->replaceAllUsesWith(SimpleV);
3973 // Gracefully handle edge cases where the instruction is not wired into any
3976 I->eraseFromParent();
3981 // Note that we must test the size on each iteration, the worklist can grow.
3982 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
3985 // See if this instruction simplifies.
3986 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC);
3992 // Stash away all the uses of the old instruction so we can check them for
3993 // recursive simplifications after a RAUW. This is cheaper than checking all
3994 // uses of To on the recursive step in most cases.
3995 for (User *U : I->users())
3996 Worklist.insert(cast<Instruction>(U));
3998 // Replace the instruction with its simplified value.
3999 I->replaceAllUsesWith(SimpleV);
4001 // Gracefully handle edge cases where the instruction is not wired into any
4004 I->eraseFromParent();
4009 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4010 const TargetLibraryInfo *TLI,
4011 const DominatorTree *DT,
4012 AssumptionCache *AC) {
4013 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4016 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4017 const TargetLibraryInfo *TLI,
4018 const DominatorTree *DT,
4019 AssumptionCache *AC) {
4020 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4021 assert(SimpleV && "Must provide a simplified value.");
4022 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);