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/ConstantFolding.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/GlobalAlias.h"
31 #include "llvm/IR/Operator.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/IR/ValueHandle.h"
35 using namespace llvm::PatternMatch;
37 #define DEBUG_TYPE "instsimplify"
39 enum { RecursionLimit = 3 };
41 STATISTIC(NumExpand, "Number of expansions");
42 STATISTIC(NumReassoc, "Number of reassociations");
46 const TargetLibraryInfo *TLI;
47 const DominatorTree *DT;
49 Query(const DataLayout *DL, const TargetLibraryInfo *tli,
50 const DominatorTree *dt) : DL(DL), TLI(tli), DT(dt) {}
53 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
54 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
56 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
58 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
59 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
60 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned);
62 /// getFalse - For a boolean type, or a vector of boolean type, return false, or
63 /// a vector with every element false, as appropriate for the type.
64 static Constant *getFalse(Type *Ty) {
65 assert(Ty->getScalarType()->isIntegerTy(1) &&
66 "Expected i1 type or a vector of i1!");
67 return Constant::getNullValue(Ty);
70 /// getTrue - For a boolean type, or a vector of boolean type, return true, or
71 /// a vector with every element true, as appropriate for the type.
72 static Constant *getTrue(Type *Ty) {
73 assert(Ty->getScalarType()->isIntegerTy(1) &&
74 "Expected i1 type or a vector of i1!");
75 return Constant::getAllOnesValue(Ty);
78 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
79 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
81 CmpInst *Cmp = dyn_cast<CmpInst>(V);
84 CmpInst::Predicate CPred = Cmp->getPredicate();
85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
86 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
92 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
93 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
94 Instruction *I = dyn_cast<Instruction>(V);
96 // Arguments and constants dominate all instructions.
99 // If we are processing instructions (and/or basic blocks) that have not been
100 // fully added to a function, the parent nodes may still be null. Simply
101 // return the conservative answer in these cases.
102 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
105 // If we have a DominatorTree then do a precise test.
107 if (!DT->isReachableFromEntry(P->getParent()))
109 if (!DT->isReachableFromEntry(I->getParent()))
111 return DT->dominates(I, P);
114 // Otherwise, if the instruction is in the entry block, and is not an invoke,
115 // then it obviously dominates all phi nodes.
116 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
123 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning
124 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
125 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
126 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
127 /// Returns the simplified value, or null if no simplification was performed.
128 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
129 unsigned OpcToExpand, const Query &Q,
130 unsigned MaxRecurse) {
131 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
132 // Recursion is always used, so bail out at once if we already hit the limit.
136 // Check whether the expression has the form "(A op' B) op C".
137 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
138 if (Op0->getOpcode() == OpcodeToExpand) {
139 // It does! Try turning it into "(A op C) op' (B op C)".
140 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
141 // Do "A op C" and "B op C" both simplify?
142 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
143 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
144 // They do! Return "L op' R" if it simplifies or is already available.
145 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
146 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
147 && L == B && R == A)) {
151 // Otherwise return "L op' R" if it simplifies.
152 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
159 // Check whether the expression has the form "A op (B op' C)".
160 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
161 if (Op1->getOpcode() == OpcodeToExpand) {
162 // It does! Try turning it into "(A op B) op' (A op C)".
163 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
164 // Do "A op B" and "A op C" both simplify?
165 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
166 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
167 // They do! Return "L op' R" if it simplifies or is already available.
168 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
169 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
170 && L == C && R == B)) {
174 // Otherwise return "L op' R" if it simplifies.
175 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
185 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
186 /// operations. Returns the simpler value, or null if none was found.
187 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
188 const Query &Q, unsigned MaxRecurse) {
189 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
190 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
192 // Recursion is always used, so bail out at once if we already hit the limit.
196 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
197 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
199 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
200 if (Op0 && Op0->getOpcode() == Opcode) {
201 Value *A = Op0->getOperand(0);
202 Value *B = Op0->getOperand(1);
205 // Does "B op C" simplify?
206 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
207 // It does! Return "A op V" if it simplifies or is already available.
208 // If V equals B then "A op V" is just the LHS.
209 if (V == B) return LHS;
210 // Otherwise return "A op V" if it simplifies.
211 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
218 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
219 if (Op1 && Op1->getOpcode() == Opcode) {
221 Value *B = Op1->getOperand(0);
222 Value *C = Op1->getOperand(1);
224 // Does "A op B" simplify?
225 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
226 // It does! Return "V op C" if it simplifies or is already available.
227 // If V equals B then "V op C" is just the RHS.
228 if (V == B) return RHS;
229 // Otherwise return "V op C" if it simplifies.
230 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
237 // The remaining transforms require commutativity as well as associativity.
238 if (!Instruction::isCommutative(Opcode))
241 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
242 if (Op0 && Op0->getOpcode() == Opcode) {
243 Value *A = Op0->getOperand(0);
244 Value *B = Op0->getOperand(1);
247 // Does "C op A" simplify?
248 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
249 // It does! Return "V op B" if it simplifies or is already available.
250 // If V equals A then "V op B" is just the LHS.
251 if (V == A) return LHS;
252 // Otherwise return "V op B" if it simplifies.
253 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
260 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
261 if (Op1 && Op1->getOpcode() == Opcode) {
263 Value *B = Op1->getOperand(0);
264 Value *C = Op1->getOperand(1);
266 // Does "C op A" simplify?
267 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
268 // It does! Return "B op V" if it simplifies or is already available.
269 // If V equals C then "B op V" is just the RHS.
270 if (V == C) return RHS;
271 // Otherwise return "B op V" if it simplifies.
272 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
282 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
283 /// instruction as an operand, try to simplify the binop by seeing whether
284 /// evaluating it on both branches of the select results in the same value.
285 /// Returns the common value if so, otherwise returns null.
286 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
287 const Query &Q, unsigned MaxRecurse) {
288 // Recursion is always used, so bail out at once if we already hit the limit.
293 if (isa<SelectInst>(LHS)) {
294 SI = cast<SelectInst>(LHS);
296 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
297 SI = cast<SelectInst>(RHS);
300 // Evaluate the BinOp on the true and false branches of the select.
304 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
305 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
307 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
308 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
311 // If they simplified to the same value, then return the common value.
312 // If they both failed to simplify then return null.
316 // If one branch simplified to undef, return the other one.
317 if (TV && isa<UndefValue>(TV))
319 if (FV && isa<UndefValue>(FV))
322 // If applying the operation did not change the true and false select values,
323 // then the result of the binop is the select itself.
324 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
327 // If one branch simplified and the other did not, and the simplified
328 // value is equal to the unsimplified one, return the simplified value.
329 // For example, select (cond, X, X & Z) & Z -> X & Z.
330 if ((FV && !TV) || (TV && !FV)) {
331 // Check that the simplified value has the form "X op Y" where "op" is the
332 // same as the original operation.
333 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
334 if (Simplified && Simplified->getOpcode() == Opcode) {
335 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
336 // We already know that "op" is the same as for the simplified value. See
337 // if the operands match too. If so, return the simplified value.
338 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
339 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
340 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
341 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
342 Simplified->getOperand(1) == UnsimplifiedRHS)
344 if (Simplified->isCommutative() &&
345 Simplified->getOperand(1) == UnsimplifiedLHS &&
346 Simplified->getOperand(0) == UnsimplifiedRHS)
354 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
355 /// try to simplify the comparison by seeing whether both branches of the select
356 /// result in the same value. Returns the common value if so, otherwise returns
358 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
359 Value *RHS, const Query &Q,
360 unsigned MaxRecurse) {
361 // Recursion is always used, so bail out at once if we already hit the limit.
365 // Make sure the select is on the LHS.
366 if (!isa<SelectInst>(LHS)) {
368 Pred = CmpInst::getSwappedPredicate(Pred);
370 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
371 SelectInst *SI = cast<SelectInst>(LHS);
372 Value *Cond = SI->getCondition();
373 Value *TV = SI->getTrueValue();
374 Value *FV = SI->getFalseValue();
376 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
377 // Does "cmp TV, RHS" simplify?
378 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
380 // It not only simplified, it simplified to the select condition. Replace
382 TCmp = getTrue(Cond->getType());
384 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
385 // condition then we can replace it with 'true'. Otherwise give up.
386 if (!isSameCompare(Cond, Pred, TV, RHS))
388 TCmp = getTrue(Cond->getType());
391 // Does "cmp FV, RHS" simplify?
392 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
394 // It not only simplified, it simplified to the select condition. Replace
396 FCmp = getFalse(Cond->getType());
398 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
399 // condition then we can replace it with 'false'. Otherwise give up.
400 if (!isSameCompare(Cond, Pred, FV, RHS))
402 FCmp = getFalse(Cond->getType());
405 // If both sides simplified to the same value, then use it as the result of
406 // the original comparison.
410 // The remaining cases only make sense if the select condition has the same
411 // type as the result of the comparison, so bail out if this is not so.
412 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
414 // If the false value simplified to false, then the result of the compare
415 // is equal to "Cond && TCmp". This also catches the case when the false
416 // value simplified to false and the true value to true, returning "Cond".
417 if (match(FCmp, m_Zero()))
418 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
420 // If the true value simplified to true, then the result of the compare
421 // is equal to "Cond || FCmp".
422 if (match(TCmp, m_One()))
423 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
425 // Finally, if the false value simplified to true and the true value to
426 // false, then the result of the compare is equal to "!Cond".
427 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
429 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
436 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
437 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
438 /// it on the incoming phi values yields the same result for every value. If so
439 /// returns the common value, otherwise returns null.
440 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
441 const Query &Q, unsigned MaxRecurse) {
442 // Recursion is always used, so bail out at once if we already hit the limit.
447 if (isa<PHINode>(LHS)) {
448 PI = cast<PHINode>(LHS);
449 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
450 if (!ValueDominatesPHI(RHS, PI, Q.DT))
453 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
454 PI = cast<PHINode>(RHS);
455 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
456 if (!ValueDominatesPHI(LHS, PI, Q.DT))
460 // Evaluate the BinOp on the incoming phi values.
461 Value *CommonValue = nullptr;
462 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
463 Value *Incoming = PI->getIncomingValue(i);
464 // If the incoming value is the phi node itself, it can safely be skipped.
465 if (Incoming == PI) continue;
466 Value *V = PI == LHS ?
467 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
468 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
469 // If the operation failed to simplify, or simplified to a different value
470 // to previously, then give up.
471 if (!V || (CommonValue && V != CommonValue))
479 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
480 /// try to simplify the comparison by seeing whether comparing with all of the
481 /// incoming phi values yields the same result every time. If so returns the
482 /// common result, otherwise returns null.
483 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
484 const Query &Q, unsigned MaxRecurse) {
485 // Recursion is always used, so bail out at once if we already hit the limit.
489 // Make sure the phi is on the LHS.
490 if (!isa<PHINode>(LHS)) {
492 Pred = CmpInst::getSwappedPredicate(Pred);
494 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
495 PHINode *PI = cast<PHINode>(LHS);
497 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
498 if (!ValueDominatesPHI(RHS, PI, Q.DT))
501 // Evaluate the BinOp on the incoming phi values.
502 Value *CommonValue = nullptr;
503 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
504 Value *Incoming = PI->getIncomingValue(i);
505 // If the incoming value is the phi node itself, it can safely be skipped.
506 if (Incoming == PI) continue;
507 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
508 // If the operation failed to simplify, or simplified to a different value
509 // to previously, then give up.
510 if (!V || (CommonValue && V != CommonValue))
518 /// SimplifyAddInst - Given operands for an Add, see if we can
519 /// fold the result. If not, this returns null.
520 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
521 const Query &Q, unsigned MaxRecurse) {
522 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
523 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
524 Constant *Ops[] = { CLHS, CRHS };
525 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
529 // Canonicalize the constant to the RHS.
533 // X + undef -> undef
534 if (match(Op1, m_Undef()))
538 if (match(Op1, m_Zero()))
545 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
546 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
549 // X + ~X -> -1 since ~X = -X-1
550 if (match(Op0, m_Not(m_Specific(Op1))) ||
551 match(Op1, m_Not(m_Specific(Op0))))
552 return Constant::getAllOnesValue(Op0->getType());
555 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
556 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
559 // Try some generic simplifications for associative operations.
560 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
564 // Threading Add over selects and phi nodes is pointless, so don't bother.
565 // Threading over the select in "A + select(cond, B, C)" means evaluating
566 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
567 // only if B and C are equal. If B and C are equal then (since we assume
568 // that operands have already been simplified) "select(cond, B, C)" should
569 // have been simplified to the common value of B and C already. Analysing
570 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
571 // for threading over phi nodes.
576 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
577 const DataLayout *DL, const TargetLibraryInfo *TLI,
578 const DominatorTree *DT) {
579 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
583 /// \brief Compute the base pointer and cumulative constant offsets for V.
585 /// This strips all constant offsets off of V, leaving it the base pointer, and
586 /// accumulates the total constant offset applied in the returned constant. It
587 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
588 /// no constant offsets applied.
590 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
591 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
593 static Constant *stripAndComputeConstantOffsets(const DataLayout *DL,
595 bool AllowNonInbounds = false) {
596 assert(V->getType()->getScalarType()->isPointerTy());
598 // Without DataLayout, just be conservative for now. Theoretically, more could
599 // be done in this case.
601 return ConstantInt::get(IntegerType::get(V->getContext(), 64), 0);
603 Type *IntPtrTy = DL->getIntPtrType(V->getType())->getScalarType();
604 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
606 // Even though we don't look through PHI nodes, we could be called on an
607 // instruction in an unreachable block, which may be on a cycle.
608 SmallPtrSet<Value *, 4> Visited;
611 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
612 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
613 !GEP->accumulateConstantOffset(*DL, Offset))
615 V = GEP->getPointerOperand();
616 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
617 V = cast<Operator>(V)->getOperand(0);
618 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
619 if (GA->mayBeOverridden())
621 V = GA->getAliasee();
625 assert(V->getType()->getScalarType()->isPointerTy() &&
626 "Unexpected operand type!");
627 } while (Visited.insert(V));
629 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
630 if (V->getType()->isVectorTy())
631 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
636 /// \brief Compute the constant difference between two pointer values.
637 /// If the difference is not a constant, returns zero.
638 static Constant *computePointerDifference(const DataLayout *DL,
639 Value *LHS, Value *RHS) {
640 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
641 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
643 // If LHS and RHS are not related via constant offsets to the same base
644 // value, there is nothing we can do here.
648 // Otherwise, the difference of LHS - RHS can be computed as:
650 // = (LHSOffset + Base) - (RHSOffset + Base)
651 // = LHSOffset - RHSOffset
652 return ConstantExpr::getSub(LHSOffset, RHSOffset);
655 /// SimplifySubInst - Given operands for a Sub, see if we can
656 /// fold the result. If not, this returns null.
657 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
658 const Query &Q, unsigned MaxRecurse) {
659 if (Constant *CLHS = dyn_cast<Constant>(Op0))
660 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
661 Constant *Ops[] = { CLHS, CRHS };
662 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
666 // X - undef -> undef
667 // undef - X -> undef
668 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
669 return UndefValue::get(Op0->getType());
672 if (match(Op1, m_Zero()))
677 return Constant::getNullValue(Op0->getType());
679 // X - (0 - Y) -> X if the second sub is NUW.
680 // If Y != 0, 0 - Y is a poison value.
681 // If Y == 0, 0 - Y simplifies to 0.
682 if (BinaryOperator::isNeg(Op1)) {
683 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
684 assert(BO->getOpcode() == Instruction::Sub &&
685 "Expected a subtraction operator!");
686 if (BO->hasNoUnsignedWrap())
691 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
692 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
693 Value *X = nullptr, *Y = nullptr, *Z = Op1;
694 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
695 // See if "V === Y - Z" simplifies.
696 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
697 // It does! Now see if "X + V" simplifies.
698 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
699 // It does, we successfully reassociated!
703 // See if "V === X - Z" simplifies.
704 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
705 // It does! Now see if "Y + V" simplifies.
706 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
707 // It does, we successfully reassociated!
713 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
714 // For example, X - (X + 1) -> -1
716 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
717 // See if "V === X - Y" simplifies.
718 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
719 // It does! Now see if "V - Z" simplifies.
720 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
721 // It does, we successfully reassociated!
725 // See if "V === X - Z" simplifies.
726 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
727 // It does! Now see if "V - Y" simplifies.
728 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
729 // It does, we successfully reassociated!
735 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
736 // For example, X - (X - Y) -> Y.
738 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
739 // See if "V === Z - X" simplifies.
740 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
741 // It does! Now see if "V + Y" simplifies.
742 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
743 // It does, we successfully reassociated!
748 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
749 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
750 match(Op1, m_Trunc(m_Value(Y))))
751 if (X->getType() == Y->getType())
752 // See if "V === X - Y" simplifies.
753 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
754 // It does! Now see if "trunc V" simplifies.
755 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
756 // It does, return the simplified "trunc V".
759 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
760 if (match(Op0, m_PtrToInt(m_Value(X))) &&
761 match(Op1, m_PtrToInt(m_Value(Y))))
762 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
763 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
766 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
767 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
770 // Threading Sub over selects and phi nodes is pointless, so don't bother.
771 // Threading over the select in "A - select(cond, B, C)" means evaluating
772 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
773 // only if B and C are equal. If B and C are equal then (since we assume
774 // that operands have already been simplified) "select(cond, B, C)" should
775 // have been simplified to the common value of B and C already. Analysing
776 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
777 // for threading over phi nodes.
782 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
783 const DataLayout *DL, const TargetLibraryInfo *TLI,
784 const DominatorTree *DT) {
785 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
789 /// Given operands for an FAdd, see if we can fold the result. If not, this
791 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
792 const Query &Q, unsigned MaxRecurse) {
793 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
794 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
795 Constant *Ops[] = { CLHS, CRHS };
796 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(),
800 // Canonicalize the constant to the RHS.
805 if (match(Op1, m_NegZero()))
808 // fadd X, 0 ==> X, when we know X is not -0
809 if (match(Op1, m_Zero()) &&
810 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
813 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
814 // where nnan and ninf have to occur at least once somewhere in this
816 Value *SubOp = nullptr;
817 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
819 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
822 Instruction *FSub = cast<Instruction>(SubOp);
823 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
824 (FMF.noInfs() || FSub->hasNoInfs()))
825 return Constant::getNullValue(Op0->getType());
831 /// Given operands for an FSub, see if we can fold the result. If not, this
833 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
834 const Query &Q, unsigned MaxRecurse) {
835 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
836 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
837 Constant *Ops[] = { CLHS, CRHS };
838 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(),
844 if (match(Op1, m_Zero()))
847 // fsub X, -0 ==> X, when we know X is not -0
848 if (match(Op1, m_NegZero()) &&
849 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
852 // fsub 0, (fsub -0.0, X) ==> X
854 if (match(Op0, m_AnyZero())) {
855 if (match(Op1, m_FSub(m_NegZero(), m_Value(X))))
857 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
861 // fsub nnan ninf x, x ==> 0.0
862 if (FMF.noNaNs() && FMF.noInfs() && Op0 == Op1)
863 return Constant::getNullValue(Op0->getType());
868 /// Given the operands for an FMul, see if we can fold the result
869 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
872 unsigned MaxRecurse) {
873 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
874 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
875 Constant *Ops[] = { CLHS, CRHS };
876 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(),
880 // Canonicalize the constant to the RHS.
885 if (match(Op1, m_FPOne()))
888 // fmul nnan nsz X, 0 ==> 0
889 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
895 /// SimplifyMulInst - Given operands for a Mul, see if we can
896 /// fold the result. If not, this returns null.
897 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
898 unsigned MaxRecurse) {
899 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
900 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
901 Constant *Ops[] = { CLHS, CRHS };
902 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(),
906 // Canonicalize the constant to the RHS.
911 if (match(Op1, m_Undef()))
912 return Constant::getNullValue(Op0->getType());
915 if (match(Op1, m_Zero()))
919 if (match(Op1, m_One()))
922 // (X / Y) * Y -> X if the division is exact.
924 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
925 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
929 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
930 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
933 // Try some generic simplifications for associative operations.
934 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
938 // Mul distributes over Add. Try some generic simplifications based on this.
939 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
943 // If the operation is with the result of a select instruction, check whether
944 // operating on either branch of the select always yields the same value.
945 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
946 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
950 // If the operation is with the result of a phi instruction, check whether
951 // operating on all incoming values of the phi always yields the same value.
952 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
953 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
960 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
961 const DataLayout *DL, const TargetLibraryInfo *TLI,
962 const DominatorTree *DT) {
963 return ::SimplifyFAddInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
966 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
967 const DataLayout *DL, const TargetLibraryInfo *TLI,
968 const DominatorTree *DT) {
969 return ::SimplifyFSubInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
972 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1,
974 const DataLayout *DL,
975 const TargetLibraryInfo *TLI,
976 const DominatorTree *DT) {
977 return ::SimplifyFMulInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
980 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout *DL,
981 const TargetLibraryInfo *TLI,
982 const DominatorTree *DT) {
983 return ::SimplifyMulInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
986 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
987 /// fold the result. If not, this returns null.
988 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
989 const Query &Q, unsigned MaxRecurse) {
990 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
991 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
992 Constant *Ops[] = { C0, C1 };
993 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
997 bool isSigned = Opcode == Instruction::SDiv;
999 // X / undef -> undef
1000 if (match(Op1, m_Undef()))
1004 if (match(Op0, m_Undef()))
1005 return Constant::getNullValue(Op0->getType());
1007 // 0 / X -> 0, we don't need to preserve faults!
1008 if (match(Op0, m_Zero()))
1012 if (match(Op1, m_One()))
1015 if (Op0->getType()->isIntegerTy(1))
1016 // It can't be division by zero, hence it must be division by one.
1021 return ConstantInt::get(Op0->getType(), 1);
1023 // (X * Y) / Y -> X if the multiplication does not overflow.
1024 Value *X = nullptr, *Y = nullptr;
1025 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1026 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1027 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1028 // If the Mul knows it does not overflow, then we are good to go.
1029 if ((isSigned && Mul->hasNoSignedWrap()) ||
1030 (!isSigned && Mul->hasNoUnsignedWrap()))
1032 // If X has the form X = A / Y then X * Y cannot overflow.
1033 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1034 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1038 // (X rem Y) / Y -> 0
1039 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1040 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1041 return Constant::getNullValue(Op0->getType());
1043 // If the operation is with the result of a select instruction, check whether
1044 // operating on either branch of the select always yields the same value.
1045 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1046 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1049 // If the operation is with the result of a phi instruction, check whether
1050 // operating on all incoming values of the phi always yields the same value.
1051 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1052 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1058 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1059 /// fold the result. If not, this returns null.
1060 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1061 unsigned MaxRecurse) {
1062 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1068 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1069 const TargetLibraryInfo *TLI,
1070 const DominatorTree *DT) {
1071 return ::SimplifySDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1074 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1075 /// fold the result. If not, this returns null.
1076 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1077 unsigned MaxRecurse) {
1078 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1084 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1085 const TargetLibraryInfo *TLI,
1086 const DominatorTree *DT) {
1087 return ::SimplifyUDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1090 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, const Query &Q,
1092 // undef / X -> undef (the undef could be a snan).
1093 if (match(Op0, m_Undef()))
1096 // X / undef -> undef
1097 if (match(Op1, m_Undef()))
1103 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1104 const TargetLibraryInfo *TLI,
1105 const DominatorTree *DT) {
1106 return ::SimplifyFDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1109 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1110 /// fold the result. If not, this returns null.
1111 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1112 const Query &Q, unsigned MaxRecurse) {
1113 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1114 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1115 Constant *Ops[] = { C0, C1 };
1116 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1120 // X % undef -> undef
1121 if (match(Op1, m_Undef()))
1125 if (match(Op0, m_Undef()))
1126 return Constant::getNullValue(Op0->getType());
1128 // 0 % X -> 0, we don't need to preserve faults!
1129 if (match(Op0, m_Zero()))
1132 // X % 0 -> undef, we don't need to preserve faults!
1133 if (match(Op1, m_Zero()))
1134 return UndefValue::get(Op0->getType());
1137 if (match(Op1, m_One()))
1138 return Constant::getNullValue(Op0->getType());
1140 if (Op0->getType()->isIntegerTy(1))
1141 // It can't be remainder by zero, hence it must be remainder by one.
1142 return Constant::getNullValue(Op0->getType());
1146 return Constant::getNullValue(Op0->getType());
1148 // If the operation is with the result of a select instruction, check whether
1149 // operating on either branch of the select always yields the same value.
1150 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1151 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1154 // If the operation is with the result of a phi instruction, check whether
1155 // operating on all incoming values of the phi always yields the same value.
1156 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1157 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1163 /// SimplifySRemInst - Given operands for an SRem, see if we can
1164 /// fold the result. If not, this returns null.
1165 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1166 unsigned MaxRecurse) {
1167 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1173 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1174 const TargetLibraryInfo *TLI,
1175 const DominatorTree *DT) {
1176 return ::SimplifySRemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1179 /// SimplifyURemInst - Given operands for a URem, see if we can
1180 /// fold the result. If not, this returns null.
1181 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1182 unsigned MaxRecurse) {
1183 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1189 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1190 const TargetLibraryInfo *TLI,
1191 const DominatorTree *DT) {
1192 return ::SimplifyURemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1195 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, const Query &,
1197 // undef % X -> undef (the undef could be a snan).
1198 if (match(Op0, m_Undef()))
1201 // X % undef -> undef
1202 if (match(Op1, m_Undef()))
1208 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1209 const TargetLibraryInfo *TLI,
1210 const DominatorTree *DT) {
1211 return ::SimplifyFRemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1214 /// isUndefShift - Returns true if a shift by \c Amount always yields undef.
1215 static bool isUndefShift(Value *Amount) {
1216 Constant *C = dyn_cast<Constant>(Amount);
1220 // X shift by undef -> undef because it may shift by the bitwidth.
1221 if (isa<UndefValue>(C))
1224 // Shifting by the bitwidth or more is undefined.
1225 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1226 if (CI->getValue().getLimitedValue() >=
1227 CI->getType()->getScalarSizeInBits())
1230 // If all lanes of a vector shift are undefined the whole shift is.
1231 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1232 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1233 if (!isUndefShift(C->getAggregateElement(I)))
1241 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1242 /// fold the result. If not, this returns null.
1243 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1244 const Query &Q, unsigned MaxRecurse) {
1245 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1246 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1247 Constant *Ops[] = { C0, C1 };
1248 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1252 // 0 shift by X -> 0
1253 if (match(Op0, m_Zero()))
1256 // X shift by 0 -> X
1257 if (match(Op1, m_Zero()))
1260 // Fold undefined shifts.
1261 if (isUndefShift(Op1))
1262 return UndefValue::get(Op0->getType());
1264 // If the operation is with the result of a select instruction, check whether
1265 // operating on either branch of the select always yields the same value.
1266 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1267 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1270 // If the operation is with the result of a phi instruction, check whether
1271 // operating on all incoming values of the phi always yields the same value.
1272 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1273 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1279 /// SimplifyShlInst - Given operands for an Shl, see if we can
1280 /// fold the result. If not, this returns null.
1281 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1282 const Query &Q, unsigned MaxRecurse) {
1283 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1287 if (match(Op0, m_Undef()))
1288 return Constant::getNullValue(Op0->getType());
1290 // (X >> A) << A -> X
1292 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1297 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1298 const DataLayout *DL, const TargetLibraryInfo *TLI,
1299 const DominatorTree *DT) {
1300 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
1304 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1305 /// fold the result. If not, this returns null.
1306 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1307 const Query &Q, unsigned MaxRecurse) {
1308 if (Value *V = SimplifyShift(Instruction::LShr, Op0, Op1, Q, MaxRecurse))
1313 return Constant::getNullValue(Op0->getType());
1316 if (match(Op0, m_Undef()))
1317 return Constant::getNullValue(Op0->getType());
1319 // (X << A) >> A -> X
1321 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1322 cast<OverflowingBinaryOperator>(Op0)->hasNoUnsignedWrap())
1328 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1329 const DataLayout *DL,
1330 const TargetLibraryInfo *TLI,
1331 const DominatorTree *DT) {
1332 return ::SimplifyLShrInst(Op0, Op1, isExact, Query (DL, TLI, DT),
1336 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1337 /// fold the result. If not, this returns null.
1338 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1339 const Query &Q, unsigned MaxRecurse) {
1340 if (Value *V = SimplifyShift(Instruction::AShr, Op0, Op1, Q, MaxRecurse))
1345 return Constant::getNullValue(Op0->getType());
1347 // all ones >>a X -> all ones
1348 if (match(Op0, m_AllOnes()))
1351 // undef >>a X -> all ones
1352 if (match(Op0, m_Undef()))
1353 return Constant::getAllOnesValue(Op0->getType());
1355 // (X << A) >> A -> X
1357 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1358 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap())
1361 // Arithmetic shifting an all-sign-bit value is a no-op.
1362 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL);
1363 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1369 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1370 const DataLayout *DL,
1371 const TargetLibraryInfo *TLI,
1372 const DominatorTree *DT) {
1373 return ::SimplifyAShrInst(Op0, Op1, isExact, Query (DL, TLI, DT),
1377 /// SimplifyAndInst - Given operands for an And, see if we can
1378 /// fold the result. If not, this returns null.
1379 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1380 unsigned MaxRecurse) {
1381 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1382 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1383 Constant *Ops[] = { CLHS, CRHS };
1384 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1388 // Canonicalize the constant to the RHS.
1389 std::swap(Op0, Op1);
1393 if (match(Op1, m_Undef()))
1394 return Constant::getNullValue(Op0->getType());
1401 if (match(Op1, m_Zero()))
1405 if (match(Op1, m_AllOnes()))
1408 // A & ~A = ~A & A = 0
1409 if (match(Op0, m_Not(m_Specific(Op1))) ||
1410 match(Op1, m_Not(m_Specific(Op0))))
1411 return Constant::getNullValue(Op0->getType());
1414 Value *A = nullptr, *B = nullptr;
1415 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1416 (A == Op1 || B == Op1))
1420 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1421 (A == Op0 || B == Op0))
1424 // A & (-A) = A if A is a power of two or zero.
1425 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1426 match(Op1, m_Neg(m_Specific(Op0)))) {
1427 if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/true))
1429 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true))
1433 // Try some generic simplifications for associative operations.
1434 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1438 // And distributes over Or. Try some generic simplifications based on this.
1439 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1443 // And distributes over Xor. Try some generic simplifications based on this.
1444 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1448 // If the operation is with the result of a select instruction, check whether
1449 // operating on either branch of the select always yields the same value.
1450 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1451 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1455 // If the operation is with the result of a phi instruction, check whether
1456 // operating on all incoming values of the phi always yields the same value.
1457 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1458 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1465 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *DL,
1466 const TargetLibraryInfo *TLI,
1467 const DominatorTree *DT) {
1468 return ::SimplifyAndInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1471 /// SimplifyOrInst - Given operands for an Or, see if we can
1472 /// fold the result. If not, this returns null.
1473 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1474 unsigned MaxRecurse) {
1475 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1476 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1477 Constant *Ops[] = { CLHS, CRHS };
1478 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1482 // Canonicalize the constant to the RHS.
1483 std::swap(Op0, Op1);
1487 if (match(Op1, m_Undef()))
1488 return Constant::getAllOnesValue(Op0->getType());
1495 if (match(Op1, m_Zero()))
1499 if (match(Op1, m_AllOnes()))
1502 // A | ~A = ~A | A = -1
1503 if (match(Op0, m_Not(m_Specific(Op1))) ||
1504 match(Op1, m_Not(m_Specific(Op0))))
1505 return Constant::getAllOnesValue(Op0->getType());
1508 Value *A = nullptr, *B = nullptr;
1509 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1510 (A == Op1 || B == Op1))
1514 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1515 (A == Op0 || B == Op0))
1518 // ~(A & ?) | A = -1
1519 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1520 (A == Op1 || B == Op1))
1521 return Constant::getAllOnesValue(Op1->getType());
1523 // A | ~(A & ?) = -1
1524 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1525 (A == Op0 || B == Op0))
1526 return Constant::getAllOnesValue(Op0->getType());
1528 // Try some generic simplifications for associative operations.
1529 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1533 // Or distributes over And. Try some generic simplifications based on this.
1534 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1538 // If the operation is with the result of a select instruction, check whether
1539 // operating on either branch of the select always yields the same value.
1540 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1541 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1546 Value *C = nullptr, *D = nullptr;
1547 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1548 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1549 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1550 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1551 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1552 // (A & C1)|(B & C2)
1553 // If we have: ((V + N) & C1) | (V & C2)
1554 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1555 // replace with V+N.
1557 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1558 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1559 // Add commutes, try both ways.
1560 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1562 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1565 // Or commutes, try both ways.
1566 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1567 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1568 // Add commutes, try both ways.
1569 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1571 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1577 // If the operation is with the result of a phi instruction, check whether
1578 // operating on all incoming values of the phi always yields the same value.
1579 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1580 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1586 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *DL,
1587 const TargetLibraryInfo *TLI,
1588 const DominatorTree *DT) {
1589 return ::SimplifyOrInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1592 /// SimplifyXorInst - Given operands for a Xor, see if we can
1593 /// fold the result. If not, this returns null.
1594 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1595 unsigned MaxRecurse) {
1596 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1597 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1598 Constant *Ops[] = { CLHS, CRHS };
1599 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1603 // Canonicalize the constant to the RHS.
1604 std::swap(Op0, Op1);
1607 // A ^ undef -> undef
1608 if (match(Op1, m_Undef()))
1612 if (match(Op1, m_Zero()))
1617 return Constant::getNullValue(Op0->getType());
1619 // A ^ ~A = ~A ^ A = -1
1620 if (match(Op0, m_Not(m_Specific(Op1))) ||
1621 match(Op1, m_Not(m_Specific(Op0))))
1622 return Constant::getAllOnesValue(Op0->getType());
1624 // Try some generic simplifications for associative operations.
1625 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1629 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1630 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1631 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1632 // only if B and C are equal. If B and C are equal then (since we assume
1633 // that operands have already been simplified) "select(cond, B, C)" should
1634 // have been simplified to the common value of B and C already. Analysing
1635 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1636 // for threading over phi nodes.
1641 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *DL,
1642 const TargetLibraryInfo *TLI,
1643 const DominatorTree *DT) {
1644 return ::SimplifyXorInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1647 static Type *GetCompareTy(Value *Op) {
1648 return CmpInst::makeCmpResultType(Op->getType());
1651 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1652 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1653 /// otherwise return null. Helper function for analyzing max/min idioms.
1654 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1655 Value *LHS, Value *RHS) {
1656 SelectInst *SI = dyn_cast<SelectInst>(V);
1659 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1662 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1663 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1665 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1666 LHS == CmpRHS && RHS == CmpLHS)
1671 // A significant optimization not implemented here is assuming that alloca
1672 // addresses are not equal to incoming argument values. They don't *alias*,
1673 // as we say, but that doesn't mean they aren't equal, so we take a
1674 // conservative approach.
1676 // This is inspired in part by C++11 5.10p1:
1677 // "Two pointers of the same type compare equal if and only if they are both
1678 // null, both point to the same function, or both represent the same
1681 // This is pretty permissive.
1683 // It's also partly due to C11 6.5.9p6:
1684 // "Two pointers compare equal if and only if both are null pointers, both are
1685 // pointers to the same object (including a pointer to an object and a
1686 // subobject at its beginning) or function, both are pointers to one past the
1687 // last element of the same array object, or one is a pointer to one past the
1688 // end of one array object and the other is a pointer to the start of a
1689 // different array object that happens to immediately follow the first array
1690 // object in the address space.)
1692 // C11's version is more restrictive, however there's no reason why an argument
1693 // couldn't be a one-past-the-end value for a stack object in the caller and be
1694 // equal to the beginning of a stack object in the callee.
1696 // If the C and C++ standards are ever made sufficiently restrictive in this
1697 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1698 // this optimization.
1699 static Constant *computePointerICmp(const DataLayout *DL,
1700 const TargetLibraryInfo *TLI,
1701 CmpInst::Predicate Pred,
1702 Value *LHS, Value *RHS) {
1703 // First, skip past any trivial no-ops.
1704 LHS = LHS->stripPointerCasts();
1705 RHS = RHS->stripPointerCasts();
1707 // A non-null pointer is not equal to a null pointer.
1708 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) &&
1709 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1710 return ConstantInt::get(GetCompareTy(LHS),
1711 !CmpInst::isTrueWhenEqual(Pred));
1713 // We can only fold certain predicates on pointer comparisons.
1718 // Equality comaprisons are easy to fold.
1719 case CmpInst::ICMP_EQ:
1720 case CmpInst::ICMP_NE:
1723 // We can only handle unsigned relational comparisons because 'inbounds' on
1724 // a GEP only protects against unsigned wrapping.
1725 case CmpInst::ICMP_UGT:
1726 case CmpInst::ICMP_UGE:
1727 case CmpInst::ICMP_ULT:
1728 case CmpInst::ICMP_ULE:
1729 // However, we have to switch them to their signed variants to handle
1730 // negative indices from the base pointer.
1731 Pred = ICmpInst::getSignedPredicate(Pred);
1735 // Strip off any constant offsets so that we can reason about them.
1736 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
1737 // here and compare base addresses like AliasAnalysis does, however there are
1738 // numerous hazards. AliasAnalysis and its utilities rely on special rules
1739 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
1740 // doesn't need to guarantee pointer inequality when it says NoAlias.
1741 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
1742 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
1744 // If LHS and RHS are related via constant offsets to the same base
1745 // value, we can replace it with an icmp which just compares the offsets.
1747 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
1749 // Various optimizations for (in)equality comparisons.
1750 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
1751 // Different non-empty allocations that exist at the same time have
1752 // different addresses (if the program can tell). Global variables always
1753 // exist, so they always exist during the lifetime of each other and all
1754 // allocas. Two different allocas usually have different addresses...
1756 // However, if there's an @llvm.stackrestore dynamically in between two
1757 // allocas, they may have the same address. It's tempting to reduce the
1758 // scope of the problem by only looking at *static* allocas here. That would
1759 // cover the majority of allocas while significantly reducing the likelihood
1760 // of having an @llvm.stackrestore pop up in the middle. However, it's not
1761 // actually impossible for an @llvm.stackrestore to pop up in the middle of
1762 // an entry block. Also, if we have a block that's not attached to a
1763 // function, we can't tell if it's "static" under the current definition.
1764 // Theoretically, this problem could be fixed by creating a new kind of
1765 // instruction kind specifically for static allocas. Such a new instruction
1766 // could be required to be at the top of the entry block, thus preventing it
1767 // from being subject to a @llvm.stackrestore. Instcombine could even
1768 // convert regular allocas into these special allocas. It'd be nifty.
1769 // However, until then, this problem remains open.
1771 // So, we'll assume that two non-empty allocas have different addresses
1774 // With all that, if the offsets are within the bounds of their allocations
1775 // (and not one-past-the-end! so we can't use inbounds!), and their
1776 // allocations aren't the same, the pointers are not equal.
1778 // Note that it's not necessary to check for LHS being a global variable
1779 // address, due to canonicalization and constant folding.
1780 if (isa<AllocaInst>(LHS) &&
1781 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
1782 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
1783 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
1784 uint64_t LHSSize, RHSSize;
1785 if (LHSOffsetCI && RHSOffsetCI &&
1786 getObjectSize(LHS, LHSSize, DL, TLI) &&
1787 getObjectSize(RHS, RHSSize, DL, TLI)) {
1788 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
1789 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
1790 if (!LHSOffsetValue.isNegative() &&
1791 !RHSOffsetValue.isNegative() &&
1792 LHSOffsetValue.ult(LHSSize) &&
1793 RHSOffsetValue.ult(RHSSize)) {
1794 return ConstantInt::get(GetCompareTy(LHS),
1795 !CmpInst::isTrueWhenEqual(Pred));
1799 // Repeat the above check but this time without depending on DataLayout
1800 // or being able to compute a precise size.
1801 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
1802 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
1803 LHSOffset->isNullValue() &&
1804 RHSOffset->isNullValue())
1805 return ConstantInt::get(GetCompareTy(LHS),
1806 !CmpInst::isTrueWhenEqual(Pred));
1809 // Even if an non-inbounds GEP occurs along the path we can still optimize
1810 // equality comparisons concerning the result. We avoid walking the whole
1811 // chain again by starting where the last calls to
1812 // stripAndComputeConstantOffsets left off and accumulate the offsets.
1813 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
1814 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
1816 return ConstantExpr::getICmp(Pred,
1817 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
1818 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
1825 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
1826 /// fold the result. If not, this returns null.
1827 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
1828 const Query &Q, unsigned MaxRecurse) {
1829 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
1830 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
1832 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
1833 if (Constant *CRHS = dyn_cast<Constant>(RHS))
1834 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
1836 // If we have a constant, make sure it is on the RHS.
1837 std::swap(LHS, RHS);
1838 Pred = CmpInst::getSwappedPredicate(Pred);
1841 Type *ITy = GetCompareTy(LHS); // The return type.
1842 Type *OpTy = LHS->getType(); // The operand type.
1844 // icmp X, X -> true/false
1845 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
1846 // because X could be 0.
1847 if (LHS == RHS || isa<UndefValue>(RHS))
1848 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
1850 // Special case logic when the operands have i1 type.
1851 if (OpTy->getScalarType()->isIntegerTy(1)) {
1854 case ICmpInst::ICMP_EQ:
1856 if (match(RHS, m_One()))
1859 case ICmpInst::ICMP_NE:
1861 if (match(RHS, m_Zero()))
1864 case ICmpInst::ICMP_UGT:
1866 if (match(RHS, m_Zero()))
1869 case ICmpInst::ICMP_UGE:
1871 if (match(RHS, m_One()))
1874 case ICmpInst::ICMP_SLT:
1876 if (match(RHS, m_Zero()))
1879 case ICmpInst::ICMP_SLE:
1881 if (match(RHS, m_One()))
1887 // If we are comparing with zero then try hard since this is a common case.
1888 if (match(RHS, m_Zero())) {
1889 bool LHSKnownNonNegative, LHSKnownNegative;
1891 default: llvm_unreachable("Unknown ICmp predicate!");
1892 case ICmpInst::ICMP_ULT:
1893 return getFalse(ITy);
1894 case ICmpInst::ICMP_UGE:
1895 return getTrue(ITy);
1896 case ICmpInst::ICMP_EQ:
1897 case ICmpInst::ICMP_ULE:
1898 if (isKnownNonZero(LHS, Q.DL))
1899 return getFalse(ITy);
1901 case ICmpInst::ICMP_NE:
1902 case ICmpInst::ICMP_UGT:
1903 if (isKnownNonZero(LHS, Q.DL))
1904 return getTrue(ITy);
1906 case ICmpInst::ICMP_SLT:
1907 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1908 if (LHSKnownNegative)
1909 return getTrue(ITy);
1910 if (LHSKnownNonNegative)
1911 return getFalse(ITy);
1913 case ICmpInst::ICMP_SLE:
1914 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1915 if (LHSKnownNegative)
1916 return getTrue(ITy);
1917 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL))
1918 return getFalse(ITy);
1920 case ICmpInst::ICMP_SGE:
1921 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1922 if (LHSKnownNegative)
1923 return getFalse(ITy);
1924 if (LHSKnownNonNegative)
1925 return getTrue(ITy);
1927 case ICmpInst::ICMP_SGT:
1928 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1929 if (LHSKnownNegative)
1930 return getFalse(ITy);
1931 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL))
1932 return getTrue(ITy);
1937 // See if we are doing a comparison with a constant integer.
1938 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1939 // Rule out tautological comparisons (eg., ult 0 or uge 0).
1940 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
1941 if (RHS_CR.isEmptySet())
1942 return ConstantInt::getFalse(CI->getContext());
1943 if (RHS_CR.isFullSet())
1944 return ConstantInt::getTrue(CI->getContext());
1946 // Many binary operators with constant RHS have easy to compute constant
1947 // range. Use them to check whether the comparison is a tautology.
1948 unsigned Width = CI->getBitWidth();
1949 APInt Lower = APInt(Width, 0);
1950 APInt Upper = APInt(Width, 0);
1952 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
1953 // 'urem x, CI2' produces [0, CI2).
1954 Upper = CI2->getValue();
1955 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
1956 // 'srem x, CI2' produces (-|CI2|, |CI2|).
1957 Upper = CI2->getValue().abs();
1958 Lower = (-Upper) + 1;
1959 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
1960 // 'udiv CI2, x' produces [0, CI2].
1961 Upper = CI2->getValue() + 1;
1962 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
1963 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
1964 APInt NegOne = APInt::getAllOnesValue(Width);
1966 Upper = NegOne.udiv(CI2->getValue()) + 1;
1967 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) {
1968 if (CI2->isMinSignedValue()) {
1969 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
1970 Lower = CI2->getValue();
1971 Upper = Lower.lshr(1) + 1;
1973 // 'sdiv CI2, x' produces [-|CI2|, |CI2|].
1974 Upper = CI2->getValue().abs() + 1;
1975 Lower = (-Upper) + 1;
1977 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
1978 APInt IntMin = APInt::getSignedMinValue(Width);
1979 APInt IntMax = APInt::getSignedMaxValue(Width);
1980 APInt Val = CI2->getValue();
1981 if (Val.isAllOnesValue()) {
1982 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
1983 // where CI2 != -1 and CI2 != 0 and CI2 != 1
1986 } else if (Val.countLeadingZeros() < Width - 1) {
1987 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2]
1988 // where CI2 != -1 and CI2 != 0 and CI2 != 1
1989 Lower = IntMin.sdiv(Val);
1990 Upper = IntMax.sdiv(Val);
1991 if (Lower.sgt(Upper))
1992 std::swap(Lower, Upper);
1994 assert(Upper != Lower && "Upper part of range has wrapped!");
1996 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
1997 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
1998 APInt NegOne = APInt::getAllOnesValue(Width);
1999 if (CI2->getValue().ult(Width))
2000 Upper = NegOne.lshr(CI2->getValue()) + 1;
2001 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) {
2002 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2].
2003 unsigned ShiftAmount = Width - 1;
2004 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2005 ShiftAmount = CI2->getValue().countTrailingZeros();
2006 Lower = CI2->getValue().lshr(ShiftAmount);
2007 Upper = CI2->getValue() + 1;
2008 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
2009 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
2010 APInt IntMin = APInt::getSignedMinValue(Width);
2011 APInt IntMax = APInt::getSignedMaxValue(Width);
2012 if (CI2->getValue().ult(Width)) {
2013 Lower = IntMin.ashr(CI2->getValue());
2014 Upper = IntMax.ashr(CI2->getValue()) + 1;
2016 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) {
2017 unsigned ShiftAmount = Width - 1;
2018 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2019 ShiftAmount = CI2->getValue().countTrailingZeros();
2020 if (CI2->isNegative()) {
2021 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)]
2022 Lower = CI2->getValue();
2023 Upper = CI2->getValue().ashr(ShiftAmount) + 1;
2025 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2]
2026 Lower = CI2->getValue().ashr(ShiftAmount);
2027 Upper = CI2->getValue() + 1;
2029 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
2030 // 'or x, CI2' produces [CI2, UINT_MAX].
2031 Lower = CI2->getValue();
2032 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
2033 // 'and x, CI2' produces [0, CI2].
2034 Upper = CI2->getValue() + 1;
2036 if (Lower != Upper) {
2037 ConstantRange LHS_CR = ConstantRange(Lower, Upper);
2038 if (RHS_CR.contains(LHS_CR))
2039 return ConstantInt::getTrue(RHS->getContext());
2040 if (RHS_CR.inverse().contains(LHS_CR))
2041 return ConstantInt::getFalse(RHS->getContext());
2045 // Compare of cast, for example (zext X) != 0 -> X != 0
2046 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
2047 Instruction *LI = cast<CastInst>(LHS);
2048 Value *SrcOp = LI->getOperand(0);
2049 Type *SrcTy = SrcOp->getType();
2050 Type *DstTy = LI->getType();
2052 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
2053 // if the integer type is the same size as the pointer type.
2054 if (MaxRecurse && Q.DL && isa<PtrToIntInst>(LI) &&
2055 Q.DL->getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
2056 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2057 // Transfer the cast to the constant.
2058 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
2059 ConstantExpr::getIntToPtr(RHSC, SrcTy),
2062 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
2063 if (RI->getOperand(0)->getType() == SrcTy)
2064 // Compare without the cast.
2065 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2071 if (isa<ZExtInst>(LHS)) {
2072 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
2074 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
2075 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2076 // Compare X and Y. Note that signed predicates become unsigned.
2077 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2078 SrcOp, RI->getOperand(0), Q,
2082 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
2083 // too. If not, then try to deduce the result of the comparison.
2084 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2085 // Compute the constant that would happen if we truncated to SrcTy then
2086 // reextended to DstTy.
2087 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2088 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
2090 // If the re-extended constant didn't change then this is effectively
2091 // also a case of comparing two zero-extended values.
2092 if (RExt == CI && MaxRecurse)
2093 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2094 SrcOp, Trunc, Q, MaxRecurse-1))
2097 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
2098 // there. Use this to work out the result of the comparison.
2101 default: llvm_unreachable("Unknown ICmp predicate!");
2103 case ICmpInst::ICMP_EQ:
2104 case ICmpInst::ICMP_UGT:
2105 case ICmpInst::ICMP_UGE:
2106 return ConstantInt::getFalse(CI->getContext());
2108 case ICmpInst::ICMP_NE:
2109 case ICmpInst::ICMP_ULT:
2110 case ICmpInst::ICMP_ULE:
2111 return ConstantInt::getTrue(CI->getContext());
2113 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
2114 // is non-negative then LHS <s RHS.
2115 case ICmpInst::ICMP_SGT:
2116 case ICmpInst::ICMP_SGE:
2117 return CI->getValue().isNegative() ?
2118 ConstantInt::getTrue(CI->getContext()) :
2119 ConstantInt::getFalse(CI->getContext());
2121 case ICmpInst::ICMP_SLT:
2122 case ICmpInst::ICMP_SLE:
2123 return CI->getValue().isNegative() ?
2124 ConstantInt::getFalse(CI->getContext()) :
2125 ConstantInt::getTrue(CI->getContext());
2131 if (isa<SExtInst>(LHS)) {
2132 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
2134 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
2135 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2136 // Compare X and Y. Note that the predicate does not change.
2137 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2141 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
2142 // too. If not, then try to deduce the result of the comparison.
2143 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2144 // Compute the constant that would happen if we truncated to SrcTy then
2145 // reextended to DstTy.
2146 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2147 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
2149 // If the re-extended constant didn't change then this is effectively
2150 // also a case of comparing two sign-extended values.
2151 if (RExt == CI && MaxRecurse)
2152 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
2155 // Otherwise the upper bits of LHS are all equal, while RHS has varying
2156 // bits there. Use this to work out the result of the comparison.
2159 default: llvm_unreachable("Unknown ICmp predicate!");
2160 case ICmpInst::ICMP_EQ:
2161 return ConstantInt::getFalse(CI->getContext());
2162 case ICmpInst::ICMP_NE:
2163 return ConstantInt::getTrue(CI->getContext());
2165 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
2167 case ICmpInst::ICMP_SGT:
2168 case ICmpInst::ICMP_SGE:
2169 return CI->getValue().isNegative() ?
2170 ConstantInt::getTrue(CI->getContext()) :
2171 ConstantInt::getFalse(CI->getContext());
2172 case ICmpInst::ICMP_SLT:
2173 case ICmpInst::ICMP_SLE:
2174 return CI->getValue().isNegative() ?
2175 ConstantInt::getFalse(CI->getContext()) :
2176 ConstantInt::getTrue(CI->getContext());
2178 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2180 case ICmpInst::ICMP_UGT:
2181 case ICmpInst::ICMP_UGE:
2182 // Comparison is true iff the LHS <s 0.
2184 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2185 Constant::getNullValue(SrcTy),
2189 case ICmpInst::ICMP_ULT:
2190 case ICmpInst::ICMP_ULE:
2191 // Comparison is true iff the LHS >=s 0.
2193 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2194 Constant::getNullValue(SrcTy),
2204 // If a bit is known to be zero for A and known to be one for B,
2205 // then A and B cannot be equal.
2206 if (ICmpInst::isEquality(Pred)) {
2207 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2208 uint32_t BitWidth = CI->getBitWidth();
2209 APInt LHSKnownZero(BitWidth, 0);
2210 APInt LHSKnownOne(BitWidth, 0);
2211 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
2212 APInt RHSKnownZero(BitWidth, 0);
2213 APInt RHSKnownOne(BitWidth, 0);
2214 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
2215 if (((LHSKnownOne & RHSKnownZero) != 0) ||
2216 ((LHSKnownZero & RHSKnownOne) != 0))
2217 return (Pred == ICmpInst::ICMP_EQ)
2218 ? ConstantInt::getFalse(CI->getContext())
2219 : ConstantInt::getTrue(CI->getContext());
2223 // Special logic for binary operators.
2224 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2225 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2226 if (MaxRecurse && (LBO || RBO)) {
2227 // Analyze the case when either LHS or RHS is an add instruction.
2228 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2229 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2230 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2231 if (LBO && LBO->getOpcode() == Instruction::Add) {
2232 A = LBO->getOperand(0); B = LBO->getOperand(1);
2233 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2234 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2235 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2237 if (RBO && RBO->getOpcode() == Instruction::Add) {
2238 C = RBO->getOperand(0); D = RBO->getOperand(1);
2239 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2240 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2241 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2244 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2245 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2246 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2247 Constant::getNullValue(RHS->getType()),
2251 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2252 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2253 if (Value *V = SimplifyICmpInst(Pred,
2254 Constant::getNullValue(LHS->getType()),
2255 C == LHS ? D : C, Q, MaxRecurse-1))
2258 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2259 if (A && C && (A == C || A == D || B == C || B == D) &&
2260 NoLHSWrapProblem && NoRHSWrapProblem) {
2261 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2264 // C + B == C + D -> B == D
2267 } else if (A == D) {
2268 // D + B == C + D -> B == C
2271 } else if (B == C) {
2272 // A + C == C + D -> A == D
2277 // A + D == C + D -> A == C
2281 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2286 // 0 - (zext X) pred C
2287 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2288 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2289 if (RHSC->getValue().isStrictlyPositive()) {
2290 if (Pred == ICmpInst::ICMP_SLT)
2291 return ConstantInt::getTrue(RHSC->getContext());
2292 if (Pred == ICmpInst::ICMP_SGE)
2293 return ConstantInt::getFalse(RHSC->getContext());
2294 if (Pred == ICmpInst::ICMP_EQ)
2295 return ConstantInt::getFalse(RHSC->getContext());
2296 if (Pred == ICmpInst::ICMP_NE)
2297 return ConstantInt::getTrue(RHSC->getContext());
2299 if (RHSC->getValue().isNonNegative()) {
2300 if (Pred == ICmpInst::ICMP_SLE)
2301 return ConstantInt::getTrue(RHSC->getContext());
2302 if (Pred == ICmpInst::ICMP_SGT)
2303 return ConstantInt::getFalse(RHSC->getContext());
2308 // icmp pred (urem X, Y), Y
2309 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2310 bool KnownNonNegative, KnownNegative;
2314 case ICmpInst::ICMP_SGT:
2315 case ICmpInst::ICMP_SGE:
2316 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL);
2317 if (!KnownNonNegative)
2320 case ICmpInst::ICMP_EQ:
2321 case ICmpInst::ICMP_UGT:
2322 case ICmpInst::ICMP_UGE:
2323 return getFalse(ITy);
2324 case ICmpInst::ICMP_SLT:
2325 case ICmpInst::ICMP_SLE:
2326 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL);
2327 if (!KnownNonNegative)
2330 case ICmpInst::ICMP_NE:
2331 case ICmpInst::ICMP_ULT:
2332 case ICmpInst::ICMP_ULE:
2333 return getTrue(ITy);
2337 // icmp pred X, (urem Y, X)
2338 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2339 bool KnownNonNegative, KnownNegative;
2343 case ICmpInst::ICMP_SGT:
2344 case ICmpInst::ICMP_SGE:
2345 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL);
2346 if (!KnownNonNegative)
2349 case ICmpInst::ICMP_NE:
2350 case ICmpInst::ICMP_UGT:
2351 case ICmpInst::ICMP_UGE:
2352 return getTrue(ITy);
2353 case ICmpInst::ICMP_SLT:
2354 case ICmpInst::ICMP_SLE:
2355 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL);
2356 if (!KnownNonNegative)
2359 case ICmpInst::ICMP_EQ:
2360 case ICmpInst::ICMP_ULT:
2361 case ICmpInst::ICMP_ULE:
2362 return getFalse(ITy);
2367 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2368 // icmp pred (X /u Y), X
2369 if (Pred == ICmpInst::ICMP_UGT)
2370 return getFalse(ITy);
2371 if (Pred == ICmpInst::ICMP_ULE)
2372 return getTrue(ITy);
2375 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2376 LBO->getOperand(1) == RBO->getOperand(1)) {
2377 switch (LBO->getOpcode()) {
2379 case Instruction::UDiv:
2380 case Instruction::LShr:
2381 if (ICmpInst::isSigned(Pred))
2384 case Instruction::SDiv:
2385 case Instruction::AShr:
2386 if (!LBO->isExact() || !RBO->isExact())
2388 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2389 RBO->getOperand(0), Q, MaxRecurse-1))
2392 case Instruction::Shl: {
2393 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2394 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2397 if (!NSW && ICmpInst::isSigned(Pred))
2399 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2400 RBO->getOperand(0), Q, MaxRecurse-1))
2407 // Simplify comparisons involving max/min.
2409 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2410 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2412 // Signed variants on "max(a,b)>=a -> true".
2413 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2414 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2415 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2416 // We analyze this as smax(A, B) pred A.
2418 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2419 (A == LHS || B == LHS)) {
2420 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2421 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2422 // We analyze this as smax(A, B) swapped-pred A.
2423 P = CmpInst::getSwappedPredicate(Pred);
2424 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2425 (A == RHS || B == RHS)) {
2426 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2427 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2428 // We analyze this as smax(-A, -B) swapped-pred -A.
2429 // Note that we do not need to actually form -A or -B thanks to EqP.
2430 P = CmpInst::getSwappedPredicate(Pred);
2431 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2432 (A == LHS || B == LHS)) {
2433 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2434 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2435 // We analyze this as smax(-A, -B) pred -A.
2436 // Note that we do not need to actually form -A or -B thanks to EqP.
2439 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2440 // Cases correspond to "max(A, B) p A".
2444 case CmpInst::ICMP_EQ:
2445 case CmpInst::ICMP_SLE:
2446 // Equivalent to "A EqP B". This may be the same as the condition tested
2447 // in the max/min; if so, we can just return that.
2448 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2450 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2452 // Otherwise, see if "A EqP B" simplifies.
2454 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2457 case CmpInst::ICMP_NE:
2458 case CmpInst::ICMP_SGT: {
2459 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2460 // Equivalent to "A InvEqP B". This may be the same as the condition
2461 // tested in the max/min; if so, we can just return that.
2462 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2464 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2466 // Otherwise, see if "A InvEqP B" simplifies.
2468 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2472 case CmpInst::ICMP_SGE:
2474 return getTrue(ITy);
2475 case CmpInst::ICMP_SLT:
2477 return getFalse(ITy);
2481 // Unsigned variants on "max(a,b)>=a -> true".
2482 P = CmpInst::BAD_ICMP_PREDICATE;
2483 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2484 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2485 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2486 // We analyze this as umax(A, B) pred A.
2488 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2489 (A == LHS || B == LHS)) {
2490 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2491 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2492 // We analyze this as umax(A, B) swapped-pred A.
2493 P = CmpInst::getSwappedPredicate(Pred);
2494 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2495 (A == RHS || B == RHS)) {
2496 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2497 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2498 // We analyze this as umax(-A, -B) swapped-pred -A.
2499 // Note that we do not need to actually form -A or -B thanks to EqP.
2500 P = CmpInst::getSwappedPredicate(Pred);
2501 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2502 (A == LHS || B == LHS)) {
2503 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2504 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2505 // We analyze this as umax(-A, -B) pred -A.
2506 // Note that we do not need to actually form -A or -B thanks to EqP.
2509 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2510 // Cases correspond to "max(A, B) p A".
2514 case CmpInst::ICMP_EQ:
2515 case CmpInst::ICMP_ULE:
2516 // Equivalent to "A EqP B". This may be the same as the condition tested
2517 // in the max/min; if so, we can just return that.
2518 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2520 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2522 // Otherwise, see if "A EqP B" simplifies.
2524 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2527 case CmpInst::ICMP_NE:
2528 case CmpInst::ICMP_UGT: {
2529 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2530 // Equivalent to "A InvEqP B". This may be the same as the condition
2531 // tested in the max/min; if so, we can just return that.
2532 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2534 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2536 // Otherwise, see if "A InvEqP B" simplifies.
2538 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2542 case CmpInst::ICMP_UGE:
2544 return getTrue(ITy);
2545 case CmpInst::ICMP_ULT:
2547 return getFalse(ITy);
2551 // Variants on "max(x,y) >= min(x,z)".
2553 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2554 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2555 (A == C || A == D || B == C || B == D)) {
2556 // max(x, ?) pred min(x, ?).
2557 if (Pred == CmpInst::ICMP_SGE)
2559 return getTrue(ITy);
2560 if (Pred == CmpInst::ICMP_SLT)
2562 return getFalse(ITy);
2563 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2564 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2565 (A == C || A == D || B == C || B == D)) {
2566 // min(x, ?) pred max(x, ?).
2567 if (Pred == CmpInst::ICMP_SLE)
2569 return getTrue(ITy);
2570 if (Pred == CmpInst::ICMP_SGT)
2572 return getFalse(ITy);
2573 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2574 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2575 (A == C || A == D || B == C || B == D)) {
2576 // max(x, ?) pred min(x, ?).
2577 if (Pred == CmpInst::ICMP_UGE)
2579 return getTrue(ITy);
2580 if (Pred == CmpInst::ICMP_ULT)
2582 return getFalse(ITy);
2583 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2584 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2585 (A == C || A == D || B == C || B == D)) {
2586 // min(x, ?) pred max(x, ?).
2587 if (Pred == CmpInst::ICMP_ULE)
2589 return getTrue(ITy);
2590 if (Pred == CmpInst::ICMP_UGT)
2592 return getFalse(ITy);
2595 // Simplify comparisons of related pointers using a powerful, recursive
2596 // GEP-walk when we have target data available..
2597 if (LHS->getType()->isPointerTy())
2598 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS))
2601 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
2602 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
2603 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
2604 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
2605 (ICmpInst::isEquality(Pred) ||
2606 (GLHS->isInBounds() && GRHS->isInBounds() &&
2607 Pred == ICmpInst::getSignedPredicate(Pred)))) {
2608 // The bases are equal and the indices are constant. Build a constant
2609 // expression GEP with the same indices and a null base pointer to see
2610 // what constant folding can make out of it.
2611 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
2612 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
2613 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS);
2615 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
2616 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS);
2617 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
2622 // If the comparison is with the result of a select instruction, check whether
2623 // comparing with either branch of the select always yields the same value.
2624 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2625 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2628 // If the comparison is with the result of a phi instruction, check whether
2629 // doing the compare with each incoming phi value yields a common result.
2630 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2631 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2637 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2638 const DataLayout *DL,
2639 const TargetLibraryInfo *TLI,
2640 const DominatorTree *DT) {
2641 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
2645 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
2646 /// fold the result. If not, this returns null.
2647 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2648 const Query &Q, unsigned MaxRecurse) {
2649 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2650 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
2652 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2653 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2654 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
2656 // If we have a constant, make sure it is on the RHS.
2657 std::swap(LHS, RHS);
2658 Pred = CmpInst::getSwappedPredicate(Pred);
2661 // Fold trivial predicates.
2662 if (Pred == FCmpInst::FCMP_FALSE)
2663 return ConstantInt::get(GetCompareTy(LHS), 0);
2664 if (Pred == FCmpInst::FCMP_TRUE)
2665 return ConstantInt::get(GetCompareTy(LHS), 1);
2667 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef
2668 return UndefValue::get(GetCompareTy(LHS));
2670 // fcmp x,x -> true/false. Not all compares are foldable.
2672 if (CmpInst::isTrueWhenEqual(Pred))
2673 return ConstantInt::get(GetCompareTy(LHS), 1);
2674 if (CmpInst::isFalseWhenEqual(Pred))
2675 return ConstantInt::get(GetCompareTy(LHS), 0);
2678 // Handle fcmp with constant RHS
2679 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2680 // If the constant is a nan, see if we can fold the comparison based on it.
2681 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2682 if (CFP->getValueAPF().isNaN()) {
2683 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
2684 return ConstantInt::getFalse(CFP->getContext());
2685 assert(FCmpInst::isUnordered(Pred) &&
2686 "Comparison must be either ordered or unordered!");
2687 // True if unordered.
2688 return ConstantInt::getTrue(CFP->getContext());
2690 // Check whether the constant is an infinity.
2691 if (CFP->getValueAPF().isInfinity()) {
2692 if (CFP->getValueAPF().isNegative()) {
2694 case FCmpInst::FCMP_OLT:
2695 // No value is ordered and less than negative infinity.
2696 return ConstantInt::getFalse(CFP->getContext());
2697 case FCmpInst::FCMP_UGE:
2698 // All values are unordered with or at least negative infinity.
2699 return ConstantInt::getTrue(CFP->getContext());
2705 case FCmpInst::FCMP_OGT:
2706 // No value is ordered and greater than infinity.
2707 return ConstantInt::getFalse(CFP->getContext());
2708 case FCmpInst::FCMP_ULE:
2709 // All values are unordered with and at most infinity.
2710 return ConstantInt::getTrue(CFP->getContext());
2719 // If the comparison is with the result of a select instruction, check whether
2720 // comparing with either branch of the select always yields the same value.
2721 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2722 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2725 // If the comparison is with the result of a phi instruction, check whether
2726 // doing the compare with each incoming phi value yields a common result.
2727 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2728 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2734 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2735 const DataLayout *DL,
2736 const TargetLibraryInfo *TLI,
2737 const DominatorTree *DT) {
2738 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
2742 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
2743 /// the result. If not, this returns null.
2744 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
2745 Value *FalseVal, const Query &Q,
2746 unsigned MaxRecurse) {
2747 // select true, X, Y -> X
2748 // select false, X, Y -> Y
2749 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
2750 if (CB->isAllOnesValue())
2752 if (CB->isNullValue())
2756 // select C, X, X -> X
2757 if (TrueVal == FalseVal)
2760 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
2761 if (isa<Constant>(TrueVal))
2765 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
2767 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
2773 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
2774 const DataLayout *DL,
2775 const TargetLibraryInfo *TLI,
2776 const DominatorTree *DT) {
2777 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Query (DL, TLI, DT),
2781 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
2782 /// fold the result. If not, this returns null.
2783 static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) {
2784 // The type of the GEP pointer operand.
2785 PointerType *PtrTy = cast<PointerType>(Ops[0]->getType()->getScalarType());
2787 // getelementptr P -> P.
2788 if (Ops.size() == 1)
2791 if (isa<UndefValue>(Ops[0])) {
2792 // Compute the (pointer) type returned by the GEP instruction.
2793 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1));
2794 Type *GEPTy = PointerType::get(LastType, PtrTy->getAddressSpace());
2795 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
2796 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
2797 return UndefValue::get(GEPTy);
2800 if (Ops.size() == 2) {
2801 // getelementptr P, 0 -> P.
2802 if (match(Ops[1], m_Zero()))
2804 // getelementptr P, N -> P if P points to a type of zero size.
2806 Type *Ty = PtrTy->getElementType();
2807 if (Ty->isSized() && Q.DL->getTypeAllocSize(Ty) == 0)
2812 // Check to see if this is constant foldable.
2813 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2814 if (!isa<Constant>(Ops[i]))
2817 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1));
2820 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *DL,
2821 const TargetLibraryInfo *TLI,
2822 const DominatorTree *DT) {
2823 return ::SimplifyGEPInst(Ops, Query (DL, TLI, DT), RecursionLimit);
2826 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
2827 /// can fold the result. If not, this returns null.
2828 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
2829 ArrayRef<unsigned> Idxs, const Query &Q,
2831 if (Constant *CAgg = dyn_cast<Constant>(Agg))
2832 if (Constant *CVal = dyn_cast<Constant>(Val))
2833 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
2835 // insertvalue x, undef, n -> x
2836 if (match(Val, m_Undef()))
2839 // insertvalue x, (extractvalue y, n), n
2840 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
2841 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
2842 EV->getIndices() == Idxs) {
2843 // insertvalue undef, (extractvalue y, n), n -> y
2844 if (match(Agg, m_Undef()))
2845 return EV->getAggregateOperand();
2847 // insertvalue y, (extractvalue y, n), n -> y
2848 if (Agg == EV->getAggregateOperand())
2855 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
2856 ArrayRef<unsigned> Idxs,
2857 const DataLayout *DL,
2858 const TargetLibraryInfo *TLI,
2859 const DominatorTree *DT) {
2860 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query (DL, TLI, DT),
2864 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
2865 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
2866 // If all of the PHI's incoming values are the same then replace the PHI node
2867 // with the common value.
2868 Value *CommonValue = nullptr;
2869 bool HasUndefInput = false;
2870 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2871 Value *Incoming = PN->getIncomingValue(i);
2872 // If the incoming value is the phi node itself, it can safely be skipped.
2873 if (Incoming == PN) continue;
2874 if (isa<UndefValue>(Incoming)) {
2875 // Remember that we saw an undef value, but otherwise ignore them.
2876 HasUndefInput = true;
2879 if (CommonValue && Incoming != CommonValue)
2880 return nullptr; // Not the same, bail out.
2881 CommonValue = Incoming;
2884 // If CommonValue is null then all of the incoming values were either undef or
2885 // equal to the phi node itself.
2887 return UndefValue::get(PN->getType());
2889 // If we have a PHI node like phi(X, undef, X), where X is defined by some
2890 // instruction, we cannot return X as the result of the PHI node unless it
2891 // dominates the PHI block.
2893 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
2898 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
2899 if (Constant *C = dyn_cast<Constant>(Op))
2900 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI);
2905 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *DL,
2906 const TargetLibraryInfo *TLI,
2907 const DominatorTree *DT) {
2908 return ::SimplifyTruncInst(Op, Ty, Query (DL, TLI, DT), RecursionLimit);
2911 //=== Helper functions for higher up the class hierarchy.
2913 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
2914 /// fold the result. If not, this returns null.
2915 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
2916 const Query &Q, unsigned MaxRecurse) {
2918 case Instruction::Add:
2919 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2921 case Instruction::FAdd:
2922 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
2924 case Instruction::Sub:
2925 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2927 case Instruction::FSub:
2928 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
2930 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
2931 case Instruction::FMul:
2932 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
2933 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
2934 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
2935 case Instruction::FDiv: return SimplifyFDivInst(LHS, RHS, Q, MaxRecurse);
2936 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
2937 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
2938 case Instruction::FRem: return SimplifyFRemInst(LHS, RHS, Q, MaxRecurse);
2939 case Instruction::Shl:
2940 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2942 case Instruction::LShr:
2943 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
2944 case Instruction::AShr:
2945 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
2946 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
2947 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
2948 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
2950 if (Constant *CLHS = dyn_cast<Constant>(LHS))
2951 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
2952 Constant *COps[] = {CLHS, CRHS};
2953 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL,
2957 // If the operation is associative, try some generic simplifications.
2958 if (Instruction::isAssociative(Opcode))
2959 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
2962 // If the operation is with the result of a select instruction check whether
2963 // operating on either branch of the select always yields the same value.
2964 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2965 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
2968 // If the operation is with the result of a phi instruction, check whether
2969 // operating on all incoming values of the phi always yields the same value.
2970 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2971 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
2978 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
2979 const DataLayout *DL, const TargetLibraryInfo *TLI,
2980 const DominatorTree *DT) {
2981 return ::SimplifyBinOp(Opcode, LHS, RHS, Query (DL, TLI, DT), RecursionLimit);
2984 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
2985 /// fold the result.
2986 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2987 const Query &Q, unsigned MaxRecurse) {
2988 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
2989 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
2990 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
2993 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2994 const DataLayout *DL, const TargetLibraryInfo *TLI,
2995 const DominatorTree *DT) {
2996 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
3000 static bool IsIdempotent(Intrinsic::ID ID) {
3002 default: return false;
3004 // Unary idempotent: f(f(x)) = f(x)
3005 case Intrinsic::fabs:
3006 case Intrinsic::floor:
3007 case Intrinsic::ceil:
3008 case Intrinsic::trunc:
3009 case Intrinsic::rint:
3010 case Intrinsic::nearbyint:
3011 case Intrinsic::round:
3016 template <typename IterTy>
3017 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd,
3018 const Query &Q, unsigned MaxRecurse) {
3019 // Perform idempotent optimizations
3020 if (!IsIdempotent(IID))
3024 if (std::distance(ArgBegin, ArgEnd) == 1)
3025 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
3026 if (II->getIntrinsicID() == IID)
3032 template <typename IterTy>
3033 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
3034 const Query &Q, unsigned MaxRecurse) {
3035 Type *Ty = V->getType();
3036 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
3037 Ty = PTy->getElementType();
3038 FunctionType *FTy = cast<FunctionType>(Ty);
3040 // call undef -> undef
3041 if (isa<UndefValue>(V))
3042 return UndefValue::get(FTy->getReturnType());
3044 Function *F = dyn_cast<Function>(V);
3048 if (unsigned IID = F->getIntrinsicID())
3050 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse))
3053 if (!canConstantFoldCallTo(F))
3056 SmallVector<Constant *, 4> ConstantArgs;
3057 ConstantArgs.reserve(ArgEnd - ArgBegin);
3058 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
3059 Constant *C = dyn_cast<Constant>(*I);
3062 ConstantArgs.push_back(C);
3065 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
3068 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
3069 User::op_iterator ArgEnd, const DataLayout *DL,
3070 const TargetLibraryInfo *TLI,
3071 const DominatorTree *DT) {
3072 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT),
3076 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
3077 const DataLayout *DL, const TargetLibraryInfo *TLI,
3078 const DominatorTree *DT) {
3079 return ::SimplifyCall(V, Args.begin(), Args.end(), Query(DL, TLI, DT),
3083 /// SimplifyInstruction - See if we can compute a simplified version of this
3084 /// instruction. If not, this returns null.
3085 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *DL,
3086 const TargetLibraryInfo *TLI,
3087 const DominatorTree *DT) {
3090 switch (I->getOpcode()) {
3092 Result = ConstantFoldInstruction(I, DL, TLI);
3094 case Instruction::FAdd:
3095 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
3096 I->getFastMathFlags(), DL, TLI, DT);
3098 case Instruction::Add:
3099 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
3100 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3101 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3104 case Instruction::FSub:
3105 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
3106 I->getFastMathFlags(), DL, TLI, DT);
3108 case Instruction::Sub:
3109 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
3110 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3111 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3114 case Instruction::FMul:
3115 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
3116 I->getFastMathFlags(), DL, TLI, DT);
3118 case Instruction::Mul:
3119 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3121 case Instruction::SDiv:
3122 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3124 case Instruction::UDiv:
3125 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3127 case Instruction::FDiv:
3128 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3130 case Instruction::SRem:
3131 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3133 case Instruction::URem:
3134 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3136 case Instruction::FRem:
3137 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3139 case Instruction::Shl:
3140 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
3141 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3142 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3145 case Instruction::LShr:
3146 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
3147 cast<BinaryOperator>(I)->isExact(),
3150 case Instruction::AShr:
3151 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
3152 cast<BinaryOperator>(I)->isExact(),
3155 case Instruction::And:
3156 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3158 case Instruction::Or:
3159 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3161 case Instruction::Xor:
3162 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3164 case Instruction::ICmp:
3165 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
3166 I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3168 case Instruction::FCmp:
3169 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
3170 I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3172 case Instruction::Select:
3173 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
3174 I->getOperand(2), DL, TLI, DT);
3176 case Instruction::GetElementPtr: {
3177 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
3178 Result = SimplifyGEPInst(Ops, DL, TLI, DT);
3181 case Instruction::InsertValue: {
3182 InsertValueInst *IV = cast<InsertValueInst>(I);
3183 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
3184 IV->getInsertedValueOperand(),
3185 IV->getIndices(), DL, TLI, DT);
3188 case Instruction::PHI:
3189 Result = SimplifyPHINode(cast<PHINode>(I), Query (DL, TLI, DT));
3191 case Instruction::Call: {
3192 CallSite CS(cast<CallInst>(I));
3193 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
3197 case Instruction::Trunc:
3198 Result = SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT);
3202 /// If called on unreachable code, the above logic may report that the
3203 /// instruction simplified to itself. Make life easier for users by
3204 /// detecting that case here, returning a safe value instead.
3205 return Result == I ? UndefValue::get(I->getType()) : Result;
3208 /// \brief Implementation of recursive simplification through an instructions
3211 /// This is the common implementation of the recursive simplification routines.
3212 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
3213 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
3214 /// instructions to process and attempt to simplify it using
3215 /// InstructionSimplify.
3217 /// This routine returns 'true' only when *it* simplifies something. The passed
3218 /// in simplified value does not count toward this.
3219 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
3220 const DataLayout *DL,
3221 const TargetLibraryInfo *TLI,
3222 const DominatorTree *DT) {
3223 bool Simplified = false;
3224 SmallSetVector<Instruction *, 8> Worklist;
3226 // If we have an explicit value to collapse to, do that round of the
3227 // simplification loop by hand initially.
3229 for (User *U : I->users())
3231 Worklist.insert(cast<Instruction>(U));
3233 // Replace the instruction with its simplified value.
3234 I->replaceAllUsesWith(SimpleV);
3236 // Gracefully handle edge cases where the instruction is not wired into any
3239 I->eraseFromParent();
3244 // Note that we must test the size on each iteration, the worklist can grow.
3245 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
3248 // See if this instruction simplifies.
3249 SimpleV = SimplifyInstruction(I, DL, TLI, DT);
3255 // Stash away all the uses of the old instruction so we can check them for
3256 // recursive simplifications after a RAUW. This is cheaper than checking all
3257 // uses of To on the recursive step in most cases.
3258 for (User *U : I->users())
3259 Worklist.insert(cast<Instruction>(U));
3261 // Replace the instruction with its simplified value.
3262 I->replaceAllUsesWith(SimpleV);
3264 // Gracefully handle edge cases where the instruction is not wired into any
3267 I->eraseFromParent();
3272 bool llvm::recursivelySimplifyInstruction(Instruction *I,
3273 const DataLayout *DL,
3274 const TargetLibraryInfo *TLI,
3275 const DominatorTree *DT) {
3276 return replaceAndRecursivelySimplifyImpl(I, nullptr, DL, TLI, DT);
3279 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
3280 const DataLayout *DL,
3281 const TargetLibraryInfo *TLI,
3282 const DominatorTree *DT) {
3283 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
3284 assert(SimpleV && "Must provide a simplified value.");
3285 return replaceAndRecursivelySimplifyImpl(I, SimpleV, DL, TLI, DT);