1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements routines for folding instructions into simpler forms
11 // that do not require creating new instructions. This does constant folding
12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15 // simplified: This is usually true and assuming it simplifies the logic (if
16 // they have not been simplified then results are correct but maybe suboptimal).
18 //===----------------------------------------------------------------------===//
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/LoopAnalysisManager.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/Analysis/VectorUtils.h"
32 #include "llvm/IR/ConstantRange.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/GetElementPtrTypeIterator.h"
36 #include "llvm/IR/GlobalAlias.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Support/KnownBits.h"
43 using namespace llvm::PatternMatch;
45 #define DEBUG_TYPE "instsimplify"
47 enum { RecursionLimit = 3 };
49 STATISTIC(NumExpand, "Number of expansions");
50 STATISTIC(NumReassoc, "Number of reassociations");
52 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
53 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
55 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
56 const SimplifyQuery &, unsigned);
57 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
59 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
60 const SimplifyQuery &Q, unsigned MaxRecurse);
61 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
62 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 static Value *SimplifyCastInst(unsigned, Value *, Type *,
64 const SimplifyQuery &, unsigned);
66 /// For a boolean type or a vector of boolean type, return false or a vector
67 /// with every element false.
68 static Constant *getFalse(Type *Ty) {
69 return ConstantInt::getFalse(Ty);
72 /// For a boolean type or a vector of boolean type, return true or a vector
73 /// with every element true.
74 static Constant *getTrue(Type *Ty) {
75 return ConstantInt::getTrue(Ty);
78 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
79 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
81 CmpInst *Cmp = dyn_cast<CmpInst>(V);
84 CmpInst::Predicate CPred = Cmp->getPredicate();
85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
86 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
92 /// Does the given value dominate the specified phi node?
93 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
94 Instruction *I = dyn_cast<Instruction>(V);
96 // Arguments and constants dominate all instructions.
99 // If we are processing instructions (and/or basic blocks) that have not been
100 // fully added to a function, the parent nodes may still be null. Simply
101 // return the conservative answer in these cases.
102 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
105 // If we have a DominatorTree then do a precise test.
107 if (!DT->isReachableFromEntry(P->getParent()))
109 if (!DT->isReachableFromEntry(I->getParent()))
111 return DT->dominates(I, P);
114 // Otherwise, if the instruction is in the entry block and is not an invoke,
115 // then it obviously dominates all phi nodes.
116 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
123 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
124 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
125 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
126 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
127 /// Returns the simplified value, or null if no simplification was performed.
128 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
129 Instruction::BinaryOps OpcodeToExpand,
130 const SimplifyQuery &Q, unsigned MaxRecurse) {
131 // Recursion is always used, so bail out at once if we already hit the limit.
135 // Check whether the expression has the form "(A op' B) op C".
136 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
137 if (Op0->getOpcode() == OpcodeToExpand) {
138 // It does! Try turning it into "(A op C) op' (B op C)".
139 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
140 // Do "A op C" and "B op C" both simplify?
141 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
142 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
143 // They do! Return "L op' R" if it simplifies or is already available.
144 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
145 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
146 && L == B && R == A)) {
150 // Otherwise return "L op' R" if it simplifies.
151 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
158 // Check whether the expression has the form "A op (B op' C)".
159 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
160 if (Op1->getOpcode() == OpcodeToExpand) {
161 // It does! Try turning it into "(A op B) op' (A op C)".
162 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
163 // Do "A op B" and "A op C" both simplify?
164 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
165 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
166 // They do! Return "L op' R" if it simplifies or is already available.
167 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
168 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
169 && L == C && R == B)) {
173 // Otherwise return "L op' R" if it simplifies.
174 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
184 /// Generic simplifications for associative binary operations.
185 /// Returns the simpler value, or null if none was found.
186 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
187 Value *LHS, Value *RHS,
188 const SimplifyQuery &Q,
189 unsigned MaxRecurse) {
190 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
192 // Recursion is always used, so bail out at once if we already hit the limit.
196 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
197 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
199 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
200 if (Op0 && Op0->getOpcode() == Opcode) {
201 Value *A = Op0->getOperand(0);
202 Value *B = Op0->getOperand(1);
205 // Does "B op C" simplify?
206 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
207 // It does! Return "A op V" if it simplifies or is already available.
208 // If V equals B then "A op V" is just the LHS.
209 if (V == B) return LHS;
210 // Otherwise return "A op V" if it simplifies.
211 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
218 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
219 if (Op1 && Op1->getOpcode() == Opcode) {
221 Value *B = Op1->getOperand(0);
222 Value *C = Op1->getOperand(1);
224 // Does "A op B" simplify?
225 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
226 // It does! Return "V op C" if it simplifies or is already available.
227 // If V equals B then "V op C" is just the RHS.
228 if (V == B) return RHS;
229 // Otherwise return "V op C" if it simplifies.
230 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
237 // The remaining transforms require commutativity as well as associativity.
238 if (!Instruction::isCommutative(Opcode))
241 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
242 if (Op0 && Op0->getOpcode() == Opcode) {
243 Value *A = Op0->getOperand(0);
244 Value *B = Op0->getOperand(1);
247 // Does "C op A" simplify?
248 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
249 // It does! Return "V op B" if it simplifies or is already available.
250 // If V equals A then "V op B" is just the LHS.
251 if (V == A) return LHS;
252 // Otherwise return "V op B" if it simplifies.
253 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
260 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
261 if (Op1 && Op1->getOpcode() == Opcode) {
263 Value *B = Op1->getOperand(0);
264 Value *C = Op1->getOperand(1);
266 // Does "C op A" simplify?
267 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
268 // It does! Return "B op V" if it simplifies or is already available.
269 // If V equals C then "B op V" is just the RHS.
270 if (V == C) return RHS;
271 // Otherwise return "B op V" if it simplifies.
272 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
282 /// In the case of a binary operation with a select instruction as an operand,
283 /// try to simplify the binop by seeing whether evaluating it on both branches
284 /// of the select results in the same value. Returns the common value if so,
285 /// otherwise returns null.
286 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
287 Value *RHS, const SimplifyQuery &Q,
288 unsigned MaxRecurse) {
289 // Recursion is always used, so bail out at once if we already hit the limit.
294 if (isa<SelectInst>(LHS)) {
295 SI = cast<SelectInst>(LHS);
297 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
298 SI = cast<SelectInst>(RHS);
301 // Evaluate the BinOp on the true and false branches of the select.
305 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
306 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
308 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
309 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
312 // If they simplified to the same value, then return the common value.
313 // If they both failed to simplify then return null.
317 // If one branch simplified to undef, return the other one.
318 if (TV && isa<UndefValue>(TV))
320 if (FV && isa<UndefValue>(FV))
323 // If applying the operation did not change the true and false select values,
324 // then the result of the binop is the select itself.
325 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
328 // If one branch simplified and the other did not, and the simplified
329 // value is equal to the unsimplified one, return the simplified value.
330 // For example, select (cond, X, X & Z) & Z -> X & Z.
331 if ((FV && !TV) || (TV && !FV)) {
332 // Check that the simplified value has the form "X op Y" where "op" is the
333 // same as the original operation.
334 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
335 if (Simplified && Simplified->getOpcode() == Opcode) {
336 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
337 // We already know that "op" is the same as for the simplified value. See
338 // if the operands match too. If so, return the simplified value.
339 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
340 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
341 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
342 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
343 Simplified->getOperand(1) == UnsimplifiedRHS)
345 if (Simplified->isCommutative() &&
346 Simplified->getOperand(1) == UnsimplifiedLHS &&
347 Simplified->getOperand(0) == UnsimplifiedRHS)
355 /// In the case of a comparison with a select instruction, try to simplify the
356 /// comparison by seeing whether both branches of the select result in the same
357 /// value. Returns the common value if so, otherwise returns null.
358 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
359 Value *RHS, const SimplifyQuery &Q,
360 unsigned MaxRecurse) {
361 // Recursion is always used, so bail out at once if we already hit the limit.
365 // Make sure the select is on the LHS.
366 if (!isa<SelectInst>(LHS)) {
368 Pred = CmpInst::getSwappedPredicate(Pred);
370 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
371 SelectInst *SI = cast<SelectInst>(LHS);
372 Value *Cond = SI->getCondition();
373 Value *TV = SI->getTrueValue();
374 Value *FV = SI->getFalseValue();
376 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
377 // Does "cmp TV, RHS" simplify?
378 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
380 // It not only simplified, it simplified to the select condition. Replace
382 TCmp = getTrue(Cond->getType());
384 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
385 // condition then we can replace it with 'true'. Otherwise give up.
386 if (!isSameCompare(Cond, Pred, TV, RHS))
388 TCmp = getTrue(Cond->getType());
391 // Does "cmp FV, RHS" simplify?
392 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
394 // It not only simplified, it simplified to the select condition. Replace
396 FCmp = getFalse(Cond->getType());
398 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
399 // condition then we can replace it with 'false'. Otherwise give up.
400 if (!isSameCompare(Cond, Pred, FV, RHS))
402 FCmp = getFalse(Cond->getType());
405 // If both sides simplified to the same value, then use it as the result of
406 // the original comparison.
410 // The remaining cases only make sense if the select condition has the same
411 // type as the result of the comparison, so bail out if this is not so.
412 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
414 // If the false value simplified to false, then the result of the compare
415 // is equal to "Cond && TCmp". This also catches the case when the false
416 // value simplified to false and the true value to true, returning "Cond".
417 if (match(FCmp, m_Zero()))
418 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
420 // If the true value simplified to true, then the result of the compare
421 // is equal to "Cond || FCmp".
422 if (match(TCmp, m_One()))
423 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
425 // Finally, if the false value simplified to true and the true value to
426 // false, then the result of the compare is equal to "!Cond".
427 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
429 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
436 /// In the case of a binary operation with an operand that is a PHI instruction,
437 /// try to simplify the binop by seeing whether evaluating it on the incoming
438 /// phi values yields the same result for every value. If so returns the common
439 /// value, otherwise returns null.
440 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
441 Value *RHS, const SimplifyQuery &Q,
442 unsigned MaxRecurse) {
443 // Recursion is always used, so bail out at once if we already hit the limit.
448 if (isa<PHINode>(LHS)) {
449 PI = cast<PHINode>(LHS);
450 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
451 if (!ValueDominatesPHI(RHS, PI, Q.DT))
454 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
455 PI = cast<PHINode>(RHS);
456 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
457 if (!ValueDominatesPHI(LHS, PI, Q.DT))
461 // Evaluate the BinOp on the incoming phi values.
462 Value *CommonValue = nullptr;
463 for (Value *Incoming : PI->incoming_values()) {
464 // If the incoming value is the phi node itself, it can safely be skipped.
465 if (Incoming == PI) continue;
466 Value *V = PI == LHS ?
467 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
468 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
469 // If the operation failed to simplify, or simplified to a different value
470 // to previously, then give up.
471 if (!V || (CommonValue && V != CommonValue))
479 /// In the case of a comparison with a PHI instruction, try to simplify the
480 /// comparison by seeing whether comparing with all of the incoming phi values
481 /// yields the same result every time. If so returns the common result,
482 /// otherwise returns null.
483 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
484 const SimplifyQuery &Q, unsigned MaxRecurse) {
485 // Recursion is always used, so bail out at once if we already hit the limit.
489 // Make sure the phi is on the LHS.
490 if (!isa<PHINode>(LHS)) {
492 Pred = CmpInst::getSwappedPredicate(Pred);
494 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
495 PHINode *PI = cast<PHINode>(LHS);
497 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
498 if (!ValueDominatesPHI(RHS, PI, Q.DT))
501 // Evaluate the BinOp on the incoming phi values.
502 Value *CommonValue = nullptr;
503 for (Value *Incoming : PI->incoming_values()) {
504 // If the incoming value is the phi node itself, it can safely be skipped.
505 if (Incoming == PI) continue;
506 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
507 // If the operation failed to simplify, or simplified to a different value
508 // to previously, then give up.
509 if (!V || (CommonValue && V != CommonValue))
517 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
518 Value *&Op0, Value *&Op1,
519 const SimplifyQuery &Q) {
520 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
521 if (auto *CRHS = dyn_cast<Constant>(Op1))
522 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
524 // Canonicalize the constant to the RHS if this is a commutative operation.
525 if (Instruction::isCommutative(Opcode))
531 /// Given operands for an Add, see if we can fold the result.
532 /// If not, this returns null.
533 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
534 const SimplifyQuery &Q, unsigned MaxRecurse) {
535 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
538 // X + undef -> undef
539 if (match(Op1, m_Undef()))
543 if (match(Op1, m_Zero()))
550 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
551 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
554 // X + ~X -> -1 since ~X = -X-1
555 Type *Ty = Op0->getType();
556 if (match(Op0, m_Not(m_Specific(Op1))) ||
557 match(Op1, m_Not(m_Specific(Op0))))
558 return Constant::getAllOnesValue(Ty);
560 // add nsw/nuw (xor Y, signmask), signmask --> Y
561 // The no-wrapping add guarantees that the top bit will be set by the add.
562 // Therefore, the xor must be clearing the already set sign bit of Y.
563 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
564 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
568 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
569 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
572 // Try some generic simplifications for associative operations.
573 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
577 // Threading Add over selects and phi nodes is pointless, so don't bother.
578 // Threading over the select in "A + select(cond, B, C)" means evaluating
579 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
580 // only if B and C are equal. If B and C are equal then (since we assume
581 // that operands have already been simplified) "select(cond, B, C)" should
582 // have been simplified to the common value of B and C already. Analysing
583 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
584 // for threading over phi nodes.
589 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
590 const SimplifyQuery &Query) {
591 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
594 /// \brief Compute the base pointer and cumulative constant offsets for V.
596 /// This strips all constant offsets off of V, leaving it the base pointer, and
597 /// accumulates the total constant offset applied in the returned constant. It
598 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
599 /// no constant offsets applied.
601 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
602 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
604 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
605 bool AllowNonInbounds = false) {
606 assert(V->getType()->getScalarType()->isPointerTy());
608 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
609 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
611 // Even though we don't look through PHI nodes, we could be called on an
612 // instruction in an unreachable block, which may be on a cycle.
613 SmallPtrSet<Value *, 4> Visited;
616 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
617 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
618 !GEP->accumulateConstantOffset(DL, Offset))
620 V = GEP->getPointerOperand();
621 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
622 V = cast<Operator>(V)->getOperand(0);
623 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
624 if (GA->isInterposable())
626 V = GA->getAliasee();
628 if (auto CS = CallSite(V))
629 if (Value *RV = CS.getReturnedArgOperand()) {
635 assert(V->getType()->getScalarType()->isPointerTy() &&
636 "Unexpected operand type!");
637 } while (Visited.insert(V).second);
639 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
640 if (V->getType()->isVectorTy())
641 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
646 /// \brief Compute the constant difference between two pointer values.
647 /// If the difference is not a constant, returns zero.
648 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
650 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
651 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
653 // If LHS and RHS are not related via constant offsets to the same base
654 // value, there is nothing we can do here.
658 // Otherwise, the difference of LHS - RHS can be computed as:
660 // = (LHSOffset + Base) - (RHSOffset + Base)
661 // = LHSOffset - RHSOffset
662 return ConstantExpr::getSub(LHSOffset, RHSOffset);
665 /// Given operands for a Sub, see if we can fold the result.
666 /// If not, this returns null.
667 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
668 const SimplifyQuery &Q, unsigned MaxRecurse) {
669 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
672 // X - undef -> undef
673 // undef - X -> undef
674 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
675 return UndefValue::get(Op0->getType());
678 if (match(Op1, m_Zero()))
683 return Constant::getNullValue(Op0->getType());
685 // Is this a negation?
686 if (match(Op0, m_Zero())) {
687 // 0 - X -> 0 if the sub is NUW.
691 unsigned BitWidth = Op1->getType()->getScalarSizeInBits();
692 KnownBits Known(BitWidth);
693 computeKnownBits(Op1, Known, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
694 if (Known.Zero.isMaxSignedValue()) {
695 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
696 // Op1 must be 0 because negating the minimum signed value is undefined.
700 // 0 - X -> X if X is 0 or the minimum signed value.
705 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
706 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
707 Value *X = nullptr, *Y = nullptr, *Z = Op1;
708 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
709 // See if "V === Y - Z" simplifies.
710 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
711 // It does! Now see if "X + V" simplifies.
712 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
713 // It does, we successfully reassociated!
717 // See if "V === X - Z" simplifies.
718 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
719 // It does! Now see if "Y + V" simplifies.
720 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
721 // It does, we successfully reassociated!
727 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
728 // For example, X - (X + 1) -> -1
730 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
731 // See if "V === X - Y" simplifies.
732 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
733 // It does! Now see if "V - Z" simplifies.
734 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
735 // It does, we successfully reassociated!
739 // See if "V === X - Z" simplifies.
740 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
741 // It does! Now see if "V - Y" simplifies.
742 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
743 // It does, we successfully reassociated!
749 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
750 // For example, X - (X - Y) -> Y.
752 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
753 // See if "V === Z - X" simplifies.
754 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
755 // It does! Now see if "V + Y" simplifies.
756 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
757 // It does, we successfully reassociated!
762 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
763 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
764 match(Op1, m_Trunc(m_Value(Y))))
765 if (X->getType() == Y->getType())
766 // See if "V === X - Y" simplifies.
767 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
768 // It does! Now see if "trunc V" simplifies.
769 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
771 // It does, return the simplified "trunc V".
774 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
775 if (match(Op0, m_PtrToInt(m_Value(X))) &&
776 match(Op1, m_PtrToInt(m_Value(Y))))
777 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
778 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
781 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
782 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
785 // Threading Sub over selects and phi nodes is pointless, so don't bother.
786 // Threading over the select in "A - select(cond, B, C)" means evaluating
787 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
788 // only if B and C are equal. If B and C are equal then (since we assume
789 // that operands have already been simplified) "select(cond, B, C)" should
790 // have been simplified to the common value of B and C already. Analysing
791 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
792 // for threading over phi nodes.
797 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
798 const SimplifyQuery &Q) {
799 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
802 /// Given operands for an FAdd, see if we can fold the result. If not, this
804 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
805 const SimplifyQuery &Q, unsigned MaxRecurse) {
806 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
810 if (match(Op1, m_NegZero()))
813 // fadd X, 0 ==> X, when we know X is not -0
814 if (match(Op1, m_Zero()) &&
815 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
818 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
819 // where nnan and ninf have to occur at least once somewhere in this
821 Value *SubOp = nullptr;
822 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
824 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
827 Instruction *FSub = cast<Instruction>(SubOp);
828 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
829 (FMF.noInfs() || FSub->hasNoInfs()))
830 return Constant::getNullValue(Op0->getType());
836 /// Given operands for an FSub, see if we can fold the result. If not, this
838 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
839 const SimplifyQuery &Q, unsigned MaxRecurse) {
840 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
844 if (match(Op1, m_Zero()))
847 // fsub X, -0 ==> X, when we know X is not -0
848 if (match(Op1, m_NegZero()) &&
849 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
852 // fsub -0.0, (fsub -0.0, X) ==> X
854 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
857 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
858 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
859 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
862 // fsub nnan x, x ==> 0.0
863 if (FMF.noNaNs() && Op0 == Op1)
864 return Constant::getNullValue(Op0->getType());
869 /// Given the operands for an FMul, see if we can fold the result
870 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
871 const SimplifyQuery &Q, unsigned MaxRecurse) {
872 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
876 if (match(Op1, m_FPOne()))
879 // fmul nnan nsz X, 0 ==> 0
880 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
886 /// Given operands for a Mul, see if we can fold the result.
887 /// If not, this returns null.
888 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
889 unsigned MaxRecurse) {
890 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
894 if (match(Op1, m_Undef()))
895 return Constant::getNullValue(Op0->getType());
898 if (match(Op1, m_Zero()))
902 if (match(Op1, m_One()))
905 // (X / Y) * Y -> X if the division is exact.
907 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
908 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
912 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
913 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
916 // Try some generic simplifications for associative operations.
917 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
921 // Mul distributes over Add. Try some generic simplifications based on this.
922 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
926 // If the operation is with the result of a select instruction, check whether
927 // operating on either branch of the select always yields the same value.
928 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
929 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
933 // If the operation is with the result of a phi instruction, check whether
934 // operating on all incoming values of the phi always yields the same value.
935 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
936 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
943 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
944 const SimplifyQuery &Q) {
945 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
949 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
950 const SimplifyQuery &Q) {
951 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
954 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
955 const SimplifyQuery &Q) {
956 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
959 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
960 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
963 /// Check for common or similar folds of integer division or integer remainder.
964 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
965 Type *Ty = Op0->getType();
967 // X / undef -> undef
968 // X % undef -> undef
969 if (match(Op1, m_Undef()))
974 // We don't need to preserve faults!
975 if (match(Op1, m_Zero()))
976 return UndefValue::get(Ty);
978 // If any element of a constant divisor vector is zero, the whole op is undef.
979 auto *Op1C = dyn_cast<Constant>(Op1);
980 if (Op1C && Ty->isVectorTy()) {
981 unsigned NumElts = Ty->getVectorNumElements();
982 for (unsigned i = 0; i != NumElts; ++i) {
983 Constant *Elt = Op1C->getAggregateElement(i);
984 if (Elt && Elt->isNullValue())
985 return UndefValue::get(Ty);
991 if (match(Op0, m_Undef()))
992 return Constant::getNullValue(Ty);
996 if (match(Op0, m_Zero()))
1002 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
1006 // If this is a boolean op (single-bit element type), we can't have
1007 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
1008 if (match(Op1, m_One()) || Ty->getScalarType()->isIntegerTy(1))
1009 return IsDiv ? Op0 : Constant::getNullValue(Ty);
1014 /// Given operands for an SDiv or UDiv, see if we can fold the result.
1015 /// If not, this returns null.
1016 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1017 const SimplifyQuery &Q, unsigned MaxRecurse) {
1018 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1021 if (Value *V = simplifyDivRem(Op0, Op1, true))
1024 bool isSigned = Opcode == Instruction::SDiv;
1026 // (X * Y) / Y -> X if the multiplication does not overflow.
1027 Value *X = nullptr, *Y = nullptr;
1028 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1029 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1030 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1031 // If the Mul knows it does not overflow, then we are good to go.
1032 if ((isSigned && Mul->hasNoSignedWrap()) ||
1033 (!isSigned && Mul->hasNoUnsignedWrap()))
1035 // If X has the form X = A / Y then X * Y cannot overflow.
1036 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1037 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1041 // (X rem Y) / Y -> 0
1042 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1043 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1044 return Constant::getNullValue(Op0->getType());
1046 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1047 ConstantInt *C1, *C2;
1048 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1049 match(Op1, m_ConstantInt(C2))) {
1051 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1053 return Constant::getNullValue(Op0->getType());
1056 // If the operation is with the result of a select instruction, check whether
1057 // operating on either branch of the select always yields the same value.
1058 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1059 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1062 // If the operation is with the result of a phi instruction, check whether
1063 // operating on all incoming values of the phi always yields the same value.
1064 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1065 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1071 /// Given operands for an SDiv, see if we can fold the result.
1072 /// If not, this returns null.
1073 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1074 unsigned MaxRecurse) {
1075 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1081 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1082 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1085 /// Given operands for a UDiv, see if we can fold the result.
1086 /// If not, this returns null.
1087 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1088 unsigned MaxRecurse) {
1089 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1092 // udiv %V, C -> 0 if %V < C
1094 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1095 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1096 if (C->isAllOnesValue()) {
1097 return Constant::getNullValue(Op0->getType());
1105 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1106 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1109 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1110 const SimplifyQuery &Q, unsigned) {
1111 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
1114 // undef / X -> undef (the undef could be a snan).
1115 if (match(Op0, m_Undef()))
1118 // X / undef -> undef
1119 if (match(Op1, m_Undef()))
1123 if (match(Op1, m_FPOne()))
1127 // Requires that NaNs are off (X could be zero) and signed zeroes are
1128 // ignored (X could be positive or negative, so the output sign is unknown).
1129 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1133 // X / X -> 1.0 is legal when NaNs are ignored.
1135 return ConstantFP::get(Op0->getType(), 1.0);
1137 // -X / X -> -1.0 and
1138 // X / -X -> -1.0 are legal when NaNs are ignored.
1139 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1140 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1141 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1142 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1143 BinaryOperator::getFNegArgument(Op1) == Op0))
1144 return ConstantFP::get(Op0->getType(), -1.0);
1150 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1151 const SimplifyQuery &Q) {
1152 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
1155 /// Given operands for an SRem or URem, see if we can fold the result.
1156 /// If not, this returns null.
1157 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1158 const SimplifyQuery &Q, unsigned MaxRecurse) {
1159 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1162 if (Value *V = simplifyDivRem(Op0, Op1, false))
1165 // (X % Y) % Y -> X % Y
1166 if ((Opcode == Instruction::SRem &&
1167 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1168 (Opcode == Instruction::URem &&
1169 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1172 // If the operation is with the result of a select instruction, check whether
1173 // operating on either branch of the select always yields the same value.
1174 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1175 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1178 // If the operation is with the result of a phi instruction, check whether
1179 // operating on all incoming values of the phi always yields the same value.
1180 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1181 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1187 /// Given operands for an SRem, see if we can fold the result.
1188 /// If not, this returns null.
1189 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1190 unsigned MaxRecurse) {
1191 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1197 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1198 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1201 /// Given operands for a URem, see if we can fold the result.
1202 /// If not, this returns null.
1203 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1204 unsigned MaxRecurse) {
1205 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1208 // urem %V, C -> %V if %V < C
1210 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1211 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1212 if (C->isAllOnesValue()) {
1221 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1222 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1225 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1226 const SimplifyQuery &Q, unsigned) {
1227 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
1230 // undef % X -> undef (the undef could be a snan).
1231 if (match(Op0, m_Undef()))
1234 // X % undef -> undef
1235 if (match(Op1, m_Undef()))
1239 // Requires that NaNs are off (X could be zero) and signed zeroes are
1240 // ignored (X could be positive or negative, so the output sign is unknown).
1241 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1247 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1248 const SimplifyQuery &Q) {
1249 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
1252 /// Returns true if a shift by \c Amount always yields undef.
1253 static bool isUndefShift(Value *Amount) {
1254 Constant *C = dyn_cast<Constant>(Amount);
1258 // X shift by undef -> undef because it may shift by the bitwidth.
1259 if (isa<UndefValue>(C))
1262 // Shifting by the bitwidth or more is undefined.
1263 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1264 if (CI->getValue().getLimitedValue() >=
1265 CI->getType()->getScalarSizeInBits())
1268 // If all lanes of a vector shift are undefined the whole shift is.
1269 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1270 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1271 if (!isUndefShift(C->getAggregateElement(I)))
1279 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1280 /// If not, this returns null.
1281 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1282 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1283 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1286 // 0 shift by X -> 0
1287 if (match(Op0, m_Zero()))
1290 // X shift by 0 -> X
1291 if (match(Op1, m_Zero()))
1294 // Fold undefined shifts.
1295 if (isUndefShift(Op1))
1296 return UndefValue::get(Op0->getType());
1298 // If the operation is with the result of a select instruction, check whether
1299 // operating on either branch of the select always yields the same value.
1300 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1301 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1304 // If the operation is with the result of a phi instruction, check whether
1305 // operating on all incoming values of the phi always yields the same value.
1306 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1307 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1310 // If any bits in the shift amount make that value greater than or equal to
1311 // the number of bits in the type, the shift is undefined.
1312 unsigned BitWidth = Op1->getType()->getScalarSizeInBits();
1313 KnownBits Known(BitWidth);
1314 computeKnownBits(Op1, Known, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1315 if (Known.One.getLimitedValue() >= BitWidth)
1316 return UndefValue::get(Op0->getType());
1318 // If all valid bits in the shift amount are known zero, the first operand is
1320 unsigned NumValidShiftBits = Log2_32_Ceil(BitWidth);
1321 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1327 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1328 /// fold the result. If not, this returns null.
1329 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1330 Value *Op1, bool isExact, const SimplifyQuery &Q,
1331 unsigned MaxRecurse) {
1332 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1337 return Constant::getNullValue(Op0->getType());
1340 // undef >> X -> undef (if it's exact)
1341 if (match(Op0, m_Undef()))
1342 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1344 // The low bit cannot be shifted out of an exact shift if it is set.
1346 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
1347 KnownBits Op0Known(BitWidth);
1348 computeKnownBits(Op0, Op0Known, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1349 if (Op0Known.One[0])
1356 /// Given operands for an Shl, see if we can fold the result.
1357 /// If not, this returns null.
1358 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1359 const SimplifyQuery &Q, unsigned MaxRecurse) {
1360 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1364 // undef << X -> undef if (if it's NSW/NUW)
1365 if (match(Op0, m_Undef()))
1366 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1368 // (X >> A) << A -> X
1370 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1375 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1376 const SimplifyQuery &Q) {
1377 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1380 /// Given operands for an LShr, see if we can fold the result.
1381 /// If not, this returns null.
1382 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1383 const SimplifyQuery &Q, unsigned MaxRecurse) {
1384 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1388 // (X << A) >> A -> X
1390 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1396 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1397 const SimplifyQuery &Q) {
1398 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1401 /// Given operands for an AShr, see if we can fold the result.
1402 /// If not, this returns null.
1403 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1404 const SimplifyQuery &Q, unsigned MaxRecurse) {
1405 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1409 // all ones >>a X -> all ones
1410 if (match(Op0, m_AllOnes()))
1413 // (X << A) >> A -> X
1415 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1418 // Arithmetic shifting an all-sign-bit value is a no-op.
1419 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1420 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1426 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1427 const SimplifyQuery &Q) {
1428 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1431 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1432 ICmpInst *UnsignedICmp, bool IsAnd) {
1435 ICmpInst::Predicate EqPred;
1436 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1437 !ICmpInst::isEquality(EqPred))
1440 ICmpInst::Predicate UnsignedPred;
1441 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1442 ICmpInst::isUnsigned(UnsignedPred))
1444 else if (match(UnsignedICmp,
1445 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1446 ICmpInst::isUnsigned(UnsignedPred))
1447 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1451 // X < Y && Y != 0 --> X < Y
1452 // X < Y || Y != 0 --> Y != 0
1453 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1454 return IsAnd ? UnsignedICmp : ZeroICmp;
1456 // X >= Y || Y != 0 --> true
1457 // X >= Y || Y == 0 --> X >= Y
1458 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1459 if (EqPred == ICmpInst::ICMP_NE)
1460 return getTrue(UnsignedICmp->getType());
1461 return UnsignedICmp;
1464 // X < Y && Y == 0 --> false
1465 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1467 return getFalse(UnsignedICmp->getType());
1472 /// Commuted variants are assumed to be handled by calling this function again
1473 /// with the parameters swapped.
1474 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1475 ICmpInst::Predicate Pred0, Pred1;
1477 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1478 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1481 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1482 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1483 // can eliminate Op1 from this 'and'.
1484 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1487 // Check for any combination of predicates that are guaranteed to be disjoint.
1488 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1489 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1490 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1491 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1492 return getFalse(Op0->getType());
1497 /// Commuted variants are assumed to be handled by calling this function again
1498 /// with the parameters swapped.
1499 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1500 ICmpInst::Predicate Pred0, Pred1;
1502 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1503 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1506 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1507 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1508 // can eliminate Op0 from this 'or'.
1509 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1512 // Check for any combination of predicates that cover the entire range of
1514 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1515 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1516 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1517 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1518 return getTrue(Op0->getType());
1523 /// Test if a pair of compares with a shared operand and 2 constants has an
1524 /// empty set intersection, full set union, or if one compare is a superset of
1526 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1528 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1529 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1532 const APInt *C0, *C1;
1533 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1534 !match(Cmp1->getOperand(1), m_APInt(C1)))
1537 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1538 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1540 // For and-of-compares, check if the intersection is empty:
1541 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1542 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1543 return getFalse(Cmp0->getType());
1545 // For or-of-compares, check if the union is full:
1546 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1547 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1548 return getTrue(Cmp0->getType());
1550 // Is one range a superset of the other?
1551 // If this is and-of-compares, take the smaller set:
1552 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1553 // If this is or-of-compares, take the larger set:
1554 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1555 if (Range0.contains(Range1))
1556 return IsAnd ? Cmp1 : Cmp0;
1557 if (Range1.contains(Range0))
1558 return IsAnd ? Cmp0 : Cmp1;
1563 /// Commuted variants are assumed to be handled by calling this function again
1564 /// with the parameters swapped.
1565 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1566 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1569 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1572 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1575 // (icmp (add V, C0), C1) & (icmp V, C0)
1576 Type *ITy = Op0->getType();
1577 ICmpInst::Predicate Pred0, Pred1;
1578 const APInt *C0, *C1;
1580 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1583 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1586 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1587 if (AddInst->getOperand(1) != Op1->getOperand(1))
1590 bool isNSW = AddInst->hasNoSignedWrap();
1591 bool isNUW = AddInst->hasNoUnsignedWrap();
1593 const APInt Delta = *C1 - *C0;
1594 if (C0->isStrictlyPositive()) {
1596 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1597 return getFalse(ITy);
1598 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1599 return getFalse(ITy);
1602 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1603 return getFalse(ITy);
1604 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1605 return getFalse(ITy);
1608 if (C0->getBoolValue() && isNUW) {
1610 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1611 return getFalse(ITy);
1613 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1614 return getFalse(ITy);
1620 /// Commuted variants are assumed to be handled by calling this function again
1621 /// with the parameters swapped.
1622 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1623 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1626 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1629 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1632 // (icmp (add V, C0), C1) | (icmp V, C0)
1633 ICmpInst::Predicate Pred0, Pred1;
1634 const APInt *C0, *C1;
1636 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1639 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1642 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1643 if (AddInst->getOperand(1) != Op1->getOperand(1))
1646 Type *ITy = Op0->getType();
1647 bool isNSW = AddInst->hasNoSignedWrap();
1648 bool isNUW = AddInst->hasNoUnsignedWrap();
1650 const APInt Delta = *C1 - *C0;
1651 if (C0->isStrictlyPositive()) {
1653 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1654 return getTrue(ITy);
1655 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1656 return getTrue(ITy);
1659 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1660 return getTrue(ITy);
1661 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1662 return getTrue(ITy);
1665 if (C0->getBoolValue() && isNUW) {
1667 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1668 return getTrue(ITy);
1670 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1671 return getTrue(ITy);
1677 static Value *simplifyPossiblyCastedAndOrOfICmps(ICmpInst *Cmp0, ICmpInst *Cmp1,
1678 bool IsAnd, CastInst *Cast) {
1680 IsAnd ? simplifyAndOfICmps(Cmp0, Cmp1) : simplifyOrOfICmps(Cmp0, Cmp1);
1686 // If we looked through casts, we can only handle a constant simplification
1687 // because we are not allowed to create a cast instruction here.
1688 if (auto *C = dyn_cast<Constant>(V))
1689 return ConstantExpr::getCast(Cast->getOpcode(), C, Cast->getType());
1694 static Value *simplifyAndOrOfICmps(Value *Op0, Value *Op1, bool IsAnd) {
1695 // Look through casts of the 'and' operands to find compares.
1696 auto *Cast0 = dyn_cast<CastInst>(Op0);
1697 auto *Cast1 = dyn_cast<CastInst>(Op1);
1698 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1699 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1700 Op0 = Cast0->getOperand(0);
1701 Op1 = Cast1->getOperand(0);
1704 auto *Cmp0 = dyn_cast<ICmpInst>(Op0);
1705 auto *Cmp1 = dyn_cast<ICmpInst>(Op1);
1709 if (Value *V = simplifyPossiblyCastedAndOrOfICmps(Cmp0, Cmp1, IsAnd, Cast0))
1711 if (Value *V = simplifyPossiblyCastedAndOrOfICmps(Cmp1, Cmp0, IsAnd, Cast0))
1717 /// Given operands for an And, see if we can fold the result.
1718 /// If not, this returns null.
1719 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1720 unsigned MaxRecurse) {
1721 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1725 if (match(Op1, m_Undef()))
1726 return Constant::getNullValue(Op0->getType());
1733 if (match(Op1, m_Zero()))
1737 if (match(Op1, m_AllOnes()))
1740 // A & ~A = ~A & A = 0
1741 if (match(Op0, m_Not(m_Specific(Op1))) ||
1742 match(Op1, m_Not(m_Specific(Op0))))
1743 return Constant::getNullValue(Op0->getType());
1746 Value *A = nullptr, *B = nullptr;
1747 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1748 (A == Op1 || B == Op1))
1752 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1753 (A == Op0 || B == Op0))
1756 // A mask that only clears known zeros of a shifted value is a no-op.
1760 if (match(Op1, m_APInt(Mask))) {
1761 // If all bits in the inverted and shifted mask are clear:
1762 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1763 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1764 (~(*Mask)).lshr(*ShAmt).isNullValue())
1767 // If all bits in the inverted and shifted mask are clear:
1768 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1769 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1770 (~(*Mask)).shl(*ShAmt).isNullValue())
1774 // A & (-A) = A if A is a power of two or zero.
1775 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1776 match(Op1, m_Neg(m_Specific(Op0)))) {
1777 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1780 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1785 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, true))
1788 // Try some generic simplifications for associative operations.
1789 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1793 // And distributes over Or. Try some generic simplifications based on this.
1794 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1798 // And distributes over Xor. Try some generic simplifications based on this.
1799 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1803 // If the operation is with the result of a select instruction, check whether
1804 // operating on either branch of the select always yields the same value.
1805 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1806 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1810 // If the operation is with the result of a phi instruction, check whether
1811 // operating on all incoming values of the phi always yields the same value.
1812 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1813 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1820 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1821 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1824 /// Given operands for an Or, see if we can fold the result.
1825 /// If not, this returns null.
1826 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1827 unsigned MaxRecurse) {
1828 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1832 if (match(Op1, m_Undef()))
1833 return Constant::getAllOnesValue(Op0->getType());
1840 if (match(Op1, m_Zero()))
1844 if (match(Op1, m_AllOnes()))
1847 // A | ~A = ~A | A = -1
1848 if (match(Op0, m_Not(m_Specific(Op1))) ||
1849 match(Op1, m_Not(m_Specific(Op0))))
1850 return Constant::getAllOnesValue(Op0->getType());
1853 Value *A = nullptr, *B = nullptr;
1854 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1855 (A == Op1 || B == Op1))
1859 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1860 (A == Op0 || B == Op0))
1863 // ~(A & ?) | A = -1
1864 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1865 (A == Op1 || B == Op1))
1866 return Constant::getAllOnesValue(Op1->getType());
1868 // A | ~(A & ?) = -1
1869 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1870 (A == Op0 || B == Op0))
1871 return Constant::getAllOnesValue(Op0->getType());
1873 // (A & ~B) | (A ^ B) -> (A ^ B)
1874 // (~B & A) | (A ^ B) -> (A ^ B)
1875 // (A & ~B) | (B ^ A) -> (B ^ A)
1876 // (~B & A) | (B ^ A) -> (B ^ A)
1877 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1878 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1879 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1882 // Commute the 'or' operands.
1883 // (A ^ B) | (A & ~B) -> (A ^ B)
1884 // (A ^ B) | (~B & A) -> (A ^ B)
1885 // (B ^ A) | (A & ~B) -> (B ^ A)
1886 // (B ^ A) | (~B & A) -> (B ^ A)
1887 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1888 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1889 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1892 // (A & B) | (~A ^ B) -> (~A ^ B)
1893 // (B & A) | (~A ^ B) -> (~A ^ B)
1894 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1895 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1896 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1897 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1898 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1901 // (~A ^ B) | (A & B) -> (~A ^ B)
1902 // (~A ^ B) | (B & A) -> (~A ^ B)
1903 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1904 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1905 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1906 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1907 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1910 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, false))
1913 // Try some generic simplifications for associative operations.
1914 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1918 // Or distributes over And. Try some generic simplifications based on this.
1919 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1923 // If the operation is with the result of a select instruction, check whether
1924 // operating on either branch of the select always yields the same value.
1925 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1926 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1931 Value *C = nullptr, *D = nullptr;
1932 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1933 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1934 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1935 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1936 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1937 // (A & C1)|(B & C2)
1938 // If we have: ((V + N) & C1) | (V & C2)
1939 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1940 // replace with V+N.
1942 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1943 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1944 // Add commutes, try both ways.
1946 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1949 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1952 // Or commutes, try both ways.
1953 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1954 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1955 // Add commutes, try both ways.
1957 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1960 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1966 // If the operation is with the result of a phi instruction, check whether
1967 // operating on all incoming values of the phi always yields the same value.
1968 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1969 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1975 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1976 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1979 /// Given operands for a Xor, see if we can fold the result.
1980 /// If not, this returns null.
1981 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1982 unsigned MaxRecurse) {
1983 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1986 // A ^ undef -> undef
1987 if (match(Op1, m_Undef()))
1991 if (match(Op1, m_Zero()))
1996 return Constant::getNullValue(Op0->getType());
1998 // A ^ ~A = ~A ^ A = -1
1999 if (match(Op0, m_Not(m_Specific(Op1))) ||
2000 match(Op1, m_Not(m_Specific(Op0))))
2001 return Constant::getAllOnesValue(Op0->getType());
2003 // Try some generic simplifications for associative operations.
2004 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2008 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2009 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2010 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2011 // only if B and C are equal. If B and C are equal then (since we assume
2012 // that operands have already been simplified) "select(cond, B, C)" should
2013 // have been simplified to the common value of B and C already. Analysing
2014 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2015 // for threading over phi nodes.
2020 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2021 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2025 static Type *GetCompareTy(Value *Op) {
2026 return CmpInst::makeCmpResultType(Op->getType());
2029 /// Rummage around inside V looking for something equivalent to the comparison
2030 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2031 /// Helper function for analyzing max/min idioms.
2032 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2033 Value *LHS, Value *RHS) {
2034 SelectInst *SI = dyn_cast<SelectInst>(V);
2037 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2040 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2041 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2043 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2044 LHS == CmpRHS && RHS == CmpLHS)
2049 // A significant optimization not implemented here is assuming that alloca
2050 // addresses are not equal to incoming argument values. They don't *alias*,
2051 // as we say, but that doesn't mean they aren't equal, so we take a
2052 // conservative approach.
2054 // This is inspired in part by C++11 5.10p1:
2055 // "Two pointers of the same type compare equal if and only if they are both
2056 // null, both point to the same function, or both represent the same
2059 // This is pretty permissive.
2061 // It's also partly due to C11 6.5.9p6:
2062 // "Two pointers compare equal if and only if both are null pointers, both are
2063 // pointers to the same object (including a pointer to an object and a
2064 // subobject at its beginning) or function, both are pointers to one past the
2065 // last element of the same array object, or one is a pointer to one past the
2066 // end of one array object and the other is a pointer to the start of a
2067 // different array object that happens to immediately follow the first array
2068 // object in the address space.)
2070 // C11's version is more restrictive, however there's no reason why an argument
2071 // couldn't be a one-past-the-end value for a stack object in the caller and be
2072 // equal to the beginning of a stack object in the callee.
2074 // If the C and C++ standards are ever made sufficiently restrictive in this
2075 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2076 // this optimization.
2078 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2079 const DominatorTree *DT, CmpInst::Predicate Pred,
2080 const Instruction *CxtI, Value *LHS, Value *RHS) {
2081 // First, skip past any trivial no-ops.
2082 LHS = LHS->stripPointerCasts();
2083 RHS = RHS->stripPointerCasts();
2085 // A non-null pointer is not equal to a null pointer.
2086 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) &&
2087 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2088 return ConstantInt::get(GetCompareTy(LHS),
2089 !CmpInst::isTrueWhenEqual(Pred));
2091 // We can only fold certain predicates on pointer comparisons.
2096 // Equality comaprisons are easy to fold.
2097 case CmpInst::ICMP_EQ:
2098 case CmpInst::ICMP_NE:
2101 // We can only handle unsigned relational comparisons because 'inbounds' on
2102 // a GEP only protects against unsigned wrapping.
2103 case CmpInst::ICMP_UGT:
2104 case CmpInst::ICMP_UGE:
2105 case CmpInst::ICMP_ULT:
2106 case CmpInst::ICMP_ULE:
2107 // However, we have to switch them to their signed variants to handle
2108 // negative indices from the base pointer.
2109 Pred = ICmpInst::getSignedPredicate(Pred);
2113 // Strip off any constant offsets so that we can reason about them.
2114 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2115 // here and compare base addresses like AliasAnalysis does, however there are
2116 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2117 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2118 // doesn't need to guarantee pointer inequality when it says NoAlias.
2119 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2120 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2122 // If LHS and RHS are related via constant offsets to the same base
2123 // value, we can replace it with an icmp which just compares the offsets.
2125 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2127 // Various optimizations for (in)equality comparisons.
2128 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2129 // Different non-empty allocations that exist at the same time have
2130 // different addresses (if the program can tell). Global variables always
2131 // exist, so they always exist during the lifetime of each other and all
2132 // allocas. Two different allocas usually have different addresses...
2134 // However, if there's an @llvm.stackrestore dynamically in between two
2135 // allocas, they may have the same address. It's tempting to reduce the
2136 // scope of the problem by only looking at *static* allocas here. That would
2137 // cover the majority of allocas while significantly reducing the likelihood
2138 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2139 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2140 // an entry block. Also, if we have a block that's not attached to a
2141 // function, we can't tell if it's "static" under the current definition.
2142 // Theoretically, this problem could be fixed by creating a new kind of
2143 // instruction kind specifically for static allocas. Such a new instruction
2144 // could be required to be at the top of the entry block, thus preventing it
2145 // from being subject to a @llvm.stackrestore. Instcombine could even
2146 // convert regular allocas into these special allocas. It'd be nifty.
2147 // However, until then, this problem remains open.
2149 // So, we'll assume that two non-empty allocas have different addresses
2152 // With all that, if the offsets are within the bounds of their allocations
2153 // (and not one-past-the-end! so we can't use inbounds!), and their
2154 // allocations aren't the same, the pointers are not equal.
2156 // Note that it's not necessary to check for LHS being a global variable
2157 // address, due to canonicalization and constant folding.
2158 if (isa<AllocaInst>(LHS) &&
2159 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2160 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2161 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2162 uint64_t LHSSize, RHSSize;
2163 if (LHSOffsetCI && RHSOffsetCI &&
2164 getObjectSize(LHS, LHSSize, DL, TLI) &&
2165 getObjectSize(RHS, RHSSize, DL, TLI)) {
2166 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2167 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2168 if (!LHSOffsetValue.isNegative() &&
2169 !RHSOffsetValue.isNegative() &&
2170 LHSOffsetValue.ult(LHSSize) &&
2171 RHSOffsetValue.ult(RHSSize)) {
2172 return ConstantInt::get(GetCompareTy(LHS),
2173 !CmpInst::isTrueWhenEqual(Pred));
2177 // Repeat the above check but this time without depending on DataLayout
2178 // or being able to compute a precise size.
2179 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2180 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2181 LHSOffset->isNullValue() &&
2182 RHSOffset->isNullValue())
2183 return ConstantInt::get(GetCompareTy(LHS),
2184 !CmpInst::isTrueWhenEqual(Pred));
2187 // Even if an non-inbounds GEP occurs along the path we can still optimize
2188 // equality comparisons concerning the result. We avoid walking the whole
2189 // chain again by starting where the last calls to
2190 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2191 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2192 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2194 return ConstantExpr::getICmp(Pred,
2195 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2196 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2198 // If one side of the equality comparison must come from a noalias call
2199 // (meaning a system memory allocation function), and the other side must
2200 // come from a pointer that cannot overlap with dynamically-allocated
2201 // memory within the lifetime of the current function (allocas, byval
2202 // arguments, globals), then determine the comparison result here.
2203 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2204 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2205 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2207 // Is the set of underlying objects all noalias calls?
2208 auto IsNAC = [](ArrayRef<Value *> Objects) {
2209 return all_of(Objects, isNoAliasCall);
2212 // Is the set of underlying objects all things which must be disjoint from
2213 // noalias calls. For allocas, we consider only static ones (dynamic
2214 // allocas might be transformed into calls to malloc not simultaneously
2215 // live with the compared-to allocation). For globals, we exclude symbols
2216 // that might be resolve lazily to symbols in another dynamically-loaded
2217 // library (and, thus, could be malloc'ed by the implementation).
2218 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2219 return all_of(Objects, [](Value *V) {
2220 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2221 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2222 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2223 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2224 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2225 !GV->isThreadLocal();
2226 if (const Argument *A = dyn_cast<Argument>(V))
2227 return A->hasByValAttr();
2232 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2233 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2234 return ConstantInt::get(GetCompareTy(LHS),
2235 !CmpInst::isTrueWhenEqual(Pred));
2237 // Fold comparisons for non-escaping pointer even if the allocation call
2238 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2239 // dynamic allocation call could be either of the operands.
2240 Value *MI = nullptr;
2241 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT))
2243 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT))
2245 // FIXME: We should also fold the compare when the pointer escapes, but the
2246 // compare dominates the pointer escape
2247 if (MI && !PointerMayBeCaptured(MI, true, true))
2248 return ConstantInt::get(GetCompareTy(LHS),
2249 CmpInst::isFalseWhenEqual(Pred));
2256 /// Fold an icmp when its operands have i1 scalar type.
2257 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2258 Value *RHS, const SimplifyQuery &Q) {
2259 Type *ITy = GetCompareTy(LHS); // The return type.
2260 Type *OpTy = LHS->getType(); // The operand type.
2261 if (!OpTy->getScalarType()->isIntegerTy(1))
2264 // A boolean compared to true/false can be simplified in 14 out of the 20
2265 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2266 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2267 if (match(RHS, m_Zero())) {
2269 case CmpInst::ICMP_NE: // X != 0 -> X
2270 case CmpInst::ICMP_UGT: // X >u 0 -> X
2271 case CmpInst::ICMP_SLT: // X <s 0 -> X
2274 case CmpInst::ICMP_ULT: // X <u 0 -> false
2275 case CmpInst::ICMP_SGT: // X >s 0 -> false
2276 return getFalse(ITy);
2278 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2279 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2280 return getTrue(ITy);
2284 } else if (match(RHS, m_One())) {
2286 case CmpInst::ICMP_EQ: // X == 1 -> X
2287 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2288 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2291 case CmpInst::ICMP_UGT: // X >u 1 -> false
2292 case CmpInst::ICMP_SLT: // X <s -1 -> false
2293 return getFalse(ITy);
2295 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2296 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2297 return getTrue(ITy);
2306 case ICmpInst::ICMP_UGE:
2307 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2308 return getTrue(ITy);
2310 case ICmpInst::ICMP_SGE:
2311 /// For signed comparison, the values for an i1 are 0 and -1
2312 /// respectively. This maps into a truth table of:
2313 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2314 /// 0 | 0 | 1 (0 >= 0) | 1
2315 /// 0 | 1 | 1 (0 >= -1) | 1
2316 /// 1 | 0 | 0 (-1 >= 0) | 0
2317 /// 1 | 1 | 1 (-1 >= -1) | 1
2318 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2319 return getTrue(ITy);
2321 case ICmpInst::ICMP_ULE:
2322 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2323 return getTrue(ITy);
2330 /// Try hard to fold icmp with zero RHS because this is a common case.
2331 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2332 Value *RHS, const SimplifyQuery &Q) {
2333 if (!match(RHS, m_Zero()))
2336 Type *ITy = GetCompareTy(LHS); // The return type.
2339 llvm_unreachable("Unknown ICmp predicate!");
2340 case ICmpInst::ICMP_ULT:
2341 return getFalse(ITy);
2342 case ICmpInst::ICMP_UGE:
2343 return getTrue(ITy);
2344 case ICmpInst::ICMP_EQ:
2345 case ICmpInst::ICMP_ULE:
2346 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2347 return getFalse(ITy);
2349 case ICmpInst::ICMP_NE:
2350 case ICmpInst::ICMP_UGT:
2351 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2352 return getTrue(ITy);
2354 case ICmpInst::ICMP_SLT: {
2355 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2356 if (LHSKnown.isNegative())
2357 return getTrue(ITy);
2358 if (LHSKnown.isNonNegative())
2359 return getFalse(ITy);
2362 case ICmpInst::ICMP_SLE: {
2363 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2364 if (LHSKnown.isNegative())
2365 return getTrue(ITy);
2366 if (LHSKnown.isNonNegative() &&
2367 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2368 return getFalse(ITy);
2371 case ICmpInst::ICMP_SGE: {
2372 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2373 if (LHSKnown.isNegative())
2374 return getFalse(ITy);
2375 if (LHSKnown.isNonNegative())
2376 return getTrue(ITy);
2379 case ICmpInst::ICMP_SGT: {
2380 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2381 if (LHSKnown.isNegative())
2382 return getFalse(ITy);
2383 if (LHSKnown.isNonNegative() &&
2384 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2385 return getTrue(ITy);
2393 /// Many binary operators with a constant operand have an easy-to-compute
2394 /// range of outputs. This can be used to fold a comparison to always true or
2396 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2397 unsigned Width = Lower.getBitWidth();
2399 switch (BO.getOpcode()) {
2400 case Instruction::Add:
2401 if (match(BO.getOperand(1), m_APInt(C)) && *C != 0) {
2402 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2403 if (BO.hasNoUnsignedWrap()) {
2404 // 'add nuw x, C' produces [C, UINT_MAX].
2406 } else if (BO.hasNoSignedWrap()) {
2407 if (C->isNegative()) {
2408 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2409 Lower = APInt::getSignedMinValue(Width);
2410 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2412 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2413 Lower = APInt::getSignedMinValue(Width) + *C;
2414 Upper = APInt::getSignedMaxValue(Width) + 1;
2420 case Instruction::And:
2421 if (match(BO.getOperand(1), m_APInt(C)))
2422 // 'and x, C' produces [0, C].
2426 case Instruction::Or:
2427 if (match(BO.getOperand(1), m_APInt(C)))
2428 // 'or x, C' produces [C, UINT_MAX].
2432 case Instruction::AShr:
2433 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2434 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2435 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2436 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2437 } else if (match(BO.getOperand(0), m_APInt(C))) {
2438 unsigned ShiftAmount = Width - 1;
2439 if (*C != 0 && BO.isExact())
2440 ShiftAmount = C->countTrailingZeros();
2441 if (C->isNegative()) {
2442 // 'ashr C, x' produces [C, C >> (Width-1)]
2444 Upper = C->ashr(ShiftAmount) + 1;
2446 // 'ashr C, x' produces [C >> (Width-1), C]
2447 Lower = C->ashr(ShiftAmount);
2453 case Instruction::LShr:
2454 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2455 // 'lshr x, C' produces [0, UINT_MAX >> C].
2456 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2457 } else if (match(BO.getOperand(0), m_APInt(C))) {
2458 // 'lshr C, x' produces [C >> (Width-1), C].
2459 unsigned ShiftAmount = Width - 1;
2460 if (*C != 0 && BO.isExact())
2461 ShiftAmount = C->countTrailingZeros();
2462 Lower = C->lshr(ShiftAmount);
2467 case Instruction::Shl:
2468 if (match(BO.getOperand(0), m_APInt(C))) {
2469 if (BO.hasNoUnsignedWrap()) {
2470 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2472 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2473 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2474 if (C->isNegative()) {
2475 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2476 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2477 Lower = C->shl(ShiftAmount);
2480 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2481 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2483 Upper = C->shl(ShiftAmount) + 1;
2489 case Instruction::SDiv:
2490 if (match(BO.getOperand(1), m_APInt(C))) {
2491 APInt IntMin = APInt::getSignedMinValue(Width);
2492 APInt IntMax = APInt::getSignedMaxValue(Width);
2493 if (C->isAllOnesValue()) {
2494 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2495 // where C != -1 and C != 0 and C != 1
2498 } else if (C->countLeadingZeros() < Width - 1) {
2499 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2500 // where C != -1 and C != 0 and C != 1
2501 Lower = IntMin.sdiv(*C);
2502 Upper = IntMax.sdiv(*C);
2503 if (Lower.sgt(Upper))
2504 std::swap(Lower, Upper);
2506 assert(Upper != Lower && "Upper part of range has wrapped!");
2508 } else if (match(BO.getOperand(0), m_APInt(C))) {
2509 if (C->isMinSignedValue()) {
2510 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2512 Upper = Lower.lshr(1) + 1;
2514 // 'sdiv C, x' produces [-|C|, |C|].
2515 Upper = C->abs() + 1;
2516 Lower = (-Upper) + 1;
2521 case Instruction::UDiv:
2522 if (match(BO.getOperand(1), m_APInt(C)) && *C != 0) {
2523 // 'udiv x, C' produces [0, UINT_MAX / C].
2524 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2525 } else if (match(BO.getOperand(0), m_APInt(C))) {
2526 // 'udiv C, x' produces [0, C].
2531 case Instruction::SRem:
2532 if (match(BO.getOperand(1), m_APInt(C))) {
2533 // 'srem x, C' produces (-|C|, |C|).
2535 Lower = (-Upper) + 1;
2539 case Instruction::URem:
2540 if (match(BO.getOperand(1), m_APInt(C)))
2541 // 'urem x, C' produces [0, C).
2550 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2553 if (!match(RHS, m_APInt(C)))
2556 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2557 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2558 if (RHS_CR.isEmptySet())
2559 return ConstantInt::getFalse(GetCompareTy(RHS));
2560 if (RHS_CR.isFullSet())
2561 return ConstantInt::getTrue(GetCompareTy(RHS));
2563 // Find the range of possible values for binary operators.
2564 unsigned Width = C->getBitWidth();
2565 APInt Lower = APInt(Width, 0);
2566 APInt Upper = APInt(Width, 0);
2567 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2568 setLimitsForBinOp(*BO, Lower, Upper);
2570 ConstantRange LHS_CR =
2571 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2573 if (auto *I = dyn_cast<Instruction>(LHS))
2574 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2575 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2577 if (!LHS_CR.isFullSet()) {
2578 if (RHS_CR.contains(LHS_CR))
2579 return ConstantInt::getTrue(GetCompareTy(RHS));
2580 if (RHS_CR.inverse().contains(LHS_CR))
2581 return ConstantInt::getFalse(GetCompareTy(RHS));
2587 /// TODO: A large part of this logic is duplicated in InstCombine's
2588 /// foldICmpBinOp(). We should be able to share that and avoid the code
2590 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2591 Value *RHS, const SimplifyQuery &Q,
2592 unsigned MaxRecurse) {
2593 Type *ITy = GetCompareTy(LHS); // The return type.
2595 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2596 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2597 if (MaxRecurse && (LBO || RBO)) {
2598 // Analyze the case when either LHS or RHS is an add instruction.
2599 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2600 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2601 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2602 if (LBO && LBO->getOpcode() == Instruction::Add) {
2603 A = LBO->getOperand(0);
2604 B = LBO->getOperand(1);
2606 ICmpInst::isEquality(Pred) ||
2607 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2608 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2610 if (RBO && RBO->getOpcode() == Instruction::Add) {
2611 C = RBO->getOperand(0);
2612 D = RBO->getOperand(1);
2614 ICmpInst::isEquality(Pred) ||
2615 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2616 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2619 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2620 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2621 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2622 Constant::getNullValue(RHS->getType()), Q,
2626 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2627 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2629 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2630 C == LHS ? D : C, Q, MaxRecurse - 1))
2633 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2634 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2636 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2639 // C + B == C + D -> B == D
2642 } else if (A == D) {
2643 // D + B == C + D -> B == C
2646 } else if (B == C) {
2647 // A + C == C + D -> A == D
2652 // A + D == C + D -> A == C
2656 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2663 // icmp pred (or X, Y), X
2664 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2665 if (Pred == ICmpInst::ICMP_ULT)
2666 return getFalse(ITy);
2667 if (Pred == ICmpInst::ICMP_UGE)
2668 return getTrue(ITy);
2670 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2671 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2672 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2673 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2674 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2675 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2676 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2679 // icmp pred X, (or X, Y)
2680 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2681 if (Pred == ICmpInst::ICMP_ULE)
2682 return getTrue(ITy);
2683 if (Pred == ICmpInst::ICMP_UGT)
2684 return getFalse(ITy);
2686 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2687 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2688 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2689 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2690 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2691 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2692 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2697 // icmp pred (and X, Y), X
2698 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)),
2699 m_And(m_Specific(RHS), m_Value())))) {
2700 if (Pred == ICmpInst::ICMP_UGT)
2701 return getFalse(ITy);
2702 if (Pred == ICmpInst::ICMP_ULE)
2703 return getTrue(ITy);
2705 // icmp pred X, (and X, Y)
2706 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)),
2707 m_And(m_Specific(LHS), m_Value())))) {
2708 if (Pred == ICmpInst::ICMP_UGE)
2709 return getTrue(ITy);
2710 if (Pred == ICmpInst::ICMP_ULT)
2711 return getFalse(ITy);
2714 // 0 - (zext X) pred C
2715 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2716 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2717 if (RHSC->getValue().isStrictlyPositive()) {
2718 if (Pred == ICmpInst::ICMP_SLT)
2719 return ConstantInt::getTrue(RHSC->getContext());
2720 if (Pred == ICmpInst::ICMP_SGE)
2721 return ConstantInt::getFalse(RHSC->getContext());
2722 if (Pred == ICmpInst::ICMP_EQ)
2723 return ConstantInt::getFalse(RHSC->getContext());
2724 if (Pred == ICmpInst::ICMP_NE)
2725 return ConstantInt::getTrue(RHSC->getContext());
2727 if (RHSC->getValue().isNonNegative()) {
2728 if (Pred == ICmpInst::ICMP_SLE)
2729 return ConstantInt::getTrue(RHSC->getContext());
2730 if (Pred == ICmpInst::ICMP_SGT)
2731 return ConstantInt::getFalse(RHSC->getContext());
2736 // icmp pred (urem X, Y), Y
2737 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2741 case ICmpInst::ICMP_SGT:
2742 case ICmpInst::ICMP_SGE: {
2743 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2744 if (!Known.isNonNegative())
2748 case ICmpInst::ICMP_EQ:
2749 case ICmpInst::ICMP_UGT:
2750 case ICmpInst::ICMP_UGE:
2751 return getFalse(ITy);
2752 case ICmpInst::ICMP_SLT:
2753 case ICmpInst::ICMP_SLE: {
2754 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2755 if (!Known.isNonNegative())
2759 case ICmpInst::ICMP_NE:
2760 case ICmpInst::ICMP_ULT:
2761 case ICmpInst::ICMP_ULE:
2762 return getTrue(ITy);
2766 // icmp pred X, (urem Y, X)
2767 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2771 case ICmpInst::ICMP_SGT:
2772 case ICmpInst::ICMP_SGE: {
2773 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2774 if (!Known.isNonNegative())
2778 case ICmpInst::ICMP_NE:
2779 case ICmpInst::ICMP_UGT:
2780 case ICmpInst::ICMP_UGE:
2781 return getTrue(ITy);
2782 case ICmpInst::ICMP_SLT:
2783 case ICmpInst::ICMP_SLE: {
2784 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2785 if (!Known.isNonNegative())
2789 case ICmpInst::ICMP_EQ:
2790 case ICmpInst::ICMP_ULT:
2791 case ICmpInst::ICMP_ULE:
2792 return getFalse(ITy);
2798 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2799 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2800 // icmp pred (X op Y), X
2801 if (Pred == ICmpInst::ICMP_UGT)
2802 return getFalse(ITy);
2803 if (Pred == ICmpInst::ICMP_ULE)
2804 return getTrue(ITy);
2809 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2810 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2811 // icmp pred X, (X op Y)
2812 if (Pred == ICmpInst::ICMP_ULT)
2813 return getFalse(ITy);
2814 if (Pred == ICmpInst::ICMP_UGE)
2815 return getTrue(ITy);
2822 // where CI2 is a power of 2 and CI isn't
2823 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2824 const APInt *CI2Val, *CIVal = &CI->getValue();
2825 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2826 CI2Val->isPowerOf2()) {
2827 if (!CIVal->isPowerOf2()) {
2828 // CI2 << X can equal zero in some circumstances,
2829 // this simplification is unsafe if CI is zero.
2831 // We know it is safe if:
2832 // - The shift is nsw, we can't shift out the one bit.
2833 // - The shift is nuw, we can't shift out the one bit.
2836 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2837 *CI2Val == 1 || !CI->isZero()) {
2838 if (Pred == ICmpInst::ICMP_EQ)
2839 return ConstantInt::getFalse(RHS->getContext());
2840 if (Pred == ICmpInst::ICMP_NE)
2841 return ConstantInt::getTrue(RHS->getContext());
2844 if (CIVal->isSignMask() && *CI2Val == 1) {
2845 if (Pred == ICmpInst::ICMP_UGT)
2846 return ConstantInt::getFalse(RHS->getContext());
2847 if (Pred == ICmpInst::ICMP_ULE)
2848 return ConstantInt::getTrue(RHS->getContext());
2853 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2854 LBO->getOperand(1) == RBO->getOperand(1)) {
2855 switch (LBO->getOpcode()) {
2858 case Instruction::UDiv:
2859 case Instruction::LShr:
2860 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2862 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2863 RBO->getOperand(0), Q, MaxRecurse - 1))
2866 case Instruction::SDiv:
2867 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2869 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2870 RBO->getOperand(0), Q, MaxRecurse - 1))
2873 case Instruction::AShr:
2874 if (!LBO->isExact() || !RBO->isExact())
2876 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2877 RBO->getOperand(0), Q, MaxRecurse - 1))
2880 case Instruction::Shl: {
2881 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2882 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2885 if (!NSW && ICmpInst::isSigned(Pred))
2887 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2888 RBO->getOperand(0), Q, MaxRecurse - 1))
2897 /// Simplify integer comparisons where at least one operand of the compare
2898 /// matches an integer min/max idiom.
2899 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2900 Value *RHS, const SimplifyQuery &Q,
2901 unsigned MaxRecurse) {
2902 Type *ITy = GetCompareTy(LHS); // The return type.
2904 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2905 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2907 // Signed variants on "max(a,b)>=a -> true".
2908 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2910 std::swap(A, B); // smax(A, B) pred A.
2911 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2912 // We analyze this as smax(A, B) pred A.
2914 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2915 (A == LHS || B == LHS)) {
2917 std::swap(A, B); // A pred smax(A, B).
2918 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2919 // We analyze this as smax(A, B) swapped-pred A.
2920 P = CmpInst::getSwappedPredicate(Pred);
2921 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2922 (A == RHS || B == RHS)) {
2924 std::swap(A, B); // smin(A, B) pred A.
2925 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2926 // We analyze this as smax(-A, -B) swapped-pred -A.
2927 // Note that we do not need to actually form -A or -B thanks to EqP.
2928 P = CmpInst::getSwappedPredicate(Pred);
2929 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2930 (A == LHS || B == LHS)) {
2932 std::swap(A, B); // A pred smin(A, B).
2933 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2934 // We analyze this as smax(-A, -B) pred -A.
2935 // Note that we do not need to actually form -A or -B thanks to EqP.
2938 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2939 // Cases correspond to "max(A, B) p A".
2943 case CmpInst::ICMP_EQ:
2944 case CmpInst::ICMP_SLE:
2945 // Equivalent to "A EqP B". This may be the same as the condition tested
2946 // in the max/min; if so, we can just return that.
2947 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2949 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2951 // Otherwise, see if "A EqP B" simplifies.
2953 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2956 case CmpInst::ICMP_NE:
2957 case CmpInst::ICMP_SGT: {
2958 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2959 // Equivalent to "A InvEqP B". This may be the same as the condition
2960 // tested in the max/min; if so, we can just return that.
2961 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2963 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2965 // Otherwise, see if "A InvEqP B" simplifies.
2967 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2971 case CmpInst::ICMP_SGE:
2973 return getTrue(ITy);
2974 case CmpInst::ICMP_SLT:
2976 return getFalse(ITy);
2980 // Unsigned variants on "max(a,b)>=a -> true".
2981 P = CmpInst::BAD_ICMP_PREDICATE;
2982 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2984 std::swap(A, B); // umax(A, B) pred A.
2985 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2986 // We analyze this as umax(A, B) pred A.
2988 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2989 (A == LHS || B == LHS)) {
2991 std::swap(A, B); // A pred umax(A, B).
2992 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2993 // We analyze this as umax(A, B) swapped-pred A.
2994 P = CmpInst::getSwappedPredicate(Pred);
2995 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2996 (A == RHS || B == RHS)) {
2998 std::swap(A, B); // umin(A, B) pred A.
2999 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3000 // We analyze this as umax(-A, -B) swapped-pred -A.
3001 // Note that we do not need to actually form -A or -B thanks to EqP.
3002 P = CmpInst::getSwappedPredicate(Pred);
3003 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3004 (A == LHS || B == LHS)) {
3006 std::swap(A, B); // A pred umin(A, B).
3007 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3008 // We analyze this as umax(-A, -B) pred -A.
3009 // Note that we do not need to actually form -A or -B thanks to EqP.
3012 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3013 // Cases correspond to "max(A, B) p A".
3017 case CmpInst::ICMP_EQ:
3018 case CmpInst::ICMP_ULE:
3019 // Equivalent to "A EqP B". This may be the same as the condition tested
3020 // in the max/min; if so, we can just return that.
3021 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3023 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3025 // Otherwise, see if "A EqP B" simplifies.
3027 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3030 case CmpInst::ICMP_NE:
3031 case CmpInst::ICMP_UGT: {
3032 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3033 // Equivalent to "A InvEqP B". This may be the same as the condition
3034 // tested in the max/min; if so, we can just return that.
3035 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3037 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3039 // Otherwise, see if "A InvEqP B" simplifies.
3041 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3045 case CmpInst::ICMP_UGE:
3047 return getTrue(ITy);
3048 case CmpInst::ICMP_ULT:
3050 return getFalse(ITy);
3054 // Variants on "max(x,y) >= min(x,z)".
3056 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3057 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3058 (A == C || A == D || B == C || B == D)) {
3059 // max(x, ?) pred min(x, ?).
3060 if (Pred == CmpInst::ICMP_SGE)
3062 return getTrue(ITy);
3063 if (Pred == CmpInst::ICMP_SLT)
3065 return getFalse(ITy);
3066 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3067 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3068 (A == C || A == D || B == C || B == D)) {
3069 // min(x, ?) pred max(x, ?).
3070 if (Pred == CmpInst::ICMP_SLE)
3072 return getTrue(ITy);
3073 if (Pred == CmpInst::ICMP_SGT)
3075 return getFalse(ITy);
3076 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3077 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3078 (A == C || A == D || B == C || B == D)) {
3079 // max(x, ?) pred min(x, ?).
3080 if (Pred == CmpInst::ICMP_UGE)
3082 return getTrue(ITy);
3083 if (Pred == CmpInst::ICMP_ULT)
3085 return getFalse(ITy);
3086 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3087 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3088 (A == C || A == D || B == C || B == D)) {
3089 // min(x, ?) pred max(x, ?).
3090 if (Pred == CmpInst::ICMP_ULE)
3092 return getTrue(ITy);
3093 if (Pred == CmpInst::ICMP_UGT)
3095 return getFalse(ITy);
3101 /// Given operands for an ICmpInst, see if we can fold the result.
3102 /// If not, this returns null.
3103 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3104 const SimplifyQuery &Q, unsigned MaxRecurse) {
3105 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3106 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3108 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3109 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3110 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3112 // If we have a constant, make sure it is on the RHS.
3113 std::swap(LHS, RHS);
3114 Pred = CmpInst::getSwappedPredicate(Pred);
3117 Type *ITy = GetCompareTy(LHS); // The return type.
3119 // icmp X, X -> true/false
3120 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3121 // because X could be 0.
3122 if (LHS == RHS || isa<UndefValue>(RHS))
3123 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3125 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3128 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3131 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3134 // If both operands have range metadata, use the metadata
3135 // to simplify the comparison.
3136 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3137 auto RHS_Instr = cast<Instruction>(RHS);
3138 auto LHS_Instr = cast<Instruction>(LHS);
3140 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3141 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3142 auto RHS_CR = getConstantRangeFromMetadata(
3143 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3144 auto LHS_CR = getConstantRangeFromMetadata(
3145 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3147 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3148 if (Satisfied_CR.contains(LHS_CR))
3149 return ConstantInt::getTrue(RHS->getContext());
3151 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3152 CmpInst::getInversePredicate(Pred), RHS_CR);
3153 if (InversedSatisfied_CR.contains(LHS_CR))
3154 return ConstantInt::getFalse(RHS->getContext());
3158 // Compare of cast, for example (zext X) != 0 -> X != 0
3159 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3160 Instruction *LI = cast<CastInst>(LHS);
3161 Value *SrcOp = LI->getOperand(0);
3162 Type *SrcTy = SrcOp->getType();
3163 Type *DstTy = LI->getType();
3165 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3166 // if the integer type is the same size as the pointer type.
3167 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3168 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3169 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3170 // Transfer the cast to the constant.
3171 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3172 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3175 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3176 if (RI->getOperand(0)->getType() == SrcTy)
3177 // Compare without the cast.
3178 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3184 if (isa<ZExtInst>(LHS)) {
3185 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3187 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3188 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3189 // Compare X and Y. Note that signed predicates become unsigned.
3190 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3191 SrcOp, RI->getOperand(0), Q,
3195 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3196 // too. If not, then try to deduce the result of the comparison.
3197 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3198 // Compute the constant that would happen if we truncated to SrcTy then
3199 // reextended to DstTy.
3200 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3201 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3203 // If the re-extended constant didn't change then this is effectively
3204 // also a case of comparing two zero-extended values.
3205 if (RExt == CI && MaxRecurse)
3206 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3207 SrcOp, Trunc, Q, MaxRecurse-1))
3210 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3211 // there. Use this to work out the result of the comparison.
3214 default: llvm_unreachable("Unknown ICmp predicate!");
3216 case ICmpInst::ICMP_EQ:
3217 case ICmpInst::ICMP_UGT:
3218 case ICmpInst::ICMP_UGE:
3219 return ConstantInt::getFalse(CI->getContext());
3221 case ICmpInst::ICMP_NE:
3222 case ICmpInst::ICMP_ULT:
3223 case ICmpInst::ICMP_ULE:
3224 return ConstantInt::getTrue(CI->getContext());
3226 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3227 // is non-negative then LHS <s RHS.
3228 case ICmpInst::ICMP_SGT:
3229 case ICmpInst::ICMP_SGE:
3230 return CI->getValue().isNegative() ?
3231 ConstantInt::getTrue(CI->getContext()) :
3232 ConstantInt::getFalse(CI->getContext());
3234 case ICmpInst::ICMP_SLT:
3235 case ICmpInst::ICMP_SLE:
3236 return CI->getValue().isNegative() ?
3237 ConstantInt::getFalse(CI->getContext()) :
3238 ConstantInt::getTrue(CI->getContext());
3244 if (isa<SExtInst>(LHS)) {
3245 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3247 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3248 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3249 // Compare X and Y. Note that the predicate does not change.
3250 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3254 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3255 // too. If not, then try to deduce the result of the comparison.
3256 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3257 // Compute the constant that would happen if we truncated to SrcTy then
3258 // reextended to DstTy.
3259 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3260 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3262 // If the re-extended constant didn't change then this is effectively
3263 // also a case of comparing two sign-extended values.
3264 if (RExt == CI && MaxRecurse)
3265 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3268 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3269 // bits there. Use this to work out the result of the comparison.
3272 default: llvm_unreachable("Unknown ICmp predicate!");
3273 case ICmpInst::ICMP_EQ:
3274 return ConstantInt::getFalse(CI->getContext());
3275 case ICmpInst::ICMP_NE:
3276 return ConstantInt::getTrue(CI->getContext());
3278 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3280 case ICmpInst::ICMP_SGT:
3281 case ICmpInst::ICMP_SGE:
3282 return CI->getValue().isNegative() ?
3283 ConstantInt::getTrue(CI->getContext()) :
3284 ConstantInt::getFalse(CI->getContext());
3285 case ICmpInst::ICMP_SLT:
3286 case ICmpInst::ICMP_SLE:
3287 return CI->getValue().isNegative() ?
3288 ConstantInt::getFalse(CI->getContext()) :
3289 ConstantInt::getTrue(CI->getContext());
3291 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3293 case ICmpInst::ICMP_UGT:
3294 case ICmpInst::ICMP_UGE:
3295 // Comparison is true iff the LHS <s 0.
3297 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3298 Constant::getNullValue(SrcTy),
3302 case ICmpInst::ICMP_ULT:
3303 case ICmpInst::ICMP_ULE:
3304 // Comparison is true iff the LHS >=s 0.
3306 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3307 Constant::getNullValue(SrcTy),
3317 // icmp eq|ne X, Y -> false|true if X != Y
3318 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
3319 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3320 LLVMContext &Ctx = LHS->getType()->getContext();
3321 return Pred == ICmpInst::ICMP_NE ?
3322 ConstantInt::getTrue(Ctx) : ConstantInt::getFalse(Ctx);
3325 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3328 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3331 // Simplify comparisons of related pointers using a powerful, recursive
3332 // GEP-walk when we have target data available..
3333 if (LHS->getType()->isPointerTy())
3334 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS))
3336 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3337 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3338 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3339 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3340 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3341 Q.DL.getTypeSizeInBits(CRHS->getType()))
3342 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI,
3343 CLHS->getPointerOperand(),
3344 CRHS->getPointerOperand()))
3347 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3348 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3349 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3350 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3351 (ICmpInst::isEquality(Pred) ||
3352 (GLHS->isInBounds() && GRHS->isInBounds() &&
3353 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3354 // The bases are equal and the indices are constant. Build a constant
3355 // expression GEP with the same indices and a null base pointer to see
3356 // what constant folding can make out of it.
3357 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3358 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3359 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3360 GLHS->getSourceElementType(), Null, IndicesLHS);
3362 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3363 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3364 GLHS->getSourceElementType(), Null, IndicesRHS);
3365 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3370 // If a bit is known to be zero for A and known to be one for B,
3371 // then A and B cannot be equal.
3372 if (ICmpInst::isEquality(Pred)) {
3373 const APInt *RHSVal;
3374 if (match(RHS, m_APInt(RHSVal))) {
3375 unsigned BitWidth = RHSVal->getBitWidth();
3376 KnownBits LHSKnown(BitWidth);
3377 computeKnownBits(LHS, LHSKnown, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
3378 if (LHSKnown.Zero.intersects(*RHSVal) ||
3379 !LHSKnown.One.isSubsetOf(*RHSVal))
3380 return Pred == ICmpInst::ICMP_EQ ? ConstantInt::getFalse(ITy)
3381 : ConstantInt::getTrue(ITy);
3385 // If the comparison is with the result of a select instruction, check whether
3386 // comparing with either branch of the select always yields the same value.
3387 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3388 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3391 // If the comparison is with the result of a phi instruction, check whether
3392 // doing the compare with each incoming phi value yields a common result.
3393 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3394 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3400 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3401 const SimplifyQuery &Q) {
3402 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3405 /// Given operands for an FCmpInst, see if we can fold the result.
3406 /// If not, this returns null.
3407 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3408 FastMathFlags FMF, const SimplifyQuery &Q,
3409 unsigned MaxRecurse) {
3410 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3411 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3413 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3414 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3415 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3417 // If we have a constant, make sure it is on the RHS.
3418 std::swap(LHS, RHS);
3419 Pred = CmpInst::getSwappedPredicate(Pred);
3422 // Fold trivial predicates.
3423 Type *RetTy = GetCompareTy(LHS);
3424 if (Pred == FCmpInst::FCMP_FALSE)
3425 return getFalse(RetTy);
3426 if (Pred == FCmpInst::FCMP_TRUE)
3427 return getTrue(RetTy);
3429 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3431 if (Pred == FCmpInst::FCMP_UNO)
3432 return getFalse(RetTy);
3433 if (Pred == FCmpInst::FCMP_ORD)
3434 return getTrue(RetTy);
3437 // fcmp pred x, undef and fcmp pred undef, x
3438 // fold to true if unordered, false if ordered
3439 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3440 // Choosing NaN for the undef will always make unordered comparison succeed
3441 // and ordered comparison fail.
3442 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3445 // fcmp x,x -> true/false. Not all compares are foldable.
3447 if (CmpInst::isTrueWhenEqual(Pred))
3448 return getTrue(RetTy);
3449 if (CmpInst::isFalseWhenEqual(Pred))
3450 return getFalse(RetTy);
3453 // Handle fcmp with constant RHS
3454 const ConstantFP *CFP = nullptr;
3455 if (const auto *RHSC = dyn_cast<Constant>(RHS)) {
3456 if (RHS->getType()->isVectorTy())
3457 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue());
3459 CFP = dyn_cast<ConstantFP>(RHSC);
3462 // If the constant is a nan, see if we can fold the comparison based on it.
3463 if (CFP->getValueAPF().isNaN()) {
3464 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3465 return getFalse(RetTy);
3466 assert(FCmpInst::isUnordered(Pred) &&
3467 "Comparison must be either ordered or unordered!");
3468 // True if unordered.
3469 return getTrue(RetTy);
3471 // Check whether the constant is an infinity.
3472 if (CFP->getValueAPF().isInfinity()) {
3473 if (CFP->getValueAPF().isNegative()) {
3475 case FCmpInst::FCMP_OLT:
3476 // No value is ordered and less than negative infinity.
3477 return getFalse(RetTy);
3478 case FCmpInst::FCMP_UGE:
3479 // All values are unordered with or at least negative infinity.
3480 return getTrue(RetTy);
3486 case FCmpInst::FCMP_OGT:
3487 // No value is ordered and greater than infinity.
3488 return getFalse(RetTy);
3489 case FCmpInst::FCMP_ULE:
3490 // All values are unordered with and at most infinity.
3491 return getTrue(RetTy);
3497 if (CFP->getValueAPF().isZero()) {
3499 case FCmpInst::FCMP_UGE:
3500 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3501 return getTrue(RetTy);
3503 case FCmpInst::FCMP_OLT:
3505 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3506 return getFalse(RetTy);
3514 // If the comparison is with the result of a select instruction, check whether
3515 // comparing with either branch of the select always yields the same value.
3516 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3517 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3520 // If the comparison is with the result of a phi instruction, check whether
3521 // doing the compare with each incoming phi value yields a common result.
3522 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3523 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3529 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3530 FastMathFlags FMF, const SimplifyQuery &Q) {
3531 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3534 /// See if V simplifies when its operand Op is replaced with RepOp.
3535 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3536 const SimplifyQuery &Q,
3537 unsigned MaxRecurse) {
3538 // Trivial replacement.
3542 auto *I = dyn_cast<Instruction>(V);
3546 // If this is a binary operator, try to simplify it with the replaced op.
3547 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3549 // %cmp = icmp eq i32 %x, 2147483647
3550 // %add = add nsw i32 %x, 1
3551 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3553 // We can't replace %sel with %add unless we strip away the flags.
3554 if (isa<OverflowingBinaryOperator>(B))
3555 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3557 if (isa<PossiblyExactOperator>(B))
3562 if (B->getOperand(0) == Op)
3563 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3565 if (B->getOperand(1) == Op)
3566 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3571 // Same for CmpInsts.
3572 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3574 if (C->getOperand(0) == Op)
3575 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3577 if (C->getOperand(1) == Op)
3578 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3583 // TODO: We could hand off more cases to instsimplify here.
3585 // If all operands are constant after substituting Op for RepOp then we can
3586 // constant fold the instruction.
3587 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3588 // Build a list of all constant operands.
3589 SmallVector<Constant *, 8> ConstOps;
3590 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3591 if (I->getOperand(i) == Op)
3592 ConstOps.push_back(CRepOp);
3593 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3594 ConstOps.push_back(COp);
3599 // All operands were constants, fold it.
3600 if (ConstOps.size() == I->getNumOperands()) {
3601 if (CmpInst *C = dyn_cast<CmpInst>(I))
3602 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3603 ConstOps[1], Q.DL, Q.TLI);
3605 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3606 if (!LI->isVolatile())
3607 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3609 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3616 /// Try to simplify a select instruction when its condition operand is an
3617 /// integer comparison where one operand of the compare is a constant.
3618 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3619 const APInt *Y, bool TrueWhenUnset) {
3622 // (X & Y) == 0 ? X & ~Y : X --> X
3623 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3624 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3626 return TrueWhenUnset ? FalseVal : TrueVal;
3628 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3629 // (X & Y) != 0 ? X : X & ~Y --> X
3630 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3632 return TrueWhenUnset ? FalseVal : TrueVal;
3634 if (Y->isPowerOf2()) {
3635 // (X & Y) == 0 ? X | Y : X --> X | Y
3636 // (X & Y) != 0 ? X | Y : X --> X
3637 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3639 return TrueWhenUnset ? TrueVal : FalseVal;
3641 // (X & Y) == 0 ? X : X | Y --> X
3642 // (X & Y) != 0 ? X : X | Y --> X | Y
3643 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3645 return TrueWhenUnset ? TrueVal : FalseVal;
3651 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3653 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal,
3655 bool TrueWhenUnset) {
3656 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
3660 APInt MinSignedValue;
3662 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) {
3663 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0
3664 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0
3665 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits();
3666 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth);
3668 // icmp slt X, 0 <--> icmp ne (and X, C), 0
3669 // icmp sgt X, -1 <--> icmp eq (and X, C), 0
3671 MinSignedValue = APInt::getSignedMinValue(BitWidth);
3674 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue,
3681 /// Try to simplify a select instruction when its condition operand is an
3682 /// integer comparison.
3683 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3684 Value *FalseVal, const SimplifyQuery &Q,
3685 unsigned MaxRecurse) {
3686 ICmpInst::Predicate Pred;
3687 Value *CmpLHS, *CmpRHS;
3688 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3691 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring
3692 // decomposeBitTestICmp() might help.
3693 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3696 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3697 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3698 Pred == ICmpInst::ICMP_EQ))
3700 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3701 // Comparing signed-less-than 0 checks if the sign bit is set.
3702 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3705 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3706 // Comparing signed-greater-than -1 checks if the sign bit is not set.
3707 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3712 if (CondVal->hasOneUse()) {
3714 if (match(CmpRHS, m_APInt(C))) {
3715 // X < MIN ? T : F --> F
3716 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3718 // X < MIN ? T : F --> F
3719 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3721 // X > MAX ? T : F --> F
3722 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3724 // X > MAX ? T : F --> F
3725 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3730 // If we have an equality comparison, then we know the value in one of the
3731 // arms of the select. See if substituting this value into the arm and
3732 // simplifying the result yields the same value as the other arm.
3733 if (Pred == ICmpInst::ICMP_EQ) {
3734 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3736 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3739 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3741 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3744 } else if (Pred == ICmpInst::ICMP_NE) {
3745 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3747 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3750 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3752 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3760 /// Given operands for a SelectInst, see if we can fold the result.
3761 /// If not, this returns null.
3762 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3763 Value *FalseVal, const SimplifyQuery &Q,
3764 unsigned MaxRecurse) {
3765 // select true, X, Y -> X
3766 // select false, X, Y -> Y
3767 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3768 if (CB->isAllOnesValue())
3770 if (CB->isNullValue())
3774 // select C, X, X -> X
3775 if (TrueVal == FalseVal)
3778 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3779 if (isa<Constant>(FalseVal))
3783 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3785 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3789 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3795 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3796 const SimplifyQuery &Q) {
3797 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3800 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3801 /// If not, this returns null.
3802 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3803 const SimplifyQuery &Q, unsigned) {
3804 // The type of the GEP pointer operand.
3806 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3808 // getelementptr P -> P.
3809 if (Ops.size() == 1)
3812 // Compute the (pointer) type returned by the GEP instruction.
3813 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3814 Type *GEPTy = PointerType::get(LastType, AS);
3815 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3816 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3817 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3818 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3820 if (isa<UndefValue>(Ops[0]))
3821 return UndefValue::get(GEPTy);
3823 if (Ops.size() == 2) {
3824 // getelementptr P, 0 -> P.
3825 if (match(Ops[1], m_Zero()))
3829 if (Ty->isSized()) {
3832 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3833 // getelementptr P, N -> P if P points to a type of zero size.
3834 if (TyAllocSize == 0)
3837 // The following transforms are only safe if the ptrtoint cast
3838 // doesn't truncate the pointers.
3839 if (Ops[1]->getType()->getScalarSizeInBits() ==
3840 Q.DL.getPointerSizeInBits(AS)) {
3841 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3842 if (match(P, m_Zero()))
3843 return Constant::getNullValue(GEPTy);
3845 if (match(P, m_PtrToInt(m_Value(Temp))))
3846 if (Temp->getType() == GEPTy)
3851 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3852 if (TyAllocSize == 1 &&
3853 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3854 if (Value *R = PtrToIntOrZero(P))
3857 // getelementptr V, (ashr (sub P, V), C) -> Q
3858 // if P points to a type of size 1 << C.
3860 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3861 m_ConstantInt(C))) &&
3862 TyAllocSize == 1ULL << C)
3863 if (Value *R = PtrToIntOrZero(P))
3866 // getelementptr V, (sdiv (sub P, V), C) -> Q
3867 // if P points to a type of size C.
3869 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3870 m_SpecificInt(TyAllocSize))))
3871 if (Value *R = PtrToIntOrZero(P))
3877 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3878 all_of(Ops.slice(1).drop_back(1),
3879 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3881 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3882 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3883 APInt BasePtrOffset(PtrWidth, 0);
3884 Value *StrippedBasePtr =
3885 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3888 // gep (gep V, C), (sub 0, V) -> C
3889 if (match(Ops.back(),
3890 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3891 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3892 return ConstantExpr::getIntToPtr(CI, GEPTy);
3894 // gep (gep V, C), (xor V, -1) -> C-1
3895 if (match(Ops.back(),
3896 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3897 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3898 return ConstantExpr::getIntToPtr(CI, GEPTy);
3903 // Check to see if this is constant foldable.
3904 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
3905 if (!isa<Constant>(Ops[i]))
3908 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3912 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3913 const SimplifyQuery &Q) {
3914 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3917 /// Given operands for an InsertValueInst, see if we can fold the result.
3918 /// If not, this returns null.
3919 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3920 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3922 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3923 if (Constant *CVal = dyn_cast<Constant>(Val))
3924 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3926 // insertvalue x, undef, n -> x
3927 if (match(Val, m_Undef()))
3930 // insertvalue x, (extractvalue y, n), n
3931 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3932 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3933 EV->getIndices() == Idxs) {
3934 // insertvalue undef, (extractvalue y, n), n -> y
3935 if (match(Agg, m_Undef()))
3936 return EV->getAggregateOperand();
3938 // insertvalue y, (extractvalue y, n), n -> y
3939 if (Agg == EV->getAggregateOperand())
3946 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3947 ArrayRef<unsigned> Idxs,
3948 const SimplifyQuery &Q) {
3949 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3952 /// Given operands for an ExtractValueInst, see if we can fold the result.
3953 /// If not, this returns null.
3954 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3955 const SimplifyQuery &, unsigned) {
3956 if (auto *CAgg = dyn_cast<Constant>(Agg))
3957 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3959 // extractvalue x, (insertvalue y, elt, n), n -> elt
3960 unsigned NumIdxs = Idxs.size();
3961 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3962 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3963 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3964 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3965 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3966 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3967 Idxs.slice(0, NumCommonIdxs)) {
3968 if (NumIdxs == NumInsertValueIdxs)
3969 return IVI->getInsertedValueOperand();
3977 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3978 const SimplifyQuery &Q) {
3979 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3982 /// Given operands for an ExtractElementInst, see if we can fold the result.
3983 /// If not, this returns null.
3984 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3986 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3987 if (auto *CIdx = dyn_cast<Constant>(Idx))
3988 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3990 // The index is not relevant if our vector is a splat.
3991 if (auto *Splat = CVec->getSplatValue())
3994 if (isa<UndefValue>(Vec))
3995 return UndefValue::get(Vec->getType()->getVectorElementType());
3998 // If extracting a specified index from the vector, see if we can recursively
3999 // find a previously computed scalar that was inserted into the vector.
4000 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
4001 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4007 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
4008 const SimplifyQuery &Q) {
4009 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
4012 /// See if we can fold the given phi. If not, returns null.
4013 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4014 // If all of the PHI's incoming values are the same then replace the PHI node
4015 // with the common value.
4016 Value *CommonValue = nullptr;
4017 bool HasUndefInput = false;
4018 for (Value *Incoming : PN->incoming_values()) {
4019 // If the incoming value is the phi node itself, it can safely be skipped.
4020 if (Incoming == PN) continue;
4021 if (isa<UndefValue>(Incoming)) {
4022 // Remember that we saw an undef value, but otherwise ignore them.
4023 HasUndefInput = true;
4026 if (CommonValue && Incoming != CommonValue)
4027 return nullptr; // Not the same, bail out.
4028 CommonValue = Incoming;
4031 // If CommonValue is null then all of the incoming values were either undef or
4032 // equal to the phi node itself.
4034 return UndefValue::get(PN->getType());
4036 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4037 // instruction, we cannot return X as the result of the PHI node unless it
4038 // dominates the PHI block.
4040 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4045 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4046 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4047 if (auto *C = dyn_cast<Constant>(Op))
4048 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4050 if (auto *CI = dyn_cast<CastInst>(Op)) {
4051 auto *Src = CI->getOperand(0);
4052 Type *SrcTy = Src->getType();
4053 Type *MidTy = CI->getType();
4055 if (Src->getType() == Ty) {
4056 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4057 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4059 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4061 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4063 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4064 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4065 SrcIntPtrTy, MidIntPtrTy,
4066 DstIntPtrTy) == Instruction::BitCast)
4072 if (CastOpc == Instruction::BitCast)
4073 if (Op->getType() == Ty)
4079 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4080 const SimplifyQuery &Q) {
4081 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4084 /// For the given destination element of a shuffle, peek through shuffles to
4085 /// match a root vector source operand that contains that element in the same
4086 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4087 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4088 int MaskVal, Value *RootVec,
4089 unsigned MaxRecurse) {
4093 // Bail out if any mask value is undefined. That kind of shuffle may be
4094 // simplified further based on demanded bits or other folds.
4098 // The mask value chooses which source operand we need to look at next.
4099 int InVecNumElts = Op0->getType()->getVectorNumElements();
4100 int RootElt = MaskVal;
4101 Value *SourceOp = Op0;
4102 if (MaskVal >= InVecNumElts) {
4103 RootElt = MaskVal - InVecNumElts;
4107 // If the source operand is a shuffle itself, look through it to find the
4108 // matching root vector.
4109 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4110 return foldIdentityShuffles(
4111 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4112 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4115 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4118 // The source operand is not a shuffle. Initialize the root vector value for
4119 // this shuffle if that has not been done yet.
4123 // Give up as soon as a source operand does not match the existing root value.
4124 if (RootVec != SourceOp)
4127 // The element must be coming from the same lane in the source vector
4128 // (although it may have crossed lanes in intermediate shuffles).
4129 if (RootElt != DestElt)
4135 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4136 Type *RetTy, const SimplifyQuery &Q,
4137 unsigned MaxRecurse) {
4138 if (isa<UndefValue>(Mask))
4139 return UndefValue::get(RetTy);
4141 Type *InVecTy = Op0->getType();
4142 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4143 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4145 SmallVector<int, 32> Indices;
4146 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4147 assert(MaskNumElts == Indices.size() &&
4148 "Size of Indices not same as number of mask elements?");
4150 // Canonicalization: If mask does not select elements from an input vector,
4151 // replace that input vector with undef.
4152 bool MaskSelects0 = false, MaskSelects1 = false;
4153 for (unsigned i = 0; i != MaskNumElts; ++i) {
4154 if (Indices[i] == -1)
4156 if ((unsigned)Indices[i] < InVecNumElts)
4157 MaskSelects0 = true;
4159 MaskSelects1 = true;
4162 Op0 = UndefValue::get(InVecTy);
4164 Op1 = UndefValue::get(InVecTy);
4166 auto *Op0Const = dyn_cast<Constant>(Op0);
4167 auto *Op1Const = dyn_cast<Constant>(Op1);
4169 // If all operands are constant, constant fold the shuffle.
4170 if (Op0Const && Op1Const)
4171 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4173 // Canonicalization: if only one input vector is constant, it shall be the
4175 if (Op0Const && !Op1Const) {
4176 std::swap(Op0, Op1);
4177 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4180 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4181 // value type is same as the input vectors' type.
4182 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4183 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4184 OpShuf->getMask()->getSplatValue())
4187 // Don't fold a shuffle with undef mask elements. This may get folded in a
4188 // better way using demanded bits or other analysis.
4189 // TODO: Should we allow this?
4190 if (find(Indices, -1) != Indices.end())
4193 // Check if every element of this shuffle can be mapped back to the
4194 // corresponding element of a single root vector. If so, we don't need this
4195 // shuffle. This handles simple identity shuffles as well as chains of
4196 // shuffles that may widen/narrow and/or move elements across lanes and back.
4197 Value *RootVec = nullptr;
4198 for (unsigned i = 0; i != MaskNumElts; ++i) {
4199 // Note that recursion is limited for each vector element, so if any element
4200 // exceeds the limit, this will fail to simplify.
4202 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4204 // We can't replace a widening/narrowing shuffle with one of its operands.
4205 if (!RootVec || RootVec->getType() != RetTy)
4211 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4212 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4213 Type *RetTy, const SimplifyQuery &Q) {
4214 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4217 //=== Helper functions for higher up the class hierarchy.
4219 /// Given operands for a BinaryOperator, see if we can fold the result.
4220 /// If not, this returns null.
4221 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4222 const SimplifyQuery &Q, unsigned MaxRecurse) {
4224 case Instruction::Add:
4225 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4226 case Instruction::FAdd:
4227 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4228 case Instruction::Sub:
4229 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4230 case Instruction::FSub:
4231 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4232 case Instruction::Mul:
4233 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4234 case Instruction::FMul:
4235 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4236 case Instruction::SDiv:
4237 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4238 case Instruction::UDiv:
4239 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4240 case Instruction::FDiv:
4241 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4242 case Instruction::SRem:
4243 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4244 case Instruction::URem:
4245 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4246 case Instruction::FRem:
4247 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4248 case Instruction::Shl:
4249 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4250 case Instruction::LShr:
4251 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4252 case Instruction::AShr:
4253 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4254 case Instruction::And:
4255 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4256 case Instruction::Or:
4257 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4258 case Instruction::Xor:
4259 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4261 llvm_unreachable("Unexpected opcode");
4265 /// Given operands for a BinaryOperator, see if we can fold the result.
4266 /// If not, this returns null.
4267 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4268 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4269 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4270 const FastMathFlags &FMF, const SimplifyQuery &Q,
4271 unsigned MaxRecurse) {
4273 case Instruction::FAdd:
4274 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4275 case Instruction::FSub:
4276 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4277 case Instruction::FMul:
4278 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4279 case Instruction::FDiv:
4280 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4282 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4286 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4287 const SimplifyQuery &Q) {
4288 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4291 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4292 FastMathFlags FMF, const SimplifyQuery &Q) {
4293 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4296 /// Given operands for a CmpInst, see if we can fold the result.
4297 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4298 const SimplifyQuery &Q, unsigned MaxRecurse) {
4299 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4300 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4301 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4304 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4305 const SimplifyQuery &Q) {
4306 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4309 static bool IsIdempotent(Intrinsic::ID ID) {
4311 default: return false;
4313 // Unary idempotent: f(f(x)) = f(x)
4314 case Intrinsic::fabs:
4315 case Intrinsic::floor:
4316 case Intrinsic::ceil:
4317 case Intrinsic::trunc:
4318 case Intrinsic::rint:
4319 case Intrinsic::nearbyint:
4320 case Intrinsic::round:
4325 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4326 const DataLayout &DL) {
4327 GlobalValue *PtrSym;
4329 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4332 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4333 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4334 Type *Int32PtrTy = Int32Ty->getPointerTo();
4335 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4337 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4338 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4341 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4342 if (OffsetInt % 4 != 0)
4345 Constant *C = ConstantExpr::getGetElementPtr(
4346 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4347 ConstantInt::get(Int64Ty, OffsetInt / 4));
4348 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4352 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4356 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4357 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4362 if (LoadedCE->getOpcode() != Instruction::Sub)
4365 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4366 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4368 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4370 Constant *LoadedRHS = LoadedCE->getOperand(1);
4371 GlobalValue *LoadedRHSSym;
4372 APInt LoadedRHSOffset;
4373 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4375 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4378 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4381 static bool maskIsAllZeroOrUndef(Value *Mask) {
4382 auto *ConstMask = dyn_cast<Constant>(Mask);
4385 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4387 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4389 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4390 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4397 template <typename IterTy>
4398 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4399 const SimplifyQuery &Q, unsigned MaxRecurse) {
4400 Intrinsic::ID IID = F->getIntrinsicID();
4401 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4404 if (NumOperands == 1) {
4405 // Perform idempotent optimizations
4406 if (IsIdempotent(IID)) {
4407 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4408 if (II->getIntrinsicID() == IID)
4414 case Intrinsic::fabs: {
4415 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4425 if (NumOperands == 2) {
4426 Value *LHS = *ArgBegin;
4427 Value *RHS = *(ArgBegin + 1);
4428 Type *ReturnType = F->getReturnType();
4431 case Intrinsic::usub_with_overflow:
4432 case Intrinsic::ssub_with_overflow: {
4433 // X - X -> { 0, false }
4435 return Constant::getNullValue(ReturnType);
4437 // X - undef -> undef
4438 // undef - X -> undef
4439 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4440 return UndefValue::get(ReturnType);
4444 case Intrinsic::uadd_with_overflow:
4445 case Intrinsic::sadd_with_overflow: {
4446 // X + undef -> undef
4447 if (isa<UndefValue>(RHS))
4448 return UndefValue::get(ReturnType);
4452 case Intrinsic::umul_with_overflow:
4453 case Intrinsic::smul_with_overflow: {
4454 // X * 0 -> { 0, false }
4455 if (match(RHS, m_Zero()))
4456 return Constant::getNullValue(ReturnType);
4458 // X * undef -> { 0, false }
4459 if (match(RHS, m_Undef()))
4460 return Constant::getNullValue(ReturnType);
4464 case Intrinsic::load_relative: {
4465 Constant *C0 = dyn_cast<Constant>(LHS);
4466 Constant *C1 = dyn_cast<Constant>(RHS);
4468 return SimplifyRelativeLoad(C0, C1, Q.DL);
4476 // Simplify calls to llvm.masked.load.*
4478 case Intrinsic::masked_load: {
4479 Value *MaskArg = ArgBegin[2];
4480 Value *PassthruArg = ArgBegin[3];
4481 // If the mask is all zeros or undef, the "passthru" argument is the result.
4482 if (maskIsAllZeroOrUndef(MaskArg))
4491 template <typename IterTy>
4492 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
4493 const SimplifyQuery &Q, unsigned MaxRecurse) {
4494 Type *Ty = V->getType();
4495 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4496 Ty = PTy->getElementType();
4497 FunctionType *FTy = cast<FunctionType>(Ty);
4499 // call undef -> undef
4500 // call null -> undef
4501 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4502 return UndefValue::get(FTy->getReturnType());
4504 Function *F = dyn_cast<Function>(V);
4508 if (F->isIntrinsic())
4509 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4512 if (!canConstantFoldCallTo(F))
4515 SmallVector<Constant *, 4> ConstantArgs;
4516 ConstantArgs.reserve(ArgEnd - ArgBegin);
4517 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4518 Constant *C = dyn_cast<Constant>(*I);
4521 ConstantArgs.push_back(C);
4524 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
4527 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
4528 User::op_iterator ArgEnd, const SimplifyQuery &Q) {
4529 return ::SimplifyCall(V, ArgBegin, ArgEnd, Q, RecursionLimit);
4532 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
4533 const SimplifyQuery &Q) {
4534 return ::SimplifyCall(V, Args.begin(), Args.end(), Q, RecursionLimit);
4537 /// See if we can compute a simplified version of this instruction.
4538 /// If not, this returns null.
4540 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4541 OptimizationRemarkEmitter *ORE) {
4542 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4545 switch (I->getOpcode()) {
4547 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4549 case Instruction::FAdd:
4550 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4551 I->getFastMathFlags(), Q);
4553 case Instruction::Add:
4554 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4555 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4556 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4558 case Instruction::FSub:
4559 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4560 I->getFastMathFlags(), Q);
4562 case Instruction::Sub:
4563 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4564 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4565 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4567 case Instruction::FMul:
4568 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4569 I->getFastMathFlags(), Q);
4571 case Instruction::Mul:
4572 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4574 case Instruction::SDiv:
4575 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4577 case Instruction::UDiv:
4578 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4580 case Instruction::FDiv:
4581 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4582 I->getFastMathFlags(), Q);
4584 case Instruction::SRem:
4585 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4587 case Instruction::URem:
4588 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4590 case Instruction::FRem:
4591 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4592 I->getFastMathFlags(), Q);
4594 case Instruction::Shl:
4595 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4596 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4597 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4599 case Instruction::LShr:
4600 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4601 cast<BinaryOperator>(I)->isExact(), Q);
4603 case Instruction::AShr:
4604 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4605 cast<BinaryOperator>(I)->isExact(), Q);
4607 case Instruction::And:
4608 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4610 case Instruction::Or:
4611 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4613 case Instruction::Xor:
4614 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4616 case Instruction::ICmp:
4617 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4618 I->getOperand(0), I->getOperand(1), Q);
4620 case Instruction::FCmp:
4622 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4623 I->getOperand(1), I->getFastMathFlags(), Q);
4625 case Instruction::Select:
4626 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4627 I->getOperand(2), Q);
4629 case Instruction::GetElementPtr: {
4630 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4631 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4635 case Instruction::InsertValue: {
4636 InsertValueInst *IV = cast<InsertValueInst>(I);
4637 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4638 IV->getInsertedValueOperand(),
4639 IV->getIndices(), Q);
4642 case Instruction::ExtractValue: {
4643 auto *EVI = cast<ExtractValueInst>(I);
4644 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4645 EVI->getIndices(), Q);
4648 case Instruction::ExtractElement: {
4649 auto *EEI = cast<ExtractElementInst>(I);
4650 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4651 EEI->getIndexOperand(), Q);
4654 case Instruction::ShuffleVector: {
4655 auto *SVI = cast<ShuffleVectorInst>(I);
4656 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4657 SVI->getMask(), SVI->getType(), Q);
4660 case Instruction::PHI:
4661 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4663 case Instruction::Call: {
4664 CallSite CS(cast<CallInst>(I));
4665 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), Q);
4668 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4669 #include "llvm/IR/Instruction.def"
4670 #undef HANDLE_CAST_INST
4672 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4674 case Instruction::Alloca:
4675 // No simplifications for Alloca and it can't be constant folded.
4680 // In general, it is possible for computeKnownBits to determine all bits in a
4681 // value even when the operands are not all constants.
4682 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4683 unsigned BitWidth = I->getType()->getScalarSizeInBits();
4684 KnownBits Known(BitWidth);
4685 computeKnownBits(I, Known, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4686 if (Known.isConstant())
4687 Result = ConstantInt::get(I->getType(), Known.getConstant());
4690 /// If called on unreachable code, the above logic may report that the
4691 /// instruction simplified to itself. Make life easier for users by
4692 /// detecting that case here, returning a safe value instead.
4693 return Result == I ? UndefValue::get(I->getType()) : Result;
4696 /// \brief Implementation of recursive simplification through an instruction's
4699 /// This is the common implementation of the recursive simplification routines.
4700 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4701 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4702 /// instructions to process and attempt to simplify it using
4703 /// InstructionSimplify.
4705 /// This routine returns 'true' only when *it* simplifies something. The passed
4706 /// in simplified value does not count toward this.
4707 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4708 const TargetLibraryInfo *TLI,
4709 const DominatorTree *DT,
4710 AssumptionCache *AC) {
4711 bool Simplified = false;
4712 SmallSetVector<Instruction *, 8> Worklist;
4713 const DataLayout &DL = I->getModule()->getDataLayout();
4715 // If we have an explicit value to collapse to, do that round of the
4716 // simplification loop by hand initially.
4718 for (User *U : I->users())
4720 Worklist.insert(cast<Instruction>(U));
4722 // Replace the instruction with its simplified value.
4723 I->replaceAllUsesWith(SimpleV);
4725 // Gracefully handle edge cases where the instruction is not wired into any
4727 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4728 !I->mayHaveSideEffects())
4729 I->eraseFromParent();
4734 // Note that we must test the size on each iteration, the worklist can grow.
4735 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4738 // See if this instruction simplifies.
4739 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4745 // Stash away all the uses of the old instruction so we can check them for
4746 // recursive simplifications after a RAUW. This is cheaper than checking all
4747 // uses of To on the recursive step in most cases.
4748 for (User *U : I->users())
4749 Worklist.insert(cast<Instruction>(U));
4751 // Replace the instruction with its simplified value.
4752 I->replaceAllUsesWith(SimpleV);
4754 // Gracefully handle edge cases where the instruction is not wired into any
4756 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4757 !I->mayHaveSideEffects())
4758 I->eraseFromParent();
4763 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4764 const TargetLibraryInfo *TLI,
4765 const DominatorTree *DT,
4766 AssumptionCache *AC) {
4767 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4770 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4771 const TargetLibraryInfo *TLI,
4772 const DominatorTree *DT,
4773 AssumptionCache *AC) {
4774 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4775 assert(SimpleV && "Must provide a simplified value.");
4776 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4780 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4781 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4782 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4783 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4784 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4785 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4786 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4787 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4790 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4791 const DataLayout &DL) {
4792 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4795 template <class T, class... TArgs>
4796 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4798 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4799 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4800 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4801 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4803 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,