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/CmpInstAnalysis.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/Analysis/VectorUtils.h"
33 #include "llvm/IR/ConstantRange.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/Dominators.h"
36 #include "llvm/IR/GetElementPtrTypeIterator.h"
37 #include "llvm/IR/GlobalAlias.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/IR/PatternMatch.h"
40 #include "llvm/IR/ValueHandle.h"
41 #include "llvm/Support/KnownBits.h"
44 using namespace llvm::PatternMatch;
46 #define DEBUG_TYPE "instsimplify"
48 enum { RecursionLimit = 3 };
50 STATISTIC(NumExpand, "Number of expansions");
51 STATISTIC(NumReassoc, "Number of reassociations");
53 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
54 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
56 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
57 const SimplifyQuery &, unsigned);
58 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
60 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
61 const SimplifyQuery &Q, unsigned MaxRecurse);
62 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
64 static Value *SimplifyCastInst(unsigned, Value *, Type *,
65 const SimplifyQuery &, unsigned);
67 /// For a boolean type or a vector of boolean type, return false or a vector
68 /// with every element false.
69 static Constant *getFalse(Type *Ty) {
70 return ConstantInt::getFalse(Ty);
73 /// For a boolean type or a vector of boolean type, return true or a vector
74 /// with every element true.
75 static Constant *getTrue(Type *Ty) {
76 return ConstantInt::getTrue(Ty);
79 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
80 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
82 CmpInst *Cmp = dyn_cast<CmpInst>(V);
85 CmpInst::Predicate CPred = Cmp->getPredicate();
86 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
87 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
89 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
93 /// Does the given value dominate the specified phi node?
94 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
95 Instruction *I = dyn_cast<Instruction>(V);
97 // Arguments and constants dominate all instructions.
100 // If we are processing instructions (and/or basic blocks) that have not been
101 // fully added to a function, the parent nodes may still be null. Simply
102 // return the conservative answer in these cases.
103 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
106 // If we have a DominatorTree then do a precise test.
108 return DT->dominates(I, P);
110 // Otherwise, if the instruction is in the entry block and is not an invoke,
111 // then it obviously dominates all phi nodes.
112 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
119 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
120 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
121 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
122 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
123 /// Returns the simplified value, or null if no simplification was performed.
124 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
125 Instruction::BinaryOps OpcodeToExpand,
126 const SimplifyQuery &Q, unsigned MaxRecurse) {
127 // Recursion is always used, so bail out at once if we already hit the limit.
131 // Check whether the expression has the form "(A op' B) op C".
132 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
133 if (Op0->getOpcode() == OpcodeToExpand) {
134 // It does! Try turning it into "(A op C) op' (B op C)".
135 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
136 // Do "A op C" and "B op C" both simplify?
137 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
138 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
139 // They do! Return "L op' R" if it simplifies or is already available.
140 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
141 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
142 && L == B && R == A)) {
146 // Otherwise return "L op' R" if it simplifies.
147 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
154 // Check whether the expression has the form "A op (B op' C)".
155 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
156 if (Op1->getOpcode() == OpcodeToExpand) {
157 // It does! Try turning it into "(A op B) op' (A op C)".
158 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
159 // Do "A op B" and "A op C" both simplify?
160 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
161 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
162 // They do! Return "L op' R" if it simplifies or is already available.
163 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
164 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
165 && L == C && R == B)) {
169 // Otherwise return "L op' R" if it simplifies.
170 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
180 /// Generic simplifications for associative binary operations.
181 /// Returns the simpler value, or null if none was found.
182 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
183 Value *LHS, Value *RHS,
184 const SimplifyQuery &Q,
185 unsigned MaxRecurse) {
186 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
188 // Recursion is always used, so bail out at once if we already hit the limit.
192 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
193 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
195 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
196 if (Op0 && Op0->getOpcode() == Opcode) {
197 Value *A = Op0->getOperand(0);
198 Value *B = Op0->getOperand(1);
201 // Does "B op C" simplify?
202 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
203 // It does! Return "A op V" if it simplifies or is already available.
204 // If V equals B then "A op V" is just the LHS.
205 if (V == B) return LHS;
206 // Otherwise return "A op V" if it simplifies.
207 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
214 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
215 if (Op1 && Op1->getOpcode() == Opcode) {
217 Value *B = Op1->getOperand(0);
218 Value *C = Op1->getOperand(1);
220 // Does "A op B" simplify?
221 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
222 // It does! Return "V op C" if it simplifies or is already available.
223 // If V equals B then "V op C" is just the RHS.
224 if (V == B) return RHS;
225 // Otherwise return "V op C" if it simplifies.
226 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
233 // The remaining transforms require commutativity as well as associativity.
234 if (!Instruction::isCommutative(Opcode))
237 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
238 if (Op0 && Op0->getOpcode() == Opcode) {
239 Value *A = Op0->getOperand(0);
240 Value *B = Op0->getOperand(1);
243 // Does "C op A" simplify?
244 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
245 // It does! Return "V op B" if it simplifies or is already available.
246 // If V equals A then "V op B" is just the LHS.
247 if (V == A) return LHS;
248 // Otherwise return "V op B" if it simplifies.
249 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
256 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
257 if (Op1 && Op1->getOpcode() == Opcode) {
259 Value *B = Op1->getOperand(0);
260 Value *C = Op1->getOperand(1);
262 // Does "C op A" simplify?
263 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
264 // It does! Return "B op V" if it simplifies or is already available.
265 // If V equals C then "B op V" is just the RHS.
266 if (V == C) return RHS;
267 // Otherwise return "B op V" if it simplifies.
268 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
278 /// In the case of a binary operation with a select instruction as an operand,
279 /// try to simplify the binop by seeing whether evaluating it on both branches
280 /// of the select results in the same value. Returns the common value if so,
281 /// otherwise returns null.
282 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
283 Value *RHS, const SimplifyQuery &Q,
284 unsigned MaxRecurse) {
285 // Recursion is always used, so bail out at once if we already hit the limit.
290 if (isa<SelectInst>(LHS)) {
291 SI = cast<SelectInst>(LHS);
293 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
294 SI = cast<SelectInst>(RHS);
297 // Evaluate the BinOp on the true and false branches of the select.
301 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
302 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
304 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
305 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
308 // If they simplified to the same value, then return the common value.
309 // If they both failed to simplify then return null.
313 // If one branch simplified to undef, return the other one.
314 if (TV && isa<UndefValue>(TV))
316 if (FV && isa<UndefValue>(FV))
319 // If applying the operation did not change the true and false select values,
320 // then the result of the binop is the select itself.
321 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
324 // If one branch simplified and the other did not, and the simplified
325 // value is equal to the unsimplified one, return the simplified value.
326 // For example, select (cond, X, X & Z) & Z -> X & Z.
327 if ((FV && !TV) || (TV && !FV)) {
328 // Check that the simplified value has the form "X op Y" where "op" is the
329 // same as the original operation.
330 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
331 if (Simplified && Simplified->getOpcode() == Opcode) {
332 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
333 // We already know that "op" is the same as for the simplified value. See
334 // if the operands match too. If so, return the simplified value.
335 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
336 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
337 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
338 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
339 Simplified->getOperand(1) == UnsimplifiedRHS)
341 if (Simplified->isCommutative() &&
342 Simplified->getOperand(1) == UnsimplifiedLHS &&
343 Simplified->getOperand(0) == UnsimplifiedRHS)
351 /// In the case of a comparison with a select instruction, try to simplify the
352 /// comparison by seeing whether both branches of the select result in the same
353 /// value. Returns the common value if so, otherwise returns null.
354 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
355 Value *RHS, const SimplifyQuery &Q,
356 unsigned MaxRecurse) {
357 // Recursion is always used, so bail out at once if we already hit the limit.
361 // Make sure the select is on the LHS.
362 if (!isa<SelectInst>(LHS)) {
364 Pred = CmpInst::getSwappedPredicate(Pred);
366 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
367 SelectInst *SI = cast<SelectInst>(LHS);
368 Value *Cond = SI->getCondition();
369 Value *TV = SI->getTrueValue();
370 Value *FV = SI->getFalseValue();
372 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
373 // Does "cmp TV, RHS" simplify?
374 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
376 // It not only simplified, it simplified to the select condition. Replace
378 TCmp = getTrue(Cond->getType());
380 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
381 // condition then we can replace it with 'true'. Otherwise give up.
382 if (!isSameCompare(Cond, Pred, TV, RHS))
384 TCmp = getTrue(Cond->getType());
387 // Does "cmp FV, RHS" simplify?
388 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
390 // It not only simplified, it simplified to the select condition. Replace
392 FCmp = getFalse(Cond->getType());
394 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
395 // condition then we can replace it with 'false'. Otherwise give up.
396 if (!isSameCompare(Cond, Pred, FV, RHS))
398 FCmp = getFalse(Cond->getType());
401 // If both sides simplified to the same value, then use it as the result of
402 // the original comparison.
406 // The remaining cases only make sense if the select condition has the same
407 // type as the result of the comparison, so bail out if this is not so.
408 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
410 // If the false value simplified to false, then the result of the compare
411 // is equal to "Cond && TCmp". This also catches the case when the false
412 // value simplified to false and the true value to true, returning "Cond".
413 if (match(FCmp, m_Zero()))
414 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
416 // If the true value simplified to true, then the result of the compare
417 // is equal to "Cond || FCmp".
418 if (match(TCmp, m_One()))
419 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
421 // Finally, if the false value simplified to true and the true value to
422 // false, then the result of the compare is equal to "!Cond".
423 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
425 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
432 /// In the case of a binary operation with an operand that is a PHI instruction,
433 /// try to simplify the binop by seeing whether evaluating it on the incoming
434 /// phi values yields the same result for every value. If so returns the common
435 /// value, otherwise returns null.
436 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
437 Value *RHS, const SimplifyQuery &Q,
438 unsigned MaxRecurse) {
439 // Recursion is always used, so bail out at once if we already hit the limit.
444 if (isa<PHINode>(LHS)) {
445 PI = cast<PHINode>(LHS);
446 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
447 if (!ValueDominatesPHI(RHS, PI, Q.DT))
450 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
451 PI = cast<PHINode>(RHS);
452 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
453 if (!ValueDominatesPHI(LHS, PI, Q.DT))
457 // Evaluate the BinOp on the incoming phi values.
458 Value *CommonValue = nullptr;
459 for (Value *Incoming : PI->incoming_values()) {
460 // If the incoming value is the phi node itself, it can safely be skipped.
461 if (Incoming == PI) continue;
462 Value *V = PI == LHS ?
463 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
464 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
465 // If the operation failed to simplify, or simplified to a different value
466 // to previously, then give up.
467 if (!V || (CommonValue && V != CommonValue))
475 /// In the case of a comparison with a PHI instruction, try to simplify the
476 /// comparison by seeing whether comparing with all of the incoming phi values
477 /// yields the same result every time. If so returns the common result,
478 /// otherwise returns null.
479 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
480 const SimplifyQuery &Q, unsigned MaxRecurse) {
481 // Recursion is always used, so bail out at once if we already hit the limit.
485 // Make sure the phi is on the LHS.
486 if (!isa<PHINode>(LHS)) {
488 Pred = CmpInst::getSwappedPredicate(Pred);
490 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
491 PHINode *PI = cast<PHINode>(LHS);
493 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
494 if (!ValueDominatesPHI(RHS, PI, Q.DT))
497 // Evaluate the BinOp on the incoming phi values.
498 Value *CommonValue = nullptr;
499 for (Value *Incoming : PI->incoming_values()) {
500 // If the incoming value is the phi node itself, it can safely be skipped.
501 if (Incoming == PI) continue;
502 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
503 // If the operation failed to simplify, or simplified to a different value
504 // to previously, then give up.
505 if (!V || (CommonValue && V != CommonValue))
513 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
514 Value *&Op0, Value *&Op1,
515 const SimplifyQuery &Q) {
516 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
517 if (auto *CRHS = dyn_cast<Constant>(Op1))
518 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
520 // Canonicalize the constant to the RHS if this is a commutative operation.
521 if (Instruction::isCommutative(Opcode))
527 /// Given operands for an Add, see if we can fold the result.
528 /// If not, this returns null.
529 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
530 const SimplifyQuery &Q, unsigned MaxRecurse) {
531 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
534 // X + undef -> undef
535 if (match(Op1, m_Undef()))
539 if (match(Op1, m_Zero()))
546 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
547 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
550 // X + ~X -> -1 since ~X = -X-1
551 Type *Ty = Op0->getType();
552 if (match(Op0, m_Not(m_Specific(Op1))) ||
553 match(Op1, m_Not(m_Specific(Op0))))
554 return Constant::getAllOnesValue(Ty);
556 // add nsw/nuw (xor Y, signmask), signmask --> Y
557 // The no-wrapping add guarantees that the top bit will be set by the add.
558 // Therefore, the xor must be clearing the already set sign bit of Y.
559 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
560 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
564 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
565 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
568 // Try some generic simplifications for associative operations.
569 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
573 // Threading Add over selects and phi nodes is pointless, so don't bother.
574 // Threading over the select in "A + select(cond, B, C)" means evaluating
575 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
576 // only if B and C are equal. If B and C are equal then (since we assume
577 // that operands have already been simplified) "select(cond, B, C)" should
578 // have been simplified to the common value of B and C already. Analysing
579 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
580 // for threading over phi nodes.
585 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
586 const SimplifyQuery &Query) {
587 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
590 /// \brief Compute the base pointer and cumulative constant offsets for V.
592 /// This strips all constant offsets off of V, leaving it the base pointer, and
593 /// accumulates the total constant offset applied in the returned constant. It
594 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
595 /// no constant offsets applied.
597 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
598 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
600 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
601 bool AllowNonInbounds = false) {
602 assert(V->getType()->isPtrOrPtrVectorTy());
604 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
605 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
607 // Even though we don't look through PHI nodes, we could be called on an
608 // instruction in an unreachable block, which may be on a cycle.
609 SmallPtrSet<Value *, 4> Visited;
612 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
613 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
614 !GEP->accumulateConstantOffset(DL, Offset))
616 V = GEP->getPointerOperand();
617 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
618 V = cast<Operator>(V)->getOperand(0);
619 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
620 if (GA->isInterposable())
622 V = GA->getAliasee();
624 if (auto CS = CallSite(V))
625 if (Value *RV = CS.getReturnedArgOperand()) {
631 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
632 } while (Visited.insert(V).second);
634 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
635 if (V->getType()->isVectorTy())
636 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
641 /// \brief Compute the constant difference between two pointer values.
642 /// If the difference is not a constant, returns zero.
643 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
645 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
646 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
648 // If LHS and RHS are not related via constant offsets to the same base
649 // value, there is nothing we can do here.
653 // Otherwise, the difference of LHS - RHS can be computed as:
655 // = (LHSOffset + Base) - (RHSOffset + Base)
656 // = LHSOffset - RHSOffset
657 return ConstantExpr::getSub(LHSOffset, RHSOffset);
660 /// Given operands for a Sub, see if we can fold the result.
661 /// If not, this returns null.
662 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
663 const SimplifyQuery &Q, unsigned MaxRecurse) {
664 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
667 // X - undef -> undef
668 // undef - X -> undef
669 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
670 return UndefValue::get(Op0->getType());
673 if (match(Op1, m_Zero()))
678 return Constant::getNullValue(Op0->getType());
680 // Is this a negation?
681 if (match(Op0, m_Zero())) {
682 // 0 - X -> 0 if the sub is NUW.
686 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
687 if (Known.Zero.isMaxSignedValue()) {
688 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
689 // Op1 must be 0 because negating the minimum signed value is undefined.
693 // 0 - X -> X if X is 0 or the minimum signed value.
698 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
699 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
700 Value *X = nullptr, *Y = nullptr, *Z = Op1;
701 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
702 // See if "V === Y - Z" simplifies.
703 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
704 // It does! Now see if "X + V" simplifies.
705 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
706 // It does, we successfully reassociated!
710 // See if "V === X - Z" simplifies.
711 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
712 // It does! Now see if "Y + V" simplifies.
713 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
714 // It does, we successfully reassociated!
720 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
721 // For example, X - (X + 1) -> -1
723 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
724 // See if "V === X - Y" simplifies.
725 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
726 // It does! Now see if "V - Z" simplifies.
727 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
728 // It does, we successfully reassociated!
732 // See if "V === X - Z" simplifies.
733 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
734 // It does! Now see if "V - Y" simplifies.
735 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
736 // It does, we successfully reassociated!
742 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
743 // For example, X - (X - Y) -> Y.
745 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
746 // See if "V === Z - X" simplifies.
747 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
748 // It does! Now see if "V + Y" simplifies.
749 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
750 // It does, we successfully reassociated!
755 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
756 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
757 match(Op1, m_Trunc(m_Value(Y))))
758 if (X->getType() == Y->getType())
759 // See if "V === X - Y" simplifies.
760 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
761 // It does! Now see if "trunc V" simplifies.
762 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
764 // It does, return the simplified "trunc V".
767 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
768 if (match(Op0, m_PtrToInt(m_Value(X))) &&
769 match(Op1, m_PtrToInt(m_Value(Y))))
770 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
771 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
774 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
775 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
778 // Threading Sub over selects and phi nodes is pointless, so don't bother.
779 // Threading over the select in "A - select(cond, B, C)" means evaluating
780 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
781 // only if B and C are equal. If B and C are equal then (since we assume
782 // that operands have already been simplified) "select(cond, B, C)" should
783 // have been simplified to the common value of B and C already. Analysing
784 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
785 // for threading over phi nodes.
790 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
791 const SimplifyQuery &Q) {
792 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
795 /// Given operands for a Mul, see if we can fold the result.
796 /// If not, this returns null.
797 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
798 unsigned MaxRecurse) {
799 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
803 if (match(Op1, m_Undef()))
804 return Constant::getNullValue(Op0->getType());
807 if (match(Op1, m_Zero()))
811 if (match(Op1, m_One()))
814 // (X / Y) * Y -> X if the division is exact.
816 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
817 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
821 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
822 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
825 // Try some generic simplifications for associative operations.
826 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
830 // Mul distributes over Add. Try some generic simplifications based on this.
831 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
835 // If the operation is with the result of a select instruction, check whether
836 // operating on either branch of the select always yields the same value.
837 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
838 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
842 // If the operation is with the result of a phi instruction, check whether
843 // operating on all incoming values of the phi always yields the same value.
844 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
845 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
852 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
853 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
856 /// Check for common or similar folds of integer division or integer remainder.
857 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
858 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
859 Type *Ty = Op0->getType();
861 // X / undef -> undef
862 // X % undef -> undef
863 if (match(Op1, m_Undef()))
868 // We don't need to preserve faults!
869 if (match(Op1, m_Zero()))
870 return UndefValue::get(Ty);
872 // If any element of a constant divisor vector is zero, the whole op is undef.
873 auto *Op1C = dyn_cast<Constant>(Op1);
874 if (Op1C && Ty->isVectorTy()) {
875 unsigned NumElts = Ty->getVectorNumElements();
876 for (unsigned i = 0; i != NumElts; ++i) {
877 Constant *Elt = Op1C->getAggregateElement(i);
878 if (Elt && Elt->isNullValue())
879 return UndefValue::get(Ty);
885 if (match(Op0, m_Undef()))
886 return Constant::getNullValue(Ty);
890 if (match(Op0, m_Zero()))
896 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
900 // If this is a boolean op (single-bit element type), we can't have
901 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
902 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1))
903 return IsDiv ? Op0 : Constant::getNullValue(Ty);
908 /// Given a predicate and two operands, return true if the comparison is true.
909 /// This is a helper for div/rem simplification where we return some other value
910 /// when we can prove a relationship between the operands.
911 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
912 const SimplifyQuery &Q, unsigned MaxRecurse) {
913 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
914 Constant *C = dyn_cast_or_null<Constant>(V);
915 return (C && C->isAllOnesValue());
918 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
919 /// to simplify X % Y to X.
920 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
921 unsigned MaxRecurse, bool IsSigned) {
922 // Recursion is always used, so bail out at once if we already hit the limit.
929 // We require that 1 operand is a simple constant. That could be extended to
930 // 2 variables if we computed the sign bit for each.
932 // Make sure that a constant is not the minimum signed value because taking
933 // the abs() of that is undefined.
934 Type *Ty = X->getType();
936 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
937 // Is the variable divisor magnitude always greater than the constant
938 // dividend magnitude?
939 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
940 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
941 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
942 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
943 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
946 if (match(Y, m_APInt(C))) {
947 // Special-case: we can't take the abs() of a minimum signed value. If
948 // that's the divisor, then all we have to do is prove that the dividend
949 // is also not the minimum signed value.
950 if (C->isMinSignedValue())
951 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
953 // Is the variable dividend magnitude always less than the constant
954 // divisor magnitude?
955 // |X| < |C| --> X > -abs(C) and X < abs(C)
956 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
957 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
958 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
959 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
965 // IsSigned == false.
966 // Is the dividend unsigned less than the divisor?
967 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
970 /// These are simplifications common to SDiv and UDiv.
971 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
972 const SimplifyQuery &Q, unsigned MaxRecurse) {
973 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
976 if (Value *V = simplifyDivRem(Op0, Op1, true))
979 bool IsSigned = Opcode == Instruction::SDiv;
981 // (X * Y) / Y -> X if the multiplication does not overflow.
982 Value *X = nullptr, *Y = nullptr;
983 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
984 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
985 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
986 // If the Mul knows it does not overflow, then we are good to go.
987 if ((IsSigned && Mul->hasNoSignedWrap()) ||
988 (!IsSigned && Mul->hasNoUnsignedWrap()))
990 // If X has the form X = A / Y then X * Y cannot overflow.
991 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
992 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
996 // (X rem Y) / Y -> 0
997 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
998 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
999 return Constant::getNullValue(Op0->getType());
1001 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1002 ConstantInt *C1, *C2;
1003 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1004 match(Op1, m_ConstantInt(C2))) {
1006 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1008 return Constant::getNullValue(Op0->getType());
1011 // If the operation is with the result of a select instruction, check whether
1012 // operating on either branch of the select always yields the same value.
1013 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1014 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1017 // If the operation is with the result of a phi instruction, check whether
1018 // operating on all incoming values of the phi always yields the same value.
1019 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1020 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1023 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1024 return Constant::getNullValue(Op0->getType());
1029 /// These are simplifications common to SRem and URem.
1030 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1031 const SimplifyQuery &Q, unsigned MaxRecurse) {
1032 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1035 if (Value *V = simplifyDivRem(Op0, Op1, false))
1038 // (X % Y) % Y -> X % Y
1039 if ((Opcode == Instruction::SRem &&
1040 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1041 (Opcode == Instruction::URem &&
1042 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1045 // If the operation is with the result of a select instruction, check whether
1046 // operating on either branch of the select always yields the same value.
1047 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1048 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1051 // If the operation is with the result of a phi instruction, check whether
1052 // operating on all incoming values of the phi always yields the same value.
1053 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1054 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1057 // If X / Y == 0, then X % Y == X.
1058 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1064 /// Given operands for an SDiv, see if we can fold the result.
1065 /// If not, this returns null.
1066 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1067 unsigned MaxRecurse) {
1068 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1071 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1072 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1075 /// Given operands for a UDiv, see if we can fold the result.
1076 /// If not, this returns null.
1077 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1078 unsigned MaxRecurse) {
1079 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1082 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1083 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1086 /// Given operands for an SRem, see if we can fold the result.
1087 /// If not, this returns null.
1088 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1089 unsigned MaxRecurse) {
1090 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1093 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1094 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1097 /// Given operands for a URem, see if we can fold the result.
1098 /// If not, this returns null.
1099 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1100 unsigned MaxRecurse) {
1101 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1104 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1105 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1108 /// Returns true if a shift by \c Amount always yields undef.
1109 static bool isUndefShift(Value *Amount) {
1110 Constant *C = dyn_cast<Constant>(Amount);
1114 // X shift by undef -> undef because it may shift by the bitwidth.
1115 if (isa<UndefValue>(C))
1118 // Shifting by the bitwidth or more is undefined.
1119 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1120 if (CI->getValue().getLimitedValue() >=
1121 CI->getType()->getScalarSizeInBits())
1124 // If all lanes of a vector shift are undefined the whole shift is.
1125 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1126 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1127 if (!isUndefShift(C->getAggregateElement(I)))
1135 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1136 /// If not, this returns null.
1137 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1138 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1139 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1142 // 0 shift by X -> 0
1143 if (match(Op0, m_Zero()))
1146 // X shift by 0 -> X
1147 if (match(Op1, m_Zero()))
1150 // Fold undefined shifts.
1151 if (isUndefShift(Op1))
1152 return UndefValue::get(Op0->getType());
1154 // If the operation is with the result of a select instruction, check whether
1155 // operating on either branch of the select always yields the same value.
1156 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1157 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1160 // If the operation is with the result of a phi instruction, check whether
1161 // operating on all incoming values of the phi always yields the same value.
1162 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1163 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1166 // If any bits in the shift amount make that value greater than or equal to
1167 // the number of bits in the type, the shift is undefined.
1168 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1169 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1170 return UndefValue::get(Op0->getType());
1172 // If all valid bits in the shift amount are known zero, the first operand is
1174 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1175 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1181 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1182 /// fold the result. If not, this returns null.
1183 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1184 Value *Op1, bool isExact, const SimplifyQuery &Q,
1185 unsigned MaxRecurse) {
1186 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1191 return Constant::getNullValue(Op0->getType());
1194 // undef >> X -> undef (if it's exact)
1195 if (match(Op0, m_Undef()))
1196 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1198 // The low bit cannot be shifted out of an exact shift if it is set.
1200 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1201 if (Op0Known.One[0])
1208 /// Given operands for an Shl, see if we can fold the result.
1209 /// If not, this returns null.
1210 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1211 const SimplifyQuery &Q, unsigned MaxRecurse) {
1212 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1216 // undef << X -> undef if (if it's NSW/NUW)
1217 if (match(Op0, m_Undef()))
1218 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1220 // (X >> A) << A -> X
1222 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1227 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1228 const SimplifyQuery &Q) {
1229 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1232 /// Given operands for an LShr, see if we can fold the result.
1233 /// If not, this returns null.
1234 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1235 const SimplifyQuery &Q, unsigned MaxRecurse) {
1236 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1240 // (X << A) >> A -> X
1242 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1248 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1249 const SimplifyQuery &Q) {
1250 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1253 /// Given operands for an AShr, see if we can fold the result.
1254 /// If not, this returns null.
1255 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1256 const SimplifyQuery &Q, unsigned MaxRecurse) {
1257 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1261 // all ones >>a X -> all ones
1262 if (match(Op0, m_AllOnes()))
1265 // (X << A) >> A -> X
1267 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1270 // Arithmetic shifting an all-sign-bit value is a no-op.
1271 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1272 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1278 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1279 const SimplifyQuery &Q) {
1280 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1283 /// Commuted variants are assumed to be handled by calling this function again
1284 /// with the parameters swapped.
1285 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1286 ICmpInst *UnsignedICmp, bool IsAnd) {
1289 ICmpInst::Predicate EqPred;
1290 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1291 !ICmpInst::isEquality(EqPred))
1294 ICmpInst::Predicate UnsignedPred;
1295 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1296 ICmpInst::isUnsigned(UnsignedPred))
1298 else if (match(UnsignedICmp,
1299 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1300 ICmpInst::isUnsigned(UnsignedPred))
1301 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1305 // X < Y && Y != 0 --> X < Y
1306 // X < Y || Y != 0 --> Y != 0
1307 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1308 return IsAnd ? UnsignedICmp : ZeroICmp;
1310 // X >= Y || Y != 0 --> true
1311 // X >= Y || Y == 0 --> X >= Y
1312 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1313 if (EqPred == ICmpInst::ICMP_NE)
1314 return getTrue(UnsignedICmp->getType());
1315 return UnsignedICmp;
1318 // X < Y && Y == 0 --> false
1319 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1321 return getFalse(UnsignedICmp->getType());
1326 /// Commuted variants are assumed to be handled by calling this function again
1327 /// with the parameters swapped.
1328 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1329 ICmpInst::Predicate Pred0, Pred1;
1331 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1332 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1335 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1336 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1337 // can eliminate Op1 from this 'and'.
1338 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1341 // Check for any combination of predicates that are guaranteed to be disjoint.
1342 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1343 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1344 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1345 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1346 return getFalse(Op0->getType());
1351 /// Commuted variants are assumed to be handled by calling this function again
1352 /// with the parameters swapped.
1353 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1354 ICmpInst::Predicate Pred0, Pred1;
1356 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1357 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1360 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1361 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1362 // can eliminate Op0 from this 'or'.
1363 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1366 // Check for any combination of predicates that cover the entire range of
1368 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1369 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1370 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1371 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1372 return getTrue(Op0->getType());
1377 /// Test if a pair of compares with a shared operand and 2 constants has an
1378 /// empty set intersection, full set union, or if one compare is a superset of
1380 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1382 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1383 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1386 const APInt *C0, *C1;
1387 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1388 !match(Cmp1->getOperand(1), m_APInt(C1)))
1391 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1392 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1394 // For and-of-compares, check if the intersection is empty:
1395 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1396 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1397 return getFalse(Cmp0->getType());
1399 // For or-of-compares, check if the union is full:
1400 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1401 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1402 return getTrue(Cmp0->getType());
1404 // Is one range a superset of the other?
1405 // If this is and-of-compares, take the smaller set:
1406 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1407 // If this is or-of-compares, take the larger set:
1408 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1409 if (Range0.contains(Range1))
1410 return IsAnd ? Cmp1 : Cmp0;
1411 if (Range1.contains(Range0))
1412 return IsAnd ? Cmp0 : Cmp1;
1417 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1418 // (icmp (add V, C0), C1) & (icmp V, C0)
1419 ICmpInst::Predicate Pred0, Pred1;
1420 const APInt *C0, *C1;
1422 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1425 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1428 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1429 if (AddInst->getOperand(1) != Op1->getOperand(1))
1432 Type *ITy = Op0->getType();
1433 bool isNSW = AddInst->hasNoSignedWrap();
1434 bool isNUW = AddInst->hasNoUnsignedWrap();
1436 const APInt Delta = *C1 - *C0;
1437 if (C0->isStrictlyPositive()) {
1439 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1440 return getFalse(ITy);
1441 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1442 return getFalse(ITy);
1445 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1446 return getFalse(ITy);
1447 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1448 return getFalse(ITy);
1451 if (C0->getBoolValue() && isNUW) {
1453 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1454 return getFalse(ITy);
1456 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1457 return getFalse(ITy);
1463 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1464 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1466 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1469 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1471 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1474 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1477 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1479 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1485 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1486 // (icmp (add V, C0), C1) | (icmp V, C0)
1487 ICmpInst::Predicate Pred0, Pred1;
1488 const APInt *C0, *C1;
1490 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1493 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1496 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1497 if (AddInst->getOperand(1) != Op1->getOperand(1))
1500 Type *ITy = Op0->getType();
1501 bool isNSW = AddInst->hasNoSignedWrap();
1502 bool isNUW = AddInst->hasNoUnsignedWrap();
1504 const APInt Delta = *C1 - *C0;
1505 if (C0->isStrictlyPositive()) {
1507 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1508 return getTrue(ITy);
1509 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1510 return getTrue(ITy);
1513 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1514 return getTrue(ITy);
1515 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1516 return getTrue(ITy);
1519 if (C0->getBoolValue() && isNUW) {
1521 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1522 return getTrue(ITy);
1524 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1525 return getTrue(ITy);
1531 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1532 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1534 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1537 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1539 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1542 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1545 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1547 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1553 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1554 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1555 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1556 if (LHS0->getType() != RHS0->getType())
1559 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1560 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1561 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1562 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1563 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1564 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1565 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1566 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1567 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1568 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1569 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1570 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1571 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1574 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1575 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1576 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1577 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1578 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1579 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1580 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1581 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1582 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1583 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1590 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1591 // Look through casts of the 'and' operands to find compares.
1592 auto *Cast0 = dyn_cast<CastInst>(Op0);
1593 auto *Cast1 = dyn_cast<CastInst>(Op1);
1594 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1595 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1596 Op0 = Cast0->getOperand(0);
1597 Op1 = Cast1->getOperand(0);
1601 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1602 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1604 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1605 simplifyOrOfICmps(ICmp0, ICmp1);
1607 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1608 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1610 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1617 // If we looked through casts, we can only handle a constant simplification
1618 // because we are not allowed to create a cast instruction here.
1619 if (auto *C = dyn_cast<Constant>(V))
1620 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1625 /// Given operands for an And, see if we can fold the result.
1626 /// If not, this returns null.
1627 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1628 unsigned MaxRecurse) {
1629 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1633 if (match(Op1, m_Undef()))
1634 return Constant::getNullValue(Op0->getType());
1641 if (match(Op1, m_Zero()))
1645 if (match(Op1, m_AllOnes()))
1648 // A & ~A = ~A & A = 0
1649 if (match(Op0, m_Not(m_Specific(Op1))) ||
1650 match(Op1, m_Not(m_Specific(Op0))))
1651 return Constant::getNullValue(Op0->getType());
1654 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1658 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1661 // A mask that only clears known zeros of a shifted value is a no-op.
1665 if (match(Op1, m_APInt(Mask))) {
1666 // If all bits in the inverted and shifted mask are clear:
1667 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1668 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1669 (~(*Mask)).lshr(*ShAmt).isNullValue())
1672 // If all bits in the inverted and shifted mask are clear:
1673 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1674 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1675 (~(*Mask)).shl(*ShAmt).isNullValue())
1679 // A & (-A) = A if A is a power of two or zero.
1680 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1681 match(Op1, m_Neg(m_Specific(Op0)))) {
1682 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1685 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1690 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1693 // Try some generic simplifications for associative operations.
1694 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1698 // And distributes over Or. Try some generic simplifications based on this.
1699 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1703 // And distributes over Xor. Try some generic simplifications based on this.
1704 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1708 // If the operation is with the result of a select instruction, check whether
1709 // operating on either branch of the select always yields the same value.
1710 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1711 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1715 // If the operation is with the result of a phi instruction, check whether
1716 // operating on all incoming values of the phi always yields the same value.
1717 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1718 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1725 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1726 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1729 /// Given operands for an Or, see if we can fold the result.
1730 /// If not, this returns null.
1731 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1732 unsigned MaxRecurse) {
1733 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1737 if (match(Op1, m_Undef()))
1738 return Constant::getAllOnesValue(Op0->getType());
1745 if (match(Op1, m_Zero()))
1749 if (match(Op1, m_AllOnes()))
1752 // A | ~A = ~A | A = -1
1753 if (match(Op0, m_Not(m_Specific(Op1))) ||
1754 match(Op1, m_Not(m_Specific(Op0))))
1755 return Constant::getAllOnesValue(Op0->getType());
1758 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1762 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1765 // ~(A & ?) | A = -1
1766 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1767 return Constant::getAllOnesValue(Op1->getType());
1769 // A | ~(A & ?) = -1
1770 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1771 return Constant::getAllOnesValue(Op0->getType());
1774 // (A & ~B) | (A ^ B) -> (A ^ B)
1775 // (~B & A) | (A ^ B) -> (A ^ B)
1776 // (A & ~B) | (B ^ A) -> (B ^ A)
1777 // (~B & A) | (B ^ A) -> (B ^ A)
1778 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1779 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1780 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1783 // Commute the 'or' operands.
1784 // (A ^ B) | (A & ~B) -> (A ^ B)
1785 // (A ^ B) | (~B & A) -> (A ^ B)
1786 // (B ^ A) | (A & ~B) -> (B ^ A)
1787 // (B ^ A) | (~B & A) -> (B ^ A)
1788 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1789 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1790 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1793 // (A & B) | (~A ^ B) -> (~A ^ B)
1794 // (B & A) | (~A ^ B) -> (~A ^ B)
1795 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1796 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1797 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1798 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1799 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1802 // (~A ^ B) | (A & B) -> (~A ^ B)
1803 // (~A ^ B) | (B & A) -> (~A ^ B)
1804 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1805 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1806 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1807 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1808 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1811 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1814 // Try some generic simplifications for associative operations.
1815 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1819 // Or distributes over And. Try some generic simplifications based on this.
1820 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1824 // If the operation is with the result of a select instruction, check whether
1825 // operating on either branch of the select always yields the same value.
1826 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1827 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1831 // (A & C1)|(B & C2)
1832 const APInt *C1, *C2;
1833 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1834 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1836 // (A & C1)|(B & C2)
1837 // If we have: ((V + N) & C1) | (V & C2)
1838 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1839 // replace with V+N.
1841 if (C2->isMask() && // C2 == 0+1+
1842 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1843 // Add commutes, try both ways.
1844 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1847 // Or commutes, try both ways.
1849 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1850 // Add commutes, try both ways.
1851 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1857 // If the operation is with the result of a phi instruction, check whether
1858 // operating on all incoming values of the phi always yields the same value.
1859 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1860 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1866 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1867 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1870 /// Given operands for a Xor, see if we can fold the result.
1871 /// If not, this returns null.
1872 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1873 unsigned MaxRecurse) {
1874 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1877 // A ^ undef -> undef
1878 if (match(Op1, m_Undef()))
1882 if (match(Op1, m_Zero()))
1887 return Constant::getNullValue(Op0->getType());
1889 // A ^ ~A = ~A ^ A = -1
1890 if (match(Op0, m_Not(m_Specific(Op1))) ||
1891 match(Op1, m_Not(m_Specific(Op0))))
1892 return Constant::getAllOnesValue(Op0->getType());
1894 // Try some generic simplifications for associative operations.
1895 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1899 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1900 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1901 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1902 // only if B and C are equal. If B and C are equal then (since we assume
1903 // that operands have already been simplified) "select(cond, B, C)" should
1904 // have been simplified to the common value of B and C already. Analysing
1905 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1906 // for threading over phi nodes.
1911 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1912 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1916 static Type *GetCompareTy(Value *Op) {
1917 return CmpInst::makeCmpResultType(Op->getType());
1920 /// Rummage around inside V looking for something equivalent to the comparison
1921 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1922 /// Helper function for analyzing max/min idioms.
1923 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1924 Value *LHS, Value *RHS) {
1925 SelectInst *SI = dyn_cast<SelectInst>(V);
1928 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1931 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1932 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1934 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1935 LHS == CmpRHS && RHS == CmpLHS)
1940 // A significant optimization not implemented here is assuming that alloca
1941 // addresses are not equal to incoming argument values. They don't *alias*,
1942 // as we say, but that doesn't mean they aren't equal, so we take a
1943 // conservative approach.
1945 // This is inspired in part by C++11 5.10p1:
1946 // "Two pointers of the same type compare equal if and only if they are both
1947 // null, both point to the same function, or both represent the same
1950 // This is pretty permissive.
1952 // It's also partly due to C11 6.5.9p6:
1953 // "Two pointers compare equal if and only if both are null pointers, both are
1954 // pointers to the same object (including a pointer to an object and a
1955 // subobject at its beginning) or function, both are pointers to one past the
1956 // last element of the same array object, or one is a pointer to one past the
1957 // end of one array object and the other is a pointer to the start of a
1958 // different array object that happens to immediately follow the first array
1959 // object in the address space.)
1961 // C11's version is more restrictive, however there's no reason why an argument
1962 // couldn't be a one-past-the-end value for a stack object in the caller and be
1963 // equal to the beginning of a stack object in the callee.
1965 // If the C and C++ standards are ever made sufficiently restrictive in this
1966 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1967 // this optimization.
1969 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
1970 const DominatorTree *DT, CmpInst::Predicate Pred,
1971 AssumptionCache *AC, const Instruction *CxtI,
1972 Value *LHS, Value *RHS) {
1973 // First, skip past any trivial no-ops.
1974 LHS = LHS->stripPointerCasts();
1975 RHS = RHS->stripPointerCasts();
1977 // A non-null pointer is not equal to a null pointer.
1978 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
1979 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1980 return ConstantInt::get(GetCompareTy(LHS),
1981 !CmpInst::isTrueWhenEqual(Pred));
1983 // We can only fold certain predicates on pointer comparisons.
1988 // Equality comaprisons are easy to fold.
1989 case CmpInst::ICMP_EQ:
1990 case CmpInst::ICMP_NE:
1993 // We can only handle unsigned relational comparisons because 'inbounds' on
1994 // a GEP only protects against unsigned wrapping.
1995 case CmpInst::ICMP_UGT:
1996 case CmpInst::ICMP_UGE:
1997 case CmpInst::ICMP_ULT:
1998 case CmpInst::ICMP_ULE:
1999 // However, we have to switch them to their signed variants to handle
2000 // negative indices from the base pointer.
2001 Pred = ICmpInst::getSignedPredicate(Pred);
2005 // Strip off any constant offsets so that we can reason about them.
2006 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2007 // here and compare base addresses like AliasAnalysis does, however there are
2008 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2009 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2010 // doesn't need to guarantee pointer inequality when it says NoAlias.
2011 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2012 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2014 // If LHS and RHS are related via constant offsets to the same base
2015 // value, we can replace it with an icmp which just compares the offsets.
2017 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2019 // Various optimizations for (in)equality comparisons.
2020 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2021 // Different non-empty allocations that exist at the same time have
2022 // different addresses (if the program can tell). Global variables always
2023 // exist, so they always exist during the lifetime of each other and all
2024 // allocas. Two different allocas usually have different addresses...
2026 // However, if there's an @llvm.stackrestore dynamically in between two
2027 // allocas, they may have the same address. It's tempting to reduce the
2028 // scope of the problem by only looking at *static* allocas here. That would
2029 // cover the majority of allocas while significantly reducing the likelihood
2030 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2031 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2032 // an entry block. Also, if we have a block that's not attached to a
2033 // function, we can't tell if it's "static" under the current definition.
2034 // Theoretically, this problem could be fixed by creating a new kind of
2035 // instruction kind specifically for static allocas. Such a new instruction
2036 // could be required to be at the top of the entry block, thus preventing it
2037 // from being subject to a @llvm.stackrestore. Instcombine could even
2038 // convert regular allocas into these special allocas. It'd be nifty.
2039 // However, until then, this problem remains open.
2041 // So, we'll assume that two non-empty allocas have different addresses
2044 // With all that, if the offsets are within the bounds of their allocations
2045 // (and not one-past-the-end! so we can't use inbounds!), and their
2046 // allocations aren't the same, the pointers are not equal.
2048 // Note that it's not necessary to check for LHS being a global variable
2049 // address, due to canonicalization and constant folding.
2050 if (isa<AllocaInst>(LHS) &&
2051 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2052 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2053 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2054 uint64_t LHSSize, RHSSize;
2055 if (LHSOffsetCI && RHSOffsetCI &&
2056 getObjectSize(LHS, LHSSize, DL, TLI) &&
2057 getObjectSize(RHS, RHSSize, DL, TLI)) {
2058 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2059 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2060 if (!LHSOffsetValue.isNegative() &&
2061 !RHSOffsetValue.isNegative() &&
2062 LHSOffsetValue.ult(LHSSize) &&
2063 RHSOffsetValue.ult(RHSSize)) {
2064 return ConstantInt::get(GetCompareTy(LHS),
2065 !CmpInst::isTrueWhenEqual(Pred));
2069 // Repeat the above check but this time without depending on DataLayout
2070 // or being able to compute a precise size.
2071 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2072 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2073 LHSOffset->isNullValue() &&
2074 RHSOffset->isNullValue())
2075 return ConstantInt::get(GetCompareTy(LHS),
2076 !CmpInst::isTrueWhenEqual(Pred));
2079 // Even if an non-inbounds GEP occurs along the path we can still optimize
2080 // equality comparisons concerning the result. We avoid walking the whole
2081 // chain again by starting where the last calls to
2082 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2083 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2084 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2086 return ConstantExpr::getICmp(Pred,
2087 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2088 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2090 // If one side of the equality comparison must come from a noalias call
2091 // (meaning a system memory allocation function), and the other side must
2092 // come from a pointer that cannot overlap with dynamically-allocated
2093 // memory within the lifetime of the current function (allocas, byval
2094 // arguments, globals), then determine the comparison result here.
2095 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2096 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2097 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2099 // Is the set of underlying objects all noalias calls?
2100 auto IsNAC = [](ArrayRef<Value *> Objects) {
2101 return all_of(Objects, isNoAliasCall);
2104 // Is the set of underlying objects all things which must be disjoint from
2105 // noalias calls. For allocas, we consider only static ones (dynamic
2106 // allocas might be transformed into calls to malloc not simultaneously
2107 // live with the compared-to allocation). For globals, we exclude symbols
2108 // that might be resolve lazily to symbols in another dynamically-loaded
2109 // library (and, thus, could be malloc'ed by the implementation).
2110 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2111 return all_of(Objects, [](Value *V) {
2112 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2113 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2114 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2115 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2116 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2117 !GV->isThreadLocal();
2118 if (const Argument *A = dyn_cast<Argument>(V))
2119 return A->hasByValAttr();
2124 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2125 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2126 return ConstantInt::get(GetCompareTy(LHS),
2127 !CmpInst::isTrueWhenEqual(Pred));
2129 // Fold comparisons for non-escaping pointer even if the allocation call
2130 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2131 // dynamic allocation call could be either of the operands.
2132 Value *MI = nullptr;
2133 if (isAllocLikeFn(LHS, TLI) &&
2134 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2136 else if (isAllocLikeFn(RHS, TLI) &&
2137 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2139 // FIXME: We should also fold the compare when the pointer escapes, but the
2140 // compare dominates the pointer escape
2141 if (MI && !PointerMayBeCaptured(MI, true, true))
2142 return ConstantInt::get(GetCompareTy(LHS),
2143 CmpInst::isFalseWhenEqual(Pred));
2150 /// Fold an icmp when its operands have i1 scalar type.
2151 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2152 Value *RHS, const SimplifyQuery &Q) {
2153 Type *ITy = GetCompareTy(LHS); // The return type.
2154 Type *OpTy = LHS->getType(); // The operand type.
2155 if (!OpTy->isIntOrIntVectorTy(1))
2158 // A boolean compared to true/false can be simplified in 14 out of the 20
2159 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2160 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2161 if (match(RHS, m_Zero())) {
2163 case CmpInst::ICMP_NE: // X != 0 -> X
2164 case CmpInst::ICMP_UGT: // X >u 0 -> X
2165 case CmpInst::ICMP_SLT: // X <s 0 -> X
2168 case CmpInst::ICMP_ULT: // X <u 0 -> false
2169 case CmpInst::ICMP_SGT: // X >s 0 -> false
2170 return getFalse(ITy);
2172 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2173 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2174 return getTrue(ITy);
2178 } else if (match(RHS, m_One())) {
2180 case CmpInst::ICMP_EQ: // X == 1 -> X
2181 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2182 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2185 case CmpInst::ICMP_UGT: // X >u 1 -> false
2186 case CmpInst::ICMP_SLT: // X <s -1 -> false
2187 return getFalse(ITy);
2189 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2190 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2191 return getTrue(ITy);
2200 case ICmpInst::ICMP_UGE:
2201 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2202 return getTrue(ITy);
2204 case ICmpInst::ICMP_SGE:
2205 /// For signed comparison, the values for an i1 are 0 and -1
2206 /// respectively. This maps into a truth table of:
2207 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2208 /// 0 | 0 | 1 (0 >= 0) | 1
2209 /// 0 | 1 | 1 (0 >= -1) | 1
2210 /// 1 | 0 | 0 (-1 >= 0) | 0
2211 /// 1 | 1 | 1 (-1 >= -1) | 1
2212 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2213 return getTrue(ITy);
2215 case ICmpInst::ICMP_ULE:
2216 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2217 return getTrue(ITy);
2224 /// Try hard to fold icmp with zero RHS because this is a common case.
2225 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2226 Value *RHS, const SimplifyQuery &Q) {
2227 if (!match(RHS, m_Zero()))
2230 Type *ITy = GetCompareTy(LHS); // The return type.
2233 llvm_unreachable("Unknown ICmp predicate!");
2234 case ICmpInst::ICMP_ULT:
2235 return getFalse(ITy);
2236 case ICmpInst::ICMP_UGE:
2237 return getTrue(ITy);
2238 case ICmpInst::ICMP_EQ:
2239 case ICmpInst::ICMP_ULE:
2240 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2241 return getFalse(ITy);
2243 case ICmpInst::ICMP_NE:
2244 case ICmpInst::ICMP_UGT:
2245 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2246 return getTrue(ITy);
2248 case ICmpInst::ICMP_SLT: {
2249 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2250 if (LHSKnown.isNegative())
2251 return getTrue(ITy);
2252 if (LHSKnown.isNonNegative())
2253 return getFalse(ITy);
2256 case ICmpInst::ICMP_SLE: {
2257 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2258 if (LHSKnown.isNegative())
2259 return getTrue(ITy);
2260 if (LHSKnown.isNonNegative() &&
2261 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2262 return getFalse(ITy);
2265 case ICmpInst::ICMP_SGE: {
2266 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2267 if (LHSKnown.isNegative())
2268 return getFalse(ITy);
2269 if (LHSKnown.isNonNegative())
2270 return getTrue(ITy);
2273 case ICmpInst::ICMP_SGT: {
2274 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2275 if (LHSKnown.isNegative())
2276 return getFalse(ITy);
2277 if (LHSKnown.isNonNegative() &&
2278 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2279 return getTrue(ITy);
2287 /// Many binary operators with a constant operand have an easy-to-compute
2288 /// range of outputs. This can be used to fold a comparison to always true or
2290 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2291 unsigned Width = Lower.getBitWidth();
2293 switch (BO.getOpcode()) {
2294 case Instruction::Add:
2295 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2296 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2297 if (BO.hasNoUnsignedWrap()) {
2298 // 'add nuw x, C' produces [C, UINT_MAX].
2300 } else if (BO.hasNoSignedWrap()) {
2301 if (C->isNegative()) {
2302 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2303 Lower = APInt::getSignedMinValue(Width);
2304 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2306 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2307 Lower = APInt::getSignedMinValue(Width) + *C;
2308 Upper = APInt::getSignedMaxValue(Width) + 1;
2314 case Instruction::And:
2315 if (match(BO.getOperand(1), m_APInt(C)))
2316 // 'and x, C' produces [0, C].
2320 case Instruction::Or:
2321 if (match(BO.getOperand(1), m_APInt(C)))
2322 // 'or x, C' produces [C, UINT_MAX].
2326 case Instruction::AShr:
2327 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2328 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2329 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2330 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2331 } else if (match(BO.getOperand(0), m_APInt(C))) {
2332 unsigned ShiftAmount = Width - 1;
2333 if (!C->isNullValue() && BO.isExact())
2334 ShiftAmount = C->countTrailingZeros();
2335 if (C->isNegative()) {
2336 // 'ashr C, x' produces [C, C >> (Width-1)]
2338 Upper = C->ashr(ShiftAmount) + 1;
2340 // 'ashr C, x' produces [C >> (Width-1), C]
2341 Lower = C->ashr(ShiftAmount);
2347 case Instruction::LShr:
2348 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2349 // 'lshr x, C' produces [0, UINT_MAX >> C].
2350 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2351 } else if (match(BO.getOperand(0), m_APInt(C))) {
2352 // 'lshr C, x' produces [C >> (Width-1), C].
2353 unsigned ShiftAmount = Width - 1;
2354 if (!C->isNullValue() && BO.isExact())
2355 ShiftAmount = C->countTrailingZeros();
2356 Lower = C->lshr(ShiftAmount);
2361 case Instruction::Shl:
2362 if (match(BO.getOperand(0), m_APInt(C))) {
2363 if (BO.hasNoUnsignedWrap()) {
2364 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2366 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2367 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2368 if (C->isNegative()) {
2369 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2370 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2371 Lower = C->shl(ShiftAmount);
2374 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2375 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2377 Upper = C->shl(ShiftAmount) + 1;
2383 case Instruction::SDiv:
2384 if (match(BO.getOperand(1), m_APInt(C))) {
2385 APInt IntMin = APInt::getSignedMinValue(Width);
2386 APInt IntMax = APInt::getSignedMaxValue(Width);
2387 if (C->isAllOnesValue()) {
2388 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2389 // where C != -1 and C != 0 and C != 1
2392 } else if (C->countLeadingZeros() < Width - 1) {
2393 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2394 // where C != -1 and C != 0 and C != 1
2395 Lower = IntMin.sdiv(*C);
2396 Upper = IntMax.sdiv(*C);
2397 if (Lower.sgt(Upper))
2398 std::swap(Lower, Upper);
2400 assert(Upper != Lower && "Upper part of range has wrapped!");
2402 } else if (match(BO.getOperand(0), m_APInt(C))) {
2403 if (C->isMinSignedValue()) {
2404 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2406 Upper = Lower.lshr(1) + 1;
2408 // 'sdiv C, x' produces [-|C|, |C|].
2409 Upper = C->abs() + 1;
2410 Lower = (-Upper) + 1;
2415 case Instruction::UDiv:
2416 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2417 // 'udiv x, C' produces [0, UINT_MAX / C].
2418 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2419 } else if (match(BO.getOperand(0), m_APInt(C))) {
2420 // 'udiv C, x' produces [0, C].
2425 case Instruction::SRem:
2426 if (match(BO.getOperand(1), m_APInt(C))) {
2427 // 'srem x, C' produces (-|C|, |C|).
2429 Lower = (-Upper) + 1;
2433 case Instruction::URem:
2434 if (match(BO.getOperand(1), m_APInt(C)))
2435 // 'urem x, C' produces [0, C).
2444 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2447 if (!match(RHS, m_APInt(C)))
2450 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2451 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2452 if (RHS_CR.isEmptySet())
2453 return ConstantInt::getFalse(GetCompareTy(RHS));
2454 if (RHS_CR.isFullSet())
2455 return ConstantInt::getTrue(GetCompareTy(RHS));
2457 // Find the range of possible values for binary operators.
2458 unsigned Width = C->getBitWidth();
2459 APInt Lower = APInt(Width, 0);
2460 APInt Upper = APInt(Width, 0);
2461 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2462 setLimitsForBinOp(*BO, Lower, Upper);
2464 ConstantRange LHS_CR =
2465 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2467 if (auto *I = dyn_cast<Instruction>(LHS))
2468 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2469 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2471 if (!LHS_CR.isFullSet()) {
2472 if (RHS_CR.contains(LHS_CR))
2473 return ConstantInt::getTrue(GetCompareTy(RHS));
2474 if (RHS_CR.inverse().contains(LHS_CR))
2475 return ConstantInt::getFalse(GetCompareTy(RHS));
2481 /// TODO: A large part of this logic is duplicated in InstCombine's
2482 /// foldICmpBinOp(). We should be able to share that and avoid the code
2484 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2485 Value *RHS, const SimplifyQuery &Q,
2486 unsigned MaxRecurse) {
2487 Type *ITy = GetCompareTy(LHS); // The return type.
2489 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2490 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2491 if (MaxRecurse && (LBO || RBO)) {
2492 // Analyze the case when either LHS or RHS is an add instruction.
2493 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2494 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2495 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2496 if (LBO && LBO->getOpcode() == Instruction::Add) {
2497 A = LBO->getOperand(0);
2498 B = LBO->getOperand(1);
2500 ICmpInst::isEquality(Pred) ||
2501 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2502 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2504 if (RBO && RBO->getOpcode() == Instruction::Add) {
2505 C = RBO->getOperand(0);
2506 D = RBO->getOperand(1);
2508 ICmpInst::isEquality(Pred) ||
2509 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2510 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2513 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2514 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2515 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2516 Constant::getNullValue(RHS->getType()), Q,
2520 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2521 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2523 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2524 C == LHS ? D : C, Q, MaxRecurse - 1))
2527 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2528 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2530 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2533 // C + B == C + D -> B == D
2536 } else if (A == D) {
2537 // D + B == C + D -> B == C
2540 } else if (B == C) {
2541 // A + C == C + D -> A == D
2546 // A + D == C + D -> A == C
2550 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2557 // icmp pred (or X, Y), X
2558 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2559 if (Pred == ICmpInst::ICMP_ULT)
2560 return getFalse(ITy);
2561 if (Pred == ICmpInst::ICMP_UGE)
2562 return getTrue(ITy);
2564 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2565 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2566 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2567 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2568 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2569 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2570 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2573 // icmp pred X, (or X, Y)
2574 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2575 if (Pred == ICmpInst::ICMP_ULE)
2576 return getTrue(ITy);
2577 if (Pred == ICmpInst::ICMP_UGT)
2578 return getFalse(ITy);
2580 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2581 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2582 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2583 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2584 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2585 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2586 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2591 // icmp pred (and X, Y), X
2592 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2593 if (Pred == ICmpInst::ICMP_UGT)
2594 return getFalse(ITy);
2595 if (Pred == ICmpInst::ICMP_ULE)
2596 return getTrue(ITy);
2598 // icmp pred X, (and X, Y)
2599 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2600 if (Pred == ICmpInst::ICMP_UGE)
2601 return getTrue(ITy);
2602 if (Pred == ICmpInst::ICMP_ULT)
2603 return getFalse(ITy);
2606 // 0 - (zext X) pred C
2607 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2608 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2609 if (RHSC->getValue().isStrictlyPositive()) {
2610 if (Pred == ICmpInst::ICMP_SLT)
2611 return ConstantInt::getTrue(RHSC->getContext());
2612 if (Pred == ICmpInst::ICMP_SGE)
2613 return ConstantInt::getFalse(RHSC->getContext());
2614 if (Pred == ICmpInst::ICMP_EQ)
2615 return ConstantInt::getFalse(RHSC->getContext());
2616 if (Pred == ICmpInst::ICMP_NE)
2617 return ConstantInt::getTrue(RHSC->getContext());
2619 if (RHSC->getValue().isNonNegative()) {
2620 if (Pred == ICmpInst::ICMP_SLE)
2621 return ConstantInt::getTrue(RHSC->getContext());
2622 if (Pred == ICmpInst::ICMP_SGT)
2623 return ConstantInt::getFalse(RHSC->getContext());
2628 // icmp pred (urem X, Y), Y
2629 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2633 case ICmpInst::ICMP_SGT:
2634 case ICmpInst::ICMP_SGE: {
2635 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2636 if (!Known.isNonNegative())
2640 case ICmpInst::ICMP_EQ:
2641 case ICmpInst::ICMP_UGT:
2642 case ICmpInst::ICMP_UGE:
2643 return getFalse(ITy);
2644 case ICmpInst::ICMP_SLT:
2645 case ICmpInst::ICMP_SLE: {
2646 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2647 if (!Known.isNonNegative())
2651 case ICmpInst::ICMP_NE:
2652 case ICmpInst::ICMP_ULT:
2653 case ICmpInst::ICMP_ULE:
2654 return getTrue(ITy);
2658 // icmp pred X, (urem Y, X)
2659 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2663 case ICmpInst::ICMP_SGT:
2664 case ICmpInst::ICMP_SGE: {
2665 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2666 if (!Known.isNonNegative())
2670 case ICmpInst::ICMP_NE:
2671 case ICmpInst::ICMP_UGT:
2672 case ICmpInst::ICMP_UGE:
2673 return getTrue(ITy);
2674 case ICmpInst::ICMP_SLT:
2675 case ICmpInst::ICMP_SLE: {
2676 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2677 if (!Known.isNonNegative())
2681 case ICmpInst::ICMP_EQ:
2682 case ICmpInst::ICMP_ULT:
2683 case ICmpInst::ICMP_ULE:
2684 return getFalse(ITy);
2690 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2691 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2692 // icmp pred (X op Y), X
2693 if (Pred == ICmpInst::ICMP_UGT)
2694 return getFalse(ITy);
2695 if (Pred == ICmpInst::ICMP_ULE)
2696 return getTrue(ITy);
2701 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2702 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2703 // icmp pred X, (X op Y)
2704 if (Pred == ICmpInst::ICMP_ULT)
2705 return getFalse(ITy);
2706 if (Pred == ICmpInst::ICMP_UGE)
2707 return getTrue(ITy);
2714 // where CI2 is a power of 2 and CI isn't
2715 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2716 const APInt *CI2Val, *CIVal = &CI->getValue();
2717 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2718 CI2Val->isPowerOf2()) {
2719 if (!CIVal->isPowerOf2()) {
2720 // CI2 << X can equal zero in some circumstances,
2721 // this simplification is unsafe if CI is zero.
2723 // We know it is safe if:
2724 // - The shift is nsw, we can't shift out the one bit.
2725 // - The shift is nuw, we can't shift out the one bit.
2728 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2729 CI2Val->isOneValue() || !CI->isZero()) {
2730 if (Pred == ICmpInst::ICMP_EQ)
2731 return ConstantInt::getFalse(RHS->getContext());
2732 if (Pred == ICmpInst::ICMP_NE)
2733 return ConstantInt::getTrue(RHS->getContext());
2736 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2737 if (Pred == ICmpInst::ICMP_UGT)
2738 return ConstantInt::getFalse(RHS->getContext());
2739 if (Pred == ICmpInst::ICMP_ULE)
2740 return ConstantInt::getTrue(RHS->getContext());
2745 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2746 LBO->getOperand(1) == RBO->getOperand(1)) {
2747 switch (LBO->getOpcode()) {
2750 case Instruction::UDiv:
2751 case Instruction::LShr:
2752 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2754 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2755 RBO->getOperand(0), Q, MaxRecurse - 1))
2758 case Instruction::SDiv:
2759 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2761 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2762 RBO->getOperand(0), Q, MaxRecurse - 1))
2765 case Instruction::AShr:
2766 if (!LBO->isExact() || !RBO->isExact())
2768 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2769 RBO->getOperand(0), Q, MaxRecurse - 1))
2772 case Instruction::Shl: {
2773 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2774 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2777 if (!NSW && ICmpInst::isSigned(Pred))
2779 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2780 RBO->getOperand(0), Q, MaxRecurse - 1))
2789 /// Simplify integer comparisons where at least one operand of the compare
2790 /// matches an integer min/max idiom.
2791 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2792 Value *RHS, const SimplifyQuery &Q,
2793 unsigned MaxRecurse) {
2794 Type *ITy = GetCompareTy(LHS); // The return type.
2796 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2797 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2799 // Signed variants on "max(a,b)>=a -> true".
2800 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2802 std::swap(A, B); // smax(A, B) pred A.
2803 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2804 // We analyze this as smax(A, B) pred A.
2806 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2807 (A == LHS || B == LHS)) {
2809 std::swap(A, B); // A pred smax(A, B).
2810 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2811 // We analyze this as smax(A, B) swapped-pred A.
2812 P = CmpInst::getSwappedPredicate(Pred);
2813 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2814 (A == RHS || B == RHS)) {
2816 std::swap(A, B); // smin(A, B) pred A.
2817 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2818 // We analyze this as smax(-A, -B) swapped-pred -A.
2819 // Note that we do not need to actually form -A or -B thanks to EqP.
2820 P = CmpInst::getSwappedPredicate(Pred);
2821 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2822 (A == LHS || B == LHS)) {
2824 std::swap(A, B); // A pred smin(A, B).
2825 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2826 // We analyze this as smax(-A, -B) pred -A.
2827 // Note that we do not need to actually form -A or -B thanks to EqP.
2830 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2831 // Cases correspond to "max(A, B) p A".
2835 case CmpInst::ICMP_EQ:
2836 case CmpInst::ICMP_SLE:
2837 // Equivalent to "A EqP B". This may be the same as the condition tested
2838 // in the max/min; if so, we can just return that.
2839 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2841 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2843 // Otherwise, see if "A EqP B" simplifies.
2845 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2848 case CmpInst::ICMP_NE:
2849 case CmpInst::ICMP_SGT: {
2850 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2851 // Equivalent to "A InvEqP B". This may be the same as the condition
2852 // tested in the max/min; if so, we can just return that.
2853 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2855 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2857 // Otherwise, see if "A InvEqP B" simplifies.
2859 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2863 case CmpInst::ICMP_SGE:
2865 return getTrue(ITy);
2866 case CmpInst::ICMP_SLT:
2868 return getFalse(ITy);
2872 // Unsigned variants on "max(a,b)>=a -> true".
2873 P = CmpInst::BAD_ICMP_PREDICATE;
2874 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2876 std::swap(A, B); // umax(A, B) pred A.
2877 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2878 // We analyze this as umax(A, B) pred A.
2880 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2881 (A == LHS || B == LHS)) {
2883 std::swap(A, B); // A pred umax(A, B).
2884 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2885 // We analyze this as umax(A, B) swapped-pred A.
2886 P = CmpInst::getSwappedPredicate(Pred);
2887 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2888 (A == RHS || B == RHS)) {
2890 std::swap(A, B); // umin(A, B) pred A.
2891 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2892 // We analyze this as umax(-A, -B) swapped-pred -A.
2893 // Note that we do not need to actually form -A or -B thanks to EqP.
2894 P = CmpInst::getSwappedPredicate(Pred);
2895 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2896 (A == LHS || B == LHS)) {
2898 std::swap(A, B); // A pred umin(A, B).
2899 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2900 // We analyze this as umax(-A, -B) pred -A.
2901 // Note that we do not need to actually form -A or -B thanks to EqP.
2904 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2905 // Cases correspond to "max(A, B) p A".
2909 case CmpInst::ICMP_EQ:
2910 case CmpInst::ICMP_ULE:
2911 // Equivalent to "A EqP B". This may be the same as the condition tested
2912 // in the max/min; if so, we can just return that.
2913 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2915 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2917 // Otherwise, see if "A EqP B" simplifies.
2919 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2922 case CmpInst::ICMP_NE:
2923 case CmpInst::ICMP_UGT: {
2924 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2925 // Equivalent to "A InvEqP B". This may be the same as the condition
2926 // tested in the max/min; if so, we can just return that.
2927 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2929 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2931 // Otherwise, see if "A InvEqP B" simplifies.
2933 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2937 case CmpInst::ICMP_UGE:
2939 return getTrue(ITy);
2940 case CmpInst::ICMP_ULT:
2942 return getFalse(ITy);
2946 // Variants on "max(x,y) >= min(x,z)".
2948 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2949 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2950 (A == C || A == D || B == C || B == D)) {
2951 // max(x, ?) pred min(x, ?).
2952 if (Pred == CmpInst::ICMP_SGE)
2954 return getTrue(ITy);
2955 if (Pred == CmpInst::ICMP_SLT)
2957 return getFalse(ITy);
2958 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2959 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2960 (A == C || A == D || B == C || B == D)) {
2961 // min(x, ?) pred max(x, ?).
2962 if (Pred == CmpInst::ICMP_SLE)
2964 return getTrue(ITy);
2965 if (Pred == CmpInst::ICMP_SGT)
2967 return getFalse(ITy);
2968 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2969 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2970 (A == C || A == D || B == C || B == D)) {
2971 // max(x, ?) pred min(x, ?).
2972 if (Pred == CmpInst::ICMP_UGE)
2974 return getTrue(ITy);
2975 if (Pred == CmpInst::ICMP_ULT)
2977 return getFalse(ITy);
2978 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2979 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2980 (A == C || A == D || B == C || B == D)) {
2981 // min(x, ?) pred max(x, ?).
2982 if (Pred == CmpInst::ICMP_ULE)
2984 return getTrue(ITy);
2985 if (Pred == CmpInst::ICMP_UGT)
2987 return getFalse(ITy);
2993 /// Given operands for an ICmpInst, see if we can fold the result.
2994 /// If not, this returns null.
2995 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2996 const SimplifyQuery &Q, unsigned MaxRecurse) {
2997 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2998 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3000 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3001 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3002 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3004 // If we have a constant, make sure it is on the RHS.
3005 std::swap(LHS, RHS);
3006 Pred = CmpInst::getSwappedPredicate(Pred);
3009 Type *ITy = GetCompareTy(LHS); // The return type.
3011 // icmp X, X -> true/false
3012 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3013 // because X could be 0.
3014 if (LHS == RHS || isa<UndefValue>(RHS))
3015 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3017 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3020 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3023 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3026 // If both operands have range metadata, use the metadata
3027 // to simplify the comparison.
3028 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3029 auto RHS_Instr = cast<Instruction>(RHS);
3030 auto LHS_Instr = cast<Instruction>(LHS);
3032 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3033 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3034 auto RHS_CR = getConstantRangeFromMetadata(
3035 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3036 auto LHS_CR = getConstantRangeFromMetadata(
3037 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3039 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3040 if (Satisfied_CR.contains(LHS_CR))
3041 return ConstantInt::getTrue(RHS->getContext());
3043 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3044 CmpInst::getInversePredicate(Pred), RHS_CR);
3045 if (InversedSatisfied_CR.contains(LHS_CR))
3046 return ConstantInt::getFalse(RHS->getContext());
3050 // Compare of cast, for example (zext X) != 0 -> X != 0
3051 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3052 Instruction *LI = cast<CastInst>(LHS);
3053 Value *SrcOp = LI->getOperand(0);
3054 Type *SrcTy = SrcOp->getType();
3055 Type *DstTy = LI->getType();
3057 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3058 // if the integer type is the same size as the pointer type.
3059 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3060 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3061 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3062 // Transfer the cast to the constant.
3063 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3064 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3067 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3068 if (RI->getOperand(0)->getType() == SrcTy)
3069 // Compare without the cast.
3070 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3076 if (isa<ZExtInst>(LHS)) {
3077 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3079 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3080 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3081 // Compare X and Y. Note that signed predicates become unsigned.
3082 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3083 SrcOp, RI->getOperand(0), Q,
3087 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3088 // too. If not, then try to deduce the result of the comparison.
3089 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3090 // Compute the constant that would happen if we truncated to SrcTy then
3091 // reextended to DstTy.
3092 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3093 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3095 // If the re-extended constant didn't change then this is effectively
3096 // also a case of comparing two zero-extended values.
3097 if (RExt == CI && MaxRecurse)
3098 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3099 SrcOp, Trunc, Q, MaxRecurse-1))
3102 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3103 // there. Use this to work out the result of the comparison.
3106 default: llvm_unreachable("Unknown ICmp predicate!");
3108 case ICmpInst::ICMP_EQ:
3109 case ICmpInst::ICMP_UGT:
3110 case ICmpInst::ICMP_UGE:
3111 return ConstantInt::getFalse(CI->getContext());
3113 case ICmpInst::ICMP_NE:
3114 case ICmpInst::ICMP_ULT:
3115 case ICmpInst::ICMP_ULE:
3116 return ConstantInt::getTrue(CI->getContext());
3118 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3119 // is non-negative then LHS <s RHS.
3120 case ICmpInst::ICMP_SGT:
3121 case ICmpInst::ICMP_SGE:
3122 return CI->getValue().isNegative() ?
3123 ConstantInt::getTrue(CI->getContext()) :
3124 ConstantInt::getFalse(CI->getContext());
3126 case ICmpInst::ICMP_SLT:
3127 case ICmpInst::ICMP_SLE:
3128 return CI->getValue().isNegative() ?
3129 ConstantInt::getFalse(CI->getContext()) :
3130 ConstantInt::getTrue(CI->getContext());
3136 if (isa<SExtInst>(LHS)) {
3137 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3139 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3140 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3141 // Compare X and Y. Note that the predicate does not change.
3142 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3146 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3147 // too. If not, then try to deduce the result of the comparison.
3148 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3149 // Compute the constant that would happen if we truncated to SrcTy then
3150 // reextended to DstTy.
3151 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3152 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3154 // If the re-extended constant didn't change then this is effectively
3155 // also a case of comparing two sign-extended values.
3156 if (RExt == CI && MaxRecurse)
3157 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3160 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3161 // bits there. Use this to work out the result of the comparison.
3164 default: llvm_unreachable("Unknown ICmp predicate!");
3165 case ICmpInst::ICMP_EQ:
3166 return ConstantInt::getFalse(CI->getContext());
3167 case ICmpInst::ICMP_NE:
3168 return ConstantInt::getTrue(CI->getContext());
3170 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3172 case ICmpInst::ICMP_SGT:
3173 case ICmpInst::ICMP_SGE:
3174 return CI->getValue().isNegative() ?
3175 ConstantInt::getTrue(CI->getContext()) :
3176 ConstantInt::getFalse(CI->getContext());
3177 case ICmpInst::ICMP_SLT:
3178 case ICmpInst::ICMP_SLE:
3179 return CI->getValue().isNegative() ?
3180 ConstantInt::getFalse(CI->getContext()) :
3181 ConstantInt::getTrue(CI->getContext());
3183 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3185 case ICmpInst::ICMP_UGT:
3186 case ICmpInst::ICMP_UGE:
3187 // Comparison is true iff the LHS <s 0.
3189 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3190 Constant::getNullValue(SrcTy),
3194 case ICmpInst::ICMP_ULT:
3195 case ICmpInst::ICMP_ULE:
3196 // Comparison is true iff the LHS >=s 0.
3198 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3199 Constant::getNullValue(SrcTy),
3209 // icmp eq|ne X, Y -> false|true if X != Y
3210 if (ICmpInst::isEquality(Pred) &&
3211 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3212 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3215 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3218 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3221 // Simplify comparisons of related pointers using a powerful, recursive
3222 // GEP-walk when we have target data available..
3223 if (LHS->getType()->isPointerTy())
3224 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3227 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3228 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3229 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3230 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3231 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3232 Q.DL.getTypeSizeInBits(CRHS->getType()))
3233 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3234 CLHS->getPointerOperand(),
3235 CRHS->getPointerOperand()))
3238 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3239 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3240 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3241 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3242 (ICmpInst::isEquality(Pred) ||
3243 (GLHS->isInBounds() && GRHS->isInBounds() &&
3244 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3245 // The bases are equal and the indices are constant. Build a constant
3246 // expression GEP with the same indices and a null base pointer to see
3247 // what constant folding can make out of it.
3248 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3249 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3250 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3251 GLHS->getSourceElementType(), Null, IndicesLHS);
3253 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3254 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3255 GLHS->getSourceElementType(), Null, IndicesRHS);
3256 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3261 // If the comparison is with the result of a select instruction, check whether
3262 // comparing with either branch of the select always yields the same value.
3263 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3264 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3267 // If the comparison is with the result of a phi instruction, check whether
3268 // doing the compare with each incoming phi value yields a common result.
3269 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3270 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3276 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3277 const SimplifyQuery &Q) {
3278 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3281 /// Given operands for an FCmpInst, see if we can fold the result.
3282 /// If not, this returns null.
3283 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3284 FastMathFlags FMF, const SimplifyQuery &Q,
3285 unsigned MaxRecurse) {
3286 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3287 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3289 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3290 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3291 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3293 // If we have a constant, make sure it is on the RHS.
3294 std::swap(LHS, RHS);
3295 Pred = CmpInst::getSwappedPredicate(Pred);
3298 // Fold trivial predicates.
3299 Type *RetTy = GetCompareTy(LHS);
3300 if (Pred == FCmpInst::FCMP_FALSE)
3301 return getFalse(RetTy);
3302 if (Pred == FCmpInst::FCMP_TRUE)
3303 return getTrue(RetTy);
3305 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3307 if (Pred == FCmpInst::FCMP_UNO)
3308 return getFalse(RetTy);
3309 if (Pred == FCmpInst::FCMP_ORD)
3310 return getTrue(RetTy);
3313 // fcmp pred x, undef and fcmp pred undef, x
3314 // fold to true if unordered, false if ordered
3315 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3316 // Choosing NaN for the undef will always make unordered comparison succeed
3317 // and ordered comparison fail.
3318 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3321 // fcmp x,x -> true/false. Not all compares are foldable.
3323 if (CmpInst::isTrueWhenEqual(Pred))
3324 return getTrue(RetTy);
3325 if (CmpInst::isFalseWhenEqual(Pred))
3326 return getFalse(RetTy);
3329 // Handle fcmp with constant RHS
3330 const ConstantFP *CFP = nullptr;
3331 if (const auto *RHSC = dyn_cast<Constant>(RHS)) {
3332 if (RHS->getType()->isVectorTy())
3333 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue());
3335 CFP = dyn_cast<ConstantFP>(RHSC);
3338 // If the constant is a nan, see if we can fold the comparison based on it.
3339 if (CFP->getValueAPF().isNaN()) {
3340 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3341 return getFalse(RetTy);
3342 assert(FCmpInst::isUnordered(Pred) &&
3343 "Comparison must be either ordered or unordered!");
3344 // True if unordered.
3345 return getTrue(RetTy);
3347 // Check whether the constant is an infinity.
3348 if (CFP->getValueAPF().isInfinity()) {
3349 if (CFP->getValueAPF().isNegative()) {
3351 case FCmpInst::FCMP_OLT:
3352 // No value is ordered and less than negative infinity.
3353 return getFalse(RetTy);
3354 case FCmpInst::FCMP_UGE:
3355 // All values are unordered with or at least negative infinity.
3356 return getTrue(RetTy);
3362 case FCmpInst::FCMP_OGT:
3363 // No value is ordered and greater than infinity.
3364 return getFalse(RetTy);
3365 case FCmpInst::FCMP_ULE:
3366 // All values are unordered with and at most infinity.
3367 return getTrue(RetTy);
3373 if (CFP->getValueAPF().isZero()) {
3375 case FCmpInst::FCMP_UGE:
3376 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3377 return getTrue(RetTy);
3379 case FCmpInst::FCMP_OLT:
3381 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3382 return getFalse(RetTy);
3390 // If the comparison is with the result of a select instruction, check whether
3391 // comparing with either branch of the select always yields the same value.
3392 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3393 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3396 // If the comparison is with the result of a phi instruction, check whether
3397 // doing the compare with each incoming phi value yields a common result.
3398 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3399 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3405 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3406 FastMathFlags FMF, const SimplifyQuery &Q) {
3407 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3410 /// See if V simplifies when its operand Op is replaced with RepOp.
3411 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3412 const SimplifyQuery &Q,
3413 unsigned MaxRecurse) {
3414 // Trivial replacement.
3418 // We cannot replace a constant, and shouldn't even try.
3419 if (isa<Constant>(Op))
3422 auto *I = dyn_cast<Instruction>(V);
3426 // If this is a binary operator, try to simplify it with the replaced op.
3427 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3429 // %cmp = icmp eq i32 %x, 2147483647
3430 // %add = add nsw i32 %x, 1
3431 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3433 // We can't replace %sel with %add unless we strip away the flags.
3434 if (isa<OverflowingBinaryOperator>(B))
3435 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3437 if (isa<PossiblyExactOperator>(B))
3442 if (B->getOperand(0) == Op)
3443 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3445 if (B->getOperand(1) == Op)
3446 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3451 // Same for CmpInsts.
3452 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3454 if (C->getOperand(0) == Op)
3455 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3457 if (C->getOperand(1) == Op)
3458 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3463 // TODO: We could hand off more cases to instsimplify here.
3465 // If all operands are constant after substituting Op for RepOp then we can
3466 // constant fold the instruction.
3467 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3468 // Build a list of all constant operands.
3469 SmallVector<Constant *, 8> ConstOps;
3470 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3471 if (I->getOperand(i) == Op)
3472 ConstOps.push_back(CRepOp);
3473 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3474 ConstOps.push_back(COp);
3479 // All operands were constants, fold it.
3480 if (ConstOps.size() == I->getNumOperands()) {
3481 if (CmpInst *C = dyn_cast<CmpInst>(I))
3482 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3483 ConstOps[1], Q.DL, Q.TLI);
3485 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3486 if (!LI->isVolatile())
3487 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3489 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3496 /// Try to simplify a select instruction when its condition operand is an
3497 /// integer comparison where one operand of the compare is a constant.
3498 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3499 const APInt *Y, bool TrueWhenUnset) {
3502 // (X & Y) == 0 ? X & ~Y : X --> X
3503 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3504 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3506 return TrueWhenUnset ? FalseVal : TrueVal;
3508 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3509 // (X & Y) != 0 ? X : X & ~Y --> X
3510 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3512 return TrueWhenUnset ? FalseVal : TrueVal;
3514 if (Y->isPowerOf2()) {
3515 // (X & Y) == 0 ? X | Y : X --> X | Y
3516 // (X & Y) != 0 ? X | Y : X --> X
3517 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3519 return TrueWhenUnset ? TrueVal : FalseVal;
3521 // (X & Y) == 0 ? X : X | Y --> X
3522 // (X & Y) != 0 ? X : X | Y --> X | Y
3523 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3525 return TrueWhenUnset ? TrueVal : FalseVal;
3531 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3533 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3534 ICmpInst::Predicate Pred,
3535 Value *TrueVal, Value *FalseVal) {
3538 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3541 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3542 Pred == ICmpInst::ICMP_EQ);
3545 /// Try to simplify a select instruction when its condition operand is an
3546 /// integer comparison.
3547 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3548 Value *FalseVal, const SimplifyQuery &Q,
3549 unsigned MaxRecurse) {
3550 ICmpInst::Predicate Pred;
3551 Value *CmpLHS, *CmpRHS;
3552 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3555 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3558 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3559 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3560 Pred == ICmpInst::ICMP_EQ))
3564 // Check for other compares that behave like bit test.
3565 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3569 if (CondVal->hasOneUse()) {
3571 if (match(CmpRHS, m_APInt(C))) {
3572 // X < MIN ? T : F --> F
3573 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3575 // X < MIN ? T : F --> F
3576 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3578 // X > MAX ? T : F --> F
3579 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3581 // X > MAX ? T : F --> F
3582 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3587 // If we have an equality comparison, then we know the value in one of the
3588 // arms of the select. See if substituting this value into the arm and
3589 // simplifying the result yields the same value as the other arm.
3590 if (Pred == ICmpInst::ICMP_EQ) {
3591 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3593 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3596 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3598 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3601 } else if (Pred == ICmpInst::ICMP_NE) {
3602 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3604 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3607 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3609 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3617 /// Given operands for a SelectInst, see if we can fold the result.
3618 /// If not, this returns null.
3619 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3620 Value *FalseVal, const SimplifyQuery &Q,
3621 unsigned MaxRecurse) {
3622 // select true, X, Y -> X
3623 // select false, X, Y -> Y
3624 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3625 if (Constant *CT = dyn_cast<Constant>(TrueVal))
3626 if (Constant *CF = dyn_cast<Constant>(FalseVal))
3627 return ConstantFoldSelectInstruction(CB, CT, CF);
3628 if (CB->isAllOnesValue())
3630 if (CB->isNullValue())
3634 // select C, X, X -> X
3635 if (TrueVal == FalseVal)
3638 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3639 if (isa<Constant>(FalseVal))
3643 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3645 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3649 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3655 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3656 const SimplifyQuery &Q) {
3657 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3660 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3661 /// If not, this returns null.
3662 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3663 const SimplifyQuery &Q, unsigned) {
3664 // The type of the GEP pointer operand.
3666 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3668 // getelementptr P -> P.
3669 if (Ops.size() == 1)
3672 // Compute the (pointer) type returned by the GEP instruction.
3673 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3674 Type *GEPTy = PointerType::get(LastType, AS);
3675 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3676 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3677 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3678 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3680 if (isa<UndefValue>(Ops[0]))
3681 return UndefValue::get(GEPTy);
3683 if (Ops.size() == 2) {
3684 // getelementptr P, 0 -> P.
3685 if (match(Ops[1], m_Zero()))
3689 if (Ty->isSized()) {
3692 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3693 // getelementptr P, N -> P if P points to a type of zero size.
3694 if (TyAllocSize == 0)
3697 // The following transforms are only safe if the ptrtoint cast
3698 // doesn't truncate the pointers.
3699 if (Ops[1]->getType()->getScalarSizeInBits() ==
3700 Q.DL.getPointerSizeInBits(AS)) {
3701 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3702 if (match(P, m_Zero()))
3703 return Constant::getNullValue(GEPTy);
3705 if (match(P, m_PtrToInt(m_Value(Temp))))
3706 if (Temp->getType() == GEPTy)
3711 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3712 if (TyAllocSize == 1 &&
3713 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3714 if (Value *R = PtrToIntOrZero(P))
3717 // getelementptr V, (ashr (sub P, V), C) -> Q
3718 // if P points to a type of size 1 << C.
3720 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3721 m_ConstantInt(C))) &&
3722 TyAllocSize == 1ULL << C)
3723 if (Value *R = PtrToIntOrZero(P))
3726 // getelementptr V, (sdiv (sub P, V), C) -> Q
3727 // if P points to a type of size C.
3729 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3730 m_SpecificInt(TyAllocSize))))
3731 if (Value *R = PtrToIntOrZero(P))
3737 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3738 all_of(Ops.slice(1).drop_back(1),
3739 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3741 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3742 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3743 APInt BasePtrOffset(PtrWidth, 0);
3744 Value *StrippedBasePtr =
3745 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3748 // gep (gep V, C), (sub 0, V) -> C
3749 if (match(Ops.back(),
3750 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3751 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3752 return ConstantExpr::getIntToPtr(CI, GEPTy);
3754 // gep (gep V, C), (xor V, -1) -> C-1
3755 if (match(Ops.back(),
3756 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3757 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3758 return ConstantExpr::getIntToPtr(CI, GEPTy);
3763 // Check to see if this is constant foldable.
3764 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3767 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3769 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3774 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3775 const SimplifyQuery &Q) {
3776 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3779 /// Given operands for an InsertValueInst, see if we can fold the result.
3780 /// If not, this returns null.
3781 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3782 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3784 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3785 if (Constant *CVal = dyn_cast<Constant>(Val))
3786 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3788 // insertvalue x, undef, n -> x
3789 if (match(Val, m_Undef()))
3792 // insertvalue x, (extractvalue y, n), n
3793 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3794 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3795 EV->getIndices() == Idxs) {
3796 // insertvalue undef, (extractvalue y, n), n -> y
3797 if (match(Agg, m_Undef()))
3798 return EV->getAggregateOperand();
3800 // insertvalue y, (extractvalue y, n), n -> y
3801 if (Agg == EV->getAggregateOperand())
3808 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3809 ArrayRef<unsigned> Idxs,
3810 const SimplifyQuery &Q) {
3811 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3814 /// Given operands for an ExtractValueInst, see if we can fold the result.
3815 /// If not, this returns null.
3816 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3817 const SimplifyQuery &, unsigned) {
3818 if (auto *CAgg = dyn_cast<Constant>(Agg))
3819 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3821 // extractvalue x, (insertvalue y, elt, n), n -> elt
3822 unsigned NumIdxs = Idxs.size();
3823 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3824 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3825 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3826 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3827 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3828 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3829 Idxs.slice(0, NumCommonIdxs)) {
3830 if (NumIdxs == NumInsertValueIdxs)
3831 return IVI->getInsertedValueOperand();
3839 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3840 const SimplifyQuery &Q) {
3841 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3844 /// Given operands for an ExtractElementInst, see if we can fold the result.
3845 /// If not, this returns null.
3846 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3848 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3849 if (auto *CIdx = dyn_cast<Constant>(Idx))
3850 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3852 // The index is not relevant if our vector is a splat.
3853 if (auto *Splat = CVec->getSplatValue())
3856 if (isa<UndefValue>(Vec))
3857 return UndefValue::get(Vec->getType()->getVectorElementType());
3860 // If extracting a specified index from the vector, see if we can recursively
3861 // find a previously computed scalar that was inserted into the vector.
3862 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3863 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3869 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3870 const SimplifyQuery &Q) {
3871 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3874 /// See if we can fold the given phi. If not, returns null.
3875 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3876 // If all of the PHI's incoming values are the same then replace the PHI node
3877 // with the common value.
3878 Value *CommonValue = nullptr;
3879 bool HasUndefInput = false;
3880 for (Value *Incoming : PN->incoming_values()) {
3881 // If the incoming value is the phi node itself, it can safely be skipped.
3882 if (Incoming == PN) continue;
3883 if (isa<UndefValue>(Incoming)) {
3884 // Remember that we saw an undef value, but otherwise ignore them.
3885 HasUndefInput = true;
3888 if (CommonValue && Incoming != CommonValue)
3889 return nullptr; // Not the same, bail out.
3890 CommonValue = Incoming;
3893 // If CommonValue is null then all of the incoming values were either undef or
3894 // equal to the phi node itself.
3896 return UndefValue::get(PN->getType());
3898 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3899 // instruction, we cannot return X as the result of the PHI node unless it
3900 // dominates the PHI block.
3902 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3907 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
3908 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
3909 if (auto *C = dyn_cast<Constant>(Op))
3910 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
3912 if (auto *CI = dyn_cast<CastInst>(Op)) {
3913 auto *Src = CI->getOperand(0);
3914 Type *SrcTy = Src->getType();
3915 Type *MidTy = CI->getType();
3917 if (Src->getType() == Ty) {
3918 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
3919 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
3921 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
3923 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
3925 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
3926 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
3927 SrcIntPtrTy, MidIntPtrTy,
3928 DstIntPtrTy) == Instruction::BitCast)
3934 if (CastOpc == Instruction::BitCast)
3935 if (Op->getType() == Ty)
3941 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
3942 const SimplifyQuery &Q) {
3943 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
3946 /// For the given destination element of a shuffle, peek through shuffles to
3947 /// match a root vector source operand that contains that element in the same
3948 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
3949 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
3950 int MaskVal, Value *RootVec,
3951 unsigned MaxRecurse) {
3955 // Bail out if any mask value is undefined. That kind of shuffle may be
3956 // simplified further based on demanded bits or other folds.
3960 // The mask value chooses which source operand we need to look at next.
3961 int InVecNumElts = Op0->getType()->getVectorNumElements();
3962 int RootElt = MaskVal;
3963 Value *SourceOp = Op0;
3964 if (MaskVal >= InVecNumElts) {
3965 RootElt = MaskVal - InVecNumElts;
3969 // If the source operand is a shuffle itself, look through it to find the
3970 // matching root vector.
3971 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
3972 return foldIdentityShuffles(
3973 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
3974 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
3977 // TODO: Look through bitcasts? What if the bitcast changes the vector element
3980 // The source operand is not a shuffle. Initialize the root vector value for
3981 // this shuffle if that has not been done yet.
3985 // Give up as soon as a source operand does not match the existing root value.
3986 if (RootVec != SourceOp)
3989 // The element must be coming from the same lane in the source vector
3990 // (although it may have crossed lanes in intermediate shuffles).
3991 if (RootElt != DestElt)
3997 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
3998 Type *RetTy, const SimplifyQuery &Q,
3999 unsigned MaxRecurse) {
4000 if (isa<UndefValue>(Mask))
4001 return UndefValue::get(RetTy);
4003 Type *InVecTy = Op0->getType();
4004 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4005 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4007 SmallVector<int, 32> Indices;
4008 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4009 assert(MaskNumElts == Indices.size() &&
4010 "Size of Indices not same as number of mask elements?");
4012 // Canonicalization: If mask does not select elements from an input vector,
4013 // replace that input vector with undef.
4014 bool MaskSelects0 = false, MaskSelects1 = false;
4015 for (unsigned i = 0; i != MaskNumElts; ++i) {
4016 if (Indices[i] == -1)
4018 if ((unsigned)Indices[i] < InVecNumElts)
4019 MaskSelects0 = true;
4021 MaskSelects1 = true;
4024 Op0 = UndefValue::get(InVecTy);
4026 Op1 = UndefValue::get(InVecTy);
4028 auto *Op0Const = dyn_cast<Constant>(Op0);
4029 auto *Op1Const = dyn_cast<Constant>(Op1);
4031 // If all operands are constant, constant fold the shuffle.
4032 if (Op0Const && Op1Const)
4033 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4035 // Canonicalization: if only one input vector is constant, it shall be the
4037 if (Op0Const && !Op1Const) {
4038 std::swap(Op0, Op1);
4039 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4042 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4043 // value type is same as the input vectors' type.
4044 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4045 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4046 OpShuf->getMask()->getSplatValue())
4049 // Don't fold a shuffle with undef mask elements. This may get folded in a
4050 // better way using demanded bits or other analysis.
4051 // TODO: Should we allow this?
4052 if (find(Indices, -1) != Indices.end())
4055 // Check if every element of this shuffle can be mapped back to the
4056 // corresponding element of a single root vector. If so, we don't need this
4057 // shuffle. This handles simple identity shuffles as well as chains of
4058 // shuffles that may widen/narrow and/or move elements across lanes and back.
4059 Value *RootVec = nullptr;
4060 for (unsigned i = 0; i != MaskNumElts; ++i) {
4061 // Note that recursion is limited for each vector element, so if any element
4062 // exceeds the limit, this will fail to simplify.
4064 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4066 // We can't replace a widening/narrowing shuffle with one of its operands.
4067 if (!RootVec || RootVec->getType() != RetTy)
4073 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4074 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4075 Type *RetTy, const SimplifyQuery &Q) {
4076 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4079 /// Given operands for an FAdd, see if we can fold the result. If not, this
4081 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4082 const SimplifyQuery &Q, unsigned MaxRecurse) {
4083 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4087 if (match(Op1, m_NegZero()))
4090 // fadd X, 0 ==> X, when we know X is not -0
4091 if (match(Op1, m_Zero()) &&
4092 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4095 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
4096 // where nnan and ninf have to occur at least once somewhere in this
4098 Value *SubOp = nullptr;
4099 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
4101 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
4104 Instruction *FSub = cast<Instruction>(SubOp);
4105 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
4106 (FMF.noInfs() || FSub->hasNoInfs()))
4107 return Constant::getNullValue(Op0->getType());
4113 /// Given operands for an FSub, see if we can fold the result. If not, this
4115 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4116 const SimplifyQuery &Q, unsigned MaxRecurse) {
4117 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4121 if (match(Op1, m_Zero()))
4124 // fsub X, -0 ==> X, when we know X is not -0
4125 if (match(Op1, m_NegZero()) &&
4126 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4129 // fsub -0.0, (fsub -0.0, X) ==> X
4131 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
4134 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4135 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
4136 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
4139 // fsub nnan x, x ==> 0.0
4140 if (FMF.noNaNs() && Op0 == Op1)
4141 return Constant::getNullValue(Op0->getType());
4146 /// Given the operands for an FMul, see if we can fold the result
4147 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4148 const SimplifyQuery &Q, unsigned MaxRecurse) {
4149 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4152 // fmul X, 1.0 ==> X
4153 if (match(Op1, m_FPOne()))
4156 // fmul nnan nsz X, 0 ==> 0
4157 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
4163 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4164 const SimplifyQuery &Q) {
4165 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4169 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4170 const SimplifyQuery &Q) {
4171 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4174 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4175 const SimplifyQuery &Q) {
4176 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4179 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4180 const SimplifyQuery &Q, unsigned) {
4181 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4184 // undef / X -> undef (the undef could be a snan).
4185 if (match(Op0, m_Undef()))
4188 // X / undef -> undef
4189 if (match(Op1, m_Undef()))
4193 if (match(Op1, m_FPOne()))
4197 // Requires that NaNs are off (X could be zero) and signed zeroes are
4198 // ignored (X could be positive or negative, so the output sign is unknown).
4199 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4203 // X / X -> 1.0 is legal when NaNs are ignored.
4205 return ConstantFP::get(Op0->getType(), 1.0);
4207 // -X / X -> -1.0 and
4208 // X / -X -> -1.0 are legal when NaNs are ignored.
4209 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4210 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4211 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4212 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4213 BinaryOperator::getFNegArgument(Op1) == Op0))
4214 return ConstantFP::get(Op0->getType(), -1.0);
4220 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4221 const SimplifyQuery &Q) {
4222 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4225 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4226 const SimplifyQuery &Q, unsigned) {
4227 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4230 // undef % X -> undef (the undef could be a snan).
4231 if (match(Op0, m_Undef()))
4234 // X % undef -> undef
4235 if (match(Op1, m_Undef()))
4239 // Requires that NaNs are off (X could be zero) and signed zeroes are
4240 // ignored (X could be positive or negative, so the output sign is unknown).
4241 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4247 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4248 const SimplifyQuery &Q) {
4249 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4252 //=== Helper functions for higher up the class hierarchy.
4254 /// Given operands for a BinaryOperator, see if we can fold the result.
4255 /// If not, this returns null.
4256 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4257 const SimplifyQuery &Q, unsigned MaxRecurse) {
4259 case Instruction::Add:
4260 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4261 case Instruction::Sub:
4262 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4263 case Instruction::Mul:
4264 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4265 case Instruction::SDiv:
4266 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4267 case Instruction::UDiv:
4268 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4269 case Instruction::SRem:
4270 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4271 case Instruction::URem:
4272 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4273 case Instruction::Shl:
4274 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4275 case Instruction::LShr:
4276 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4277 case Instruction::AShr:
4278 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4279 case Instruction::And:
4280 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4281 case Instruction::Or:
4282 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4283 case Instruction::Xor:
4284 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4285 case Instruction::FAdd:
4286 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4287 case Instruction::FSub:
4288 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4289 case Instruction::FMul:
4290 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4291 case Instruction::FDiv:
4292 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4293 case Instruction::FRem:
4294 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4296 llvm_unreachable("Unexpected opcode");
4300 /// Given operands for a BinaryOperator, see if we can fold the result.
4301 /// If not, this returns null.
4302 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4303 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4304 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4305 const FastMathFlags &FMF, const SimplifyQuery &Q,
4306 unsigned MaxRecurse) {
4308 case Instruction::FAdd:
4309 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4310 case Instruction::FSub:
4311 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4312 case Instruction::FMul:
4313 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4314 case Instruction::FDiv:
4315 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4317 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4321 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4322 const SimplifyQuery &Q) {
4323 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4326 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4327 FastMathFlags FMF, const SimplifyQuery &Q) {
4328 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4331 /// Given operands for a CmpInst, see if we can fold the result.
4332 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4333 const SimplifyQuery &Q, unsigned MaxRecurse) {
4334 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4335 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4336 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4339 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4340 const SimplifyQuery &Q) {
4341 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4344 static bool IsIdempotent(Intrinsic::ID ID) {
4346 default: return false;
4348 // Unary idempotent: f(f(x)) = f(x)
4349 case Intrinsic::fabs:
4350 case Intrinsic::floor:
4351 case Intrinsic::ceil:
4352 case Intrinsic::trunc:
4353 case Intrinsic::rint:
4354 case Intrinsic::nearbyint:
4355 case Intrinsic::round:
4356 case Intrinsic::canonicalize:
4361 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4362 const DataLayout &DL) {
4363 GlobalValue *PtrSym;
4365 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4368 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4369 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4370 Type *Int32PtrTy = Int32Ty->getPointerTo();
4371 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4373 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4374 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4377 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4378 if (OffsetInt % 4 != 0)
4381 Constant *C = ConstantExpr::getGetElementPtr(
4382 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4383 ConstantInt::get(Int64Ty, OffsetInt / 4));
4384 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4388 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4392 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4393 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4398 if (LoadedCE->getOpcode() != Instruction::Sub)
4401 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4402 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4404 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4406 Constant *LoadedRHS = LoadedCE->getOperand(1);
4407 GlobalValue *LoadedRHSSym;
4408 APInt LoadedRHSOffset;
4409 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4411 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4414 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4417 static bool maskIsAllZeroOrUndef(Value *Mask) {
4418 auto *ConstMask = dyn_cast<Constant>(Mask);
4421 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4423 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4425 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4426 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4433 template <typename IterTy>
4434 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4435 const SimplifyQuery &Q, unsigned MaxRecurse) {
4436 Intrinsic::ID IID = F->getIntrinsicID();
4437 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4440 if (NumOperands == 1) {
4441 // Perform idempotent optimizations
4442 if (IsIdempotent(IID)) {
4443 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4444 if (II->getIntrinsicID() == IID)
4450 case Intrinsic::fabs: {
4451 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4461 if (NumOperands == 2) {
4462 Value *LHS = *ArgBegin;
4463 Value *RHS = *(ArgBegin + 1);
4464 Type *ReturnType = F->getReturnType();
4467 case Intrinsic::usub_with_overflow:
4468 case Intrinsic::ssub_with_overflow: {
4469 // X - X -> { 0, false }
4471 return Constant::getNullValue(ReturnType);
4473 // X - undef -> undef
4474 // undef - X -> undef
4475 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4476 return UndefValue::get(ReturnType);
4480 case Intrinsic::uadd_with_overflow:
4481 case Intrinsic::sadd_with_overflow: {
4482 // X + undef -> undef
4483 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4484 return UndefValue::get(ReturnType);
4488 case Intrinsic::umul_with_overflow:
4489 case Intrinsic::smul_with_overflow: {
4490 // 0 * X -> { 0, false }
4491 // X * 0 -> { 0, false }
4492 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4493 return Constant::getNullValue(ReturnType);
4495 // undef * X -> { 0, false }
4496 // X * undef -> { 0, false }
4497 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4498 return Constant::getNullValue(ReturnType);
4502 case Intrinsic::load_relative: {
4503 Constant *C0 = dyn_cast<Constant>(LHS);
4504 Constant *C1 = dyn_cast<Constant>(RHS);
4506 return SimplifyRelativeLoad(C0, C1, Q.DL);
4514 // Simplify calls to llvm.masked.load.*
4516 case Intrinsic::masked_load: {
4517 Value *MaskArg = ArgBegin[2];
4518 Value *PassthruArg = ArgBegin[3];
4519 // If the mask is all zeros or undef, the "passthru" argument is the result.
4520 if (maskIsAllZeroOrUndef(MaskArg))
4529 template <typename IterTy>
4530 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4531 IterTy ArgEnd, const SimplifyQuery &Q,
4532 unsigned MaxRecurse) {
4533 Type *Ty = V->getType();
4534 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4535 Ty = PTy->getElementType();
4536 FunctionType *FTy = cast<FunctionType>(Ty);
4538 // call undef -> undef
4539 // call null -> undef
4540 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4541 return UndefValue::get(FTy->getReturnType());
4543 Function *F = dyn_cast<Function>(V);
4547 if (F->isIntrinsic())
4548 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4551 if (!canConstantFoldCallTo(CS, F))
4554 SmallVector<Constant *, 4> ConstantArgs;
4555 ConstantArgs.reserve(ArgEnd - ArgBegin);
4556 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4557 Constant *C = dyn_cast<Constant>(*I);
4560 ConstantArgs.push_back(C);
4563 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4566 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4567 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4568 const SimplifyQuery &Q) {
4569 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4572 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4573 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4574 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4577 /// See if we can compute a simplified version of this instruction.
4578 /// If not, this returns null.
4580 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4581 OptimizationRemarkEmitter *ORE) {
4582 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4585 switch (I->getOpcode()) {
4587 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4589 case Instruction::FAdd:
4590 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4591 I->getFastMathFlags(), Q);
4593 case Instruction::Add:
4594 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4595 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4596 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4598 case Instruction::FSub:
4599 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4600 I->getFastMathFlags(), Q);
4602 case Instruction::Sub:
4603 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4604 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4605 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4607 case Instruction::FMul:
4608 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4609 I->getFastMathFlags(), Q);
4611 case Instruction::Mul:
4612 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4614 case Instruction::SDiv:
4615 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4617 case Instruction::UDiv:
4618 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4620 case Instruction::FDiv:
4621 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4622 I->getFastMathFlags(), Q);
4624 case Instruction::SRem:
4625 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4627 case Instruction::URem:
4628 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4630 case Instruction::FRem:
4631 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4632 I->getFastMathFlags(), Q);
4634 case Instruction::Shl:
4635 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4636 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4637 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4639 case Instruction::LShr:
4640 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4641 cast<BinaryOperator>(I)->isExact(), Q);
4643 case Instruction::AShr:
4644 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4645 cast<BinaryOperator>(I)->isExact(), Q);
4647 case Instruction::And:
4648 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4650 case Instruction::Or:
4651 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4653 case Instruction::Xor:
4654 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4656 case Instruction::ICmp:
4657 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4658 I->getOperand(0), I->getOperand(1), Q);
4660 case Instruction::FCmp:
4662 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4663 I->getOperand(1), I->getFastMathFlags(), Q);
4665 case Instruction::Select:
4666 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4667 I->getOperand(2), Q);
4669 case Instruction::GetElementPtr: {
4670 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4671 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4675 case Instruction::InsertValue: {
4676 InsertValueInst *IV = cast<InsertValueInst>(I);
4677 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4678 IV->getInsertedValueOperand(),
4679 IV->getIndices(), Q);
4682 case Instruction::ExtractValue: {
4683 auto *EVI = cast<ExtractValueInst>(I);
4684 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4685 EVI->getIndices(), Q);
4688 case Instruction::ExtractElement: {
4689 auto *EEI = cast<ExtractElementInst>(I);
4690 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4691 EEI->getIndexOperand(), Q);
4694 case Instruction::ShuffleVector: {
4695 auto *SVI = cast<ShuffleVectorInst>(I);
4696 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4697 SVI->getMask(), SVI->getType(), Q);
4700 case Instruction::PHI:
4701 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4703 case Instruction::Call: {
4704 CallSite CS(cast<CallInst>(I));
4705 Result = SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4709 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4710 #include "llvm/IR/Instruction.def"
4711 #undef HANDLE_CAST_INST
4713 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4715 case Instruction::Alloca:
4716 // No simplifications for Alloca and it can't be constant folded.
4721 // In general, it is possible for computeKnownBits to determine all bits in a
4722 // value even when the operands are not all constants.
4723 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4724 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4725 if (Known.isConstant())
4726 Result = ConstantInt::get(I->getType(), Known.getConstant());
4729 /// If called on unreachable code, the above logic may report that the
4730 /// instruction simplified to itself. Make life easier for users by
4731 /// detecting that case here, returning a safe value instead.
4732 return Result == I ? UndefValue::get(I->getType()) : Result;
4735 /// \brief Implementation of recursive simplification through an instruction's
4738 /// This is the common implementation of the recursive simplification routines.
4739 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4740 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4741 /// instructions to process and attempt to simplify it using
4742 /// InstructionSimplify.
4744 /// This routine returns 'true' only when *it* simplifies something. The passed
4745 /// in simplified value does not count toward this.
4746 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4747 const TargetLibraryInfo *TLI,
4748 const DominatorTree *DT,
4749 AssumptionCache *AC) {
4750 bool Simplified = false;
4751 SmallSetVector<Instruction *, 8> Worklist;
4752 const DataLayout &DL = I->getModule()->getDataLayout();
4754 // If we have an explicit value to collapse to, do that round of the
4755 // simplification loop by hand initially.
4757 for (User *U : I->users())
4759 Worklist.insert(cast<Instruction>(U));
4761 // Replace the instruction with its simplified value.
4762 I->replaceAllUsesWith(SimpleV);
4764 // Gracefully handle edge cases where the instruction is not wired into any
4766 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4767 !I->mayHaveSideEffects())
4768 I->eraseFromParent();
4773 // Note that we must test the size on each iteration, the worklist can grow.
4774 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4777 // See if this instruction simplifies.
4778 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4784 // Stash away all the uses of the old instruction so we can check them for
4785 // recursive simplifications after a RAUW. This is cheaper than checking all
4786 // uses of To on the recursive step in most cases.
4787 for (User *U : I->users())
4788 Worklist.insert(cast<Instruction>(U));
4790 // Replace the instruction with its simplified value.
4791 I->replaceAllUsesWith(SimpleV);
4793 // Gracefully handle edge cases where the instruction is not wired into any
4795 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4796 !I->mayHaveSideEffects())
4797 I->eraseFromParent();
4802 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4803 const TargetLibraryInfo *TLI,
4804 const DominatorTree *DT,
4805 AssumptionCache *AC) {
4806 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4809 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4810 const TargetLibraryInfo *TLI,
4811 const DominatorTree *DT,
4812 AssumptionCache *AC) {
4813 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4814 assert(SimpleV && "Must provide a simplified value.");
4815 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4819 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4820 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4821 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4822 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4823 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4824 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4825 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4826 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4829 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4830 const DataLayout &DL) {
4831 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4834 template <class T, class... TArgs>
4835 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4837 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4838 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4839 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4840 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4842 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,