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/Analysis/VectorUtils.h"
28 #include "llvm/IR/ConstantRange.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/GlobalAlias.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/ValueHandle.h"
38 using namespace llvm::PatternMatch;
40 #define DEBUG_TYPE "instsimplify"
42 enum { RecursionLimit = 3 };
44 STATISTIC(NumExpand, "Number of expansions");
45 STATISTIC(NumReassoc, "Number of reassociations");
50 const TargetLibraryInfo *TLI;
51 const DominatorTree *DT;
53 const Instruction *CxtI;
55 Query(const DataLayout &DL, const TargetLibraryInfo *tli,
56 const DominatorTree *dt, AssumptionCache *ac = nullptr,
57 const Instruction *cxti = nullptr)
58 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {}
60 } // end anonymous namespace
62 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
63 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
65 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
66 const Query &, unsigned);
67 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
69 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
70 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
71 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned);
73 /// getFalse - For a boolean type, or a vector of boolean type, return false, or
74 /// a vector with every element false, as appropriate for the type.
75 static Constant *getFalse(Type *Ty) {
76 assert(Ty->getScalarType()->isIntegerTy(1) &&
77 "Expected i1 type or a vector of i1!");
78 return Constant::getNullValue(Ty);
81 /// getTrue - For a boolean type, or a vector of boolean type, return true, or
82 /// a vector with every element true, as appropriate for the type.
83 static Constant *getTrue(Type *Ty) {
84 assert(Ty->getScalarType()->isIntegerTy(1) &&
85 "Expected i1 type or a vector of i1!");
86 return Constant::getAllOnesValue(Ty);
89 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
90 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
92 CmpInst *Cmp = dyn_cast<CmpInst>(V);
95 CmpInst::Predicate CPred = Cmp->getPredicate();
96 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
97 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
99 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
103 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
104 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
105 Instruction *I = dyn_cast<Instruction>(V);
107 // Arguments and constants dominate all instructions.
110 // If we are processing instructions (and/or basic blocks) that have not been
111 // fully added to a function, the parent nodes may still be null. Simply
112 // return the conservative answer in these cases.
113 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
116 // If we have a DominatorTree then do a precise test.
118 if (!DT->isReachableFromEntry(P->getParent()))
120 if (!DT->isReachableFromEntry(I->getParent()))
122 return DT->dominates(I, P);
125 // Otherwise, if the instruction is in the entry block, and is not an invoke,
126 // and is not a catchpad, then it obviously dominates all phi nodes.
127 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
128 !isa<InvokeInst>(I) && !isa<CatchPadInst>(I))
134 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning
135 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
136 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
137 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
138 /// Returns the simplified value, or null if no simplification was performed.
139 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
140 unsigned OpcToExpand, const Query &Q,
141 unsigned MaxRecurse) {
142 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
143 // Recursion is always used, so bail out at once if we already hit the limit.
147 // Check whether the expression has the form "(A op' B) op C".
148 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
149 if (Op0->getOpcode() == OpcodeToExpand) {
150 // It does! Try turning it into "(A op C) op' (B op C)".
151 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
152 // Do "A op C" and "B op C" both simplify?
153 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
154 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
155 // They do! Return "L op' R" if it simplifies or is already available.
156 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
157 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
158 && L == B && R == A)) {
162 // Otherwise return "L op' R" if it simplifies.
163 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
170 // Check whether the expression has the form "A op (B op' C)".
171 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
172 if (Op1->getOpcode() == OpcodeToExpand) {
173 // It does! Try turning it into "(A op B) op' (A op C)".
174 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
175 // Do "A op B" and "A op C" both simplify?
176 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
177 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
178 // They do! Return "L op' R" if it simplifies or is already available.
179 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
180 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
181 && L == C && R == B)) {
185 // Otherwise return "L op' R" if it simplifies.
186 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
196 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
197 /// operations. Returns the simpler value, or null if none was found.
198 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
199 const Query &Q, unsigned MaxRecurse) {
200 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
201 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
203 // Recursion is always used, so bail out at once if we already hit the limit.
207 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
208 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
210 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
211 if (Op0 && Op0->getOpcode() == Opcode) {
212 Value *A = Op0->getOperand(0);
213 Value *B = Op0->getOperand(1);
216 // Does "B op C" simplify?
217 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
218 // It does! Return "A op V" if it simplifies or is already available.
219 // If V equals B then "A op V" is just the LHS.
220 if (V == B) return LHS;
221 // Otherwise return "A op V" if it simplifies.
222 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
229 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
230 if (Op1 && Op1->getOpcode() == Opcode) {
232 Value *B = Op1->getOperand(0);
233 Value *C = Op1->getOperand(1);
235 // Does "A op B" simplify?
236 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
237 // It does! Return "V op C" if it simplifies or is already available.
238 // If V equals B then "V op C" is just the RHS.
239 if (V == B) return RHS;
240 // Otherwise return "V op C" if it simplifies.
241 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
248 // The remaining transforms require commutativity as well as associativity.
249 if (!Instruction::isCommutative(Opcode))
252 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
253 if (Op0 && Op0->getOpcode() == Opcode) {
254 Value *A = Op0->getOperand(0);
255 Value *B = Op0->getOperand(1);
258 // Does "C op A" simplify?
259 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
260 // It does! Return "V op B" if it simplifies or is already available.
261 // If V equals A then "V op B" is just the LHS.
262 if (V == A) return LHS;
263 // Otherwise return "V op B" if it simplifies.
264 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
271 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
272 if (Op1 && Op1->getOpcode() == Opcode) {
274 Value *B = Op1->getOperand(0);
275 Value *C = Op1->getOperand(1);
277 // Does "C op A" simplify?
278 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
279 // It does! Return "B op V" if it simplifies or is already available.
280 // If V equals C then "B op V" is just the RHS.
281 if (V == C) return RHS;
282 // Otherwise return "B op V" if it simplifies.
283 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
293 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
294 /// instruction as an operand, try to simplify the binop by seeing whether
295 /// evaluating it on both branches of the select results in the same value.
296 /// Returns the common value if so, otherwise returns null.
297 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
298 const Query &Q, unsigned MaxRecurse) {
299 // Recursion is always used, so bail out at once if we already hit the limit.
304 if (isa<SelectInst>(LHS)) {
305 SI = cast<SelectInst>(LHS);
307 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
308 SI = cast<SelectInst>(RHS);
311 // Evaluate the BinOp on the true and false branches of the select.
315 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
316 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
318 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
319 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
322 // If they simplified to the same value, then return the common value.
323 // If they both failed to simplify then return null.
327 // If one branch simplified to undef, return the other one.
328 if (TV && isa<UndefValue>(TV))
330 if (FV && isa<UndefValue>(FV))
333 // If applying the operation did not change the true and false select values,
334 // then the result of the binop is the select itself.
335 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
338 // If one branch simplified and the other did not, and the simplified
339 // value is equal to the unsimplified one, return the simplified value.
340 // For example, select (cond, X, X & Z) & Z -> X & Z.
341 if ((FV && !TV) || (TV && !FV)) {
342 // Check that the simplified value has the form "X op Y" where "op" is the
343 // same as the original operation.
344 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
345 if (Simplified && Simplified->getOpcode() == Opcode) {
346 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
347 // We already know that "op" is the same as for the simplified value. See
348 // if the operands match too. If so, return the simplified value.
349 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
350 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
351 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
352 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
353 Simplified->getOperand(1) == UnsimplifiedRHS)
355 if (Simplified->isCommutative() &&
356 Simplified->getOperand(1) == UnsimplifiedLHS &&
357 Simplified->getOperand(0) == UnsimplifiedRHS)
365 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
366 /// try to simplify the comparison by seeing whether both branches of the select
367 /// result in the same value. Returns the common value if so, otherwise returns
369 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
370 Value *RHS, const Query &Q,
371 unsigned MaxRecurse) {
372 // Recursion is always used, so bail out at once if we already hit the limit.
376 // Make sure the select is on the LHS.
377 if (!isa<SelectInst>(LHS)) {
379 Pred = CmpInst::getSwappedPredicate(Pred);
381 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
382 SelectInst *SI = cast<SelectInst>(LHS);
383 Value *Cond = SI->getCondition();
384 Value *TV = SI->getTrueValue();
385 Value *FV = SI->getFalseValue();
387 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
388 // Does "cmp TV, RHS" simplify?
389 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
391 // It not only simplified, it simplified to the select condition. Replace
393 TCmp = getTrue(Cond->getType());
395 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
396 // condition then we can replace it with 'true'. Otherwise give up.
397 if (!isSameCompare(Cond, Pred, TV, RHS))
399 TCmp = getTrue(Cond->getType());
402 // Does "cmp FV, RHS" simplify?
403 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
405 // It not only simplified, it simplified to the select condition. Replace
407 FCmp = getFalse(Cond->getType());
409 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
410 // condition then we can replace it with 'false'. Otherwise give up.
411 if (!isSameCompare(Cond, Pred, FV, RHS))
413 FCmp = getFalse(Cond->getType());
416 // If both sides simplified to the same value, then use it as the result of
417 // the original comparison.
421 // The remaining cases only make sense if the select condition has the same
422 // type as the result of the comparison, so bail out if this is not so.
423 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
425 // If the false value simplified to false, then the result of the compare
426 // is equal to "Cond && TCmp". This also catches the case when the false
427 // value simplified to false and the true value to true, returning "Cond".
428 if (match(FCmp, m_Zero()))
429 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
431 // If the true value simplified to true, then the result of the compare
432 // is equal to "Cond || FCmp".
433 if (match(TCmp, m_One()))
434 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
436 // Finally, if the false value simplified to true and the true value to
437 // false, then the result of the compare is equal to "!Cond".
438 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
440 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
447 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
448 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
449 /// it on the incoming phi values yields the same result for every value. If so
450 /// returns the common value, otherwise returns null.
451 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
452 const Query &Q, unsigned MaxRecurse) {
453 // Recursion is always used, so bail out at once if we already hit the limit.
458 if (isa<PHINode>(LHS)) {
459 PI = cast<PHINode>(LHS);
460 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
461 if (!ValueDominatesPHI(RHS, PI, Q.DT))
464 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
465 PI = cast<PHINode>(RHS);
466 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
467 if (!ValueDominatesPHI(LHS, PI, Q.DT))
471 // Evaluate the BinOp on the incoming phi values.
472 Value *CommonValue = nullptr;
473 for (Value *Incoming : PI->incoming_values()) {
474 // If the incoming value is the phi node itself, it can safely be skipped.
475 if (Incoming == PI) continue;
476 Value *V = PI == LHS ?
477 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
478 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
479 // If the operation failed to simplify, or simplified to a different value
480 // to previously, then give up.
481 if (!V || (CommonValue && V != CommonValue))
489 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
490 /// try to simplify the comparison by seeing whether comparing with all of the
491 /// incoming phi values yields the same result every time. If so returns the
492 /// common result, otherwise returns null.
493 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
494 const Query &Q, unsigned MaxRecurse) {
495 // Recursion is always used, so bail out at once if we already hit the limit.
499 // Make sure the phi is on the LHS.
500 if (!isa<PHINode>(LHS)) {
502 Pred = CmpInst::getSwappedPredicate(Pred);
504 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
505 PHINode *PI = cast<PHINode>(LHS);
507 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
508 if (!ValueDominatesPHI(RHS, PI, Q.DT))
511 // Evaluate the BinOp on the incoming phi values.
512 Value *CommonValue = nullptr;
513 for (Value *Incoming : PI->incoming_values()) {
514 // If the incoming value is the phi node itself, it can safely be skipped.
515 if (Incoming == PI) continue;
516 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
517 // If the operation failed to simplify, or simplified to a different value
518 // to previously, then give up.
519 if (!V || (CommonValue && V != CommonValue))
527 /// SimplifyAddInst - Given operands for an Add, see if we can
528 /// fold the result. If not, this returns null.
529 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
530 const Query &Q, unsigned MaxRecurse) {
531 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
532 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
533 Constant *Ops[] = { CLHS, CRHS };
534 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
538 // Canonicalize the constant to the RHS.
542 // X + undef -> undef
543 if (match(Op1, m_Undef()))
547 if (match(Op1, m_Zero()))
554 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
555 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
558 // X + ~X -> -1 since ~X = -X-1
559 if (match(Op0, m_Not(m_Specific(Op1))) ||
560 match(Op1, m_Not(m_Specific(Op0))))
561 return Constant::getAllOnesValue(Op0->getType());
564 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
565 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
568 // Try some generic simplifications for associative operations.
569 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
573 // Threading Add over selects and phi nodes is pointless, so don't bother.
574 // Threading over the select in "A + select(cond, B, C)" means evaluating
575 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
576 // only if B and C are equal. If B and C are equal then (since we assume
577 // that operands have already been simplified) "select(cond, B, C)" should
578 // have been simplified to the common value of B and C already. Analysing
579 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
580 // for threading over phi nodes.
585 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
586 const DataLayout &DL, const TargetLibraryInfo *TLI,
587 const DominatorTree *DT, AssumptionCache *AC,
588 const Instruction *CxtI) {
589 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
593 /// \brief Compute the base pointer and cumulative constant offsets for V.
595 /// This strips all constant offsets off of V, leaving it the base pointer, and
596 /// accumulates the total constant offset applied in the returned constant. It
597 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
598 /// no constant offsets applied.
600 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
601 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
603 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
604 bool AllowNonInbounds = false) {
605 assert(V->getType()->getScalarType()->isPointerTy());
607 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
608 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
610 // Even though we don't look through PHI nodes, we could be called on an
611 // instruction in an unreachable block, which may be on a cycle.
612 SmallPtrSet<Value *, 4> Visited;
615 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
616 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
617 !GEP->accumulateConstantOffset(DL, Offset))
619 V = GEP->getPointerOperand();
620 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
621 V = cast<Operator>(V)->getOperand(0);
622 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
623 if (GA->mayBeOverridden())
625 V = GA->getAliasee();
629 assert(V->getType()->getScalarType()->isPointerTy() &&
630 "Unexpected operand type!");
631 } while (Visited.insert(V).second);
633 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
634 if (V->getType()->isVectorTy())
635 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
640 /// \brief Compute the constant difference between two pointer values.
641 /// If the difference is not a constant, returns zero.
642 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
644 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
645 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
647 // If LHS and RHS are not related via constant offsets to the same base
648 // value, there is nothing we can do here.
652 // Otherwise, the difference of LHS - RHS can be computed as:
654 // = (LHSOffset + Base) - (RHSOffset + Base)
655 // = LHSOffset - RHSOffset
656 return ConstantExpr::getSub(LHSOffset, RHSOffset);
659 /// SimplifySubInst - Given operands for a Sub, see if we can
660 /// fold the result. If not, this returns null.
661 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
662 const Query &Q, unsigned MaxRecurse) {
663 if (Constant *CLHS = dyn_cast<Constant>(Op0))
664 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
665 Constant *Ops[] = { CLHS, CRHS };
666 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
670 // X - undef -> undef
671 // undef - X -> undef
672 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
673 return UndefValue::get(Op0->getType());
676 if (match(Op1, m_Zero()))
681 return Constant::getNullValue(Op0->getType());
683 // 0 - X -> 0 if the sub is NUW.
684 if (isNUW && match(Op0, m_Zero()))
687 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
688 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
689 Value *X = nullptr, *Y = nullptr, *Z = Op1;
690 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
691 // See if "V === Y - Z" simplifies.
692 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
693 // It does! Now see if "X + V" simplifies.
694 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
695 // It does, we successfully reassociated!
699 // See if "V === X - Z" simplifies.
700 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
701 // It does! Now see if "Y + V" simplifies.
702 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
703 // It does, we successfully reassociated!
709 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
710 // For example, X - (X + 1) -> -1
712 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
713 // See if "V === X - Y" simplifies.
714 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
715 // It does! Now see if "V - Z" simplifies.
716 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
717 // It does, we successfully reassociated!
721 // See if "V === X - Z" simplifies.
722 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
723 // It does! Now see if "V - Y" simplifies.
724 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
725 // It does, we successfully reassociated!
731 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
732 // For example, X - (X - Y) -> Y.
734 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
735 // See if "V === Z - X" simplifies.
736 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
737 // It does! Now see if "V + Y" simplifies.
738 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
739 // It does, we successfully reassociated!
744 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
745 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
746 match(Op1, m_Trunc(m_Value(Y))))
747 if (X->getType() == Y->getType())
748 // See if "V === X - Y" simplifies.
749 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
750 // It does! Now see if "trunc V" simplifies.
751 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
752 // It does, return the simplified "trunc V".
755 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
756 if (match(Op0, m_PtrToInt(m_Value(X))) &&
757 match(Op1, m_PtrToInt(m_Value(Y))))
758 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
759 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
762 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
763 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
766 // Threading Sub over selects and phi nodes is pointless, so don't bother.
767 // Threading over the select in "A - select(cond, B, C)" means evaluating
768 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
769 // only if B and C are equal. If B and C are equal then (since we assume
770 // that operands have already been simplified) "select(cond, B, C)" should
771 // have been simplified to the common value of B and C already. Analysing
772 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
773 // for threading over phi nodes.
778 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
779 const DataLayout &DL, const TargetLibraryInfo *TLI,
780 const DominatorTree *DT, AssumptionCache *AC,
781 const Instruction *CxtI) {
782 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
786 /// Given operands for an FAdd, see if we can fold the result. If not, this
788 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
789 const Query &Q, unsigned MaxRecurse) {
790 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
791 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
792 Constant *Ops[] = { CLHS, CRHS };
793 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(),
797 // Canonicalize the constant to the RHS.
802 if (match(Op1, m_NegZero()))
805 // fadd X, 0 ==> X, when we know X is not -0
806 if (match(Op1, m_Zero()) &&
807 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
810 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
811 // where nnan and ninf have to occur at least once somewhere in this
813 Value *SubOp = nullptr;
814 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
816 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
819 Instruction *FSub = cast<Instruction>(SubOp);
820 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
821 (FMF.noInfs() || FSub->hasNoInfs()))
822 return Constant::getNullValue(Op0->getType());
828 /// Given operands for an FSub, see if we can fold the result. If not, this
830 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
831 const Query &Q, unsigned MaxRecurse) {
832 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
833 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
834 Constant *Ops[] = { CLHS, CRHS };
835 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(),
841 if (match(Op1, m_Zero()))
844 // fsub X, -0 ==> X, when we know X is not -0
845 if (match(Op1, m_NegZero()) &&
846 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
849 // fsub 0, (fsub -0.0, X) ==> X
851 if (match(Op0, m_AnyZero())) {
852 if (match(Op1, m_FSub(m_NegZero(), m_Value(X))))
854 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
858 // fsub nnan x, x ==> 0.0
859 if (FMF.noNaNs() && Op0 == Op1)
860 return Constant::getNullValue(Op0->getType());
865 /// Given the operands for an FMul, see if we can fold the result
866 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
869 unsigned MaxRecurse) {
870 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
871 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
872 Constant *Ops[] = { CLHS, CRHS };
873 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(),
877 // Canonicalize the constant to the RHS.
882 if (match(Op1, m_FPOne()))
885 // fmul nnan nsz X, 0 ==> 0
886 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
892 /// SimplifyMulInst - Given operands for a Mul, see if we can
893 /// fold the result. If not, this returns null.
894 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
895 unsigned MaxRecurse) {
896 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
897 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
898 Constant *Ops[] = { CLHS, CRHS };
899 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(),
903 // Canonicalize the constant to the RHS.
908 if (match(Op1, m_Undef()))
909 return Constant::getNullValue(Op0->getType());
912 if (match(Op1, m_Zero()))
916 if (match(Op1, m_One()))
919 // (X / Y) * Y -> X if the division is exact.
921 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
922 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
926 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
927 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
930 // Try some generic simplifications for associative operations.
931 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
935 // Mul distributes over Add. Try some generic simplifications based on this.
936 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
940 // If the operation is with the result of a select instruction, check whether
941 // operating on either branch of the select always yields the same value.
942 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
943 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
947 // If the operation is with the result of a phi instruction, check whether
948 // operating on all incoming values of the phi always yields the same value.
949 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
950 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
957 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
958 const DataLayout &DL,
959 const TargetLibraryInfo *TLI,
960 const DominatorTree *DT, AssumptionCache *AC,
961 const Instruction *CxtI) {
962 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
966 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
967 const DataLayout &DL,
968 const TargetLibraryInfo *TLI,
969 const DominatorTree *DT, AssumptionCache *AC,
970 const Instruction *CxtI) {
971 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
975 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
976 const DataLayout &DL,
977 const TargetLibraryInfo *TLI,
978 const DominatorTree *DT, AssumptionCache *AC,
979 const Instruction *CxtI) {
980 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
984 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout &DL,
985 const TargetLibraryInfo *TLI,
986 const DominatorTree *DT, AssumptionCache *AC,
987 const Instruction *CxtI) {
988 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
992 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
993 /// fold the result. If not, this returns null.
994 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
995 const Query &Q, unsigned MaxRecurse) {
996 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
997 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
998 Constant *Ops[] = { C0, C1 };
999 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1003 bool isSigned = Opcode == Instruction::SDiv;
1005 // X / undef -> undef
1006 if (match(Op1, m_Undef()))
1009 // X / 0 -> undef, we don't need to preserve faults!
1010 if (match(Op1, m_Zero()))
1011 return UndefValue::get(Op1->getType());
1014 if (match(Op0, m_Undef()))
1015 return Constant::getNullValue(Op0->getType());
1017 // 0 / X -> 0, we don't need to preserve faults!
1018 if (match(Op0, m_Zero()))
1022 if (match(Op1, m_One()))
1025 if (Op0->getType()->isIntegerTy(1))
1026 // It can't be division by zero, hence it must be division by one.
1031 return ConstantInt::get(Op0->getType(), 1);
1033 // (X * Y) / Y -> X if the multiplication does not overflow.
1034 Value *X = nullptr, *Y = nullptr;
1035 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1036 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1037 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1038 // If the Mul knows it does not overflow, then we are good to go.
1039 if ((isSigned && Mul->hasNoSignedWrap()) ||
1040 (!isSigned && Mul->hasNoUnsignedWrap()))
1042 // If X has the form X = A / Y then X * Y cannot overflow.
1043 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1044 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1048 // (X rem Y) / Y -> 0
1049 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1050 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1051 return Constant::getNullValue(Op0->getType());
1053 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1054 ConstantInt *C1, *C2;
1055 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1056 match(Op1, m_ConstantInt(C2))) {
1058 C1->getValue().umul_ov(C2->getValue(), Overflow);
1060 return Constant::getNullValue(Op0->getType());
1063 // If the operation is with the result of a select instruction, check whether
1064 // operating on either branch of the select always yields the same value.
1065 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1066 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1069 // If the operation is with the result of a phi instruction, check whether
1070 // operating on all incoming values of the phi always yields the same value.
1071 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1072 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1078 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1079 /// fold the result. If not, this returns null.
1080 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1081 unsigned MaxRecurse) {
1082 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1088 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1089 const TargetLibraryInfo *TLI,
1090 const DominatorTree *DT, AssumptionCache *AC,
1091 const Instruction *CxtI) {
1092 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1096 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1097 /// fold the result. If not, this returns null.
1098 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1099 unsigned MaxRecurse) {
1100 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1106 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1107 const TargetLibraryInfo *TLI,
1108 const DominatorTree *DT, AssumptionCache *AC,
1109 const Instruction *CxtI) {
1110 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1114 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1115 const Query &Q, unsigned) {
1116 // undef / X -> undef (the undef could be a snan).
1117 if (match(Op0, m_Undef()))
1120 // X / undef -> undef
1121 if (match(Op1, m_Undef()))
1125 // Requires that NaNs are off (X could be zero) and signed zeroes are
1126 // ignored (X could be positive or negative, so the output sign is unknown).
1127 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1131 // X / X -> 1.0 is legal when NaNs are ignored.
1133 return ConstantFP::get(Op0->getType(), 1.0);
1135 // -X / X -> -1.0 and
1136 // X / -X -> -1.0 are legal when NaNs are ignored.
1137 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1138 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1139 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1140 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1141 BinaryOperator::getFNegArgument(Op1) == Op0))
1142 return ConstantFP::get(Op0->getType(), -1.0);
1148 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1149 const DataLayout &DL,
1150 const TargetLibraryInfo *TLI,
1151 const DominatorTree *DT, AssumptionCache *AC,
1152 const Instruction *CxtI) {
1153 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1157 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1158 /// fold the result. If not, this returns null.
1159 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1160 const Query &Q, unsigned MaxRecurse) {
1161 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1162 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1163 Constant *Ops[] = { C0, C1 };
1164 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1168 // X % undef -> undef
1169 if (match(Op1, m_Undef()))
1173 if (match(Op0, m_Undef()))
1174 return Constant::getNullValue(Op0->getType());
1176 // 0 % X -> 0, we don't need to preserve faults!
1177 if (match(Op0, m_Zero()))
1180 // X % 0 -> undef, we don't need to preserve faults!
1181 if (match(Op1, m_Zero()))
1182 return UndefValue::get(Op0->getType());
1185 if (match(Op1, m_One()))
1186 return Constant::getNullValue(Op0->getType());
1188 if (Op0->getType()->isIntegerTy(1))
1189 // It can't be remainder by zero, hence it must be remainder by one.
1190 return Constant::getNullValue(Op0->getType());
1194 return Constant::getNullValue(Op0->getType());
1196 // (X % Y) % Y -> X % Y
1197 if ((Opcode == Instruction::SRem &&
1198 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1199 (Opcode == Instruction::URem &&
1200 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1203 // If the operation is with the result of a select instruction, check whether
1204 // operating on either branch of the select always yields the same value.
1205 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1206 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1209 // If the operation is with the result of a phi instruction, check whether
1210 // operating on all incoming values of the phi always yields the same value.
1211 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1212 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1218 /// SimplifySRemInst - Given operands for an SRem, see if we can
1219 /// fold the result. If not, this returns null.
1220 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1221 unsigned MaxRecurse) {
1222 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1228 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1229 const TargetLibraryInfo *TLI,
1230 const DominatorTree *DT, AssumptionCache *AC,
1231 const Instruction *CxtI) {
1232 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1236 /// SimplifyURemInst - Given operands for a URem, see if we can
1237 /// fold the result. If not, this returns null.
1238 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1239 unsigned MaxRecurse) {
1240 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1246 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1247 const TargetLibraryInfo *TLI,
1248 const DominatorTree *DT, AssumptionCache *AC,
1249 const Instruction *CxtI) {
1250 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1254 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1255 const Query &, unsigned) {
1256 // undef % X -> undef (the undef could be a snan).
1257 if (match(Op0, m_Undef()))
1260 // X % undef -> undef
1261 if (match(Op1, m_Undef()))
1265 // Requires that NaNs are off (X could be zero) and signed zeroes are
1266 // ignored (X could be positive or negative, so the output sign is unknown).
1267 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1273 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1274 const DataLayout &DL,
1275 const TargetLibraryInfo *TLI,
1276 const DominatorTree *DT, AssumptionCache *AC,
1277 const Instruction *CxtI) {
1278 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1282 /// isUndefShift - Returns true if a shift by \c Amount always yields undef.
1283 static bool isUndefShift(Value *Amount) {
1284 Constant *C = dyn_cast<Constant>(Amount);
1288 // X shift by undef -> undef because it may shift by the bitwidth.
1289 if (isa<UndefValue>(C))
1292 // Shifting by the bitwidth or more is undefined.
1293 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1294 if (CI->getValue().getLimitedValue() >=
1295 CI->getType()->getScalarSizeInBits())
1298 // If all lanes of a vector shift are undefined the whole shift is.
1299 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1300 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1301 if (!isUndefShift(C->getAggregateElement(I)))
1309 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1310 /// fold the result. If not, this returns null.
1311 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1312 const Query &Q, unsigned MaxRecurse) {
1313 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1314 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1315 Constant *Ops[] = { C0, C1 };
1316 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1320 // 0 shift by X -> 0
1321 if (match(Op0, m_Zero()))
1324 // X shift by 0 -> X
1325 if (match(Op1, m_Zero()))
1328 // Fold undefined shifts.
1329 if (isUndefShift(Op1))
1330 return UndefValue::get(Op0->getType());
1332 // If the operation is with the result of a select instruction, check whether
1333 // operating on either branch of the select always yields the same value.
1334 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1335 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1338 // If the operation is with the result of a phi instruction, check whether
1339 // operating on all incoming values of the phi always yields the same value.
1340 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1341 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1347 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1348 /// fold the result. If not, this returns null.
1349 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1,
1350 bool isExact, const Query &Q,
1351 unsigned MaxRecurse) {
1352 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1357 return Constant::getNullValue(Op0->getType());
1360 // undef >> X -> undef (if it's exact)
1361 if (match(Op0, m_Undef()))
1362 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1364 // The low bit cannot be shifted out of an exact shift if it is set.
1366 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
1367 APInt Op0KnownZero(BitWidth, 0);
1368 APInt Op0KnownOne(BitWidth, 0);
1369 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC,
1378 /// SimplifyShlInst - Given operands for an Shl, see if we can
1379 /// fold the result. If not, this returns null.
1380 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1381 const Query &Q, unsigned MaxRecurse) {
1382 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1386 // undef << X -> undef if (if it's NSW/NUW)
1387 if (match(Op0, m_Undef()))
1388 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1390 // (X >> A) << A -> X
1392 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1397 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1398 const DataLayout &DL, const TargetLibraryInfo *TLI,
1399 const DominatorTree *DT, AssumptionCache *AC,
1400 const Instruction *CxtI) {
1401 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
1405 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1406 /// fold the result. If not, this returns null.
1407 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1408 const Query &Q, unsigned MaxRecurse) {
1409 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1413 // (X << A) >> A -> X
1415 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1421 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1422 const DataLayout &DL,
1423 const TargetLibraryInfo *TLI,
1424 const DominatorTree *DT, AssumptionCache *AC,
1425 const Instruction *CxtI) {
1426 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1430 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1431 /// fold the result. If not, this returns null.
1432 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1433 const Query &Q, unsigned MaxRecurse) {
1434 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1438 // all ones >>a X -> all ones
1439 if (match(Op0, m_AllOnes()))
1442 // (X << A) >> A -> X
1444 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1447 // Arithmetic shifting an all-sign-bit value is a no-op.
1448 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1449 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1455 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1456 const DataLayout &DL,
1457 const TargetLibraryInfo *TLI,
1458 const DominatorTree *DT, AssumptionCache *AC,
1459 const Instruction *CxtI) {
1460 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1464 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1465 ICmpInst *UnsignedICmp, bool IsAnd) {
1468 ICmpInst::Predicate EqPred;
1469 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1470 !ICmpInst::isEquality(EqPred))
1473 ICmpInst::Predicate UnsignedPred;
1474 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1475 ICmpInst::isUnsigned(UnsignedPred))
1477 else if (match(UnsignedICmp,
1478 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1479 ICmpInst::isUnsigned(UnsignedPred))
1480 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1484 // X < Y && Y != 0 --> X < Y
1485 // X < Y || Y != 0 --> Y != 0
1486 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1487 return IsAnd ? UnsignedICmp : ZeroICmp;
1489 // X >= Y || Y != 0 --> true
1490 // X >= Y || Y == 0 --> X >= Y
1491 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1492 if (EqPred == ICmpInst::ICMP_NE)
1493 return getTrue(UnsignedICmp->getType());
1494 return UnsignedICmp;
1497 // X < Y && Y == 0 --> false
1498 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1500 return getFalse(UnsignedICmp->getType());
1505 // Simplify (and (icmp ...) (icmp ...)) to true when we can tell that the range
1506 // of possible values cannot be satisfied.
1507 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1508 ICmpInst::Predicate Pred0, Pred1;
1509 ConstantInt *CI1, *CI2;
1512 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1515 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)),
1516 m_ConstantInt(CI2))))
1519 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1))))
1522 Type *ITy = Op0->getType();
1524 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1525 bool isNSW = AddInst->hasNoSignedWrap();
1526 bool isNUW = AddInst->hasNoUnsignedWrap();
1528 const APInt &CI1V = CI1->getValue();
1529 const APInt &CI2V = CI2->getValue();
1530 const APInt Delta = CI2V - CI1V;
1531 if (CI1V.isStrictlyPositive()) {
1533 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1534 return getFalse(ITy);
1535 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1536 return getFalse(ITy);
1539 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1540 return getFalse(ITy);
1541 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1542 return getFalse(ITy);
1545 if (CI1V.getBoolValue() && isNUW) {
1547 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1548 return getFalse(ITy);
1550 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1551 return getFalse(ITy);
1557 /// SimplifyAndInst - Given operands for an And, see if we can
1558 /// fold the result. If not, this returns null.
1559 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1560 unsigned MaxRecurse) {
1561 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1562 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1563 Constant *Ops[] = { CLHS, CRHS };
1564 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1568 // Canonicalize the constant to the RHS.
1569 std::swap(Op0, Op1);
1573 if (match(Op1, m_Undef()))
1574 return Constant::getNullValue(Op0->getType());
1581 if (match(Op1, m_Zero()))
1585 if (match(Op1, m_AllOnes()))
1588 // A & ~A = ~A & A = 0
1589 if (match(Op0, m_Not(m_Specific(Op1))) ||
1590 match(Op1, m_Not(m_Specific(Op0))))
1591 return Constant::getNullValue(Op0->getType());
1594 Value *A = nullptr, *B = nullptr;
1595 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1596 (A == Op1 || B == Op1))
1600 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1601 (A == Op0 || B == Op0))
1604 // A & (-A) = A if A is a power of two or zero.
1605 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1606 match(Op1, m_Neg(m_Specific(Op0)))) {
1607 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1610 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1615 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1616 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1617 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS))
1619 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS))
1624 // Try some generic simplifications for associative operations.
1625 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1629 // And distributes over Or. Try some generic simplifications based on this.
1630 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1634 // And distributes over Xor. Try some generic simplifications based on this.
1635 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1639 // If the operation is with the result of a select instruction, check whether
1640 // operating on either branch of the select always yields the same value.
1641 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1642 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1646 // If the operation is with the result of a phi instruction, check whether
1647 // operating on all incoming values of the phi always yields the same value.
1648 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1649 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1656 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout &DL,
1657 const TargetLibraryInfo *TLI,
1658 const DominatorTree *DT, AssumptionCache *AC,
1659 const Instruction *CxtI) {
1660 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1664 // Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union
1665 // contains all possible values.
1666 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1667 ICmpInst::Predicate Pred0, Pred1;
1668 ConstantInt *CI1, *CI2;
1671 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1674 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)),
1675 m_ConstantInt(CI2))))
1678 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1))))
1681 Type *ITy = Op0->getType();
1683 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1684 bool isNSW = AddInst->hasNoSignedWrap();
1685 bool isNUW = AddInst->hasNoUnsignedWrap();
1687 const APInt &CI1V = CI1->getValue();
1688 const APInt &CI2V = CI2->getValue();
1689 const APInt Delta = CI2V - CI1V;
1690 if (CI1V.isStrictlyPositive()) {
1692 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1693 return getTrue(ITy);
1694 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1695 return getTrue(ITy);
1698 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1699 return getTrue(ITy);
1700 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1701 return getTrue(ITy);
1704 if (CI1V.getBoolValue() && isNUW) {
1706 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1707 return getTrue(ITy);
1709 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1710 return getTrue(ITy);
1716 /// SimplifyOrInst - Given operands for an Or, see if we can
1717 /// fold the result. If not, this returns null.
1718 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1719 unsigned MaxRecurse) {
1720 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1721 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1722 Constant *Ops[] = { CLHS, CRHS };
1723 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1727 // Canonicalize the constant to the RHS.
1728 std::swap(Op0, Op1);
1732 if (match(Op1, m_Undef()))
1733 return Constant::getAllOnesValue(Op0->getType());
1740 if (match(Op1, m_Zero()))
1744 if (match(Op1, m_AllOnes()))
1747 // A | ~A = ~A | A = -1
1748 if (match(Op0, m_Not(m_Specific(Op1))) ||
1749 match(Op1, m_Not(m_Specific(Op0))))
1750 return Constant::getAllOnesValue(Op0->getType());
1753 Value *A = nullptr, *B = nullptr;
1754 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1755 (A == Op1 || B == Op1))
1759 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1760 (A == Op0 || B == Op0))
1763 // ~(A & ?) | A = -1
1764 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1765 (A == Op1 || B == Op1))
1766 return Constant::getAllOnesValue(Op1->getType());
1768 // A | ~(A & ?) = -1
1769 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1770 (A == Op0 || B == Op0))
1771 return Constant::getAllOnesValue(Op0->getType());
1773 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1774 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1775 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS))
1777 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS))
1782 // Try some generic simplifications for associative operations.
1783 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1787 // Or distributes over And. Try some generic simplifications based on this.
1788 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1792 // If the operation is with the result of a select instruction, check whether
1793 // operating on either branch of the select always yields the same value.
1794 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1795 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1800 Value *C = nullptr, *D = nullptr;
1801 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1802 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1803 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1804 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1805 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1806 // (A & C1)|(B & C2)
1807 // If we have: ((V + N) & C1) | (V & C2)
1808 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1809 // replace with V+N.
1811 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1812 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1813 // Add commutes, try both ways.
1815 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1818 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1821 // Or commutes, try both ways.
1822 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1823 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1824 // Add commutes, try both ways.
1826 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1829 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1835 // If the operation is with the result of a phi instruction, check whether
1836 // operating on all incoming values of the phi always yields the same value.
1837 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1838 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1844 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout &DL,
1845 const TargetLibraryInfo *TLI,
1846 const DominatorTree *DT, AssumptionCache *AC,
1847 const Instruction *CxtI) {
1848 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1852 /// SimplifyXorInst - Given operands for a Xor, see if we can
1853 /// fold the result. If not, this returns null.
1854 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1855 unsigned MaxRecurse) {
1856 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1857 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1858 Constant *Ops[] = { CLHS, CRHS };
1859 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1863 // Canonicalize the constant to the RHS.
1864 std::swap(Op0, Op1);
1867 // A ^ undef -> undef
1868 if (match(Op1, m_Undef()))
1872 if (match(Op1, m_Zero()))
1877 return Constant::getNullValue(Op0->getType());
1879 // A ^ ~A = ~A ^ A = -1
1880 if (match(Op0, m_Not(m_Specific(Op1))) ||
1881 match(Op1, m_Not(m_Specific(Op0))))
1882 return Constant::getAllOnesValue(Op0->getType());
1884 // Try some generic simplifications for associative operations.
1885 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1889 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1890 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1891 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1892 // only if B and C are equal. If B and C are equal then (since we assume
1893 // that operands have already been simplified) "select(cond, B, C)" should
1894 // have been simplified to the common value of B and C already. Analysing
1895 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1896 // for threading over phi nodes.
1901 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout &DL,
1902 const TargetLibraryInfo *TLI,
1903 const DominatorTree *DT, AssumptionCache *AC,
1904 const Instruction *CxtI) {
1905 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1909 static Type *GetCompareTy(Value *Op) {
1910 return CmpInst::makeCmpResultType(Op->getType());
1913 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1914 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1915 /// otherwise return null. Helper function for analyzing max/min idioms.
1916 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1917 Value *LHS, Value *RHS) {
1918 SelectInst *SI = dyn_cast<SelectInst>(V);
1921 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1924 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1925 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1927 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1928 LHS == CmpRHS && RHS == CmpLHS)
1933 // A significant optimization not implemented here is assuming that alloca
1934 // addresses are not equal to incoming argument values. They don't *alias*,
1935 // as we say, but that doesn't mean they aren't equal, so we take a
1936 // conservative approach.
1938 // This is inspired in part by C++11 5.10p1:
1939 // "Two pointers of the same type compare equal if and only if they are both
1940 // null, both point to the same function, or both represent the same
1943 // This is pretty permissive.
1945 // It's also partly due to C11 6.5.9p6:
1946 // "Two pointers compare equal if and only if both are null pointers, both are
1947 // pointers to the same object (including a pointer to an object and a
1948 // subobject at its beginning) or function, both are pointers to one past the
1949 // last element of the same array object, or one is a pointer to one past the
1950 // end of one array object and the other is a pointer to the start of a
1951 // different array object that happens to immediately follow the first array
1952 // object in the address space.)
1954 // C11's version is more restrictive, however there's no reason why an argument
1955 // couldn't be a one-past-the-end value for a stack object in the caller and be
1956 // equal to the beginning of a stack object in the callee.
1958 // If the C and C++ standards are ever made sufficiently restrictive in this
1959 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1960 // this optimization.
1961 static Constant *computePointerICmp(const DataLayout &DL,
1962 const TargetLibraryInfo *TLI,
1963 CmpInst::Predicate Pred, Value *LHS,
1965 // First, skip past any trivial no-ops.
1966 LHS = LHS->stripPointerCasts();
1967 RHS = RHS->stripPointerCasts();
1969 // A non-null pointer is not equal to a null pointer.
1970 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) &&
1971 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1972 return ConstantInt::get(GetCompareTy(LHS),
1973 !CmpInst::isTrueWhenEqual(Pred));
1975 // We can only fold certain predicates on pointer comparisons.
1980 // Equality comaprisons are easy to fold.
1981 case CmpInst::ICMP_EQ:
1982 case CmpInst::ICMP_NE:
1985 // We can only handle unsigned relational comparisons because 'inbounds' on
1986 // a GEP only protects against unsigned wrapping.
1987 case CmpInst::ICMP_UGT:
1988 case CmpInst::ICMP_UGE:
1989 case CmpInst::ICMP_ULT:
1990 case CmpInst::ICMP_ULE:
1991 // However, we have to switch them to their signed variants to handle
1992 // negative indices from the base pointer.
1993 Pred = ICmpInst::getSignedPredicate(Pred);
1997 // Strip off any constant offsets so that we can reason about them.
1998 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
1999 // here and compare base addresses like AliasAnalysis does, however there are
2000 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2001 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2002 // doesn't need to guarantee pointer inequality when it says NoAlias.
2003 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2004 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2006 // If LHS and RHS are related via constant offsets to the same base
2007 // value, we can replace it with an icmp which just compares the offsets.
2009 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2011 // Various optimizations for (in)equality comparisons.
2012 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2013 // Different non-empty allocations that exist at the same time have
2014 // different addresses (if the program can tell). Global variables always
2015 // exist, so they always exist during the lifetime of each other and all
2016 // allocas. Two different allocas usually have different addresses...
2018 // However, if there's an @llvm.stackrestore dynamically in between two
2019 // allocas, they may have the same address. It's tempting to reduce the
2020 // scope of the problem by only looking at *static* allocas here. That would
2021 // cover the majority of allocas while significantly reducing the likelihood
2022 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2023 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2024 // an entry block. Also, if we have a block that's not attached to a
2025 // function, we can't tell if it's "static" under the current definition.
2026 // Theoretically, this problem could be fixed by creating a new kind of
2027 // instruction kind specifically for static allocas. Such a new instruction
2028 // could be required to be at the top of the entry block, thus preventing it
2029 // from being subject to a @llvm.stackrestore. Instcombine could even
2030 // convert regular allocas into these special allocas. It'd be nifty.
2031 // However, until then, this problem remains open.
2033 // So, we'll assume that two non-empty allocas have different addresses
2036 // With all that, if the offsets are within the bounds of their allocations
2037 // (and not one-past-the-end! so we can't use inbounds!), and their
2038 // allocations aren't the same, the pointers are not equal.
2040 // Note that it's not necessary to check for LHS being a global variable
2041 // address, due to canonicalization and constant folding.
2042 if (isa<AllocaInst>(LHS) &&
2043 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2044 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2045 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2046 uint64_t LHSSize, RHSSize;
2047 if (LHSOffsetCI && RHSOffsetCI &&
2048 getObjectSize(LHS, LHSSize, DL, TLI) &&
2049 getObjectSize(RHS, RHSSize, DL, TLI)) {
2050 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2051 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2052 if (!LHSOffsetValue.isNegative() &&
2053 !RHSOffsetValue.isNegative() &&
2054 LHSOffsetValue.ult(LHSSize) &&
2055 RHSOffsetValue.ult(RHSSize)) {
2056 return ConstantInt::get(GetCompareTy(LHS),
2057 !CmpInst::isTrueWhenEqual(Pred));
2061 // Repeat the above check but this time without depending on DataLayout
2062 // or being able to compute a precise size.
2063 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2064 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2065 LHSOffset->isNullValue() &&
2066 RHSOffset->isNullValue())
2067 return ConstantInt::get(GetCompareTy(LHS),
2068 !CmpInst::isTrueWhenEqual(Pred));
2071 // Even if an non-inbounds GEP occurs along the path we can still optimize
2072 // equality comparisons concerning the result. We avoid walking the whole
2073 // chain again by starting where the last calls to
2074 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2075 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2076 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2078 return ConstantExpr::getICmp(Pred,
2079 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2080 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2082 // If one side of the equality comparison must come from a noalias call
2083 // (meaning a system memory allocation function), and the other side must
2084 // come from a pointer that cannot overlap with dynamically-allocated
2085 // memory within the lifetime of the current function (allocas, byval
2086 // arguments, globals), then determine the comparison result here.
2087 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2088 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2089 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2091 // Is the set of underlying objects all noalias calls?
2092 auto IsNAC = [](SmallVectorImpl<Value *> &Objects) {
2093 return std::all_of(Objects.begin(), Objects.end(),
2094 [](Value *V){ return isNoAliasCall(V); });
2097 // Is the set of underlying objects all things which must be disjoint from
2098 // noalias calls. For allocas, we consider only static ones (dynamic
2099 // allocas might be transformed into calls to malloc not simultaneously
2100 // live with the compared-to allocation). For globals, we exclude symbols
2101 // that might be resolve lazily to symbols in another dynamically-loaded
2102 // library (and, thus, could be malloc'ed by the implementation).
2103 auto IsAllocDisjoint = [](SmallVectorImpl<Value *> &Objects) {
2104 return std::all_of(Objects.begin(), Objects.end(),
2106 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2107 return AI->getParent() && AI->getParent()->getParent() &&
2108 AI->isStaticAlloca();
2109 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2110 return (GV->hasLocalLinkage() ||
2111 GV->hasHiddenVisibility() ||
2112 GV->hasProtectedVisibility() ||
2113 GV->hasUnnamedAddr()) &&
2114 !GV->isThreadLocal();
2115 if (const Argument *A = dyn_cast<Argument>(V))
2116 return A->hasByValAttr();
2121 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2122 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2123 return ConstantInt::get(GetCompareTy(LHS),
2124 !CmpInst::isTrueWhenEqual(Pred));
2131 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
2132 /// fold the result. If not, this returns null.
2133 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2134 const Query &Q, unsigned MaxRecurse) {
2135 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2136 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
2138 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2139 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2140 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
2142 // If we have a constant, make sure it is on the RHS.
2143 std::swap(LHS, RHS);
2144 Pred = CmpInst::getSwappedPredicate(Pred);
2147 Type *ITy = GetCompareTy(LHS); // The return type.
2148 Type *OpTy = LHS->getType(); // The operand type.
2150 // icmp X, X -> true/false
2151 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
2152 // because X could be 0.
2153 if (LHS == RHS || isa<UndefValue>(RHS))
2154 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
2156 // Special case logic when the operands have i1 type.
2157 if (OpTy->getScalarType()->isIntegerTy(1)) {
2160 case ICmpInst::ICMP_EQ:
2162 if (match(RHS, m_One()))
2165 case ICmpInst::ICMP_NE:
2167 if (match(RHS, m_Zero()))
2170 case ICmpInst::ICMP_UGT:
2172 if (match(RHS, m_Zero()))
2175 case ICmpInst::ICMP_UGE:
2177 if (match(RHS, m_One()))
2180 case ICmpInst::ICMP_SLT:
2182 if (match(RHS, m_Zero()))
2185 case ICmpInst::ICMP_SLE:
2187 if (match(RHS, m_One()))
2193 // If we are comparing with zero then try hard since this is a common case.
2194 if (match(RHS, m_Zero())) {
2195 bool LHSKnownNonNegative, LHSKnownNegative;
2197 default: llvm_unreachable("Unknown ICmp predicate!");
2198 case ICmpInst::ICMP_ULT:
2199 return getFalse(ITy);
2200 case ICmpInst::ICMP_UGE:
2201 return getTrue(ITy);
2202 case ICmpInst::ICMP_EQ:
2203 case ICmpInst::ICMP_ULE:
2204 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2205 return getFalse(ITy);
2207 case ICmpInst::ICMP_NE:
2208 case ICmpInst::ICMP_UGT:
2209 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2210 return getTrue(ITy);
2212 case ICmpInst::ICMP_SLT:
2213 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2215 if (LHSKnownNegative)
2216 return getTrue(ITy);
2217 if (LHSKnownNonNegative)
2218 return getFalse(ITy);
2220 case ICmpInst::ICMP_SLE:
2221 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2223 if (LHSKnownNegative)
2224 return getTrue(ITy);
2225 if (LHSKnownNonNegative &&
2226 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2227 return getFalse(ITy);
2229 case ICmpInst::ICMP_SGE:
2230 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2232 if (LHSKnownNegative)
2233 return getFalse(ITy);
2234 if (LHSKnownNonNegative)
2235 return getTrue(ITy);
2237 case ICmpInst::ICMP_SGT:
2238 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2240 if (LHSKnownNegative)
2241 return getFalse(ITy);
2242 if (LHSKnownNonNegative &&
2243 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2244 return getTrue(ITy);
2249 // See if we are doing a comparison with a constant integer.
2250 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2251 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2252 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
2253 if (RHS_CR.isEmptySet())
2254 return ConstantInt::getFalse(CI->getContext());
2255 if (RHS_CR.isFullSet())
2256 return ConstantInt::getTrue(CI->getContext());
2258 // Many binary operators with constant RHS have easy to compute constant
2259 // range. Use them to check whether the comparison is a tautology.
2260 unsigned Width = CI->getBitWidth();
2261 APInt Lower = APInt(Width, 0);
2262 APInt Upper = APInt(Width, 0);
2264 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
2265 // 'urem x, CI2' produces [0, CI2).
2266 Upper = CI2->getValue();
2267 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
2268 // 'srem x, CI2' produces (-|CI2|, |CI2|).
2269 Upper = CI2->getValue().abs();
2270 Lower = (-Upper) + 1;
2271 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
2272 // 'udiv CI2, x' produces [0, CI2].
2273 Upper = CI2->getValue() + 1;
2274 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
2275 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
2276 APInt NegOne = APInt::getAllOnesValue(Width);
2278 Upper = NegOne.udiv(CI2->getValue()) + 1;
2279 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) {
2280 if (CI2->isMinSignedValue()) {
2281 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2282 Lower = CI2->getValue();
2283 Upper = Lower.lshr(1) + 1;
2285 // 'sdiv CI2, x' produces [-|CI2|, |CI2|].
2286 Upper = CI2->getValue().abs() + 1;
2287 Lower = (-Upper) + 1;
2289 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
2290 APInt IntMin = APInt::getSignedMinValue(Width);
2291 APInt IntMax = APInt::getSignedMaxValue(Width);
2292 APInt Val = CI2->getValue();
2293 if (Val.isAllOnesValue()) {
2294 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2295 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2298 } else if (Val.countLeadingZeros() < Width - 1) {
2299 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2]
2300 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2301 Lower = IntMin.sdiv(Val);
2302 Upper = IntMax.sdiv(Val);
2303 if (Lower.sgt(Upper))
2304 std::swap(Lower, Upper);
2306 assert(Upper != Lower && "Upper part of range has wrapped!");
2308 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) {
2309 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)]
2310 Lower = CI2->getValue();
2311 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2312 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) {
2313 if (CI2->isNegative()) {
2314 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2]
2315 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1;
2316 Lower = CI2->getValue().shl(ShiftAmount);
2317 Upper = CI2->getValue() + 1;
2319 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1]
2320 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1;
2321 Lower = CI2->getValue();
2322 Upper = CI2->getValue().shl(ShiftAmount) + 1;
2324 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
2325 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
2326 APInt NegOne = APInt::getAllOnesValue(Width);
2327 if (CI2->getValue().ult(Width))
2328 Upper = NegOne.lshr(CI2->getValue()) + 1;
2329 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) {
2330 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2].
2331 unsigned ShiftAmount = Width - 1;
2332 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2333 ShiftAmount = CI2->getValue().countTrailingZeros();
2334 Lower = CI2->getValue().lshr(ShiftAmount);
2335 Upper = CI2->getValue() + 1;
2336 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
2337 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
2338 APInt IntMin = APInt::getSignedMinValue(Width);
2339 APInt IntMax = APInt::getSignedMaxValue(Width);
2340 if (CI2->getValue().ult(Width)) {
2341 Lower = IntMin.ashr(CI2->getValue());
2342 Upper = IntMax.ashr(CI2->getValue()) + 1;
2344 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) {
2345 unsigned ShiftAmount = Width - 1;
2346 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2347 ShiftAmount = CI2->getValue().countTrailingZeros();
2348 if (CI2->isNegative()) {
2349 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)]
2350 Lower = CI2->getValue();
2351 Upper = CI2->getValue().ashr(ShiftAmount) + 1;
2353 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2]
2354 Lower = CI2->getValue().ashr(ShiftAmount);
2355 Upper = CI2->getValue() + 1;
2357 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
2358 // 'or x, CI2' produces [CI2, UINT_MAX].
2359 Lower = CI2->getValue();
2360 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
2361 // 'and x, CI2' produces [0, CI2].
2362 Upper = CI2->getValue() + 1;
2363 } else if (match(LHS, m_NUWAdd(m_Value(), m_ConstantInt(CI2)))) {
2364 // 'add nuw x, CI2' produces [CI2, UINT_MAX].
2365 Lower = CI2->getValue();
2367 if (Lower != Upper) {
2368 ConstantRange LHS_CR = ConstantRange(Lower, Upper);
2369 if (RHS_CR.contains(LHS_CR))
2370 return ConstantInt::getTrue(RHS->getContext());
2371 if (RHS_CR.inverse().contains(LHS_CR))
2372 return ConstantInt::getFalse(RHS->getContext());
2376 // Compare of cast, for example (zext X) != 0 -> X != 0
2377 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
2378 Instruction *LI = cast<CastInst>(LHS);
2379 Value *SrcOp = LI->getOperand(0);
2380 Type *SrcTy = SrcOp->getType();
2381 Type *DstTy = LI->getType();
2383 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
2384 // if the integer type is the same size as the pointer type.
2385 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
2386 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
2387 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2388 // Transfer the cast to the constant.
2389 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
2390 ConstantExpr::getIntToPtr(RHSC, SrcTy),
2393 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
2394 if (RI->getOperand(0)->getType() == SrcTy)
2395 // Compare without the cast.
2396 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2402 if (isa<ZExtInst>(LHS)) {
2403 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
2405 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
2406 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2407 // Compare X and Y. Note that signed predicates become unsigned.
2408 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2409 SrcOp, RI->getOperand(0), Q,
2413 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
2414 // too. If not, then try to deduce the result of the comparison.
2415 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2416 // Compute the constant that would happen if we truncated to SrcTy then
2417 // reextended to DstTy.
2418 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2419 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
2421 // If the re-extended constant didn't change then this is effectively
2422 // also a case of comparing two zero-extended values.
2423 if (RExt == CI && MaxRecurse)
2424 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2425 SrcOp, Trunc, Q, MaxRecurse-1))
2428 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
2429 // there. Use this to work out the result of the comparison.
2432 default: llvm_unreachable("Unknown ICmp predicate!");
2434 case ICmpInst::ICMP_EQ:
2435 case ICmpInst::ICMP_UGT:
2436 case ICmpInst::ICMP_UGE:
2437 return ConstantInt::getFalse(CI->getContext());
2439 case ICmpInst::ICMP_NE:
2440 case ICmpInst::ICMP_ULT:
2441 case ICmpInst::ICMP_ULE:
2442 return ConstantInt::getTrue(CI->getContext());
2444 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
2445 // is non-negative then LHS <s RHS.
2446 case ICmpInst::ICMP_SGT:
2447 case ICmpInst::ICMP_SGE:
2448 return CI->getValue().isNegative() ?
2449 ConstantInt::getTrue(CI->getContext()) :
2450 ConstantInt::getFalse(CI->getContext());
2452 case ICmpInst::ICMP_SLT:
2453 case ICmpInst::ICMP_SLE:
2454 return CI->getValue().isNegative() ?
2455 ConstantInt::getFalse(CI->getContext()) :
2456 ConstantInt::getTrue(CI->getContext());
2462 if (isa<SExtInst>(LHS)) {
2463 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
2465 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
2466 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2467 // Compare X and Y. Note that the predicate does not change.
2468 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2472 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
2473 // too. If not, then try to deduce the result of the comparison.
2474 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2475 // Compute the constant that would happen if we truncated to SrcTy then
2476 // reextended to DstTy.
2477 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2478 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
2480 // If the re-extended constant didn't change then this is effectively
2481 // also a case of comparing two sign-extended values.
2482 if (RExt == CI && MaxRecurse)
2483 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
2486 // Otherwise the upper bits of LHS are all equal, while RHS has varying
2487 // bits there. Use this to work out the result of the comparison.
2490 default: llvm_unreachable("Unknown ICmp predicate!");
2491 case ICmpInst::ICMP_EQ:
2492 return ConstantInt::getFalse(CI->getContext());
2493 case ICmpInst::ICMP_NE:
2494 return ConstantInt::getTrue(CI->getContext());
2496 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
2498 case ICmpInst::ICMP_SGT:
2499 case ICmpInst::ICMP_SGE:
2500 return CI->getValue().isNegative() ?
2501 ConstantInt::getTrue(CI->getContext()) :
2502 ConstantInt::getFalse(CI->getContext());
2503 case ICmpInst::ICMP_SLT:
2504 case ICmpInst::ICMP_SLE:
2505 return CI->getValue().isNegative() ?
2506 ConstantInt::getFalse(CI->getContext()) :
2507 ConstantInt::getTrue(CI->getContext());
2509 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2511 case ICmpInst::ICMP_UGT:
2512 case ICmpInst::ICMP_UGE:
2513 // Comparison is true iff the LHS <s 0.
2515 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2516 Constant::getNullValue(SrcTy),
2520 case ICmpInst::ICMP_ULT:
2521 case ICmpInst::ICMP_ULE:
2522 // Comparison is true iff the LHS >=s 0.
2524 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2525 Constant::getNullValue(SrcTy),
2535 // Special logic for binary operators.
2536 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2537 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2538 if (MaxRecurse && (LBO || RBO)) {
2539 // Analyze the case when either LHS or RHS is an add instruction.
2540 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2541 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2542 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2543 if (LBO && LBO->getOpcode() == Instruction::Add) {
2544 A = LBO->getOperand(0); B = LBO->getOperand(1);
2545 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2546 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2547 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2549 if (RBO && RBO->getOpcode() == Instruction::Add) {
2550 C = RBO->getOperand(0); D = RBO->getOperand(1);
2551 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2552 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2553 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2556 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2557 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2558 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2559 Constant::getNullValue(RHS->getType()),
2563 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2564 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2565 if (Value *V = SimplifyICmpInst(Pred,
2566 Constant::getNullValue(LHS->getType()),
2567 C == LHS ? D : C, Q, MaxRecurse-1))
2570 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2571 if (A && C && (A == C || A == D || B == C || B == D) &&
2572 NoLHSWrapProblem && NoRHSWrapProblem) {
2573 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2576 // C + B == C + D -> B == D
2579 } else if (A == D) {
2580 // D + B == C + D -> B == C
2583 } else if (B == C) {
2584 // A + C == C + D -> A == D
2589 // A + D == C + D -> A == C
2593 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2598 // icmp pred (or X, Y), X
2599 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)),
2600 m_Or(m_Specific(RHS), m_Value())))) {
2601 if (Pred == ICmpInst::ICMP_ULT)
2602 return getFalse(ITy);
2603 if (Pred == ICmpInst::ICMP_UGE)
2604 return getTrue(ITy);
2606 // icmp pred X, (or X, Y)
2607 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)),
2608 m_Or(m_Specific(LHS), m_Value())))) {
2609 if (Pred == ICmpInst::ICMP_ULE)
2610 return getTrue(ITy);
2611 if (Pred == ICmpInst::ICMP_UGT)
2612 return getFalse(ITy);
2615 // icmp pred (and X, Y), X
2616 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)),
2617 m_And(m_Specific(RHS), m_Value())))) {
2618 if (Pred == ICmpInst::ICMP_UGT)
2619 return getFalse(ITy);
2620 if (Pred == ICmpInst::ICMP_ULE)
2621 return getTrue(ITy);
2623 // icmp pred X, (and X, Y)
2624 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)),
2625 m_And(m_Specific(LHS), m_Value())))) {
2626 if (Pred == ICmpInst::ICMP_UGE)
2627 return getTrue(ITy);
2628 if (Pred == ICmpInst::ICMP_ULT)
2629 return getFalse(ITy);
2632 // 0 - (zext X) pred C
2633 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2634 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2635 if (RHSC->getValue().isStrictlyPositive()) {
2636 if (Pred == ICmpInst::ICMP_SLT)
2637 return ConstantInt::getTrue(RHSC->getContext());
2638 if (Pred == ICmpInst::ICMP_SGE)
2639 return ConstantInt::getFalse(RHSC->getContext());
2640 if (Pred == ICmpInst::ICMP_EQ)
2641 return ConstantInt::getFalse(RHSC->getContext());
2642 if (Pred == ICmpInst::ICMP_NE)
2643 return ConstantInt::getTrue(RHSC->getContext());
2645 if (RHSC->getValue().isNonNegative()) {
2646 if (Pred == ICmpInst::ICMP_SLE)
2647 return ConstantInt::getTrue(RHSC->getContext());
2648 if (Pred == ICmpInst::ICMP_SGT)
2649 return ConstantInt::getFalse(RHSC->getContext());
2654 // icmp pred (urem X, Y), Y
2655 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2656 bool KnownNonNegative, KnownNegative;
2660 case ICmpInst::ICMP_SGT:
2661 case ICmpInst::ICMP_SGE:
2662 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2664 if (!KnownNonNegative)
2667 case ICmpInst::ICMP_EQ:
2668 case ICmpInst::ICMP_UGT:
2669 case ICmpInst::ICMP_UGE:
2670 return getFalse(ITy);
2671 case ICmpInst::ICMP_SLT:
2672 case ICmpInst::ICMP_SLE:
2673 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2675 if (!KnownNonNegative)
2678 case ICmpInst::ICMP_NE:
2679 case ICmpInst::ICMP_ULT:
2680 case ICmpInst::ICMP_ULE:
2681 return getTrue(ITy);
2685 // icmp pred X, (urem Y, X)
2686 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2687 bool KnownNonNegative, KnownNegative;
2691 case ICmpInst::ICMP_SGT:
2692 case ICmpInst::ICMP_SGE:
2693 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2695 if (!KnownNonNegative)
2698 case ICmpInst::ICMP_NE:
2699 case ICmpInst::ICMP_UGT:
2700 case ICmpInst::ICMP_UGE:
2701 return getTrue(ITy);
2702 case ICmpInst::ICMP_SLT:
2703 case ICmpInst::ICMP_SLE:
2704 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2706 if (!KnownNonNegative)
2709 case ICmpInst::ICMP_EQ:
2710 case ICmpInst::ICMP_ULT:
2711 case ICmpInst::ICMP_ULE:
2712 return getFalse(ITy);
2717 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2718 // icmp pred (X /u Y), X
2719 if (Pred == ICmpInst::ICMP_UGT)
2720 return getFalse(ITy);
2721 if (Pred == ICmpInst::ICMP_ULE)
2722 return getTrue(ITy);
2729 // where CI2 is a power of 2 and CI isn't
2730 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2731 const APInt *CI2Val, *CIVal = &CI->getValue();
2732 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2733 CI2Val->isPowerOf2()) {
2734 if (!CIVal->isPowerOf2()) {
2735 // CI2 << X can equal zero in some circumstances,
2736 // this simplification is unsafe if CI is zero.
2738 // We know it is safe if:
2739 // - The shift is nsw, we can't shift out the one bit.
2740 // - The shift is nuw, we can't shift out the one bit.
2743 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2744 *CI2Val == 1 || !CI->isZero()) {
2745 if (Pred == ICmpInst::ICMP_EQ)
2746 return ConstantInt::getFalse(RHS->getContext());
2747 if (Pred == ICmpInst::ICMP_NE)
2748 return ConstantInt::getTrue(RHS->getContext());
2751 if (CIVal->isSignBit() && *CI2Val == 1) {
2752 if (Pred == ICmpInst::ICMP_UGT)
2753 return ConstantInt::getFalse(RHS->getContext());
2754 if (Pred == ICmpInst::ICMP_ULE)
2755 return ConstantInt::getTrue(RHS->getContext());
2760 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2761 LBO->getOperand(1) == RBO->getOperand(1)) {
2762 switch (LBO->getOpcode()) {
2764 case Instruction::UDiv:
2765 case Instruction::LShr:
2766 if (ICmpInst::isSigned(Pred))
2769 case Instruction::SDiv:
2770 case Instruction::AShr:
2771 if (!LBO->isExact() || !RBO->isExact())
2773 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2774 RBO->getOperand(0), Q, MaxRecurse-1))
2777 case Instruction::Shl: {
2778 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2779 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2782 if (!NSW && ICmpInst::isSigned(Pred))
2784 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2785 RBO->getOperand(0), Q, MaxRecurse-1))
2792 // Simplify comparisons involving max/min.
2794 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2795 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2797 // Signed variants on "max(a,b)>=a -> true".
2798 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2799 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2800 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2801 // We analyze this as smax(A, B) pred A.
2803 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2804 (A == LHS || B == LHS)) {
2805 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2806 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2807 // We analyze this as smax(A, B) swapped-pred A.
2808 P = CmpInst::getSwappedPredicate(Pred);
2809 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2810 (A == RHS || B == RHS)) {
2811 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2812 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2813 // We analyze this as smax(-A, -B) swapped-pred -A.
2814 // Note that we do not need to actually form -A or -B thanks to EqP.
2815 P = CmpInst::getSwappedPredicate(Pred);
2816 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2817 (A == LHS || B == LHS)) {
2818 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2819 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2820 // We analyze this as smax(-A, -B) pred -A.
2821 // Note that we do not need to actually form -A or -B thanks to EqP.
2824 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2825 // Cases correspond to "max(A, B) p A".
2829 case CmpInst::ICMP_EQ:
2830 case CmpInst::ICMP_SLE:
2831 // Equivalent to "A EqP B". This may be the same as the condition tested
2832 // in the max/min; if so, we can just return that.
2833 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2835 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2837 // Otherwise, see if "A EqP B" simplifies.
2839 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2842 case CmpInst::ICMP_NE:
2843 case CmpInst::ICMP_SGT: {
2844 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2845 // Equivalent to "A InvEqP B". This may be the same as the condition
2846 // tested in the max/min; if so, we can just return that.
2847 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2849 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2851 // Otherwise, see if "A InvEqP B" simplifies.
2853 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2857 case CmpInst::ICMP_SGE:
2859 return getTrue(ITy);
2860 case CmpInst::ICMP_SLT:
2862 return getFalse(ITy);
2866 // Unsigned variants on "max(a,b)>=a -> true".
2867 P = CmpInst::BAD_ICMP_PREDICATE;
2868 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2869 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2870 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2871 // We analyze this as umax(A, B) pred A.
2873 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2874 (A == LHS || B == LHS)) {
2875 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2876 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2877 // We analyze this as umax(A, B) swapped-pred A.
2878 P = CmpInst::getSwappedPredicate(Pred);
2879 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2880 (A == RHS || B == RHS)) {
2881 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2882 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2883 // We analyze this as umax(-A, -B) swapped-pred -A.
2884 // Note that we do not need to actually form -A or -B thanks to EqP.
2885 P = CmpInst::getSwappedPredicate(Pred);
2886 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2887 (A == LHS || B == LHS)) {
2888 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2889 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2890 // We analyze this as umax(-A, -B) pred -A.
2891 // Note that we do not need to actually form -A or -B thanks to EqP.
2894 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2895 // Cases correspond to "max(A, B) p A".
2899 case CmpInst::ICMP_EQ:
2900 case CmpInst::ICMP_ULE:
2901 // Equivalent to "A EqP B". This may be the same as the condition tested
2902 // in the max/min; if so, we can just return that.
2903 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2905 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2907 // Otherwise, see if "A EqP B" simplifies.
2909 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2912 case CmpInst::ICMP_NE:
2913 case CmpInst::ICMP_UGT: {
2914 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2915 // Equivalent to "A InvEqP B". This may be the same as the condition
2916 // tested in the max/min; if so, we can just return that.
2917 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2919 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2921 // Otherwise, see if "A InvEqP B" simplifies.
2923 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2927 case CmpInst::ICMP_UGE:
2929 return getTrue(ITy);
2930 case CmpInst::ICMP_ULT:
2932 return getFalse(ITy);
2936 // Variants on "max(x,y) >= min(x,z)".
2938 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2939 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2940 (A == C || A == D || B == C || B == D)) {
2941 // max(x, ?) pred min(x, ?).
2942 if (Pred == CmpInst::ICMP_SGE)
2944 return getTrue(ITy);
2945 if (Pred == CmpInst::ICMP_SLT)
2947 return getFalse(ITy);
2948 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2949 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2950 (A == C || A == D || B == C || B == D)) {
2951 // min(x, ?) pred max(x, ?).
2952 if (Pred == CmpInst::ICMP_SLE)
2954 return getTrue(ITy);
2955 if (Pred == CmpInst::ICMP_SGT)
2957 return getFalse(ITy);
2958 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2959 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2960 (A == C || A == D || B == C || B == D)) {
2961 // max(x, ?) pred min(x, ?).
2962 if (Pred == CmpInst::ICMP_UGE)
2964 return getTrue(ITy);
2965 if (Pred == CmpInst::ICMP_ULT)
2967 return getFalse(ITy);
2968 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2969 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2970 (A == C || A == D || B == C || B == D)) {
2971 // min(x, ?) pred max(x, ?).
2972 if (Pred == CmpInst::ICMP_ULE)
2974 return getTrue(ITy);
2975 if (Pred == CmpInst::ICMP_UGT)
2977 return getFalse(ITy);
2980 // Simplify comparisons of related pointers using a powerful, recursive
2981 // GEP-walk when we have target data available..
2982 if (LHS->getType()->isPointerTy())
2983 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS))
2986 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
2987 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
2988 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
2989 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
2990 (ICmpInst::isEquality(Pred) ||
2991 (GLHS->isInBounds() && GRHS->isInBounds() &&
2992 Pred == ICmpInst::getSignedPredicate(Pred)))) {
2993 // The bases are equal and the indices are constant. Build a constant
2994 // expression GEP with the same indices and a null base pointer to see
2995 // what constant folding can make out of it.
2996 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
2997 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
2998 Constant *NewLHS = ConstantExpr::getGetElementPtr(
2999 GLHS->getSourceElementType(), Null, IndicesLHS);
3001 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3002 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3003 GLHS->getSourceElementType(), Null, IndicesRHS);
3004 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3009 // If a bit is known to be zero for A and known to be one for B,
3010 // then A and B cannot be equal.
3011 if (ICmpInst::isEquality(Pred)) {
3012 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3013 uint32_t BitWidth = CI->getBitWidth();
3014 APInt LHSKnownZero(BitWidth, 0);
3015 APInt LHSKnownOne(BitWidth, 0);
3016 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC,
3018 const APInt &RHSVal = CI->getValue();
3019 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0))
3020 return Pred == ICmpInst::ICMP_EQ
3021 ? ConstantInt::getFalse(CI->getContext())
3022 : ConstantInt::getTrue(CI->getContext());
3026 // If the comparison is with the result of a select instruction, check whether
3027 // comparing with either branch of the select always yields the same value.
3028 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3029 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3032 // If the comparison is with the result of a phi instruction, check whether
3033 // doing the compare with each incoming phi value yields a common result.
3034 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3035 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3041 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3042 const DataLayout &DL,
3043 const TargetLibraryInfo *TLI,
3044 const DominatorTree *DT, AssumptionCache *AC,
3045 Instruction *CxtI) {
3046 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3050 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
3051 /// fold the result. If not, this returns null.
3052 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3053 FastMathFlags FMF, const Query &Q,
3054 unsigned MaxRecurse) {
3055 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3056 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3058 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3059 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3060 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3062 // If we have a constant, make sure it is on the RHS.
3063 std::swap(LHS, RHS);
3064 Pred = CmpInst::getSwappedPredicate(Pred);
3067 // Fold trivial predicates.
3068 if (Pred == FCmpInst::FCMP_FALSE)
3069 return ConstantInt::get(GetCompareTy(LHS), 0);
3070 if (Pred == FCmpInst::FCMP_TRUE)
3071 return ConstantInt::get(GetCompareTy(LHS), 1);
3073 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3075 if (Pred == FCmpInst::FCMP_UNO)
3076 return ConstantInt::get(GetCompareTy(LHS), 0);
3077 if (Pred == FCmpInst::FCMP_ORD)
3078 return ConstantInt::get(GetCompareTy(LHS), 1);
3081 // fcmp pred x, undef and fcmp pred undef, x
3082 // fold to true if unordered, false if ordered
3083 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3084 // Choosing NaN for the undef will always make unordered comparison succeed
3085 // and ordered comparison fail.
3086 return ConstantInt::get(GetCompareTy(LHS), CmpInst::isUnordered(Pred));
3089 // fcmp x,x -> true/false. Not all compares are foldable.
3091 if (CmpInst::isTrueWhenEqual(Pred))
3092 return ConstantInt::get(GetCompareTy(LHS), 1);
3093 if (CmpInst::isFalseWhenEqual(Pred))
3094 return ConstantInt::get(GetCompareTy(LHS), 0);
3097 // Handle fcmp with constant RHS
3098 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
3099 // If the constant is a nan, see if we can fold the comparison based on it.
3100 if (CFP->getValueAPF().isNaN()) {
3101 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3102 return ConstantInt::getFalse(CFP->getContext());
3103 assert(FCmpInst::isUnordered(Pred) &&
3104 "Comparison must be either ordered or unordered!");
3105 // True if unordered.
3106 return ConstantInt::getTrue(CFP->getContext());
3108 // Check whether the constant is an infinity.
3109 if (CFP->getValueAPF().isInfinity()) {
3110 if (CFP->getValueAPF().isNegative()) {
3112 case FCmpInst::FCMP_OLT:
3113 // No value is ordered and less than negative infinity.
3114 return ConstantInt::getFalse(CFP->getContext());
3115 case FCmpInst::FCMP_UGE:
3116 // All values are unordered with or at least negative infinity.
3117 return ConstantInt::getTrue(CFP->getContext());
3123 case FCmpInst::FCMP_OGT:
3124 // No value is ordered and greater than infinity.
3125 return ConstantInt::getFalse(CFP->getContext());
3126 case FCmpInst::FCMP_ULE:
3127 // All values are unordered with and at most infinity.
3128 return ConstantInt::getTrue(CFP->getContext());
3134 if (CFP->getValueAPF().isZero()) {
3136 case FCmpInst::FCMP_UGE:
3137 if (CannotBeOrderedLessThanZero(LHS))
3138 return ConstantInt::getTrue(CFP->getContext());
3140 case FCmpInst::FCMP_OLT:
3142 if (CannotBeOrderedLessThanZero(LHS))
3143 return ConstantInt::getFalse(CFP->getContext());
3151 // If the comparison is with the result of a select instruction, check whether
3152 // comparing with either branch of the select always yields the same value.
3153 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3154 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3157 // If the comparison is with the result of a phi instruction, check whether
3158 // doing the compare with each incoming phi value yields a common result.
3159 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3160 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3166 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3167 FastMathFlags FMF, const DataLayout &DL,
3168 const TargetLibraryInfo *TLI,
3169 const DominatorTree *DT, AssumptionCache *AC,
3170 const Instruction *CxtI) {
3171 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF,
3172 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3175 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is
3176 /// replaced with RepOp.
3177 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3179 unsigned MaxRecurse) {
3180 // Trivial replacement.
3184 auto *I = dyn_cast<Instruction>(V);
3188 // If this is a binary operator, try to simplify it with the replaced op.
3189 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3191 // %cmp = icmp eq i32 %x, 2147483647
3192 // %add = add nsw i32 %x, 1
3193 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3195 // We can't replace %sel with %add unless we strip away the flags.
3196 if (isa<OverflowingBinaryOperator>(B))
3197 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3199 if (isa<PossiblyExactOperator>(B))
3204 if (B->getOperand(0) == Op)
3205 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3207 if (B->getOperand(1) == Op)
3208 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3213 // Same for CmpInsts.
3214 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3216 if (C->getOperand(0) == Op)
3217 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3219 if (C->getOperand(1) == Op)
3220 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3225 // TODO: We could hand off more cases to instsimplify here.
3227 // If all operands are constant after substituting Op for RepOp then we can
3228 // constant fold the instruction.
3229 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3230 // Build a list of all constant operands.
3231 SmallVector<Constant *, 8> ConstOps;
3232 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3233 if (I->getOperand(i) == Op)
3234 ConstOps.push_back(CRepOp);
3235 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3236 ConstOps.push_back(COp);
3241 // All operands were constants, fold it.
3242 if (ConstOps.size() == I->getNumOperands()) {
3243 if (CmpInst *C = dyn_cast<CmpInst>(I))
3244 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3245 ConstOps[1], Q.DL, Q.TLI);
3247 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3248 if (!LI->isVolatile())
3249 return ConstantFoldLoadFromConstPtr(ConstOps[0], Q.DL);
3251 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), ConstOps,
3259 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
3260 /// the result. If not, this returns null.
3261 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3262 Value *FalseVal, const Query &Q,
3263 unsigned MaxRecurse) {
3264 // select true, X, Y -> X
3265 // select false, X, Y -> Y
3266 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3267 if (CB->isAllOnesValue())
3269 if (CB->isNullValue())
3273 // select C, X, X -> X
3274 if (TrueVal == FalseVal)
3277 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3278 if (isa<Constant>(TrueVal))
3282 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3284 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3287 if (const auto *ICI = dyn_cast<ICmpInst>(CondVal)) {
3288 unsigned BitWidth = Q.DL.getTypeSizeInBits(TrueVal->getType());
3289 ICmpInst::Predicate Pred = ICI->getPredicate();
3290 Value *CmpLHS = ICI->getOperand(0);
3291 Value *CmpRHS = ICI->getOperand(1);
3292 APInt MinSignedValue = APInt::getSignBit(BitWidth);
3296 bool IsBitTest = false;
3297 if (ICmpInst::isEquality(Pred) &&
3298 match(CmpLHS, m_And(m_Value(X), m_APInt(Y))) &&
3299 match(CmpRHS, m_Zero())) {
3301 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
3302 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3304 Y = &MinSignedValue;
3306 TrueWhenUnset = false;
3307 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3309 Y = &MinSignedValue;
3311 TrueWhenUnset = true;
3315 // (X & Y) == 0 ? X & ~Y : X --> X
3316 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3317 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3319 return TrueWhenUnset ? FalseVal : TrueVal;
3320 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3321 // (X & Y) != 0 ? X : X & ~Y --> X
3322 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3324 return TrueWhenUnset ? FalseVal : TrueVal;
3326 if (Y->isPowerOf2()) {
3327 // (X & Y) == 0 ? X | Y : X --> X | Y
3328 // (X & Y) != 0 ? X | Y : X --> X
3329 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3331 return TrueWhenUnset ? TrueVal : FalseVal;
3332 // (X & Y) == 0 ? X : X | Y --> X
3333 // (X & Y) != 0 ? X : X | Y --> X | Y
3334 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3336 return TrueWhenUnset ? TrueVal : FalseVal;
3339 if (ICI->hasOneUse()) {
3341 if (match(CmpRHS, m_APInt(C))) {
3342 // X < MIN ? T : F --> F
3343 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3345 // X < MIN ? T : F --> F
3346 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3348 // X > MAX ? T : F --> F
3349 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3351 // X > MAX ? T : F --> F
3352 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3357 // If we have an equality comparison then we know the value in one of the
3358 // arms of the select. See if substituting this value into the arm and
3359 // simplifying the result yields the same value as the other arm.
3360 if (Pred == ICmpInst::ICMP_EQ) {
3361 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3363 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3366 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3368 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3371 } else if (Pred == ICmpInst::ICMP_NE) {
3372 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3374 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3377 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3379 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3388 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3389 const DataLayout &DL,
3390 const TargetLibraryInfo *TLI,
3391 const DominatorTree *DT, AssumptionCache *AC,
3392 const Instruction *CxtI) {
3393 return ::SimplifySelectInst(Cond, TrueVal, FalseVal,
3394 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3397 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
3398 /// fold the result. If not, this returns null.
3399 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3400 const Query &Q, unsigned) {
3401 // The type of the GEP pointer operand.
3403 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3405 // getelementptr P -> P.
3406 if (Ops.size() == 1)
3409 // Compute the (pointer) type returned by the GEP instruction.
3410 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3411 Type *GEPTy = PointerType::get(LastType, AS);
3412 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3413 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3415 if (isa<UndefValue>(Ops[0]))
3416 return UndefValue::get(GEPTy);
3418 if (Ops.size() == 2) {
3419 // getelementptr P, 0 -> P.
3420 if (match(Ops[1], m_Zero()))
3424 if (Ty->isSized()) {
3427 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3428 // getelementptr P, N -> P if P points to a type of zero size.
3429 if (TyAllocSize == 0)
3432 // The following transforms are only safe if the ptrtoint cast
3433 // doesn't truncate the pointers.
3434 if (Ops[1]->getType()->getScalarSizeInBits() ==
3435 Q.DL.getPointerSizeInBits(AS)) {
3436 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3437 if (match(P, m_Zero()))
3438 return Constant::getNullValue(GEPTy);
3440 if (match(P, m_PtrToInt(m_Value(Temp))))
3441 if (Temp->getType() == GEPTy)
3446 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3447 if (TyAllocSize == 1 &&
3448 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3449 if (Value *R = PtrToIntOrZero(P))
3452 // getelementptr V, (ashr (sub P, V), C) -> Q
3453 // if P points to a type of size 1 << C.
3455 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3456 m_ConstantInt(C))) &&
3457 TyAllocSize == 1ULL << C)
3458 if (Value *R = PtrToIntOrZero(P))
3461 // getelementptr V, (sdiv (sub P, V), C) -> Q
3462 // if P points to a type of size C.
3464 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3465 m_SpecificInt(TyAllocSize))))
3466 if (Value *R = PtrToIntOrZero(P))
3472 // Check to see if this is constant foldable.
3473 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
3474 if (!isa<Constant>(Ops[i]))
3477 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3481 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout &DL,
3482 const TargetLibraryInfo *TLI,
3483 const DominatorTree *DT, AssumptionCache *AC,
3484 const Instruction *CxtI) {
3485 return ::SimplifyGEPInst(
3486 cast<PointerType>(Ops[0]->getType()->getScalarType())->getElementType(),
3487 Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3490 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
3491 /// can fold the result. If not, this returns null.
3492 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3493 ArrayRef<unsigned> Idxs, const Query &Q,
3495 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3496 if (Constant *CVal = dyn_cast<Constant>(Val))
3497 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3499 // insertvalue x, undef, n -> x
3500 if (match(Val, m_Undef()))
3503 // insertvalue x, (extractvalue y, n), n
3504 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3505 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3506 EV->getIndices() == Idxs) {
3507 // insertvalue undef, (extractvalue y, n), n -> y
3508 if (match(Agg, m_Undef()))
3509 return EV->getAggregateOperand();
3511 // insertvalue y, (extractvalue y, n), n -> y
3512 if (Agg == EV->getAggregateOperand())
3519 Value *llvm::SimplifyInsertValueInst(
3520 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL,
3521 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC,
3522 const Instruction *CxtI) {
3523 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI),
3527 /// SimplifyExtractValueInst - Given operands for an ExtractValueInst, see if we
3528 /// can fold the result. If not, this returns null.
3529 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3530 const Query &, unsigned) {
3531 if (auto *CAgg = dyn_cast<Constant>(Agg))
3532 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3534 // extractvalue x, (insertvalue y, elt, n), n -> elt
3535 unsigned NumIdxs = Idxs.size();
3536 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3537 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3538 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3539 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3540 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3541 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3542 Idxs.slice(0, NumCommonIdxs)) {
3543 if (NumIdxs == NumInsertValueIdxs)
3544 return IVI->getInsertedValueOperand();
3552 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3553 const DataLayout &DL,
3554 const TargetLibraryInfo *TLI,
3555 const DominatorTree *DT,
3556 AssumptionCache *AC,
3557 const Instruction *CxtI) {
3558 return ::SimplifyExtractValueInst(Agg, Idxs, Query(DL, TLI, DT, AC, CxtI),
3562 /// SimplifyExtractElementInst - Given operands for an ExtractElementInst, see if we
3563 /// can fold the result. If not, this returns null.
3564 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const Query &,
3566 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3567 if (auto *CIdx = dyn_cast<Constant>(Idx))
3568 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3570 // The index is not relevant if our vector is a splat.
3571 if (auto *Splat = CVec->getSplatValue())
3574 if (isa<UndefValue>(Vec))
3575 return UndefValue::get(Vec->getType()->getVectorElementType());
3578 // If extracting a specified index from the vector, see if we can recursively
3579 // find a previously computed scalar that was inserted into the vector.
3580 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3581 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3587 Value *llvm::SimplifyExtractElementInst(
3588 Value *Vec, Value *Idx, const DataLayout &DL, const TargetLibraryInfo *TLI,
3589 const DominatorTree *DT, AssumptionCache *AC, const Instruction *CxtI) {
3590 return ::SimplifyExtractElementInst(Vec, Idx, Query(DL, TLI, DT, AC, CxtI),
3594 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
3595 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
3596 // If all of the PHI's incoming values are the same then replace the PHI node
3597 // with the common value.
3598 Value *CommonValue = nullptr;
3599 bool HasUndefInput = false;
3600 for (Value *Incoming : PN->incoming_values()) {
3601 // If the incoming value is the phi node itself, it can safely be skipped.
3602 if (Incoming == PN) continue;
3603 if (isa<UndefValue>(Incoming)) {
3604 // Remember that we saw an undef value, but otherwise ignore them.
3605 HasUndefInput = true;
3608 if (CommonValue && Incoming != CommonValue)
3609 return nullptr; // Not the same, bail out.
3610 CommonValue = Incoming;
3613 // If CommonValue is null then all of the incoming values were either undef or
3614 // equal to the phi node itself.
3616 return UndefValue::get(PN->getType());
3618 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3619 // instruction, we cannot return X as the result of the PHI node unless it
3620 // dominates the PHI block.
3622 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3627 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
3628 if (Constant *C = dyn_cast<Constant>(Op))
3629 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI);
3634 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout &DL,
3635 const TargetLibraryInfo *TLI,
3636 const DominatorTree *DT, AssumptionCache *AC,
3637 const Instruction *CxtI) {
3638 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI),
3642 //=== Helper functions for higher up the class hierarchy.
3644 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
3645 /// fold the result. If not, this returns null.
3646 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3647 const Query &Q, unsigned MaxRecurse) {
3649 case Instruction::Add:
3650 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3652 case Instruction::FAdd:
3653 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3655 case Instruction::Sub:
3656 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3658 case Instruction::FSub:
3659 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3661 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
3662 case Instruction::FMul:
3663 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3664 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
3665 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
3666 case Instruction::FDiv:
3667 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3668 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
3669 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
3670 case Instruction::FRem:
3671 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3672 case Instruction::Shl:
3673 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3675 case Instruction::LShr:
3676 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3677 case Instruction::AShr:
3678 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3679 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
3680 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
3681 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
3683 if (Constant *CLHS = dyn_cast<Constant>(LHS))
3684 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
3685 Constant *COps[] = {CLHS, CRHS};
3686 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL,
3690 // If the operation is associative, try some generic simplifications.
3691 if (Instruction::isAssociative(Opcode))
3692 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
3695 // If the operation is with the result of a select instruction check whether
3696 // operating on either branch of the select always yields the same value.
3697 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3698 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
3701 // If the operation is with the result of a phi instruction, check whether
3702 // operating on all incoming values of the phi always yields the same value.
3703 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3704 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
3711 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can
3712 /// fold the result. If not, this returns null.
3713 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
3714 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
3715 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3716 const FastMathFlags &FMF, const Query &Q,
3717 unsigned MaxRecurse) {
3719 case Instruction::FAdd:
3720 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
3721 case Instruction::FSub:
3722 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
3723 case Instruction::FMul:
3724 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
3726 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
3730 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3731 const DataLayout &DL, const TargetLibraryInfo *TLI,
3732 const DominatorTree *DT, AssumptionCache *AC,
3733 const Instruction *CxtI) {
3734 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3738 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3739 const FastMathFlags &FMF, const DataLayout &DL,
3740 const TargetLibraryInfo *TLI,
3741 const DominatorTree *DT, AssumptionCache *AC,
3742 const Instruction *CxtI) {
3743 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI),
3747 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
3748 /// fold the result.
3749 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3750 const Query &Q, unsigned MaxRecurse) {
3751 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
3752 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3753 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3756 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3757 const DataLayout &DL, const TargetLibraryInfo *TLI,
3758 const DominatorTree *DT, AssumptionCache *AC,
3759 const Instruction *CxtI) {
3760 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3764 static bool IsIdempotent(Intrinsic::ID ID) {
3766 default: return false;
3768 // Unary idempotent: f(f(x)) = f(x)
3769 case Intrinsic::fabs:
3770 case Intrinsic::floor:
3771 case Intrinsic::ceil:
3772 case Intrinsic::trunc:
3773 case Intrinsic::rint:
3774 case Intrinsic::nearbyint:
3775 case Intrinsic::round:
3780 template <typename IterTy>
3781 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
3782 const Query &Q, unsigned MaxRecurse) {
3783 Intrinsic::ID IID = F->getIntrinsicID();
3784 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
3785 Type *ReturnType = F->getReturnType();
3788 if (NumOperands == 2) {
3789 Value *LHS = *ArgBegin;
3790 Value *RHS = *(ArgBegin + 1);
3791 if (IID == Intrinsic::usub_with_overflow ||
3792 IID == Intrinsic::ssub_with_overflow) {
3793 // X - X -> { 0, false }
3795 return Constant::getNullValue(ReturnType);
3797 // X - undef -> undef
3798 // undef - X -> undef
3799 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
3800 return UndefValue::get(ReturnType);
3803 if (IID == Intrinsic::uadd_with_overflow ||
3804 IID == Intrinsic::sadd_with_overflow) {
3805 // X + undef -> undef
3806 if (isa<UndefValue>(RHS))
3807 return UndefValue::get(ReturnType);
3810 if (IID == Intrinsic::umul_with_overflow ||
3811 IID == Intrinsic::smul_with_overflow) {
3812 // X * 0 -> { 0, false }
3813 if (match(RHS, m_Zero()))
3814 return Constant::getNullValue(ReturnType);
3816 // X * undef -> { 0, false }
3817 if (match(RHS, m_Undef()))
3818 return Constant::getNullValue(ReturnType);
3822 // Perform idempotent optimizations
3823 if (!IsIdempotent(IID))
3827 if (NumOperands == 1)
3828 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
3829 if (II->getIntrinsicID() == IID)
3835 template <typename IterTy>
3836 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
3837 const Query &Q, unsigned MaxRecurse) {
3838 Type *Ty = V->getType();
3839 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
3840 Ty = PTy->getElementType();
3841 FunctionType *FTy = cast<FunctionType>(Ty);
3843 // call undef -> undef
3844 if (isa<UndefValue>(V))
3845 return UndefValue::get(FTy->getReturnType());
3847 Function *F = dyn_cast<Function>(V);
3851 if (F->isIntrinsic())
3852 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
3855 if (!canConstantFoldCallTo(F))
3858 SmallVector<Constant *, 4> ConstantArgs;
3859 ConstantArgs.reserve(ArgEnd - ArgBegin);
3860 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
3861 Constant *C = dyn_cast<Constant>(*I);
3864 ConstantArgs.push_back(C);
3867 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
3870 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
3871 User::op_iterator ArgEnd, const DataLayout &DL,
3872 const TargetLibraryInfo *TLI, const DominatorTree *DT,
3873 AssumptionCache *AC, const Instruction *CxtI) {
3874 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI),
3878 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
3879 const DataLayout &DL, const TargetLibraryInfo *TLI,
3880 const DominatorTree *DT, AssumptionCache *AC,
3881 const Instruction *CxtI) {
3882 return ::SimplifyCall(V, Args.begin(), Args.end(),
3883 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3886 /// SimplifyInstruction - See if we can compute a simplified version of this
3887 /// instruction. If not, this returns null.
3888 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL,
3889 const TargetLibraryInfo *TLI,
3890 const DominatorTree *DT, AssumptionCache *AC) {
3893 switch (I->getOpcode()) {
3895 Result = ConstantFoldInstruction(I, DL, TLI);
3897 case Instruction::FAdd:
3898 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
3899 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3901 case Instruction::Add:
3902 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
3903 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3904 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3907 case Instruction::FSub:
3908 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
3909 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3911 case Instruction::Sub:
3912 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
3913 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3914 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3917 case Instruction::FMul:
3918 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
3919 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3921 case Instruction::Mul:
3923 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3925 case Instruction::SDiv:
3926 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3929 case Instruction::UDiv:
3930 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3933 case Instruction::FDiv:
3934 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
3935 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3937 case Instruction::SRem:
3938 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3941 case Instruction::URem:
3942 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3945 case Instruction::FRem:
3946 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
3947 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3949 case Instruction::Shl:
3950 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
3951 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3952 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3955 case Instruction::LShr:
3956 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
3957 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
3960 case Instruction::AShr:
3961 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
3962 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
3965 case Instruction::And:
3967 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3969 case Instruction::Or:
3971 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3973 case Instruction::Xor:
3975 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3977 case Instruction::ICmp:
3979 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0),
3980 I->getOperand(1), DL, TLI, DT, AC, I);
3982 case Instruction::FCmp:
3983 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
3984 I->getOperand(0), I->getOperand(1),
3985 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3987 case Instruction::Select:
3988 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
3989 I->getOperand(2), DL, TLI, DT, AC, I);
3991 case Instruction::GetElementPtr: {
3992 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
3993 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I);
3996 case Instruction::InsertValue: {
3997 InsertValueInst *IV = cast<InsertValueInst>(I);
3998 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
3999 IV->getInsertedValueOperand(),
4000 IV->getIndices(), DL, TLI, DT, AC, I);
4003 case Instruction::ExtractValue: {
4004 auto *EVI = cast<ExtractValueInst>(I);
4005 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4006 EVI->getIndices(), DL, TLI, DT, AC, I);
4009 case Instruction::ExtractElement: {
4010 auto *EEI = cast<ExtractElementInst>(I);
4011 Result = SimplifyExtractElementInst(
4012 EEI->getVectorOperand(), EEI->getIndexOperand(), DL, TLI, DT, AC, I);
4015 case Instruction::PHI:
4016 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I));
4018 case Instruction::Call: {
4019 CallSite CS(cast<CallInst>(I));
4020 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL,
4024 case Instruction::Trunc:
4026 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I);
4030 /// If called on unreachable code, the above logic may report that the
4031 /// instruction simplified to itself. Make life easier for users by
4032 /// detecting that case here, returning a safe value instead.
4033 return Result == I ? UndefValue::get(I->getType()) : Result;
4036 /// \brief Implementation of recursive simplification through an instructions
4039 /// This is the common implementation of the recursive simplification routines.
4040 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4041 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4042 /// instructions to process and attempt to simplify it using
4043 /// InstructionSimplify.
4045 /// This routine returns 'true' only when *it* simplifies something. The passed
4046 /// in simplified value does not count toward this.
4047 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4048 const TargetLibraryInfo *TLI,
4049 const DominatorTree *DT,
4050 AssumptionCache *AC) {
4051 bool Simplified = false;
4052 SmallSetVector<Instruction *, 8> Worklist;
4053 const DataLayout &DL = I->getModule()->getDataLayout();
4055 // If we have an explicit value to collapse to, do that round of the
4056 // simplification loop by hand initially.
4058 for (User *U : I->users())
4060 Worklist.insert(cast<Instruction>(U));
4062 // Replace the instruction with its simplified value.
4063 I->replaceAllUsesWith(SimpleV);
4065 // Gracefully handle edge cases where the instruction is not wired into any
4068 I->eraseFromParent();
4073 // Note that we must test the size on each iteration, the worklist can grow.
4074 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4077 // See if this instruction simplifies.
4078 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC);
4084 // Stash away all the uses of the old instruction so we can check them for
4085 // recursive simplifications after a RAUW. This is cheaper than checking all
4086 // uses of To on the recursive step in most cases.
4087 for (User *U : I->users())
4088 Worklist.insert(cast<Instruction>(U));
4090 // Replace the instruction with its simplified value.
4091 I->replaceAllUsesWith(SimpleV);
4093 // Gracefully handle edge cases where the instruction is not wired into any
4096 I->eraseFromParent();
4101 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4102 const TargetLibraryInfo *TLI,
4103 const DominatorTree *DT,
4104 AssumptionCache *AC) {
4105 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4108 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4109 const TargetLibraryInfo *TLI,
4110 const DominatorTree *DT,
4111 AssumptionCache *AC) {
4112 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4113 assert(SimpleV && "Must provide a simplified value.");
4114 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);