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 return DT->dominates(I, P);
109 // Otherwise, if the instruction is in the entry block and is not an invoke,
110 // then it obviously dominates all phi nodes.
111 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
118 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
119 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
120 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
121 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
122 /// Returns the simplified value, or null if no simplification was performed.
123 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
124 Instruction::BinaryOps OpcodeToExpand,
125 const SimplifyQuery &Q, unsigned MaxRecurse) {
126 // Recursion is always used, so bail out at once if we already hit the limit.
130 // Check whether the expression has the form "(A op' B) op C".
131 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
132 if (Op0->getOpcode() == OpcodeToExpand) {
133 // It does! Try turning it into "(A op C) op' (B op C)".
134 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
135 // Do "A op C" and "B op C" both simplify?
136 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
137 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
138 // They do! Return "L op' R" if it simplifies or is already available.
139 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
140 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
141 && L == B && R == A)) {
145 // Otherwise return "L op' R" if it simplifies.
146 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
153 // Check whether the expression has the form "A op (B op' C)".
154 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
155 if (Op1->getOpcode() == OpcodeToExpand) {
156 // It does! Try turning it into "(A op B) op' (A op C)".
157 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
158 // Do "A op B" and "A op C" both simplify?
159 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
160 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
161 // They do! Return "L op' R" if it simplifies or is already available.
162 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
163 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
164 && L == C && R == B)) {
168 // Otherwise return "L op' R" if it simplifies.
169 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
179 /// Generic simplifications for associative binary operations.
180 /// Returns the simpler value, or null if none was found.
181 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
182 Value *LHS, Value *RHS,
183 const SimplifyQuery &Q,
184 unsigned MaxRecurse) {
185 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
187 // Recursion is always used, so bail out at once if we already hit the limit.
191 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
192 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
194 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
195 if (Op0 && Op0->getOpcode() == Opcode) {
196 Value *A = Op0->getOperand(0);
197 Value *B = Op0->getOperand(1);
200 // Does "B op C" simplify?
201 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
202 // It does! Return "A op V" if it simplifies or is already available.
203 // If V equals B then "A op V" is just the LHS.
204 if (V == B) return LHS;
205 // Otherwise return "A op V" if it simplifies.
206 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
213 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
214 if (Op1 && Op1->getOpcode() == Opcode) {
216 Value *B = Op1->getOperand(0);
217 Value *C = Op1->getOperand(1);
219 // Does "A op B" simplify?
220 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
221 // It does! Return "V op C" if it simplifies or is already available.
222 // If V equals B then "V op C" is just the RHS.
223 if (V == B) return RHS;
224 // Otherwise return "V op C" if it simplifies.
225 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
232 // The remaining transforms require commutativity as well as associativity.
233 if (!Instruction::isCommutative(Opcode))
236 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
237 if (Op0 && Op0->getOpcode() == Opcode) {
238 Value *A = Op0->getOperand(0);
239 Value *B = Op0->getOperand(1);
242 // Does "C op A" simplify?
243 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
244 // It does! Return "V op B" if it simplifies or is already available.
245 // If V equals A then "V op B" is just the LHS.
246 if (V == A) return LHS;
247 // Otherwise return "V op B" if it simplifies.
248 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
255 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
256 if (Op1 && Op1->getOpcode() == Opcode) {
258 Value *B = Op1->getOperand(0);
259 Value *C = Op1->getOperand(1);
261 // Does "C op A" simplify?
262 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
263 // It does! Return "B op V" if it simplifies or is already available.
264 // If V equals C then "B op V" is just the RHS.
265 if (V == C) return RHS;
266 // Otherwise return "B op V" if it simplifies.
267 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
277 /// In the case of a binary operation with a select instruction as an operand,
278 /// try to simplify the binop by seeing whether evaluating it on both branches
279 /// of the select results in the same value. Returns the common value if so,
280 /// otherwise returns null.
281 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
282 Value *RHS, const SimplifyQuery &Q,
283 unsigned MaxRecurse) {
284 // Recursion is always used, so bail out at once if we already hit the limit.
289 if (isa<SelectInst>(LHS)) {
290 SI = cast<SelectInst>(LHS);
292 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
293 SI = cast<SelectInst>(RHS);
296 // Evaluate the BinOp on the true and false branches of the select.
300 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
301 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
303 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
304 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
307 // If they simplified to the same value, then return the common value.
308 // If they both failed to simplify then return null.
312 // If one branch simplified to undef, return the other one.
313 if (TV && isa<UndefValue>(TV))
315 if (FV && isa<UndefValue>(FV))
318 // If applying the operation did not change the true and false select values,
319 // then the result of the binop is the select itself.
320 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
323 // If one branch simplified and the other did not, and the simplified
324 // value is equal to the unsimplified one, return the simplified value.
325 // For example, select (cond, X, X & Z) & Z -> X & Z.
326 if ((FV && !TV) || (TV && !FV)) {
327 // Check that the simplified value has the form "X op Y" where "op" is the
328 // same as the original operation.
329 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
330 if (Simplified && Simplified->getOpcode() == Opcode) {
331 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
332 // We already know that "op" is the same as for the simplified value. See
333 // if the operands match too. If so, return the simplified value.
334 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
335 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
336 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
337 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
338 Simplified->getOperand(1) == UnsimplifiedRHS)
340 if (Simplified->isCommutative() &&
341 Simplified->getOperand(1) == UnsimplifiedLHS &&
342 Simplified->getOperand(0) == UnsimplifiedRHS)
350 /// In the case of a comparison with a select instruction, try to simplify the
351 /// comparison by seeing whether both branches of the select result in the same
352 /// value. Returns the common value if so, otherwise returns null.
353 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
354 Value *RHS, const SimplifyQuery &Q,
355 unsigned MaxRecurse) {
356 // Recursion is always used, so bail out at once if we already hit the limit.
360 // Make sure the select is on the LHS.
361 if (!isa<SelectInst>(LHS)) {
363 Pred = CmpInst::getSwappedPredicate(Pred);
365 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
366 SelectInst *SI = cast<SelectInst>(LHS);
367 Value *Cond = SI->getCondition();
368 Value *TV = SI->getTrueValue();
369 Value *FV = SI->getFalseValue();
371 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
372 // Does "cmp TV, RHS" simplify?
373 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
375 // It not only simplified, it simplified to the select condition. Replace
377 TCmp = getTrue(Cond->getType());
379 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
380 // condition then we can replace it with 'true'. Otherwise give up.
381 if (!isSameCompare(Cond, Pred, TV, RHS))
383 TCmp = getTrue(Cond->getType());
386 // Does "cmp FV, RHS" simplify?
387 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
389 // It not only simplified, it simplified to the select condition. Replace
391 FCmp = getFalse(Cond->getType());
393 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
394 // condition then we can replace it with 'false'. Otherwise give up.
395 if (!isSameCompare(Cond, Pred, FV, RHS))
397 FCmp = getFalse(Cond->getType());
400 // If both sides simplified to the same value, then use it as the result of
401 // the original comparison.
405 // The remaining cases only make sense if the select condition has the same
406 // type as the result of the comparison, so bail out if this is not so.
407 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
409 // If the false value simplified to false, then the result of the compare
410 // is equal to "Cond && TCmp". This also catches the case when the false
411 // value simplified to false and the true value to true, returning "Cond".
412 if (match(FCmp, m_Zero()))
413 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
415 // If the true value simplified to true, then the result of the compare
416 // is equal to "Cond || FCmp".
417 if (match(TCmp, m_One()))
418 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
420 // Finally, if the false value simplified to true and the true value to
421 // false, then the result of the compare is equal to "!Cond".
422 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
424 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
431 /// In the case of a binary operation with an operand that is a PHI instruction,
432 /// try to simplify the binop by seeing whether evaluating it on the incoming
433 /// phi values yields the same result for every value. If so returns the common
434 /// value, otherwise returns null.
435 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
436 Value *RHS, const SimplifyQuery &Q,
437 unsigned MaxRecurse) {
438 // Recursion is always used, so bail out at once if we already hit the limit.
443 if (isa<PHINode>(LHS)) {
444 PI = cast<PHINode>(LHS);
445 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
446 if (!ValueDominatesPHI(RHS, PI, Q.DT))
449 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
450 PI = cast<PHINode>(RHS);
451 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
452 if (!ValueDominatesPHI(LHS, PI, Q.DT))
456 // Evaluate the BinOp on the incoming phi values.
457 Value *CommonValue = nullptr;
458 for (Value *Incoming : PI->incoming_values()) {
459 // If the incoming value is the phi node itself, it can safely be skipped.
460 if (Incoming == PI) continue;
461 Value *V = PI == LHS ?
462 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
463 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
464 // If the operation failed to simplify, or simplified to a different value
465 // to previously, then give up.
466 if (!V || (CommonValue && V != CommonValue))
474 /// In the case of a comparison with a PHI instruction, try to simplify the
475 /// comparison by seeing whether comparing with all of the incoming phi values
476 /// yields the same result every time. If so returns the common result,
477 /// otherwise returns null.
478 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
479 const SimplifyQuery &Q, unsigned MaxRecurse) {
480 // Recursion is always used, so bail out at once if we already hit the limit.
484 // Make sure the phi is on the LHS.
485 if (!isa<PHINode>(LHS)) {
487 Pred = CmpInst::getSwappedPredicate(Pred);
489 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
490 PHINode *PI = cast<PHINode>(LHS);
492 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
493 if (!ValueDominatesPHI(RHS, PI, Q.DT))
496 // Evaluate the BinOp on the incoming phi values.
497 Value *CommonValue = nullptr;
498 for (Value *Incoming : PI->incoming_values()) {
499 // If the incoming value is the phi node itself, it can safely be skipped.
500 if (Incoming == PI) continue;
501 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
502 // If the operation failed to simplify, or simplified to a different value
503 // to previously, then give up.
504 if (!V || (CommonValue && V != CommonValue))
512 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
513 Value *&Op0, Value *&Op1,
514 const SimplifyQuery &Q) {
515 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
516 if (auto *CRHS = dyn_cast<Constant>(Op1))
517 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
519 // Canonicalize the constant to the RHS if this is a commutative operation.
520 if (Instruction::isCommutative(Opcode))
526 /// Given operands for an Add, see if we can fold the result.
527 /// If not, this returns null.
528 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
529 const SimplifyQuery &Q, unsigned MaxRecurse) {
530 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
533 // X + undef -> undef
534 if (match(Op1, m_Undef()))
538 if (match(Op1, m_Zero()))
545 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
546 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
549 // X + ~X -> -1 since ~X = -X-1
550 Type *Ty = Op0->getType();
551 if (match(Op0, m_Not(m_Specific(Op1))) ||
552 match(Op1, m_Not(m_Specific(Op0))))
553 return Constant::getAllOnesValue(Ty);
555 // add nsw/nuw (xor Y, signmask), signmask --> Y
556 // The no-wrapping add guarantees that the top bit will be set by the add.
557 // Therefore, the xor must be clearing the already set sign bit of Y.
558 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
559 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
563 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
564 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
567 // Try some generic simplifications for associative operations.
568 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
572 // Threading Add over selects and phi nodes is pointless, so don't bother.
573 // Threading over the select in "A + select(cond, B, C)" means evaluating
574 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
575 // only if B and C are equal. If B and C are equal then (since we assume
576 // that operands have already been simplified) "select(cond, B, C)" should
577 // have been simplified to the common value of B and C already. Analysing
578 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
579 // for threading over phi nodes.
584 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
585 const SimplifyQuery &Query) {
586 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
589 /// \brief Compute the base pointer and cumulative constant offsets for V.
591 /// This strips all constant offsets off of V, leaving it the base pointer, and
592 /// accumulates the total constant offset applied in the returned constant. It
593 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
594 /// no constant offsets applied.
596 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
597 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
599 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
600 bool AllowNonInbounds = false) {
601 assert(V->getType()->isPtrOrPtrVectorTy());
603 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
604 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
606 // Even though we don't look through PHI nodes, we could be called on an
607 // instruction in an unreachable block, which may be on a cycle.
608 SmallPtrSet<Value *, 4> Visited;
611 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
612 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
613 !GEP->accumulateConstantOffset(DL, Offset))
615 V = GEP->getPointerOperand();
616 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
617 V = cast<Operator>(V)->getOperand(0);
618 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
619 if (GA->isInterposable())
621 V = GA->getAliasee();
623 if (auto CS = CallSite(V))
624 if (Value *RV = CS.getReturnedArgOperand()) {
630 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
631 } while (Visited.insert(V).second);
633 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
634 if (V->getType()->isVectorTy())
635 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
640 /// \brief Compute the constant difference between two pointer values.
641 /// If the difference is not a constant, returns zero.
642 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
644 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
645 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
647 // If LHS and RHS are not related via constant offsets to the same base
648 // value, there is nothing we can do here.
652 // Otherwise, the difference of LHS - RHS can be computed as:
654 // = (LHSOffset + Base) - (RHSOffset + Base)
655 // = LHSOffset - RHSOffset
656 return ConstantExpr::getSub(LHSOffset, RHSOffset);
659 /// Given operands for a Sub, see if we can fold the result.
660 /// If not, this returns null.
661 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
662 const SimplifyQuery &Q, unsigned MaxRecurse) {
663 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
666 // X - undef -> undef
667 // undef - X -> undef
668 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
669 return UndefValue::get(Op0->getType());
672 if (match(Op1, m_Zero()))
677 return Constant::getNullValue(Op0->getType());
679 // Is this a negation?
680 if (match(Op0, m_Zero())) {
681 // 0 - X -> 0 if the sub is NUW.
685 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
686 if (Known.Zero.isMaxSignedValue()) {
687 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
688 // Op1 must be 0 because negating the minimum signed value is undefined.
692 // 0 - X -> X if X is 0 or the minimum signed value.
697 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
698 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
699 Value *X = nullptr, *Y = nullptr, *Z = Op1;
700 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
701 // See if "V === Y - Z" simplifies.
702 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
703 // It does! Now see if "X + V" simplifies.
704 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
705 // It does, we successfully reassociated!
709 // See if "V === X - Z" simplifies.
710 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
711 // It does! Now see if "Y + V" simplifies.
712 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
713 // It does, we successfully reassociated!
719 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
720 // For example, X - (X + 1) -> -1
722 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
723 // See if "V === X - Y" simplifies.
724 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
725 // It does! Now see if "V - Z" simplifies.
726 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
727 // It does, we successfully reassociated!
731 // See if "V === X - Z" simplifies.
732 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
733 // It does! Now see if "V - Y" simplifies.
734 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
735 // It does, we successfully reassociated!
741 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
742 // For example, X - (X - Y) -> Y.
744 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
745 // See if "V === Z - X" simplifies.
746 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
747 // It does! Now see if "V + Y" simplifies.
748 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
749 // It does, we successfully reassociated!
754 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
755 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
756 match(Op1, m_Trunc(m_Value(Y))))
757 if (X->getType() == Y->getType())
758 // See if "V === X - Y" simplifies.
759 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
760 // It does! Now see if "trunc V" simplifies.
761 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
763 // It does, return the simplified "trunc V".
766 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
767 if (match(Op0, m_PtrToInt(m_Value(X))) &&
768 match(Op1, m_PtrToInt(m_Value(Y))))
769 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
770 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
773 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
774 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
777 // Threading Sub over selects and phi nodes is pointless, so don't bother.
778 // Threading over the select in "A - select(cond, B, C)" means evaluating
779 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
780 // only if B and C are equal. If B and C are equal then (since we assume
781 // that operands have already been simplified) "select(cond, B, C)" should
782 // have been simplified to the common value of B and C already. Analysing
783 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
784 // for threading over phi nodes.
789 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
790 const SimplifyQuery &Q) {
791 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
794 /// Given operands for an FAdd, see if we can fold the result. If not, this
796 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
797 const SimplifyQuery &Q, unsigned MaxRecurse) {
798 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
802 if (match(Op1, m_NegZero()))
805 // fadd X, 0 ==> X, when we know X is not -0
806 if (match(Op1, m_Zero()) &&
807 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
810 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
811 // where nnan and ninf have to occur at least once somewhere in this
813 Value *SubOp = nullptr;
814 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
816 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
819 Instruction *FSub = cast<Instruction>(SubOp);
820 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
821 (FMF.noInfs() || FSub->hasNoInfs()))
822 return Constant::getNullValue(Op0->getType());
828 /// Given operands for an FSub, see if we can fold the result. If not, this
830 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
831 const SimplifyQuery &Q, unsigned MaxRecurse) {
832 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
836 if (match(Op1, m_Zero()))
839 // fsub X, -0 ==> X, when we know X is not -0
840 if (match(Op1, m_NegZero()) &&
841 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
844 // fsub -0.0, (fsub -0.0, X) ==> X
846 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
849 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
850 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
851 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
854 // fsub nnan x, x ==> 0.0
855 if (FMF.noNaNs() && Op0 == Op1)
856 return Constant::getNullValue(Op0->getType());
861 /// Given the operands for an FMul, see if we can fold the result
862 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
863 const SimplifyQuery &Q, unsigned MaxRecurse) {
864 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
868 if (match(Op1, m_FPOne()))
871 // fmul nnan nsz X, 0 ==> 0
872 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
878 /// Given operands for a Mul, see if we can fold the result.
879 /// If not, this returns null.
880 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
881 unsigned MaxRecurse) {
882 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
886 if (match(Op1, m_Undef()))
887 return Constant::getNullValue(Op0->getType());
890 if (match(Op1, m_Zero()))
894 if (match(Op1, m_One()))
897 // (X / Y) * Y -> X if the division is exact.
899 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
900 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
904 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
905 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
908 // Try some generic simplifications for associative operations.
909 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
913 // Mul distributes over Add. Try some generic simplifications based on this.
914 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
918 // If the operation is with the result of a select instruction, check whether
919 // operating on either branch of the select always yields the same value.
920 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
921 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
925 // If the operation is with the result of a phi instruction, check whether
926 // operating on all incoming values of the phi always yields the same value.
927 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
928 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
935 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
936 const SimplifyQuery &Q) {
937 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
941 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
942 const SimplifyQuery &Q) {
943 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
946 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
947 const SimplifyQuery &Q) {
948 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
951 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
952 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
955 /// Check for common or similar folds of integer division or integer remainder.
956 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
957 Type *Ty = Op0->getType();
959 // X / undef -> undef
960 // X % undef -> undef
961 if (match(Op1, m_Undef()))
966 // We don't need to preserve faults!
967 if (match(Op1, m_Zero()))
968 return UndefValue::get(Ty);
970 // If any element of a constant divisor vector is zero, the whole op is undef.
971 auto *Op1C = dyn_cast<Constant>(Op1);
972 if (Op1C && Ty->isVectorTy()) {
973 unsigned NumElts = Ty->getVectorNumElements();
974 for (unsigned i = 0; i != NumElts; ++i) {
975 Constant *Elt = Op1C->getAggregateElement(i);
976 if (Elt && Elt->isNullValue())
977 return UndefValue::get(Ty);
983 if (match(Op0, m_Undef()))
984 return Constant::getNullValue(Ty);
988 if (match(Op0, m_Zero()))
994 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
998 // If this is a boolean op (single-bit element type), we can't have
999 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
1000 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1))
1001 return IsDiv ? Op0 : Constant::getNullValue(Ty);
1006 /// Given operands for an SDiv or UDiv, see if we can fold the result.
1007 /// If not, this returns null.
1008 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1009 const SimplifyQuery &Q, unsigned MaxRecurse) {
1010 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1013 if (Value *V = simplifyDivRem(Op0, Op1, true))
1016 bool isSigned = Opcode == Instruction::SDiv;
1018 // (X * Y) / Y -> X if the multiplication does not overflow.
1019 Value *X = nullptr, *Y = nullptr;
1020 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1021 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1022 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1023 // If the Mul knows it does not overflow, then we are good to go.
1024 if ((isSigned && Mul->hasNoSignedWrap()) ||
1025 (!isSigned && Mul->hasNoUnsignedWrap()))
1027 // If X has the form X = A / Y then X * Y cannot overflow.
1028 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1029 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1033 // (X rem Y) / Y -> 0
1034 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1035 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1036 return Constant::getNullValue(Op0->getType());
1038 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1039 ConstantInt *C1, *C2;
1040 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1041 match(Op1, m_ConstantInt(C2))) {
1043 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1045 return Constant::getNullValue(Op0->getType());
1048 // If the operation is with the result of a select instruction, check whether
1049 // operating on either branch of the select always yields the same value.
1050 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1051 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1054 // If the operation is with the result of a phi instruction, check whether
1055 // operating on all incoming values of the phi always yields the same value.
1056 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1057 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1063 /// Given operands for an SDiv, see if we can fold the result.
1064 /// If not, this returns null.
1065 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1066 unsigned MaxRecurse) {
1067 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1073 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1074 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1077 /// Given operands for a UDiv, see if we can fold the result.
1078 /// If not, this returns null.
1079 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1080 unsigned MaxRecurse) {
1081 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1084 // udiv %V, C -> 0 if %V < C
1086 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1087 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1088 if (C->isAllOnesValue()) {
1089 return Constant::getNullValue(Op0->getType());
1097 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1098 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1101 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1102 const SimplifyQuery &Q, unsigned) {
1103 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
1106 // undef / X -> undef (the undef could be a snan).
1107 if (match(Op0, m_Undef()))
1110 // X / undef -> undef
1111 if (match(Op1, m_Undef()))
1115 if (match(Op1, m_FPOne()))
1119 // Requires that NaNs are off (X could be zero) and signed zeroes are
1120 // ignored (X could be positive or negative, so the output sign is unknown).
1121 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1125 // X / X -> 1.0 is legal when NaNs are ignored.
1127 return ConstantFP::get(Op0->getType(), 1.0);
1129 // -X / X -> -1.0 and
1130 // X / -X -> -1.0 are legal when NaNs are ignored.
1131 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1132 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1133 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1134 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1135 BinaryOperator::getFNegArgument(Op1) == Op0))
1136 return ConstantFP::get(Op0->getType(), -1.0);
1142 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1143 const SimplifyQuery &Q) {
1144 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
1147 /// Given operands for an SRem or URem, see if we can fold the result.
1148 /// If not, this returns null.
1149 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1150 const SimplifyQuery &Q, unsigned MaxRecurse) {
1151 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1154 if (Value *V = simplifyDivRem(Op0, Op1, false))
1157 // (X % Y) % Y -> X % Y
1158 if ((Opcode == Instruction::SRem &&
1159 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1160 (Opcode == Instruction::URem &&
1161 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1164 // If the operation is with the result of a select instruction, check whether
1165 // operating on either branch of the select always yields the same value.
1166 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1167 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1170 // If the operation is with the result of a phi instruction, check whether
1171 // operating on all incoming values of the phi always yields the same value.
1172 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1173 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1179 /// Given operands for an SRem, see if we can fold the result.
1180 /// If not, this returns null.
1181 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1182 unsigned MaxRecurse) {
1183 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1189 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1190 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1193 /// Given operands for a URem, see if we can fold the result.
1194 /// If not, this returns null.
1195 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1196 unsigned MaxRecurse) {
1197 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1200 // urem %V, C -> %V if %V < C
1202 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1203 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1204 if (C->isAllOnesValue()) {
1213 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1214 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1217 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1218 const SimplifyQuery &Q, unsigned) {
1219 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
1222 // undef % X -> undef (the undef could be a snan).
1223 if (match(Op0, m_Undef()))
1226 // X % undef -> undef
1227 if (match(Op1, m_Undef()))
1231 // Requires that NaNs are off (X could be zero) and signed zeroes are
1232 // ignored (X could be positive or negative, so the output sign is unknown).
1233 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1239 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1240 const SimplifyQuery &Q) {
1241 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
1244 /// Returns true if a shift by \c Amount always yields undef.
1245 static bool isUndefShift(Value *Amount) {
1246 Constant *C = dyn_cast<Constant>(Amount);
1250 // X shift by undef -> undef because it may shift by the bitwidth.
1251 if (isa<UndefValue>(C))
1254 // Shifting by the bitwidth or more is undefined.
1255 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1256 if (CI->getValue().getLimitedValue() >=
1257 CI->getType()->getScalarSizeInBits())
1260 // If all lanes of a vector shift are undefined the whole shift is.
1261 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1262 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1263 if (!isUndefShift(C->getAggregateElement(I)))
1271 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1272 /// If not, this returns null.
1273 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1274 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1275 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1278 // 0 shift by X -> 0
1279 if (match(Op0, m_Zero()))
1282 // X shift by 0 -> X
1283 if (match(Op1, m_Zero()))
1286 // Fold undefined shifts.
1287 if (isUndefShift(Op1))
1288 return UndefValue::get(Op0->getType());
1290 // If the operation is with the result of a select instruction, check whether
1291 // operating on either branch of the select always yields the same value.
1292 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1293 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1296 // If the operation is with the result of a phi instruction, check whether
1297 // operating on all incoming values of the phi always yields the same value.
1298 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1299 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1302 // If any bits in the shift amount make that value greater than or equal to
1303 // the number of bits in the type, the shift is undefined.
1304 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1305 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1306 return UndefValue::get(Op0->getType());
1308 // If all valid bits in the shift amount are known zero, the first operand is
1310 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1311 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1317 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1318 /// fold the result. If not, this returns null.
1319 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1320 Value *Op1, bool isExact, const SimplifyQuery &Q,
1321 unsigned MaxRecurse) {
1322 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1327 return Constant::getNullValue(Op0->getType());
1330 // undef >> X -> undef (if it's exact)
1331 if (match(Op0, m_Undef()))
1332 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1334 // The low bit cannot be shifted out of an exact shift if it is set.
1336 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1337 if (Op0Known.One[0])
1344 /// Given operands for an Shl, see if we can fold the result.
1345 /// If not, this returns null.
1346 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1347 const SimplifyQuery &Q, unsigned MaxRecurse) {
1348 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1352 // undef << X -> undef if (if it's NSW/NUW)
1353 if (match(Op0, m_Undef()))
1354 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1356 // (X >> A) << A -> X
1358 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1363 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1364 const SimplifyQuery &Q) {
1365 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1368 /// Given operands for an LShr, see if we can fold the result.
1369 /// If not, this returns null.
1370 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1371 const SimplifyQuery &Q, unsigned MaxRecurse) {
1372 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1376 // (X << A) >> A -> X
1378 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1384 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1385 const SimplifyQuery &Q) {
1386 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1389 /// Given operands for an AShr, see if we can fold the result.
1390 /// If not, this returns null.
1391 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1392 const SimplifyQuery &Q, unsigned MaxRecurse) {
1393 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1397 // all ones >>a X -> all ones
1398 if (match(Op0, m_AllOnes()))
1401 // (X << A) >> A -> X
1403 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1406 // Arithmetic shifting an all-sign-bit value is a no-op.
1407 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1408 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1414 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1415 const SimplifyQuery &Q) {
1416 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1419 /// Commuted variants are assumed to be handled by calling this function again
1420 /// with the parameters swapped.
1421 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1422 ICmpInst *UnsignedICmp, bool IsAnd) {
1425 ICmpInst::Predicate EqPred;
1426 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1427 !ICmpInst::isEquality(EqPred))
1430 ICmpInst::Predicate UnsignedPred;
1431 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1432 ICmpInst::isUnsigned(UnsignedPred))
1434 else if (match(UnsignedICmp,
1435 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1436 ICmpInst::isUnsigned(UnsignedPred))
1437 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1441 // X < Y && Y != 0 --> X < Y
1442 // X < Y || Y != 0 --> Y != 0
1443 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1444 return IsAnd ? UnsignedICmp : ZeroICmp;
1446 // X >= Y || Y != 0 --> true
1447 // X >= Y || Y == 0 --> X >= Y
1448 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1449 if (EqPred == ICmpInst::ICMP_NE)
1450 return getTrue(UnsignedICmp->getType());
1451 return UnsignedICmp;
1454 // X < Y && Y == 0 --> false
1455 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1457 return getFalse(UnsignedICmp->getType());
1462 /// Commuted variants are assumed to be handled by calling this function again
1463 /// with the parameters swapped.
1464 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1465 ICmpInst::Predicate Pred0, Pred1;
1467 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1468 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1471 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1472 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1473 // can eliminate Op1 from this 'and'.
1474 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1477 // Check for any combination of predicates that are guaranteed to be disjoint.
1478 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1479 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1480 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1481 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1482 return getFalse(Op0->getType());
1487 /// Commuted variants are assumed to be handled by calling this function again
1488 /// with the parameters swapped.
1489 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1490 ICmpInst::Predicate Pred0, Pred1;
1492 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1493 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1496 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1497 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1498 // can eliminate Op0 from this 'or'.
1499 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1502 // Check for any combination of predicates that cover the entire range of
1504 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1505 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1506 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1507 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1508 return getTrue(Op0->getType());
1513 /// Test if a pair of compares with a shared operand and 2 constants has an
1514 /// empty set intersection, full set union, or if one compare is a superset of
1516 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1518 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1519 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1522 const APInt *C0, *C1;
1523 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1524 !match(Cmp1->getOperand(1), m_APInt(C1)))
1527 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1528 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1530 // For and-of-compares, check if the intersection is empty:
1531 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1532 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1533 return getFalse(Cmp0->getType());
1535 // For or-of-compares, check if the union is full:
1536 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1537 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1538 return getTrue(Cmp0->getType());
1540 // Is one range a superset of the other?
1541 // If this is and-of-compares, take the smaller set:
1542 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1543 // If this is or-of-compares, take the larger set:
1544 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1545 if (Range0.contains(Range1))
1546 return IsAnd ? Cmp1 : Cmp0;
1547 if (Range1.contains(Range0))
1548 return IsAnd ? Cmp0 : Cmp1;
1553 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1554 // (icmp (add V, C0), C1) & (icmp V, C0)
1555 ICmpInst::Predicate Pred0, Pred1;
1556 const APInt *C0, *C1;
1558 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1561 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1564 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1565 if (AddInst->getOperand(1) != Op1->getOperand(1))
1568 Type *ITy = Op0->getType();
1569 bool isNSW = AddInst->hasNoSignedWrap();
1570 bool isNUW = AddInst->hasNoUnsignedWrap();
1572 const APInt Delta = *C1 - *C0;
1573 if (C0->isStrictlyPositive()) {
1575 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1576 return getFalse(ITy);
1577 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1578 return getFalse(ITy);
1581 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1582 return getFalse(ITy);
1583 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1584 return getFalse(ITy);
1587 if (C0->getBoolValue() && isNUW) {
1589 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1590 return getFalse(ITy);
1592 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1593 return getFalse(ITy);
1599 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1600 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1602 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1605 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1607 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1610 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1613 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1615 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1621 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1622 // (icmp (add V, C0), C1) | (icmp V, C0)
1623 ICmpInst::Predicate Pred0, Pred1;
1624 const APInt *C0, *C1;
1626 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1629 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1632 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1633 if (AddInst->getOperand(1) != Op1->getOperand(1))
1636 Type *ITy = Op0->getType();
1637 bool isNSW = AddInst->hasNoSignedWrap();
1638 bool isNUW = AddInst->hasNoUnsignedWrap();
1640 const APInt Delta = *C1 - *C0;
1641 if (C0->isStrictlyPositive()) {
1643 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1644 return getTrue(ITy);
1645 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1646 return getTrue(ITy);
1649 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1650 return getTrue(ITy);
1651 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1652 return getTrue(ITy);
1655 if (C0->getBoolValue() && isNUW) {
1657 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1658 return getTrue(ITy);
1660 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1661 return getTrue(ITy);
1667 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1668 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1670 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1673 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1675 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1678 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1681 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1683 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1689 static Value *simplifyAndOrOfICmps(Value *Op0, Value *Op1, bool IsAnd) {
1690 // Look through casts of the 'and' operands to find compares.
1691 auto *Cast0 = dyn_cast<CastInst>(Op0);
1692 auto *Cast1 = dyn_cast<CastInst>(Op1);
1693 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1694 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1695 Op0 = Cast0->getOperand(0);
1696 Op1 = Cast1->getOperand(0);
1699 auto *Cmp0 = dyn_cast<ICmpInst>(Op0);
1700 auto *Cmp1 = dyn_cast<ICmpInst>(Op1);
1705 IsAnd ? simplifyAndOfICmps(Cmp0, Cmp1) : simplifyOrOfICmps(Cmp0, Cmp1);
1711 // If we looked through casts, we can only handle a constant simplification
1712 // because we are not allowed to create a cast instruction here.
1713 if (auto *C = dyn_cast<Constant>(V))
1714 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1719 /// Given operands for an And, see if we can fold the result.
1720 /// If not, this returns null.
1721 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1722 unsigned MaxRecurse) {
1723 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1727 if (match(Op1, m_Undef()))
1728 return Constant::getNullValue(Op0->getType());
1735 if (match(Op1, m_Zero()))
1739 if (match(Op1, m_AllOnes()))
1742 // A & ~A = ~A & A = 0
1743 if (match(Op0, m_Not(m_Specific(Op1))) ||
1744 match(Op1, m_Not(m_Specific(Op0))))
1745 return Constant::getNullValue(Op0->getType());
1748 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1752 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
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 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1856 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1859 // ~(A & ?) | A = -1
1860 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1861 return Constant::getAllOnesValue(Op1->getType());
1863 // A | ~(A & ?) = -1
1864 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1865 return Constant::getAllOnesValue(Op0->getType());
1868 // (A & ~B) | (A ^ B) -> (A ^ B)
1869 // (~B & A) | (A ^ B) -> (A ^ B)
1870 // (A & ~B) | (B ^ A) -> (B ^ A)
1871 // (~B & A) | (B ^ A) -> (B ^ A)
1872 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1873 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1874 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1877 // Commute the 'or' operands.
1878 // (A ^ B) | (A & ~B) -> (A ^ B)
1879 // (A ^ B) | (~B & A) -> (A ^ B)
1880 // (B ^ A) | (A & ~B) -> (B ^ A)
1881 // (B ^ A) | (~B & A) -> (B ^ A)
1882 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1883 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1884 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1887 // (A & B) | (~A ^ B) -> (~A ^ B)
1888 // (B & A) | (~A ^ B) -> (~A ^ B)
1889 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1890 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1891 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1892 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1893 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1896 // (~A ^ B) | (A & B) -> (~A ^ B)
1897 // (~A ^ B) | (B & A) -> (~A ^ B)
1898 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1899 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1900 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1901 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1902 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1905 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, false))
1908 // Try some generic simplifications for associative operations.
1909 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1913 // Or distributes over And. Try some generic simplifications based on this.
1914 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1918 // If the operation is with the result of a select instruction, check whether
1919 // operating on either branch of the select always yields the same value.
1920 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1921 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1925 // (A & C1)|(B & C2)
1926 const APInt *C1, *C2;
1927 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1928 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1930 // (A & C1)|(B & C2)
1931 // If we have: ((V + N) & C1) | (V & C2)
1932 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1933 // replace with V+N.
1935 if (C2->isMask() && // C2 == 0+1+
1936 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1937 // Add commutes, try both ways.
1938 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1941 // Or commutes, try both ways.
1943 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1944 // Add commutes, try both ways.
1945 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1951 // If the operation is with the result of a phi instruction, check whether
1952 // operating on all incoming values of the phi always yields the same value.
1953 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1954 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1960 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1961 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1964 /// Given operands for a Xor, see if we can fold the result.
1965 /// If not, this returns null.
1966 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1967 unsigned MaxRecurse) {
1968 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1971 // A ^ undef -> undef
1972 if (match(Op1, m_Undef()))
1976 if (match(Op1, m_Zero()))
1981 return Constant::getNullValue(Op0->getType());
1983 // A ^ ~A = ~A ^ A = -1
1984 if (match(Op0, m_Not(m_Specific(Op1))) ||
1985 match(Op1, m_Not(m_Specific(Op0))))
1986 return Constant::getAllOnesValue(Op0->getType());
1988 // Try some generic simplifications for associative operations.
1989 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1993 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1994 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1995 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1996 // only if B and C are equal. If B and C are equal then (since we assume
1997 // that operands have already been simplified) "select(cond, B, C)" should
1998 // have been simplified to the common value of B and C already. Analysing
1999 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2000 // for threading over phi nodes.
2005 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2006 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2010 static Type *GetCompareTy(Value *Op) {
2011 return CmpInst::makeCmpResultType(Op->getType());
2014 /// Rummage around inside V looking for something equivalent to the comparison
2015 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2016 /// Helper function for analyzing max/min idioms.
2017 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2018 Value *LHS, Value *RHS) {
2019 SelectInst *SI = dyn_cast<SelectInst>(V);
2022 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2025 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2026 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2028 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2029 LHS == CmpRHS && RHS == CmpLHS)
2034 // A significant optimization not implemented here is assuming that alloca
2035 // addresses are not equal to incoming argument values. They don't *alias*,
2036 // as we say, but that doesn't mean they aren't equal, so we take a
2037 // conservative approach.
2039 // This is inspired in part by C++11 5.10p1:
2040 // "Two pointers of the same type compare equal if and only if they are both
2041 // null, both point to the same function, or both represent the same
2044 // This is pretty permissive.
2046 // It's also partly due to C11 6.5.9p6:
2047 // "Two pointers compare equal if and only if both are null pointers, both are
2048 // pointers to the same object (including a pointer to an object and a
2049 // subobject at its beginning) or function, both are pointers to one past the
2050 // last element of the same array object, or one is a pointer to one past the
2051 // end of one array object and the other is a pointer to the start of a
2052 // different array object that happens to immediately follow the first array
2053 // object in the address space.)
2055 // C11's version is more restrictive, however there's no reason why an argument
2056 // couldn't be a one-past-the-end value for a stack object in the caller and be
2057 // equal to the beginning of a stack object in the callee.
2059 // If the C and C++ standards are ever made sufficiently restrictive in this
2060 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2061 // this optimization.
2063 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2064 const DominatorTree *DT, CmpInst::Predicate Pred,
2065 const Instruction *CxtI, Value *LHS, Value *RHS) {
2066 // First, skip past any trivial no-ops.
2067 LHS = LHS->stripPointerCasts();
2068 RHS = RHS->stripPointerCasts();
2070 // A non-null pointer is not equal to a null pointer.
2071 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) &&
2072 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2073 return ConstantInt::get(GetCompareTy(LHS),
2074 !CmpInst::isTrueWhenEqual(Pred));
2076 // We can only fold certain predicates on pointer comparisons.
2081 // Equality comaprisons are easy to fold.
2082 case CmpInst::ICMP_EQ:
2083 case CmpInst::ICMP_NE:
2086 // We can only handle unsigned relational comparisons because 'inbounds' on
2087 // a GEP only protects against unsigned wrapping.
2088 case CmpInst::ICMP_UGT:
2089 case CmpInst::ICMP_UGE:
2090 case CmpInst::ICMP_ULT:
2091 case CmpInst::ICMP_ULE:
2092 // However, we have to switch them to their signed variants to handle
2093 // negative indices from the base pointer.
2094 Pred = ICmpInst::getSignedPredicate(Pred);
2098 // Strip off any constant offsets so that we can reason about them.
2099 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2100 // here and compare base addresses like AliasAnalysis does, however there are
2101 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2102 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2103 // doesn't need to guarantee pointer inequality when it says NoAlias.
2104 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2105 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2107 // If LHS and RHS are related via constant offsets to the same base
2108 // value, we can replace it with an icmp which just compares the offsets.
2110 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2112 // Various optimizations for (in)equality comparisons.
2113 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2114 // Different non-empty allocations that exist at the same time have
2115 // different addresses (if the program can tell). Global variables always
2116 // exist, so they always exist during the lifetime of each other and all
2117 // allocas. Two different allocas usually have different addresses...
2119 // However, if there's an @llvm.stackrestore dynamically in between two
2120 // allocas, they may have the same address. It's tempting to reduce the
2121 // scope of the problem by only looking at *static* allocas here. That would
2122 // cover the majority of allocas while significantly reducing the likelihood
2123 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2124 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2125 // an entry block. Also, if we have a block that's not attached to a
2126 // function, we can't tell if it's "static" under the current definition.
2127 // Theoretically, this problem could be fixed by creating a new kind of
2128 // instruction kind specifically for static allocas. Such a new instruction
2129 // could be required to be at the top of the entry block, thus preventing it
2130 // from being subject to a @llvm.stackrestore. Instcombine could even
2131 // convert regular allocas into these special allocas. It'd be nifty.
2132 // However, until then, this problem remains open.
2134 // So, we'll assume that two non-empty allocas have different addresses
2137 // With all that, if the offsets are within the bounds of their allocations
2138 // (and not one-past-the-end! so we can't use inbounds!), and their
2139 // allocations aren't the same, the pointers are not equal.
2141 // Note that it's not necessary to check for LHS being a global variable
2142 // address, due to canonicalization and constant folding.
2143 if (isa<AllocaInst>(LHS) &&
2144 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2145 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2146 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2147 uint64_t LHSSize, RHSSize;
2148 if (LHSOffsetCI && RHSOffsetCI &&
2149 getObjectSize(LHS, LHSSize, DL, TLI) &&
2150 getObjectSize(RHS, RHSSize, DL, TLI)) {
2151 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2152 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2153 if (!LHSOffsetValue.isNegative() &&
2154 !RHSOffsetValue.isNegative() &&
2155 LHSOffsetValue.ult(LHSSize) &&
2156 RHSOffsetValue.ult(RHSSize)) {
2157 return ConstantInt::get(GetCompareTy(LHS),
2158 !CmpInst::isTrueWhenEqual(Pred));
2162 // Repeat the above check but this time without depending on DataLayout
2163 // or being able to compute a precise size.
2164 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2165 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2166 LHSOffset->isNullValue() &&
2167 RHSOffset->isNullValue())
2168 return ConstantInt::get(GetCompareTy(LHS),
2169 !CmpInst::isTrueWhenEqual(Pred));
2172 // Even if an non-inbounds GEP occurs along the path we can still optimize
2173 // equality comparisons concerning the result. We avoid walking the whole
2174 // chain again by starting where the last calls to
2175 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2176 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2177 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2179 return ConstantExpr::getICmp(Pred,
2180 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2181 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2183 // If one side of the equality comparison must come from a noalias call
2184 // (meaning a system memory allocation function), and the other side must
2185 // come from a pointer that cannot overlap with dynamically-allocated
2186 // memory within the lifetime of the current function (allocas, byval
2187 // arguments, globals), then determine the comparison result here.
2188 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2189 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2190 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2192 // Is the set of underlying objects all noalias calls?
2193 auto IsNAC = [](ArrayRef<Value *> Objects) {
2194 return all_of(Objects, isNoAliasCall);
2197 // Is the set of underlying objects all things which must be disjoint from
2198 // noalias calls. For allocas, we consider only static ones (dynamic
2199 // allocas might be transformed into calls to malloc not simultaneously
2200 // live with the compared-to allocation). For globals, we exclude symbols
2201 // that might be resolve lazily to symbols in another dynamically-loaded
2202 // library (and, thus, could be malloc'ed by the implementation).
2203 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2204 return all_of(Objects, [](Value *V) {
2205 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2206 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2207 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2208 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2209 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2210 !GV->isThreadLocal();
2211 if (const Argument *A = dyn_cast<Argument>(V))
2212 return A->hasByValAttr();
2217 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2218 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2219 return ConstantInt::get(GetCompareTy(LHS),
2220 !CmpInst::isTrueWhenEqual(Pred));
2222 // Fold comparisons for non-escaping pointer even if the allocation call
2223 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2224 // dynamic allocation call could be either of the operands.
2225 Value *MI = nullptr;
2226 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT))
2228 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT))
2230 // FIXME: We should also fold the compare when the pointer escapes, but the
2231 // compare dominates the pointer escape
2232 if (MI && !PointerMayBeCaptured(MI, true, true))
2233 return ConstantInt::get(GetCompareTy(LHS),
2234 CmpInst::isFalseWhenEqual(Pred));
2241 /// Fold an icmp when its operands have i1 scalar type.
2242 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2243 Value *RHS, const SimplifyQuery &Q) {
2244 Type *ITy = GetCompareTy(LHS); // The return type.
2245 Type *OpTy = LHS->getType(); // The operand type.
2246 if (!OpTy->isIntOrIntVectorTy(1))
2249 // A boolean compared to true/false can be simplified in 14 out of the 20
2250 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2251 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2252 if (match(RHS, m_Zero())) {
2254 case CmpInst::ICMP_NE: // X != 0 -> X
2255 case CmpInst::ICMP_UGT: // X >u 0 -> X
2256 case CmpInst::ICMP_SLT: // X <s 0 -> X
2259 case CmpInst::ICMP_ULT: // X <u 0 -> false
2260 case CmpInst::ICMP_SGT: // X >s 0 -> false
2261 return getFalse(ITy);
2263 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2264 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2265 return getTrue(ITy);
2269 } else if (match(RHS, m_One())) {
2271 case CmpInst::ICMP_EQ: // X == 1 -> X
2272 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2273 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2276 case CmpInst::ICMP_UGT: // X >u 1 -> false
2277 case CmpInst::ICMP_SLT: // X <s -1 -> false
2278 return getFalse(ITy);
2280 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2281 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2282 return getTrue(ITy);
2291 case ICmpInst::ICMP_UGE:
2292 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2293 return getTrue(ITy);
2295 case ICmpInst::ICMP_SGE:
2296 /// For signed comparison, the values for an i1 are 0 and -1
2297 /// respectively. This maps into a truth table of:
2298 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2299 /// 0 | 0 | 1 (0 >= 0) | 1
2300 /// 0 | 1 | 1 (0 >= -1) | 1
2301 /// 1 | 0 | 0 (-1 >= 0) | 0
2302 /// 1 | 1 | 1 (-1 >= -1) | 1
2303 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2304 return getTrue(ITy);
2306 case ICmpInst::ICMP_ULE:
2307 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2308 return getTrue(ITy);
2315 /// Try hard to fold icmp with zero RHS because this is a common case.
2316 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2317 Value *RHS, const SimplifyQuery &Q) {
2318 if (!match(RHS, m_Zero()))
2321 Type *ITy = GetCompareTy(LHS); // The return type.
2324 llvm_unreachable("Unknown ICmp predicate!");
2325 case ICmpInst::ICMP_ULT:
2326 return getFalse(ITy);
2327 case ICmpInst::ICMP_UGE:
2328 return getTrue(ITy);
2329 case ICmpInst::ICMP_EQ:
2330 case ICmpInst::ICMP_ULE:
2331 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2332 return getFalse(ITy);
2334 case ICmpInst::ICMP_NE:
2335 case ICmpInst::ICMP_UGT:
2336 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2337 return getTrue(ITy);
2339 case ICmpInst::ICMP_SLT: {
2340 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2341 if (LHSKnown.isNegative())
2342 return getTrue(ITy);
2343 if (LHSKnown.isNonNegative())
2344 return getFalse(ITy);
2347 case ICmpInst::ICMP_SLE: {
2348 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2349 if (LHSKnown.isNegative())
2350 return getTrue(ITy);
2351 if (LHSKnown.isNonNegative() &&
2352 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2353 return getFalse(ITy);
2356 case ICmpInst::ICMP_SGE: {
2357 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2358 if (LHSKnown.isNegative())
2359 return getFalse(ITy);
2360 if (LHSKnown.isNonNegative())
2361 return getTrue(ITy);
2364 case ICmpInst::ICMP_SGT: {
2365 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2366 if (LHSKnown.isNegative())
2367 return getFalse(ITy);
2368 if (LHSKnown.isNonNegative() &&
2369 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2370 return getTrue(ITy);
2378 /// Many binary operators with a constant operand have an easy-to-compute
2379 /// range of outputs. This can be used to fold a comparison to always true or
2381 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2382 unsigned Width = Lower.getBitWidth();
2384 switch (BO.getOpcode()) {
2385 case Instruction::Add:
2386 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2387 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2388 if (BO.hasNoUnsignedWrap()) {
2389 // 'add nuw x, C' produces [C, UINT_MAX].
2391 } else if (BO.hasNoSignedWrap()) {
2392 if (C->isNegative()) {
2393 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2394 Lower = APInt::getSignedMinValue(Width);
2395 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2397 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2398 Lower = APInt::getSignedMinValue(Width) + *C;
2399 Upper = APInt::getSignedMaxValue(Width) + 1;
2405 case Instruction::And:
2406 if (match(BO.getOperand(1), m_APInt(C)))
2407 // 'and x, C' produces [0, C].
2411 case Instruction::Or:
2412 if (match(BO.getOperand(1), m_APInt(C)))
2413 // 'or x, C' produces [C, UINT_MAX].
2417 case Instruction::AShr:
2418 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2419 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2420 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2421 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2422 } else if (match(BO.getOperand(0), m_APInt(C))) {
2423 unsigned ShiftAmount = Width - 1;
2424 if (!C->isNullValue() && BO.isExact())
2425 ShiftAmount = C->countTrailingZeros();
2426 if (C->isNegative()) {
2427 // 'ashr C, x' produces [C, C >> (Width-1)]
2429 Upper = C->ashr(ShiftAmount) + 1;
2431 // 'ashr C, x' produces [C >> (Width-1), C]
2432 Lower = C->ashr(ShiftAmount);
2438 case Instruction::LShr:
2439 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2440 // 'lshr x, C' produces [0, UINT_MAX >> C].
2441 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2442 } else if (match(BO.getOperand(0), m_APInt(C))) {
2443 // 'lshr C, x' produces [C >> (Width-1), C].
2444 unsigned ShiftAmount = Width - 1;
2445 if (!C->isNullValue() && BO.isExact())
2446 ShiftAmount = C->countTrailingZeros();
2447 Lower = C->lshr(ShiftAmount);
2452 case Instruction::Shl:
2453 if (match(BO.getOperand(0), m_APInt(C))) {
2454 if (BO.hasNoUnsignedWrap()) {
2455 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2457 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2458 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2459 if (C->isNegative()) {
2460 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2461 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2462 Lower = C->shl(ShiftAmount);
2465 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2466 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2468 Upper = C->shl(ShiftAmount) + 1;
2474 case Instruction::SDiv:
2475 if (match(BO.getOperand(1), m_APInt(C))) {
2476 APInt IntMin = APInt::getSignedMinValue(Width);
2477 APInt IntMax = APInt::getSignedMaxValue(Width);
2478 if (C->isAllOnesValue()) {
2479 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2480 // where C != -1 and C != 0 and C != 1
2483 } else if (C->countLeadingZeros() < Width - 1) {
2484 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2485 // where C != -1 and C != 0 and C != 1
2486 Lower = IntMin.sdiv(*C);
2487 Upper = IntMax.sdiv(*C);
2488 if (Lower.sgt(Upper))
2489 std::swap(Lower, Upper);
2491 assert(Upper != Lower && "Upper part of range has wrapped!");
2493 } else if (match(BO.getOperand(0), m_APInt(C))) {
2494 if (C->isMinSignedValue()) {
2495 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2497 Upper = Lower.lshr(1) + 1;
2499 // 'sdiv C, x' produces [-|C|, |C|].
2500 Upper = C->abs() + 1;
2501 Lower = (-Upper) + 1;
2506 case Instruction::UDiv:
2507 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2508 // 'udiv x, C' produces [0, UINT_MAX / C].
2509 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2510 } else if (match(BO.getOperand(0), m_APInt(C))) {
2511 // 'udiv C, x' produces [0, C].
2516 case Instruction::SRem:
2517 if (match(BO.getOperand(1), m_APInt(C))) {
2518 // 'srem x, C' produces (-|C|, |C|).
2520 Lower = (-Upper) + 1;
2524 case Instruction::URem:
2525 if (match(BO.getOperand(1), m_APInt(C)))
2526 // 'urem x, C' produces [0, C).
2535 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2538 if (!match(RHS, m_APInt(C)))
2541 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2542 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2543 if (RHS_CR.isEmptySet())
2544 return ConstantInt::getFalse(GetCompareTy(RHS));
2545 if (RHS_CR.isFullSet())
2546 return ConstantInt::getTrue(GetCompareTy(RHS));
2548 // Find the range of possible values for binary operators.
2549 unsigned Width = C->getBitWidth();
2550 APInt Lower = APInt(Width, 0);
2551 APInt Upper = APInt(Width, 0);
2552 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2553 setLimitsForBinOp(*BO, Lower, Upper);
2555 ConstantRange LHS_CR =
2556 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2558 if (auto *I = dyn_cast<Instruction>(LHS))
2559 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2560 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2562 if (!LHS_CR.isFullSet()) {
2563 if (RHS_CR.contains(LHS_CR))
2564 return ConstantInt::getTrue(GetCompareTy(RHS));
2565 if (RHS_CR.inverse().contains(LHS_CR))
2566 return ConstantInt::getFalse(GetCompareTy(RHS));
2572 /// TODO: A large part of this logic is duplicated in InstCombine's
2573 /// foldICmpBinOp(). We should be able to share that and avoid the code
2575 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2576 Value *RHS, const SimplifyQuery &Q,
2577 unsigned MaxRecurse) {
2578 Type *ITy = GetCompareTy(LHS); // The return type.
2580 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2581 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2582 if (MaxRecurse && (LBO || RBO)) {
2583 // Analyze the case when either LHS or RHS is an add instruction.
2584 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2585 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2586 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2587 if (LBO && LBO->getOpcode() == Instruction::Add) {
2588 A = LBO->getOperand(0);
2589 B = LBO->getOperand(1);
2591 ICmpInst::isEquality(Pred) ||
2592 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2593 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2595 if (RBO && RBO->getOpcode() == Instruction::Add) {
2596 C = RBO->getOperand(0);
2597 D = RBO->getOperand(1);
2599 ICmpInst::isEquality(Pred) ||
2600 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2601 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2604 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2605 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2606 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2607 Constant::getNullValue(RHS->getType()), Q,
2611 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2612 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2614 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2615 C == LHS ? D : C, Q, MaxRecurse - 1))
2618 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2619 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2621 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2624 // C + B == C + D -> B == D
2627 } else if (A == D) {
2628 // D + B == C + D -> B == C
2631 } else if (B == C) {
2632 // A + C == C + D -> A == D
2637 // A + D == C + D -> A == C
2641 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2648 // icmp pred (or X, Y), X
2649 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2650 if (Pred == ICmpInst::ICMP_ULT)
2651 return getFalse(ITy);
2652 if (Pred == ICmpInst::ICMP_UGE)
2653 return getTrue(ITy);
2655 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2656 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2657 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2658 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2659 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2660 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2661 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2664 // icmp pred X, (or X, Y)
2665 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2666 if (Pred == ICmpInst::ICMP_ULE)
2667 return getTrue(ITy);
2668 if (Pred == ICmpInst::ICMP_UGT)
2669 return getFalse(ITy);
2671 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2672 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2673 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2674 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2675 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2676 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2677 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2682 // icmp pred (and X, Y), X
2683 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2684 if (Pred == ICmpInst::ICMP_UGT)
2685 return getFalse(ITy);
2686 if (Pred == ICmpInst::ICMP_ULE)
2687 return getTrue(ITy);
2689 // icmp pred X, (and X, Y)
2690 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2691 if (Pred == ICmpInst::ICMP_UGE)
2692 return getTrue(ITy);
2693 if (Pred == ICmpInst::ICMP_ULT)
2694 return getFalse(ITy);
2697 // 0 - (zext X) pred C
2698 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2699 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2700 if (RHSC->getValue().isStrictlyPositive()) {
2701 if (Pred == ICmpInst::ICMP_SLT)
2702 return ConstantInt::getTrue(RHSC->getContext());
2703 if (Pred == ICmpInst::ICMP_SGE)
2704 return ConstantInt::getFalse(RHSC->getContext());
2705 if (Pred == ICmpInst::ICMP_EQ)
2706 return ConstantInt::getFalse(RHSC->getContext());
2707 if (Pred == ICmpInst::ICMP_NE)
2708 return ConstantInt::getTrue(RHSC->getContext());
2710 if (RHSC->getValue().isNonNegative()) {
2711 if (Pred == ICmpInst::ICMP_SLE)
2712 return ConstantInt::getTrue(RHSC->getContext());
2713 if (Pred == ICmpInst::ICMP_SGT)
2714 return ConstantInt::getFalse(RHSC->getContext());
2719 // icmp pred (urem X, Y), Y
2720 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2724 case ICmpInst::ICMP_SGT:
2725 case ICmpInst::ICMP_SGE: {
2726 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2727 if (!Known.isNonNegative())
2731 case ICmpInst::ICMP_EQ:
2732 case ICmpInst::ICMP_UGT:
2733 case ICmpInst::ICMP_UGE:
2734 return getFalse(ITy);
2735 case ICmpInst::ICMP_SLT:
2736 case ICmpInst::ICMP_SLE: {
2737 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2738 if (!Known.isNonNegative())
2742 case ICmpInst::ICMP_NE:
2743 case ICmpInst::ICMP_ULT:
2744 case ICmpInst::ICMP_ULE:
2745 return getTrue(ITy);
2749 // icmp pred X, (urem Y, X)
2750 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2754 case ICmpInst::ICMP_SGT:
2755 case ICmpInst::ICMP_SGE: {
2756 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2757 if (!Known.isNonNegative())
2761 case ICmpInst::ICMP_NE:
2762 case ICmpInst::ICMP_UGT:
2763 case ICmpInst::ICMP_UGE:
2764 return getTrue(ITy);
2765 case ICmpInst::ICMP_SLT:
2766 case ICmpInst::ICMP_SLE: {
2767 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2768 if (!Known.isNonNegative())
2772 case ICmpInst::ICMP_EQ:
2773 case ICmpInst::ICMP_ULT:
2774 case ICmpInst::ICMP_ULE:
2775 return getFalse(ITy);
2781 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2782 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2783 // icmp pred (X op Y), X
2784 if (Pred == ICmpInst::ICMP_UGT)
2785 return getFalse(ITy);
2786 if (Pred == ICmpInst::ICMP_ULE)
2787 return getTrue(ITy);
2792 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2793 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2794 // icmp pred X, (X op Y)
2795 if (Pred == ICmpInst::ICMP_ULT)
2796 return getFalse(ITy);
2797 if (Pred == ICmpInst::ICMP_UGE)
2798 return getTrue(ITy);
2805 // where CI2 is a power of 2 and CI isn't
2806 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2807 const APInt *CI2Val, *CIVal = &CI->getValue();
2808 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2809 CI2Val->isPowerOf2()) {
2810 if (!CIVal->isPowerOf2()) {
2811 // CI2 << X can equal zero in some circumstances,
2812 // this simplification is unsafe if CI is zero.
2814 // We know it is safe if:
2815 // - The shift is nsw, we can't shift out the one bit.
2816 // - The shift is nuw, we can't shift out the one bit.
2819 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2820 CI2Val->isOneValue() || !CI->isZero()) {
2821 if (Pred == ICmpInst::ICMP_EQ)
2822 return ConstantInt::getFalse(RHS->getContext());
2823 if (Pred == ICmpInst::ICMP_NE)
2824 return ConstantInt::getTrue(RHS->getContext());
2827 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2828 if (Pred == ICmpInst::ICMP_UGT)
2829 return ConstantInt::getFalse(RHS->getContext());
2830 if (Pred == ICmpInst::ICMP_ULE)
2831 return ConstantInt::getTrue(RHS->getContext());
2836 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2837 LBO->getOperand(1) == RBO->getOperand(1)) {
2838 switch (LBO->getOpcode()) {
2841 case Instruction::UDiv:
2842 case Instruction::LShr:
2843 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2845 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2846 RBO->getOperand(0), Q, MaxRecurse - 1))
2849 case Instruction::SDiv:
2850 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2852 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2853 RBO->getOperand(0), Q, MaxRecurse - 1))
2856 case Instruction::AShr:
2857 if (!LBO->isExact() || !RBO->isExact())
2859 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2860 RBO->getOperand(0), Q, MaxRecurse - 1))
2863 case Instruction::Shl: {
2864 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2865 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2868 if (!NSW && ICmpInst::isSigned(Pred))
2870 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2871 RBO->getOperand(0), Q, MaxRecurse - 1))
2880 /// Simplify integer comparisons where at least one operand of the compare
2881 /// matches an integer min/max idiom.
2882 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2883 Value *RHS, const SimplifyQuery &Q,
2884 unsigned MaxRecurse) {
2885 Type *ITy = GetCompareTy(LHS); // The return type.
2887 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2888 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2890 // Signed variants on "max(a,b)>=a -> true".
2891 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2893 std::swap(A, B); // smax(A, B) pred A.
2894 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2895 // We analyze this as smax(A, B) pred A.
2897 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2898 (A == LHS || B == LHS)) {
2900 std::swap(A, B); // A pred smax(A, B).
2901 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2902 // We analyze this as smax(A, B) swapped-pred A.
2903 P = CmpInst::getSwappedPredicate(Pred);
2904 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2905 (A == RHS || B == RHS)) {
2907 std::swap(A, B); // smin(A, B) pred A.
2908 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2909 // We analyze this as smax(-A, -B) swapped-pred -A.
2910 // Note that we do not need to actually form -A or -B thanks to EqP.
2911 P = CmpInst::getSwappedPredicate(Pred);
2912 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2913 (A == LHS || B == LHS)) {
2915 std::swap(A, B); // A pred smin(A, B).
2916 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2917 // We analyze this as smax(-A, -B) pred -A.
2918 // Note that we do not need to actually form -A or -B thanks to EqP.
2921 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2922 // Cases correspond to "max(A, B) p A".
2926 case CmpInst::ICMP_EQ:
2927 case CmpInst::ICMP_SLE:
2928 // Equivalent to "A EqP B". This may be the same as the condition tested
2929 // in the max/min; if so, we can just return that.
2930 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2932 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2934 // Otherwise, see if "A EqP B" simplifies.
2936 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2939 case CmpInst::ICMP_NE:
2940 case CmpInst::ICMP_SGT: {
2941 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2942 // Equivalent to "A InvEqP B". This may be the same as the condition
2943 // tested in the max/min; if so, we can just return that.
2944 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2946 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2948 // Otherwise, see if "A InvEqP B" simplifies.
2950 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2954 case CmpInst::ICMP_SGE:
2956 return getTrue(ITy);
2957 case CmpInst::ICMP_SLT:
2959 return getFalse(ITy);
2963 // Unsigned variants on "max(a,b)>=a -> true".
2964 P = CmpInst::BAD_ICMP_PREDICATE;
2965 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2967 std::swap(A, B); // umax(A, B) pred A.
2968 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2969 // We analyze this as umax(A, B) pred A.
2971 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2972 (A == LHS || B == LHS)) {
2974 std::swap(A, B); // A pred umax(A, B).
2975 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2976 // We analyze this as umax(A, B) swapped-pred A.
2977 P = CmpInst::getSwappedPredicate(Pred);
2978 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2979 (A == RHS || B == RHS)) {
2981 std::swap(A, B); // umin(A, B) pred A.
2982 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2983 // We analyze this as umax(-A, -B) swapped-pred -A.
2984 // Note that we do not need to actually form -A or -B thanks to EqP.
2985 P = CmpInst::getSwappedPredicate(Pred);
2986 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2987 (A == LHS || B == LHS)) {
2989 std::swap(A, B); // A pred umin(A, B).
2990 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2991 // We analyze this as umax(-A, -B) pred -A.
2992 // Note that we do not need to actually form -A or -B thanks to EqP.
2995 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2996 // Cases correspond to "max(A, B) p A".
3000 case CmpInst::ICMP_EQ:
3001 case CmpInst::ICMP_ULE:
3002 // Equivalent to "A EqP B". This may be the same as the condition tested
3003 // in the max/min; if so, we can just return that.
3004 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3006 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3008 // Otherwise, see if "A EqP B" simplifies.
3010 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3013 case CmpInst::ICMP_NE:
3014 case CmpInst::ICMP_UGT: {
3015 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3016 // Equivalent to "A InvEqP B". This may be the same as the condition
3017 // tested in the max/min; if so, we can just return that.
3018 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3020 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3022 // Otherwise, see if "A InvEqP B" simplifies.
3024 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3028 case CmpInst::ICMP_UGE:
3030 return getTrue(ITy);
3031 case CmpInst::ICMP_ULT:
3033 return getFalse(ITy);
3037 // Variants on "max(x,y) >= min(x,z)".
3039 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3040 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3041 (A == C || A == D || B == C || B == D)) {
3042 // max(x, ?) pred min(x, ?).
3043 if (Pred == CmpInst::ICMP_SGE)
3045 return getTrue(ITy);
3046 if (Pred == CmpInst::ICMP_SLT)
3048 return getFalse(ITy);
3049 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3050 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3051 (A == C || A == D || B == C || B == D)) {
3052 // min(x, ?) pred max(x, ?).
3053 if (Pred == CmpInst::ICMP_SLE)
3055 return getTrue(ITy);
3056 if (Pred == CmpInst::ICMP_SGT)
3058 return getFalse(ITy);
3059 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3060 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3061 (A == C || A == D || B == C || B == D)) {
3062 // max(x, ?) pred min(x, ?).
3063 if (Pred == CmpInst::ICMP_UGE)
3065 return getTrue(ITy);
3066 if (Pred == CmpInst::ICMP_ULT)
3068 return getFalse(ITy);
3069 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3070 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3071 (A == C || A == D || B == C || B == D)) {
3072 // min(x, ?) pred max(x, ?).
3073 if (Pred == CmpInst::ICMP_ULE)
3075 return getTrue(ITy);
3076 if (Pred == CmpInst::ICMP_UGT)
3078 return getFalse(ITy);
3084 /// Given operands for an ICmpInst, see if we can fold the result.
3085 /// If not, this returns null.
3086 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3087 const SimplifyQuery &Q, unsigned MaxRecurse) {
3088 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3089 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3091 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3092 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3093 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3095 // If we have a constant, make sure it is on the RHS.
3096 std::swap(LHS, RHS);
3097 Pred = CmpInst::getSwappedPredicate(Pred);
3100 Type *ITy = GetCompareTy(LHS); // The return type.
3102 // icmp X, X -> true/false
3103 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3104 // because X could be 0.
3105 if (LHS == RHS || isa<UndefValue>(RHS))
3106 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3108 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3111 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3114 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3117 // If both operands have range metadata, use the metadata
3118 // to simplify the comparison.
3119 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3120 auto RHS_Instr = cast<Instruction>(RHS);
3121 auto LHS_Instr = cast<Instruction>(LHS);
3123 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3124 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3125 auto RHS_CR = getConstantRangeFromMetadata(
3126 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3127 auto LHS_CR = getConstantRangeFromMetadata(
3128 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3130 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3131 if (Satisfied_CR.contains(LHS_CR))
3132 return ConstantInt::getTrue(RHS->getContext());
3134 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3135 CmpInst::getInversePredicate(Pred), RHS_CR);
3136 if (InversedSatisfied_CR.contains(LHS_CR))
3137 return ConstantInt::getFalse(RHS->getContext());
3141 // Compare of cast, for example (zext X) != 0 -> X != 0
3142 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3143 Instruction *LI = cast<CastInst>(LHS);
3144 Value *SrcOp = LI->getOperand(0);
3145 Type *SrcTy = SrcOp->getType();
3146 Type *DstTy = LI->getType();
3148 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3149 // if the integer type is the same size as the pointer type.
3150 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3151 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3152 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3153 // Transfer the cast to the constant.
3154 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3155 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3158 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3159 if (RI->getOperand(0)->getType() == SrcTy)
3160 // Compare without the cast.
3161 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3167 if (isa<ZExtInst>(LHS)) {
3168 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3170 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3171 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3172 // Compare X and Y. Note that signed predicates become unsigned.
3173 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3174 SrcOp, RI->getOperand(0), Q,
3178 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3179 // too. If not, then try to deduce the result of the comparison.
3180 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3181 // Compute the constant that would happen if we truncated to SrcTy then
3182 // reextended to DstTy.
3183 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3184 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3186 // If the re-extended constant didn't change then this is effectively
3187 // also a case of comparing two zero-extended values.
3188 if (RExt == CI && MaxRecurse)
3189 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3190 SrcOp, Trunc, Q, MaxRecurse-1))
3193 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3194 // there. Use this to work out the result of the comparison.
3197 default: llvm_unreachable("Unknown ICmp predicate!");
3199 case ICmpInst::ICMP_EQ:
3200 case ICmpInst::ICMP_UGT:
3201 case ICmpInst::ICMP_UGE:
3202 return ConstantInt::getFalse(CI->getContext());
3204 case ICmpInst::ICMP_NE:
3205 case ICmpInst::ICMP_ULT:
3206 case ICmpInst::ICMP_ULE:
3207 return ConstantInt::getTrue(CI->getContext());
3209 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3210 // is non-negative then LHS <s RHS.
3211 case ICmpInst::ICMP_SGT:
3212 case ICmpInst::ICMP_SGE:
3213 return CI->getValue().isNegative() ?
3214 ConstantInt::getTrue(CI->getContext()) :
3215 ConstantInt::getFalse(CI->getContext());
3217 case ICmpInst::ICMP_SLT:
3218 case ICmpInst::ICMP_SLE:
3219 return CI->getValue().isNegative() ?
3220 ConstantInt::getFalse(CI->getContext()) :
3221 ConstantInt::getTrue(CI->getContext());
3227 if (isa<SExtInst>(LHS)) {
3228 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3230 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3231 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3232 // Compare X and Y. Note that the predicate does not change.
3233 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3237 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3238 // too. If not, then try to deduce the result of the comparison.
3239 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3240 // Compute the constant that would happen if we truncated to SrcTy then
3241 // reextended to DstTy.
3242 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3243 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3245 // If the re-extended constant didn't change then this is effectively
3246 // also a case of comparing two sign-extended values.
3247 if (RExt == CI && MaxRecurse)
3248 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3251 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3252 // bits there. Use this to work out the result of the comparison.
3255 default: llvm_unreachable("Unknown ICmp predicate!");
3256 case ICmpInst::ICMP_EQ:
3257 return ConstantInt::getFalse(CI->getContext());
3258 case ICmpInst::ICMP_NE:
3259 return ConstantInt::getTrue(CI->getContext());
3261 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3263 case ICmpInst::ICMP_SGT:
3264 case ICmpInst::ICMP_SGE:
3265 return CI->getValue().isNegative() ?
3266 ConstantInt::getTrue(CI->getContext()) :
3267 ConstantInt::getFalse(CI->getContext());
3268 case ICmpInst::ICMP_SLT:
3269 case ICmpInst::ICMP_SLE:
3270 return CI->getValue().isNegative() ?
3271 ConstantInt::getFalse(CI->getContext()) :
3272 ConstantInt::getTrue(CI->getContext());
3274 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3276 case ICmpInst::ICMP_UGT:
3277 case ICmpInst::ICMP_UGE:
3278 // Comparison is true iff the LHS <s 0.
3280 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3281 Constant::getNullValue(SrcTy),
3285 case ICmpInst::ICMP_ULT:
3286 case ICmpInst::ICMP_ULE:
3287 // Comparison is true iff the LHS >=s 0.
3289 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3290 Constant::getNullValue(SrcTy),
3300 // icmp eq|ne X, Y -> false|true if X != Y
3301 if (ICmpInst::isEquality(Pred) &&
3302 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3303 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3306 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3309 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3312 // Simplify comparisons of related pointers using a powerful, recursive
3313 // GEP-walk when we have target data available..
3314 if (LHS->getType()->isPointerTy())
3315 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS))
3317 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3318 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3319 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3320 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3321 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3322 Q.DL.getTypeSizeInBits(CRHS->getType()))
3323 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI,
3324 CLHS->getPointerOperand(),
3325 CRHS->getPointerOperand()))
3328 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3329 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3330 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3331 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3332 (ICmpInst::isEquality(Pred) ||
3333 (GLHS->isInBounds() && GRHS->isInBounds() &&
3334 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3335 // The bases are equal and the indices are constant. Build a constant
3336 // expression GEP with the same indices and a null base pointer to see
3337 // what constant folding can make out of it.
3338 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3339 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3340 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3341 GLHS->getSourceElementType(), Null, IndicesLHS);
3343 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3344 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3345 GLHS->getSourceElementType(), Null, IndicesRHS);
3346 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3351 // If the comparison is with the result of a select instruction, check whether
3352 // comparing with either branch of the select always yields the same value.
3353 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3354 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3357 // If the comparison is with the result of a phi instruction, check whether
3358 // doing the compare with each incoming phi value yields a common result.
3359 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3360 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3366 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3367 const SimplifyQuery &Q) {
3368 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3371 /// Given operands for an FCmpInst, see if we can fold the result.
3372 /// If not, this returns null.
3373 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3374 FastMathFlags FMF, const SimplifyQuery &Q,
3375 unsigned MaxRecurse) {
3376 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3377 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3379 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3380 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3381 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3383 // If we have a constant, make sure it is on the RHS.
3384 std::swap(LHS, RHS);
3385 Pred = CmpInst::getSwappedPredicate(Pred);
3388 // Fold trivial predicates.
3389 Type *RetTy = GetCompareTy(LHS);
3390 if (Pred == FCmpInst::FCMP_FALSE)
3391 return getFalse(RetTy);
3392 if (Pred == FCmpInst::FCMP_TRUE)
3393 return getTrue(RetTy);
3395 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3397 if (Pred == FCmpInst::FCMP_UNO)
3398 return getFalse(RetTy);
3399 if (Pred == FCmpInst::FCMP_ORD)
3400 return getTrue(RetTy);
3403 // fcmp pred x, undef and fcmp pred undef, x
3404 // fold to true if unordered, false if ordered
3405 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3406 // Choosing NaN for the undef will always make unordered comparison succeed
3407 // and ordered comparison fail.
3408 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3411 // fcmp x,x -> true/false. Not all compares are foldable.
3413 if (CmpInst::isTrueWhenEqual(Pred))
3414 return getTrue(RetTy);
3415 if (CmpInst::isFalseWhenEqual(Pred))
3416 return getFalse(RetTy);
3419 // Handle fcmp with constant RHS
3420 const ConstantFP *CFP = nullptr;
3421 if (const auto *RHSC = dyn_cast<Constant>(RHS)) {
3422 if (RHS->getType()->isVectorTy())
3423 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue());
3425 CFP = dyn_cast<ConstantFP>(RHSC);
3428 // If the constant is a nan, see if we can fold the comparison based on it.
3429 if (CFP->getValueAPF().isNaN()) {
3430 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3431 return getFalse(RetTy);
3432 assert(FCmpInst::isUnordered(Pred) &&
3433 "Comparison must be either ordered or unordered!");
3434 // True if unordered.
3435 return getTrue(RetTy);
3437 // Check whether the constant is an infinity.
3438 if (CFP->getValueAPF().isInfinity()) {
3439 if (CFP->getValueAPF().isNegative()) {
3441 case FCmpInst::FCMP_OLT:
3442 // No value is ordered and less than negative infinity.
3443 return getFalse(RetTy);
3444 case FCmpInst::FCMP_UGE:
3445 // All values are unordered with or at least negative infinity.
3446 return getTrue(RetTy);
3452 case FCmpInst::FCMP_OGT:
3453 // No value is ordered and greater than infinity.
3454 return getFalse(RetTy);
3455 case FCmpInst::FCMP_ULE:
3456 // All values are unordered with and at most infinity.
3457 return getTrue(RetTy);
3463 if (CFP->getValueAPF().isZero()) {
3465 case FCmpInst::FCMP_UGE:
3466 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3467 return getTrue(RetTy);
3469 case FCmpInst::FCMP_OLT:
3471 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3472 return getFalse(RetTy);
3480 // If the comparison is with the result of a select instruction, check whether
3481 // comparing with either branch of the select always yields the same value.
3482 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3483 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3486 // If the comparison is with the result of a phi instruction, check whether
3487 // doing the compare with each incoming phi value yields a common result.
3488 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3489 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3495 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3496 FastMathFlags FMF, const SimplifyQuery &Q) {
3497 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3500 /// See if V simplifies when its operand Op is replaced with RepOp.
3501 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3502 const SimplifyQuery &Q,
3503 unsigned MaxRecurse) {
3504 // Trivial replacement.
3508 // We cannot replace a constant, and shouldn't even try.
3509 if (isa<Constant>(Op))
3512 auto *I = dyn_cast<Instruction>(V);
3516 // If this is a binary operator, try to simplify it with the replaced op.
3517 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3519 // %cmp = icmp eq i32 %x, 2147483647
3520 // %add = add nsw i32 %x, 1
3521 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3523 // We can't replace %sel with %add unless we strip away the flags.
3524 if (isa<OverflowingBinaryOperator>(B))
3525 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3527 if (isa<PossiblyExactOperator>(B))
3532 if (B->getOperand(0) == Op)
3533 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3535 if (B->getOperand(1) == Op)
3536 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3541 // Same for CmpInsts.
3542 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3544 if (C->getOperand(0) == Op)
3545 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3547 if (C->getOperand(1) == Op)
3548 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3553 // TODO: We could hand off more cases to instsimplify here.
3555 // If all operands are constant after substituting Op for RepOp then we can
3556 // constant fold the instruction.
3557 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3558 // Build a list of all constant operands.
3559 SmallVector<Constant *, 8> ConstOps;
3560 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3561 if (I->getOperand(i) == Op)
3562 ConstOps.push_back(CRepOp);
3563 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3564 ConstOps.push_back(COp);
3569 // All operands were constants, fold it.
3570 if (ConstOps.size() == I->getNumOperands()) {
3571 if (CmpInst *C = dyn_cast<CmpInst>(I))
3572 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3573 ConstOps[1], Q.DL, Q.TLI);
3575 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3576 if (!LI->isVolatile())
3577 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3579 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3586 /// Try to simplify a select instruction when its condition operand is an
3587 /// integer comparison where one operand of the compare is a constant.
3588 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3589 const APInt *Y, bool TrueWhenUnset) {
3592 // (X & Y) == 0 ? X & ~Y : X --> X
3593 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3594 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3596 return TrueWhenUnset ? FalseVal : TrueVal;
3598 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3599 // (X & Y) != 0 ? X : X & ~Y --> X
3600 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3602 return TrueWhenUnset ? FalseVal : TrueVal;
3604 if (Y->isPowerOf2()) {
3605 // (X & Y) == 0 ? X | Y : X --> X | Y
3606 // (X & Y) != 0 ? X | Y : X --> X
3607 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3609 return TrueWhenUnset ? TrueVal : FalseVal;
3611 // (X & Y) == 0 ? X : X | Y --> X
3612 // (X & Y) != 0 ? X : X | Y --> X | Y
3613 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3615 return TrueWhenUnset ? TrueVal : FalseVal;
3621 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3623 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal,
3625 bool TrueWhenUnset) {
3626 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
3630 APInt MinSignedValue;
3632 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) {
3633 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0
3634 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0
3635 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits();
3636 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth);
3638 // icmp slt X, 0 <--> icmp ne (and X, C), 0
3639 // icmp sgt X, -1 <--> icmp eq (and X, C), 0
3641 MinSignedValue = APInt::getSignedMinValue(BitWidth);
3644 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue,
3651 /// Try to simplify a select instruction when its condition operand is an
3652 /// integer comparison.
3653 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3654 Value *FalseVal, const SimplifyQuery &Q,
3655 unsigned MaxRecurse) {
3656 ICmpInst::Predicate Pred;
3657 Value *CmpLHS, *CmpRHS;
3658 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3661 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring
3662 // decomposeBitTestICmp() might help.
3663 // FIXME this should support ICMP_SLE/SGE forms as well
3664 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3667 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3668 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3669 Pred == ICmpInst::ICMP_EQ))
3671 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3672 // Comparing signed-less-than 0 checks if the sign bit is set.
3673 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3676 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3677 // Comparing signed-greater-than -1 checks if the sign bit is not set.
3678 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3683 if (CondVal->hasOneUse()) {
3685 if (match(CmpRHS, m_APInt(C))) {
3686 // X < MIN ? T : F --> F
3687 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3689 // X < MIN ? T : F --> F
3690 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3692 // X > MAX ? T : F --> F
3693 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3695 // X > MAX ? T : F --> F
3696 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3701 // If we have an equality comparison, then we know the value in one of the
3702 // arms of the select. See if substituting this value into the arm and
3703 // simplifying the result yields the same value as the other arm.
3704 if (Pred == ICmpInst::ICMP_EQ) {
3705 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3707 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3710 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3712 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3715 } else if (Pred == ICmpInst::ICMP_NE) {
3716 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3718 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3721 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3723 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3731 /// Given operands for a SelectInst, see if we can fold the result.
3732 /// If not, this returns null.
3733 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3734 Value *FalseVal, const SimplifyQuery &Q,
3735 unsigned MaxRecurse) {
3736 // select true, X, Y -> X
3737 // select false, X, Y -> Y
3738 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3739 if (CB->isAllOnesValue())
3741 if (CB->isNullValue())
3745 // select C, X, X -> X
3746 if (TrueVal == FalseVal)
3749 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3750 if (isa<Constant>(FalseVal))
3754 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3756 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3760 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3766 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3767 const SimplifyQuery &Q) {
3768 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3771 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3772 /// If not, this returns null.
3773 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3774 const SimplifyQuery &Q, unsigned) {
3775 // The type of the GEP pointer operand.
3777 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3779 // getelementptr P -> P.
3780 if (Ops.size() == 1)
3783 // Compute the (pointer) type returned by the GEP instruction.
3784 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3785 Type *GEPTy = PointerType::get(LastType, AS);
3786 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3787 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3788 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3789 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3791 if (isa<UndefValue>(Ops[0]))
3792 return UndefValue::get(GEPTy);
3794 if (Ops.size() == 2) {
3795 // getelementptr P, 0 -> P.
3796 if (match(Ops[1], m_Zero()))
3800 if (Ty->isSized()) {
3803 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3804 // getelementptr P, N -> P if P points to a type of zero size.
3805 if (TyAllocSize == 0)
3808 // The following transforms are only safe if the ptrtoint cast
3809 // doesn't truncate the pointers.
3810 if (Ops[1]->getType()->getScalarSizeInBits() ==
3811 Q.DL.getPointerSizeInBits(AS)) {
3812 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3813 if (match(P, m_Zero()))
3814 return Constant::getNullValue(GEPTy);
3816 if (match(P, m_PtrToInt(m_Value(Temp))))
3817 if (Temp->getType() == GEPTy)
3822 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3823 if (TyAllocSize == 1 &&
3824 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3825 if (Value *R = PtrToIntOrZero(P))
3828 // getelementptr V, (ashr (sub P, V), C) -> Q
3829 // if P points to a type of size 1 << C.
3831 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3832 m_ConstantInt(C))) &&
3833 TyAllocSize == 1ULL << C)
3834 if (Value *R = PtrToIntOrZero(P))
3837 // getelementptr V, (sdiv (sub P, V), C) -> Q
3838 // if P points to a type of size C.
3840 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3841 m_SpecificInt(TyAllocSize))))
3842 if (Value *R = PtrToIntOrZero(P))
3848 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3849 all_of(Ops.slice(1).drop_back(1),
3850 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3852 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3853 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3854 APInt BasePtrOffset(PtrWidth, 0);
3855 Value *StrippedBasePtr =
3856 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3859 // gep (gep V, C), (sub 0, V) -> C
3860 if (match(Ops.back(),
3861 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3862 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3863 return ConstantExpr::getIntToPtr(CI, GEPTy);
3865 // gep (gep V, C), (xor V, -1) -> C-1
3866 if (match(Ops.back(),
3867 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3868 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3869 return ConstantExpr::getIntToPtr(CI, GEPTy);
3874 // Check to see if this is constant foldable.
3875 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3878 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3880 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3885 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3886 const SimplifyQuery &Q) {
3887 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3890 /// Given operands for an InsertValueInst, see if we can fold the result.
3891 /// If not, this returns null.
3892 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3893 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3895 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3896 if (Constant *CVal = dyn_cast<Constant>(Val))
3897 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3899 // insertvalue x, undef, n -> x
3900 if (match(Val, m_Undef()))
3903 // insertvalue x, (extractvalue y, n), n
3904 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3905 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3906 EV->getIndices() == Idxs) {
3907 // insertvalue undef, (extractvalue y, n), n -> y
3908 if (match(Agg, m_Undef()))
3909 return EV->getAggregateOperand();
3911 // insertvalue y, (extractvalue y, n), n -> y
3912 if (Agg == EV->getAggregateOperand())
3919 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3920 ArrayRef<unsigned> Idxs,
3921 const SimplifyQuery &Q) {
3922 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3925 /// Given operands for an ExtractValueInst, see if we can fold the result.
3926 /// If not, this returns null.
3927 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3928 const SimplifyQuery &, unsigned) {
3929 if (auto *CAgg = dyn_cast<Constant>(Agg))
3930 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3932 // extractvalue x, (insertvalue y, elt, n), n -> elt
3933 unsigned NumIdxs = Idxs.size();
3934 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3935 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3936 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3937 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3938 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3939 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3940 Idxs.slice(0, NumCommonIdxs)) {
3941 if (NumIdxs == NumInsertValueIdxs)
3942 return IVI->getInsertedValueOperand();
3950 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3951 const SimplifyQuery &Q) {
3952 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3955 /// Given operands for an ExtractElementInst, see if we can fold the result.
3956 /// If not, this returns null.
3957 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3959 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3960 if (auto *CIdx = dyn_cast<Constant>(Idx))
3961 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3963 // The index is not relevant if our vector is a splat.
3964 if (auto *Splat = CVec->getSplatValue())
3967 if (isa<UndefValue>(Vec))
3968 return UndefValue::get(Vec->getType()->getVectorElementType());
3971 // If extracting a specified index from the vector, see if we can recursively
3972 // find a previously computed scalar that was inserted into the vector.
3973 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3974 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3980 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3981 const SimplifyQuery &Q) {
3982 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3985 /// See if we can fold the given phi. If not, returns null.
3986 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3987 // If all of the PHI's incoming values are the same then replace the PHI node
3988 // with the common value.
3989 Value *CommonValue = nullptr;
3990 bool HasUndefInput = false;
3991 for (Value *Incoming : PN->incoming_values()) {
3992 // If the incoming value is the phi node itself, it can safely be skipped.
3993 if (Incoming == PN) continue;
3994 if (isa<UndefValue>(Incoming)) {
3995 // Remember that we saw an undef value, but otherwise ignore them.
3996 HasUndefInput = true;
3999 if (CommonValue && Incoming != CommonValue)
4000 return nullptr; // Not the same, bail out.
4001 CommonValue = Incoming;
4004 // If CommonValue is null then all of the incoming values were either undef or
4005 // equal to the phi node itself.
4007 return UndefValue::get(PN->getType());
4009 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4010 // instruction, we cannot return X as the result of the PHI node unless it
4011 // dominates the PHI block.
4013 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4018 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4019 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4020 if (auto *C = dyn_cast<Constant>(Op))
4021 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4023 if (auto *CI = dyn_cast<CastInst>(Op)) {
4024 auto *Src = CI->getOperand(0);
4025 Type *SrcTy = Src->getType();
4026 Type *MidTy = CI->getType();
4028 if (Src->getType() == Ty) {
4029 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4030 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4032 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4034 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4036 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4037 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4038 SrcIntPtrTy, MidIntPtrTy,
4039 DstIntPtrTy) == Instruction::BitCast)
4045 if (CastOpc == Instruction::BitCast)
4046 if (Op->getType() == Ty)
4052 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4053 const SimplifyQuery &Q) {
4054 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4057 /// For the given destination element of a shuffle, peek through shuffles to
4058 /// match a root vector source operand that contains that element in the same
4059 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4060 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4061 int MaskVal, Value *RootVec,
4062 unsigned MaxRecurse) {
4066 // Bail out if any mask value is undefined. That kind of shuffle may be
4067 // simplified further based on demanded bits or other folds.
4071 // The mask value chooses which source operand we need to look at next.
4072 int InVecNumElts = Op0->getType()->getVectorNumElements();
4073 int RootElt = MaskVal;
4074 Value *SourceOp = Op0;
4075 if (MaskVal >= InVecNumElts) {
4076 RootElt = MaskVal - InVecNumElts;
4080 // If the source operand is a shuffle itself, look through it to find the
4081 // matching root vector.
4082 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4083 return foldIdentityShuffles(
4084 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4085 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4088 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4091 // The source operand is not a shuffle. Initialize the root vector value for
4092 // this shuffle if that has not been done yet.
4096 // Give up as soon as a source operand does not match the existing root value.
4097 if (RootVec != SourceOp)
4100 // The element must be coming from the same lane in the source vector
4101 // (although it may have crossed lanes in intermediate shuffles).
4102 if (RootElt != DestElt)
4108 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4109 Type *RetTy, const SimplifyQuery &Q,
4110 unsigned MaxRecurse) {
4111 if (isa<UndefValue>(Mask))
4112 return UndefValue::get(RetTy);
4114 Type *InVecTy = Op0->getType();
4115 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4116 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4118 SmallVector<int, 32> Indices;
4119 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4120 assert(MaskNumElts == Indices.size() &&
4121 "Size of Indices not same as number of mask elements?");
4123 // Canonicalization: If mask does not select elements from an input vector,
4124 // replace that input vector with undef.
4125 bool MaskSelects0 = false, MaskSelects1 = false;
4126 for (unsigned i = 0; i != MaskNumElts; ++i) {
4127 if (Indices[i] == -1)
4129 if ((unsigned)Indices[i] < InVecNumElts)
4130 MaskSelects0 = true;
4132 MaskSelects1 = true;
4135 Op0 = UndefValue::get(InVecTy);
4137 Op1 = UndefValue::get(InVecTy);
4139 auto *Op0Const = dyn_cast<Constant>(Op0);
4140 auto *Op1Const = dyn_cast<Constant>(Op1);
4142 // If all operands are constant, constant fold the shuffle.
4143 if (Op0Const && Op1Const)
4144 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4146 // Canonicalization: if only one input vector is constant, it shall be the
4148 if (Op0Const && !Op1Const) {
4149 std::swap(Op0, Op1);
4150 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4153 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4154 // value type is same as the input vectors' type.
4155 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4156 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4157 OpShuf->getMask()->getSplatValue())
4160 // Don't fold a shuffle with undef mask elements. This may get folded in a
4161 // better way using demanded bits or other analysis.
4162 // TODO: Should we allow this?
4163 if (find(Indices, -1) != Indices.end())
4166 // Check if every element of this shuffle can be mapped back to the
4167 // corresponding element of a single root vector. If so, we don't need this
4168 // shuffle. This handles simple identity shuffles as well as chains of
4169 // shuffles that may widen/narrow and/or move elements across lanes and back.
4170 Value *RootVec = nullptr;
4171 for (unsigned i = 0; i != MaskNumElts; ++i) {
4172 // Note that recursion is limited for each vector element, so if any element
4173 // exceeds the limit, this will fail to simplify.
4175 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4177 // We can't replace a widening/narrowing shuffle with one of its operands.
4178 if (!RootVec || RootVec->getType() != RetTy)
4184 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4185 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4186 Type *RetTy, const SimplifyQuery &Q) {
4187 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4190 //=== Helper functions for higher up the class hierarchy.
4192 /// Given operands for a BinaryOperator, see if we can fold the result.
4193 /// If not, this returns null.
4194 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4195 const SimplifyQuery &Q, unsigned MaxRecurse) {
4197 case Instruction::Add:
4198 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4199 case Instruction::FAdd:
4200 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4201 case Instruction::Sub:
4202 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4203 case Instruction::FSub:
4204 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4205 case Instruction::Mul:
4206 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4207 case Instruction::FMul:
4208 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4209 case Instruction::SDiv:
4210 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4211 case Instruction::UDiv:
4212 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4213 case Instruction::FDiv:
4214 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4215 case Instruction::SRem:
4216 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4217 case Instruction::URem:
4218 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4219 case Instruction::FRem:
4220 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4221 case Instruction::Shl:
4222 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4223 case Instruction::LShr:
4224 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4225 case Instruction::AShr:
4226 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4227 case Instruction::And:
4228 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4229 case Instruction::Or:
4230 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4231 case Instruction::Xor:
4232 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4234 llvm_unreachable("Unexpected opcode");
4238 /// Given operands for a BinaryOperator, see if we can fold the result.
4239 /// If not, this returns null.
4240 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4241 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4242 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4243 const FastMathFlags &FMF, const SimplifyQuery &Q,
4244 unsigned MaxRecurse) {
4246 case Instruction::FAdd:
4247 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4248 case Instruction::FSub:
4249 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4250 case Instruction::FMul:
4251 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4252 case Instruction::FDiv:
4253 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4255 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4259 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4260 const SimplifyQuery &Q) {
4261 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4264 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4265 FastMathFlags FMF, const SimplifyQuery &Q) {
4266 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4269 /// Given operands for a CmpInst, see if we can fold the result.
4270 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4271 const SimplifyQuery &Q, unsigned MaxRecurse) {
4272 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4273 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4274 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4277 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4278 const SimplifyQuery &Q) {
4279 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4282 static bool IsIdempotent(Intrinsic::ID ID) {
4284 default: return false;
4286 // Unary idempotent: f(f(x)) = f(x)
4287 case Intrinsic::fabs:
4288 case Intrinsic::floor:
4289 case Intrinsic::ceil:
4290 case Intrinsic::trunc:
4291 case Intrinsic::rint:
4292 case Intrinsic::nearbyint:
4293 case Intrinsic::round:
4298 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4299 const DataLayout &DL) {
4300 GlobalValue *PtrSym;
4302 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4305 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4306 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4307 Type *Int32PtrTy = Int32Ty->getPointerTo();
4308 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4310 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4311 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4314 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4315 if (OffsetInt % 4 != 0)
4318 Constant *C = ConstantExpr::getGetElementPtr(
4319 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4320 ConstantInt::get(Int64Ty, OffsetInt / 4));
4321 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4325 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4329 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4330 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4335 if (LoadedCE->getOpcode() != Instruction::Sub)
4338 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4339 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4341 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4343 Constant *LoadedRHS = LoadedCE->getOperand(1);
4344 GlobalValue *LoadedRHSSym;
4345 APInt LoadedRHSOffset;
4346 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4348 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4351 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4354 static bool maskIsAllZeroOrUndef(Value *Mask) {
4355 auto *ConstMask = dyn_cast<Constant>(Mask);
4358 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4360 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4362 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4363 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4370 template <typename IterTy>
4371 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4372 const SimplifyQuery &Q, unsigned MaxRecurse) {
4373 Intrinsic::ID IID = F->getIntrinsicID();
4374 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4377 if (NumOperands == 1) {
4378 // Perform idempotent optimizations
4379 if (IsIdempotent(IID)) {
4380 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4381 if (II->getIntrinsicID() == IID)
4387 case Intrinsic::fabs: {
4388 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4398 if (NumOperands == 2) {
4399 Value *LHS = *ArgBegin;
4400 Value *RHS = *(ArgBegin + 1);
4401 Type *ReturnType = F->getReturnType();
4404 case Intrinsic::usub_with_overflow:
4405 case Intrinsic::ssub_with_overflow: {
4406 // X - X -> { 0, false }
4408 return Constant::getNullValue(ReturnType);
4410 // X - undef -> undef
4411 // undef - X -> undef
4412 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4413 return UndefValue::get(ReturnType);
4417 case Intrinsic::uadd_with_overflow:
4418 case Intrinsic::sadd_with_overflow: {
4419 // X + undef -> undef
4420 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4421 return UndefValue::get(ReturnType);
4425 case Intrinsic::umul_with_overflow:
4426 case Intrinsic::smul_with_overflow: {
4427 // 0 * X -> { 0, false }
4428 // X * 0 -> { 0, false }
4429 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4430 return Constant::getNullValue(ReturnType);
4432 // undef * X -> { 0, false }
4433 // X * undef -> { 0, false }
4434 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4435 return Constant::getNullValue(ReturnType);
4439 case Intrinsic::load_relative: {
4440 Constant *C0 = dyn_cast<Constant>(LHS);
4441 Constant *C1 = dyn_cast<Constant>(RHS);
4443 return SimplifyRelativeLoad(C0, C1, Q.DL);
4451 // Simplify calls to llvm.masked.load.*
4453 case Intrinsic::masked_load: {
4454 Value *MaskArg = ArgBegin[2];
4455 Value *PassthruArg = ArgBegin[3];
4456 // If the mask is all zeros or undef, the "passthru" argument is the result.
4457 if (maskIsAllZeroOrUndef(MaskArg))
4466 template <typename IterTy>
4467 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4468 IterTy ArgEnd, const SimplifyQuery &Q,
4469 unsigned MaxRecurse) {
4470 Type *Ty = V->getType();
4471 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4472 Ty = PTy->getElementType();
4473 FunctionType *FTy = cast<FunctionType>(Ty);
4475 // call undef -> undef
4476 // call null -> undef
4477 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4478 return UndefValue::get(FTy->getReturnType());
4480 Function *F = dyn_cast<Function>(V);
4484 if (F->isIntrinsic())
4485 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4488 if (!canConstantFoldCallTo(CS, F))
4491 SmallVector<Constant *, 4> ConstantArgs;
4492 ConstantArgs.reserve(ArgEnd - ArgBegin);
4493 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4494 Constant *C = dyn_cast<Constant>(*I);
4497 ConstantArgs.push_back(C);
4500 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4503 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4504 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4505 const SimplifyQuery &Q) {
4506 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4509 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4510 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4511 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4514 /// See if we can compute a simplified version of this instruction.
4515 /// If not, this returns null.
4517 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4518 OptimizationRemarkEmitter *ORE) {
4519 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4522 switch (I->getOpcode()) {
4524 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4526 case Instruction::FAdd:
4527 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4528 I->getFastMathFlags(), Q);
4530 case Instruction::Add:
4531 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4532 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4533 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4535 case Instruction::FSub:
4536 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4537 I->getFastMathFlags(), Q);
4539 case Instruction::Sub:
4540 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4541 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4542 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4544 case Instruction::FMul:
4545 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4546 I->getFastMathFlags(), Q);
4548 case Instruction::Mul:
4549 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4551 case Instruction::SDiv:
4552 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4554 case Instruction::UDiv:
4555 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4557 case Instruction::FDiv:
4558 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4559 I->getFastMathFlags(), Q);
4561 case Instruction::SRem:
4562 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4564 case Instruction::URem:
4565 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4567 case Instruction::FRem:
4568 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4569 I->getFastMathFlags(), Q);
4571 case Instruction::Shl:
4572 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4573 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4574 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4576 case Instruction::LShr:
4577 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4578 cast<BinaryOperator>(I)->isExact(), Q);
4580 case Instruction::AShr:
4581 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4582 cast<BinaryOperator>(I)->isExact(), Q);
4584 case Instruction::And:
4585 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4587 case Instruction::Or:
4588 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4590 case Instruction::Xor:
4591 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4593 case Instruction::ICmp:
4594 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4595 I->getOperand(0), I->getOperand(1), Q);
4597 case Instruction::FCmp:
4599 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4600 I->getOperand(1), I->getFastMathFlags(), Q);
4602 case Instruction::Select:
4603 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4604 I->getOperand(2), Q);
4606 case Instruction::GetElementPtr: {
4607 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4608 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4612 case Instruction::InsertValue: {
4613 InsertValueInst *IV = cast<InsertValueInst>(I);
4614 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4615 IV->getInsertedValueOperand(),
4616 IV->getIndices(), Q);
4619 case Instruction::ExtractValue: {
4620 auto *EVI = cast<ExtractValueInst>(I);
4621 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4622 EVI->getIndices(), Q);
4625 case Instruction::ExtractElement: {
4626 auto *EEI = cast<ExtractElementInst>(I);
4627 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4628 EEI->getIndexOperand(), Q);
4631 case Instruction::ShuffleVector: {
4632 auto *SVI = cast<ShuffleVectorInst>(I);
4633 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4634 SVI->getMask(), SVI->getType(), Q);
4637 case Instruction::PHI:
4638 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4640 case Instruction::Call: {
4641 CallSite CS(cast<CallInst>(I));
4642 Result = SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4646 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4647 #include "llvm/IR/Instruction.def"
4648 #undef HANDLE_CAST_INST
4650 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4652 case Instruction::Alloca:
4653 // No simplifications for Alloca and it can't be constant folded.
4658 // In general, it is possible for computeKnownBits to determine all bits in a
4659 // value even when the operands are not all constants.
4660 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4661 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4662 if (Known.isConstant())
4663 Result = ConstantInt::get(I->getType(), Known.getConstant());
4666 /// If called on unreachable code, the above logic may report that the
4667 /// instruction simplified to itself. Make life easier for users by
4668 /// detecting that case here, returning a safe value instead.
4669 return Result == I ? UndefValue::get(I->getType()) : Result;
4672 /// \brief Implementation of recursive simplification through an instruction's
4675 /// This is the common implementation of the recursive simplification routines.
4676 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4677 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4678 /// instructions to process and attempt to simplify it using
4679 /// InstructionSimplify.
4681 /// This routine returns 'true' only when *it* simplifies something. The passed
4682 /// in simplified value does not count toward this.
4683 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4684 const TargetLibraryInfo *TLI,
4685 const DominatorTree *DT,
4686 AssumptionCache *AC) {
4687 bool Simplified = false;
4688 SmallSetVector<Instruction *, 8> Worklist;
4689 const DataLayout &DL = I->getModule()->getDataLayout();
4691 // If we have an explicit value to collapse to, do that round of the
4692 // simplification loop by hand initially.
4694 for (User *U : I->users())
4696 Worklist.insert(cast<Instruction>(U));
4698 // Replace the instruction with its simplified value.
4699 I->replaceAllUsesWith(SimpleV);
4701 // Gracefully handle edge cases where the instruction is not wired into any
4703 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4704 !I->mayHaveSideEffects())
4705 I->eraseFromParent();
4710 // Note that we must test the size on each iteration, the worklist can grow.
4711 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4714 // See if this instruction simplifies.
4715 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4721 // Stash away all the uses of the old instruction so we can check them for
4722 // recursive simplifications after a RAUW. This is cheaper than checking all
4723 // uses of To on the recursive step in most cases.
4724 for (User *U : I->users())
4725 Worklist.insert(cast<Instruction>(U));
4727 // Replace the instruction with its simplified value.
4728 I->replaceAllUsesWith(SimpleV);
4730 // Gracefully handle edge cases where the instruction is not wired into any
4732 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4733 !I->mayHaveSideEffects())
4734 I->eraseFromParent();
4739 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4740 const TargetLibraryInfo *TLI,
4741 const DominatorTree *DT,
4742 AssumptionCache *AC) {
4743 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4746 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4747 const TargetLibraryInfo *TLI,
4748 const DominatorTree *DT,
4749 AssumptionCache *AC) {
4750 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4751 assert(SimpleV && "Must provide a simplified value.");
4752 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4756 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4757 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4758 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4759 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4760 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4761 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4762 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4763 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4766 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4767 const DataLayout &DL) {
4768 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4771 template <class T, class... TArgs>
4772 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4774 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4775 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4776 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4777 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4779 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,