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");
47 const TargetLibraryInfo *TLI;
48 const DominatorTree *DT;
49 AssumptionTracker *AT;
50 const Instruction *CxtI;
52 Query(const DataLayout *DL, const TargetLibraryInfo *tli,
53 const DominatorTree *dt, AssumptionTracker *at = nullptr,
54 const Instruction *cxti = nullptr)
55 : DL(DL), TLI(tli), DT(dt), AT(at), CxtI(cxti) {}
57 } // end anonymous namespace
59 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
60 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
62 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
64 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
65 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
66 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned);
68 /// getFalse - For a boolean type, or a vector of boolean type, return false, or
69 /// a vector with every element false, as appropriate for the type.
70 static Constant *getFalse(Type *Ty) {
71 assert(Ty->getScalarType()->isIntegerTy(1) &&
72 "Expected i1 type or a vector of i1!");
73 return Constant::getNullValue(Ty);
76 /// getTrue - For a boolean type, or a vector of boolean type, return true, or
77 /// a vector with every element true, as appropriate for the type.
78 static Constant *getTrue(Type *Ty) {
79 assert(Ty->getScalarType()->isIntegerTy(1) &&
80 "Expected i1 type or a vector of i1!");
81 return Constant::getAllOnesValue(Ty);
84 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
85 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
87 CmpInst *Cmp = dyn_cast<CmpInst>(V);
90 CmpInst::Predicate CPred = Cmp->getPredicate();
91 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
92 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
94 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
98 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
99 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
100 Instruction *I = dyn_cast<Instruction>(V);
102 // Arguments and constants dominate all instructions.
105 // If we are processing instructions (and/or basic blocks) that have not been
106 // fully added to a function, the parent nodes may still be null. Simply
107 // return the conservative answer in these cases.
108 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
111 // If we have a DominatorTree then do a precise test.
113 if (!DT->isReachableFromEntry(P->getParent()))
115 if (!DT->isReachableFromEntry(I->getParent()))
117 return DT->dominates(I, P);
120 // Otherwise, if the instruction is in the entry block, and is not an invoke,
121 // then it obviously dominates all phi nodes.
122 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
129 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning
130 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
131 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
132 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
133 /// Returns the simplified value, or null if no simplification was performed.
134 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
135 unsigned OpcToExpand, const Query &Q,
136 unsigned MaxRecurse) {
137 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
138 // Recursion is always used, so bail out at once if we already hit the limit.
142 // Check whether the expression has the form "(A op' B) op C".
143 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
144 if (Op0->getOpcode() == OpcodeToExpand) {
145 // It does! Try turning it into "(A op C) op' (B op C)".
146 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
147 // Do "A op C" and "B op C" both simplify?
148 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
149 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
150 // They do! Return "L op' R" if it simplifies or is already available.
151 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
152 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
153 && L == B && R == A)) {
157 // Otherwise return "L op' R" if it simplifies.
158 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
165 // Check whether the expression has the form "A op (B op' C)".
166 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
167 if (Op1->getOpcode() == OpcodeToExpand) {
168 // It does! Try turning it into "(A op B) op' (A op C)".
169 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
170 // Do "A op B" and "A op C" both simplify?
171 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
172 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
173 // They do! Return "L op' R" if it simplifies or is already available.
174 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
175 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
176 && L == C && R == B)) {
180 // Otherwise return "L op' R" if it simplifies.
181 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
191 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
192 /// operations. Returns the simpler value, or null if none was found.
193 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
194 const Query &Q, unsigned MaxRecurse) {
195 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
196 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
198 // Recursion is always used, so bail out at once if we already hit the limit.
202 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
203 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
205 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
206 if (Op0 && Op0->getOpcode() == Opcode) {
207 Value *A = Op0->getOperand(0);
208 Value *B = Op0->getOperand(1);
211 // Does "B op C" simplify?
212 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
213 // It does! Return "A op V" if it simplifies or is already available.
214 // If V equals B then "A op V" is just the LHS.
215 if (V == B) return LHS;
216 // Otherwise return "A op V" if it simplifies.
217 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
224 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
225 if (Op1 && Op1->getOpcode() == Opcode) {
227 Value *B = Op1->getOperand(0);
228 Value *C = Op1->getOperand(1);
230 // Does "A op B" simplify?
231 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
232 // It does! Return "V op C" if it simplifies or is already available.
233 // If V equals B then "V op C" is just the RHS.
234 if (V == B) return RHS;
235 // Otherwise return "V op C" if it simplifies.
236 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
243 // The remaining transforms require commutativity as well as associativity.
244 if (!Instruction::isCommutative(Opcode))
247 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
248 if (Op0 && Op0->getOpcode() == Opcode) {
249 Value *A = Op0->getOperand(0);
250 Value *B = Op0->getOperand(1);
253 // Does "C op A" simplify?
254 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
255 // It does! Return "V op B" if it simplifies or is already available.
256 // If V equals A then "V op B" is just the LHS.
257 if (V == A) return LHS;
258 // Otherwise return "V op B" if it simplifies.
259 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
266 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
267 if (Op1 && Op1->getOpcode() == Opcode) {
269 Value *B = Op1->getOperand(0);
270 Value *C = Op1->getOperand(1);
272 // Does "C op A" simplify?
273 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
274 // It does! Return "B op V" if it simplifies or is already available.
275 // If V equals C then "B op V" is just the RHS.
276 if (V == C) return RHS;
277 // Otherwise return "B op V" if it simplifies.
278 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
288 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
289 /// instruction as an operand, try to simplify the binop by seeing whether
290 /// evaluating it on both branches of the select results in the same value.
291 /// Returns the common value if so, otherwise returns null.
292 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
293 const Query &Q, unsigned MaxRecurse) {
294 // Recursion is always used, so bail out at once if we already hit the limit.
299 if (isa<SelectInst>(LHS)) {
300 SI = cast<SelectInst>(LHS);
302 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
303 SI = cast<SelectInst>(RHS);
306 // Evaluate the BinOp on the true and false branches of the select.
310 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
311 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
313 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
314 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
317 // If they simplified to the same value, then return the common value.
318 // If they both failed to simplify then return null.
322 // If one branch simplified to undef, return the other one.
323 if (TV && isa<UndefValue>(TV))
325 if (FV && isa<UndefValue>(FV))
328 // If applying the operation did not change the true and false select values,
329 // then the result of the binop is the select itself.
330 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
333 // If one branch simplified and the other did not, and the simplified
334 // value is equal to the unsimplified one, return the simplified value.
335 // For example, select (cond, X, X & Z) & Z -> X & Z.
336 if ((FV && !TV) || (TV && !FV)) {
337 // Check that the simplified value has the form "X op Y" where "op" is the
338 // same as the original operation.
339 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
340 if (Simplified && Simplified->getOpcode() == Opcode) {
341 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
342 // We already know that "op" is the same as for the simplified value. See
343 // if the operands match too. If so, return the simplified value.
344 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
345 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
346 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
347 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
348 Simplified->getOperand(1) == UnsimplifiedRHS)
350 if (Simplified->isCommutative() &&
351 Simplified->getOperand(1) == UnsimplifiedLHS &&
352 Simplified->getOperand(0) == UnsimplifiedRHS)
360 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
361 /// try to simplify the comparison by seeing whether both branches of the select
362 /// result in the same value. Returns the common value if so, otherwise returns
364 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
365 Value *RHS, const Query &Q,
366 unsigned MaxRecurse) {
367 // Recursion is always used, so bail out at once if we already hit the limit.
371 // Make sure the select is on the LHS.
372 if (!isa<SelectInst>(LHS)) {
374 Pred = CmpInst::getSwappedPredicate(Pred);
376 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
377 SelectInst *SI = cast<SelectInst>(LHS);
378 Value *Cond = SI->getCondition();
379 Value *TV = SI->getTrueValue();
380 Value *FV = SI->getFalseValue();
382 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
383 // Does "cmp TV, RHS" simplify?
384 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
386 // It not only simplified, it simplified to the select condition. Replace
388 TCmp = getTrue(Cond->getType());
390 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
391 // condition then we can replace it with 'true'. Otherwise give up.
392 if (!isSameCompare(Cond, Pred, TV, RHS))
394 TCmp = getTrue(Cond->getType());
397 // Does "cmp FV, RHS" simplify?
398 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
400 // It not only simplified, it simplified to the select condition. Replace
402 FCmp = getFalse(Cond->getType());
404 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
405 // condition then we can replace it with 'false'. Otherwise give up.
406 if (!isSameCompare(Cond, Pred, FV, RHS))
408 FCmp = getFalse(Cond->getType());
411 // If both sides simplified to the same value, then use it as the result of
412 // the original comparison.
416 // The remaining cases only make sense if the select condition has the same
417 // type as the result of the comparison, so bail out if this is not so.
418 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
420 // If the false value simplified to false, then the result of the compare
421 // is equal to "Cond && TCmp". This also catches the case when the false
422 // value simplified to false and the true value to true, returning "Cond".
423 if (match(FCmp, m_Zero()))
424 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
426 // If the true value simplified to true, then the result of the compare
427 // is equal to "Cond || FCmp".
428 if (match(TCmp, m_One()))
429 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
431 // Finally, if the false value simplified to true and the true value to
432 // false, then the result of the compare is equal to "!Cond".
433 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
435 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
442 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
443 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
444 /// it on the incoming phi values yields the same result for every value. If so
445 /// returns the common value, otherwise returns null.
446 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
447 const Query &Q, unsigned MaxRecurse) {
448 // Recursion is always used, so bail out at once if we already hit the limit.
453 if (isa<PHINode>(LHS)) {
454 PI = cast<PHINode>(LHS);
455 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
456 if (!ValueDominatesPHI(RHS, PI, Q.DT))
459 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
460 PI = cast<PHINode>(RHS);
461 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
462 if (!ValueDominatesPHI(LHS, PI, Q.DT))
466 // Evaluate the BinOp on the incoming phi values.
467 Value *CommonValue = nullptr;
468 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
469 Value *Incoming = PI->getIncomingValue(i);
470 // If the incoming value is the phi node itself, it can safely be skipped.
471 if (Incoming == PI) continue;
472 Value *V = PI == LHS ?
473 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
474 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
475 // If the operation failed to simplify, or simplified to a different value
476 // to previously, then give up.
477 if (!V || (CommonValue && V != CommonValue))
485 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
486 /// try to simplify the comparison by seeing whether comparing with all of the
487 /// incoming phi values yields the same result every time. If so returns the
488 /// common result, otherwise returns null.
489 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
490 const Query &Q, unsigned MaxRecurse) {
491 // Recursion is always used, so bail out at once if we already hit the limit.
495 // Make sure the phi is on the LHS.
496 if (!isa<PHINode>(LHS)) {
498 Pred = CmpInst::getSwappedPredicate(Pred);
500 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
501 PHINode *PI = cast<PHINode>(LHS);
503 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
504 if (!ValueDominatesPHI(RHS, PI, Q.DT))
507 // Evaluate the BinOp on the incoming phi values.
508 Value *CommonValue = nullptr;
509 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
510 Value *Incoming = PI->getIncomingValue(i);
511 // If the incoming value is the phi node itself, it can safely be skipped.
512 if (Incoming == PI) continue;
513 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
514 // If the operation failed to simplify, or simplified to a different value
515 // to previously, then give up.
516 if (!V || (CommonValue && V != CommonValue))
524 /// SimplifyAddInst - Given operands for an Add, see if we can
525 /// fold the result. If not, this returns null.
526 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
527 const Query &Q, unsigned MaxRecurse) {
528 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
529 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
530 Constant *Ops[] = { CLHS, CRHS };
531 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
535 // Canonicalize the constant to the RHS.
539 // X + undef -> undef
540 if (match(Op1, m_Undef()))
544 if (match(Op1, m_Zero()))
551 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
552 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
555 // X + ~X -> -1 since ~X = -X-1
556 if (match(Op0, m_Not(m_Specific(Op1))) ||
557 match(Op1, m_Not(m_Specific(Op0))))
558 return Constant::getAllOnesValue(Op0->getType());
561 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
562 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
565 // Try some generic simplifications for associative operations.
566 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
570 // Threading Add over selects and phi nodes is pointless, so don't bother.
571 // Threading over the select in "A + select(cond, B, C)" means evaluating
572 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
573 // only if B and C are equal. If B and C are equal then (since we assume
574 // that operands have already been simplified) "select(cond, B, C)" should
575 // have been simplified to the common value of B and C already. Analysing
576 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
577 // for threading over phi nodes.
582 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
583 const DataLayout *DL, const TargetLibraryInfo *TLI,
584 const DominatorTree *DT, AssumptionTracker *AT,
585 const Instruction *CxtI) {
586 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW,
587 Query (DL, TLI, DT, AT, CxtI), RecursionLimit);
590 /// \brief Compute the base pointer and cumulative constant offsets for V.
592 /// This strips all constant offsets off of V, leaving it the base pointer, and
593 /// accumulates the total constant offset applied in the returned constant. It
594 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
595 /// no constant offsets applied.
597 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
598 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
600 static Constant *stripAndComputeConstantOffsets(const DataLayout *DL,
602 bool AllowNonInbounds = false) {
603 assert(V->getType()->getScalarType()->isPointerTy());
605 // Without DataLayout, just be conservative for now. Theoretically, more could
606 // be done in this case.
608 return ConstantInt::get(IntegerType::get(V->getContext(), 64), 0);
610 Type *IntPtrTy = DL->getIntPtrType(V->getType())->getScalarType();
611 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
613 // Even though we don't look through PHI nodes, we could be called on an
614 // instruction in an unreachable block, which may be on a cycle.
615 SmallPtrSet<Value *, 4> Visited;
618 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
619 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
620 !GEP->accumulateConstantOffset(*DL, Offset))
622 V = GEP->getPointerOperand();
623 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
624 V = cast<Operator>(V)->getOperand(0);
625 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
626 if (GA->mayBeOverridden())
628 V = GA->getAliasee();
632 assert(V->getType()->getScalarType()->isPointerTy() &&
633 "Unexpected operand type!");
634 } while (Visited.insert(V));
636 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
637 if (V->getType()->isVectorTy())
638 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
643 /// \brief Compute the constant difference between two pointer values.
644 /// If the difference is not a constant, returns zero.
645 static Constant *computePointerDifference(const DataLayout *DL,
646 Value *LHS, Value *RHS) {
647 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
648 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
650 // If LHS and RHS are not related via constant offsets to the same base
651 // value, there is nothing we can do here.
655 // Otherwise, the difference of LHS - RHS can be computed as:
657 // = (LHSOffset + Base) - (RHSOffset + Base)
658 // = LHSOffset - RHSOffset
659 return ConstantExpr::getSub(LHSOffset, RHSOffset);
662 /// SimplifySubInst - Given operands for a Sub, see if we can
663 /// fold the result. If not, this returns null.
664 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
665 const Query &Q, unsigned MaxRecurse) {
666 if (Constant *CLHS = dyn_cast<Constant>(Op0))
667 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
668 Constant *Ops[] = { CLHS, CRHS };
669 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
673 // X - undef -> undef
674 // undef - X -> undef
675 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
676 return UndefValue::get(Op0->getType());
679 if (match(Op1, m_Zero()))
684 return Constant::getNullValue(Op0->getType());
686 // X - (0 - Y) -> X if the second sub is NUW.
687 // If Y != 0, 0 - Y is a poison value.
688 // If Y == 0, 0 - Y simplifies to 0.
689 if (BinaryOperator::isNeg(Op1)) {
690 if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
691 assert(BO->getOpcode() == Instruction::Sub &&
692 "Expected a subtraction operator!");
693 if (BO->hasNoUnsignedWrap())
698 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
699 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
700 Value *X = nullptr, *Y = nullptr, *Z = Op1;
701 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
702 // See if "V === Y - Z" simplifies.
703 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
704 // It does! Now see if "X + V" simplifies.
705 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
706 // It does, we successfully reassociated!
710 // See if "V === X - Z" simplifies.
711 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
712 // It does! Now see if "Y + V" simplifies.
713 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
714 // It does, we successfully reassociated!
720 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
721 // For example, X - (X + 1) -> -1
723 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
724 // See if "V === X - Y" simplifies.
725 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
726 // It does! Now see if "V - Z" simplifies.
727 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
728 // It does, we successfully reassociated!
732 // See if "V === X - Z" simplifies.
733 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
734 // It does! Now see if "V - Y" simplifies.
735 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
736 // It does, we successfully reassociated!
742 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
743 // For example, X - (X - Y) -> Y.
745 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
746 // See if "V === Z - X" simplifies.
747 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
748 // It does! Now see if "V + Y" simplifies.
749 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
750 // It does, we successfully reassociated!
755 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
756 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
757 match(Op1, m_Trunc(m_Value(Y))))
758 if (X->getType() == Y->getType())
759 // See if "V === X - Y" simplifies.
760 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
761 // It does! Now see if "trunc V" simplifies.
762 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
763 // It does, return the simplified "trunc V".
766 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
767 if (match(Op0, m_PtrToInt(m_Value(X))) &&
768 match(Op1, m_PtrToInt(m_Value(Y))))
769 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
770 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
773 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
774 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
777 // Threading Sub over selects and phi nodes is pointless, so don't bother.
778 // Threading over the select in "A - select(cond, B, C)" means evaluating
779 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
780 // only if B and C are equal. If B and C are equal then (since we assume
781 // that operands have already been simplified) "select(cond, B, C)" should
782 // have been simplified to the common value of B and C already. Analysing
783 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
784 // for threading over phi nodes.
789 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
790 const DataLayout *DL, const TargetLibraryInfo *TLI,
791 const DominatorTree *DT, AssumptionTracker *AT,
792 const Instruction *CxtI) {
793 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW,
794 Query (DL, TLI, DT, AT, CxtI), RecursionLimit);
797 /// Given operands for an FAdd, see if we can fold the result. If not, this
799 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
800 const Query &Q, unsigned MaxRecurse) {
801 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
802 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
803 Constant *Ops[] = { CLHS, CRHS };
804 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(),
808 // Canonicalize the constant to the RHS.
813 if (match(Op1, m_NegZero()))
816 // fadd X, 0 ==> X, when we know X is not -0
817 if (match(Op1, m_Zero()) &&
818 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
821 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
822 // where nnan and ninf have to occur at least once somewhere in this
824 Value *SubOp = nullptr;
825 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
827 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
830 Instruction *FSub = cast<Instruction>(SubOp);
831 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
832 (FMF.noInfs() || FSub->hasNoInfs()))
833 return Constant::getNullValue(Op0->getType());
839 /// Given operands for an FSub, see if we can fold the result. If not, this
841 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
842 const Query &Q, unsigned MaxRecurse) {
843 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
844 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
845 Constant *Ops[] = { CLHS, CRHS };
846 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(),
852 if (match(Op1, m_Zero()))
855 // fsub X, -0 ==> X, when we know X is not -0
856 if (match(Op1, m_NegZero()) &&
857 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
860 // fsub 0, (fsub -0.0, X) ==> X
862 if (match(Op0, m_AnyZero())) {
863 if (match(Op1, m_FSub(m_NegZero(), m_Value(X))))
865 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
869 // fsub nnan ninf x, x ==> 0.0
870 if (FMF.noNaNs() && FMF.noInfs() && Op0 == Op1)
871 return Constant::getNullValue(Op0->getType());
876 /// Given the operands for an FMul, see if we can fold the result
877 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
880 unsigned MaxRecurse) {
881 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
882 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
883 Constant *Ops[] = { CLHS, CRHS };
884 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(),
888 // Canonicalize the constant to the RHS.
893 if (match(Op1, m_FPOne()))
896 // fmul nnan nsz X, 0 ==> 0
897 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
903 /// SimplifyMulInst - Given operands for a Mul, see if we can
904 /// fold the result. If not, this returns null.
905 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
906 unsigned MaxRecurse) {
907 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
908 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
909 Constant *Ops[] = { CLHS, CRHS };
910 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(),
914 // Canonicalize the constant to the RHS.
919 if (match(Op1, m_Undef()))
920 return Constant::getNullValue(Op0->getType());
923 if (match(Op1, m_Zero()))
927 if (match(Op1, m_One()))
930 // (X / Y) * Y -> X if the division is exact.
932 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
933 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
937 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
938 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
941 // Try some generic simplifications for associative operations.
942 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
946 // Mul distributes over Add. Try some generic simplifications based on this.
947 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
951 // If the operation is with the result of a select instruction, check whether
952 // operating on either branch of the select always yields the same value.
953 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
954 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
958 // If the operation is with the result of a phi instruction, check whether
959 // operating on all incoming values of the phi always yields the same value.
960 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
961 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
968 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
969 const DataLayout *DL, const TargetLibraryInfo *TLI,
970 const DominatorTree *DT, AssumptionTracker *AT,
971 const Instruction *CxtI) {
972 return ::SimplifyFAddInst(Op0, Op1, FMF, Query (DL, TLI, DT, AT, CxtI),
976 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
977 const DataLayout *DL, const TargetLibraryInfo *TLI,
978 const DominatorTree *DT, AssumptionTracker *AT,
979 const Instruction *CxtI) {
980 return ::SimplifyFSubInst(Op0, Op1, FMF, Query (DL, TLI, DT, AT, CxtI),
984 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1,
986 const DataLayout *DL,
987 const TargetLibraryInfo *TLI,
988 const DominatorTree *DT,
989 AssumptionTracker *AT,
990 const Instruction *CxtI) {
991 return ::SimplifyFMulInst(Op0, Op1, FMF, Query (DL, TLI, DT, AT, CxtI),
995 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout *DL,
996 const TargetLibraryInfo *TLI,
997 const DominatorTree *DT, AssumptionTracker *AT,
998 const Instruction *CxtI) {
999 return ::SimplifyMulInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1003 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
1004 /// fold the result. If not, this returns null.
1005 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1006 const Query &Q, unsigned MaxRecurse) {
1007 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1008 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1009 Constant *Ops[] = { C0, C1 };
1010 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1014 bool isSigned = Opcode == Instruction::SDiv;
1016 // X / undef -> undef
1017 if (match(Op1, m_Undef()))
1021 if (match(Op0, m_Undef()))
1022 return Constant::getNullValue(Op0->getType());
1024 // 0 / X -> 0, we don't need to preserve faults!
1025 if (match(Op0, m_Zero()))
1029 if (match(Op1, m_One()))
1032 if (Op0->getType()->isIntegerTy(1))
1033 // It can't be division by zero, hence it must be division by one.
1038 return ConstantInt::get(Op0->getType(), 1);
1040 // (X * Y) / Y -> X if the multiplication does not overflow.
1041 Value *X = nullptr, *Y = nullptr;
1042 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1043 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1044 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1045 // If the Mul knows it does not overflow, then we are good to go.
1046 if ((isSigned && Mul->hasNoSignedWrap()) ||
1047 (!isSigned && Mul->hasNoUnsignedWrap()))
1049 // If X has the form X = A / Y then X * Y cannot overflow.
1050 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1051 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1055 // (X rem Y) / Y -> 0
1056 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1057 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1058 return Constant::getNullValue(Op0->getType());
1060 // If the operation is with the result of a select instruction, check whether
1061 // operating on either branch of the select always yields the same value.
1062 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1063 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1066 // If the operation is with the result of a phi instruction, check whether
1067 // operating on all incoming values of the phi always yields the same value.
1068 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1069 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1075 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1076 /// fold the result. If not, this returns null.
1077 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1078 unsigned MaxRecurse) {
1079 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1085 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1086 const TargetLibraryInfo *TLI,
1087 const DominatorTree *DT,
1088 AssumptionTracker *AT,
1089 const Instruction *CxtI) {
1090 return ::SimplifySDivInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1094 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1095 /// fold the result. If not, this returns null.
1096 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1097 unsigned MaxRecurse) {
1098 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1104 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1105 const TargetLibraryInfo *TLI,
1106 const DominatorTree *DT,
1107 AssumptionTracker *AT,
1108 const Instruction *CxtI) {
1109 return ::SimplifyUDivInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1113 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, const Query &Q,
1115 // undef / X -> undef (the undef could be a snan).
1116 if (match(Op0, m_Undef()))
1119 // X / undef -> undef
1120 if (match(Op1, m_Undef()))
1126 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1127 const TargetLibraryInfo *TLI,
1128 const DominatorTree *DT,
1129 AssumptionTracker *AT,
1130 const Instruction *CxtI) {
1131 return ::SimplifyFDivInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1135 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1136 /// fold the result. If not, this returns null.
1137 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1138 const Query &Q, unsigned MaxRecurse) {
1139 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1140 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1141 Constant *Ops[] = { C0, C1 };
1142 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1146 // X % undef -> undef
1147 if (match(Op1, m_Undef()))
1151 if (match(Op0, m_Undef()))
1152 return Constant::getNullValue(Op0->getType());
1154 // 0 % X -> 0, we don't need to preserve faults!
1155 if (match(Op0, m_Zero()))
1158 // X % 0 -> undef, we don't need to preserve faults!
1159 if (match(Op1, m_Zero()))
1160 return UndefValue::get(Op0->getType());
1163 if (match(Op1, m_One()))
1164 return Constant::getNullValue(Op0->getType());
1166 if (Op0->getType()->isIntegerTy(1))
1167 // It can't be remainder by zero, hence it must be remainder by one.
1168 return Constant::getNullValue(Op0->getType());
1172 return Constant::getNullValue(Op0->getType());
1174 // If the operation is with the result of a select instruction, check whether
1175 // operating on either branch of the select always yields the same value.
1176 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1177 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1180 // If the operation is with the result of a phi instruction, check whether
1181 // operating on all incoming values of the phi always yields the same value.
1182 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1183 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1189 /// SimplifySRemInst - Given operands for an SRem, see if we can
1190 /// fold the result. If not, this returns null.
1191 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1192 unsigned MaxRecurse) {
1193 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1199 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1200 const TargetLibraryInfo *TLI,
1201 const DominatorTree *DT,
1202 AssumptionTracker *AT,
1203 const Instruction *CxtI) {
1204 return ::SimplifySRemInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1208 /// SimplifyURemInst - Given operands for a URem, see if we can
1209 /// fold the result. If not, this returns null.
1210 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1211 unsigned MaxRecurse) {
1212 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1218 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1219 const TargetLibraryInfo *TLI,
1220 const DominatorTree *DT,
1221 AssumptionTracker *AT,
1222 const Instruction *CxtI) {
1223 return ::SimplifyURemInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1227 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, const Query &,
1229 // undef % X -> undef (the undef could be a snan).
1230 if (match(Op0, m_Undef()))
1233 // X % undef -> undef
1234 if (match(Op1, m_Undef()))
1240 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1241 const TargetLibraryInfo *TLI,
1242 const DominatorTree *DT,
1243 AssumptionTracker *AT,
1244 const Instruction *CxtI) {
1245 return ::SimplifyFRemInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1249 /// isUndefShift - Returns true if a shift by \c Amount always yields undef.
1250 static bool isUndefShift(Value *Amount) {
1251 Constant *C = dyn_cast<Constant>(Amount);
1255 // X shift by undef -> undef because it may shift by the bitwidth.
1256 if (isa<UndefValue>(C))
1259 // Shifting by the bitwidth or more is undefined.
1260 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1261 if (CI->getValue().getLimitedValue() >=
1262 CI->getType()->getScalarSizeInBits())
1265 // If all lanes of a vector shift are undefined the whole shift is.
1266 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1267 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1268 if (!isUndefShift(C->getAggregateElement(I)))
1276 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1277 /// fold the result. If not, this returns null.
1278 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1279 const Query &Q, unsigned MaxRecurse) {
1280 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1281 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1282 Constant *Ops[] = { C0, C1 };
1283 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1287 // 0 shift by X -> 0
1288 if (match(Op0, m_Zero()))
1291 // X shift by 0 -> X
1292 if (match(Op1, m_Zero()))
1295 // Fold undefined shifts.
1296 if (isUndefShift(Op1))
1297 return UndefValue::get(Op0->getType());
1299 // If the operation is with the result of a select instruction, check whether
1300 // operating on either branch of the select always yields the same value.
1301 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1302 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1305 // If the operation is with the result of a phi instruction, check whether
1306 // operating on all incoming values of the phi always yields the same value.
1307 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1308 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1314 /// SimplifyShlInst - Given operands for an Shl, see if we can
1315 /// fold the result. If not, this returns null.
1316 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1317 const Query &Q, unsigned MaxRecurse) {
1318 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1322 if (match(Op0, m_Undef()))
1323 return Constant::getNullValue(Op0->getType());
1325 // (X >> A) << A -> X
1327 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1332 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1333 const DataLayout *DL, const TargetLibraryInfo *TLI,
1334 const DominatorTree *DT, AssumptionTracker *AT,
1335 const Instruction *CxtI) {
1336 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT, AT, CxtI),
1340 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1341 /// fold the result. If not, this returns null.
1342 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1343 const Query &Q, unsigned MaxRecurse) {
1344 if (Value *V = SimplifyShift(Instruction::LShr, Op0, Op1, Q, MaxRecurse))
1349 return Constant::getNullValue(Op0->getType());
1352 if (match(Op0, m_Undef()))
1353 return Constant::getNullValue(Op0->getType());
1355 // (X << A) >> A -> X
1357 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1358 cast<OverflowingBinaryOperator>(Op0)->hasNoUnsignedWrap())
1364 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1365 const DataLayout *DL,
1366 const TargetLibraryInfo *TLI,
1367 const DominatorTree *DT,
1368 AssumptionTracker *AT,
1369 const Instruction *CxtI) {
1370 return ::SimplifyLShrInst(Op0, Op1, isExact, Query (DL, TLI, DT, AT, CxtI),
1374 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1375 /// fold the result. If not, this returns null.
1376 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1377 const Query &Q, unsigned MaxRecurse) {
1378 if (Value *V = SimplifyShift(Instruction::AShr, Op0, Op1, Q, MaxRecurse))
1383 return Constant::getNullValue(Op0->getType());
1385 // all ones >>a X -> all ones
1386 if (match(Op0, m_AllOnes()))
1389 // undef >>a X -> all ones
1390 if (match(Op0, m_Undef()))
1391 return Constant::getAllOnesValue(Op0->getType());
1393 // (X << A) >> A -> X
1395 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1396 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap())
1399 // Arithmetic shifting an all-sign-bit value is a no-op.
1400 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AT, Q.CxtI, Q.DT);
1401 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1407 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1408 const DataLayout *DL,
1409 const TargetLibraryInfo *TLI,
1410 const DominatorTree *DT,
1411 AssumptionTracker *AT,
1412 const Instruction *CxtI) {
1413 return ::SimplifyAShrInst(Op0, Op1, isExact, Query (DL, TLI, DT, AT, CxtI),
1417 /// SimplifyAndInst - Given operands for an And, see if we can
1418 /// fold the result. If not, this returns null.
1419 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1420 unsigned MaxRecurse) {
1421 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1422 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1423 Constant *Ops[] = { CLHS, CRHS };
1424 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1428 // Canonicalize the constant to the RHS.
1429 std::swap(Op0, Op1);
1433 if (match(Op1, m_Undef()))
1434 return Constant::getNullValue(Op0->getType());
1441 if (match(Op1, m_Zero()))
1445 if (match(Op1, m_AllOnes()))
1448 // A & ~A = ~A & A = 0
1449 if (match(Op0, m_Not(m_Specific(Op1))) ||
1450 match(Op1, m_Not(m_Specific(Op0))))
1451 return Constant::getNullValue(Op0->getType());
1454 Value *A = nullptr, *B = nullptr;
1455 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1456 (A == Op1 || B == Op1))
1460 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1461 (A == Op0 || B == Op0))
1464 // A & (-A) = A if A is a power of two or zero.
1465 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1466 match(Op1, m_Neg(m_Specific(Op0)))) {
1467 if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/true, 0, Q.AT, Q.CxtI, Q.DT))
1469 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true, 0, Q.AT, Q.CxtI, Q.DT))
1473 // Try some generic simplifications for associative operations.
1474 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1478 // And distributes over Or. Try some generic simplifications based on this.
1479 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1483 // And distributes over Xor. Try some generic simplifications based on this.
1484 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1488 // If the operation is with the result of a select instruction, check whether
1489 // operating on either branch of the select always yields the same value.
1490 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1491 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1495 // If the operation is with the result of a phi instruction, check whether
1496 // operating on all incoming values of the phi always yields the same value.
1497 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1498 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1505 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *DL,
1506 const TargetLibraryInfo *TLI,
1507 const DominatorTree *DT, AssumptionTracker *AT,
1508 const Instruction *CxtI) {
1509 return ::SimplifyAndInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1513 /// SimplifyOrInst - Given operands for an Or, see if we can
1514 /// fold the result. If not, this returns null.
1515 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1516 unsigned MaxRecurse) {
1517 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1518 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1519 Constant *Ops[] = { CLHS, CRHS };
1520 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1524 // Canonicalize the constant to the RHS.
1525 std::swap(Op0, Op1);
1529 if (match(Op1, m_Undef()))
1530 return Constant::getAllOnesValue(Op0->getType());
1537 if (match(Op1, m_Zero()))
1541 if (match(Op1, m_AllOnes()))
1544 // A | ~A = ~A | A = -1
1545 if (match(Op0, m_Not(m_Specific(Op1))) ||
1546 match(Op1, m_Not(m_Specific(Op0))))
1547 return Constant::getAllOnesValue(Op0->getType());
1550 Value *A = nullptr, *B = nullptr;
1551 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1552 (A == Op1 || B == Op1))
1556 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1557 (A == Op0 || B == Op0))
1560 // ~(A & ?) | A = -1
1561 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1562 (A == Op1 || B == Op1))
1563 return Constant::getAllOnesValue(Op1->getType());
1565 // A | ~(A & ?) = -1
1566 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1567 (A == Op0 || B == Op0))
1568 return Constant::getAllOnesValue(Op0->getType());
1570 // Try some generic simplifications for associative operations.
1571 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1575 // Or distributes over And. Try some generic simplifications based on this.
1576 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1580 // If the operation is with the result of a select instruction, check whether
1581 // operating on either branch of the select always yields the same value.
1582 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1583 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1588 Value *C = nullptr, *D = nullptr;
1589 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1590 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1591 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1592 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1593 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1594 // (A & C1)|(B & C2)
1595 // If we have: ((V + N) & C1) | (V & C2)
1596 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1597 // replace with V+N.
1599 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1600 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1601 // Add commutes, try both ways.
1602 if (V1 == B && MaskedValueIsZero(V2, C2->getValue(), Q.DL,
1603 0, Q.AT, Q.CxtI, Q.DT))
1605 if (V2 == B && MaskedValueIsZero(V1, C2->getValue(), Q.DL,
1606 0, Q.AT, Q.CxtI, Q.DT))
1609 // Or commutes, try both ways.
1610 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1611 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1612 // Add commutes, try both ways.
1613 if (V1 == A && MaskedValueIsZero(V2, C1->getValue(), Q.DL,
1614 0, Q.AT, Q.CxtI, Q.DT))
1616 if (V2 == A && MaskedValueIsZero(V1, C1->getValue(), Q.DL,
1617 0, Q.AT, Q.CxtI, Q.DT))
1623 // If the operation is with the result of a phi instruction, check whether
1624 // operating on all incoming values of the phi always yields the same value.
1625 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1626 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1632 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *DL,
1633 const TargetLibraryInfo *TLI,
1634 const DominatorTree *DT, AssumptionTracker *AT,
1635 const Instruction *CxtI) {
1636 return ::SimplifyOrInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1640 /// SimplifyXorInst - Given operands for a Xor, see if we can
1641 /// fold the result. If not, this returns null.
1642 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1643 unsigned MaxRecurse) {
1644 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1645 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1646 Constant *Ops[] = { CLHS, CRHS };
1647 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1651 // Canonicalize the constant to the RHS.
1652 std::swap(Op0, Op1);
1655 // A ^ undef -> undef
1656 if (match(Op1, m_Undef()))
1660 if (match(Op1, m_Zero()))
1665 return Constant::getNullValue(Op0->getType());
1667 // A ^ ~A = ~A ^ A = -1
1668 if (match(Op0, m_Not(m_Specific(Op1))) ||
1669 match(Op1, m_Not(m_Specific(Op0))))
1670 return Constant::getAllOnesValue(Op0->getType());
1672 // Try some generic simplifications for associative operations.
1673 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1677 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1678 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1679 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1680 // only if B and C are equal. If B and C are equal then (since we assume
1681 // that operands have already been simplified) "select(cond, B, C)" should
1682 // have been simplified to the common value of B and C already. Analysing
1683 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1684 // for threading over phi nodes.
1689 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *DL,
1690 const TargetLibraryInfo *TLI,
1691 const DominatorTree *DT, AssumptionTracker *AT,
1692 const Instruction *CxtI) {
1693 return ::SimplifyXorInst(Op0, Op1, Query (DL, TLI, DT, AT, CxtI),
1697 static Type *GetCompareTy(Value *Op) {
1698 return CmpInst::makeCmpResultType(Op->getType());
1701 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1702 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1703 /// otherwise return null. Helper function for analyzing max/min idioms.
1704 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1705 Value *LHS, Value *RHS) {
1706 SelectInst *SI = dyn_cast<SelectInst>(V);
1709 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1712 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1713 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1715 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1716 LHS == CmpRHS && RHS == CmpLHS)
1721 // A significant optimization not implemented here is assuming that alloca
1722 // addresses are not equal to incoming argument values. They don't *alias*,
1723 // as we say, but that doesn't mean they aren't equal, so we take a
1724 // conservative approach.
1726 // This is inspired in part by C++11 5.10p1:
1727 // "Two pointers of the same type compare equal if and only if they are both
1728 // null, both point to the same function, or both represent the same
1731 // This is pretty permissive.
1733 // It's also partly due to C11 6.5.9p6:
1734 // "Two pointers compare equal if and only if both are null pointers, both are
1735 // pointers to the same object (including a pointer to an object and a
1736 // subobject at its beginning) or function, both are pointers to one past the
1737 // last element of the same array object, or one is a pointer to one past the
1738 // end of one array object and the other is a pointer to the start of a
1739 // different array object that happens to immediately follow the first array
1740 // object in the address space.)
1742 // C11's version is more restrictive, however there's no reason why an argument
1743 // couldn't be a one-past-the-end value for a stack object in the caller and be
1744 // equal to the beginning of a stack object in the callee.
1746 // If the C and C++ standards are ever made sufficiently restrictive in this
1747 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1748 // this optimization.
1749 static Constant *computePointerICmp(const DataLayout *DL,
1750 const TargetLibraryInfo *TLI,
1751 CmpInst::Predicate Pred,
1752 Value *LHS, Value *RHS) {
1753 // First, skip past any trivial no-ops.
1754 LHS = LHS->stripPointerCasts();
1755 RHS = RHS->stripPointerCasts();
1757 // A non-null pointer is not equal to a null pointer.
1758 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) &&
1759 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1760 return ConstantInt::get(GetCompareTy(LHS),
1761 !CmpInst::isTrueWhenEqual(Pred));
1763 // We can only fold certain predicates on pointer comparisons.
1768 // Equality comaprisons are easy to fold.
1769 case CmpInst::ICMP_EQ:
1770 case CmpInst::ICMP_NE:
1773 // We can only handle unsigned relational comparisons because 'inbounds' on
1774 // a GEP only protects against unsigned wrapping.
1775 case CmpInst::ICMP_UGT:
1776 case CmpInst::ICMP_UGE:
1777 case CmpInst::ICMP_ULT:
1778 case CmpInst::ICMP_ULE:
1779 // However, we have to switch them to their signed variants to handle
1780 // negative indices from the base pointer.
1781 Pred = ICmpInst::getSignedPredicate(Pred);
1785 // Strip off any constant offsets so that we can reason about them.
1786 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
1787 // here and compare base addresses like AliasAnalysis does, however there are
1788 // numerous hazards. AliasAnalysis and its utilities rely on special rules
1789 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
1790 // doesn't need to guarantee pointer inequality when it says NoAlias.
1791 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
1792 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
1794 // If LHS and RHS are related via constant offsets to the same base
1795 // value, we can replace it with an icmp which just compares the offsets.
1797 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
1799 // Various optimizations for (in)equality comparisons.
1800 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
1801 // Different non-empty allocations that exist at the same time have
1802 // different addresses (if the program can tell). Global variables always
1803 // exist, so they always exist during the lifetime of each other and all
1804 // allocas. Two different allocas usually have different addresses...
1806 // However, if there's an @llvm.stackrestore dynamically in between two
1807 // allocas, they may have the same address. It's tempting to reduce the
1808 // scope of the problem by only looking at *static* allocas here. That would
1809 // cover the majority of allocas while significantly reducing the likelihood
1810 // of having an @llvm.stackrestore pop up in the middle. However, it's not
1811 // actually impossible for an @llvm.stackrestore to pop up in the middle of
1812 // an entry block. Also, if we have a block that's not attached to a
1813 // function, we can't tell if it's "static" under the current definition.
1814 // Theoretically, this problem could be fixed by creating a new kind of
1815 // instruction kind specifically for static allocas. Such a new instruction
1816 // could be required to be at the top of the entry block, thus preventing it
1817 // from being subject to a @llvm.stackrestore. Instcombine could even
1818 // convert regular allocas into these special allocas. It'd be nifty.
1819 // However, until then, this problem remains open.
1821 // So, we'll assume that two non-empty allocas have different addresses
1824 // With all that, if the offsets are within the bounds of their allocations
1825 // (and not one-past-the-end! so we can't use inbounds!), and their
1826 // allocations aren't the same, the pointers are not equal.
1828 // Note that it's not necessary to check for LHS being a global variable
1829 // address, due to canonicalization and constant folding.
1830 if (isa<AllocaInst>(LHS) &&
1831 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
1832 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
1833 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
1834 uint64_t LHSSize, RHSSize;
1835 if (LHSOffsetCI && RHSOffsetCI &&
1836 getObjectSize(LHS, LHSSize, DL, TLI) &&
1837 getObjectSize(RHS, RHSSize, DL, TLI)) {
1838 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
1839 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
1840 if (!LHSOffsetValue.isNegative() &&
1841 !RHSOffsetValue.isNegative() &&
1842 LHSOffsetValue.ult(LHSSize) &&
1843 RHSOffsetValue.ult(RHSSize)) {
1844 return ConstantInt::get(GetCompareTy(LHS),
1845 !CmpInst::isTrueWhenEqual(Pred));
1849 // Repeat the above check but this time without depending on DataLayout
1850 // or being able to compute a precise size.
1851 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
1852 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
1853 LHSOffset->isNullValue() &&
1854 RHSOffset->isNullValue())
1855 return ConstantInt::get(GetCompareTy(LHS),
1856 !CmpInst::isTrueWhenEqual(Pred));
1859 // Even if an non-inbounds GEP occurs along the path we can still optimize
1860 // equality comparisons concerning the result. We avoid walking the whole
1861 // chain again by starting where the last calls to
1862 // stripAndComputeConstantOffsets left off and accumulate the offsets.
1863 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
1864 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
1866 return ConstantExpr::getICmp(Pred,
1867 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
1868 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
1875 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
1876 /// fold the result. If not, this returns null.
1877 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
1878 const Query &Q, unsigned MaxRecurse) {
1879 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
1880 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
1882 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
1883 if (Constant *CRHS = dyn_cast<Constant>(RHS))
1884 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
1886 // If we have a constant, make sure it is on the RHS.
1887 std::swap(LHS, RHS);
1888 Pred = CmpInst::getSwappedPredicate(Pred);
1891 Type *ITy = GetCompareTy(LHS); // The return type.
1892 Type *OpTy = LHS->getType(); // The operand type.
1894 // icmp X, X -> true/false
1895 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
1896 // because X could be 0.
1897 if (LHS == RHS || isa<UndefValue>(RHS))
1898 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
1900 // Special case logic when the operands have i1 type.
1901 if (OpTy->getScalarType()->isIntegerTy(1)) {
1904 case ICmpInst::ICMP_EQ:
1906 if (match(RHS, m_One()))
1909 case ICmpInst::ICMP_NE:
1911 if (match(RHS, m_Zero()))
1914 case ICmpInst::ICMP_UGT:
1916 if (match(RHS, m_Zero()))
1919 case ICmpInst::ICMP_UGE:
1921 if (match(RHS, m_One()))
1924 case ICmpInst::ICMP_SLT:
1926 if (match(RHS, m_Zero()))
1929 case ICmpInst::ICMP_SLE:
1931 if (match(RHS, m_One()))
1937 // If we are comparing with zero then try hard since this is a common case.
1938 if (match(RHS, m_Zero())) {
1939 bool LHSKnownNonNegative, LHSKnownNegative;
1941 default: llvm_unreachable("Unknown ICmp predicate!");
1942 case ICmpInst::ICMP_ULT:
1943 return getFalse(ITy);
1944 case ICmpInst::ICMP_UGE:
1945 return getTrue(ITy);
1946 case ICmpInst::ICMP_EQ:
1947 case ICmpInst::ICMP_ULE:
1948 if (isKnownNonZero(LHS, Q.DL, 0, Q.AT, Q.CxtI, Q.DT))
1949 return getFalse(ITy);
1951 case ICmpInst::ICMP_NE:
1952 case ICmpInst::ICMP_UGT:
1953 if (isKnownNonZero(LHS, Q.DL, 0, Q.AT, Q.CxtI, Q.DT))
1954 return getTrue(ITy);
1956 case ICmpInst::ICMP_SLT:
1957 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL,
1958 0, Q.AT, Q.CxtI, Q.DT);
1959 if (LHSKnownNegative)
1960 return getTrue(ITy);
1961 if (LHSKnownNonNegative)
1962 return getFalse(ITy);
1964 case ICmpInst::ICMP_SLE:
1965 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL,
1966 0, Q.AT, Q.CxtI, Q.DT);
1967 if (LHSKnownNegative)
1968 return getTrue(ITy);
1969 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL,
1970 0, Q.AT, Q.CxtI, Q.DT))
1971 return getFalse(ITy);
1973 case ICmpInst::ICMP_SGE:
1974 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL,
1975 0, Q.AT, Q.CxtI, Q.DT);
1976 if (LHSKnownNegative)
1977 return getFalse(ITy);
1978 if (LHSKnownNonNegative)
1979 return getTrue(ITy);
1981 case ICmpInst::ICMP_SGT:
1982 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL,
1983 0, Q.AT, Q.CxtI, Q.DT);
1984 if (LHSKnownNegative)
1985 return getFalse(ITy);
1986 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL,
1987 0, Q.AT, Q.CxtI, Q.DT))
1988 return getTrue(ITy);
1993 // See if we are doing a comparison with a constant integer.
1994 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1995 // Rule out tautological comparisons (eg., ult 0 or uge 0).
1996 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
1997 if (RHS_CR.isEmptySet())
1998 return ConstantInt::getFalse(CI->getContext());
1999 if (RHS_CR.isFullSet())
2000 return ConstantInt::getTrue(CI->getContext());
2002 // Many binary operators with constant RHS have easy to compute constant
2003 // range. Use them to check whether the comparison is a tautology.
2004 unsigned Width = CI->getBitWidth();
2005 APInt Lower = APInt(Width, 0);
2006 APInt Upper = APInt(Width, 0);
2008 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
2009 // 'urem x, CI2' produces [0, CI2).
2010 Upper = CI2->getValue();
2011 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
2012 // 'srem x, CI2' produces (-|CI2|, |CI2|).
2013 Upper = CI2->getValue().abs();
2014 Lower = (-Upper) + 1;
2015 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
2016 // 'udiv CI2, x' produces [0, CI2].
2017 Upper = CI2->getValue() + 1;
2018 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
2019 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
2020 APInt NegOne = APInt::getAllOnesValue(Width);
2022 Upper = NegOne.udiv(CI2->getValue()) + 1;
2023 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) {
2024 if (CI2->isMinSignedValue()) {
2025 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2026 Lower = CI2->getValue();
2027 Upper = Lower.lshr(1) + 1;
2029 // 'sdiv CI2, x' produces [-|CI2|, |CI2|].
2030 Upper = CI2->getValue().abs() + 1;
2031 Lower = (-Upper) + 1;
2033 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
2034 APInt IntMin = APInt::getSignedMinValue(Width);
2035 APInt IntMax = APInt::getSignedMaxValue(Width);
2036 APInt Val = CI2->getValue();
2037 if (Val.isAllOnesValue()) {
2038 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2039 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2042 } else if (Val.countLeadingZeros() < Width - 1) {
2043 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2]
2044 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2045 Lower = IntMin.sdiv(Val);
2046 Upper = IntMax.sdiv(Val);
2047 if (Lower.sgt(Upper))
2048 std::swap(Lower, Upper);
2050 assert(Upper != Lower && "Upper part of range has wrapped!");
2052 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) {
2053 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)]
2054 Lower = CI2->getValue();
2055 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2056 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) {
2057 if (CI2->isNegative()) {
2058 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2]
2059 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1;
2060 Lower = CI2->getValue().shl(ShiftAmount);
2061 Upper = CI2->getValue() + 1;
2063 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1]
2064 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1;
2065 Lower = CI2->getValue();
2066 Upper = CI2->getValue().shl(ShiftAmount) + 1;
2068 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
2069 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
2070 APInt NegOne = APInt::getAllOnesValue(Width);
2071 if (CI2->getValue().ult(Width))
2072 Upper = NegOne.lshr(CI2->getValue()) + 1;
2073 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) {
2074 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2].
2075 unsigned ShiftAmount = Width - 1;
2076 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2077 ShiftAmount = CI2->getValue().countTrailingZeros();
2078 Lower = CI2->getValue().lshr(ShiftAmount);
2079 Upper = CI2->getValue() + 1;
2080 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
2081 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
2082 APInt IntMin = APInt::getSignedMinValue(Width);
2083 APInt IntMax = APInt::getSignedMaxValue(Width);
2084 if (CI2->getValue().ult(Width)) {
2085 Lower = IntMin.ashr(CI2->getValue());
2086 Upper = IntMax.ashr(CI2->getValue()) + 1;
2088 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) {
2089 unsigned ShiftAmount = Width - 1;
2090 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2091 ShiftAmount = CI2->getValue().countTrailingZeros();
2092 if (CI2->isNegative()) {
2093 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)]
2094 Lower = CI2->getValue();
2095 Upper = CI2->getValue().ashr(ShiftAmount) + 1;
2097 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2]
2098 Lower = CI2->getValue().ashr(ShiftAmount);
2099 Upper = CI2->getValue() + 1;
2101 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
2102 // 'or x, CI2' produces [CI2, UINT_MAX].
2103 Lower = CI2->getValue();
2104 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
2105 // 'and x, CI2' produces [0, CI2].
2106 Upper = CI2->getValue() + 1;
2108 if (Lower != Upper) {
2109 ConstantRange LHS_CR = ConstantRange(Lower, Upper);
2110 if (RHS_CR.contains(LHS_CR))
2111 return ConstantInt::getTrue(RHS->getContext());
2112 if (RHS_CR.inverse().contains(LHS_CR))
2113 return ConstantInt::getFalse(RHS->getContext());
2117 // Compare of cast, for example (zext X) != 0 -> X != 0
2118 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
2119 Instruction *LI = cast<CastInst>(LHS);
2120 Value *SrcOp = LI->getOperand(0);
2121 Type *SrcTy = SrcOp->getType();
2122 Type *DstTy = LI->getType();
2124 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
2125 // if the integer type is the same size as the pointer type.
2126 if (MaxRecurse && Q.DL && isa<PtrToIntInst>(LI) &&
2127 Q.DL->getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
2128 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2129 // Transfer the cast to the constant.
2130 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
2131 ConstantExpr::getIntToPtr(RHSC, SrcTy),
2134 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
2135 if (RI->getOperand(0)->getType() == SrcTy)
2136 // Compare without the cast.
2137 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2143 if (isa<ZExtInst>(LHS)) {
2144 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
2146 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
2147 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2148 // Compare X and Y. Note that signed predicates become unsigned.
2149 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2150 SrcOp, RI->getOperand(0), Q,
2154 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
2155 // too. If not, then try to deduce the result of the comparison.
2156 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2157 // Compute the constant that would happen if we truncated to SrcTy then
2158 // reextended to DstTy.
2159 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2160 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
2162 // If the re-extended constant didn't change then this is effectively
2163 // also a case of comparing two zero-extended values.
2164 if (RExt == CI && MaxRecurse)
2165 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2166 SrcOp, Trunc, Q, MaxRecurse-1))
2169 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
2170 // there. Use this to work out the result of the comparison.
2173 default: llvm_unreachable("Unknown ICmp predicate!");
2175 case ICmpInst::ICMP_EQ:
2176 case ICmpInst::ICMP_UGT:
2177 case ICmpInst::ICMP_UGE:
2178 return ConstantInt::getFalse(CI->getContext());
2180 case ICmpInst::ICMP_NE:
2181 case ICmpInst::ICMP_ULT:
2182 case ICmpInst::ICMP_ULE:
2183 return ConstantInt::getTrue(CI->getContext());
2185 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
2186 // is non-negative then LHS <s RHS.
2187 case ICmpInst::ICMP_SGT:
2188 case ICmpInst::ICMP_SGE:
2189 return CI->getValue().isNegative() ?
2190 ConstantInt::getTrue(CI->getContext()) :
2191 ConstantInt::getFalse(CI->getContext());
2193 case ICmpInst::ICMP_SLT:
2194 case ICmpInst::ICMP_SLE:
2195 return CI->getValue().isNegative() ?
2196 ConstantInt::getFalse(CI->getContext()) :
2197 ConstantInt::getTrue(CI->getContext());
2203 if (isa<SExtInst>(LHS)) {
2204 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
2206 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
2207 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2208 // Compare X and Y. Note that the predicate does not change.
2209 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2213 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
2214 // too. If not, then try to deduce the result of the comparison.
2215 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2216 // Compute the constant that would happen if we truncated to SrcTy then
2217 // reextended to DstTy.
2218 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2219 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
2221 // If the re-extended constant didn't change then this is effectively
2222 // also a case of comparing two sign-extended values.
2223 if (RExt == CI && MaxRecurse)
2224 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
2227 // Otherwise the upper bits of LHS are all equal, while RHS has varying
2228 // bits there. Use this to work out the result of the comparison.
2231 default: llvm_unreachable("Unknown ICmp predicate!");
2232 case ICmpInst::ICMP_EQ:
2233 return ConstantInt::getFalse(CI->getContext());
2234 case ICmpInst::ICMP_NE:
2235 return ConstantInt::getTrue(CI->getContext());
2237 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
2239 case ICmpInst::ICMP_SGT:
2240 case ICmpInst::ICMP_SGE:
2241 return CI->getValue().isNegative() ?
2242 ConstantInt::getTrue(CI->getContext()) :
2243 ConstantInt::getFalse(CI->getContext());
2244 case ICmpInst::ICMP_SLT:
2245 case ICmpInst::ICMP_SLE:
2246 return CI->getValue().isNegative() ?
2247 ConstantInt::getFalse(CI->getContext()) :
2248 ConstantInt::getTrue(CI->getContext());
2250 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2252 case ICmpInst::ICMP_UGT:
2253 case ICmpInst::ICMP_UGE:
2254 // Comparison is true iff the LHS <s 0.
2256 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2257 Constant::getNullValue(SrcTy),
2261 case ICmpInst::ICMP_ULT:
2262 case ICmpInst::ICMP_ULE:
2263 // Comparison is true iff the LHS >=s 0.
2265 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2266 Constant::getNullValue(SrcTy),
2276 // If a bit is known to be zero for A and known to be one for B,
2277 // then A and B cannot be equal.
2278 if (ICmpInst::isEquality(Pred)) {
2279 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2280 uint32_t BitWidth = CI->getBitWidth();
2281 APInt LHSKnownZero(BitWidth, 0);
2282 APInt LHSKnownOne(BitWidth, 0);
2283 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL,
2284 0, Q.AT, Q.CxtI, Q.DT);
2285 APInt RHSKnownZero(BitWidth, 0);
2286 APInt RHSKnownOne(BitWidth, 0);
2287 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne, Q.DL,
2288 0, Q.AT, Q.CxtI, Q.DT);
2289 if (((LHSKnownOne & RHSKnownZero) != 0) ||
2290 ((LHSKnownZero & RHSKnownOne) != 0))
2291 return (Pred == ICmpInst::ICMP_EQ)
2292 ? ConstantInt::getFalse(CI->getContext())
2293 : ConstantInt::getTrue(CI->getContext());
2297 // Special logic for binary operators.
2298 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2299 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2300 if (MaxRecurse && (LBO || RBO)) {
2301 // Analyze the case when either LHS or RHS is an add instruction.
2302 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2303 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2304 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2305 if (LBO && LBO->getOpcode() == Instruction::Add) {
2306 A = LBO->getOperand(0); B = LBO->getOperand(1);
2307 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2308 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2309 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2311 if (RBO && RBO->getOpcode() == Instruction::Add) {
2312 C = RBO->getOperand(0); D = RBO->getOperand(1);
2313 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2314 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2315 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2318 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2319 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2320 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2321 Constant::getNullValue(RHS->getType()),
2325 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2326 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2327 if (Value *V = SimplifyICmpInst(Pred,
2328 Constant::getNullValue(LHS->getType()),
2329 C == LHS ? D : C, Q, MaxRecurse-1))
2332 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2333 if (A && C && (A == C || A == D || B == C || B == D) &&
2334 NoLHSWrapProblem && NoRHSWrapProblem) {
2335 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2338 // C + B == C + D -> B == D
2341 } else if (A == D) {
2342 // D + B == C + D -> B == C
2345 } else if (B == C) {
2346 // A + C == C + D -> A == D
2351 // A + D == C + D -> A == C
2355 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2360 // 0 - (zext X) pred C
2361 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2362 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2363 if (RHSC->getValue().isStrictlyPositive()) {
2364 if (Pred == ICmpInst::ICMP_SLT)
2365 return ConstantInt::getTrue(RHSC->getContext());
2366 if (Pred == ICmpInst::ICMP_SGE)
2367 return ConstantInt::getFalse(RHSC->getContext());
2368 if (Pred == ICmpInst::ICMP_EQ)
2369 return ConstantInt::getFalse(RHSC->getContext());
2370 if (Pred == ICmpInst::ICMP_NE)
2371 return ConstantInt::getTrue(RHSC->getContext());
2373 if (RHSC->getValue().isNonNegative()) {
2374 if (Pred == ICmpInst::ICMP_SLE)
2375 return ConstantInt::getTrue(RHSC->getContext());
2376 if (Pred == ICmpInst::ICMP_SGT)
2377 return ConstantInt::getFalse(RHSC->getContext());
2382 // icmp pred (urem X, Y), Y
2383 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2384 bool KnownNonNegative, KnownNegative;
2388 case ICmpInst::ICMP_SGT:
2389 case ICmpInst::ICMP_SGE:
2390 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL,
2391 0, Q.AT, Q.CxtI, Q.DT);
2392 if (!KnownNonNegative)
2395 case ICmpInst::ICMP_EQ:
2396 case ICmpInst::ICMP_UGT:
2397 case ICmpInst::ICMP_UGE:
2398 return getFalse(ITy);
2399 case ICmpInst::ICMP_SLT:
2400 case ICmpInst::ICMP_SLE:
2401 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL,
2402 0, Q.AT, Q.CxtI, Q.DT);
2403 if (!KnownNonNegative)
2406 case ICmpInst::ICMP_NE:
2407 case ICmpInst::ICMP_ULT:
2408 case ICmpInst::ICMP_ULE:
2409 return getTrue(ITy);
2413 // icmp pred X, (urem Y, X)
2414 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2415 bool KnownNonNegative, KnownNegative;
2419 case ICmpInst::ICMP_SGT:
2420 case ICmpInst::ICMP_SGE:
2421 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL,
2422 0, Q.AT, Q.CxtI, Q.DT);
2423 if (!KnownNonNegative)
2426 case ICmpInst::ICMP_NE:
2427 case ICmpInst::ICMP_UGT:
2428 case ICmpInst::ICMP_UGE:
2429 return getTrue(ITy);
2430 case ICmpInst::ICMP_SLT:
2431 case ICmpInst::ICMP_SLE:
2432 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL,
2433 0, Q.AT, Q.CxtI, Q.DT);
2434 if (!KnownNonNegative)
2437 case ICmpInst::ICMP_EQ:
2438 case ICmpInst::ICMP_ULT:
2439 case ICmpInst::ICMP_ULE:
2440 return getFalse(ITy);
2445 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2446 // icmp pred (X /u Y), X
2447 if (Pred == ICmpInst::ICMP_UGT)
2448 return getFalse(ITy);
2449 if (Pred == ICmpInst::ICMP_ULE)
2450 return getTrue(ITy);
2457 // where CI2 is a power of 2 and CI isn't
2458 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2459 const APInt *CI2Val, *CIVal = &CI->getValue();
2460 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2461 CI2Val->isPowerOf2()) {
2462 if (!CIVal->isPowerOf2()) {
2463 // CI2 << X can equal zero in some circumstances,
2464 // this simplification is unsafe if CI is zero.
2466 // We know it is safe if:
2467 // - The shift is nsw, we can't shift out the one bit.
2468 // - The shift is nuw, we can't shift out the one bit.
2471 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2472 *CI2Val == 1 || !CI->isZero()) {
2473 if (Pred == ICmpInst::ICMP_EQ)
2474 return ConstantInt::getFalse(RHS->getContext());
2475 if (Pred == ICmpInst::ICMP_NE)
2476 return ConstantInt::getTrue(RHS->getContext());
2479 if (CIVal->isSignBit() && *CI2Val == 1) {
2480 if (Pred == ICmpInst::ICMP_UGT)
2481 return ConstantInt::getFalse(RHS->getContext());
2482 if (Pred == ICmpInst::ICMP_ULE)
2483 return ConstantInt::getTrue(RHS->getContext());
2488 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2489 LBO->getOperand(1) == RBO->getOperand(1)) {
2490 switch (LBO->getOpcode()) {
2492 case Instruction::UDiv:
2493 case Instruction::LShr:
2494 if (ICmpInst::isSigned(Pred))
2497 case Instruction::SDiv:
2498 case Instruction::AShr:
2499 if (!LBO->isExact() || !RBO->isExact())
2501 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2502 RBO->getOperand(0), Q, MaxRecurse-1))
2505 case Instruction::Shl: {
2506 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2507 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2510 if (!NSW && ICmpInst::isSigned(Pred))
2512 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2513 RBO->getOperand(0), Q, MaxRecurse-1))
2520 // Simplify comparisons involving max/min.
2522 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2523 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2525 // Signed variants on "max(a,b)>=a -> true".
2526 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2527 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2528 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2529 // We analyze this as smax(A, B) pred A.
2531 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2532 (A == LHS || B == LHS)) {
2533 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2534 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2535 // We analyze this as smax(A, B) swapped-pred A.
2536 P = CmpInst::getSwappedPredicate(Pred);
2537 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2538 (A == RHS || B == RHS)) {
2539 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2540 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2541 // We analyze this as smax(-A, -B) swapped-pred -A.
2542 // Note that we do not need to actually form -A or -B thanks to EqP.
2543 P = CmpInst::getSwappedPredicate(Pred);
2544 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2545 (A == LHS || B == LHS)) {
2546 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2547 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2548 // We analyze this as smax(-A, -B) pred -A.
2549 // Note that we do not need to actually form -A or -B thanks to EqP.
2552 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2553 // Cases correspond to "max(A, B) p A".
2557 case CmpInst::ICMP_EQ:
2558 case CmpInst::ICMP_SLE:
2559 // Equivalent to "A EqP B". This may be the same as the condition tested
2560 // in the max/min; if so, we can just return that.
2561 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2563 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2565 // Otherwise, see if "A EqP B" simplifies.
2567 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2570 case CmpInst::ICMP_NE:
2571 case CmpInst::ICMP_SGT: {
2572 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2573 // Equivalent to "A InvEqP B". This may be the same as the condition
2574 // tested in the max/min; if so, we can just return that.
2575 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2577 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2579 // Otherwise, see if "A InvEqP B" simplifies.
2581 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2585 case CmpInst::ICMP_SGE:
2587 return getTrue(ITy);
2588 case CmpInst::ICMP_SLT:
2590 return getFalse(ITy);
2594 // Unsigned variants on "max(a,b)>=a -> true".
2595 P = CmpInst::BAD_ICMP_PREDICATE;
2596 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2597 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2598 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2599 // We analyze this as umax(A, B) pred A.
2601 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2602 (A == LHS || B == LHS)) {
2603 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2604 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2605 // We analyze this as umax(A, B) swapped-pred A.
2606 P = CmpInst::getSwappedPredicate(Pred);
2607 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2608 (A == RHS || B == RHS)) {
2609 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2610 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2611 // We analyze this as umax(-A, -B) swapped-pred -A.
2612 // Note that we do not need to actually form -A or -B thanks to EqP.
2613 P = CmpInst::getSwappedPredicate(Pred);
2614 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2615 (A == LHS || B == LHS)) {
2616 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2617 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2618 // We analyze this as umax(-A, -B) pred -A.
2619 // Note that we do not need to actually form -A or -B thanks to EqP.
2622 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2623 // Cases correspond to "max(A, B) p A".
2627 case CmpInst::ICMP_EQ:
2628 case CmpInst::ICMP_ULE:
2629 // Equivalent to "A EqP B". This may be the same as the condition tested
2630 // in the max/min; if so, we can just return that.
2631 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2633 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2635 // Otherwise, see if "A EqP B" simplifies.
2637 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2640 case CmpInst::ICMP_NE:
2641 case CmpInst::ICMP_UGT: {
2642 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2643 // Equivalent to "A InvEqP B". This may be the same as the condition
2644 // tested in the max/min; if so, we can just return that.
2645 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2647 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2649 // Otherwise, see if "A InvEqP B" simplifies.
2651 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2655 case CmpInst::ICMP_UGE:
2657 return getTrue(ITy);
2658 case CmpInst::ICMP_ULT:
2660 return getFalse(ITy);
2664 // Variants on "max(x,y) >= min(x,z)".
2666 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2667 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2668 (A == C || A == D || B == C || B == D)) {
2669 // max(x, ?) pred min(x, ?).
2670 if (Pred == CmpInst::ICMP_SGE)
2672 return getTrue(ITy);
2673 if (Pred == CmpInst::ICMP_SLT)
2675 return getFalse(ITy);
2676 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2677 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2678 (A == C || A == D || B == C || B == D)) {
2679 // min(x, ?) pred max(x, ?).
2680 if (Pred == CmpInst::ICMP_SLE)
2682 return getTrue(ITy);
2683 if (Pred == CmpInst::ICMP_SGT)
2685 return getFalse(ITy);
2686 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2687 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2688 (A == C || A == D || B == C || B == D)) {
2689 // max(x, ?) pred min(x, ?).
2690 if (Pred == CmpInst::ICMP_UGE)
2692 return getTrue(ITy);
2693 if (Pred == CmpInst::ICMP_ULT)
2695 return getFalse(ITy);
2696 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2697 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2698 (A == C || A == D || B == C || B == D)) {
2699 // min(x, ?) pred max(x, ?).
2700 if (Pred == CmpInst::ICMP_ULE)
2702 return getTrue(ITy);
2703 if (Pred == CmpInst::ICMP_UGT)
2705 return getFalse(ITy);
2708 // Simplify comparisons of related pointers using a powerful, recursive
2709 // GEP-walk when we have target data available..
2710 if (LHS->getType()->isPointerTy())
2711 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS))
2714 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
2715 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
2716 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
2717 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
2718 (ICmpInst::isEquality(Pred) ||
2719 (GLHS->isInBounds() && GRHS->isInBounds() &&
2720 Pred == ICmpInst::getSignedPredicate(Pred)))) {
2721 // The bases are equal and the indices are constant. Build a constant
2722 // expression GEP with the same indices and a null base pointer to see
2723 // what constant folding can make out of it.
2724 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
2725 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
2726 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS);
2728 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
2729 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS);
2730 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
2735 // If the comparison is with the result of a select instruction, check whether
2736 // comparing with either branch of the select always yields the same value.
2737 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2738 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2741 // If the comparison is with the result of a phi instruction, check whether
2742 // doing the compare with each incoming phi value yields a common result.
2743 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2744 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2750 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2751 const DataLayout *DL,
2752 const TargetLibraryInfo *TLI,
2753 const DominatorTree *DT,
2754 AssumptionTracker *AT,
2755 Instruction *CxtI) {
2756 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT, AT, CxtI),
2760 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
2761 /// fold the result. If not, this returns null.
2762 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2763 const Query &Q, unsigned MaxRecurse) {
2764 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2765 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
2767 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2768 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2769 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
2771 // If we have a constant, make sure it is on the RHS.
2772 std::swap(LHS, RHS);
2773 Pred = CmpInst::getSwappedPredicate(Pred);
2776 // Fold trivial predicates.
2777 if (Pred == FCmpInst::FCMP_FALSE)
2778 return ConstantInt::get(GetCompareTy(LHS), 0);
2779 if (Pred == FCmpInst::FCMP_TRUE)
2780 return ConstantInt::get(GetCompareTy(LHS), 1);
2782 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef
2783 return UndefValue::get(GetCompareTy(LHS));
2785 // fcmp x,x -> true/false. Not all compares are foldable.
2787 if (CmpInst::isTrueWhenEqual(Pred))
2788 return ConstantInt::get(GetCompareTy(LHS), 1);
2789 if (CmpInst::isFalseWhenEqual(Pred))
2790 return ConstantInt::get(GetCompareTy(LHS), 0);
2793 // Handle fcmp with constant RHS
2794 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2795 // If the constant is a nan, see if we can fold the comparison based on it.
2796 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2797 if (CFP->getValueAPF().isNaN()) {
2798 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
2799 return ConstantInt::getFalse(CFP->getContext());
2800 assert(FCmpInst::isUnordered(Pred) &&
2801 "Comparison must be either ordered or unordered!");
2802 // True if unordered.
2803 return ConstantInt::getTrue(CFP->getContext());
2805 // Check whether the constant is an infinity.
2806 if (CFP->getValueAPF().isInfinity()) {
2807 if (CFP->getValueAPF().isNegative()) {
2809 case FCmpInst::FCMP_OLT:
2810 // No value is ordered and less than negative infinity.
2811 return ConstantInt::getFalse(CFP->getContext());
2812 case FCmpInst::FCMP_UGE:
2813 // All values are unordered with or at least negative infinity.
2814 return ConstantInt::getTrue(CFP->getContext());
2820 case FCmpInst::FCMP_OGT:
2821 // No value is ordered and greater than infinity.
2822 return ConstantInt::getFalse(CFP->getContext());
2823 case FCmpInst::FCMP_ULE:
2824 // All values are unordered with and at most infinity.
2825 return ConstantInt::getTrue(CFP->getContext());
2834 // If the comparison is with the result of a select instruction, check whether
2835 // comparing with either branch of the select always yields the same value.
2836 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2837 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2840 // If the comparison is with the result of a phi instruction, check whether
2841 // doing the compare with each incoming phi value yields a common result.
2842 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2843 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2849 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2850 const DataLayout *DL,
2851 const TargetLibraryInfo *TLI,
2852 const DominatorTree *DT,
2853 AssumptionTracker *AT,
2854 const Instruction *CxtI) {
2855 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT, AT, CxtI),
2859 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
2860 /// the result. If not, this returns null.
2861 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
2862 Value *FalseVal, const Query &Q,
2863 unsigned MaxRecurse) {
2864 // select true, X, Y -> X
2865 // select false, X, Y -> Y
2866 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
2867 if (CB->isAllOnesValue())
2869 if (CB->isNullValue())
2873 // select C, X, X -> X
2874 if (TrueVal == FalseVal)
2877 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
2878 if (isa<Constant>(TrueVal))
2882 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
2884 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
2890 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
2891 const DataLayout *DL,
2892 const TargetLibraryInfo *TLI,
2893 const DominatorTree *DT,
2894 AssumptionTracker *AT,
2895 const Instruction *CxtI) {
2896 return ::SimplifySelectInst(Cond, TrueVal, FalseVal,
2897 Query (DL, TLI, DT, AT, CxtI), RecursionLimit);
2900 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
2901 /// fold the result. If not, this returns null.
2902 static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) {
2903 // The type of the GEP pointer operand.
2904 PointerType *PtrTy = cast<PointerType>(Ops[0]->getType()->getScalarType());
2905 unsigned AS = PtrTy->getAddressSpace();
2907 // getelementptr P -> P.
2908 if (Ops.size() == 1)
2911 // Compute the (pointer) type returned by the GEP instruction.
2912 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1));
2913 Type *GEPTy = PointerType::get(LastType, AS);
2914 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
2915 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
2917 if (isa<UndefValue>(Ops[0]))
2918 return UndefValue::get(GEPTy);
2920 if (Ops.size() == 2) {
2921 // getelementptr P, 0 -> P.
2922 if (match(Ops[1], m_Zero()))
2925 Type *Ty = PtrTy->getElementType();
2926 if (Q.DL && Ty->isSized()) {
2929 uint64_t TyAllocSize = Q.DL->getTypeAllocSize(Ty);
2930 // getelementptr P, N -> P if P points to a type of zero size.
2931 if (TyAllocSize == 0)
2934 // The following transforms are only safe if the ptrtoint cast
2935 // doesn't truncate the pointers.
2936 if (Ops[1]->getType()->getScalarSizeInBits() ==
2937 Q.DL->getPointerSizeInBits(AS)) {
2938 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
2939 if (match(P, m_Zero()))
2940 return Constant::getNullValue(GEPTy);
2942 if (match(P, m_PtrToInt(m_Value(Temp))))
2943 if (Temp->getType() == GEPTy)
2948 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
2949 if (TyAllocSize == 1 &&
2950 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
2951 if (Value *R = PtrToIntOrZero(P))
2954 // getelementptr V, (ashr (sub P, V), C) -> Q
2955 // if P points to a type of size 1 << C.
2957 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
2958 m_ConstantInt(C))) &&
2959 TyAllocSize == 1ULL << C)
2960 if (Value *R = PtrToIntOrZero(P))
2963 // getelementptr V, (sdiv (sub P, V), C) -> Q
2964 // if P points to a type of size C.
2966 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
2967 m_SpecificInt(TyAllocSize))))
2968 if (Value *R = PtrToIntOrZero(P))
2974 // Check to see if this is constant foldable.
2975 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2976 if (!isa<Constant>(Ops[i]))
2979 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1));
2982 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *DL,
2983 const TargetLibraryInfo *TLI,
2984 const DominatorTree *DT, AssumptionTracker *AT,
2985 const Instruction *CxtI) {
2986 return ::SimplifyGEPInst(Ops, Query (DL, TLI, DT, AT, CxtI), RecursionLimit);
2989 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
2990 /// can fold the result. If not, this returns null.
2991 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
2992 ArrayRef<unsigned> Idxs, const Query &Q,
2994 if (Constant *CAgg = dyn_cast<Constant>(Agg))
2995 if (Constant *CVal = dyn_cast<Constant>(Val))
2996 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
2998 // insertvalue x, undef, n -> x
2999 if (match(Val, m_Undef()))
3002 // insertvalue x, (extractvalue y, n), n
3003 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3004 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3005 EV->getIndices() == Idxs) {
3006 // insertvalue undef, (extractvalue y, n), n -> y
3007 if (match(Agg, m_Undef()))
3008 return EV->getAggregateOperand();
3010 // insertvalue y, (extractvalue y, n), n -> y
3011 if (Agg == EV->getAggregateOperand())
3018 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3019 ArrayRef<unsigned> Idxs,
3020 const DataLayout *DL,
3021 const TargetLibraryInfo *TLI,
3022 const DominatorTree *DT,
3023 AssumptionTracker *AT,
3024 const Instruction *CxtI) {
3025 return ::SimplifyInsertValueInst(Agg, Val, Idxs,
3026 Query (DL, TLI, DT, AT, CxtI),
3030 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
3031 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
3032 // If all of the PHI's incoming values are the same then replace the PHI node
3033 // with the common value.
3034 Value *CommonValue = nullptr;
3035 bool HasUndefInput = false;
3036 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
3037 Value *Incoming = PN->getIncomingValue(i);
3038 // If the incoming value is the phi node itself, it can safely be skipped.
3039 if (Incoming == PN) continue;
3040 if (isa<UndefValue>(Incoming)) {
3041 // Remember that we saw an undef value, but otherwise ignore them.
3042 HasUndefInput = true;
3045 if (CommonValue && Incoming != CommonValue)
3046 return nullptr; // Not the same, bail out.
3047 CommonValue = Incoming;
3050 // If CommonValue is null then all of the incoming values were either undef or
3051 // equal to the phi node itself.
3053 return UndefValue::get(PN->getType());
3055 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3056 // instruction, we cannot return X as the result of the PHI node unless it
3057 // dominates the PHI block.
3059 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3064 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
3065 if (Constant *C = dyn_cast<Constant>(Op))
3066 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI);
3071 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *DL,
3072 const TargetLibraryInfo *TLI,
3073 const DominatorTree *DT,
3074 AssumptionTracker *AT,
3075 const Instruction *CxtI) {
3076 return ::SimplifyTruncInst(Op, Ty, Query (DL, TLI, DT, AT, CxtI),
3080 //=== Helper functions for higher up the class hierarchy.
3082 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
3083 /// fold the result. If not, this returns null.
3084 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3085 const Query &Q, unsigned MaxRecurse) {
3087 case Instruction::Add:
3088 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3090 case Instruction::FAdd:
3091 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3093 case Instruction::Sub:
3094 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3096 case Instruction::FSub:
3097 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3099 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
3100 case Instruction::FMul:
3101 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3102 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
3103 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
3104 case Instruction::FDiv: return SimplifyFDivInst(LHS, RHS, Q, MaxRecurse);
3105 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
3106 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
3107 case Instruction::FRem: return SimplifyFRemInst(LHS, RHS, Q, MaxRecurse);
3108 case Instruction::Shl:
3109 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3111 case Instruction::LShr:
3112 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3113 case Instruction::AShr:
3114 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3115 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
3116 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
3117 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
3119 if (Constant *CLHS = dyn_cast<Constant>(LHS))
3120 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
3121 Constant *COps[] = {CLHS, CRHS};
3122 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL,
3126 // If the operation is associative, try some generic simplifications.
3127 if (Instruction::isAssociative(Opcode))
3128 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
3131 // If the operation is with the result of a select instruction check whether
3132 // operating on either branch of the select always yields the same value.
3133 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3134 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
3137 // If the operation is with the result of a phi instruction, check whether
3138 // operating on all incoming values of the phi always yields the same value.
3139 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3140 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
3147 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3148 const DataLayout *DL, const TargetLibraryInfo *TLI,
3149 const DominatorTree *DT, AssumptionTracker *AT,
3150 const Instruction *CxtI) {
3151 return ::SimplifyBinOp(Opcode, LHS, RHS, Query (DL, TLI, DT, AT, CxtI),
3155 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
3156 /// fold the result.
3157 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3158 const Query &Q, unsigned MaxRecurse) {
3159 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
3160 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3161 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3164 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3165 const DataLayout *DL, const TargetLibraryInfo *TLI,
3166 const DominatorTree *DT, AssumptionTracker *AT,
3167 const Instruction *CxtI) {
3168 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT, AT, CxtI),
3172 static bool IsIdempotent(Intrinsic::ID ID) {
3174 default: return false;
3176 // Unary idempotent: f(f(x)) = f(x)
3177 case Intrinsic::fabs:
3178 case Intrinsic::floor:
3179 case Intrinsic::ceil:
3180 case Intrinsic::trunc:
3181 case Intrinsic::rint:
3182 case Intrinsic::nearbyint:
3183 case Intrinsic::round:
3188 template <typename IterTy>
3189 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd,
3190 const Query &Q, unsigned MaxRecurse) {
3191 // Perform idempotent optimizations
3192 if (!IsIdempotent(IID))
3196 if (std::distance(ArgBegin, ArgEnd) == 1)
3197 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
3198 if (II->getIntrinsicID() == IID)
3204 template <typename IterTy>
3205 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
3206 const Query &Q, unsigned MaxRecurse) {
3207 Type *Ty = V->getType();
3208 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
3209 Ty = PTy->getElementType();
3210 FunctionType *FTy = cast<FunctionType>(Ty);
3212 // call undef -> undef
3213 if (isa<UndefValue>(V))
3214 return UndefValue::get(FTy->getReturnType());
3216 Function *F = dyn_cast<Function>(V);
3220 if (unsigned IID = F->getIntrinsicID())
3222 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse))
3225 if (!canConstantFoldCallTo(F))
3228 SmallVector<Constant *, 4> ConstantArgs;
3229 ConstantArgs.reserve(ArgEnd - ArgBegin);
3230 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
3231 Constant *C = dyn_cast<Constant>(*I);
3234 ConstantArgs.push_back(C);
3237 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
3240 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
3241 User::op_iterator ArgEnd, const DataLayout *DL,
3242 const TargetLibraryInfo *TLI,
3243 const DominatorTree *DT, AssumptionTracker *AT,
3244 const Instruction *CxtI) {
3245 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AT, CxtI),
3249 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
3250 const DataLayout *DL, const TargetLibraryInfo *TLI,
3251 const DominatorTree *DT, AssumptionTracker *AT,
3252 const Instruction *CxtI) {
3253 return ::SimplifyCall(V, Args.begin(), Args.end(),
3254 Query(DL, TLI, DT, AT, CxtI), RecursionLimit);
3257 /// SimplifyInstruction - See if we can compute a simplified version of this
3258 /// instruction. If not, this returns null.
3259 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *DL,
3260 const TargetLibraryInfo *TLI,
3261 const DominatorTree *DT,
3262 AssumptionTracker *AT) {
3265 switch (I->getOpcode()) {
3267 Result = ConstantFoldInstruction(I, DL, TLI);
3269 case Instruction::FAdd:
3270 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
3271 I->getFastMathFlags(), DL, TLI, DT, AT, I);
3273 case Instruction::Add:
3274 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
3275 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3276 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3277 DL, TLI, DT, AT, I);
3279 case Instruction::FSub:
3280 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
3281 I->getFastMathFlags(), DL, TLI, DT, AT, I);
3283 case Instruction::Sub:
3284 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
3285 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3286 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3287 DL, TLI, DT, AT, I);
3289 case Instruction::FMul:
3290 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
3291 I->getFastMathFlags(), DL, TLI, DT, AT, I);
3293 case Instruction::Mul:
3294 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1),
3295 DL, TLI, DT, AT, I);
3297 case Instruction::SDiv:
3298 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1),
3299 DL, TLI, DT, AT, I);
3301 case Instruction::UDiv:
3302 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1),
3303 DL, TLI, DT, AT, I);
3305 case Instruction::FDiv:
3306 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
3307 DL, TLI, DT, AT, I);
3309 case Instruction::SRem:
3310 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1),
3311 DL, TLI, DT, AT, I);
3313 case Instruction::URem:
3314 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1),
3315 DL, TLI, DT, AT, I);
3317 case Instruction::FRem:
3318 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
3319 DL, TLI, DT, AT, I);
3321 case Instruction::Shl:
3322 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
3323 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3324 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3325 DL, TLI, DT, AT, I);
3327 case Instruction::LShr:
3328 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
3329 cast<BinaryOperator>(I)->isExact(),
3330 DL, TLI, DT, AT, I);
3332 case Instruction::AShr:
3333 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
3334 cast<BinaryOperator>(I)->isExact(),
3335 DL, TLI, DT, AT, I);
3337 case Instruction::And:
3338 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1),
3339 DL, TLI, DT, AT, I);
3341 case Instruction::Or:
3342 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3345 case Instruction::Xor:
3346 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1),
3347 DL, TLI, DT, AT, I);
3349 case Instruction::ICmp:
3350 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
3351 I->getOperand(0), I->getOperand(1),
3352 DL, TLI, DT, AT, I);
3354 case Instruction::FCmp:
3355 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
3356 I->getOperand(0), I->getOperand(1),
3357 DL, TLI, DT, AT, I);
3359 case Instruction::Select:
3360 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
3361 I->getOperand(2), DL, TLI, DT, AT, I);
3363 case Instruction::GetElementPtr: {
3364 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
3365 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AT, I);
3368 case Instruction::InsertValue: {
3369 InsertValueInst *IV = cast<InsertValueInst>(I);
3370 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
3371 IV->getInsertedValueOperand(),
3372 IV->getIndices(), DL, TLI, DT, AT, I);
3375 case Instruction::PHI:
3376 Result = SimplifyPHINode(cast<PHINode>(I), Query (DL, TLI, DT, AT, I));
3378 case Instruction::Call: {
3379 CallSite CS(cast<CallInst>(I));
3380 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
3381 DL, TLI, DT, AT, I);
3384 case Instruction::Trunc:
3385 Result = SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT,
3390 /// If called on unreachable code, the above logic may report that the
3391 /// instruction simplified to itself. Make life easier for users by
3392 /// detecting that case here, returning a safe value instead.
3393 return Result == I ? UndefValue::get(I->getType()) : Result;
3396 /// \brief Implementation of recursive simplification through an instructions
3399 /// This is the common implementation of the recursive simplification routines.
3400 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
3401 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
3402 /// instructions to process and attempt to simplify it using
3403 /// InstructionSimplify.
3405 /// This routine returns 'true' only when *it* simplifies something. The passed
3406 /// in simplified value does not count toward this.
3407 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
3408 const DataLayout *DL,
3409 const TargetLibraryInfo *TLI,
3410 const DominatorTree *DT,
3411 AssumptionTracker *AT) {
3412 bool Simplified = false;
3413 SmallSetVector<Instruction *, 8> Worklist;
3415 // If we have an explicit value to collapse to, do that round of the
3416 // simplification loop by hand initially.
3418 for (User *U : I->users())
3420 Worklist.insert(cast<Instruction>(U));
3422 // Replace the instruction with its simplified value.
3423 I->replaceAllUsesWith(SimpleV);
3425 // Gracefully handle edge cases where the instruction is not wired into any
3428 I->eraseFromParent();
3433 // Note that we must test the size on each iteration, the worklist can grow.
3434 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
3437 // See if this instruction simplifies.
3438 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AT);
3444 // Stash away all the uses of the old instruction so we can check them for
3445 // recursive simplifications after a RAUW. This is cheaper than checking all
3446 // uses of To on the recursive step in most cases.
3447 for (User *U : I->users())
3448 Worklist.insert(cast<Instruction>(U));
3450 // Replace the instruction with its simplified value.
3451 I->replaceAllUsesWith(SimpleV);
3453 // Gracefully handle edge cases where the instruction is not wired into any
3456 I->eraseFromParent();
3461 bool llvm::recursivelySimplifyInstruction(Instruction *I,
3462 const DataLayout *DL,
3463 const TargetLibraryInfo *TLI,
3464 const DominatorTree *DT,
3465 AssumptionTracker *AT) {
3466 return replaceAndRecursivelySimplifyImpl(I, nullptr, DL, TLI, DT, AT);
3469 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
3470 const DataLayout *DL,
3471 const TargetLibraryInfo *TLI,
3472 const DominatorTree *DT,
3473 AssumptionTracker *AT) {
3474 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
3475 assert(SimpleV && "Must provide a simplified value.");
3476 return replaceAndRecursivelySimplifyImpl(I, SimpleV, DL, TLI, DT, AT);