1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements routines for folding instructions into simpler forms
11 // that do not require creating new instructions. This does constant folding
12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15 // simplified: This is usually true and assuming it simplifies the logic (if
16 // they have not been simplified then results are correct but maybe suboptimal).
18 //===----------------------------------------------------------------------===//
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/LoopAnalysisManager.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/Analysis/VectorUtils.h"
32 #include "llvm/IR/ConstantRange.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/GetElementPtrTypeIterator.h"
36 #include "llvm/IR/GlobalAlias.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Support/KnownBits.h"
43 using namespace llvm::PatternMatch;
45 #define DEBUG_TYPE "instsimplify"
47 enum { RecursionLimit = 3 };
49 STATISTIC(NumExpand, "Number of expansions");
50 STATISTIC(NumReassoc, "Number of reassociations");
52 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
53 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
55 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
56 const SimplifyQuery &, unsigned);
57 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
59 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
60 const SimplifyQuery &Q, unsigned MaxRecurse);
61 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
62 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 static Value *SimplifyCastInst(unsigned, Value *, Type *,
64 const SimplifyQuery &, unsigned);
66 /// For a boolean type or a vector of boolean type, return false or a vector
67 /// with every element false.
68 static Constant *getFalse(Type *Ty) {
69 return ConstantInt::getFalse(Ty);
72 /// For a boolean type or a vector of boolean type, return true or a vector
73 /// with every element true.
74 static Constant *getTrue(Type *Ty) {
75 return ConstantInt::getTrue(Ty);
78 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
79 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
81 CmpInst *Cmp = dyn_cast<CmpInst>(V);
84 CmpInst::Predicate CPred = Cmp->getPredicate();
85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
86 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
92 /// Does the given value dominate the specified phi node?
93 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
94 Instruction *I = dyn_cast<Instruction>(V);
96 // Arguments and constants dominate all instructions.
99 // If we are processing instructions (and/or basic blocks) that have not been
100 // fully added to a function, the parent nodes may still be null. Simply
101 // return the conservative answer in these cases.
102 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
105 // If we have a DominatorTree then do a precise test.
107 if (!DT->isReachableFromEntry(P->getParent()))
109 if (!DT->isReachableFromEntry(I->getParent()))
111 return DT->dominates(I, P);
114 // Otherwise, if the instruction is in the entry block and is not an invoke,
115 // then it obviously dominates all phi nodes.
116 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
123 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
124 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
125 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
126 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
127 /// Returns the simplified value, or null if no simplification was performed.
128 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
129 Instruction::BinaryOps OpcodeToExpand, const SimplifyQuery &Q,
130 unsigned MaxRecurse) {
131 // Recursion is always used, so bail out at once if we already hit the limit.
135 // Check whether the expression has the form "(A op' B) op C".
136 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
137 if (Op0->getOpcode() == OpcodeToExpand) {
138 // It does! Try turning it into "(A op C) op' (B op C)".
139 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
140 // Do "A op C" and "B op C" both simplify?
141 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
142 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
143 // They do! Return "L op' R" if it simplifies or is already available.
144 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
145 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
146 && L == B && R == A)) {
150 // Otherwise return "L op' R" if it simplifies.
151 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
158 // Check whether the expression has the form "A op (B op' C)".
159 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
160 if (Op1->getOpcode() == OpcodeToExpand) {
161 // It does! Try turning it into "(A op B) op' (A op C)".
162 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
163 // Do "A op B" and "A op C" both simplify?
164 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
165 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
166 // They do! Return "L op' R" if it simplifies or is already available.
167 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
168 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
169 && L == C && R == B)) {
173 // Otherwise return "L op' R" if it simplifies.
174 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
184 /// Generic simplifications for associative binary operations.
185 /// Returns the simpler value, or null if none was found.
186 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
187 Value *LHS, Value *RHS, const SimplifyQuery &Q,
188 unsigned MaxRecurse) {
189 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
191 // Recursion is always used, so bail out at once if we already hit the limit.
195 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
196 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
198 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
199 if (Op0 && Op0->getOpcode() == Opcode) {
200 Value *A = Op0->getOperand(0);
201 Value *B = Op0->getOperand(1);
204 // Does "B op C" simplify?
205 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
206 // It does! Return "A op V" if it simplifies or is already available.
207 // If V equals B then "A op V" is just the LHS.
208 if (V == B) return LHS;
209 // Otherwise return "A op V" if it simplifies.
210 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
217 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
218 if (Op1 && Op1->getOpcode() == Opcode) {
220 Value *B = Op1->getOperand(0);
221 Value *C = Op1->getOperand(1);
223 // Does "A op B" simplify?
224 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
225 // It does! Return "V op C" if it simplifies or is already available.
226 // If V equals B then "V op C" is just the RHS.
227 if (V == B) return RHS;
228 // Otherwise return "V op C" if it simplifies.
229 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
236 // The remaining transforms require commutativity as well as associativity.
237 if (!Instruction::isCommutative(Opcode))
240 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
241 if (Op0 && Op0->getOpcode() == Opcode) {
242 Value *A = Op0->getOperand(0);
243 Value *B = Op0->getOperand(1);
246 // Does "C op A" simplify?
247 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
248 // It does! Return "V op B" if it simplifies or is already available.
249 // If V equals A then "V op B" is just the LHS.
250 if (V == A) return LHS;
251 // Otherwise return "V op B" if it simplifies.
252 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
259 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
260 if (Op1 && Op1->getOpcode() == Opcode) {
262 Value *B = Op1->getOperand(0);
263 Value *C = Op1->getOperand(1);
265 // Does "C op A" simplify?
266 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
267 // It does! Return "B op V" if it simplifies or is already available.
268 // If V equals C then "B op V" is just the RHS.
269 if (V == C) return RHS;
270 // Otherwise return "B op V" if it simplifies.
271 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
281 /// In the case of a binary operation with a select instruction as an operand,
282 /// try to simplify the binop by seeing whether evaluating it on both branches
283 /// of the select results in the same value. Returns the common value if so,
284 /// otherwise returns null.
285 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
286 Value *RHS, const SimplifyQuery &Q,
287 unsigned MaxRecurse) {
288 // Recursion is always used, so bail out at once if we already hit the limit.
293 if (isa<SelectInst>(LHS)) {
294 SI = cast<SelectInst>(LHS);
296 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
297 SI = cast<SelectInst>(RHS);
300 // Evaluate the BinOp on the true and false branches of the select.
304 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
305 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
307 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
308 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
311 // If they simplified to the same value, then return the common value.
312 // If they both failed to simplify then return null.
316 // If one branch simplified to undef, return the other one.
317 if (TV && isa<UndefValue>(TV))
319 if (FV && isa<UndefValue>(FV))
322 // If applying the operation did not change the true and false select values,
323 // then the result of the binop is the select itself.
324 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
327 // If one branch simplified and the other did not, and the simplified
328 // value is equal to the unsimplified one, return the simplified value.
329 // For example, select (cond, X, X & Z) & Z -> X & Z.
330 if ((FV && !TV) || (TV && !FV)) {
331 // Check that the simplified value has the form "X op Y" where "op" is the
332 // same as the original operation.
333 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
334 if (Simplified && Simplified->getOpcode() == Opcode) {
335 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
336 // We already know that "op" is the same as for the simplified value. See
337 // if the operands match too. If so, return the simplified value.
338 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
339 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
340 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
341 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
342 Simplified->getOperand(1) == UnsimplifiedRHS)
344 if (Simplified->isCommutative() &&
345 Simplified->getOperand(1) == UnsimplifiedLHS &&
346 Simplified->getOperand(0) == UnsimplifiedRHS)
354 /// In the case of a comparison with a select instruction, try to simplify the
355 /// comparison by seeing whether both branches of the select result in the same
356 /// value. Returns the common value if so, otherwise returns null.
357 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
358 Value *RHS, const SimplifyQuery &Q,
359 unsigned MaxRecurse) {
360 // Recursion is always used, so bail out at once if we already hit the limit.
364 // Make sure the select is on the LHS.
365 if (!isa<SelectInst>(LHS)) {
367 Pred = CmpInst::getSwappedPredicate(Pred);
369 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
370 SelectInst *SI = cast<SelectInst>(LHS);
371 Value *Cond = SI->getCondition();
372 Value *TV = SI->getTrueValue();
373 Value *FV = SI->getFalseValue();
375 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
376 // Does "cmp TV, RHS" simplify?
377 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
379 // It not only simplified, it simplified to the select condition. Replace
381 TCmp = getTrue(Cond->getType());
383 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
384 // condition then we can replace it with 'true'. Otherwise give up.
385 if (!isSameCompare(Cond, Pred, TV, RHS))
387 TCmp = getTrue(Cond->getType());
390 // Does "cmp FV, RHS" simplify?
391 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
393 // It not only simplified, it simplified to the select condition. Replace
395 FCmp = getFalse(Cond->getType());
397 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
398 // condition then we can replace it with 'false'. Otherwise give up.
399 if (!isSameCompare(Cond, Pred, FV, RHS))
401 FCmp = getFalse(Cond->getType());
404 // If both sides simplified to the same value, then use it as the result of
405 // the original comparison.
409 // The remaining cases only make sense if the select condition has the same
410 // type as the result of the comparison, so bail out if this is not so.
411 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
413 // If the false value simplified to false, then the result of the compare
414 // is equal to "Cond && TCmp". This also catches the case when the false
415 // value simplified to false and the true value to true, returning "Cond".
416 if (match(FCmp, m_Zero()))
417 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
419 // If the true value simplified to true, then the result of the compare
420 // is equal to "Cond || FCmp".
421 if (match(TCmp, m_One()))
422 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
424 // Finally, if the false value simplified to true and the true value to
425 // false, then the result of the compare is equal to "!Cond".
426 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
428 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
435 /// In the case of a binary operation with an operand that is a PHI instruction,
436 /// try to simplify the binop by seeing whether evaluating it on the incoming
437 /// phi values yields the same result for every value. If so returns the common
438 /// value, otherwise returns null.
439 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
440 Value *RHS, const SimplifyQuery &Q,
441 unsigned MaxRecurse) {
442 // Recursion is always used, so bail out at once if we already hit the limit.
447 if (isa<PHINode>(LHS)) {
448 PI = cast<PHINode>(LHS);
449 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
450 if (!ValueDominatesPHI(RHS, PI, Q.DT))
453 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
454 PI = cast<PHINode>(RHS);
455 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
456 if (!ValueDominatesPHI(LHS, PI, Q.DT))
460 // Evaluate the BinOp on the incoming phi values.
461 Value *CommonValue = nullptr;
462 for (Value *Incoming : PI->incoming_values()) {
463 // If the incoming value is the phi node itself, it can safely be skipped.
464 if (Incoming == PI) continue;
465 Value *V = PI == LHS ?
466 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
467 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
468 // If the operation failed to simplify, or simplified to a different value
469 // to previously, then give up.
470 if (!V || (CommonValue && V != CommonValue))
478 /// In the case of a comparison with a PHI instruction, try to simplify the
479 /// comparison by seeing whether comparing with all of the incoming phi values
480 /// yields the same result every time. If so returns the common result,
481 /// otherwise returns null.
482 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
483 const SimplifyQuery &Q, unsigned MaxRecurse) {
484 // Recursion is always used, so bail out at once if we already hit the limit.
488 // Make sure the phi is on the LHS.
489 if (!isa<PHINode>(LHS)) {
491 Pred = CmpInst::getSwappedPredicate(Pred);
493 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
494 PHINode *PI = cast<PHINode>(LHS);
496 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
497 if (!ValueDominatesPHI(RHS, PI, Q.DT))
500 // Evaluate the BinOp on the incoming phi values.
501 Value *CommonValue = nullptr;
502 for (Value *Incoming : PI->incoming_values()) {
503 // If the incoming value is the phi node itself, it can safely be skipped.
504 if (Incoming == PI) continue;
505 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
506 // If the operation failed to simplify, or simplified to a different value
507 // to previously, then give up.
508 if (!V || (CommonValue && V != CommonValue))
516 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
517 Value *&Op0, Value *&Op1,
518 const SimplifyQuery &Q) {
519 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
520 if (auto *CRHS = dyn_cast<Constant>(Op1))
521 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
523 // Canonicalize the constant to the RHS if this is a commutative operation.
524 if (Instruction::isCommutative(Opcode))
530 /// Given operands for an Add, see if we can fold the result.
531 /// If not, this returns null.
532 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
533 const SimplifyQuery &Q, unsigned MaxRecurse) {
534 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
537 // X + undef -> undef
538 if (match(Op1, m_Undef()))
542 if (match(Op1, m_Zero()))
549 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
550 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
553 // X + ~X -> -1 since ~X = -X-1
554 Type *Ty = Op0->getType();
555 if (match(Op0, m_Not(m_Specific(Op1))) ||
556 match(Op1, m_Not(m_Specific(Op0))))
557 return Constant::getAllOnesValue(Ty);
559 // add nsw/nuw (xor Y, signmask), signmask --> Y
560 // The no-wrapping add guarantees that the top bit will be set by the add.
561 // Therefore, the xor must be clearing the already set sign bit of Y.
562 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
563 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
567 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
568 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
571 // Try some generic simplifications for associative operations.
572 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
576 // Threading Add over selects and phi nodes is pointless, so don't bother.
577 // Threading over the select in "A + select(cond, B, C)" means evaluating
578 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
579 // only if B and C are equal. If B and C are equal then (since we assume
580 // that operands have already been simplified) "select(cond, B, C)" should
581 // have been simplified to the common value of B and C already. Analysing
582 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
583 // for threading over phi nodes.
588 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
589 const SimplifyQuery &Query) {
590 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
593 /// \brief Compute the base pointer and cumulative constant offsets for V.
595 /// This strips all constant offsets off of V, leaving it the base pointer, and
596 /// accumulates the total constant offset applied in the returned constant. It
597 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
598 /// no constant offsets applied.
600 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
601 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
603 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
604 bool AllowNonInbounds = false) {
605 assert(V->getType()->getScalarType()->isPointerTy());
607 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
608 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
610 // Even though we don't look through PHI nodes, we could be called on an
611 // instruction in an unreachable block, which may be on a cycle.
612 SmallPtrSet<Value *, 4> Visited;
615 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
616 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
617 !GEP->accumulateConstantOffset(DL, Offset))
619 V = GEP->getPointerOperand();
620 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
621 V = cast<Operator>(V)->getOperand(0);
622 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
623 if (GA->isInterposable())
625 V = GA->getAliasee();
627 if (auto CS = CallSite(V))
628 if (Value *RV = CS.getReturnedArgOperand()) {
634 assert(V->getType()->getScalarType()->isPointerTy() &&
635 "Unexpected operand type!");
636 } while (Visited.insert(V).second);
638 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
639 if (V->getType()->isVectorTy())
640 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
645 /// \brief Compute the constant difference between two pointer values.
646 /// If the difference is not a constant, returns zero.
647 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
649 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
650 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
652 // If LHS and RHS are not related via constant offsets to the same base
653 // value, there is nothing we can do here.
657 // Otherwise, the difference of LHS - RHS can be computed as:
659 // = (LHSOffset + Base) - (RHSOffset + Base)
660 // = LHSOffset - RHSOffset
661 return ConstantExpr::getSub(LHSOffset, RHSOffset);
664 /// Given operands for a Sub, see if we can fold the result.
665 /// If not, this returns null.
666 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
667 const SimplifyQuery &Q, unsigned MaxRecurse) {
668 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
671 // X - undef -> undef
672 // undef - X -> undef
673 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
674 return UndefValue::get(Op0->getType());
677 if (match(Op1, m_Zero()))
682 return Constant::getNullValue(Op0->getType());
684 // Is this a negation?
685 if (match(Op0, m_Zero())) {
686 // 0 - X -> 0 if the sub is NUW.
690 unsigned BitWidth = Op1->getType()->getScalarSizeInBits();
691 KnownBits Known(BitWidth);
692 computeKnownBits(Op1, Known, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
693 if (Known.Zero.isMaxSignedValue()) {
694 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
695 // Op1 must be 0 because negating the minimum signed value is undefined.
699 // 0 - X -> X if X is 0 or the minimum signed value.
704 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
705 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
706 Value *X = nullptr, *Y = nullptr, *Z = Op1;
707 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
708 // See if "V === Y - Z" simplifies.
709 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
710 // It does! Now see if "X + V" simplifies.
711 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
712 // It does, we successfully reassociated!
716 // See if "V === X - Z" simplifies.
717 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
718 // It does! Now see if "Y + V" simplifies.
719 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
720 // It does, we successfully reassociated!
726 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
727 // For example, X - (X + 1) -> -1
729 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
730 // See if "V === X - Y" simplifies.
731 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
732 // It does! Now see if "V - Z" simplifies.
733 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
734 // It does, we successfully reassociated!
738 // See if "V === X - Z" simplifies.
739 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
740 // It does! Now see if "V - Y" simplifies.
741 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
742 // It does, we successfully reassociated!
748 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
749 // For example, X - (X - Y) -> Y.
751 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
752 // See if "V === Z - X" simplifies.
753 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
754 // It does! Now see if "V + Y" simplifies.
755 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
756 // It does, we successfully reassociated!
761 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
762 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
763 match(Op1, m_Trunc(m_Value(Y))))
764 if (X->getType() == Y->getType())
765 // See if "V === X - Y" simplifies.
766 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
767 // It does! Now see if "trunc V" simplifies.
768 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
770 // It does, return the simplified "trunc V".
773 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
774 if (match(Op0, m_PtrToInt(m_Value(X))) &&
775 match(Op1, m_PtrToInt(m_Value(Y))))
776 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
777 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
780 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
781 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
784 // Threading Sub over selects and phi nodes is pointless, so don't bother.
785 // Threading over the select in "A - select(cond, B, C)" means evaluating
786 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
787 // only if B and C are equal. If B and C are equal then (since we assume
788 // that operands have already been simplified) "select(cond, B, C)" should
789 // have been simplified to the common value of B and C already. Analysing
790 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
791 // for threading over phi nodes.
796 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
797 const SimplifyQuery &Q) {
798 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
801 /// Given operands for an FAdd, see if we can fold the result. If not, this
803 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
804 const SimplifyQuery &Q, unsigned MaxRecurse) {
805 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
809 if (match(Op1, m_NegZero()))
812 // fadd X, 0 ==> X, when we know X is not -0
813 if (match(Op1, m_Zero()) &&
814 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
817 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
818 // where nnan and ninf have to occur at least once somewhere in this
820 Value *SubOp = nullptr;
821 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
823 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
826 Instruction *FSub = cast<Instruction>(SubOp);
827 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
828 (FMF.noInfs() || FSub->hasNoInfs()))
829 return Constant::getNullValue(Op0->getType());
835 /// Given operands for an FSub, see if we can fold the result. If not, this
837 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
838 const SimplifyQuery &Q, unsigned MaxRecurse) {
839 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
843 if (match(Op1, m_Zero()))
846 // fsub X, -0 ==> X, when we know X is not -0
847 if (match(Op1, m_NegZero()) &&
848 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
851 // fsub -0.0, (fsub -0.0, X) ==> X
853 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
856 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
857 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
858 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
861 // fsub nnan x, x ==> 0.0
862 if (FMF.noNaNs() && Op0 == Op1)
863 return Constant::getNullValue(Op0->getType());
868 /// Given the operands for an FMul, see if we can fold the result
869 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
870 const SimplifyQuery &Q, unsigned MaxRecurse) {
871 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
875 if (match(Op1, m_FPOne()))
878 // fmul nnan nsz X, 0 ==> 0
879 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
885 /// Given operands for a Mul, see if we can fold the result.
886 /// If not, this returns null.
887 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
888 unsigned MaxRecurse) {
889 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
893 if (match(Op1, m_Undef()))
894 return Constant::getNullValue(Op0->getType());
897 if (match(Op1, m_Zero()))
901 if (match(Op1, m_One()))
904 // (X / Y) * Y -> X if the division is exact.
906 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
907 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
911 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
912 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
915 // Try some generic simplifications for associative operations.
916 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
920 // Mul distributes over Add. Try some generic simplifications based on this.
921 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
925 // If the operation is with the result of a select instruction, check whether
926 // operating on either branch of the select always yields the same value.
927 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
928 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
932 // If the operation is with the result of a phi instruction, check whether
933 // operating on all incoming values of the phi always yields the same value.
934 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
935 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
942 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
943 const SimplifyQuery &Q) {
944 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
948 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
949 const SimplifyQuery &Q) {
950 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
953 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
954 const SimplifyQuery &Q) {
955 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
958 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
959 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
962 /// Check for common or similar folds of integer division or integer remainder.
963 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
964 Type *Ty = Op0->getType();
966 // X / undef -> undef
967 // X % undef -> undef
968 if (match(Op1, m_Undef()))
973 // We don't need to preserve faults!
974 if (match(Op1, m_Zero()))
975 return UndefValue::get(Ty);
977 // If any element of a constant divisor vector is zero, the whole op is undef.
978 auto *Op1C = dyn_cast<Constant>(Op1);
979 if (Op1C && Ty->isVectorTy()) {
980 unsigned NumElts = Ty->getVectorNumElements();
981 for (unsigned i = 0; i != NumElts; ++i) {
982 Constant *Elt = Op1C->getAggregateElement(i);
983 if (Elt && Elt->isNullValue())
984 return UndefValue::get(Ty);
990 if (match(Op0, m_Undef()))
991 return Constant::getNullValue(Ty);
995 if (match(Op0, m_Zero()))
1001 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
1005 // If this is a boolean op (single-bit element type), we can't have
1006 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
1007 if (match(Op1, m_One()) || Ty->getScalarType()->isIntegerTy(1))
1008 return IsDiv ? Op0 : Constant::getNullValue(Ty);
1013 /// Given operands for an SDiv or UDiv, see if we can fold the result.
1014 /// If not, this returns null.
1015 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1016 const SimplifyQuery &Q, unsigned MaxRecurse) {
1017 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1020 if (Value *V = simplifyDivRem(Op0, Op1, true))
1023 bool isSigned = Opcode == Instruction::SDiv;
1025 // (X * Y) / Y -> X if the multiplication does not overflow.
1026 Value *X = nullptr, *Y = nullptr;
1027 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1028 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1029 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1030 // If the Mul knows it does not overflow, then we are good to go.
1031 if ((isSigned && Mul->hasNoSignedWrap()) ||
1032 (!isSigned && Mul->hasNoUnsignedWrap()))
1034 // If X has the form X = A / Y then X * Y cannot overflow.
1035 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1036 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1040 // (X rem Y) / Y -> 0
1041 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1042 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1043 return Constant::getNullValue(Op0->getType());
1045 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1046 ConstantInt *C1, *C2;
1047 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1048 match(Op1, m_ConstantInt(C2))) {
1050 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1052 return Constant::getNullValue(Op0->getType());
1055 // If the operation is with the result of a select instruction, check whether
1056 // operating on either branch of the select always yields the same value.
1057 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1058 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1061 // If the operation is with the result of a phi instruction, check whether
1062 // operating on all incoming values of the phi always yields the same value.
1063 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1064 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1070 /// Given operands for an SDiv, see if we can fold the result.
1071 /// If not, this returns null.
1072 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1073 unsigned MaxRecurse) {
1074 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1080 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1081 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1084 /// Given operands for a UDiv, see if we can fold the result.
1085 /// If not, this returns null.
1086 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1087 unsigned MaxRecurse) {
1088 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1091 // udiv %V, C -> 0 if %V < C
1093 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1094 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1095 if (C->isAllOnesValue()) {
1096 return Constant::getNullValue(Op0->getType());
1104 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1105 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1108 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1109 const SimplifyQuery &Q, unsigned) {
1110 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
1113 // undef / X -> undef (the undef could be a snan).
1114 if (match(Op0, m_Undef()))
1117 // X / undef -> undef
1118 if (match(Op1, m_Undef()))
1122 if (match(Op1, m_FPOne()))
1126 // Requires that NaNs are off (X could be zero) and signed zeroes are
1127 // ignored (X could be positive or negative, so the output sign is unknown).
1128 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1132 // X / X -> 1.0 is legal when NaNs are ignored.
1134 return ConstantFP::get(Op0->getType(), 1.0);
1136 // -X / X -> -1.0 and
1137 // X / -X -> -1.0 are legal when NaNs are ignored.
1138 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1139 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1140 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1141 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1142 BinaryOperator::getFNegArgument(Op1) == Op0))
1143 return ConstantFP::get(Op0->getType(), -1.0);
1149 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1150 const SimplifyQuery &Q) {
1151 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
1154 /// Given operands for an SRem or URem, see if we can fold the result.
1155 /// If not, this returns null.
1156 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1157 const SimplifyQuery &Q, unsigned MaxRecurse) {
1158 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1161 if (Value *V = simplifyDivRem(Op0, Op1, false))
1164 // (X % Y) % Y -> X % Y
1165 if ((Opcode == Instruction::SRem &&
1166 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1167 (Opcode == Instruction::URem &&
1168 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1171 // If the operation is with the result of a select instruction, check whether
1172 // operating on either branch of the select always yields the same value.
1173 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1174 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1177 // If the operation is with the result of a phi instruction, check whether
1178 // operating on all incoming values of the phi always yields the same value.
1179 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1180 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1186 /// Given operands for an SRem, see if we can fold the result.
1187 /// If not, this returns null.
1188 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1189 unsigned MaxRecurse) {
1190 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1196 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1197 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1200 /// Given operands for a URem, see if we can fold the result.
1201 /// If not, this returns null.
1202 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1203 unsigned MaxRecurse) {
1204 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1207 // urem %V, C -> %V if %V < C
1209 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1210 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1211 if (C->isAllOnesValue()) {
1220 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1221 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1224 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1225 const SimplifyQuery &Q, unsigned) {
1226 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
1229 // undef % X -> undef (the undef could be a snan).
1230 if (match(Op0, m_Undef()))
1233 // X % undef -> undef
1234 if (match(Op1, m_Undef()))
1238 // Requires that NaNs are off (X could be zero) and signed zeroes are
1239 // ignored (X could be positive or negative, so the output sign is unknown).
1240 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1246 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1247 const SimplifyQuery &Q) {
1248 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
1251 /// Returns true if a shift by \c Amount always yields undef.
1252 static bool isUndefShift(Value *Amount) {
1253 Constant *C = dyn_cast<Constant>(Amount);
1257 // X shift by undef -> undef because it may shift by the bitwidth.
1258 if (isa<UndefValue>(C))
1261 // Shifting by the bitwidth or more is undefined.
1262 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1263 if (CI->getValue().getLimitedValue() >=
1264 CI->getType()->getScalarSizeInBits())
1267 // If all lanes of a vector shift are undefined the whole shift is.
1268 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1269 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1270 if (!isUndefShift(C->getAggregateElement(I)))
1278 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1279 /// If not, this returns null.
1280 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1281 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1282 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1285 // 0 shift by X -> 0
1286 if (match(Op0, m_Zero()))
1289 // X shift by 0 -> X
1290 if (match(Op1, m_Zero()))
1293 // Fold undefined shifts.
1294 if (isUndefShift(Op1))
1295 return UndefValue::get(Op0->getType());
1297 // If the operation is with the result of a select instruction, check whether
1298 // operating on either branch of the select always yields the same value.
1299 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1300 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1303 // If the operation is with the result of a phi instruction, check whether
1304 // operating on all incoming values of the phi always yields the same value.
1305 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1306 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1309 // If any bits in the shift amount make that value greater than or equal to
1310 // the number of bits in the type, the shift is undefined.
1311 unsigned BitWidth = Op1->getType()->getScalarSizeInBits();
1312 KnownBits Known(BitWidth);
1313 computeKnownBits(Op1, Known, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1314 if (Known.One.getLimitedValue() >= BitWidth)
1315 return UndefValue::get(Op0->getType());
1317 // If all valid bits in the shift amount are known zero, the first operand is
1319 unsigned NumValidShiftBits = Log2_32_Ceil(BitWidth);
1320 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1326 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1327 /// fold the result. If not, this returns null.
1328 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1329 Value *Op1, bool isExact, const SimplifyQuery &Q,
1330 unsigned MaxRecurse) {
1331 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1336 return Constant::getNullValue(Op0->getType());
1339 // undef >> X -> undef (if it's exact)
1340 if (match(Op0, m_Undef()))
1341 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1343 // The low bit cannot be shifted out of an exact shift if it is set.
1345 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
1346 KnownBits Op0Known(BitWidth);
1347 computeKnownBits(Op0, Op0Known, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1348 if (Op0Known.One[0])
1355 /// Given operands for an Shl, see if we can fold the result.
1356 /// If not, this returns null.
1357 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1358 const SimplifyQuery &Q, unsigned MaxRecurse) {
1359 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1363 // undef << X -> undef if (if it's NSW/NUW)
1364 if (match(Op0, m_Undef()))
1365 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1367 // (X >> A) << A -> X
1369 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1374 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1375 const SimplifyQuery &Q) {
1376 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1379 /// Given operands for an LShr, see if we can fold the result.
1380 /// If not, this returns null.
1381 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1382 const SimplifyQuery &Q, unsigned MaxRecurse) {
1383 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1387 // (X << A) >> A -> X
1389 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1395 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1396 const SimplifyQuery &Q) {
1397 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1400 /// Given operands for an AShr, see if we can fold the result.
1401 /// If not, this returns null.
1402 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1403 const SimplifyQuery &Q, unsigned MaxRecurse) {
1404 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1408 // all ones >>a X -> all ones
1409 if (match(Op0, m_AllOnes()))
1412 // (X << A) >> A -> X
1414 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1417 // Arithmetic shifting an all-sign-bit value is a no-op.
1418 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1419 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1425 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1426 const SimplifyQuery &Q) {
1427 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1430 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1431 ICmpInst *UnsignedICmp, bool IsAnd) {
1434 ICmpInst::Predicate EqPred;
1435 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1436 !ICmpInst::isEquality(EqPred))
1439 ICmpInst::Predicate UnsignedPred;
1440 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1441 ICmpInst::isUnsigned(UnsignedPred))
1443 else if (match(UnsignedICmp,
1444 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1445 ICmpInst::isUnsigned(UnsignedPred))
1446 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1450 // X < Y && Y != 0 --> X < Y
1451 // X < Y || Y != 0 --> Y != 0
1452 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1453 return IsAnd ? UnsignedICmp : ZeroICmp;
1455 // X >= Y || Y != 0 --> true
1456 // X >= Y || Y == 0 --> X >= Y
1457 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1458 if (EqPred == ICmpInst::ICMP_NE)
1459 return getTrue(UnsignedICmp->getType());
1460 return UnsignedICmp;
1463 // X < Y && Y == 0 --> false
1464 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1466 return getFalse(UnsignedICmp->getType());
1471 /// Commuted variants are assumed to be handled by calling this function again
1472 /// with the parameters swapped.
1473 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1474 ICmpInst::Predicate Pred0, Pred1;
1476 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1477 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1480 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1481 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1482 // can eliminate Op1 from this 'and'.
1483 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1486 // Check for any combination of predicates that are guaranteed to be disjoint.
1487 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1488 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1489 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1490 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1491 return getFalse(Op0->getType());
1496 /// Commuted variants are assumed to be handled by calling this function again
1497 /// with the parameters swapped.
1498 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1499 ICmpInst::Predicate Pred0, Pred1;
1501 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1502 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1505 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1506 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1507 // can eliminate Op0 from this 'or'.
1508 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1511 // Check for any combination of predicates that cover the entire range of
1513 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1514 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1515 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1516 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1517 return getTrue(Op0->getType());
1522 /// Test if a pair of compares with a shared operand and 2 constants has an
1523 /// empty set intersection, full set union, or if one compare is a superset of
1525 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1527 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1528 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1531 const APInt *C0, *C1;
1532 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1533 !match(Cmp1->getOperand(1), m_APInt(C1)))
1536 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1537 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1539 // For and-of-compares, check if the intersection is empty:
1540 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1541 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1542 return getFalse(Cmp0->getType());
1544 // For or-of-compares, check if the union is full:
1545 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1546 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1547 return getTrue(Cmp0->getType());
1549 // Is one range a superset of the other?
1550 // If this is and-of-compares, take the smaller set:
1551 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1552 // If this is or-of-compares, take the larger set:
1553 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1554 if (Range0.contains(Range1))
1555 return IsAnd ? Cmp1 : Cmp0;
1556 if (Range1.contains(Range0))
1557 return IsAnd ? Cmp0 : Cmp1;
1562 /// Commuted variants are assumed to be handled by calling this function again
1563 /// with the parameters swapped.
1564 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1565 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1568 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1571 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1574 // (icmp (add V, C0), C1) & (icmp V, C0)
1575 Type *ITy = Op0->getType();
1576 ICmpInst::Predicate Pred0, Pred1;
1577 const APInt *C0, *C1;
1579 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1582 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1585 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1586 if (AddInst->getOperand(1) != Op1->getOperand(1))
1589 bool isNSW = AddInst->hasNoSignedWrap();
1590 bool isNUW = AddInst->hasNoUnsignedWrap();
1592 const APInt Delta = *C1 - *C0;
1593 if (C0->isStrictlyPositive()) {
1595 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1596 return getFalse(ITy);
1597 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1598 return getFalse(ITy);
1601 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1602 return getFalse(ITy);
1603 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1604 return getFalse(ITy);
1607 if (C0->getBoolValue() && isNUW) {
1609 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1610 return getFalse(ITy);
1612 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1613 return getFalse(ITy);
1619 /// Commuted variants are assumed to be handled by calling this function again
1620 /// with the parameters swapped.
1621 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1622 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1625 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1628 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1631 // (icmp (add V, C0), C1) | (icmp V, C0)
1632 ICmpInst::Predicate Pred0, Pred1;
1633 const APInt *C0, *C1;
1635 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1638 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1641 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1642 if (AddInst->getOperand(1) != Op1->getOperand(1))
1645 Type *ITy = Op0->getType();
1646 bool isNSW = AddInst->hasNoSignedWrap();
1647 bool isNUW = AddInst->hasNoUnsignedWrap();
1649 const APInt Delta = *C1 - *C0;
1650 if (C0->isStrictlyPositive()) {
1652 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1653 return getTrue(ITy);
1654 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1655 return getTrue(ITy);
1658 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1659 return getTrue(ITy);
1660 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1661 return getTrue(ITy);
1664 if (C0->getBoolValue() && isNUW) {
1666 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1667 return getTrue(ITy);
1669 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1670 return getTrue(ITy);
1676 static Value *simplifyPossiblyCastedAndOrOfICmps(ICmpInst *Cmp0, ICmpInst *Cmp1,
1677 bool IsAnd, CastInst *Cast) {
1679 IsAnd ? simplifyAndOfICmps(Cmp0, Cmp1) : simplifyOrOfICmps(Cmp0, Cmp1);
1685 // If we looked through casts, we can only handle a constant simplification
1686 // because we are not allowed to create a cast instruction here.
1687 if (auto *C = dyn_cast<Constant>(V))
1688 return ConstantExpr::getCast(Cast->getOpcode(), C, Cast->getType());
1693 static Value *simplifyAndOrOfICmps(Value *Op0, Value *Op1, bool IsAnd) {
1694 // Look through casts of the 'and' operands to find compares.
1695 auto *Cast0 = dyn_cast<CastInst>(Op0);
1696 auto *Cast1 = dyn_cast<CastInst>(Op1);
1697 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1698 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1699 Op0 = Cast0->getOperand(0);
1700 Op1 = Cast1->getOperand(0);
1703 auto *Cmp0 = dyn_cast<ICmpInst>(Op0);
1704 auto *Cmp1 = dyn_cast<ICmpInst>(Op1);
1708 if (Value *V = simplifyPossiblyCastedAndOrOfICmps(Cmp0, Cmp1, IsAnd, Cast0))
1710 if (Value *V = simplifyPossiblyCastedAndOrOfICmps(Cmp1, Cmp0, IsAnd, Cast0))
1716 /// Given operands for an And, see if we can fold the result.
1717 /// If not, this returns null.
1718 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1719 unsigned MaxRecurse) {
1720 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1724 if (match(Op1, m_Undef()))
1725 return Constant::getNullValue(Op0->getType());
1732 if (match(Op1, m_Zero()))
1736 if (match(Op1, m_AllOnes()))
1739 // A & ~A = ~A & A = 0
1740 if (match(Op0, m_Not(m_Specific(Op1))) ||
1741 match(Op1, m_Not(m_Specific(Op0))))
1742 return Constant::getNullValue(Op0->getType());
1745 Value *A = nullptr, *B = nullptr;
1746 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1747 (A == Op1 || B == Op1))
1751 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1752 (A == Op0 || B == Op0))
1755 // A mask that only clears known zeros of a shifted value is a no-op.
1759 if (match(Op1, m_APInt(Mask))) {
1760 // If all bits in the inverted and shifted mask are clear:
1761 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1762 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1763 (~(*Mask)).lshr(*ShAmt).isNullValue())
1766 // If all bits in the inverted and shifted mask are clear:
1767 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1768 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1769 (~(*Mask)).shl(*ShAmt).isNullValue())
1773 // A & (-A) = A if A is a power of two or zero.
1774 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1775 match(Op1, m_Neg(m_Specific(Op0)))) {
1776 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1779 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1784 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, true))
1787 // Try some generic simplifications for associative operations.
1788 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1792 // And distributes over Or. Try some generic simplifications based on this.
1793 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1797 // And distributes over Xor. Try some generic simplifications based on this.
1798 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1802 // If the operation is with the result of a select instruction, check whether
1803 // operating on either branch of the select always yields the same value.
1804 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1805 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1809 // If the operation is with the result of a phi instruction, check whether
1810 // operating on all incoming values of the phi always yields the same value.
1811 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1812 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1819 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1820 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1823 /// Given operands for an Or, see if we can fold the result.
1824 /// If not, this returns null.
1825 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1826 unsigned MaxRecurse) {
1827 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1831 if (match(Op1, m_Undef()))
1832 return Constant::getAllOnesValue(Op0->getType());
1839 if (match(Op1, m_Zero()))
1843 if (match(Op1, m_AllOnes()))
1846 // A | ~A = ~A | A = -1
1847 if (match(Op0, m_Not(m_Specific(Op1))) ||
1848 match(Op1, m_Not(m_Specific(Op0))))
1849 return Constant::getAllOnesValue(Op0->getType());
1852 Value *A = nullptr, *B = nullptr;
1853 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1854 (A == Op1 || B == Op1))
1858 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1859 (A == Op0 || B == Op0))
1862 // ~(A & ?) | A = -1
1863 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1864 (A == Op1 || B == Op1))
1865 return Constant::getAllOnesValue(Op1->getType());
1867 // A | ~(A & ?) = -1
1868 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1869 (A == Op0 || B == Op0))
1870 return Constant::getAllOnesValue(Op0->getType());
1872 // (A & ~B) | (A ^ B) -> (A ^ B)
1873 // (~B & A) | (A ^ B) -> (A ^ B)
1874 // (A & ~B) | (B ^ A) -> (B ^ A)
1875 // (~B & A) | (B ^ A) -> (B ^ A)
1876 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1877 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1878 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1881 // Commute the 'or' operands.
1882 // (A ^ B) | (A & ~B) -> (A ^ B)
1883 // (A ^ B) | (~B & A) -> (A ^ B)
1884 // (B ^ A) | (A & ~B) -> (B ^ A)
1885 // (B ^ A) | (~B & A) -> (B ^ A)
1886 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1887 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1888 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1891 // (A & B) | (~A ^ B) -> (~A ^ B)
1892 // (B & A) | (~A ^ B) -> (~A ^ B)
1893 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1894 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1895 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1896 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1897 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1900 // (~A ^ B) | (A & B) -> (~A ^ B)
1901 // (~A ^ B) | (B & A) -> (~A ^ B)
1902 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1903 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1904 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1905 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1906 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1909 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, false))
1912 // Try some generic simplifications for associative operations.
1913 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1917 // Or distributes over And. Try some generic simplifications based on this.
1918 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1922 // If the operation is with the result of a select instruction, check whether
1923 // operating on either branch of the select always yields the same value.
1924 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1925 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1930 Value *C = nullptr, *D = nullptr;
1931 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1932 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1933 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1934 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1935 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1936 // (A & C1)|(B & C2)
1937 // If we have: ((V + N) & C1) | (V & C2)
1938 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1939 // replace with V+N.
1941 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1942 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1943 // Add commutes, try both ways.
1945 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1948 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1951 // Or commutes, try both ways.
1952 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1953 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1954 // Add commutes, try both ways.
1956 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1959 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1965 // If the operation is with the result of a phi instruction, check whether
1966 // operating on all incoming values of the phi always yields the same value.
1967 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1968 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1974 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1975 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1978 /// Given operands for a Xor, see if we can fold the result.
1979 /// If not, this returns null.
1980 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1981 unsigned MaxRecurse) {
1982 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1985 // A ^ undef -> undef
1986 if (match(Op1, m_Undef()))
1990 if (match(Op1, m_Zero()))
1995 return Constant::getNullValue(Op0->getType());
1997 // A ^ ~A = ~A ^ A = -1
1998 if (match(Op0, m_Not(m_Specific(Op1))) ||
1999 match(Op1, m_Not(m_Specific(Op0))))
2000 return Constant::getAllOnesValue(Op0->getType());
2002 // Try some generic simplifications for associative operations.
2003 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2007 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2008 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2009 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2010 // only if B and C are equal. If B and C are equal then (since we assume
2011 // that operands have already been simplified) "select(cond, B, C)" should
2012 // have been simplified to the common value of B and C already. Analysing
2013 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2014 // for threading over phi nodes.
2019 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2020 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2024 static Type *GetCompareTy(Value *Op) {
2025 return CmpInst::makeCmpResultType(Op->getType());
2028 /// Rummage around inside V looking for something equivalent to the comparison
2029 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2030 /// Helper function for analyzing max/min idioms.
2031 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2032 Value *LHS, Value *RHS) {
2033 SelectInst *SI = dyn_cast<SelectInst>(V);
2036 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2039 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2040 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2042 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2043 LHS == CmpRHS && RHS == CmpLHS)
2048 // A significant optimization not implemented here is assuming that alloca
2049 // addresses are not equal to incoming argument values. They don't *alias*,
2050 // as we say, but that doesn't mean they aren't equal, so we take a
2051 // conservative approach.
2053 // This is inspired in part by C++11 5.10p1:
2054 // "Two pointers of the same type compare equal if and only if they are both
2055 // null, both point to the same function, or both represent the same
2058 // This is pretty permissive.
2060 // It's also partly due to C11 6.5.9p6:
2061 // "Two pointers compare equal if and only if both are null pointers, both are
2062 // pointers to the same object (including a pointer to an object and a
2063 // subobject at its beginning) or function, both are pointers to one past the
2064 // last element of the same array object, or one is a pointer to one past the
2065 // end of one array object and the other is a pointer to the start of a
2066 // different array object that happens to immediately follow the first array
2067 // object in the address space.)
2069 // C11's version is more restrictive, however there's no reason why an argument
2070 // couldn't be a one-past-the-end value for a stack object in the caller and be
2071 // equal to the beginning of a stack object in the callee.
2073 // If the C and C++ standards are ever made sufficiently restrictive in this
2074 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2075 // this optimization.
2077 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2078 const DominatorTree *DT, CmpInst::Predicate Pred,
2079 const Instruction *CxtI, Value *LHS, Value *RHS) {
2080 // First, skip past any trivial no-ops.
2081 LHS = LHS->stripPointerCasts();
2082 RHS = RHS->stripPointerCasts();
2084 // A non-null pointer is not equal to a null pointer.
2085 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) &&
2086 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2087 return ConstantInt::get(GetCompareTy(LHS),
2088 !CmpInst::isTrueWhenEqual(Pred));
2090 // We can only fold certain predicates on pointer comparisons.
2095 // Equality comaprisons are easy to fold.
2096 case CmpInst::ICMP_EQ:
2097 case CmpInst::ICMP_NE:
2100 // We can only handle unsigned relational comparisons because 'inbounds' on
2101 // a GEP only protects against unsigned wrapping.
2102 case CmpInst::ICMP_UGT:
2103 case CmpInst::ICMP_UGE:
2104 case CmpInst::ICMP_ULT:
2105 case CmpInst::ICMP_ULE:
2106 // However, we have to switch them to their signed variants to handle
2107 // negative indices from the base pointer.
2108 Pred = ICmpInst::getSignedPredicate(Pred);
2112 // Strip off any constant offsets so that we can reason about them.
2113 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2114 // here and compare base addresses like AliasAnalysis does, however there are
2115 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2116 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2117 // doesn't need to guarantee pointer inequality when it says NoAlias.
2118 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2119 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2121 // If LHS and RHS are related via constant offsets to the same base
2122 // value, we can replace it with an icmp which just compares the offsets.
2124 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2126 // Various optimizations for (in)equality comparisons.
2127 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2128 // Different non-empty allocations that exist at the same time have
2129 // different addresses (if the program can tell). Global variables always
2130 // exist, so they always exist during the lifetime of each other and all
2131 // allocas. Two different allocas usually have different addresses...
2133 // However, if there's an @llvm.stackrestore dynamically in between two
2134 // allocas, they may have the same address. It's tempting to reduce the
2135 // scope of the problem by only looking at *static* allocas here. That would
2136 // cover the majority of allocas while significantly reducing the likelihood
2137 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2138 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2139 // an entry block. Also, if we have a block that's not attached to a
2140 // function, we can't tell if it's "static" under the current definition.
2141 // Theoretically, this problem could be fixed by creating a new kind of
2142 // instruction kind specifically for static allocas. Such a new instruction
2143 // could be required to be at the top of the entry block, thus preventing it
2144 // from being subject to a @llvm.stackrestore. Instcombine could even
2145 // convert regular allocas into these special allocas. It'd be nifty.
2146 // However, until then, this problem remains open.
2148 // So, we'll assume that two non-empty allocas have different addresses
2151 // With all that, if the offsets are within the bounds of their allocations
2152 // (and not one-past-the-end! so we can't use inbounds!), and their
2153 // allocations aren't the same, the pointers are not equal.
2155 // Note that it's not necessary to check for LHS being a global variable
2156 // address, due to canonicalization and constant folding.
2157 if (isa<AllocaInst>(LHS) &&
2158 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2159 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2160 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2161 uint64_t LHSSize, RHSSize;
2162 if (LHSOffsetCI && RHSOffsetCI &&
2163 getObjectSize(LHS, LHSSize, DL, TLI) &&
2164 getObjectSize(RHS, RHSSize, DL, TLI)) {
2165 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2166 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2167 if (!LHSOffsetValue.isNegative() &&
2168 !RHSOffsetValue.isNegative() &&
2169 LHSOffsetValue.ult(LHSSize) &&
2170 RHSOffsetValue.ult(RHSSize)) {
2171 return ConstantInt::get(GetCompareTy(LHS),
2172 !CmpInst::isTrueWhenEqual(Pred));
2176 // Repeat the above check but this time without depending on DataLayout
2177 // or being able to compute a precise size.
2178 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2179 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2180 LHSOffset->isNullValue() &&
2181 RHSOffset->isNullValue())
2182 return ConstantInt::get(GetCompareTy(LHS),
2183 !CmpInst::isTrueWhenEqual(Pred));
2186 // Even if an non-inbounds GEP occurs along the path we can still optimize
2187 // equality comparisons concerning the result. We avoid walking the whole
2188 // chain again by starting where the last calls to
2189 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2190 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2191 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2193 return ConstantExpr::getICmp(Pred,
2194 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2195 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2197 // If one side of the equality comparison must come from a noalias call
2198 // (meaning a system memory allocation function), and the other side must
2199 // come from a pointer that cannot overlap with dynamically-allocated
2200 // memory within the lifetime of the current function (allocas, byval
2201 // arguments, globals), then determine the comparison result here.
2202 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2203 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2204 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2206 // Is the set of underlying objects all noalias calls?
2207 auto IsNAC = [](ArrayRef<Value *> Objects) {
2208 return all_of(Objects, isNoAliasCall);
2211 // Is the set of underlying objects all things which must be disjoint from
2212 // noalias calls. For allocas, we consider only static ones (dynamic
2213 // allocas might be transformed into calls to malloc not simultaneously
2214 // live with the compared-to allocation). For globals, we exclude symbols
2215 // that might be resolve lazily to symbols in another dynamically-loaded
2216 // library (and, thus, could be malloc'ed by the implementation).
2217 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2218 return all_of(Objects, [](Value *V) {
2219 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2220 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2221 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2222 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2223 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2224 !GV->isThreadLocal();
2225 if (const Argument *A = dyn_cast<Argument>(V))
2226 return A->hasByValAttr();
2231 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2232 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2233 return ConstantInt::get(GetCompareTy(LHS),
2234 !CmpInst::isTrueWhenEqual(Pred));
2236 // Fold comparisons for non-escaping pointer even if the allocation call
2237 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2238 // dynamic allocation call could be either of the operands.
2239 Value *MI = nullptr;
2240 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT))
2242 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT))
2244 // FIXME: We should also fold the compare when the pointer escapes, but the
2245 // compare dominates the pointer escape
2246 if (MI && !PointerMayBeCaptured(MI, true, true))
2247 return ConstantInt::get(GetCompareTy(LHS),
2248 CmpInst::isFalseWhenEqual(Pred));
2255 /// Fold an icmp when its operands have i1 scalar type.
2256 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2257 Value *RHS, const SimplifyQuery &Q) {
2258 Type *ITy = GetCompareTy(LHS); // The return type.
2259 Type *OpTy = LHS->getType(); // The operand type.
2260 if (!OpTy->getScalarType()->isIntegerTy(1))
2266 case ICmpInst::ICMP_EQ:
2268 if (match(RHS, m_One()))
2271 case ICmpInst::ICMP_NE:
2273 if (match(RHS, m_Zero()))
2276 case ICmpInst::ICMP_UGT:
2278 if (match(RHS, m_Zero()))
2281 case ICmpInst::ICMP_UGE:
2283 if (match(RHS, m_One()))
2285 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2286 return getTrue(ITy);
2288 case ICmpInst::ICMP_SGE:
2289 /// For signed comparison, the values for an i1 are 0 and -1
2290 /// respectively. This maps into a truth table of:
2291 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2292 /// 0 | 0 | 1 (0 >= 0) | 1
2293 /// 0 | 1 | 1 (0 >= -1) | 1
2294 /// 1 | 0 | 0 (-1 >= 0) | 0
2295 /// 1 | 1 | 1 (-1 >= -1) | 1
2296 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2297 return getTrue(ITy);
2299 case ICmpInst::ICMP_SLT:
2301 if (match(RHS, m_Zero()))
2304 case ICmpInst::ICMP_SLE:
2306 if (match(RHS, m_One()))
2309 case ICmpInst::ICMP_ULE:
2310 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2311 return getTrue(ITy);
2318 /// Try hard to fold icmp with zero RHS because this is a common case.
2319 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2320 Value *RHS, const SimplifyQuery &Q) {
2321 if (!match(RHS, m_Zero()))
2324 Type *ITy = GetCompareTy(LHS); // The return type.
2327 llvm_unreachable("Unknown ICmp predicate!");
2328 case ICmpInst::ICMP_ULT:
2329 return getFalse(ITy);
2330 case ICmpInst::ICMP_UGE:
2331 return getTrue(ITy);
2332 case ICmpInst::ICMP_EQ:
2333 case ICmpInst::ICMP_ULE:
2334 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2335 return getFalse(ITy);
2337 case ICmpInst::ICMP_NE:
2338 case ICmpInst::ICMP_UGT:
2339 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2340 return getTrue(ITy);
2342 case ICmpInst::ICMP_SLT: {
2343 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2344 if (LHSKnown.isNegative())
2345 return getTrue(ITy);
2346 if (LHSKnown.isNonNegative())
2347 return getFalse(ITy);
2350 case ICmpInst::ICMP_SLE: {
2351 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2352 if (LHSKnown.isNegative())
2353 return getTrue(ITy);
2354 if (LHSKnown.isNonNegative() &&
2355 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2356 return getFalse(ITy);
2359 case ICmpInst::ICMP_SGE: {
2360 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2361 if (LHSKnown.isNegative())
2362 return getFalse(ITy);
2363 if (LHSKnown.isNonNegative())
2364 return getTrue(ITy);
2367 case ICmpInst::ICMP_SGT: {
2368 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2369 if (LHSKnown.isNegative())
2370 return getFalse(ITy);
2371 if (LHSKnown.isNonNegative() &&
2372 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2373 return getTrue(ITy);
2381 /// Many binary operators with a constant operand have an easy-to-compute
2382 /// range of outputs. This can be used to fold a comparison to always true or
2384 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2385 unsigned Width = Lower.getBitWidth();
2387 switch (BO.getOpcode()) {
2388 case Instruction::Add:
2389 if (match(BO.getOperand(1), m_APInt(C)) && *C != 0) {
2390 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2391 if (BO.hasNoUnsignedWrap()) {
2392 // 'add nuw x, C' produces [C, UINT_MAX].
2394 } else if (BO.hasNoSignedWrap()) {
2395 if (C->isNegative()) {
2396 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2397 Lower = APInt::getSignedMinValue(Width);
2398 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2400 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2401 Lower = APInt::getSignedMinValue(Width) + *C;
2402 Upper = APInt::getSignedMaxValue(Width) + 1;
2408 case Instruction::And:
2409 if (match(BO.getOperand(1), m_APInt(C)))
2410 // 'and x, C' produces [0, C].
2414 case Instruction::Or:
2415 if (match(BO.getOperand(1), m_APInt(C)))
2416 // 'or x, C' produces [C, UINT_MAX].
2420 case Instruction::AShr:
2421 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2422 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2423 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2424 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2425 } else if (match(BO.getOperand(0), m_APInt(C))) {
2426 unsigned ShiftAmount = Width - 1;
2427 if (*C != 0 && BO.isExact())
2428 ShiftAmount = C->countTrailingZeros();
2429 if (C->isNegative()) {
2430 // 'ashr C, x' produces [C, C >> (Width-1)]
2432 Upper = C->ashr(ShiftAmount) + 1;
2434 // 'ashr C, x' produces [C >> (Width-1), C]
2435 Lower = C->ashr(ShiftAmount);
2441 case Instruction::LShr:
2442 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2443 // 'lshr x, C' produces [0, UINT_MAX >> C].
2444 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2445 } else if (match(BO.getOperand(0), m_APInt(C))) {
2446 // 'lshr C, x' produces [C >> (Width-1), C].
2447 unsigned ShiftAmount = Width - 1;
2448 if (*C != 0 && BO.isExact())
2449 ShiftAmount = C->countTrailingZeros();
2450 Lower = C->lshr(ShiftAmount);
2455 case Instruction::Shl:
2456 if (match(BO.getOperand(0), m_APInt(C))) {
2457 if (BO.hasNoUnsignedWrap()) {
2458 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2460 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2461 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2462 if (C->isNegative()) {
2463 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2464 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2465 Lower = C->shl(ShiftAmount);
2468 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2469 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2471 Upper = C->shl(ShiftAmount) + 1;
2477 case Instruction::SDiv:
2478 if (match(BO.getOperand(1), m_APInt(C))) {
2479 APInt IntMin = APInt::getSignedMinValue(Width);
2480 APInt IntMax = APInt::getSignedMaxValue(Width);
2481 if (C->isAllOnesValue()) {
2482 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2483 // where C != -1 and C != 0 and C != 1
2486 } else if (C->countLeadingZeros() < Width - 1) {
2487 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2488 // where C != -1 and C != 0 and C != 1
2489 Lower = IntMin.sdiv(*C);
2490 Upper = IntMax.sdiv(*C);
2491 if (Lower.sgt(Upper))
2492 std::swap(Lower, Upper);
2494 assert(Upper != Lower && "Upper part of range has wrapped!");
2496 } else if (match(BO.getOperand(0), m_APInt(C))) {
2497 if (C->isMinSignedValue()) {
2498 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2500 Upper = Lower.lshr(1) + 1;
2502 // 'sdiv C, x' produces [-|C|, |C|].
2503 Upper = C->abs() + 1;
2504 Lower = (-Upper) + 1;
2509 case Instruction::UDiv:
2510 if (match(BO.getOperand(1), m_APInt(C)) && *C != 0) {
2511 // 'udiv x, C' produces [0, UINT_MAX / C].
2512 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2513 } else if (match(BO.getOperand(0), m_APInt(C))) {
2514 // 'udiv C, x' produces [0, C].
2519 case Instruction::SRem:
2520 if (match(BO.getOperand(1), m_APInt(C))) {
2521 // 'srem x, C' produces (-|C|, |C|).
2523 Lower = (-Upper) + 1;
2527 case Instruction::URem:
2528 if (match(BO.getOperand(1), m_APInt(C)))
2529 // 'urem x, C' produces [0, C).
2538 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2541 if (!match(RHS, m_APInt(C)))
2544 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2545 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2546 if (RHS_CR.isEmptySet())
2547 return ConstantInt::getFalse(GetCompareTy(RHS));
2548 if (RHS_CR.isFullSet())
2549 return ConstantInt::getTrue(GetCompareTy(RHS));
2551 // Find the range of possible values for binary operators.
2552 unsigned Width = C->getBitWidth();
2553 APInt Lower = APInt(Width, 0);
2554 APInt Upper = APInt(Width, 0);
2555 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2556 setLimitsForBinOp(*BO, Lower, Upper);
2558 ConstantRange LHS_CR =
2559 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2561 if (auto *I = dyn_cast<Instruction>(LHS))
2562 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2563 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2565 if (!LHS_CR.isFullSet()) {
2566 if (RHS_CR.contains(LHS_CR))
2567 return ConstantInt::getTrue(GetCompareTy(RHS));
2568 if (RHS_CR.inverse().contains(LHS_CR))
2569 return ConstantInt::getFalse(GetCompareTy(RHS));
2575 /// TODO: A large part of this logic is duplicated in InstCombine's
2576 /// foldICmpBinOp(). We should be able to share that and avoid the code
2578 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2579 Value *RHS, const SimplifyQuery &Q,
2580 unsigned MaxRecurse) {
2581 Type *ITy = GetCompareTy(LHS); // The return type.
2583 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2584 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2585 if (MaxRecurse && (LBO || RBO)) {
2586 // Analyze the case when either LHS or RHS is an add instruction.
2587 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2588 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2589 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2590 if (LBO && LBO->getOpcode() == Instruction::Add) {
2591 A = LBO->getOperand(0);
2592 B = LBO->getOperand(1);
2594 ICmpInst::isEquality(Pred) ||
2595 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2596 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2598 if (RBO && RBO->getOpcode() == Instruction::Add) {
2599 C = RBO->getOperand(0);
2600 D = RBO->getOperand(1);
2602 ICmpInst::isEquality(Pred) ||
2603 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2604 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2607 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2608 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2609 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2610 Constant::getNullValue(RHS->getType()), Q,
2614 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2615 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2617 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2618 C == LHS ? D : C, Q, MaxRecurse - 1))
2621 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2622 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2624 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2627 // C + B == C + D -> B == D
2630 } else if (A == D) {
2631 // D + B == C + D -> B == C
2634 } else if (B == C) {
2635 // A + C == C + D -> A == D
2640 // A + D == C + D -> A == C
2644 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2651 // icmp pred (or X, Y), X
2652 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2653 if (Pred == ICmpInst::ICMP_ULT)
2654 return getFalse(ITy);
2655 if (Pred == ICmpInst::ICMP_UGE)
2656 return getTrue(ITy);
2658 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2659 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2660 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2661 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2662 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2663 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2664 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2667 // icmp pred X, (or X, Y)
2668 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2669 if (Pred == ICmpInst::ICMP_ULE)
2670 return getTrue(ITy);
2671 if (Pred == ICmpInst::ICMP_UGT)
2672 return getFalse(ITy);
2674 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2675 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2676 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2677 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2678 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2679 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2680 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2685 // icmp pred (and X, Y), X
2686 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)),
2687 m_And(m_Specific(RHS), m_Value())))) {
2688 if (Pred == ICmpInst::ICMP_UGT)
2689 return getFalse(ITy);
2690 if (Pred == ICmpInst::ICMP_ULE)
2691 return getTrue(ITy);
2693 // icmp pred X, (and X, Y)
2694 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)),
2695 m_And(m_Specific(LHS), m_Value())))) {
2696 if (Pred == ICmpInst::ICMP_UGE)
2697 return getTrue(ITy);
2698 if (Pred == ICmpInst::ICMP_ULT)
2699 return getFalse(ITy);
2702 // 0 - (zext X) pred C
2703 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2704 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2705 if (RHSC->getValue().isStrictlyPositive()) {
2706 if (Pred == ICmpInst::ICMP_SLT)
2707 return ConstantInt::getTrue(RHSC->getContext());
2708 if (Pred == ICmpInst::ICMP_SGE)
2709 return ConstantInt::getFalse(RHSC->getContext());
2710 if (Pred == ICmpInst::ICMP_EQ)
2711 return ConstantInt::getFalse(RHSC->getContext());
2712 if (Pred == ICmpInst::ICMP_NE)
2713 return ConstantInt::getTrue(RHSC->getContext());
2715 if (RHSC->getValue().isNonNegative()) {
2716 if (Pred == ICmpInst::ICMP_SLE)
2717 return ConstantInt::getTrue(RHSC->getContext());
2718 if (Pred == ICmpInst::ICMP_SGT)
2719 return ConstantInt::getFalse(RHSC->getContext());
2724 // icmp pred (urem X, Y), Y
2725 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2729 case ICmpInst::ICMP_SGT:
2730 case ICmpInst::ICMP_SGE: {
2731 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2732 if (!Known.isNonNegative())
2736 case ICmpInst::ICMP_EQ:
2737 case ICmpInst::ICMP_UGT:
2738 case ICmpInst::ICMP_UGE:
2739 return getFalse(ITy);
2740 case ICmpInst::ICMP_SLT:
2741 case ICmpInst::ICMP_SLE: {
2742 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2743 if (!Known.isNonNegative())
2747 case ICmpInst::ICMP_NE:
2748 case ICmpInst::ICMP_ULT:
2749 case ICmpInst::ICMP_ULE:
2750 return getTrue(ITy);
2754 // icmp pred X, (urem Y, X)
2755 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2759 case ICmpInst::ICMP_SGT:
2760 case ICmpInst::ICMP_SGE: {
2761 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2762 if (!Known.isNonNegative())
2766 case ICmpInst::ICMP_NE:
2767 case ICmpInst::ICMP_UGT:
2768 case ICmpInst::ICMP_UGE:
2769 return getTrue(ITy);
2770 case ICmpInst::ICMP_SLT:
2771 case ICmpInst::ICMP_SLE: {
2772 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2773 if (!Known.isNonNegative())
2777 case ICmpInst::ICMP_EQ:
2778 case ICmpInst::ICMP_ULT:
2779 case ICmpInst::ICMP_ULE:
2780 return getFalse(ITy);
2786 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2787 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2788 // icmp pred (X op Y), X
2789 if (Pred == ICmpInst::ICMP_UGT)
2790 return getFalse(ITy);
2791 if (Pred == ICmpInst::ICMP_ULE)
2792 return getTrue(ITy);
2797 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2798 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2799 // icmp pred X, (X op Y)
2800 if (Pred == ICmpInst::ICMP_ULT)
2801 return getFalse(ITy);
2802 if (Pred == ICmpInst::ICMP_UGE)
2803 return getTrue(ITy);
2810 // where CI2 is a power of 2 and CI isn't
2811 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2812 const APInt *CI2Val, *CIVal = &CI->getValue();
2813 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2814 CI2Val->isPowerOf2()) {
2815 if (!CIVal->isPowerOf2()) {
2816 // CI2 << X can equal zero in some circumstances,
2817 // this simplification is unsafe if CI is zero.
2819 // We know it is safe if:
2820 // - The shift is nsw, we can't shift out the one bit.
2821 // - The shift is nuw, we can't shift out the one bit.
2824 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2825 *CI2Val == 1 || !CI->isZero()) {
2826 if (Pred == ICmpInst::ICMP_EQ)
2827 return ConstantInt::getFalse(RHS->getContext());
2828 if (Pred == ICmpInst::ICMP_NE)
2829 return ConstantInt::getTrue(RHS->getContext());
2832 if (CIVal->isSignMask() && *CI2Val == 1) {
2833 if (Pred == ICmpInst::ICMP_UGT)
2834 return ConstantInt::getFalse(RHS->getContext());
2835 if (Pred == ICmpInst::ICMP_ULE)
2836 return ConstantInt::getTrue(RHS->getContext());
2841 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2842 LBO->getOperand(1) == RBO->getOperand(1)) {
2843 switch (LBO->getOpcode()) {
2846 case Instruction::UDiv:
2847 case Instruction::LShr:
2848 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2850 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2851 RBO->getOperand(0), Q, MaxRecurse - 1))
2854 case Instruction::SDiv:
2855 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2857 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2858 RBO->getOperand(0), Q, MaxRecurse - 1))
2861 case Instruction::AShr:
2862 if (!LBO->isExact() || !RBO->isExact())
2864 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2865 RBO->getOperand(0), Q, MaxRecurse - 1))
2868 case Instruction::Shl: {
2869 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2870 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2873 if (!NSW && ICmpInst::isSigned(Pred))
2875 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2876 RBO->getOperand(0), Q, MaxRecurse - 1))
2885 /// Simplify integer comparisons where at least one operand of the compare
2886 /// matches an integer min/max idiom.
2887 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2888 Value *RHS, const SimplifyQuery &Q,
2889 unsigned MaxRecurse) {
2890 Type *ITy = GetCompareTy(LHS); // The return type.
2892 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2893 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2895 // Signed variants on "max(a,b)>=a -> true".
2896 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2898 std::swap(A, B); // smax(A, B) pred A.
2899 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2900 // We analyze this as smax(A, B) pred A.
2902 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2903 (A == LHS || B == LHS)) {
2905 std::swap(A, B); // A pred smax(A, B).
2906 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2907 // We analyze this as smax(A, B) swapped-pred A.
2908 P = CmpInst::getSwappedPredicate(Pred);
2909 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2910 (A == RHS || B == RHS)) {
2912 std::swap(A, B); // smin(A, B) pred A.
2913 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2914 // We analyze this as smax(-A, -B) swapped-pred -A.
2915 // Note that we do not need to actually form -A or -B thanks to EqP.
2916 P = CmpInst::getSwappedPredicate(Pred);
2917 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2918 (A == LHS || B == LHS)) {
2920 std::swap(A, B); // A pred smin(A, B).
2921 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2922 // We analyze this as smax(-A, -B) pred -A.
2923 // Note that we do not need to actually form -A or -B thanks to EqP.
2926 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2927 // Cases correspond to "max(A, B) p A".
2931 case CmpInst::ICMP_EQ:
2932 case CmpInst::ICMP_SLE:
2933 // Equivalent to "A EqP B". This may be the same as the condition tested
2934 // in the max/min; if so, we can just return that.
2935 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2937 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2939 // Otherwise, see if "A EqP B" simplifies.
2941 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2944 case CmpInst::ICMP_NE:
2945 case CmpInst::ICMP_SGT: {
2946 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2947 // Equivalent to "A InvEqP B". This may be the same as the condition
2948 // tested in the max/min; if so, we can just return that.
2949 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2951 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2953 // Otherwise, see if "A InvEqP B" simplifies.
2955 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2959 case CmpInst::ICMP_SGE:
2961 return getTrue(ITy);
2962 case CmpInst::ICMP_SLT:
2964 return getFalse(ITy);
2968 // Unsigned variants on "max(a,b)>=a -> true".
2969 P = CmpInst::BAD_ICMP_PREDICATE;
2970 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2972 std::swap(A, B); // umax(A, B) pred A.
2973 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2974 // We analyze this as umax(A, B) pred A.
2976 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2977 (A == LHS || B == LHS)) {
2979 std::swap(A, B); // A pred umax(A, B).
2980 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2981 // We analyze this as umax(A, B) swapped-pred A.
2982 P = CmpInst::getSwappedPredicate(Pred);
2983 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2984 (A == RHS || B == RHS)) {
2986 std::swap(A, B); // umin(A, B) pred A.
2987 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2988 // We analyze this as umax(-A, -B) swapped-pred -A.
2989 // Note that we do not need to actually form -A or -B thanks to EqP.
2990 P = CmpInst::getSwappedPredicate(Pred);
2991 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2992 (A == LHS || B == LHS)) {
2994 std::swap(A, B); // A pred umin(A, B).
2995 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2996 // We analyze this as umax(-A, -B) pred -A.
2997 // Note that we do not need to actually form -A or -B thanks to EqP.
3000 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3001 // Cases correspond to "max(A, B) p A".
3005 case CmpInst::ICMP_EQ:
3006 case CmpInst::ICMP_ULE:
3007 // Equivalent to "A EqP B". This may be the same as the condition tested
3008 // in the max/min; if so, we can just return that.
3009 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3011 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3013 // Otherwise, see if "A EqP B" simplifies.
3015 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3018 case CmpInst::ICMP_NE:
3019 case CmpInst::ICMP_UGT: {
3020 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3021 // Equivalent to "A InvEqP B". This may be the same as the condition
3022 // tested in the max/min; if so, we can just return that.
3023 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3025 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3027 // Otherwise, see if "A InvEqP B" simplifies.
3029 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3033 case CmpInst::ICMP_UGE:
3035 return getTrue(ITy);
3036 case CmpInst::ICMP_ULT:
3038 return getFalse(ITy);
3042 // Variants on "max(x,y) >= min(x,z)".
3044 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3045 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3046 (A == C || A == D || B == C || B == D)) {
3047 // max(x, ?) pred min(x, ?).
3048 if (Pred == CmpInst::ICMP_SGE)
3050 return getTrue(ITy);
3051 if (Pred == CmpInst::ICMP_SLT)
3053 return getFalse(ITy);
3054 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3055 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3056 (A == C || A == D || B == C || B == D)) {
3057 // min(x, ?) pred max(x, ?).
3058 if (Pred == CmpInst::ICMP_SLE)
3060 return getTrue(ITy);
3061 if (Pred == CmpInst::ICMP_SGT)
3063 return getFalse(ITy);
3064 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3065 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3066 (A == C || A == D || B == C || B == D)) {
3067 // max(x, ?) pred min(x, ?).
3068 if (Pred == CmpInst::ICMP_UGE)
3070 return getTrue(ITy);
3071 if (Pred == CmpInst::ICMP_ULT)
3073 return getFalse(ITy);
3074 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3075 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3076 (A == C || A == D || B == C || B == D)) {
3077 // min(x, ?) pred max(x, ?).
3078 if (Pred == CmpInst::ICMP_ULE)
3080 return getTrue(ITy);
3081 if (Pred == CmpInst::ICMP_UGT)
3083 return getFalse(ITy);
3089 /// Given operands for an ICmpInst, see if we can fold the result.
3090 /// If not, this returns null.
3091 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3092 const SimplifyQuery &Q, unsigned MaxRecurse) {
3093 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3094 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3096 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3097 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3098 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3100 // If we have a constant, make sure it is on the RHS.
3101 std::swap(LHS, RHS);
3102 Pred = CmpInst::getSwappedPredicate(Pred);
3105 Type *ITy = GetCompareTy(LHS); // The return type.
3107 // icmp X, X -> true/false
3108 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3109 // because X could be 0.
3110 if (LHS == RHS || isa<UndefValue>(RHS))
3111 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3113 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3116 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3119 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3122 // If both operands have range metadata, use the metadata
3123 // to simplify the comparison.
3124 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3125 auto RHS_Instr = cast<Instruction>(RHS);
3126 auto LHS_Instr = cast<Instruction>(LHS);
3128 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3129 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3130 auto RHS_CR = getConstantRangeFromMetadata(
3131 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3132 auto LHS_CR = getConstantRangeFromMetadata(
3133 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3135 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3136 if (Satisfied_CR.contains(LHS_CR))
3137 return ConstantInt::getTrue(RHS->getContext());
3139 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3140 CmpInst::getInversePredicate(Pred), RHS_CR);
3141 if (InversedSatisfied_CR.contains(LHS_CR))
3142 return ConstantInt::getFalse(RHS->getContext());
3146 // Compare of cast, for example (zext X) != 0 -> X != 0
3147 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3148 Instruction *LI = cast<CastInst>(LHS);
3149 Value *SrcOp = LI->getOperand(0);
3150 Type *SrcTy = SrcOp->getType();
3151 Type *DstTy = LI->getType();
3153 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3154 // if the integer type is the same size as the pointer type.
3155 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3156 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3157 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3158 // Transfer the cast to the constant.
3159 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3160 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3163 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3164 if (RI->getOperand(0)->getType() == SrcTy)
3165 // Compare without the cast.
3166 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3172 if (isa<ZExtInst>(LHS)) {
3173 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3175 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3176 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3177 // Compare X and Y. Note that signed predicates become unsigned.
3178 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3179 SrcOp, RI->getOperand(0), Q,
3183 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3184 // too. If not, then try to deduce the result of the comparison.
3185 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3186 // Compute the constant that would happen if we truncated to SrcTy then
3187 // reextended to DstTy.
3188 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3189 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3191 // If the re-extended constant didn't change then this is effectively
3192 // also a case of comparing two zero-extended values.
3193 if (RExt == CI && MaxRecurse)
3194 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3195 SrcOp, Trunc, Q, MaxRecurse-1))
3198 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3199 // there. Use this to work out the result of the comparison.
3202 default: llvm_unreachable("Unknown ICmp predicate!");
3204 case ICmpInst::ICMP_EQ:
3205 case ICmpInst::ICMP_UGT:
3206 case ICmpInst::ICMP_UGE:
3207 return ConstantInt::getFalse(CI->getContext());
3209 case ICmpInst::ICMP_NE:
3210 case ICmpInst::ICMP_ULT:
3211 case ICmpInst::ICMP_ULE:
3212 return ConstantInt::getTrue(CI->getContext());
3214 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3215 // is non-negative then LHS <s RHS.
3216 case ICmpInst::ICMP_SGT:
3217 case ICmpInst::ICMP_SGE:
3218 return CI->getValue().isNegative() ?
3219 ConstantInt::getTrue(CI->getContext()) :
3220 ConstantInt::getFalse(CI->getContext());
3222 case ICmpInst::ICMP_SLT:
3223 case ICmpInst::ICMP_SLE:
3224 return CI->getValue().isNegative() ?
3225 ConstantInt::getFalse(CI->getContext()) :
3226 ConstantInt::getTrue(CI->getContext());
3232 if (isa<SExtInst>(LHS)) {
3233 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3235 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3236 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3237 // Compare X and Y. Note that the predicate does not change.
3238 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3242 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3243 // too. If not, then try to deduce the result of the comparison.
3244 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3245 // Compute the constant that would happen if we truncated to SrcTy then
3246 // reextended to DstTy.
3247 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3248 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3250 // If the re-extended constant didn't change then this is effectively
3251 // also a case of comparing two sign-extended values.
3252 if (RExt == CI && MaxRecurse)
3253 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3256 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3257 // bits there. Use this to work out the result of the comparison.
3260 default: llvm_unreachable("Unknown ICmp predicate!");
3261 case ICmpInst::ICMP_EQ:
3262 return ConstantInt::getFalse(CI->getContext());
3263 case ICmpInst::ICMP_NE:
3264 return ConstantInt::getTrue(CI->getContext());
3266 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3268 case ICmpInst::ICMP_SGT:
3269 case ICmpInst::ICMP_SGE:
3270 return CI->getValue().isNegative() ?
3271 ConstantInt::getTrue(CI->getContext()) :
3272 ConstantInt::getFalse(CI->getContext());
3273 case ICmpInst::ICMP_SLT:
3274 case ICmpInst::ICMP_SLE:
3275 return CI->getValue().isNegative() ?
3276 ConstantInt::getFalse(CI->getContext()) :
3277 ConstantInt::getTrue(CI->getContext());
3279 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3281 case ICmpInst::ICMP_UGT:
3282 case ICmpInst::ICMP_UGE:
3283 // Comparison is true iff the LHS <s 0.
3285 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3286 Constant::getNullValue(SrcTy),
3290 case ICmpInst::ICMP_ULT:
3291 case ICmpInst::ICMP_ULE:
3292 // Comparison is true iff the LHS >=s 0.
3294 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3295 Constant::getNullValue(SrcTy),
3305 // icmp eq|ne X, Y -> false|true if X != Y
3306 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
3307 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3308 LLVMContext &Ctx = LHS->getType()->getContext();
3309 return Pred == ICmpInst::ICMP_NE ?
3310 ConstantInt::getTrue(Ctx) : ConstantInt::getFalse(Ctx);
3313 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3316 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3319 // Simplify comparisons of related pointers using a powerful, recursive
3320 // GEP-walk when we have target data available..
3321 if (LHS->getType()->isPointerTy())
3322 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS))
3324 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3325 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3326 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3327 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3328 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3329 Q.DL.getTypeSizeInBits(CRHS->getType()))
3330 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI,
3331 CLHS->getPointerOperand(),
3332 CRHS->getPointerOperand()))
3335 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3336 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3337 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3338 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3339 (ICmpInst::isEquality(Pred) ||
3340 (GLHS->isInBounds() && GRHS->isInBounds() &&
3341 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3342 // The bases are equal and the indices are constant. Build a constant
3343 // expression GEP with the same indices and a null base pointer to see
3344 // what constant folding can make out of it.
3345 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3346 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3347 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3348 GLHS->getSourceElementType(), Null, IndicesLHS);
3350 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3351 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3352 GLHS->getSourceElementType(), Null, IndicesRHS);
3353 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3358 // If a bit is known to be zero for A and known to be one for B,
3359 // then A and B cannot be equal.
3360 if (ICmpInst::isEquality(Pred)) {
3361 const APInt *RHSVal;
3362 if (match(RHS, m_APInt(RHSVal))) {
3363 unsigned BitWidth = RHSVal->getBitWidth();
3364 KnownBits LHSKnown(BitWidth);
3365 computeKnownBits(LHS, LHSKnown, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
3366 if (LHSKnown.Zero.intersects(*RHSVal) ||
3367 !LHSKnown.One.isSubsetOf(*RHSVal))
3368 return Pred == ICmpInst::ICMP_EQ ? ConstantInt::getFalse(ITy)
3369 : ConstantInt::getTrue(ITy);
3373 // If the comparison is with the result of a select instruction, check whether
3374 // comparing with either branch of the select always yields the same value.
3375 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3376 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3379 // If the comparison is with the result of a phi instruction, check whether
3380 // doing the compare with each incoming phi value yields a common result.
3381 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3382 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3388 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3389 const SimplifyQuery &Q) {
3390 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3393 /// Given operands for an FCmpInst, see if we can fold the result.
3394 /// If not, this returns null.
3395 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3396 FastMathFlags FMF, const SimplifyQuery &Q,
3397 unsigned MaxRecurse) {
3398 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3399 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3401 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3402 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3403 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3405 // If we have a constant, make sure it is on the RHS.
3406 std::swap(LHS, RHS);
3407 Pred = CmpInst::getSwappedPredicate(Pred);
3410 // Fold trivial predicates.
3411 Type *RetTy = GetCompareTy(LHS);
3412 if (Pred == FCmpInst::FCMP_FALSE)
3413 return getFalse(RetTy);
3414 if (Pred == FCmpInst::FCMP_TRUE)
3415 return getTrue(RetTy);
3417 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3419 if (Pred == FCmpInst::FCMP_UNO)
3420 return getFalse(RetTy);
3421 if (Pred == FCmpInst::FCMP_ORD)
3422 return getTrue(RetTy);
3425 // fcmp pred x, undef and fcmp pred undef, x
3426 // fold to true if unordered, false if ordered
3427 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3428 // Choosing NaN for the undef will always make unordered comparison succeed
3429 // and ordered comparison fail.
3430 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3433 // fcmp x,x -> true/false. Not all compares are foldable.
3435 if (CmpInst::isTrueWhenEqual(Pred))
3436 return getTrue(RetTy);
3437 if (CmpInst::isFalseWhenEqual(Pred))
3438 return getFalse(RetTy);
3441 // Handle fcmp with constant RHS
3442 const ConstantFP *CFP = nullptr;
3443 if (const auto *RHSC = dyn_cast<Constant>(RHS)) {
3444 if (RHS->getType()->isVectorTy())
3445 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue());
3447 CFP = dyn_cast<ConstantFP>(RHSC);
3450 // If the constant is a nan, see if we can fold the comparison based on it.
3451 if (CFP->getValueAPF().isNaN()) {
3452 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3453 return getFalse(RetTy);
3454 assert(FCmpInst::isUnordered(Pred) &&
3455 "Comparison must be either ordered or unordered!");
3456 // True if unordered.
3457 return getTrue(RetTy);
3459 // Check whether the constant is an infinity.
3460 if (CFP->getValueAPF().isInfinity()) {
3461 if (CFP->getValueAPF().isNegative()) {
3463 case FCmpInst::FCMP_OLT:
3464 // No value is ordered and less than negative infinity.
3465 return getFalse(RetTy);
3466 case FCmpInst::FCMP_UGE:
3467 // All values are unordered with or at least negative infinity.
3468 return getTrue(RetTy);
3474 case FCmpInst::FCMP_OGT:
3475 // No value is ordered and greater than infinity.
3476 return getFalse(RetTy);
3477 case FCmpInst::FCMP_ULE:
3478 // All values are unordered with and at most infinity.
3479 return getTrue(RetTy);
3485 if (CFP->getValueAPF().isZero()) {
3487 case FCmpInst::FCMP_UGE:
3488 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3489 return getTrue(RetTy);
3491 case FCmpInst::FCMP_OLT:
3493 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3494 return getFalse(RetTy);
3502 // If the comparison is with the result of a select instruction, check whether
3503 // comparing with either branch of the select always yields the same value.
3504 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3505 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3508 // If the comparison is with the result of a phi instruction, check whether
3509 // doing the compare with each incoming phi value yields a common result.
3510 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3511 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3517 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3518 FastMathFlags FMF, const SimplifyQuery &Q) {
3519 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3522 /// See if V simplifies when its operand Op is replaced with RepOp.
3523 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3524 const SimplifyQuery &Q,
3525 unsigned MaxRecurse) {
3526 // Trivial replacement.
3530 auto *I = dyn_cast<Instruction>(V);
3534 // If this is a binary operator, try to simplify it with the replaced op.
3535 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3537 // %cmp = icmp eq i32 %x, 2147483647
3538 // %add = add nsw i32 %x, 1
3539 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3541 // We can't replace %sel with %add unless we strip away the flags.
3542 if (isa<OverflowingBinaryOperator>(B))
3543 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3545 if (isa<PossiblyExactOperator>(B))
3550 if (B->getOperand(0) == Op)
3551 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3553 if (B->getOperand(1) == Op)
3554 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3559 // Same for CmpInsts.
3560 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3562 if (C->getOperand(0) == Op)
3563 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3565 if (C->getOperand(1) == Op)
3566 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3571 // TODO: We could hand off more cases to instsimplify here.
3573 // If all operands are constant after substituting Op for RepOp then we can
3574 // constant fold the instruction.
3575 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3576 // Build a list of all constant operands.
3577 SmallVector<Constant *, 8> ConstOps;
3578 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3579 if (I->getOperand(i) == Op)
3580 ConstOps.push_back(CRepOp);
3581 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3582 ConstOps.push_back(COp);
3587 // All operands were constants, fold it.
3588 if (ConstOps.size() == I->getNumOperands()) {
3589 if (CmpInst *C = dyn_cast<CmpInst>(I))
3590 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3591 ConstOps[1], Q.DL, Q.TLI);
3593 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3594 if (!LI->isVolatile())
3595 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3597 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3604 /// Try to simplify a select instruction when its condition operand is an
3605 /// integer comparison where one operand of the compare is a constant.
3606 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3607 const APInt *Y, bool TrueWhenUnset) {
3610 // (X & Y) == 0 ? X & ~Y : X --> X
3611 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3612 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3614 return TrueWhenUnset ? FalseVal : TrueVal;
3616 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3617 // (X & Y) != 0 ? X : X & ~Y --> X
3618 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3620 return TrueWhenUnset ? FalseVal : TrueVal;
3622 if (Y->isPowerOf2()) {
3623 // (X & Y) == 0 ? X | Y : X --> X | Y
3624 // (X & Y) != 0 ? X | Y : X --> X
3625 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3627 return TrueWhenUnset ? TrueVal : FalseVal;
3629 // (X & Y) == 0 ? X : X | Y --> X
3630 // (X & Y) != 0 ? X : X | Y --> X | Y
3631 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3633 return TrueWhenUnset ? TrueVal : FalseVal;
3639 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3641 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal,
3643 bool TrueWhenUnset) {
3644 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
3648 APInt MinSignedValue;
3650 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) {
3651 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0
3652 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0
3653 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits();
3654 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth);
3656 // icmp slt X, 0 <--> icmp ne (and X, C), 0
3657 // icmp sgt X, -1 <--> icmp eq (and X, C), 0
3659 MinSignedValue = APInt::getSignedMinValue(BitWidth);
3662 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue,
3669 /// Try to simplify a select instruction when its condition operand is an
3670 /// integer comparison.
3671 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3672 Value *FalseVal, const SimplifyQuery &Q,
3673 unsigned MaxRecurse) {
3674 ICmpInst::Predicate Pred;
3675 Value *CmpLHS, *CmpRHS;
3676 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3679 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring
3680 // decomposeBitTestICmp() might help.
3681 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3684 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3685 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3686 Pred == ICmpInst::ICMP_EQ))
3688 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3689 // Comparing signed-less-than 0 checks if the sign bit is set.
3690 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3693 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3694 // Comparing signed-greater-than -1 checks if the sign bit is not set.
3695 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3700 if (CondVal->hasOneUse()) {
3702 if (match(CmpRHS, m_APInt(C))) {
3703 // X < MIN ? T : F --> F
3704 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3706 // X < MIN ? T : F --> F
3707 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3709 // X > MAX ? T : F --> F
3710 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3712 // X > MAX ? T : F --> F
3713 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3718 // If we have an equality comparison, then we know the value in one of the
3719 // arms of the select. See if substituting this value into the arm and
3720 // simplifying the result yields the same value as the other arm.
3721 if (Pred == ICmpInst::ICMP_EQ) {
3722 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3724 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3727 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3729 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3732 } else if (Pred == ICmpInst::ICMP_NE) {
3733 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3735 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3738 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3740 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3748 /// Given operands for a SelectInst, see if we can fold the result.
3749 /// If not, this returns null.
3750 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3751 Value *FalseVal, const SimplifyQuery &Q,
3752 unsigned MaxRecurse) {
3753 // select true, X, Y -> X
3754 // select false, X, Y -> Y
3755 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3756 if (CB->isAllOnesValue())
3758 if (CB->isNullValue())
3762 // select C, X, X -> X
3763 if (TrueVal == FalseVal)
3766 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3767 if (isa<Constant>(FalseVal))
3771 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3773 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3777 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3783 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3784 const SimplifyQuery &Q) {
3785 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3788 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3789 /// If not, this returns null.
3790 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3791 const SimplifyQuery &Q, unsigned) {
3792 // The type of the GEP pointer operand.
3794 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3796 // getelementptr P -> P.
3797 if (Ops.size() == 1)
3800 // Compute the (pointer) type returned by the GEP instruction.
3801 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3802 Type *GEPTy = PointerType::get(LastType, AS);
3803 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3804 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3805 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3806 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3808 if (isa<UndefValue>(Ops[0]))
3809 return UndefValue::get(GEPTy);
3811 if (Ops.size() == 2) {
3812 // getelementptr P, 0 -> P.
3813 if (match(Ops[1], m_Zero()))
3817 if (Ty->isSized()) {
3820 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3821 // getelementptr P, N -> P if P points to a type of zero size.
3822 if (TyAllocSize == 0)
3825 // The following transforms are only safe if the ptrtoint cast
3826 // doesn't truncate the pointers.
3827 if (Ops[1]->getType()->getScalarSizeInBits() ==
3828 Q.DL.getPointerSizeInBits(AS)) {
3829 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3830 if (match(P, m_Zero()))
3831 return Constant::getNullValue(GEPTy);
3833 if (match(P, m_PtrToInt(m_Value(Temp))))
3834 if (Temp->getType() == GEPTy)
3839 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3840 if (TyAllocSize == 1 &&
3841 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3842 if (Value *R = PtrToIntOrZero(P))
3845 // getelementptr V, (ashr (sub P, V), C) -> Q
3846 // if P points to a type of size 1 << C.
3848 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3849 m_ConstantInt(C))) &&
3850 TyAllocSize == 1ULL << C)
3851 if (Value *R = PtrToIntOrZero(P))
3854 // getelementptr V, (sdiv (sub P, V), C) -> Q
3855 // if P points to a type of size C.
3857 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3858 m_SpecificInt(TyAllocSize))))
3859 if (Value *R = PtrToIntOrZero(P))
3865 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3866 all_of(Ops.slice(1).drop_back(1),
3867 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3869 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3870 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3871 APInt BasePtrOffset(PtrWidth, 0);
3872 Value *StrippedBasePtr =
3873 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3876 // gep (gep V, C), (sub 0, V) -> C
3877 if (match(Ops.back(),
3878 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3879 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3880 return ConstantExpr::getIntToPtr(CI, GEPTy);
3882 // gep (gep V, C), (xor V, -1) -> C-1
3883 if (match(Ops.back(),
3884 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3885 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3886 return ConstantExpr::getIntToPtr(CI, GEPTy);
3891 // Check to see if this is constant foldable.
3892 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
3893 if (!isa<Constant>(Ops[i]))
3896 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3900 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3901 const SimplifyQuery &Q) {
3902 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3905 /// Given operands for an InsertValueInst, see if we can fold the result.
3906 /// If not, this returns null.
3907 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3908 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3910 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3911 if (Constant *CVal = dyn_cast<Constant>(Val))
3912 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3914 // insertvalue x, undef, n -> x
3915 if (match(Val, m_Undef()))
3918 // insertvalue x, (extractvalue y, n), n
3919 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3920 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3921 EV->getIndices() == Idxs) {
3922 // insertvalue undef, (extractvalue y, n), n -> y
3923 if (match(Agg, m_Undef()))
3924 return EV->getAggregateOperand();
3926 // insertvalue y, (extractvalue y, n), n -> y
3927 if (Agg == EV->getAggregateOperand())
3934 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3935 ArrayRef<unsigned> Idxs,
3936 const SimplifyQuery &Q) {
3937 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3940 /// Given operands for an ExtractValueInst, see if we can fold the result.
3941 /// If not, this returns null.
3942 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3943 const SimplifyQuery &, unsigned) {
3944 if (auto *CAgg = dyn_cast<Constant>(Agg))
3945 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3947 // extractvalue x, (insertvalue y, elt, n), n -> elt
3948 unsigned NumIdxs = Idxs.size();
3949 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3950 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3951 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3952 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3953 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3954 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3955 Idxs.slice(0, NumCommonIdxs)) {
3956 if (NumIdxs == NumInsertValueIdxs)
3957 return IVI->getInsertedValueOperand();
3965 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3966 const SimplifyQuery &Q) {
3967 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3970 /// Given operands for an ExtractElementInst, see if we can fold the result.
3971 /// If not, this returns null.
3972 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3974 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3975 if (auto *CIdx = dyn_cast<Constant>(Idx))
3976 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3978 // The index is not relevant if our vector is a splat.
3979 if (auto *Splat = CVec->getSplatValue())
3982 if (isa<UndefValue>(Vec))
3983 return UndefValue::get(Vec->getType()->getVectorElementType());
3986 // If extracting a specified index from the vector, see if we can recursively
3987 // find a previously computed scalar that was inserted into the vector.
3988 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3989 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3995 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3996 const SimplifyQuery &Q) {
3997 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
4000 /// See if we can fold the given phi. If not, returns null.
4001 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4002 // If all of the PHI's incoming values are the same then replace the PHI node
4003 // with the common value.
4004 Value *CommonValue = nullptr;
4005 bool HasUndefInput = false;
4006 for (Value *Incoming : PN->incoming_values()) {
4007 // If the incoming value is the phi node itself, it can safely be skipped.
4008 if (Incoming == PN) continue;
4009 if (isa<UndefValue>(Incoming)) {
4010 // Remember that we saw an undef value, but otherwise ignore them.
4011 HasUndefInput = true;
4014 if (CommonValue && Incoming != CommonValue)
4015 return nullptr; // Not the same, bail out.
4016 CommonValue = Incoming;
4019 // If CommonValue is null then all of the incoming values were either undef or
4020 // equal to the phi node itself.
4022 return UndefValue::get(PN->getType());
4024 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4025 // instruction, we cannot return X as the result of the PHI node unless it
4026 // dominates the PHI block.
4028 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4033 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4034 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4035 if (auto *C = dyn_cast<Constant>(Op))
4036 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4038 if (auto *CI = dyn_cast<CastInst>(Op)) {
4039 auto *Src = CI->getOperand(0);
4040 Type *SrcTy = Src->getType();
4041 Type *MidTy = CI->getType();
4043 if (Src->getType() == Ty) {
4044 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4045 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4047 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4049 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4051 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4052 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4053 SrcIntPtrTy, MidIntPtrTy,
4054 DstIntPtrTy) == Instruction::BitCast)
4060 if (CastOpc == Instruction::BitCast)
4061 if (Op->getType() == Ty)
4067 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4068 const SimplifyQuery &Q) {
4069 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4072 /// For the given destination element of a shuffle, peek through shuffles to
4073 /// match a root vector source operand that contains that element in the same
4074 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4075 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4076 int MaskVal, Value *RootVec,
4077 unsigned MaxRecurse) {
4081 // Bail out if any mask value is undefined. That kind of shuffle may be
4082 // simplified further based on demanded bits or other folds.
4086 // The mask value chooses which source operand we need to look at next.
4087 int InVecNumElts = Op0->getType()->getVectorNumElements();
4088 int RootElt = MaskVal;
4089 Value *SourceOp = Op0;
4090 if (MaskVal >= InVecNumElts) {
4091 RootElt = MaskVal - InVecNumElts;
4095 // If the source operand is a shuffle itself, look through it to find the
4096 // matching root vector.
4097 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4098 return foldIdentityShuffles(
4099 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4100 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4103 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4106 // The source operand is not a shuffle. Initialize the root vector value for
4107 // this shuffle if that has not been done yet.
4111 // Give up as soon as a source operand does not match the existing root value.
4112 if (RootVec != SourceOp)
4115 // The element must be coming from the same lane in the source vector
4116 // (although it may have crossed lanes in intermediate shuffles).
4117 if (RootElt != DestElt)
4123 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4124 Type *RetTy, const SimplifyQuery &Q,
4125 unsigned MaxRecurse) {
4126 if (isa<UndefValue>(Mask))
4127 return UndefValue::get(RetTy);
4129 Type *InVecTy = Op0->getType();
4130 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4131 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4133 SmallVector<int, 32> Indices;
4134 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4135 assert(MaskNumElts == Indices.size() &&
4136 "Size of Indices not same as number of mask elements?");
4138 // Canonicalization: If mask does not select elements from an input vector,
4139 // replace that input vector with undef.
4140 bool MaskSelects0 = false, MaskSelects1 = false;
4141 for (unsigned i = 0; i != MaskNumElts; ++i) {
4142 if (Indices[i] == -1)
4144 if ((unsigned)Indices[i] < InVecNumElts)
4145 MaskSelects0 = true;
4147 MaskSelects1 = true;
4150 Op0 = UndefValue::get(InVecTy);
4152 Op1 = UndefValue::get(InVecTy);
4154 auto *Op0Const = dyn_cast<Constant>(Op0);
4155 auto *Op1Const = dyn_cast<Constant>(Op1);
4157 // If all operands are constant, constant fold the shuffle.
4158 if (Op0Const && Op1Const)
4159 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4161 // Canonicalization: if only one input vector is constant, it shall be the
4163 if (Op0Const && !Op1Const) {
4164 std::swap(Op0, Op1);
4165 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4168 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4169 // value type is same as the input vectors' type.
4170 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4171 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4172 OpShuf->getMask()->getSplatValue())
4175 // Don't fold a shuffle with undef mask elements. This may get folded in a
4176 // better way using demanded bits or other analysis.
4177 // TODO: Should we allow this?
4178 if (find(Indices, -1) != Indices.end())
4181 // Check if every element of this shuffle can be mapped back to the
4182 // corresponding element of a single root vector. If so, we don't need this
4183 // shuffle. This handles simple identity shuffles as well as chains of
4184 // shuffles that may widen/narrow and/or move elements across lanes and back.
4185 Value *RootVec = nullptr;
4186 for (unsigned i = 0; i != MaskNumElts; ++i) {
4187 // Note that recursion is limited for each vector element, so if any element
4188 // exceeds the limit, this will fail to simplify.
4190 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4192 // We can't replace a widening/narrowing shuffle with one of its operands.
4193 if (!RootVec || RootVec->getType() != RetTy)
4199 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4200 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4201 Type *RetTy, const SimplifyQuery &Q) {
4202 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4205 //=== Helper functions for higher up the class hierarchy.
4207 /// Given operands for a BinaryOperator, see if we can fold the result.
4208 /// If not, this returns null.
4209 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4210 const SimplifyQuery &Q, unsigned MaxRecurse) {
4212 case Instruction::Add:
4213 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4214 case Instruction::FAdd:
4215 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4216 case Instruction::Sub:
4217 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4218 case Instruction::FSub:
4219 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4220 case Instruction::Mul:
4221 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4222 case Instruction::FMul:
4223 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4224 case Instruction::SDiv:
4225 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4226 case Instruction::UDiv:
4227 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4228 case Instruction::FDiv:
4229 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4230 case Instruction::SRem:
4231 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4232 case Instruction::URem:
4233 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4234 case Instruction::FRem:
4235 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4236 case Instruction::Shl:
4237 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4238 case Instruction::LShr:
4239 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4240 case Instruction::AShr:
4241 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4242 case Instruction::And:
4243 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4244 case Instruction::Or:
4245 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4246 case Instruction::Xor:
4247 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4249 llvm_unreachable("Unexpected opcode");
4253 /// Given operands for a BinaryOperator, see if we can fold the result.
4254 /// If not, this returns null.
4255 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4256 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4257 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4258 const FastMathFlags &FMF, const SimplifyQuery &Q,
4259 unsigned MaxRecurse) {
4261 case Instruction::FAdd:
4262 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4263 case Instruction::FSub:
4264 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4265 case Instruction::FMul:
4266 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4267 case Instruction::FDiv:
4268 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4270 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4274 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4275 const SimplifyQuery &Q) {
4276 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4279 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4280 FastMathFlags FMF, const SimplifyQuery &Q) {
4281 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4284 /// Given operands for a CmpInst, see if we can fold the result.
4285 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4286 const SimplifyQuery &Q, unsigned MaxRecurse) {
4287 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4288 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4289 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4292 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4293 const SimplifyQuery &Q) {
4294 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4297 static bool IsIdempotent(Intrinsic::ID ID) {
4299 default: return false;
4301 // Unary idempotent: f(f(x)) = f(x)
4302 case Intrinsic::fabs:
4303 case Intrinsic::floor:
4304 case Intrinsic::ceil:
4305 case Intrinsic::trunc:
4306 case Intrinsic::rint:
4307 case Intrinsic::nearbyint:
4308 case Intrinsic::round:
4313 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4314 const DataLayout &DL) {
4315 GlobalValue *PtrSym;
4317 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4320 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4321 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4322 Type *Int32PtrTy = Int32Ty->getPointerTo();
4323 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4325 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4326 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4329 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4330 if (OffsetInt % 4 != 0)
4333 Constant *C = ConstantExpr::getGetElementPtr(
4334 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4335 ConstantInt::get(Int64Ty, OffsetInt / 4));
4336 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4340 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4344 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4345 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4350 if (LoadedCE->getOpcode() != Instruction::Sub)
4353 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4354 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4356 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4358 Constant *LoadedRHS = LoadedCE->getOperand(1);
4359 GlobalValue *LoadedRHSSym;
4360 APInt LoadedRHSOffset;
4361 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4363 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4366 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4369 static bool maskIsAllZeroOrUndef(Value *Mask) {
4370 auto *ConstMask = dyn_cast<Constant>(Mask);
4373 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4375 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4377 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4378 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4385 template <typename IterTy>
4386 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4387 const SimplifyQuery &Q, unsigned MaxRecurse) {
4388 Intrinsic::ID IID = F->getIntrinsicID();
4389 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4392 if (NumOperands == 1) {
4393 // Perform idempotent optimizations
4394 if (IsIdempotent(IID)) {
4395 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4396 if (II->getIntrinsicID() == IID)
4402 case Intrinsic::fabs: {
4403 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4413 if (NumOperands == 2) {
4414 Value *LHS = *ArgBegin;
4415 Value *RHS = *(ArgBegin + 1);
4416 Type *ReturnType = F->getReturnType();
4419 case Intrinsic::usub_with_overflow:
4420 case Intrinsic::ssub_with_overflow: {
4421 // X - X -> { 0, false }
4423 return Constant::getNullValue(ReturnType);
4425 // X - undef -> undef
4426 // undef - X -> undef
4427 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4428 return UndefValue::get(ReturnType);
4432 case Intrinsic::uadd_with_overflow:
4433 case Intrinsic::sadd_with_overflow: {
4434 // X + undef -> undef
4435 if (isa<UndefValue>(RHS))
4436 return UndefValue::get(ReturnType);
4440 case Intrinsic::umul_with_overflow:
4441 case Intrinsic::smul_with_overflow: {
4442 // X * 0 -> { 0, false }
4443 if (match(RHS, m_Zero()))
4444 return Constant::getNullValue(ReturnType);
4446 // X * undef -> { 0, false }
4447 if (match(RHS, m_Undef()))
4448 return Constant::getNullValue(ReturnType);
4452 case Intrinsic::load_relative: {
4453 Constant *C0 = dyn_cast<Constant>(LHS);
4454 Constant *C1 = dyn_cast<Constant>(RHS);
4456 return SimplifyRelativeLoad(C0, C1, Q.DL);
4464 // Simplify calls to llvm.masked.load.*
4466 case Intrinsic::masked_load: {
4467 Value *MaskArg = ArgBegin[2];
4468 Value *PassthruArg = ArgBegin[3];
4469 // If the mask is all zeros or undef, the "passthru" argument is the result.
4470 if (maskIsAllZeroOrUndef(MaskArg))
4479 template <typename IterTy>
4480 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
4481 const SimplifyQuery &Q, unsigned MaxRecurse) {
4482 Type *Ty = V->getType();
4483 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4484 Ty = PTy->getElementType();
4485 FunctionType *FTy = cast<FunctionType>(Ty);
4487 // call undef -> undef
4488 // call null -> undef
4489 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4490 return UndefValue::get(FTy->getReturnType());
4492 Function *F = dyn_cast<Function>(V);
4496 if (F->isIntrinsic())
4497 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4500 if (!canConstantFoldCallTo(F))
4503 SmallVector<Constant *, 4> ConstantArgs;
4504 ConstantArgs.reserve(ArgEnd - ArgBegin);
4505 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4506 Constant *C = dyn_cast<Constant>(*I);
4509 ConstantArgs.push_back(C);
4512 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
4515 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
4516 User::op_iterator ArgEnd, const SimplifyQuery &Q) {
4517 return ::SimplifyCall(V, ArgBegin, ArgEnd, Q, RecursionLimit);
4520 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
4521 const SimplifyQuery &Q) {
4522 return ::SimplifyCall(V, Args.begin(), Args.end(), Q, RecursionLimit);
4525 /// See if we can compute a simplified version of this instruction.
4526 /// If not, this returns null.
4528 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4529 OptimizationRemarkEmitter *ORE) {
4530 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4533 switch (I->getOpcode()) {
4535 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4537 case Instruction::FAdd:
4538 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4539 I->getFastMathFlags(), Q);
4541 case Instruction::Add:
4542 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4543 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4544 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4546 case Instruction::FSub:
4547 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4548 I->getFastMathFlags(), Q);
4550 case Instruction::Sub:
4551 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4552 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4553 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4555 case Instruction::FMul:
4556 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4557 I->getFastMathFlags(), Q);
4559 case Instruction::Mul:
4560 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4562 case Instruction::SDiv:
4563 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4565 case Instruction::UDiv:
4566 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4568 case Instruction::FDiv:
4569 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4570 I->getFastMathFlags(), Q);
4572 case Instruction::SRem:
4573 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4575 case Instruction::URem:
4576 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4578 case Instruction::FRem:
4579 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4580 I->getFastMathFlags(), Q);
4582 case Instruction::Shl:
4583 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4584 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4585 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4587 case Instruction::LShr:
4588 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4589 cast<BinaryOperator>(I)->isExact(), Q);
4591 case Instruction::AShr:
4592 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4593 cast<BinaryOperator>(I)->isExact(), Q);
4595 case Instruction::And:
4596 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4598 case Instruction::Or:
4599 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4601 case Instruction::Xor:
4602 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4604 case Instruction::ICmp:
4605 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4606 I->getOperand(0), I->getOperand(1), Q);
4608 case Instruction::FCmp:
4610 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4611 I->getOperand(1), I->getFastMathFlags(), Q);
4613 case Instruction::Select:
4614 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4615 I->getOperand(2), Q);
4617 case Instruction::GetElementPtr: {
4618 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4619 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4623 case Instruction::InsertValue: {
4624 InsertValueInst *IV = cast<InsertValueInst>(I);
4625 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4626 IV->getInsertedValueOperand(),
4627 IV->getIndices(), Q);
4630 case Instruction::ExtractValue: {
4631 auto *EVI = cast<ExtractValueInst>(I);
4632 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4633 EVI->getIndices(), Q);
4636 case Instruction::ExtractElement: {
4637 auto *EEI = cast<ExtractElementInst>(I);
4638 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4639 EEI->getIndexOperand(), Q);
4642 case Instruction::ShuffleVector: {
4643 auto *SVI = cast<ShuffleVectorInst>(I);
4644 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4645 SVI->getMask(), SVI->getType(), Q);
4648 case Instruction::PHI:
4649 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4651 case Instruction::Call: {
4652 CallSite CS(cast<CallInst>(I));
4653 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), Q);
4656 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4657 #include "llvm/IR/Instruction.def"
4658 #undef HANDLE_CAST_INST
4660 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4662 case Instruction::Alloca:
4663 // No simplifications for Alloca and it can't be constant folded.
4668 // In general, it is possible for computeKnownBits to determine all bits in a
4669 // value even when the operands are not all constants.
4670 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4671 unsigned BitWidth = I->getType()->getScalarSizeInBits();
4672 KnownBits Known(BitWidth);
4673 computeKnownBits(I, Known, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4674 if (Known.isConstant())
4675 Result = ConstantInt::get(I->getType(), Known.getConstant());
4678 /// If called on unreachable code, the above logic may report that the
4679 /// instruction simplified to itself. Make life easier for users by
4680 /// detecting that case here, returning a safe value instead.
4681 return Result == I ? UndefValue::get(I->getType()) : Result;
4684 /// \brief Implementation of recursive simplification through an instruction's
4687 /// This is the common implementation of the recursive simplification routines.
4688 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4689 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4690 /// instructions to process and attempt to simplify it using
4691 /// InstructionSimplify.
4693 /// This routine returns 'true' only when *it* simplifies something. The passed
4694 /// in simplified value does not count toward this.
4695 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4696 const TargetLibraryInfo *TLI,
4697 const DominatorTree *DT,
4698 AssumptionCache *AC) {
4699 bool Simplified = false;
4700 SmallSetVector<Instruction *, 8> Worklist;
4701 const DataLayout &DL = I->getModule()->getDataLayout();
4703 // If we have an explicit value to collapse to, do that round of the
4704 // simplification loop by hand initially.
4706 for (User *U : I->users())
4708 Worklist.insert(cast<Instruction>(U));
4710 // Replace the instruction with its simplified value.
4711 I->replaceAllUsesWith(SimpleV);
4713 // Gracefully handle edge cases where the instruction is not wired into any
4715 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4716 !I->mayHaveSideEffects())
4717 I->eraseFromParent();
4722 // Note that we must test the size on each iteration, the worklist can grow.
4723 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4726 // See if this instruction simplifies.
4727 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4733 // Stash away all the uses of the old instruction so we can check them for
4734 // recursive simplifications after a RAUW. This is cheaper than checking all
4735 // uses of To on the recursive step in most cases.
4736 for (User *U : I->users())
4737 Worklist.insert(cast<Instruction>(U));
4739 // Replace the instruction with its simplified value.
4740 I->replaceAllUsesWith(SimpleV);
4742 // Gracefully handle edge cases where the instruction is not wired into any
4744 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4745 !I->mayHaveSideEffects())
4746 I->eraseFromParent();
4751 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4752 const TargetLibraryInfo *TLI,
4753 const DominatorTree *DT,
4754 AssumptionCache *AC) {
4755 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4758 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4759 const TargetLibraryInfo *TLI,
4760 const DominatorTree *DT,
4761 AssumptionCache *AC) {
4762 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4763 assert(SimpleV && "Must provide a simplified value.");
4764 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4768 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4769 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4770 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4771 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4772 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4773 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4774 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4775 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4778 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4779 const DataLayout &DL) {
4780 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4783 template <class T, class... TArgs>
4784 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4786 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4787 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4788 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4789 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4791 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,