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()->getScalarType()->isIntegerTy(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()->getScalarType()->isPointerTy());
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()->getScalarType()->isPointerTy() &&
631 "Unexpected operand type!");
632 } while (Visited.insert(V).second);
634 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
635 if (V->getType()->isVectorTy())
636 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
641 /// \brief Compute the constant difference between two pointer values.
642 /// If the difference is not a constant, returns zero.
643 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
645 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
646 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
648 // If LHS and RHS are not related via constant offsets to the same base
649 // value, there is nothing we can do here.
653 // Otherwise, the difference of LHS - RHS can be computed as:
655 // = (LHSOffset + Base) - (RHSOffset + Base)
656 // = LHSOffset - RHSOffset
657 return ConstantExpr::getSub(LHSOffset, RHSOffset);
660 /// Given operands for a Sub, see if we can fold the result.
661 /// If not, this returns null.
662 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
663 const SimplifyQuery &Q, unsigned MaxRecurse) {
664 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
667 // X - undef -> undef
668 // undef - X -> undef
669 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
670 return UndefValue::get(Op0->getType());
673 if (match(Op1, m_Zero()))
678 return Constant::getNullValue(Op0->getType());
680 // Is this a negation?
681 if (match(Op0, m_Zero())) {
682 // 0 - X -> 0 if the sub is NUW.
686 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
687 if (Known.Zero.isMaxSignedValue()) {
688 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
689 // Op1 must be 0 because negating the minimum signed value is undefined.
693 // 0 - X -> X if X is 0 or the minimum signed value.
698 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
699 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
700 Value *X = nullptr, *Y = nullptr, *Z = Op1;
701 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
702 // See if "V === Y - Z" simplifies.
703 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
704 // It does! Now see if "X + V" simplifies.
705 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
706 // It does, we successfully reassociated!
710 // See if "V === X - Z" simplifies.
711 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
712 // It does! Now see if "Y + V" simplifies.
713 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
714 // It does, we successfully reassociated!
720 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
721 // For example, X - (X + 1) -> -1
723 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
724 // See if "V === X - Y" simplifies.
725 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
726 // It does! Now see if "V - Z" simplifies.
727 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
728 // It does, we successfully reassociated!
732 // See if "V === X - Z" simplifies.
733 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
734 // It does! Now see if "V - Y" simplifies.
735 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
736 // It does, we successfully reassociated!
742 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
743 // For example, X - (X - Y) -> Y.
745 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
746 // See if "V === Z - X" simplifies.
747 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
748 // It does! Now see if "V + Y" simplifies.
749 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
750 // It does, we successfully reassociated!
755 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
756 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
757 match(Op1, m_Trunc(m_Value(Y))))
758 if (X->getType() == Y->getType())
759 // See if "V === X - Y" simplifies.
760 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
761 // It does! Now see if "trunc V" simplifies.
762 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
764 // It does, return the simplified "trunc V".
767 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
768 if (match(Op0, m_PtrToInt(m_Value(X))) &&
769 match(Op1, m_PtrToInt(m_Value(Y))))
770 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
771 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
774 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
775 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
778 // Threading Sub over selects and phi nodes is pointless, so don't bother.
779 // Threading over the select in "A - select(cond, B, C)" means evaluating
780 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
781 // only if B and C are equal. If B and C are equal then (since we assume
782 // that operands have already been simplified) "select(cond, B, C)" should
783 // have been simplified to the common value of B and C already. Analysing
784 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
785 // for threading over phi nodes.
790 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
791 const SimplifyQuery &Q) {
792 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
795 /// Given operands for an FAdd, see if we can fold the result. If not, this
797 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
798 const SimplifyQuery &Q, unsigned MaxRecurse) {
799 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
803 if (match(Op1, m_NegZero()))
806 // fadd X, 0 ==> X, when we know X is not -0
807 if (match(Op1, m_Zero()) &&
808 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
811 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
812 // where nnan and ninf have to occur at least once somewhere in this
814 Value *SubOp = nullptr;
815 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
817 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
820 Instruction *FSub = cast<Instruction>(SubOp);
821 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
822 (FMF.noInfs() || FSub->hasNoInfs()))
823 return Constant::getNullValue(Op0->getType());
829 /// Given operands for an FSub, see if we can fold the result. If not, this
831 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
832 const SimplifyQuery &Q, unsigned MaxRecurse) {
833 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
837 if (match(Op1, m_Zero()))
840 // fsub X, -0 ==> X, when we know X is not -0
841 if (match(Op1, m_NegZero()) &&
842 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
845 // fsub -0.0, (fsub -0.0, X) ==> X
847 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
850 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
851 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
852 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
855 // fsub nnan x, x ==> 0.0
856 if (FMF.noNaNs() && Op0 == Op1)
857 return Constant::getNullValue(Op0->getType());
862 /// Given the operands for an FMul, see if we can fold the result
863 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
864 const SimplifyQuery &Q, unsigned MaxRecurse) {
865 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
869 if (match(Op1, m_FPOne()))
872 // fmul nnan nsz X, 0 ==> 0
873 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
879 /// Given operands for a Mul, see if we can fold the result.
880 /// If not, this returns null.
881 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
882 unsigned MaxRecurse) {
883 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
887 if (match(Op1, m_Undef()))
888 return Constant::getNullValue(Op0->getType());
891 if (match(Op1, m_Zero()))
895 if (match(Op1, m_One()))
898 // (X / Y) * Y -> X if the division is exact.
900 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
901 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
905 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
906 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
909 // Try some generic simplifications for associative operations.
910 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
914 // Mul distributes over Add. Try some generic simplifications based on this.
915 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
919 // If the operation is with the result of a select instruction, check whether
920 // operating on either branch of the select always yields the same value.
921 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
922 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
926 // If the operation is with the result of a phi instruction, check whether
927 // operating on all incoming values of the phi always yields the same value.
928 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
929 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
936 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
937 const SimplifyQuery &Q) {
938 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
942 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
943 const SimplifyQuery &Q) {
944 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
947 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
948 const SimplifyQuery &Q) {
949 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
952 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
953 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
956 /// Check for common or similar folds of integer division or integer remainder.
957 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
958 Type *Ty = Op0->getType();
960 // X / undef -> undef
961 // X % undef -> undef
962 if (match(Op1, m_Undef()))
967 // We don't need to preserve faults!
968 if (match(Op1, m_Zero()))
969 return UndefValue::get(Ty);
971 // If any element of a constant divisor vector is zero, the whole op is undef.
972 auto *Op1C = dyn_cast<Constant>(Op1);
973 if (Op1C && Ty->isVectorTy()) {
974 unsigned NumElts = Ty->getVectorNumElements();
975 for (unsigned i = 0; i != NumElts; ++i) {
976 Constant *Elt = Op1C->getAggregateElement(i);
977 if (Elt && Elt->isNullValue())
978 return UndefValue::get(Ty);
984 if (match(Op0, m_Undef()))
985 return Constant::getNullValue(Ty);
989 if (match(Op0, m_Zero()))
995 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
999 // If this is a boolean op (single-bit element type), we can't have
1000 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
1001 if (match(Op1, m_One()) || Ty->getScalarType()->isIntegerTy(1))
1002 return IsDiv ? Op0 : Constant::getNullValue(Ty);
1007 /// Given operands for an SDiv or UDiv, see if we can fold the result.
1008 /// If not, this returns null.
1009 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1010 const SimplifyQuery &Q, unsigned MaxRecurse) {
1011 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1014 if (Value *V = simplifyDivRem(Op0, Op1, true))
1017 bool isSigned = Opcode == Instruction::SDiv;
1019 // (X * Y) / Y -> X if the multiplication does not overflow.
1020 Value *X = nullptr, *Y = nullptr;
1021 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1022 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1023 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1024 // If the Mul knows it does not overflow, then we are good to go.
1025 if ((isSigned && Mul->hasNoSignedWrap()) ||
1026 (!isSigned && Mul->hasNoUnsignedWrap()))
1028 // If X has the form X = A / Y then X * Y cannot overflow.
1029 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1030 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1034 // (X rem Y) / Y -> 0
1035 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1036 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1037 return Constant::getNullValue(Op0->getType());
1039 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1040 ConstantInt *C1, *C2;
1041 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1042 match(Op1, m_ConstantInt(C2))) {
1044 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1046 return Constant::getNullValue(Op0->getType());
1049 // If the operation is with the result of a select instruction, check whether
1050 // operating on either branch of the select always yields the same value.
1051 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1052 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1055 // If the operation is with the result of a phi instruction, check whether
1056 // operating on all incoming values of the phi always yields the same value.
1057 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1058 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1064 /// Given operands for an SDiv, see if we can fold the result.
1065 /// If not, this returns null.
1066 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1067 unsigned MaxRecurse) {
1068 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1074 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1075 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1078 /// Given operands for a UDiv, see if we can fold the result.
1079 /// If not, this returns null.
1080 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1081 unsigned MaxRecurse) {
1082 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1085 // udiv %V, C -> 0 if %V < C
1087 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1088 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1089 if (C->isAllOnesValue()) {
1090 return Constant::getNullValue(Op0->getType());
1098 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1099 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1102 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1103 const SimplifyQuery &Q, unsigned) {
1104 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
1107 // undef / X -> undef (the undef could be a snan).
1108 if (match(Op0, m_Undef()))
1111 // X / undef -> undef
1112 if (match(Op1, m_Undef()))
1116 if (match(Op1, m_FPOne()))
1120 // Requires that NaNs are off (X could be zero) and signed zeroes are
1121 // ignored (X could be positive or negative, so the output sign is unknown).
1122 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1126 // X / X -> 1.0 is legal when NaNs are ignored.
1128 return ConstantFP::get(Op0->getType(), 1.0);
1130 // -X / X -> -1.0 and
1131 // X / -X -> -1.0 are legal when NaNs are ignored.
1132 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1133 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1134 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1135 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1136 BinaryOperator::getFNegArgument(Op1) == Op0))
1137 return ConstantFP::get(Op0->getType(), -1.0);
1143 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1144 const SimplifyQuery &Q) {
1145 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
1148 /// Given operands for an SRem or URem, see if we can fold the result.
1149 /// If not, this returns null.
1150 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1151 const SimplifyQuery &Q, unsigned MaxRecurse) {
1152 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1155 if (Value *V = simplifyDivRem(Op0, Op1, false))
1158 // (X % Y) % Y -> X % Y
1159 if ((Opcode == Instruction::SRem &&
1160 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1161 (Opcode == Instruction::URem &&
1162 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1165 // If the operation is with the result of a select instruction, check whether
1166 // operating on either branch of the select always yields the same value.
1167 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1168 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1171 // If the operation is with the result of a phi instruction, check whether
1172 // operating on all incoming values of the phi always yields the same value.
1173 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1174 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1180 /// Given operands for an SRem, see if we can fold the result.
1181 /// If not, this returns null.
1182 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1183 unsigned MaxRecurse) {
1184 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1190 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1191 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1194 /// Given operands for a URem, see if we can fold the result.
1195 /// If not, this returns null.
1196 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1197 unsigned MaxRecurse) {
1198 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1201 // urem %V, C -> %V if %V < C
1203 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1204 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1205 if (C->isAllOnesValue()) {
1214 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1215 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1218 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1219 const SimplifyQuery &Q, unsigned) {
1220 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
1223 // undef % X -> undef (the undef could be a snan).
1224 if (match(Op0, m_Undef()))
1227 // X % undef -> undef
1228 if (match(Op1, m_Undef()))
1232 // Requires that NaNs are off (X could be zero) and signed zeroes are
1233 // ignored (X could be positive or negative, so the output sign is unknown).
1234 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1240 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1241 const SimplifyQuery &Q) {
1242 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
1245 /// Returns true if a shift by \c Amount always yields undef.
1246 static bool isUndefShift(Value *Amount) {
1247 Constant *C = dyn_cast<Constant>(Amount);
1251 // X shift by undef -> undef because it may shift by the bitwidth.
1252 if (isa<UndefValue>(C))
1255 // Shifting by the bitwidth or more is undefined.
1256 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1257 if (CI->getValue().getLimitedValue() >=
1258 CI->getType()->getScalarSizeInBits())
1261 // If all lanes of a vector shift are undefined the whole shift is.
1262 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1263 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1264 if (!isUndefShift(C->getAggregateElement(I)))
1272 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1273 /// If not, this returns null.
1274 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1275 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1276 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1279 // 0 shift by X -> 0
1280 if (match(Op0, m_Zero()))
1283 // X shift by 0 -> X
1284 if (match(Op1, m_Zero()))
1287 // Fold undefined shifts.
1288 if (isUndefShift(Op1))
1289 return UndefValue::get(Op0->getType());
1291 // If the operation is with the result of a select instruction, check whether
1292 // operating on either branch of the select always yields the same value.
1293 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1294 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1297 // If the operation is with the result of a phi instruction, check whether
1298 // operating on all incoming values of the phi always yields the same value.
1299 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1300 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1303 // If any bits in the shift amount make that value greater than or equal to
1304 // the number of bits in the type, the shift is undefined.
1305 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1306 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1307 return UndefValue::get(Op0->getType());
1309 // If all valid bits in the shift amount are known zero, the first operand is
1311 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1312 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1318 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1319 /// fold the result. If not, this returns null.
1320 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1321 Value *Op1, bool isExact, const SimplifyQuery &Q,
1322 unsigned MaxRecurse) {
1323 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1328 return Constant::getNullValue(Op0->getType());
1331 // undef >> X -> undef (if it's exact)
1332 if (match(Op0, m_Undef()))
1333 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1335 // The low bit cannot be shifted out of an exact shift if it is set.
1337 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1338 if (Op0Known.One[0])
1345 /// Given operands for an Shl, see if we can fold the result.
1346 /// If not, this returns null.
1347 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1348 const SimplifyQuery &Q, unsigned MaxRecurse) {
1349 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1353 // undef << X -> undef if (if it's NSW/NUW)
1354 if (match(Op0, m_Undef()))
1355 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1357 // (X >> A) << A -> X
1359 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1364 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1365 const SimplifyQuery &Q) {
1366 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1369 /// Given operands for an LShr, see if we can fold the result.
1370 /// If not, this returns null.
1371 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1372 const SimplifyQuery &Q, unsigned MaxRecurse) {
1373 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1377 // (X << A) >> A -> X
1379 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1385 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1386 const SimplifyQuery &Q) {
1387 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1390 /// Given operands for an AShr, see if we can fold the result.
1391 /// If not, this returns null.
1392 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1393 const SimplifyQuery &Q, unsigned MaxRecurse) {
1394 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1398 // all ones >>a X -> all ones
1399 if (match(Op0, m_AllOnes()))
1402 // (X << A) >> A -> X
1404 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1407 // Arithmetic shifting an all-sign-bit value is a no-op.
1408 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1409 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1415 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1416 const SimplifyQuery &Q) {
1417 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1420 /// Commuted variants are assumed to be handled by calling this function again
1421 /// with the parameters swapped.
1422 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1423 ICmpInst *UnsignedICmp, bool IsAnd) {
1426 ICmpInst::Predicate EqPred;
1427 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1428 !ICmpInst::isEquality(EqPred))
1431 ICmpInst::Predicate UnsignedPred;
1432 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1433 ICmpInst::isUnsigned(UnsignedPred))
1435 else if (match(UnsignedICmp,
1436 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1437 ICmpInst::isUnsigned(UnsignedPred))
1438 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1442 // X < Y && Y != 0 --> X < Y
1443 // X < Y || Y != 0 --> Y != 0
1444 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1445 return IsAnd ? UnsignedICmp : ZeroICmp;
1447 // X >= Y || Y != 0 --> true
1448 // X >= Y || Y == 0 --> X >= Y
1449 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1450 if (EqPred == ICmpInst::ICMP_NE)
1451 return getTrue(UnsignedICmp->getType());
1452 return UnsignedICmp;
1455 // X < Y && Y == 0 --> false
1456 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1458 return getFalse(UnsignedICmp->getType());
1463 /// Commuted variants are assumed to be handled by calling this function again
1464 /// with the parameters swapped.
1465 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1466 ICmpInst::Predicate Pred0, Pred1;
1468 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1469 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1472 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1473 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1474 // can eliminate Op1 from this 'and'.
1475 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1478 // Check for any combination of predicates that are guaranteed to be disjoint.
1479 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1480 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1481 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1482 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1483 return getFalse(Op0->getType());
1488 /// Commuted variants are assumed to be handled by calling this function again
1489 /// with the parameters swapped.
1490 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1491 ICmpInst::Predicate Pred0, Pred1;
1493 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1494 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1497 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1498 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1499 // can eliminate Op0 from this 'or'.
1500 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1503 // Check for any combination of predicates that cover the entire range of
1505 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1506 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1507 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1508 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1509 return getTrue(Op0->getType());
1514 /// Test if a pair of compares with a shared operand and 2 constants has an
1515 /// empty set intersection, full set union, or if one compare is a superset of
1517 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1519 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1520 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1523 const APInt *C0, *C1;
1524 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1525 !match(Cmp1->getOperand(1), m_APInt(C1)))
1528 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1529 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1531 // For and-of-compares, check if the intersection is empty:
1532 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1533 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1534 return getFalse(Cmp0->getType());
1536 // For or-of-compares, check if the union is full:
1537 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1538 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1539 return getTrue(Cmp0->getType());
1541 // Is one range a superset of the other?
1542 // If this is and-of-compares, take the smaller set:
1543 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1544 // If this is or-of-compares, take the larger set:
1545 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1546 if (Range0.contains(Range1))
1547 return IsAnd ? Cmp1 : Cmp0;
1548 if (Range1.contains(Range0))
1549 return IsAnd ? Cmp0 : Cmp1;
1554 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1555 // (icmp (add V, C0), C1) & (icmp V, C0)
1556 ICmpInst::Predicate Pred0, Pred1;
1557 const APInt *C0, *C1;
1559 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1562 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1565 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1566 if (AddInst->getOperand(1) != Op1->getOperand(1))
1569 Type *ITy = Op0->getType();
1570 bool isNSW = AddInst->hasNoSignedWrap();
1571 bool isNUW = AddInst->hasNoUnsignedWrap();
1573 const APInt Delta = *C1 - *C0;
1574 if (C0->isStrictlyPositive()) {
1576 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1577 return getFalse(ITy);
1578 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1579 return getFalse(ITy);
1582 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1583 return getFalse(ITy);
1584 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1585 return getFalse(ITy);
1588 if (C0->getBoolValue() && isNUW) {
1590 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1591 return getFalse(ITy);
1593 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1594 return getFalse(ITy);
1600 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1601 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1603 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1606 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1608 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1611 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1614 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1616 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1622 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1623 // (icmp (add V, C0), C1) | (icmp V, C0)
1624 ICmpInst::Predicate Pred0, Pred1;
1625 const APInt *C0, *C1;
1627 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1630 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1633 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1634 if (AddInst->getOperand(1) != Op1->getOperand(1))
1637 Type *ITy = Op0->getType();
1638 bool isNSW = AddInst->hasNoSignedWrap();
1639 bool isNUW = AddInst->hasNoUnsignedWrap();
1641 const APInt Delta = *C1 - *C0;
1642 if (C0->isStrictlyPositive()) {
1644 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1645 return getTrue(ITy);
1646 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1647 return getTrue(ITy);
1650 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1651 return getTrue(ITy);
1652 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1653 return getTrue(ITy);
1656 if (C0->getBoolValue() && isNUW) {
1658 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1659 return getTrue(ITy);
1661 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1662 return getTrue(ITy);
1668 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1669 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1671 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1674 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1676 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1679 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1682 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1684 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1690 static Value *simplifyAndOrOfICmps(Value *Op0, Value *Op1, bool IsAnd) {
1691 // Look through casts of the 'and' operands to find compares.
1692 auto *Cast0 = dyn_cast<CastInst>(Op0);
1693 auto *Cast1 = dyn_cast<CastInst>(Op1);
1694 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1695 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1696 Op0 = Cast0->getOperand(0);
1697 Op1 = Cast1->getOperand(0);
1700 auto *Cmp0 = dyn_cast<ICmpInst>(Op0);
1701 auto *Cmp1 = dyn_cast<ICmpInst>(Op1);
1706 IsAnd ? simplifyAndOfICmps(Cmp0, Cmp1) : simplifyOrOfICmps(Cmp0, Cmp1);
1712 // If we looked through casts, we can only handle a constant simplification
1713 // because we are not allowed to create a cast instruction here.
1714 if (auto *C = dyn_cast<Constant>(V))
1715 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1720 /// Given operands for an And, see if we can fold the result.
1721 /// If not, this returns null.
1722 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1723 unsigned MaxRecurse) {
1724 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1728 if (match(Op1, m_Undef()))
1729 return Constant::getNullValue(Op0->getType());
1736 if (match(Op1, m_Zero()))
1740 if (match(Op1, m_AllOnes()))
1743 // A & ~A = ~A & A = 0
1744 if (match(Op0, m_Not(m_Specific(Op1))) ||
1745 match(Op1, m_Not(m_Specific(Op0))))
1746 return Constant::getNullValue(Op0->getType());
1749 Value *A = nullptr, *B = nullptr;
1750 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1751 (A == Op1 || B == Op1))
1755 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1756 (A == Op0 || B == Op0))
1759 // A mask that only clears known zeros of a shifted value is a no-op.
1763 if (match(Op1, m_APInt(Mask))) {
1764 // If all bits in the inverted and shifted mask are clear:
1765 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1766 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1767 (~(*Mask)).lshr(*ShAmt).isNullValue())
1770 // If all bits in the inverted and shifted mask are clear:
1771 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1772 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1773 (~(*Mask)).shl(*ShAmt).isNullValue())
1777 // A & (-A) = A if A is a power of two or zero.
1778 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1779 match(Op1, m_Neg(m_Specific(Op0)))) {
1780 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1783 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1788 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, true))
1791 // Try some generic simplifications for associative operations.
1792 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1796 // And distributes over Or. Try some generic simplifications based on this.
1797 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1801 // And distributes over Xor. Try some generic simplifications based on this.
1802 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1806 // If the operation is with the result of a select instruction, check whether
1807 // operating on either branch of the select always yields the same value.
1808 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1809 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1813 // If the operation is with the result of a phi instruction, check whether
1814 // operating on all incoming values of the phi always yields the same value.
1815 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1816 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1823 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1824 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1827 /// Given operands for an Or, see if we can fold the result.
1828 /// If not, this returns null.
1829 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1830 unsigned MaxRecurse) {
1831 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1835 if (match(Op1, m_Undef()))
1836 return Constant::getAllOnesValue(Op0->getType());
1843 if (match(Op1, m_Zero()))
1847 if (match(Op1, m_AllOnes()))
1850 // A | ~A = ~A | A = -1
1851 if (match(Op0, m_Not(m_Specific(Op1))) ||
1852 match(Op1, m_Not(m_Specific(Op0))))
1853 return Constant::getAllOnesValue(Op0->getType());
1856 Value *A = nullptr, *B = nullptr;
1857 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1858 (A == Op1 || B == Op1))
1862 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1863 (A == Op0 || B == Op0))
1866 // ~(A & ?) | A = -1
1867 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1868 (A == Op1 || B == Op1))
1869 return Constant::getAllOnesValue(Op1->getType());
1871 // A | ~(A & ?) = -1
1872 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1873 (A == Op0 || B == Op0))
1874 return Constant::getAllOnesValue(Op0->getType());
1876 // (A & ~B) | (A ^ B) -> (A ^ B)
1877 // (~B & A) | (A ^ B) -> (A ^ B)
1878 // (A & ~B) | (B ^ A) -> (B ^ A)
1879 // (~B & A) | (B ^ A) -> (B ^ A)
1880 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1881 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1882 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1885 // Commute the 'or' operands.
1886 // (A ^ B) | (A & ~B) -> (A ^ B)
1887 // (A ^ B) | (~B & A) -> (A ^ B)
1888 // (B ^ A) | (A & ~B) -> (B ^ A)
1889 // (B ^ A) | (~B & A) -> (B ^ A)
1890 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1891 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1892 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1895 // (A & B) | (~A ^ B) -> (~A ^ B)
1896 // (B & A) | (~A ^ B) -> (~A ^ B)
1897 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1898 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1899 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1900 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1901 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1904 // (~A ^ B) | (A & B) -> (~A ^ B)
1905 // (~A ^ B) | (B & A) -> (~A ^ B)
1906 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1907 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1908 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1909 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1910 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1913 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, false))
1916 // Try some generic simplifications for associative operations.
1917 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1921 // Or distributes over And. Try some generic simplifications based on this.
1922 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1926 // If the operation is with the result of a select instruction, check whether
1927 // operating on either branch of the select always yields the same value.
1928 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1929 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1933 // (A & C1)|(B & C2)
1934 const APInt *C1, *C2;
1935 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1936 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1938 // (A & C1)|(B & C2)
1939 // If we have: ((V + N) & C1) | (V & C2)
1940 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1941 // replace with V+N.
1943 if (C2->isMask() && // C2 == 0+1+
1944 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1945 // Add commutes, try both ways.
1946 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1949 // Or commutes, try both ways.
1951 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1952 // Add commutes, try both ways.
1953 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1959 // If the operation is with the result of a phi instruction, check whether
1960 // operating on all incoming values of the phi always yields the same value.
1961 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1962 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1968 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1969 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1972 /// Given operands for a Xor, see if we can fold the result.
1973 /// If not, this returns null.
1974 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1975 unsigned MaxRecurse) {
1976 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1979 // A ^ undef -> undef
1980 if (match(Op1, m_Undef()))
1984 if (match(Op1, m_Zero()))
1989 return Constant::getNullValue(Op0->getType());
1991 // A ^ ~A = ~A ^ A = -1
1992 if (match(Op0, m_Not(m_Specific(Op1))) ||
1993 match(Op1, m_Not(m_Specific(Op0))))
1994 return Constant::getAllOnesValue(Op0->getType());
1996 // Try some generic simplifications for associative operations.
1997 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2001 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2002 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2003 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2004 // only if B and C are equal. If B and C are equal then (since we assume
2005 // that operands have already been simplified) "select(cond, B, C)" should
2006 // have been simplified to the common value of B and C already. Analysing
2007 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2008 // for threading over phi nodes.
2013 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2014 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2018 static Type *GetCompareTy(Value *Op) {
2019 return CmpInst::makeCmpResultType(Op->getType());
2022 /// Rummage around inside V looking for something equivalent to the comparison
2023 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2024 /// Helper function for analyzing max/min idioms.
2025 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2026 Value *LHS, Value *RHS) {
2027 SelectInst *SI = dyn_cast<SelectInst>(V);
2030 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2033 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2034 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2036 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2037 LHS == CmpRHS && RHS == CmpLHS)
2042 // A significant optimization not implemented here is assuming that alloca
2043 // addresses are not equal to incoming argument values. They don't *alias*,
2044 // as we say, but that doesn't mean they aren't equal, so we take a
2045 // conservative approach.
2047 // This is inspired in part by C++11 5.10p1:
2048 // "Two pointers of the same type compare equal if and only if they are both
2049 // null, both point to the same function, or both represent the same
2052 // This is pretty permissive.
2054 // It's also partly due to C11 6.5.9p6:
2055 // "Two pointers compare equal if and only if both are null pointers, both are
2056 // pointers to the same object (including a pointer to an object and a
2057 // subobject at its beginning) or function, both are pointers to one past the
2058 // last element of the same array object, or one is a pointer to one past the
2059 // end of one array object and the other is a pointer to the start of a
2060 // different array object that happens to immediately follow the first array
2061 // object in the address space.)
2063 // C11's version is more restrictive, however there's no reason why an argument
2064 // couldn't be a one-past-the-end value for a stack object in the caller and be
2065 // equal to the beginning of a stack object in the callee.
2067 // If the C and C++ standards are ever made sufficiently restrictive in this
2068 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2069 // this optimization.
2071 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2072 const DominatorTree *DT, CmpInst::Predicate Pred,
2073 const Instruction *CxtI, Value *LHS, Value *RHS) {
2074 // First, skip past any trivial no-ops.
2075 LHS = LHS->stripPointerCasts();
2076 RHS = RHS->stripPointerCasts();
2078 // A non-null pointer is not equal to a null pointer.
2079 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) &&
2080 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2081 return ConstantInt::get(GetCompareTy(LHS),
2082 !CmpInst::isTrueWhenEqual(Pred));
2084 // We can only fold certain predicates on pointer comparisons.
2089 // Equality comaprisons are easy to fold.
2090 case CmpInst::ICMP_EQ:
2091 case CmpInst::ICMP_NE:
2094 // We can only handle unsigned relational comparisons because 'inbounds' on
2095 // a GEP only protects against unsigned wrapping.
2096 case CmpInst::ICMP_UGT:
2097 case CmpInst::ICMP_UGE:
2098 case CmpInst::ICMP_ULT:
2099 case CmpInst::ICMP_ULE:
2100 // However, we have to switch them to their signed variants to handle
2101 // negative indices from the base pointer.
2102 Pred = ICmpInst::getSignedPredicate(Pred);
2106 // Strip off any constant offsets so that we can reason about them.
2107 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2108 // here and compare base addresses like AliasAnalysis does, however there are
2109 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2110 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2111 // doesn't need to guarantee pointer inequality when it says NoAlias.
2112 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2113 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2115 // If LHS and RHS are related via constant offsets to the same base
2116 // value, we can replace it with an icmp which just compares the offsets.
2118 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2120 // Various optimizations for (in)equality comparisons.
2121 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2122 // Different non-empty allocations that exist at the same time have
2123 // different addresses (if the program can tell). Global variables always
2124 // exist, so they always exist during the lifetime of each other and all
2125 // allocas. Two different allocas usually have different addresses...
2127 // However, if there's an @llvm.stackrestore dynamically in between two
2128 // allocas, they may have the same address. It's tempting to reduce the
2129 // scope of the problem by only looking at *static* allocas here. That would
2130 // cover the majority of allocas while significantly reducing the likelihood
2131 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2132 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2133 // an entry block. Also, if we have a block that's not attached to a
2134 // function, we can't tell if it's "static" under the current definition.
2135 // Theoretically, this problem could be fixed by creating a new kind of
2136 // instruction kind specifically for static allocas. Such a new instruction
2137 // could be required to be at the top of the entry block, thus preventing it
2138 // from being subject to a @llvm.stackrestore. Instcombine could even
2139 // convert regular allocas into these special allocas. It'd be nifty.
2140 // However, until then, this problem remains open.
2142 // So, we'll assume that two non-empty allocas have different addresses
2145 // With all that, if the offsets are within the bounds of their allocations
2146 // (and not one-past-the-end! so we can't use inbounds!), and their
2147 // allocations aren't the same, the pointers are not equal.
2149 // Note that it's not necessary to check for LHS being a global variable
2150 // address, due to canonicalization and constant folding.
2151 if (isa<AllocaInst>(LHS) &&
2152 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2153 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2154 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2155 uint64_t LHSSize, RHSSize;
2156 if (LHSOffsetCI && RHSOffsetCI &&
2157 getObjectSize(LHS, LHSSize, DL, TLI) &&
2158 getObjectSize(RHS, RHSSize, DL, TLI)) {
2159 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2160 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2161 if (!LHSOffsetValue.isNegative() &&
2162 !RHSOffsetValue.isNegative() &&
2163 LHSOffsetValue.ult(LHSSize) &&
2164 RHSOffsetValue.ult(RHSSize)) {
2165 return ConstantInt::get(GetCompareTy(LHS),
2166 !CmpInst::isTrueWhenEqual(Pred));
2170 // Repeat the above check but this time without depending on DataLayout
2171 // or being able to compute a precise size.
2172 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2173 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2174 LHSOffset->isNullValue() &&
2175 RHSOffset->isNullValue())
2176 return ConstantInt::get(GetCompareTy(LHS),
2177 !CmpInst::isTrueWhenEqual(Pred));
2180 // Even if an non-inbounds GEP occurs along the path we can still optimize
2181 // equality comparisons concerning the result. We avoid walking the whole
2182 // chain again by starting where the last calls to
2183 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2184 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2185 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2187 return ConstantExpr::getICmp(Pred,
2188 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2189 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2191 // If one side of the equality comparison must come from a noalias call
2192 // (meaning a system memory allocation function), and the other side must
2193 // come from a pointer that cannot overlap with dynamically-allocated
2194 // memory within the lifetime of the current function (allocas, byval
2195 // arguments, globals), then determine the comparison result here.
2196 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2197 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2198 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2200 // Is the set of underlying objects all noalias calls?
2201 auto IsNAC = [](ArrayRef<Value *> Objects) {
2202 return all_of(Objects, isNoAliasCall);
2205 // Is the set of underlying objects all things which must be disjoint from
2206 // noalias calls. For allocas, we consider only static ones (dynamic
2207 // allocas might be transformed into calls to malloc not simultaneously
2208 // live with the compared-to allocation). For globals, we exclude symbols
2209 // that might be resolve lazily to symbols in another dynamically-loaded
2210 // library (and, thus, could be malloc'ed by the implementation).
2211 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2212 return all_of(Objects, [](Value *V) {
2213 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2214 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2215 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2216 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2217 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2218 !GV->isThreadLocal();
2219 if (const Argument *A = dyn_cast<Argument>(V))
2220 return A->hasByValAttr();
2225 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2226 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2227 return ConstantInt::get(GetCompareTy(LHS),
2228 !CmpInst::isTrueWhenEqual(Pred));
2230 // Fold comparisons for non-escaping pointer even if the allocation call
2231 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2232 // dynamic allocation call could be either of the operands.
2233 Value *MI = nullptr;
2234 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT))
2236 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT))
2238 // FIXME: We should also fold the compare when the pointer escapes, but the
2239 // compare dominates the pointer escape
2240 if (MI && !PointerMayBeCaptured(MI, true, true))
2241 return ConstantInt::get(GetCompareTy(LHS),
2242 CmpInst::isFalseWhenEqual(Pred));
2249 /// Fold an icmp when its operands have i1 scalar type.
2250 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2251 Value *RHS, const SimplifyQuery &Q) {
2252 Type *ITy = GetCompareTy(LHS); // The return type.
2253 Type *OpTy = LHS->getType(); // The operand type.
2254 if (!OpTy->getScalarType()->isIntegerTy(1))
2257 // A boolean compared to true/false can be simplified in 14 out of the 20
2258 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2259 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2260 if (match(RHS, m_Zero())) {
2262 case CmpInst::ICMP_NE: // X != 0 -> X
2263 case CmpInst::ICMP_UGT: // X >u 0 -> X
2264 case CmpInst::ICMP_SLT: // X <s 0 -> X
2267 case CmpInst::ICMP_ULT: // X <u 0 -> false
2268 case CmpInst::ICMP_SGT: // X >s 0 -> false
2269 return getFalse(ITy);
2271 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2272 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2273 return getTrue(ITy);
2277 } else if (match(RHS, m_One())) {
2279 case CmpInst::ICMP_EQ: // X == 1 -> X
2280 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2281 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2284 case CmpInst::ICMP_UGT: // X >u 1 -> false
2285 case CmpInst::ICMP_SLT: // X <s -1 -> false
2286 return getFalse(ITy);
2288 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2289 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2290 return getTrue(ITy);
2299 case ICmpInst::ICMP_UGE:
2300 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2301 return getTrue(ITy);
2303 case ICmpInst::ICMP_SGE:
2304 /// For signed comparison, the values for an i1 are 0 and -1
2305 /// respectively. This maps into a truth table of:
2306 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2307 /// 0 | 0 | 1 (0 >= 0) | 1
2308 /// 0 | 1 | 1 (0 >= -1) | 1
2309 /// 1 | 0 | 0 (-1 >= 0) | 0
2310 /// 1 | 1 | 1 (-1 >= -1) | 1
2311 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2312 return getTrue(ITy);
2314 case ICmpInst::ICMP_ULE:
2315 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2316 return getTrue(ITy);
2323 /// Try hard to fold icmp with zero RHS because this is a common case.
2324 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2325 Value *RHS, const SimplifyQuery &Q) {
2326 if (!match(RHS, m_Zero()))
2329 Type *ITy = GetCompareTy(LHS); // The return type.
2332 llvm_unreachable("Unknown ICmp predicate!");
2333 case ICmpInst::ICMP_ULT:
2334 return getFalse(ITy);
2335 case ICmpInst::ICMP_UGE:
2336 return getTrue(ITy);
2337 case ICmpInst::ICMP_EQ:
2338 case ICmpInst::ICMP_ULE:
2339 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2340 return getFalse(ITy);
2342 case ICmpInst::ICMP_NE:
2343 case ICmpInst::ICMP_UGT:
2344 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2345 return getTrue(ITy);
2347 case ICmpInst::ICMP_SLT: {
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 return getFalse(ITy);
2355 case ICmpInst::ICMP_SLE: {
2356 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2357 if (LHSKnown.isNegative())
2358 return getTrue(ITy);
2359 if (LHSKnown.isNonNegative() &&
2360 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2361 return getFalse(ITy);
2364 case ICmpInst::ICMP_SGE: {
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 return getTrue(ITy);
2372 case ICmpInst::ICMP_SGT: {
2373 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2374 if (LHSKnown.isNegative())
2375 return getFalse(ITy);
2376 if (LHSKnown.isNonNegative() &&
2377 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2378 return getTrue(ITy);
2386 /// Many binary operators with a constant operand have an easy-to-compute
2387 /// range of outputs. This can be used to fold a comparison to always true or
2389 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2390 unsigned Width = Lower.getBitWidth();
2392 switch (BO.getOpcode()) {
2393 case Instruction::Add:
2394 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2395 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2396 if (BO.hasNoUnsignedWrap()) {
2397 // 'add nuw x, C' produces [C, UINT_MAX].
2399 } else if (BO.hasNoSignedWrap()) {
2400 if (C->isNegative()) {
2401 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2402 Lower = APInt::getSignedMinValue(Width);
2403 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2405 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2406 Lower = APInt::getSignedMinValue(Width) + *C;
2407 Upper = APInt::getSignedMaxValue(Width) + 1;
2413 case Instruction::And:
2414 if (match(BO.getOperand(1), m_APInt(C)))
2415 // 'and x, C' produces [0, C].
2419 case Instruction::Or:
2420 if (match(BO.getOperand(1), m_APInt(C)))
2421 // 'or x, C' produces [C, UINT_MAX].
2425 case Instruction::AShr:
2426 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2427 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2428 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2429 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2430 } else if (match(BO.getOperand(0), m_APInt(C))) {
2431 unsigned ShiftAmount = Width - 1;
2432 if (!C->isNullValue() && BO.isExact())
2433 ShiftAmount = C->countTrailingZeros();
2434 if (C->isNegative()) {
2435 // 'ashr C, x' produces [C, C >> (Width-1)]
2437 Upper = C->ashr(ShiftAmount) + 1;
2439 // 'ashr C, x' produces [C >> (Width-1), C]
2440 Lower = C->ashr(ShiftAmount);
2446 case Instruction::LShr:
2447 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2448 // 'lshr x, C' produces [0, UINT_MAX >> C].
2449 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2450 } else if (match(BO.getOperand(0), m_APInt(C))) {
2451 // 'lshr C, x' produces [C >> (Width-1), C].
2452 unsigned ShiftAmount = Width - 1;
2453 if (!C->isNullValue() && BO.isExact())
2454 ShiftAmount = C->countTrailingZeros();
2455 Lower = C->lshr(ShiftAmount);
2460 case Instruction::Shl:
2461 if (match(BO.getOperand(0), m_APInt(C))) {
2462 if (BO.hasNoUnsignedWrap()) {
2463 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2465 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2466 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2467 if (C->isNegative()) {
2468 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2469 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2470 Lower = C->shl(ShiftAmount);
2473 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2474 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2476 Upper = C->shl(ShiftAmount) + 1;
2482 case Instruction::SDiv:
2483 if (match(BO.getOperand(1), m_APInt(C))) {
2484 APInt IntMin = APInt::getSignedMinValue(Width);
2485 APInt IntMax = APInt::getSignedMaxValue(Width);
2486 if (C->isAllOnesValue()) {
2487 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2488 // where C != -1 and C != 0 and C != 1
2491 } else if (C->countLeadingZeros() < Width - 1) {
2492 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2493 // where C != -1 and C != 0 and C != 1
2494 Lower = IntMin.sdiv(*C);
2495 Upper = IntMax.sdiv(*C);
2496 if (Lower.sgt(Upper))
2497 std::swap(Lower, Upper);
2499 assert(Upper != Lower && "Upper part of range has wrapped!");
2501 } else if (match(BO.getOperand(0), m_APInt(C))) {
2502 if (C->isMinSignedValue()) {
2503 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2505 Upper = Lower.lshr(1) + 1;
2507 // 'sdiv C, x' produces [-|C|, |C|].
2508 Upper = C->abs() + 1;
2509 Lower = (-Upper) + 1;
2514 case Instruction::UDiv:
2515 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2516 // 'udiv x, C' produces [0, UINT_MAX / C].
2517 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2518 } else if (match(BO.getOperand(0), m_APInt(C))) {
2519 // 'udiv C, x' produces [0, C].
2524 case Instruction::SRem:
2525 if (match(BO.getOperand(1), m_APInt(C))) {
2526 // 'srem x, C' produces (-|C|, |C|).
2528 Lower = (-Upper) + 1;
2532 case Instruction::URem:
2533 if (match(BO.getOperand(1), m_APInt(C)))
2534 // 'urem x, C' produces [0, C).
2543 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2546 if (!match(RHS, m_APInt(C)))
2549 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2550 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2551 if (RHS_CR.isEmptySet())
2552 return ConstantInt::getFalse(GetCompareTy(RHS));
2553 if (RHS_CR.isFullSet())
2554 return ConstantInt::getTrue(GetCompareTy(RHS));
2556 // Find the range of possible values for binary operators.
2557 unsigned Width = C->getBitWidth();
2558 APInt Lower = APInt(Width, 0);
2559 APInt Upper = APInt(Width, 0);
2560 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2561 setLimitsForBinOp(*BO, Lower, Upper);
2563 ConstantRange LHS_CR =
2564 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2566 if (auto *I = dyn_cast<Instruction>(LHS))
2567 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2568 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2570 if (!LHS_CR.isFullSet()) {
2571 if (RHS_CR.contains(LHS_CR))
2572 return ConstantInt::getTrue(GetCompareTy(RHS));
2573 if (RHS_CR.inverse().contains(LHS_CR))
2574 return ConstantInt::getFalse(GetCompareTy(RHS));
2580 /// TODO: A large part of this logic is duplicated in InstCombine's
2581 /// foldICmpBinOp(). We should be able to share that and avoid the code
2583 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2584 Value *RHS, const SimplifyQuery &Q,
2585 unsigned MaxRecurse) {
2586 Type *ITy = GetCompareTy(LHS); // The return type.
2588 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2589 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2590 if (MaxRecurse && (LBO || RBO)) {
2591 // Analyze the case when either LHS or RHS is an add instruction.
2592 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2593 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2594 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2595 if (LBO && LBO->getOpcode() == Instruction::Add) {
2596 A = LBO->getOperand(0);
2597 B = LBO->getOperand(1);
2599 ICmpInst::isEquality(Pred) ||
2600 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2601 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2603 if (RBO && RBO->getOpcode() == Instruction::Add) {
2604 C = RBO->getOperand(0);
2605 D = RBO->getOperand(1);
2607 ICmpInst::isEquality(Pred) ||
2608 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2609 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2612 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2613 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2614 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2615 Constant::getNullValue(RHS->getType()), Q,
2619 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2620 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2622 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2623 C == LHS ? D : C, Q, MaxRecurse - 1))
2626 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2627 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2629 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2632 // C + B == C + D -> B == D
2635 } else if (A == D) {
2636 // D + B == C + D -> B == C
2639 } else if (B == C) {
2640 // A + C == C + D -> A == D
2645 // A + D == C + D -> A == C
2649 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2656 // icmp pred (or X, Y), X
2657 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2658 if (Pred == ICmpInst::ICMP_ULT)
2659 return getFalse(ITy);
2660 if (Pred == ICmpInst::ICMP_UGE)
2661 return getTrue(ITy);
2663 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2664 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2665 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2666 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2667 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2668 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2669 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2672 // icmp pred X, (or X, Y)
2673 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2674 if (Pred == ICmpInst::ICMP_ULE)
2675 return getTrue(ITy);
2676 if (Pred == ICmpInst::ICMP_UGT)
2677 return getFalse(ITy);
2679 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2680 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2681 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2682 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2683 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2684 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2685 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2690 // icmp pred (and X, Y), X
2691 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)),
2692 m_And(m_Specific(RHS), m_Value())))) {
2693 if (Pred == ICmpInst::ICMP_UGT)
2694 return getFalse(ITy);
2695 if (Pred == ICmpInst::ICMP_ULE)
2696 return getTrue(ITy);
2698 // icmp pred X, (and X, Y)
2699 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)),
2700 m_And(m_Specific(LHS), m_Value())))) {
2701 if (Pred == ICmpInst::ICMP_UGE)
2702 return getTrue(ITy);
2703 if (Pred == ICmpInst::ICMP_ULT)
2704 return getFalse(ITy);
2707 // 0 - (zext X) pred C
2708 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2709 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2710 if (RHSC->getValue().isStrictlyPositive()) {
2711 if (Pred == ICmpInst::ICMP_SLT)
2712 return ConstantInt::getTrue(RHSC->getContext());
2713 if (Pred == ICmpInst::ICMP_SGE)
2714 return ConstantInt::getFalse(RHSC->getContext());
2715 if (Pred == ICmpInst::ICMP_EQ)
2716 return ConstantInt::getFalse(RHSC->getContext());
2717 if (Pred == ICmpInst::ICMP_NE)
2718 return ConstantInt::getTrue(RHSC->getContext());
2720 if (RHSC->getValue().isNonNegative()) {
2721 if (Pred == ICmpInst::ICMP_SLE)
2722 return ConstantInt::getTrue(RHSC->getContext());
2723 if (Pred == ICmpInst::ICMP_SGT)
2724 return ConstantInt::getFalse(RHSC->getContext());
2729 // icmp pred (urem X, Y), Y
2730 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2734 case ICmpInst::ICMP_SGT:
2735 case ICmpInst::ICMP_SGE: {
2736 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2737 if (!Known.isNonNegative())
2741 case ICmpInst::ICMP_EQ:
2742 case ICmpInst::ICMP_UGT:
2743 case ICmpInst::ICMP_UGE:
2744 return getFalse(ITy);
2745 case ICmpInst::ICMP_SLT:
2746 case ICmpInst::ICMP_SLE: {
2747 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2748 if (!Known.isNonNegative())
2752 case ICmpInst::ICMP_NE:
2753 case ICmpInst::ICMP_ULT:
2754 case ICmpInst::ICMP_ULE:
2755 return getTrue(ITy);
2759 // icmp pred X, (urem Y, X)
2760 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2764 case ICmpInst::ICMP_SGT:
2765 case ICmpInst::ICMP_SGE: {
2766 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2767 if (!Known.isNonNegative())
2771 case ICmpInst::ICMP_NE:
2772 case ICmpInst::ICMP_UGT:
2773 case ICmpInst::ICMP_UGE:
2774 return getTrue(ITy);
2775 case ICmpInst::ICMP_SLT:
2776 case ICmpInst::ICMP_SLE: {
2777 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2778 if (!Known.isNonNegative())
2782 case ICmpInst::ICMP_EQ:
2783 case ICmpInst::ICMP_ULT:
2784 case ICmpInst::ICMP_ULE:
2785 return getFalse(ITy);
2791 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2792 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2793 // icmp pred (X op Y), X
2794 if (Pred == ICmpInst::ICMP_UGT)
2795 return getFalse(ITy);
2796 if (Pred == ICmpInst::ICMP_ULE)
2797 return getTrue(ITy);
2802 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2803 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2804 // icmp pred X, (X op Y)
2805 if (Pred == ICmpInst::ICMP_ULT)
2806 return getFalse(ITy);
2807 if (Pred == ICmpInst::ICMP_UGE)
2808 return getTrue(ITy);
2815 // where CI2 is a power of 2 and CI isn't
2816 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2817 const APInt *CI2Val, *CIVal = &CI->getValue();
2818 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2819 CI2Val->isPowerOf2()) {
2820 if (!CIVal->isPowerOf2()) {
2821 // CI2 << X can equal zero in some circumstances,
2822 // this simplification is unsafe if CI is zero.
2824 // We know it is safe if:
2825 // - The shift is nsw, we can't shift out the one bit.
2826 // - The shift is nuw, we can't shift out the one bit.
2829 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2830 CI2Val->isOneValue() || !CI->isZero()) {
2831 if (Pred == ICmpInst::ICMP_EQ)
2832 return ConstantInt::getFalse(RHS->getContext());
2833 if (Pred == ICmpInst::ICMP_NE)
2834 return ConstantInt::getTrue(RHS->getContext());
2837 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2838 if (Pred == ICmpInst::ICMP_UGT)
2839 return ConstantInt::getFalse(RHS->getContext());
2840 if (Pred == ICmpInst::ICMP_ULE)
2841 return ConstantInt::getTrue(RHS->getContext());
2846 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2847 LBO->getOperand(1) == RBO->getOperand(1)) {
2848 switch (LBO->getOpcode()) {
2851 case Instruction::UDiv:
2852 case Instruction::LShr:
2853 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2855 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2856 RBO->getOperand(0), Q, MaxRecurse - 1))
2859 case Instruction::SDiv:
2860 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2862 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2863 RBO->getOperand(0), Q, MaxRecurse - 1))
2866 case Instruction::AShr:
2867 if (!LBO->isExact() || !RBO->isExact())
2869 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2870 RBO->getOperand(0), Q, MaxRecurse - 1))
2873 case Instruction::Shl: {
2874 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2875 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2878 if (!NSW && ICmpInst::isSigned(Pred))
2880 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2881 RBO->getOperand(0), Q, MaxRecurse - 1))
2890 /// Simplify integer comparisons where at least one operand of the compare
2891 /// matches an integer min/max idiom.
2892 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2893 Value *RHS, const SimplifyQuery &Q,
2894 unsigned MaxRecurse) {
2895 Type *ITy = GetCompareTy(LHS); // The return type.
2897 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2898 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2900 // Signed variants on "max(a,b)>=a -> true".
2901 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2903 std::swap(A, B); // smax(A, B) pred A.
2904 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2905 // We analyze this as smax(A, B) pred A.
2907 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2908 (A == LHS || B == LHS)) {
2910 std::swap(A, B); // A pred smax(A, B).
2911 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2912 // We analyze this as smax(A, B) swapped-pred A.
2913 P = CmpInst::getSwappedPredicate(Pred);
2914 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2915 (A == RHS || B == RHS)) {
2917 std::swap(A, B); // smin(A, B) pred A.
2918 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2919 // We analyze this as smax(-A, -B) swapped-pred -A.
2920 // Note that we do not need to actually form -A or -B thanks to EqP.
2921 P = CmpInst::getSwappedPredicate(Pred);
2922 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2923 (A == LHS || B == LHS)) {
2925 std::swap(A, B); // A pred smin(A, B).
2926 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2927 // We analyze this as smax(-A, -B) pred -A.
2928 // Note that we do not need to actually form -A or -B thanks to EqP.
2931 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2932 // Cases correspond to "max(A, B) p A".
2936 case CmpInst::ICMP_EQ:
2937 case CmpInst::ICMP_SLE:
2938 // Equivalent to "A EqP B". This may be the same as the condition tested
2939 // in the max/min; if so, we can just return that.
2940 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2942 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2944 // Otherwise, see if "A EqP B" simplifies.
2946 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2949 case CmpInst::ICMP_NE:
2950 case CmpInst::ICMP_SGT: {
2951 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2952 // Equivalent to "A InvEqP B". This may be the same as the condition
2953 // tested in the max/min; if so, we can just return that.
2954 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2956 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2958 // Otherwise, see if "A InvEqP B" simplifies.
2960 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2964 case CmpInst::ICMP_SGE:
2966 return getTrue(ITy);
2967 case CmpInst::ICMP_SLT:
2969 return getFalse(ITy);
2973 // Unsigned variants on "max(a,b)>=a -> true".
2974 P = CmpInst::BAD_ICMP_PREDICATE;
2975 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2977 std::swap(A, B); // umax(A, B) pred A.
2978 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2979 // We analyze this as umax(A, B) pred A.
2981 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2982 (A == LHS || B == LHS)) {
2984 std::swap(A, B); // A pred umax(A, B).
2985 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2986 // We analyze this as umax(A, B) swapped-pred A.
2987 P = CmpInst::getSwappedPredicate(Pred);
2988 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2989 (A == RHS || B == RHS)) {
2991 std::swap(A, B); // umin(A, B) pred A.
2992 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2993 // We analyze this as umax(-A, -B) swapped-pred -A.
2994 // Note that we do not need to actually form -A or -B thanks to EqP.
2995 P = CmpInst::getSwappedPredicate(Pred);
2996 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2997 (A == LHS || B == LHS)) {
2999 std::swap(A, B); // A pred umin(A, B).
3000 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3001 // We analyze this as umax(-A, -B) pred -A.
3002 // Note that we do not need to actually form -A or -B thanks to EqP.
3005 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3006 // Cases correspond to "max(A, B) p A".
3010 case CmpInst::ICMP_EQ:
3011 case CmpInst::ICMP_ULE:
3012 // Equivalent to "A EqP B". This may be the same as the condition tested
3013 // in the max/min; if so, we can just return that.
3014 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3016 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3018 // Otherwise, see if "A EqP B" simplifies.
3020 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3023 case CmpInst::ICMP_NE:
3024 case CmpInst::ICMP_UGT: {
3025 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3026 // Equivalent to "A InvEqP B". This may be the same as the condition
3027 // tested in the max/min; if so, we can just return that.
3028 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3030 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3032 // Otherwise, see if "A InvEqP B" simplifies.
3034 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3038 case CmpInst::ICMP_UGE:
3040 return getTrue(ITy);
3041 case CmpInst::ICMP_ULT:
3043 return getFalse(ITy);
3047 // Variants on "max(x,y) >= min(x,z)".
3049 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3050 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3051 (A == C || A == D || B == C || B == D)) {
3052 // max(x, ?) pred min(x, ?).
3053 if (Pred == CmpInst::ICMP_SGE)
3055 return getTrue(ITy);
3056 if (Pred == CmpInst::ICMP_SLT)
3058 return getFalse(ITy);
3059 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3060 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3061 (A == C || A == D || B == C || B == D)) {
3062 // min(x, ?) pred max(x, ?).
3063 if (Pred == CmpInst::ICMP_SLE)
3065 return getTrue(ITy);
3066 if (Pred == CmpInst::ICMP_SGT)
3068 return getFalse(ITy);
3069 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3070 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3071 (A == C || A == D || B == C || B == D)) {
3072 // max(x, ?) pred min(x, ?).
3073 if (Pred == CmpInst::ICMP_UGE)
3075 return getTrue(ITy);
3076 if (Pred == CmpInst::ICMP_ULT)
3078 return getFalse(ITy);
3079 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3080 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3081 (A == C || A == D || B == C || B == D)) {
3082 // min(x, ?) pred max(x, ?).
3083 if (Pred == CmpInst::ICMP_ULE)
3085 return getTrue(ITy);
3086 if (Pred == CmpInst::ICMP_UGT)
3088 return getFalse(ITy);
3094 /// Given operands for an ICmpInst, see if we can fold the result.
3095 /// If not, this returns null.
3096 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3097 const SimplifyQuery &Q, unsigned MaxRecurse) {
3098 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3099 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3101 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3102 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3103 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3105 // If we have a constant, make sure it is on the RHS.
3106 std::swap(LHS, RHS);
3107 Pred = CmpInst::getSwappedPredicate(Pred);
3110 Type *ITy = GetCompareTy(LHS); // The return type.
3112 // icmp X, X -> true/false
3113 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3114 // because X could be 0.
3115 if (LHS == RHS || isa<UndefValue>(RHS))
3116 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3118 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3121 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3124 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3127 // If both operands have range metadata, use the metadata
3128 // to simplify the comparison.
3129 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3130 auto RHS_Instr = cast<Instruction>(RHS);
3131 auto LHS_Instr = cast<Instruction>(LHS);
3133 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3134 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3135 auto RHS_CR = getConstantRangeFromMetadata(
3136 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3137 auto LHS_CR = getConstantRangeFromMetadata(
3138 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3140 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3141 if (Satisfied_CR.contains(LHS_CR))
3142 return ConstantInt::getTrue(RHS->getContext());
3144 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3145 CmpInst::getInversePredicate(Pred), RHS_CR);
3146 if (InversedSatisfied_CR.contains(LHS_CR))
3147 return ConstantInt::getFalse(RHS->getContext());
3151 // Compare of cast, for example (zext X) != 0 -> X != 0
3152 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3153 Instruction *LI = cast<CastInst>(LHS);
3154 Value *SrcOp = LI->getOperand(0);
3155 Type *SrcTy = SrcOp->getType();
3156 Type *DstTy = LI->getType();
3158 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3159 // if the integer type is the same size as the pointer type.
3160 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3161 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3162 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3163 // Transfer the cast to the constant.
3164 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3165 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3168 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3169 if (RI->getOperand(0)->getType() == SrcTy)
3170 // Compare without the cast.
3171 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3177 if (isa<ZExtInst>(LHS)) {
3178 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3180 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3181 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3182 // Compare X and Y. Note that signed predicates become unsigned.
3183 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3184 SrcOp, RI->getOperand(0), Q,
3188 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3189 // too. If not, then try to deduce the result of the comparison.
3190 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3191 // Compute the constant that would happen if we truncated to SrcTy then
3192 // reextended to DstTy.
3193 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3194 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3196 // If the re-extended constant didn't change then this is effectively
3197 // also a case of comparing two zero-extended values.
3198 if (RExt == CI && MaxRecurse)
3199 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3200 SrcOp, Trunc, Q, MaxRecurse-1))
3203 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3204 // there. Use this to work out the result of the comparison.
3207 default: llvm_unreachable("Unknown ICmp predicate!");
3209 case ICmpInst::ICMP_EQ:
3210 case ICmpInst::ICMP_UGT:
3211 case ICmpInst::ICMP_UGE:
3212 return ConstantInt::getFalse(CI->getContext());
3214 case ICmpInst::ICMP_NE:
3215 case ICmpInst::ICMP_ULT:
3216 case ICmpInst::ICMP_ULE:
3217 return ConstantInt::getTrue(CI->getContext());
3219 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3220 // is non-negative then LHS <s RHS.
3221 case ICmpInst::ICMP_SGT:
3222 case ICmpInst::ICMP_SGE:
3223 return CI->getValue().isNegative() ?
3224 ConstantInt::getTrue(CI->getContext()) :
3225 ConstantInt::getFalse(CI->getContext());
3227 case ICmpInst::ICMP_SLT:
3228 case ICmpInst::ICMP_SLE:
3229 return CI->getValue().isNegative() ?
3230 ConstantInt::getFalse(CI->getContext()) :
3231 ConstantInt::getTrue(CI->getContext());
3237 if (isa<SExtInst>(LHS)) {
3238 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3240 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3241 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3242 // Compare X and Y. Note that the predicate does not change.
3243 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3247 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3248 // too. If not, then try to deduce the result of the comparison.
3249 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3250 // Compute the constant that would happen if we truncated to SrcTy then
3251 // reextended to DstTy.
3252 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3253 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3255 // If the re-extended constant didn't change then this is effectively
3256 // also a case of comparing two sign-extended values.
3257 if (RExt == CI && MaxRecurse)
3258 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3261 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3262 // bits there. Use this to work out the result of the comparison.
3265 default: llvm_unreachable("Unknown ICmp predicate!");
3266 case ICmpInst::ICMP_EQ:
3267 return ConstantInt::getFalse(CI->getContext());
3268 case ICmpInst::ICMP_NE:
3269 return ConstantInt::getTrue(CI->getContext());
3271 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3273 case ICmpInst::ICMP_SGT:
3274 case ICmpInst::ICMP_SGE:
3275 return CI->getValue().isNegative() ?
3276 ConstantInt::getTrue(CI->getContext()) :
3277 ConstantInt::getFalse(CI->getContext());
3278 case ICmpInst::ICMP_SLT:
3279 case ICmpInst::ICMP_SLE:
3280 return CI->getValue().isNegative() ?
3281 ConstantInt::getFalse(CI->getContext()) :
3282 ConstantInt::getTrue(CI->getContext());
3284 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3286 case ICmpInst::ICMP_UGT:
3287 case ICmpInst::ICMP_UGE:
3288 // Comparison is true iff the LHS <s 0.
3290 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3291 Constant::getNullValue(SrcTy),
3295 case ICmpInst::ICMP_ULT:
3296 case ICmpInst::ICMP_ULE:
3297 // Comparison is true iff the LHS >=s 0.
3299 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3300 Constant::getNullValue(SrcTy),
3310 // icmp eq|ne X, Y -> false|true if X != Y
3311 if (ICmpInst::isEquality(Pred) &&
3312 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3313 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3316 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3319 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3322 // Simplify comparisons of related pointers using a powerful, recursive
3323 // GEP-walk when we have target data available..
3324 if (LHS->getType()->isPointerTy())
3325 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS))
3327 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3328 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3329 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3330 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3331 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3332 Q.DL.getTypeSizeInBits(CRHS->getType()))
3333 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI,
3334 CLHS->getPointerOperand(),
3335 CRHS->getPointerOperand()))
3338 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3339 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3340 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3341 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3342 (ICmpInst::isEquality(Pred) ||
3343 (GLHS->isInBounds() && GRHS->isInBounds() &&
3344 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3345 // The bases are equal and the indices are constant. Build a constant
3346 // expression GEP with the same indices and a null base pointer to see
3347 // what constant folding can make out of it.
3348 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3349 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3350 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3351 GLHS->getSourceElementType(), Null, IndicesLHS);
3353 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3354 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3355 GLHS->getSourceElementType(), Null, IndicesRHS);
3356 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3361 // If the comparison is with the result of a select instruction, check whether
3362 // comparing with either branch of the select always yields the same value.
3363 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3364 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3367 // If the comparison is with the result of a phi instruction, check whether
3368 // doing the compare with each incoming phi value yields a common result.
3369 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3370 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3376 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3377 const SimplifyQuery &Q) {
3378 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3381 /// Given operands for an FCmpInst, see if we can fold the result.
3382 /// If not, this returns null.
3383 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3384 FastMathFlags FMF, const SimplifyQuery &Q,
3385 unsigned MaxRecurse) {
3386 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3387 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3389 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3390 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3391 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3393 // If we have a constant, make sure it is on the RHS.
3394 std::swap(LHS, RHS);
3395 Pred = CmpInst::getSwappedPredicate(Pred);
3398 // Fold trivial predicates.
3399 Type *RetTy = GetCompareTy(LHS);
3400 if (Pred == FCmpInst::FCMP_FALSE)
3401 return getFalse(RetTy);
3402 if (Pred == FCmpInst::FCMP_TRUE)
3403 return getTrue(RetTy);
3405 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3407 if (Pred == FCmpInst::FCMP_UNO)
3408 return getFalse(RetTy);
3409 if (Pred == FCmpInst::FCMP_ORD)
3410 return getTrue(RetTy);
3413 // fcmp pred x, undef and fcmp pred undef, x
3414 // fold to true if unordered, false if ordered
3415 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3416 // Choosing NaN for the undef will always make unordered comparison succeed
3417 // and ordered comparison fail.
3418 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3421 // fcmp x,x -> true/false. Not all compares are foldable.
3423 if (CmpInst::isTrueWhenEqual(Pred))
3424 return getTrue(RetTy);
3425 if (CmpInst::isFalseWhenEqual(Pred))
3426 return getFalse(RetTy);
3429 // Handle fcmp with constant RHS
3430 const ConstantFP *CFP = nullptr;
3431 if (const auto *RHSC = dyn_cast<Constant>(RHS)) {
3432 if (RHS->getType()->isVectorTy())
3433 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue());
3435 CFP = dyn_cast<ConstantFP>(RHSC);
3438 // If the constant is a nan, see if we can fold the comparison based on it.
3439 if (CFP->getValueAPF().isNaN()) {
3440 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3441 return getFalse(RetTy);
3442 assert(FCmpInst::isUnordered(Pred) &&
3443 "Comparison must be either ordered or unordered!");
3444 // True if unordered.
3445 return getTrue(RetTy);
3447 // Check whether the constant is an infinity.
3448 if (CFP->getValueAPF().isInfinity()) {
3449 if (CFP->getValueAPF().isNegative()) {
3451 case FCmpInst::FCMP_OLT:
3452 // No value is ordered and less than negative infinity.
3453 return getFalse(RetTy);
3454 case FCmpInst::FCMP_UGE:
3455 // All values are unordered with or at least negative infinity.
3456 return getTrue(RetTy);
3462 case FCmpInst::FCMP_OGT:
3463 // No value is ordered and greater than infinity.
3464 return getFalse(RetTy);
3465 case FCmpInst::FCMP_ULE:
3466 // All values are unordered with and at most infinity.
3467 return getTrue(RetTy);
3473 if (CFP->getValueAPF().isZero()) {
3475 case FCmpInst::FCMP_UGE:
3476 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3477 return getTrue(RetTy);
3479 case FCmpInst::FCMP_OLT:
3481 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3482 return getFalse(RetTy);
3490 // If the comparison is with the result of a select instruction, check whether
3491 // comparing with either branch of the select always yields the same value.
3492 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3493 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3496 // If the comparison is with the result of a phi instruction, check whether
3497 // doing the compare with each incoming phi value yields a common result.
3498 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3499 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3505 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3506 FastMathFlags FMF, const SimplifyQuery &Q) {
3507 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3510 /// See if V simplifies when its operand Op is replaced with RepOp.
3511 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3512 const SimplifyQuery &Q,
3513 unsigned MaxRecurse) {
3514 // Trivial replacement.
3518 // We cannot replace a constant, and shouldn't even try.
3519 if (isa<Constant>(Op))
3522 auto *I = dyn_cast<Instruction>(V);
3526 // If this is a binary operator, try to simplify it with the replaced op.
3527 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3529 // %cmp = icmp eq i32 %x, 2147483647
3530 // %add = add nsw i32 %x, 1
3531 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3533 // We can't replace %sel with %add unless we strip away the flags.
3534 if (isa<OverflowingBinaryOperator>(B))
3535 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3537 if (isa<PossiblyExactOperator>(B))
3542 if (B->getOperand(0) == Op)
3543 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3545 if (B->getOperand(1) == Op)
3546 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3551 // Same for CmpInsts.
3552 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3554 if (C->getOperand(0) == Op)
3555 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3557 if (C->getOperand(1) == Op)
3558 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3563 // TODO: We could hand off more cases to instsimplify here.
3565 // If all operands are constant after substituting Op for RepOp then we can
3566 // constant fold the instruction.
3567 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3568 // Build a list of all constant operands.
3569 SmallVector<Constant *, 8> ConstOps;
3570 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3571 if (I->getOperand(i) == Op)
3572 ConstOps.push_back(CRepOp);
3573 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3574 ConstOps.push_back(COp);
3579 // All operands were constants, fold it.
3580 if (ConstOps.size() == I->getNumOperands()) {
3581 if (CmpInst *C = dyn_cast<CmpInst>(I))
3582 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3583 ConstOps[1], Q.DL, Q.TLI);
3585 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3586 if (!LI->isVolatile())
3587 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3589 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3596 /// Try to simplify a select instruction when its condition operand is an
3597 /// integer comparison where one operand of the compare is a constant.
3598 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3599 const APInt *Y, bool TrueWhenUnset) {
3602 // (X & Y) == 0 ? X & ~Y : X --> X
3603 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3604 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3606 return TrueWhenUnset ? FalseVal : TrueVal;
3608 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3609 // (X & Y) != 0 ? X : X & ~Y --> X
3610 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3612 return TrueWhenUnset ? FalseVal : TrueVal;
3614 if (Y->isPowerOf2()) {
3615 // (X & Y) == 0 ? X | Y : X --> X | Y
3616 // (X & Y) != 0 ? X | Y : X --> X
3617 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3619 return TrueWhenUnset ? TrueVal : FalseVal;
3621 // (X & Y) == 0 ? X : X | Y --> X
3622 // (X & Y) != 0 ? X : X | Y --> X | Y
3623 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3625 return TrueWhenUnset ? TrueVal : FalseVal;
3631 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3633 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal,
3635 bool TrueWhenUnset) {
3636 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
3640 APInt MinSignedValue;
3642 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) {
3643 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0
3644 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0
3645 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits();
3646 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth);
3648 // icmp slt X, 0 <--> icmp ne (and X, C), 0
3649 // icmp sgt X, -1 <--> icmp eq (and X, C), 0
3651 MinSignedValue = APInt::getSignedMinValue(BitWidth);
3654 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue,
3661 /// Try to simplify a select instruction when its condition operand is an
3662 /// integer comparison.
3663 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3664 Value *FalseVal, const SimplifyQuery &Q,
3665 unsigned MaxRecurse) {
3666 ICmpInst::Predicate Pred;
3667 Value *CmpLHS, *CmpRHS;
3668 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3671 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring
3672 // decomposeBitTestICmp() might help.
3673 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3676 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3677 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3678 Pred == ICmpInst::ICMP_EQ))
3680 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3681 // Comparing signed-less-than 0 checks if the sign bit is set.
3682 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3685 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3686 // Comparing signed-greater-than -1 checks if the sign bit is not set.
3687 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3692 if (CondVal->hasOneUse()) {
3694 if (match(CmpRHS, m_APInt(C))) {
3695 // X < MIN ? T : F --> F
3696 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3698 // X < MIN ? T : F --> F
3699 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3701 // X > MAX ? T : F --> F
3702 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3704 // X > MAX ? T : F --> F
3705 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3710 // If we have an equality comparison, then we know the value in one of the
3711 // arms of the select. See if substituting this value into the arm and
3712 // simplifying the result yields the same value as the other arm.
3713 if (Pred == ICmpInst::ICMP_EQ) {
3714 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3716 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3719 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3721 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3724 } else if (Pred == ICmpInst::ICMP_NE) {
3725 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3727 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3730 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3732 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3740 /// Given operands for a SelectInst, see if we can fold the result.
3741 /// If not, this returns null.
3742 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3743 Value *FalseVal, const SimplifyQuery &Q,
3744 unsigned MaxRecurse) {
3745 // select true, X, Y -> X
3746 // select false, X, Y -> Y
3747 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3748 if (CB->isAllOnesValue())
3750 if (CB->isNullValue())
3754 // select C, X, X -> X
3755 if (TrueVal == FalseVal)
3758 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3759 if (isa<Constant>(FalseVal))
3763 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3765 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3769 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3775 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3776 const SimplifyQuery &Q) {
3777 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3780 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3781 /// If not, this returns null.
3782 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3783 const SimplifyQuery &Q, unsigned) {
3784 // The type of the GEP pointer operand.
3786 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3788 // getelementptr P -> P.
3789 if (Ops.size() == 1)
3792 // Compute the (pointer) type returned by the GEP instruction.
3793 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3794 Type *GEPTy = PointerType::get(LastType, AS);
3795 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3796 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3797 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3798 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3800 if (isa<UndefValue>(Ops[0]))
3801 return UndefValue::get(GEPTy);
3803 if (Ops.size() == 2) {
3804 // getelementptr P, 0 -> P.
3805 if (match(Ops[1], m_Zero()))
3809 if (Ty->isSized()) {
3812 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3813 // getelementptr P, N -> P if P points to a type of zero size.
3814 if (TyAllocSize == 0)
3817 // The following transforms are only safe if the ptrtoint cast
3818 // doesn't truncate the pointers.
3819 if (Ops[1]->getType()->getScalarSizeInBits() ==
3820 Q.DL.getPointerSizeInBits(AS)) {
3821 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3822 if (match(P, m_Zero()))
3823 return Constant::getNullValue(GEPTy);
3825 if (match(P, m_PtrToInt(m_Value(Temp))))
3826 if (Temp->getType() == GEPTy)
3831 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3832 if (TyAllocSize == 1 &&
3833 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3834 if (Value *R = PtrToIntOrZero(P))
3837 // getelementptr V, (ashr (sub P, V), C) -> Q
3838 // if P points to a type of size 1 << C.
3840 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3841 m_ConstantInt(C))) &&
3842 TyAllocSize == 1ULL << C)
3843 if (Value *R = PtrToIntOrZero(P))
3846 // getelementptr V, (sdiv (sub P, V), C) -> Q
3847 // if P points to a type of size C.
3849 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3850 m_SpecificInt(TyAllocSize))))
3851 if (Value *R = PtrToIntOrZero(P))
3857 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3858 all_of(Ops.slice(1).drop_back(1),
3859 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3861 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3862 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3863 APInt BasePtrOffset(PtrWidth, 0);
3864 Value *StrippedBasePtr =
3865 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3868 // gep (gep V, C), (sub 0, V) -> C
3869 if (match(Ops.back(),
3870 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3871 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3872 return ConstantExpr::getIntToPtr(CI, GEPTy);
3874 // gep (gep V, C), (xor V, -1) -> C-1
3875 if (match(Ops.back(),
3876 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3877 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3878 return ConstantExpr::getIntToPtr(CI, GEPTy);
3883 // Check to see if this is constant foldable.
3884 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3887 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3889 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3894 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3895 const SimplifyQuery &Q) {
3896 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3899 /// Given operands for an InsertValueInst, see if we can fold the result.
3900 /// If not, this returns null.
3901 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3902 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3904 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3905 if (Constant *CVal = dyn_cast<Constant>(Val))
3906 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3908 // insertvalue x, undef, n -> x
3909 if (match(Val, m_Undef()))
3912 // insertvalue x, (extractvalue y, n), n
3913 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3914 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3915 EV->getIndices() == Idxs) {
3916 // insertvalue undef, (extractvalue y, n), n -> y
3917 if (match(Agg, m_Undef()))
3918 return EV->getAggregateOperand();
3920 // insertvalue y, (extractvalue y, n), n -> y
3921 if (Agg == EV->getAggregateOperand())
3928 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3929 ArrayRef<unsigned> Idxs,
3930 const SimplifyQuery &Q) {
3931 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3934 /// Given operands for an ExtractValueInst, see if we can fold the result.
3935 /// If not, this returns null.
3936 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3937 const SimplifyQuery &, unsigned) {
3938 if (auto *CAgg = dyn_cast<Constant>(Agg))
3939 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3941 // extractvalue x, (insertvalue y, elt, n), n -> elt
3942 unsigned NumIdxs = Idxs.size();
3943 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3944 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3945 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3946 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3947 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3948 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3949 Idxs.slice(0, NumCommonIdxs)) {
3950 if (NumIdxs == NumInsertValueIdxs)
3951 return IVI->getInsertedValueOperand();
3959 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3960 const SimplifyQuery &Q) {
3961 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3964 /// Given operands for an ExtractElementInst, see if we can fold the result.
3965 /// If not, this returns null.
3966 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3968 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3969 if (auto *CIdx = dyn_cast<Constant>(Idx))
3970 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3972 // The index is not relevant if our vector is a splat.
3973 if (auto *Splat = CVec->getSplatValue())
3976 if (isa<UndefValue>(Vec))
3977 return UndefValue::get(Vec->getType()->getVectorElementType());
3980 // If extracting a specified index from the vector, see if we can recursively
3981 // find a previously computed scalar that was inserted into the vector.
3982 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3983 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3989 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3990 const SimplifyQuery &Q) {
3991 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3994 /// See if we can fold the given phi. If not, returns null.
3995 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3996 // If all of the PHI's incoming values are the same then replace the PHI node
3997 // with the common value.
3998 Value *CommonValue = nullptr;
3999 bool HasUndefInput = false;
4000 for (Value *Incoming : PN->incoming_values()) {
4001 // If the incoming value is the phi node itself, it can safely be skipped.
4002 if (Incoming == PN) continue;
4003 if (isa<UndefValue>(Incoming)) {
4004 // Remember that we saw an undef value, but otherwise ignore them.
4005 HasUndefInput = true;
4008 if (CommonValue && Incoming != CommonValue)
4009 return nullptr; // Not the same, bail out.
4010 CommonValue = Incoming;
4013 // If CommonValue is null then all of the incoming values were either undef or
4014 // equal to the phi node itself.
4016 return UndefValue::get(PN->getType());
4018 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4019 // instruction, we cannot return X as the result of the PHI node unless it
4020 // dominates the PHI block.
4022 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4027 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4028 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4029 if (auto *C = dyn_cast<Constant>(Op))
4030 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4032 if (auto *CI = dyn_cast<CastInst>(Op)) {
4033 auto *Src = CI->getOperand(0);
4034 Type *SrcTy = Src->getType();
4035 Type *MidTy = CI->getType();
4037 if (Src->getType() == Ty) {
4038 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4039 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4041 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4043 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4045 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4046 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4047 SrcIntPtrTy, MidIntPtrTy,
4048 DstIntPtrTy) == Instruction::BitCast)
4054 if (CastOpc == Instruction::BitCast)
4055 if (Op->getType() == Ty)
4061 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4062 const SimplifyQuery &Q) {
4063 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4066 /// For the given destination element of a shuffle, peek through shuffles to
4067 /// match a root vector source operand that contains that element in the same
4068 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4069 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4070 int MaskVal, Value *RootVec,
4071 unsigned MaxRecurse) {
4075 // Bail out if any mask value is undefined. That kind of shuffle may be
4076 // simplified further based on demanded bits or other folds.
4080 // The mask value chooses which source operand we need to look at next.
4081 int InVecNumElts = Op0->getType()->getVectorNumElements();
4082 int RootElt = MaskVal;
4083 Value *SourceOp = Op0;
4084 if (MaskVal >= InVecNumElts) {
4085 RootElt = MaskVal - InVecNumElts;
4089 // If the source operand is a shuffle itself, look through it to find the
4090 // matching root vector.
4091 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4092 return foldIdentityShuffles(
4093 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4094 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4097 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4100 // The source operand is not a shuffle. Initialize the root vector value for
4101 // this shuffle if that has not been done yet.
4105 // Give up as soon as a source operand does not match the existing root value.
4106 if (RootVec != SourceOp)
4109 // The element must be coming from the same lane in the source vector
4110 // (although it may have crossed lanes in intermediate shuffles).
4111 if (RootElt != DestElt)
4117 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4118 Type *RetTy, const SimplifyQuery &Q,
4119 unsigned MaxRecurse) {
4120 if (isa<UndefValue>(Mask))
4121 return UndefValue::get(RetTy);
4123 Type *InVecTy = Op0->getType();
4124 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4125 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4127 SmallVector<int, 32> Indices;
4128 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4129 assert(MaskNumElts == Indices.size() &&
4130 "Size of Indices not same as number of mask elements?");
4132 // Canonicalization: If mask does not select elements from an input vector,
4133 // replace that input vector with undef.
4134 bool MaskSelects0 = false, MaskSelects1 = false;
4135 for (unsigned i = 0; i != MaskNumElts; ++i) {
4136 if (Indices[i] == -1)
4138 if ((unsigned)Indices[i] < InVecNumElts)
4139 MaskSelects0 = true;
4141 MaskSelects1 = true;
4144 Op0 = UndefValue::get(InVecTy);
4146 Op1 = UndefValue::get(InVecTy);
4148 auto *Op0Const = dyn_cast<Constant>(Op0);
4149 auto *Op1Const = dyn_cast<Constant>(Op1);
4151 // If all operands are constant, constant fold the shuffle.
4152 if (Op0Const && Op1Const)
4153 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4155 // Canonicalization: if only one input vector is constant, it shall be the
4157 if (Op0Const && !Op1Const) {
4158 std::swap(Op0, Op1);
4159 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4162 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4163 // value type is same as the input vectors' type.
4164 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4165 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4166 OpShuf->getMask()->getSplatValue())
4169 // Don't fold a shuffle with undef mask elements. This may get folded in a
4170 // better way using demanded bits or other analysis.
4171 // TODO: Should we allow this?
4172 if (find(Indices, -1) != Indices.end())
4175 // Check if every element of this shuffle can be mapped back to the
4176 // corresponding element of a single root vector. If so, we don't need this
4177 // shuffle. This handles simple identity shuffles as well as chains of
4178 // shuffles that may widen/narrow and/or move elements across lanes and back.
4179 Value *RootVec = nullptr;
4180 for (unsigned i = 0; i != MaskNumElts; ++i) {
4181 // Note that recursion is limited for each vector element, so if any element
4182 // exceeds the limit, this will fail to simplify.
4184 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4186 // We can't replace a widening/narrowing shuffle with one of its operands.
4187 if (!RootVec || RootVec->getType() != RetTy)
4193 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4194 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4195 Type *RetTy, const SimplifyQuery &Q) {
4196 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4199 //=== Helper functions for higher up the class hierarchy.
4201 /// Given operands for a BinaryOperator, see if we can fold the result.
4202 /// If not, this returns null.
4203 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4204 const SimplifyQuery &Q, unsigned MaxRecurse) {
4206 case Instruction::Add:
4207 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4208 case Instruction::FAdd:
4209 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4210 case Instruction::Sub:
4211 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4212 case Instruction::FSub:
4213 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4214 case Instruction::Mul:
4215 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4216 case Instruction::FMul:
4217 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4218 case Instruction::SDiv:
4219 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4220 case Instruction::UDiv:
4221 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4222 case Instruction::FDiv:
4223 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4224 case Instruction::SRem:
4225 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4226 case Instruction::URem:
4227 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4228 case Instruction::FRem:
4229 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4230 case Instruction::Shl:
4231 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4232 case Instruction::LShr:
4233 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4234 case Instruction::AShr:
4235 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4236 case Instruction::And:
4237 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4238 case Instruction::Or:
4239 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4240 case Instruction::Xor:
4241 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4243 llvm_unreachable("Unexpected opcode");
4247 /// Given operands for a BinaryOperator, see if we can fold the result.
4248 /// If not, this returns null.
4249 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4250 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4251 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4252 const FastMathFlags &FMF, const SimplifyQuery &Q,
4253 unsigned MaxRecurse) {
4255 case Instruction::FAdd:
4256 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4257 case Instruction::FSub:
4258 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4259 case Instruction::FMul:
4260 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4261 case Instruction::FDiv:
4262 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4264 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4268 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4269 const SimplifyQuery &Q) {
4270 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4273 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4274 FastMathFlags FMF, const SimplifyQuery &Q) {
4275 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4278 /// Given operands for a CmpInst, see if we can fold the result.
4279 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4280 const SimplifyQuery &Q, unsigned MaxRecurse) {
4281 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4282 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4283 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4286 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4287 const SimplifyQuery &Q) {
4288 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4291 static bool IsIdempotent(Intrinsic::ID ID) {
4293 default: return false;
4295 // Unary idempotent: f(f(x)) = f(x)
4296 case Intrinsic::fabs:
4297 case Intrinsic::floor:
4298 case Intrinsic::ceil:
4299 case Intrinsic::trunc:
4300 case Intrinsic::rint:
4301 case Intrinsic::nearbyint:
4302 case Intrinsic::round:
4307 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4308 const DataLayout &DL) {
4309 GlobalValue *PtrSym;
4311 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4314 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4315 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4316 Type *Int32PtrTy = Int32Ty->getPointerTo();
4317 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4319 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4320 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4323 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4324 if (OffsetInt % 4 != 0)
4327 Constant *C = ConstantExpr::getGetElementPtr(
4328 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4329 ConstantInt::get(Int64Ty, OffsetInt / 4));
4330 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4334 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4338 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4339 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4344 if (LoadedCE->getOpcode() != Instruction::Sub)
4347 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4348 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4350 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4352 Constant *LoadedRHS = LoadedCE->getOperand(1);
4353 GlobalValue *LoadedRHSSym;
4354 APInt LoadedRHSOffset;
4355 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4357 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4360 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4363 static bool maskIsAllZeroOrUndef(Value *Mask) {
4364 auto *ConstMask = dyn_cast<Constant>(Mask);
4367 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4369 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4371 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4372 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4379 template <typename IterTy>
4380 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4381 const SimplifyQuery &Q, unsigned MaxRecurse) {
4382 Intrinsic::ID IID = F->getIntrinsicID();
4383 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4386 if (NumOperands == 1) {
4387 // Perform idempotent optimizations
4388 if (IsIdempotent(IID)) {
4389 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4390 if (II->getIntrinsicID() == IID)
4396 case Intrinsic::fabs: {
4397 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4407 if (NumOperands == 2) {
4408 Value *LHS = *ArgBegin;
4409 Value *RHS = *(ArgBegin + 1);
4410 Type *ReturnType = F->getReturnType();
4413 case Intrinsic::usub_with_overflow:
4414 case Intrinsic::ssub_with_overflow: {
4415 // X - X -> { 0, false }
4417 return Constant::getNullValue(ReturnType);
4419 // X - undef -> undef
4420 // undef - X -> undef
4421 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4422 return UndefValue::get(ReturnType);
4426 case Intrinsic::uadd_with_overflow:
4427 case Intrinsic::sadd_with_overflow: {
4428 // X + undef -> undef
4429 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4430 return UndefValue::get(ReturnType);
4434 case Intrinsic::umul_with_overflow:
4435 case Intrinsic::smul_with_overflow: {
4436 // 0 * X -> { 0, false }
4437 // X * 0 -> { 0, false }
4438 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4439 return Constant::getNullValue(ReturnType);
4441 // undef * X -> { 0, false }
4442 // X * undef -> { 0, false }
4443 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4444 return Constant::getNullValue(ReturnType);
4448 case Intrinsic::load_relative: {
4449 Constant *C0 = dyn_cast<Constant>(LHS);
4450 Constant *C1 = dyn_cast<Constant>(RHS);
4452 return SimplifyRelativeLoad(C0, C1, Q.DL);
4460 // Simplify calls to llvm.masked.load.*
4462 case Intrinsic::masked_load: {
4463 Value *MaskArg = ArgBegin[2];
4464 Value *PassthruArg = ArgBegin[3];
4465 // If the mask is all zeros or undef, the "passthru" argument is the result.
4466 if (maskIsAllZeroOrUndef(MaskArg))
4475 template <typename IterTy>
4476 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4477 IterTy ArgEnd, const SimplifyQuery &Q,
4478 unsigned MaxRecurse) {
4479 Type *Ty = V->getType();
4480 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4481 Ty = PTy->getElementType();
4482 FunctionType *FTy = cast<FunctionType>(Ty);
4484 // call undef -> undef
4485 // call null -> undef
4486 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4487 return UndefValue::get(FTy->getReturnType());
4489 Function *F = dyn_cast<Function>(V);
4493 if (F->isIntrinsic())
4494 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4497 if (!canConstantFoldCallTo(CS, F))
4500 SmallVector<Constant *, 4> ConstantArgs;
4501 ConstantArgs.reserve(ArgEnd - ArgBegin);
4502 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4503 Constant *C = dyn_cast<Constant>(*I);
4506 ConstantArgs.push_back(C);
4509 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4512 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4513 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4514 const SimplifyQuery &Q) {
4515 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4518 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4519 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4520 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4523 /// See if we can compute a simplified version of this instruction.
4524 /// If not, this returns null.
4526 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4527 OptimizationRemarkEmitter *ORE) {
4528 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4531 switch (I->getOpcode()) {
4533 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4535 case Instruction::FAdd:
4536 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4537 I->getFastMathFlags(), Q);
4539 case Instruction::Add:
4540 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4541 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4542 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4544 case Instruction::FSub:
4545 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4546 I->getFastMathFlags(), Q);
4548 case Instruction::Sub:
4549 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4550 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4551 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4553 case Instruction::FMul:
4554 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4555 I->getFastMathFlags(), Q);
4557 case Instruction::Mul:
4558 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4560 case Instruction::SDiv:
4561 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4563 case Instruction::UDiv:
4564 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4566 case Instruction::FDiv:
4567 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4568 I->getFastMathFlags(), Q);
4570 case Instruction::SRem:
4571 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4573 case Instruction::URem:
4574 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4576 case Instruction::FRem:
4577 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4578 I->getFastMathFlags(), Q);
4580 case Instruction::Shl:
4581 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4582 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4583 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4585 case Instruction::LShr:
4586 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4587 cast<BinaryOperator>(I)->isExact(), Q);
4589 case Instruction::AShr:
4590 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4591 cast<BinaryOperator>(I)->isExact(), Q);
4593 case Instruction::And:
4594 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4596 case Instruction::Or:
4597 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4599 case Instruction::Xor:
4600 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4602 case Instruction::ICmp:
4603 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4604 I->getOperand(0), I->getOperand(1), Q);
4606 case Instruction::FCmp:
4608 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4609 I->getOperand(1), I->getFastMathFlags(), Q);
4611 case Instruction::Select:
4612 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4613 I->getOperand(2), Q);
4615 case Instruction::GetElementPtr: {
4616 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4617 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4621 case Instruction::InsertValue: {
4622 InsertValueInst *IV = cast<InsertValueInst>(I);
4623 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4624 IV->getInsertedValueOperand(),
4625 IV->getIndices(), Q);
4628 case Instruction::ExtractValue: {
4629 auto *EVI = cast<ExtractValueInst>(I);
4630 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4631 EVI->getIndices(), Q);
4634 case Instruction::ExtractElement: {
4635 auto *EEI = cast<ExtractElementInst>(I);
4636 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4637 EEI->getIndexOperand(), Q);
4640 case Instruction::ShuffleVector: {
4641 auto *SVI = cast<ShuffleVectorInst>(I);
4642 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4643 SVI->getMask(), SVI->getType(), Q);
4646 case Instruction::PHI:
4647 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4649 case Instruction::Call: {
4650 CallSite CS(cast<CallInst>(I));
4651 Result = SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4655 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4656 #include "llvm/IR/Instruction.def"
4657 #undef HANDLE_CAST_INST
4659 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4661 case Instruction::Alloca:
4662 // No simplifications for Alloca and it can't be constant folded.
4667 // In general, it is possible for computeKnownBits to determine all bits in a
4668 // value even when the operands are not all constants.
4669 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4670 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4671 if (Known.isConstant())
4672 Result = ConstantInt::get(I->getType(), Known.getConstant());
4675 /// If called on unreachable code, the above logic may report that the
4676 /// instruction simplified to itself. Make life easier for users by
4677 /// detecting that case here, returning a safe value instead.
4678 return Result == I ? UndefValue::get(I->getType()) : Result;
4681 /// \brief Implementation of recursive simplification through an instruction's
4684 /// This is the common implementation of the recursive simplification routines.
4685 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4686 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4687 /// instructions to process and attempt to simplify it using
4688 /// InstructionSimplify.
4690 /// This routine returns 'true' only when *it* simplifies something. The passed
4691 /// in simplified value does not count toward this.
4692 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4693 const TargetLibraryInfo *TLI,
4694 const DominatorTree *DT,
4695 AssumptionCache *AC) {
4696 bool Simplified = false;
4697 SmallSetVector<Instruction *, 8> Worklist;
4698 const DataLayout &DL = I->getModule()->getDataLayout();
4700 // If we have an explicit value to collapse to, do that round of the
4701 // simplification loop by hand initially.
4703 for (User *U : I->users())
4705 Worklist.insert(cast<Instruction>(U));
4707 // Replace the instruction with its simplified value.
4708 I->replaceAllUsesWith(SimpleV);
4710 // Gracefully handle edge cases where the instruction is not wired into any
4712 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4713 !I->mayHaveSideEffects())
4714 I->eraseFromParent();
4719 // Note that we must test the size on each iteration, the worklist can grow.
4720 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4723 // See if this instruction simplifies.
4724 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4730 // Stash away all the uses of the old instruction so we can check them for
4731 // recursive simplifications after a RAUW. This is cheaper than checking all
4732 // uses of To on the recursive step in most cases.
4733 for (User *U : I->users())
4734 Worklist.insert(cast<Instruction>(U));
4736 // Replace the instruction with its simplified value.
4737 I->replaceAllUsesWith(SimpleV);
4739 // Gracefully handle edge cases where the instruction is not wired into any
4741 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4742 !I->mayHaveSideEffects())
4743 I->eraseFromParent();
4748 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4749 const TargetLibraryInfo *TLI,
4750 const DominatorTree *DT,
4751 AssumptionCache *AC) {
4752 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4755 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4756 const TargetLibraryInfo *TLI,
4757 const DominatorTree *DT,
4758 AssumptionCache *AC) {
4759 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4760 assert(SimpleV && "Must provide a simplified value.");
4761 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4765 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4766 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4767 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4768 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4769 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4770 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4771 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4772 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4775 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4776 const DataLayout &DL) {
4777 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4780 template <class T, class... TArgs>
4781 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4783 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4784 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4785 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4786 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4788 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,