1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements routines for folding instructions into simpler forms
11 // that do not require creating new instructions. This does constant folding
12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15 // simplified: This is usually true and assuming it simplifies the logic (if
16 // they have not been simplified then results are correct but maybe suboptimal).
18 //===----------------------------------------------------------------------===//
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/CmpInstAnalysis.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/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() == unsigned(Opcode)) {
331 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
332 // We already know that "op" is the same as for the simplified value. See
333 // if the operands match too. If so, return the simplified value.
334 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
335 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
336 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
337 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
338 Simplified->getOperand(1) == UnsimplifiedRHS)
340 if (Simplified->isCommutative() &&
341 Simplified->getOperand(1) == UnsimplifiedLHS &&
342 Simplified->getOperand(0) == UnsimplifiedRHS)
350 /// In the case of a comparison with a select instruction, try to simplify the
351 /// comparison by seeing whether both branches of the select result in the same
352 /// value. Returns the common value if so, otherwise returns null.
353 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
354 Value *RHS, const SimplifyQuery &Q,
355 unsigned MaxRecurse) {
356 // Recursion is always used, so bail out at once if we already hit the limit.
360 // Make sure the select is on the LHS.
361 if (!isa<SelectInst>(LHS)) {
363 Pred = CmpInst::getSwappedPredicate(Pred);
365 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
366 SelectInst *SI = cast<SelectInst>(LHS);
367 Value *Cond = SI->getCondition();
368 Value *TV = SI->getTrueValue();
369 Value *FV = SI->getFalseValue();
371 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
372 // Does "cmp TV, RHS" simplify?
373 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
375 // It not only simplified, it simplified to the select condition. Replace
377 TCmp = getTrue(Cond->getType());
379 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
380 // condition then we can replace it with 'true'. Otherwise give up.
381 if (!isSameCompare(Cond, Pred, TV, RHS))
383 TCmp = getTrue(Cond->getType());
386 // Does "cmp FV, RHS" simplify?
387 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
389 // It not only simplified, it simplified to the select condition. Replace
391 FCmp = getFalse(Cond->getType());
393 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
394 // condition then we can replace it with 'false'. Otherwise give up.
395 if (!isSameCompare(Cond, Pred, FV, RHS))
397 FCmp = getFalse(Cond->getType());
400 // If both sides simplified to the same value, then use it as the result of
401 // the original comparison.
405 // The remaining cases only make sense if the select condition has the same
406 // type as the result of the comparison, so bail out if this is not so.
407 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
409 // If the false value simplified to false, then the result of the compare
410 // is equal to "Cond && TCmp". This also catches the case when the false
411 // value simplified to false and the true value to true, returning "Cond".
412 if (match(FCmp, m_Zero()))
413 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
415 // If the true value simplified to true, then the result of the compare
416 // is equal to "Cond || FCmp".
417 if (match(TCmp, m_One()))
418 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
420 // Finally, if the false value simplified to true and the true value to
421 // false, then the result of the compare is equal to "!Cond".
422 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
424 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
431 /// In the case of a binary operation with an operand that is a PHI instruction,
432 /// try to simplify the binop by seeing whether evaluating it on the incoming
433 /// phi values yields the same result for every value. If so returns the common
434 /// value, otherwise returns null.
435 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
436 Value *RHS, const SimplifyQuery &Q,
437 unsigned MaxRecurse) {
438 // Recursion is always used, so bail out at once if we already hit the limit.
443 if (isa<PHINode>(LHS)) {
444 PI = cast<PHINode>(LHS);
445 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
446 if (!ValueDominatesPHI(RHS, PI, Q.DT))
449 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
450 PI = cast<PHINode>(RHS);
451 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
452 if (!ValueDominatesPHI(LHS, PI, Q.DT))
456 // Evaluate the BinOp on the incoming phi values.
457 Value *CommonValue = nullptr;
458 for (Value *Incoming : PI->incoming_values()) {
459 // If the incoming value is the phi node itself, it can safely be skipped.
460 if (Incoming == PI) continue;
461 Value *V = PI == LHS ?
462 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
463 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
464 // If the operation failed to simplify, or simplified to a different value
465 // to previously, then give up.
466 if (!V || (CommonValue && V != CommonValue))
474 /// In the case of a comparison with a PHI instruction, try to simplify the
475 /// comparison by seeing whether comparing with all of the incoming phi values
476 /// yields the same result every time. If so returns the common result,
477 /// otherwise returns null.
478 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
479 const SimplifyQuery &Q, unsigned MaxRecurse) {
480 // Recursion is always used, so bail out at once if we already hit the limit.
484 // Make sure the phi is on the LHS.
485 if (!isa<PHINode>(LHS)) {
487 Pred = CmpInst::getSwappedPredicate(Pred);
489 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
490 PHINode *PI = cast<PHINode>(LHS);
492 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
493 if (!ValueDominatesPHI(RHS, PI, Q.DT))
496 // Evaluate the BinOp on the incoming phi values.
497 Value *CommonValue = nullptr;
498 for (Value *Incoming : PI->incoming_values()) {
499 // If the incoming value is the phi node itself, it can safely be skipped.
500 if (Incoming == PI) continue;
501 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
502 // If the operation failed to simplify, or simplified to a different value
503 // to previously, then give up.
504 if (!V || (CommonValue && V != CommonValue))
512 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
513 Value *&Op0, Value *&Op1,
514 const SimplifyQuery &Q) {
515 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
516 if (auto *CRHS = dyn_cast<Constant>(Op1))
517 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
519 // Canonicalize the constant to the RHS if this is a commutative operation.
520 if (Instruction::isCommutative(Opcode))
526 /// Given operands for an Add, see if we can fold the result.
527 /// If not, this returns null.
528 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
529 const SimplifyQuery &Q, unsigned MaxRecurse) {
530 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
533 // X + undef -> undef
534 if (match(Op1, m_Undef()))
538 if (match(Op1, m_Zero()))
545 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
546 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
549 // X + ~X -> -1 since ~X = -X-1
550 Type *Ty = Op0->getType();
551 if (match(Op0, m_Not(m_Specific(Op1))) ||
552 match(Op1, m_Not(m_Specific(Op0))))
553 return Constant::getAllOnesValue(Ty);
555 // add nsw/nuw (xor Y, signmask), signmask --> Y
556 // The no-wrapping add guarantees that the top bit will be set by the add.
557 // Therefore, the xor must be clearing the already set sign bit of Y.
558 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
559 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
563 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
564 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
567 // Try some generic simplifications for associative operations.
568 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
572 // Threading Add over selects and phi nodes is pointless, so don't bother.
573 // Threading over the select in "A + select(cond, B, C)" means evaluating
574 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
575 // only if B and C are equal. If B and C are equal then (since we assume
576 // that operands have already been simplified) "select(cond, B, C)" should
577 // have been simplified to the common value of B and C already. Analysing
578 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
579 // for threading over phi nodes.
584 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
585 const SimplifyQuery &Query) {
586 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
589 /// \brief Compute the base pointer and cumulative constant offsets for V.
591 /// This strips all constant offsets off of V, leaving it the base pointer, and
592 /// accumulates the total constant offset applied in the returned constant. It
593 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
594 /// no constant offsets applied.
596 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
597 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
599 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
600 bool AllowNonInbounds = false) {
601 assert(V->getType()->isPtrOrPtrVectorTy());
603 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
604 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
606 // Even though we don't look through PHI nodes, we could be called on an
607 // instruction in an unreachable block, which may be on a cycle.
608 SmallPtrSet<Value *, 4> Visited;
611 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
612 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
613 !GEP->accumulateConstantOffset(DL, Offset))
615 V = GEP->getPointerOperand();
616 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
617 V = cast<Operator>(V)->getOperand(0);
618 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
619 if (GA->isInterposable())
621 V = GA->getAliasee();
623 if (auto CS = CallSite(V))
624 if (Value *RV = CS.getReturnedArgOperand()) {
630 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
631 } while (Visited.insert(V).second);
633 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
634 if (V->getType()->isVectorTy())
635 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
640 /// \brief Compute the constant difference between two pointer values.
641 /// If the difference is not a constant, returns zero.
642 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
644 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
645 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
647 // If LHS and RHS are not related via constant offsets to the same base
648 // value, there is nothing we can do here.
652 // Otherwise, the difference of LHS - RHS can be computed as:
654 // = (LHSOffset + Base) - (RHSOffset + Base)
655 // = LHSOffset - RHSOffset
656 return ConstantExpr::getSub(LHSOffset, RHSOffset);
659 /// Given operands for a Sub, see if we can fold the result.
660 /// If not, this returns null.
661 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
662 const SimplifyQuery &Q, unsigned MaxRecurse) {
663 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
666 // X - undef -> undef
667 // undef - X -> undef
668 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
669 return UndefValue::get(Op0->getType());
672 if (match(Op1, m_Zero()))
677 return Constant::getNullValue(Op0->getType());
679 // Is this a negation?
680 if (match(Op0, m_Zero())) {
681 // 0 - X -> 0 if the sub is NUW.
685 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
686 if (Known.Zero.isMaxSignedValue()) {
687 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
688 // Op1 must be 0 because negating the minimum signed value is undefined.
692 // 0 - X -> X if X is 0 or the minimum signed value.
697 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
698 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
699 Value *X = nullptr, *Y = nullptr, *Z = Op1;
700 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
701 // See if "V === Y - Z" simplifies.
702 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
703 // It does! Now see if "X + V" simplifies.
704 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
705 // It does, we successfully reassociated!
709 // See if "V === X - Z" simplifies.
710 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
711 // It does! Now see if "Y + V" simplifies.
712 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
713 // It does, we successfully reassociated!
719 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
720 // For example, X - (X + 1) -> -1
722 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
723 // See if "V === X - Y" simplifies.
724 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
725 // It does! Now see if "V - Z" simplifies.
726 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
727 // It does, we successfully reassociated!
731 // See if "V === X - Z" simplifies.
732 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
733 // It does! Now see if "V - Y" simplifies.
734 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
735 // It does, we successfully reassociated!
741 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
742 // For example, X - (X - Y) -> Y.
744 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
745 // See if "V === Z - X" simplifies.
746 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
747 // It does! Now see if "V + Y" simplifies.
748 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
749 // It does, we successfully reassociated!
754 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
755 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
756 match(Op1, m_Trunc(m_Value(Y))))
757 if (X->getType() == Y->getType())
758 // See if "V === X - Y" simplifies.
759 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
760 // It does! Now see if "trunc V" simplifies.
761 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
763 // It does, return the simplified "trunc V".
766 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
767 if (match(Op0, m_PtrToInt(m_Value(X))) &&
768 match(Op1, m_PtrToInt(m_Value(Y))))
769 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
770 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
773 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
774 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
777 // Threading Sub over selects and phi nodes is pointless, so don't bother.
778 // Threading over the select in "A - select(cond, B, C)" means evaluating
779 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
780 // only if B and C are equal. If B and C are equal then (since we assume
781 // that operands have already been simplified) "select(cond, B, C)" should
782 // have been simplified to the common value of B and C already. Analysing
783 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
784 // for threading over phi nodes.
789 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
790 const SimplifyQuery &Q) {
791 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
794 /// Given operands for a Mul, see if we can fold the result.
795 /// If not, this returns null.
796 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
797 unsigned MaxRecurse) {
798 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
802 if (match(Op1, m_Undef()))
803 return Constant::getNullValue(Op0->getType());
806 if (match(Op1, m_Zero()))
810 if (match(Op1, m_One()))
813 // (X / Y) * Y -> X if the division is exact.
815 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
816 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
820 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
821 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
824 // Try some generic simplifications for associative operations.
825 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
829 // Mul distributes over Add. Try some generic simplifications based on this.
830 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
834 // If the operation is with the result of a select instruction, check whether
835 // operating on either branch of the select always yields the same value.
836 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
837 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
841 // If the operation is with the result of a phi instruction, check whether
842 // operating on all incoming values of the phi always yields the same value.
843 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
844 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
851 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
852 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
855 /// Check for common or similar folds of integer division or integer remainder.
856 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
857 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
858 Type *Ty = Op0->getType();
860 // X / undef -> undef
861 // X % undef -> undef
862 if (match(Op1, m_Undef()))
867 // We don't need to preserve faults!
868 if (match(Op1, m_Zero()))
869 return UndefValue::get(Ty);
871 // If any element of a constant divisor vector is zero, the whole op is undef.
872 auto *Op1C = dyn_cast<Constant>(Op1);
873 if (Op1C && Ty->isVectorTy()) {
874 unsigned NumElts = Ty->getVectorNumElements();
875 for (unsigned i = 0; i != NumElts; ++i) {
876 Constant *Elt = Op1C->getAggregateElement(i);
877 if (Elt && Elt->isNullValue())
878 return UndefValue::get(Ty);
884 if (match(Op0, m_Undef()))
885 return Constant::getNullValue(Ty);
889 if (match(Op0, m_Zero()))
895 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
899 // If this is a boolean op (single-bit element type), we can't have
900 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
901 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1))
902 return IsDiv ? Op0 : Constant::getNullValue(Ty);
907 /// Given a predicate and two operands, return true if the comparison is true.
908 /// This is a helper for div/rem simplification where we return some other value
909 /// when we can prove a relationship between the operands.
910 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
911 const SimplifyQuery &Q, unsigned MaxRecurse) {
912 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
913 Constant *C = dyn_cast_or_null<Constant>(V);
914 return (C && C->isAllOnesValue());
917 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
918 /// to simplify X % Y to X.
919 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
920 unsigned MaxRecurse, bool IsSigned) {
921 // Recursion is always used, so bail out at once if we already hit the limit.
928 // We require that 1 operand is a simple constant. That could be extended to
929 // 2 variables if we computed the sign bit for each.
931 // Make sure that a constant is not the minimum signed value because taking
932 // the abs() of that is undefined.
933 Type *Ty = X->getType();
935 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
936 // Is the variable divisor magnitude always greater than the constant
937 // dividend magnitude?
938 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
939 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
940 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
941 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
942 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
945 if (match(Y, m_APInt(C))) {
946 // Special-case: we can't take the abs() of a minimum signed value. If
947 // that's the divisor, then all we have to do is prove that the dividend
948 // is also not the minimum signed value.
949 if (C->isMinSignedValue())
950 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
952 // Is the variable dividend magnitude always less than the constant
953 // divisor magnitude?
954 // |X| < |C| --> X > -abs(C) and X < abs(C)
955 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
956 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
957 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
958 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
964 // IsSigned == false.
965 // Is the dividend unsigned less than the divisor?
966 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
969 /// These are simplifications common to SDiv and UDiv.
970 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
971 const SimplifyQuery &Q, unsigned MaxRecurse) {
972 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
975 if (Value *V = simplifyDivRem(Op0, Op1, true))
978 bool IsSigned = Opcode == Instruction::SDiv;
980 // (X * Y) / Y -> X if the multiplication does not overflow.
981 Value *X = nullptr, *Y = nullptr;
982 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
983 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
984 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
985 // If the Mul knows it does not overflow, then we are good to go.
986 if ((IsSigned && Mul->hasNoSignedWrap()) ||
987 (!IsSigned && Mul->hasNoUnsignedWrap()))
989 // If X has the form X = A / Y then X * Y cannot overflow.
990 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
991 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
995 // (X rem Y) / Y -> 0
996 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
997 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
998 return Constant::getNullValue(Op0->getType());
1000 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1001 ConstantInt *C1, *C2;
1002 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1003 match(Op1, m_ConstantInt(C2))) {
1005 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1007 return Constant::getNullValue(Op0->getType());
1010 // If the operation is with the result of a select instruction, check whether
1011 // operating on either branch of the select always yields the same value.
1012 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1013 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1016 // If the operation is with the result of a phi instruction, check whether
1017 // operating on all incoming values of the phi always yields the same value.
1018 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1019 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1022 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1023 return Constant::getNullValue(Op0->getType());
1028 /// These are simplifications common to SRem and URem.
1029 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1030 const SimplifyQuery &Q, unsigned MaxRecurse) {
1031 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1034 if (Value *V = simplifyDivRem(Op0, Op1, false))
1037 // (X % Y) % Y -> X % Y
1038 if ((Opcode == Instruction::SRem &&
1039 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1040 (Opcode == Instruction::URem &&
1041 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1044 // If the operation is with the result of a select instruction, check whether
1045 // operating on either branch of the select always yields the same value.
1046 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1047 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1050 // If the operation is with the result of a phi instruction, check whether
1051 // operating on all incoming values of the phi always yields the same value.
1052 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1053 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1056 // If X / Y == 0, then X % Y == X.
1057 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1063 /// Given operands for an SDiv, see if we can fold the result.
1064 /// If not, this returns null.
1065 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1066 unsigned MaxRecurse) {
1067 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1070 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1071 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1074 /// Given operands for a UDiv, see if we can fold the result.
1075 /// If not, this returns null.
1076 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1077 unsigned MaxRecurse) {
1078 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1081 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1082 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1085 /// Given operands for an SRem, see if we can fold the result.
1086 /// If not, this returns null.
1087 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1088 unsigned MaxRecurse) {
1089 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1092 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1093 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1096 /// Given operands for a URem, see if we can fold the result.
1097 /// If not, this returns null.
1098 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1099 unsigned MaxRecurse) {
1100 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1103 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1104 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1107 /// Returns true if a shift by \c Amount always yields undef.
1108 static bool isUndefShift(Value *Amount) {
1109 Constant *C = dyn_cast<Constant>(Amount);
1113 // X shift by undef -> undef because it may shift by the bitwidth.
1114 if (isa<UndefValue>(C))
1117 // Shifting by the bitwidth or more is undefined.
1118 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1119 if (CI->getValue().getLimitedValue() >=
1120 CI->getType()->getScalarSizeInBits())
1123 // If all lanes of a vector shift are undefined the whole shift is.
1124 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1125 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1126 if (!isUndefShift(C->getAggregateElement(I)))
1134 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1135 /// If not, this returns null.
1136 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1137 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1138 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1141 // 0 shift by X -> 0
1142 if (match(Op0, m_Zero()))
1145 // X shift by 0 -> X
1146 if (match(Op1, m_Zero()))
1149 // Fold undefined shifts.
1150 if (isUndefShift(Op1))
1151 return UndefValue::get(Op0->getType());
1153 // If the operation is with the result of a select instruction, check whether
1154 // operating on either branch of the select always yields the same value.
1155 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1156 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1159 // If the operation is with the result of a phi instruction, check whether
1160 // operating on all incoming values of the phi always yields the same value.
1161 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1162 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1165 // If any bits in the shift amount make that value greater than or equal to
1166 // the number of bits in the type, the shift is undefined.
1167 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1168 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1169 return UndefValue::get(Op0->getType());
1171 // If all valid bits in the shift amount are known zero, the first operand is
1173 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1174 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1180 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1181 /// fold the result. If not, this returns null.
1182 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1183 Value *Op1, bool isExact, const SimplifyQuery &Q,
1184 unsigned MaxRecurse) {
1185 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1190 return Constant::getNullValue(Op0->getType());
1193 // undef >> X -> undef (if it's exact)
1194 if (match(Op0, m_Undef()))
1195 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1197 // The low bit cannot be shifted out of an exact shift if it is set.
1199 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1200 if (Op0Known.One[0])
1207 /// Given operands for an Shl, see if we can fold the result.
1208 /// If not, this returns null.
1209 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1210 const SimplifyQuery &Q, unsigned MaxRecurse) {
1211 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1215 // undef << X -> undef if (if it's NSW/NUW)
1216 if (match(Op0, m_Undef()))
1217 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1219 // (X >> A) << A -> X
1221 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1226 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1227 const SimplifyQuery &Q) {
1228 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1231 /// Given operands for an LShr, see if we can fold the result.
1232 /// If not, this returns null.
1233 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1234 const SimplifyQuery &Q, unsigned MaxRecurse) {
1235 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1239 // (X << A) >> A -> X
1241 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1247 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1248 const SimplifyQuery &Q) {
1249 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1252 /// Given operands for an AShr, see if we can fold the result.
1253 /// If not, this returns null.
1254 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1255 const SimplifyQuery &Q, unsigned MaxRecurse) {
1256 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1260 // all ones >>a X -> all ones
1261 if (match(Op0, m_AllOnes()))
1264 // (X << A) >> A -> X
1266 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1269 // Arithmetic shifting an all-sign-bit value is a no-op.
1270 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1271 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1277 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1278 const SimplifyQuery &Q) {
1279 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1282 /// Commuted variants are assumed to be handled by calling this function again
1283 /// with the parameters swapped.
1284 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1285 ICmpInst *UnsignedICmp, bool IsAnd) {
1288 ICmpInst::Predicate EqPred;
1289 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1290 !ICmpInst::isEquality(EqPred))
1293 ICmpInst::Predicate UnsignedPred;
1294 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1295 ICmpInst::isUnsigned(UnsignedPred))
1297 else if (match(UnsignedICmp,
1298 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1299 ICmpInst::isUnsigned(UnsignedPred))
1300 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1304 // X < Y && Y != 0 --> X < Y
1305 // X < Y || Y != 0 --> Y != 0
1306 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1307 return IsAnd ? UnsignedICmp : ZeroICmp;
1309 // X >= Y || Y != 0 --> true
1310 // X >= Y || Y == 0 --> X >= Y
1311 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1312 if (EqPred == ICmpInst::ICMP_NE)
1313 return getTrue(UnsignedICmp->getType());
1314 return UnsignedICmp;
1317 // X < Y && Y == 0 --> false
1318 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1320 return getFalse(UnsignedICmp->getType());
1325 /// Commuted variants are assumed to be handled by calling this function again
1326 /// with the parameters swapped.
1327 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1328 ICmpInst::Predicate Pred0, Pred1;
1330 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1331 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1334 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1335 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1336 // can eliminate Op1 from this 'and'.
1337 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1340 // Check for any combination of predicates that are guaranteed to be disjoint.
1341 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1342 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1343 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1344 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1345 return getFalse(Op0->getType());
1350 /// Commuted variants are assumed to be handled by calling this function again
1351 /// with the parameters swapped.
1352 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1353 ICmpInst::Predicate Pred0, Pred1;
1355 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1356 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1359 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1360 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1361 // can eliminate Op0 from this 'or'.
1362 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1365 // Check for any combination of predicates that cover the entire range of
1367 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1368 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1369 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1370 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1371 return getTrue(Op0->getType());
1376 /// Test if a pair of compares with a shared operand and 2 constants has an
1377 /// empty set intersection, full set union, or if one compare is a superset of
1379 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1381 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1382 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1385 const APInt *C0, *C1;
1386 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1387 !match(Cmp1->getOperand(1), m_APInt(C1)))
1390 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1391 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1393 // For and-of-compares, check if the intersection is empty:
1394 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1395 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1396 return getFalse(Cmp0->getType());
1398 // For or-of-compares, check if the union is full:
1399 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1400 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1401 return getTrue(Cmp0->getType());
1403 // Is one range a superset of the other?
1404 // If this is and-of-compares, take the smaller set:
1405 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1406 // If this is or-of-compares, take the larger set:
1407 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1408 if (Range0.contains(Range1))
1409 return IsAnd ? Cmp1 : Cmp0;
1410 if (Range1.contains(Range0))
1411 return IsAnd ? Cmp0 : Cmp1;
1416 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1417 // (icmp (add V, C0), C1) & (icmp V, C0)
1418 ICmpInst::Predicate Pred0, Pred1;
1419 const APInt *C0, *C1;
1421 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1424 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1427 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1428 if (AddInst->getOperand(1) != Op1->getOperand(1))
1431 Type *ITy = Op0->getType();
1432 bool isNSW = AddInst->hasNoSignedWrap();
1433 bool isNUW = AddInst->hasNoUnsignedWrap();
1435 const APInt Delta = *C1 - *C0;
1436 if (C0->isStrictlyPositive()) {
1438 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1439 return getFalse(ITy);
1440 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1441 return getFalse(ITy);
1444 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1445 return getFalse(ITy);
1446 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1447 return getFalse(ITy);
1450 if (C0->getBoolValue() && isNUW) {
1452 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1453 return getFalse(ITy);
1455 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1456 return getFalse(ITy);
1462 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1463 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1465 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1468 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1470 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1473 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1476 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1478 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1484 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1485 // (icmp (add V, C0), C1) | (icmp V, C0)
1486 ICmpInst::Predicate Pred0, Pred1;
1487 const APInt *C0, *C1;
1489 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1492 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1495 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1496 if (AddInst->getOperand(1) != Op1->getOperand(1))
1499 Type *ITy = Op0->getType();
1500 bool isNSW = AddInst->hasNoSignedWrap();
1501 bool isNUW = AddInst->hasNoUnsignedWrap();
1503 const APInt Delta = *C1 - *C0;
1504 if (C0->isStrictlyPositive()) {
1506 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1507 return getTrue(ITy);
1508 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1509 return getTrue(ITy);
1512 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1513 return getTrue(ITy);
1514 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1515 return getTrue(ITy);
1518 if (C0->getBoolValue() && isNUW) {
1520 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1521 return getTrue(ITy);
1523 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1524 return getTrue(ITy);
1530 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1531 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1533 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1536 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1538 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1541 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1544 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1546 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1552 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1553 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1554 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1555 if (LHS0->getType() != RHS0->getType())
1558 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1559 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1560 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1561 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1562 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1563 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1564 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1565 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1566 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1567 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1568 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1569 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1570 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1573 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1574 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1575 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1576 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1577 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1578 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1579 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1580 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1581 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1582 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1589 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1590 // Look through casts of the 'and' operands to find compares.
1591 auto *Cast0 = dyn_cast<CastInst>(Op0);
1592 auto *Cast1 = dyn_cast<CastInst>(Op1);
1593 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1594 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1595 Op0 = Cast0->getOperand(0);
1596 Op1 = Cast1->getOperand(0);
1600 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1601 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1603 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1604 simplifyOrOfICmps(ICmp0, ICmp1);
1606 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1607 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1609 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1616 // If we looked through casts, we can only handle a constant simplification
1617 // because we are not allowed to create a cast instruction here.
1618 if (auto *C = dyn_cast<Constant>(V))
1619 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1624 /// Given operands for an And, see if we can fold the result.
1625 /// If not, this returns null.
1626 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1627 unsigned MaxRecurse) {
1628 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1632 if (match(Op1, m_Undef()))
1633 return Constant::getNullValue(Op0->getType());
1640 if (match(Op1, m_Zero()))
1644 if (match(Op1, m_AllOnes()))
1647 // A & ~A = ~A & A = 0
1648 if (match(Op0, m_Not(m_Specific(Op1))) ||
1649 match(Op1, m_Not(m_Specific(Op0))))
1650 return Constant::getNullValue(Op0->getType());
1653 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1657 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1660 // A mask that only clears known zeros of a shifted value is a no-op.
1664 if (match(Op1, m_APInt(Mask))) {
1665 // If all bits in the inverted and shifted mask are clear:
1666 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1667 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1668 (~(*Mask)).lshr(*ShAmt).isNullValue())
1671 // If all bits in the inverted and shifted mask are clear:
1672 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1673 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1674 (~(*Mask)).shl(*ShAmt).isNullValue())
1678 // A & (-A) = A if A is a power of two or zero.
1679 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1680 match(Op1, m_Neg(m_Specific(Op0)))) {
1681 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1684 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1689 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1692 // Try some generic simplifications for associative operations.
1693 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1697 // And distributes over Or. Try some generic simplifications based on this.
1698 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1702 // And distributes over Xor. Try some generic simplifications based on this.
1703 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1707 // If the operation is with the result of a select instruction, check whether
1708 // operating on either branch of the select always yields the same value.
1709 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1710 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1714 // If the operation is with the result of a phi instruction, check whether
1715 // operating on all incoming values of the phi always yields the same value.
1716 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1717 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1724 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1725 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1728 /// Given operands for an Or, see if we can fold the result.
1729 /// If not, this returns null.
1730 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1731 unsigned MaxRecurse) {
1732 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1736 if (match(Op1, m_Undef()))
1737 return Constant::getAllOnesValue(Op0->getType());
1744 if (match(Op1, m_Zero()))
1748 if (match(Op1, m_AllOnes()))
1751 // A | ~A = ~A | A = -1
1752 if (match(Op0, m_Not(m_Specific(Op1))) ||
1753 match(Op1, m_Not(m_Specific(Op0))))
1754 return Constant::getAllOnesValue(Op0->getType());
1757 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1761 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1764 // ~(A & ?) | A = -1
1765 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1766 return Constant::getAllOnesValue(Op1->getType());
1768 // A | ~(A & ?) = -1
1769 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1770 return Constant::getAllOnesValue(Op0->getType());
1773 // (A & ~B) | (A ^ B) -> (A ^ B)
1774 // (~B & A) | (A ^ B) -> (A ^ B)
1775 // (A & ~B) | (B ^ A) -> (B ^ A)
1776 // (~B & A) | (B ^ A) -> (B ^ A)
1777 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1778 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1779 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1782 // Commute the 'or' operands.
1783 // (A ^ B) | (A & ~B) -> (A ^ B)
1784 // (A ^ B) | (~B & A) -> (A ^ B)
1785 // (B ^ A) | (A & ~B) -> (B ^ A)
1786 // (B ^ A) | (~B & A) -> (B ^ A)
1787 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1788 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1789 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1792 // (A & B) | (~A ^ B) -> (~A ^ B)
1793 // (B & A) | (~A ^ B) -> (~A ^ B)
1794 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1795 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1796 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1797 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1798 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1801 // (~A ^ B) | (A & B) -> (~A ^ B)
1802 // (~A ^ B) | (B & A) -> (~A ^ B)
1803 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1804 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1805 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1806 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1807 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1810 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1813 // Try some generic simplifications for associative operations.
1814 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1818 // Or distributes over And. Try some generic simplifications based on this.
1819 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1823 // If the operation is with the result of a select instruction, check whether
1824 // operating on either branch of the select always yields the same value.
1825 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1826 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1830 // (A & C1)|(B & C2)
1831 const APInt *C1, *C2;
1832 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1833 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1835 // (A & C1)|(B & C2)
1836 // If we have: ((V + N) & C1) | (V & C2)
1837 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1838 // replace with V+N.
1840 if (C2->isMask() && // C2 == 0+1+
1841 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1842 // Add commutes, try both ways.
1843 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1846 // Or commutes, try both ways.
1848 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1849 // Add commutes, try both ways.
1850 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1856 // If the operation is with the result of a phi instruction, check whether
1857 // operating on all incoming values of the phi always yields the same value.
1858 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1859 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1865 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1866 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1869 /// Given operands for a Xor, see if we can fold the result.
1870 /// If not, this returns null.
1871 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1872 unsigned MaxRecurse) {
1873 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1876 // A ^ undef -> undef
1877 if (match(Op1, m_Undef()))
1881 if (match(Op1, m_Zero()))
1886 return Constant::getNullValue(Op0->getType());
1888 // A ^ ~A = ~A ^ A = -1
1889 if (match(Op0, m_Not(m_Specific(Op1))) ||
1890 match(Op1, m_Not(m_Specific(Op0))))
1891 return Constant::getAllOnesValue(Op0->getType());
1893 // Try some generic simplifications for associative operations.
1894 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1898 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1899 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1900 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1901 // only if B and C are equal. If B and C are equal then (since we assume
1902 // that operands have already been simplified) "select(cond, B, C)" should
1903 // have been simplified to the common value of B and C already. Analysing
1904 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1905 // for threading over phi nodes.
1910 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1911 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1915 static Type *GetCompareTy(Value *Op) {
1916 return CmpInst::makeCmpResultType(Op->getType());
1919 /// Rummage around inside V looking for something equivalent to the comparison
1920 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1921 /// Helper function for analyzing max/min idioms.
1922 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1923 Value *LHS, Value *RHS) {
1924 SelectInst *SI = dyn_cast<SelectInst>(V);
1927 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1930 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1931 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1933 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1934 LHS == CmpRHS && RHS == CmpLHS)
1939 // A significant optimization not implemented here is assuming that alloca
1940 // addresses are not equal to incoming argument values. They don't *alias*,
1941 // as we say, but that doesn't mean they aren't equal, so we take a
1942 // conservative approach.
1944 // This is inspired in part by C++11 5.10p1:
1945 // "Two pointers of the same type compare equal if and only if they are both
1946 // null, both point to the same function, or both represent the same
1949 // This is pretty permissive.
1951 // It's also partly due to C11 6.5.9p6:
1952 // "Two pointers compare equal if and only if both are null pointers, both are
1953 // pointers to the same object (including a pointer to an object and a
1954 // subobject at its beginning) or function, both are pointers to one past the
1955 // last element of the same array object, or one is a pointer to one past the
1956 // end of one array object and the other is a pointer to the start of a
1957 // different array object that happens to immediately follow the first array
1958 // object in the address space.)
1960 // C11's version is more restrictive, however there's no reason why an argument
1961 // couldn't be a one-past-the-end value for a stack object in the caller and be
1962 // equal to the beginning of a stack object in the callee.
1964 // If the C and C++ standards are ever made sufficiently restrictive in this
1965 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1966 // this optimization.
1968 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
1969 const DominatorTree *DT, CmpInst::Predicate Pred,
1970 AssumptionCache *AC, const Instruction *CxtI,
1971 Value *LHS, Value *RHS) {
1972 // First, skip past any trivial no-ops.
1973 LHS = LHS->stripPointerCasts();
1974 RHS = RHS->stripPointerCasts();
1976 // A non-null pointer is not equal to a null pointer.
1977 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
1978 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1979 return ConstantInt::get(GetCompareTy(LHS),
1980 !CmpInst::isTrueWhenEqual(Pred));
1982 // We can only fold certain predicates on pointer comparisons.
1987 // Equality comaprisons are easy to fold.
1988 case CmpInst::ICMP_EQ:
1989 case CmpInst::ICMP_NE:
1992 // We can only handle unsigned relational comparisons because 'inbounds' on
1993 // a GEP only protects against unsigned wrapping.
1994 case CmpInst::ICMP_UGT:
1995 case CmpInst::ICMP_UGE:
1996 case CmpInst::ICMP_ULT:
1997 case CmpInst::ICMP_ULE:
1998 // However, we have to switch them to their signed variants to handle
1999 // negative indices from the base pointer.
2000 Pred = ICmpInst::getSignedPredicate(Pred);
2004 // Strip off any constant offsets so that we can reason about them.
2005 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2006 // here and compare base addresses like AliasAnalysis does, however there are
2007 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2008 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2009 // doesn't need to guarantee pointer inequality when it says NoAlias.
2010 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2011 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2013 // If LHS and RHS are related via constant offsets to the same base
2014 // value, we can replace it with an icmp which just compares the offsets.
2016 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2018 // Various optimizations for (in)equality comparisons.
2019 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2020 // Different non-empty allocations that exist at the same time have
2021 // different addresses (if the program can tell). Global variables always
2022 // exist, so they always exist during the lifetime of each other and all
2023 // allocas. Two different allocas usually have different addresses...
2025 // However, if there's an @llvm.stackrestore dynamically in between two
2026 // allocas, they may have the same address. It's tempting to reduce the
2027 // scope of the problem by only looking at *static* allocas here. That would
2028 // cover the majority of allocas while significantly reducing the likelihood
2029 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2030 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2031 // an entry block. Also, if we have a block that's not attached to a
2032 // function, we can't tell if it's "static" under the current definition.
2033 // Theoretically, this problem could be fixed by creating a new kind of
2034 // instruction kind specifically for static allocas. Such a new instruction
2035 // could be required to be at the top of the entry block, thus preventing it
2036 // from being subject to a @llvm.stackrestore. Instcombine could even
2037 // convert regular allocas into these special allocas. It'd be nifty.
2038 // However, until then, this problem remains open.
2040 // So, we'll assume that two non-empty allocas have different addresses
2043 // With all that, if the offsets are within the bounds of their allocations
2044 // (and not one-past-the-end! so we can't use inbounds!), and their
2045 // allocations aren't the same, the pointers are not equal.
2047 // Note that it's not necessary to check for LHS being a global variable
2048 // address, due to canonicalization and constant folding.
2049 if (isa<AllocaInst>(LHS) &&
2050 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2051 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2052 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2053 uint64_t LHSSize, RHSSize;
2054 if (LHSOffsetCI && RHSOffsetCI &&
2055 getObjectSize(LHS, LHSSize, DL, TLI) &&
2056 getObjectSize(RHS, RHSSize, DL, TLI)) {
2057 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2058 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2059 if (!LHSOffsetValue.isNegative() &&
2060 !RHSOffsetValue.isNegative() &&
2061 LHSOffsetValue.ult(LHSSize) &&
2062 RHSOffsetValue.ult(RHSSize)) {
2063 return ConstantInt::get(GetCompareTy(LHS),
2064 !CmpInst::isTrueWhenEqual(Pred));
2068 // Repeat the above check but this time without depending on DataLayout
2069 // or being able to compute a precise size.
2070 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2071 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2072 LHSOffset->isNullValue() &&
2073 RHSOffset->isNullValue())
2074 return ConstantInt::get(GetCompareTy(LHS),
2075 !CmpInst::isTrueWhenEqual(Pred));
2078 // Even if an non-inbounds GEP occurs along the path we can still optimize
2079 // equality comparisons concerning the result. We avoid walking the whole
2080 // chain again by starting where the last calls to
2081 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2082 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2083 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2085 return ConstantExpr::getICmp(Pred,
2086 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2087 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2089 // If one side of the equality comparison must come from a noalias call
2090 // (meaning a system memory allocation function), and the other side must
2091 // come from a pointer that cannot overlap with dynamically-allocated
2092 // memory within the lifetime of the current function (allocas, byval
2093 // arguments, globals), then determine the comparison result here.
2094 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2095 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2096 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2098 // Is the set of underlying objects all noalias calls?
2099 auto IsNAC = [](ArrayRef<Value *> Objects) {
2100 return all_of(Objects, isNoAliasCall);
2103 // Is the set of underlying objects all things which must be disjoint from
2104 // noalias calls. For allocas, we consider only static ones (dynamic
2105 // allocas might be transformed into calls to malloc not simultaneously
2106 // live with the compared-to allocation). For globals, we exclude symbols
2107 // that might be resolve lazily to symbols in another dynamically-loaded
2108 // library (and, thus, could be malloc'ed by the implementation).
2109 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2110 return all_of(Objects, [](Value *V) {
2111 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2112 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2113 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2114 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2115 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2116 !GV->isThreadLocal();
2117 if (const Argument *A = dyn_cast<Argument>(V))
2118 return A->hasByValAttr();
2123 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2124 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2125 return ConstantInt::get(GetCompareTy(LHS),
2126 !CmpInst::isTrueWhenEqual(Pred));
2128 // Fold comparisons for non-escaping pointer even if the allocation call
2129 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2130 // dynamic allocation call could be either of the operands.
2131 Value *MI = nullptr;
2132 if (isAllocLikeFn(LHS, TLI) &&
2133 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2135 else if (isAllocLikeFn(RHS, TLI) &&
2136 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2138 // FIXME: We should also fold the compare when the pointer escapes, but the
2139 // compare dominates the pointer escape
2140 if (MI && !PointerMayBeCaptured(MI, true, true))
2141 return ConstantInt::get(GetCompareTy(LHS),
2142 CmpInst::isFalseWhenEqual(Pred));
2149 /// Fold an icmp when its operands have i1 scalar type.
2150 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2151 Value *RHS, const SimplifyQuery &Q) {
2152 Type *ITy = GetCompareTy(LHS); // The return type.
2153 Type *OpTy = LHS->getType(); // The operand type.
2154 if (!OpTy->isIntOrIntVectorTy(1))
2157 // A boolean compared to true/false can be simplified in 14 out of the 20
2158 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2159 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2160 if (match(RHS, m_Zero())) {
2162 case CmpInst::ICMP_NE: // X != 0 -> X
2163 case CmpInst::ICMP_UGT: // X >u 0 -> X
2164 case CmpInst::ICMP_SLT: // X <s 0 -> X
2167 case CmpInst::ICMP_ULT: // X <u 0 -> false
2168 case CmpInst::ICMP_SGT: // X >s 0 -> false
2169 return getFalse(ITy);
2171 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2172 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2173 return getTrue(ITy);
2177 } else if (match(RHS, m_One())) {
2179 case CmpInst::ICMP_EQ: // X == 1 -> X
2180 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2181 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2184 case CmpInst::ICMP_UGT: // X >u 1 -> false
2185 case CmpInst::ICMP_SLT: // X <s -1 -> false
2186 return getFalse(ITy);
2188 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2189 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2190 return getTrue(ITy);
2199 case ICmpInst::ICMP_UGE:
2200 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2201 return getTrue(ITy);
2203 case ICmpInst::ICMP_SGE:
2204 /// For signed comparison, the values for an i1 are 0 and -1
2205 /// respectively. This maps into a truth table of:
2206 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2207 /// 0 | 0 | 1 (0 >= 0) | 1
2208 /// 0 | 1 | 1 (0 >= -1) | 1
2209 /// 1 | 0 | 0 (-1 >= 0) | 0
2210 /// 1 | 1 | 1 (-1 >= -1) | 1
2211 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2212 return getTrue(ITy);
2214 case ICmpInst::ICMP_ULE:
2215 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2216 return getTrue(ITy);
2223 /// Try hard to fold icmp with zero RHS because this is a common case.
2224 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2225 Value *RHS, const SimplifyQuery &Q) {
2226 if (!match(RHS, m_Zero()))
2229 Type *ITy = GetCompareTy(LHS); // The return type.
2232 llvm_unreachable("Unknown ICmp predicate!");
2233 case ICmpInst::ICMP_ULT:
2234 return getFalse(ITy);
2235 case ICmpInst::ICMP_UGE:
2236 return getTrue(ITy);
2237 case ICmpInst::ICMP_EQ:
2238 case ICmpInst::ICMP_ULE:
2239 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2240 return getFalse(ITy);
2242 case ICmpInst::ICMP_NE:
2243 case ICmpInst::ICMP_UGT:
2244 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2245 return getTrue(ITy);
2247 case ICmpInst::ICMP_SLT: {
2248 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2249 if (LHSKnown.isNegative())
2250 return getTrue(ITy);
2251 if (LHSKnown.isNonNegative())
2252 return getFalse(ITy);
2255 case ICmpInst::ICMP_SLE: {
2256 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2257 if (LHSKnown.isNegative())
2258 return getTrue(ITy);
2259 if (LHSKnown.isNonNegative() &&
2260 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2261 return getFalse(ITy);
2264 case ICmpInst::ICMP_SGE: {
2265 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2266 if (LHSKnown.isNegative())
2267 return getFalse(ITy);
2268 if (LHSKnown.isNonNegative())
2269 return getTrue(ITy);
2272 case ICmpInst::ICMP_SGT: {
2273 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2274 if (LHSKnown.isNegative())
2275 return getFalse(ITy);
2276 if (LHSKnown.isNonNegative() &&
2277 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2278 return getTrue(ITy);
2286 /// Many binary operators with a constant operand have an easy-to-compute
2287 /// range of outputs. This can be used to fold a comparison to always true or
2289 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2290 unsigned Width = Lower.getBitWidth();
2292 switch (BO.getOpcode()) {
2293 case Instruction::Add:
2294 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2295 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2296 if (BO.hasNoUnsignedWrap()) {
2297 // 'add nuw x, C' produces [C, UINT_MAX].
2299 } else if (BO.hasNoSignedWrap()) {
2300 if (C->isNegative()) {
2301 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2302 Lower = APInt::getSignedMinValue(Width);
2303 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2305 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2306 Lower = APInt::getSignedMinValue(Width) + *C;
2307 Upper = APInt::getSignedMaxValue(Width) + 1;
2313 case Instruction::And:
2314 if (match(BO.getOperand(1), m_APInt(C)))
2315 // 'and x, C' produces [0, C].
2319 case Instruction::Or:
2320 if (match(BO.getOperand(1), m_APInt(C)))
2321 // 'or x, C' produces [C, UINT_MAX].
2325 case Instruction::AShr:
2326 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2327 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2328 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2329 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2330 } else if (match(BO.getOperand(0), m_APInt(C))) {
2331 unsigned ShiftAmount = Width - 1;
2332 if (!C->isNullValue() && BO.isExact())
2333 ShiftAmount = C->countTrailingZeros();
2334 if (C->isNegative()) {
2335 // 'ashr C, x' produces [C, C >> (Width-1)]
2337 Upper = C->ashr(ShiftAmount) + 1;
2339 // 'ashr C, x' produces [C >> (Width-1), C]
2340 Lower = C->ashr(ShiftAmount);
2346 case Instruction::LShr:
2347 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2348 // 'lshr x, C' produces [0, UINT_MAX >> C].
2349 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2350 } else if (match(BO.getOperand(0), m_APInt(C))) {
2351 // 'lshr C, x' produces [C >> (Width-1), C].
2352 unsigned ShiftAmount = Width - 1;
2353 if (!C->isNullValue() && BO.isExact())
2354 ShiftAmount = C->countTrailingZeros();
2355 Lower = C->lshr(ShiftAmount);
2360 case Instruction::Shl:
2361 if (match(BO.getOperand(0), m_APInt(C))) {
2362 if (BO.hasNoUnsignedWrap()) {
2363 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2365 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2366 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2367 if (C->isNegative()) {
2368 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2369 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2370 Lower = C->shl(ShiftAmount);
2373 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2374 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2376 Upper = C->shl(ShiftAmount) + 1;
2382 case Instruction::SDiv:
2383 if (match(BO.getOperand(1), m_APInt(C))) {
2384 APInt IntMin = APInt::getSignedMinValue(Width);
2385 APInt IntMax = APInt::getSignedMaxValue(Width);
2386 if (C->isAllOnesValue()) {
2387 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2388 // where C != -1 and C != 0 and C != 1
2391 } else if (C->countLeadingZeros() < Width - 1) {
2392 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2393 // where C != -1 and C != 0 and C != 1
2394 Lower = IntMin.sdiv(*C);
2395 Upper = IntMax.sdiv(*C);
2396 if (Lower.sgt(Upper))
2397 std::swap(Lower, Upper);
2399 assert(Upper != Lower && "Upper part of range has wrapped!");
2401 } else if (match(BO.getOperand(0), m_APInt(C))) {
2402 if (C->isMinSignedValue()) {
2403 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2405 Upper = Lower.lshr(1) + 1;
2407 // 'sdiv C, x' produces [-|C|, |C|].
2408 Upper = C->abs() + 1;
2409 Lower = (-Upper) + 1;
2414 case Instruction::UDiv:
2415 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2416 // 'udiv x, C' produces [0, UINT_MAX / C].
2417 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2418 } else if (match(BO.getOperand(0), m_APInt(C))) {
2419 // 'udiv C, x' produces [0, C].
2424 case Instruction::SRem:
2425 if (match(BO.getOperand(1), m_APInt(C))) {
2426 // 'srem x, C' produces (-|C|, |C|).
2428 Lower = (-Upper) + 1;
2432 case Instruction::URem:
2433 if (match(BO.getOperand(1), m_APInt(C)))
2434 // 'urem x, C' produces [0, C).
2443 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2446 if (!match(RHS, m_APInt(C)))
2449 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2450 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2451 if (RHS_CR.isEmptySet())
2452 return ConstantInt::getFalse(GetCompareTy(RHS));
2453 if (RHS_CR.isFullSet())
2454 return ConstantInt::getTrue(GetCompareTy(RHS));
2456 // Find the range of possible values for binary operators.
2457 unsigned Width = C->getBitWidth();
2458 APInt Lower = APInt(Width, 0);
2459 APInt Upper = APInt(Width, 0);
2460 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2461 setLimitsForBinOp(*BO, Lower, Upper);
2463 ConstantRange LHS_CR =
2464 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2466 if (auto *I = dyn_cast<Instruction>(LHS))
2467 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2468 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2470 if (!LHS_CR.isFullSet()) {
2471 if (RHS_CR.contains(LHS_CR))
2472 return ConstantInt::getTrue(GetCompareTy(RHS));
2473 if (RHS_CR.inverse().contains(LHS_CR))
2474 return ConstantInt::getFalse(GetCompareTy(RHS));
2480 /// TODO: A large part of this logic is duplicated in InstCombine's
2481 /// foldICmpBinOp(). We should be able to share that and avoid the code
2483 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2484 Value *RHS, const SimplifyQuery &Q,
2485 unsigned MaxRecurse) {
2486 Type *ITy = GetCompareTy(LHS); // The return type.
2488 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2489 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2490 if (MaxRecurse && (LBO || RBO)) {
2491 // Analyze the case when either LHS or RHS is an add instruction.
2492 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2493 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2494 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2495 if (LBO && LBO->getOpcode() == Instruction::Add) {
2496 A = LBO->getOperand(0);
2497 B = LBO->getOperand(1);
2499 ICmpInst::isEquality(Pred) ||
2500 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2501 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2503 if (RBO && RBO->getOpcode() == Instruction::Add) {
2504 C = RBO->getOperand(0);
2505 D = RBO->getOperand(1);
2507 ICmpInst::isEquality(Pred) ||
2508 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2509 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2512 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2513 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2514 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2515 Constant::getNullValue(RHS->getType()), Q,
2519 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2520 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2522 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2523 C == LHS ? D : C, Q, MaxRecurse - 1))
2526 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2527 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2529 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2532 // C + B == C + D -> B == D
2535 } else if (A == D) {
2536 // D + B == C + D -> B == C
2539 } else if (B == C) {
2540 // A + C == C + D -> A == D
2545 // A + D == C + D -> A == C
2549 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2556 // icmp pred (or X, Y), X
2557 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2558 if (Pred == ICmpInst::ICMP_ULT)
2559 return getFalse(ITy);
2560 if (Pred == ICmpInst::ICMP_UGE)
2561 return getTrue(ITy);
2563 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2564 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2565 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2566 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2567 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2568 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2569 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2572 // icmp pred X, (or X, Y)
2573 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2574 if (Pred == ICmpInst::ICMP_ULE)
2575 return getTrue(ITy);
2576 if (Pred == ICmpInst::ICMP_UGT)
2577 return getFalse(ITy);
2579 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2580 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2581 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2582 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2583 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2584 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2585 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2590 // icmp pred (and X, Y), X
2591 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2592 if (Pred == ICmpInst::ICMP_UGT)
2593 return getFalse(ITy);
2594 if (Pred == ICmpInst::ICMP_ULE)
2595 return getTrue(ITy);
2597 // icmp pred X, (and X, Y)
2598 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2599 if (Pred == ICmpInst::ICMP_UGE)
2600 return getTrue(ITy);
2601 if (Pred == ICmpInst::ICMP_ULT)
2602 return getFalse(ITy);
2605 // 0 - (zext X) pred C
2606 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2607 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2608 if (RHSC->getValue().isStrictlyPositive()) {
2609 if (Pred == ICmpInst::ICMP_SLT)
2610 return ConstantInt::getTrue(RHSC->getContext());
2611 if (Pred == ICmpInst::ICMP_SGE)
2612 return ConstantInt::getFalse(RHSC->getContext());
2613 if (Pred == ICmpInst::ICMP_EQ)
2614 return ConstantInt::getFalse(RHSC->getContext());
2615 if (Pred == ICmpInst::ICMP_NE)
2616 return ConstantInt::getTrue(RHSC->getContext());
2618 if (RHSC->getValue().isNonNegative()) {
2619 if (Pred == ICmpInst::ICMP_SLE)
2620 return ConstantInt::getTrue(RHSC->getContext());
2621 if (Pred == ICmpInst::ICMP_SGT)
2622 return ConstantInt::getFalse(RHSC->getContext());
2627 // icmp pred (urem X, Y), Y
2628 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2632 case ICmpInst::ICMP_SGT:
2633 case ICmpInst::ICMP_SGE: {
2634 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2635 if (!Known.isNonNegative())
2639 case ICmpInst::ICMP_EQ:
2640 case ICmpInst::ICMP_UGT:
2641 case ICmpInst::ICMP_UGE:
2642 return getFalse(ITy);
2643 case ICmpInst::ICMP_SLT:
2644 case ICmpInst::ICMP_SLE: {
2645 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2646 if (!Known.isNonNegative())
2650 case ICmpInst::ICMP_NE:
2651 case ICmpInst::ICMP_ULT:
2652 case ICmpInst::ICMP_ULE:
2653 return getTrue(ITy);
2657 // icmp pred X, (urem Y, X)
2658 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2662 case ICmpInst::ICMP_SGT:
2663 case ICmpInst::ICMP_SGE: {
2664 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2665 if (!Known.isNonNegative())
2669 case ICmpInst::ICMP_NE:
2670 case ICmpInst::ICMP_UGT:
2671 case ICmpInst::ICMP_UGE:
2672 return getTrue(ITy);
2673 case ICmpInst::ICMP_SLT:
2674 case ICmpInst::ICMP_SLE: {
2675 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2676 if (!Known.isNonNegative())
2680 case ICmpInst::ICMP_EQ:
2681 case ICmpInst::ICMP_ULT:
2682 case ICmpInst::ICMP_ULE:
2683 return getFalse(ITy);
2689 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2690 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2691 // icmp pred (X op Y), X
2692 if (Pred == ICmpInst::ICMP_UGT)
2693 return getFalse(ITy);
2694 if (Pred == ICmpInst::ICMP_ULE)
2695 return getTrue(ITy);
2700 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2701 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2702 // icmp pred X, (X op Y)
2703 if (Pred == ICmpInst::ICMP_ULT)
2704 return getFalse(ITy);
2705 if (Pred == ICmpInst::ICMP_UGE)
2706 return getTrue(ITy);
2713 // where CI2 is a power of 2 and CI isn't
2714 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2715 const APInt *CI2Val, *CIVal = &CI->getValue();
2716 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2717 CI2Val->isPowerOf2()) {
2718 if (!CIVal->isPowerOf2()) {
2719 // CI2 << X can equal zero in some circumstances,
2720 // this simplification is unsafe if CI is zero.
2722 // We know it is safe if:
2723 // - The shift is nsw, we can't shift out the one bit.
2724 // - The shift is nuw, we can't shift out the one bit.
2727 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2728 CI2Val->isOneValue() || !CI->isZero()) {
2729 if (Pred == ICmpInst::ICMP_EQ)
2730 return ConstantInt::getFalse(RHS->getContext());
2731 if (Pred == ICmpInst::ICMP_NE)
2732 return ConstantInt::getTrue(RHS->getContext());
2735 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2736 if (Pred == ICmpInst::ICMP_UGT)
2737 return ConstantInt::getFalse(RHS->getContext());
2738 if (Pred == ICmpInst::ICMP_ULE)
2739 return ConstantInt::getTrue(RHS->getContext());
2744 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2745 LBO->getOperand(1) == RBO->getOperand(1)) {
2746 switch (LBO->getOpcode()) {
2749 case Instruction::UDiv:
2750 case Instruction::LShr:
2751 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2753 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2754 RBO->getOperand(0), Q, MaxRecurse - 1))
2757 case Instruction::SDiv:
2758 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2760 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2761 RBO->getOperand(0), Q, MaxRecurse - 1))
2764 case Instruction::AShr:
2765 if (!LBO->isExact() || !RBO->isExact())
2767 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2768 RBO->getOperand(0), Q, MaxRecurse - 1))
2771 case Instruction::Shl: {
2772 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2773 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2776 if (!NSW && ICmpInst::isSigned(Pred))
2778 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2779 RBO->getOperand(0), Q, MaxRecurse - 1))
2788 /// Simplify integer comparisons where at least one operand of the compare
2789 /// matches an integer min/max idiom.
2790 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2791 Value *RHS, const SimplifyQuery &Q,
2792 unsigned MaxRecurse) {
2793 Type *ITy = GetCompareTy(LHS); // The return type.
2795 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2796 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2798 // Signed variants on "max(a,b)>=a -> true".
2799 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2801 std::swap(A, B); // smax(A, B) pred A.
2802 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2803 // We analyze this as smax(A, B) pred A.
2805 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2806 (A == LHS || B == LHS)) {
2808 std::swap(A, B); // A pred smax(A, B).
2809 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2810 // We analyze this as smax(A, B) swapped-pred A.
2811 P = CmpInst::getSwappedPredicate(Pred);
2812 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2813 (A == RHS || B == RHS)) {
2815 std::swap(A, B); // smin(A, B) pred A.
2816 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2817 // We analyze this as smax(-A, -B) swapped-pred -A.
2818 // Note that we do not need to actually form -A or -B thanks to EqP.
2819 P = CmpInst::getSwappedPredicate(Pred);
2820 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2821 (A == LHS || B == LHS)) {
2823 std::swap(A, B); // A pred smin(A, B).
2824 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2825 // We analyze this as smax(-A, -B) pred -A.
2826 // Note that we do not need to actually form -A or -B thanks to EqP.
2829 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2830 // Cases correspond to "max(A, B) p A".
2834 case CmpInst::ICMP_EQ:
2835 case CmpInst::ICMP_SLE:
2836 // Equivalent to "A EqP B". This may be the same as the condition tested
2837 // in the max/min; if so, we can just return that.
2838 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2840 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2842 // Otherwise, see if "A EqP B" simplifies.
2844 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2847 case CmpInst::ICMP_NE:
2848 case CmpInst::ICMP_SGT: {
2849 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2850 // Equivalent to "A InvEqP B". This may be the same as the condition
2851 // tested in the max/min; if so, we can just return that.
2852 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2854 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2856 // Otherwise, see if "A InvEqP B" simplifies.
2858 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2862 case CmpInst::ICMP_SGE:
2864 return getTrue(ITy);
2865 case CmpInst::ICMP_SLT:
2867 return getFalse(ITy);
2871 // Unsigned variants on "max(a,b)>=a -> true".
2872 P = CmpInst::BAD_ICMP_PREDICATE;
2873 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2875 std::swap(A, B); // umax(A, B) pred A.
2876 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2877 // We analyze this as umax(A, B) pred A.
2879 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2880 (A == LHS || B == LHS)) {
2882 std::swap(A, B); // A pred umax(A, B).
2883 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2884 // We analyze this as umax(A, B) swapped-pred A.
2885 P = CmpInst::getSwappedPredicate(Pred);
2886 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2887 (A == RHS || B == RHS)) {
2889 std::swap(A, B); // umin(A, B) pred A.
2890 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2891 // We analyze this as umax(-A, -B) swapped-pred -A.
2892 // Note that we do not need to actually form -A or -B thanks to EqP.
2893 P = CmpInst::getSwappedPredicate(Pred);
2894 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2895 (A == LHS || B == LHS)) {
2897 std::swap(A, B); // A pred umin(A, B).
2898 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2899 // We analyze this as umax(-A, -B) pred -A.
2900 // Note that we do not need to actually form -A or -B thanks to EqP.
2903 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2904 // Cases correspond to "max(A, B) p A".
2908 case CmpInst::ICMP_EQ:
2909 case CmpInst::ICMP_ULE:
2910 // Equivalent to "A EqP B". This may be the same as the condition tested
2911 // in the max/min; if so, we can just return that.
2912 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2914 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2916 // Otherwise, see if "A EqP B" simplifies.
2918 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2921 case CmpInst::ICMP_NE:
2922 case CmpInst::ICMP_UGT: {
2923 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2924 // Equivalent to "A InvEqP B". This may be the same as the condition
2925 // tested in the max/min; if so, we can just return that.
2926 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2928 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2930 // Otherwise, see if "A InvEqP B" simplifies.
2932 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2936 case CmpInst::ICMP_UGE:
2938 return getTrue(ITy);
2939 case CmpInst::ICMP_ULT:
2941 return getFalse(ITy);
2945 // Variants on "max(x,y) >= min(x,z)".
2947 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2948 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2949 (A == C || A == D || B == C || B == D)) {
2950 // max(x, ?) pred min(x, ?).
2951 if (Pred == CmpInst::ICMP_SGE)
2953 return getTrue(ITy);
2954 if (Pred == CmpInst::ICMP_SLT)
2956 return getFalse(ITy);
2957 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2958 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2959 (A == C || A == D || B == C || B == D)) {
2960 // min(x, ?) pred max(x, ?).
2961 if (Pred == CmpInst::ICMP_SLE)
2963 return getTrue(ITy);
2964 if (Pred == CmpInst::ICMP_SGT)
2966 return getFalse(ITy);
2967 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2968 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2969 (A == C || A == D || B == C || B == D)) {
2970 // max(x, ?) pred min(x, ?).
2971 if (Pred == CmpInst::ICMP_UGE)
2973 return getTrue(ITy);
2974 if (Pred == CmpInst::ICMP_ULT)
2976 return getFalse(ITy);
2977 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2978 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2979 (A == C || A == D || B == C || B == D)) {
2980 // min(x, ?) pred max(x, ?).
2981 if (Pred == CmpInst::ICMP_ULE)
2983 return getTrue(ITy);
2984 if (Pred == CmpInst::ICMP_UGT)
2986 return getFalse(ITy);
2992 /// Given operands for an ICmpInst, see if we can fold the result.
2993 /// If not, this returns null.
2994 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2995 const SimplifyQuery &Q, unsigned MaxRecurse) {
2996 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2997 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
2999 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3000 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3001 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3003 // If we have a constant, make sure it is on the RHS.
3004 std::swap(LHS, RHS);
3005 Pred = CmpInst::getSwappedPredicate(Pred);
3008 Type *ITy = GetCompareTy(LHS); // The return type.
3010 // icmp X, X -> true/false
3011 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3012 // because X could be 0.
3013 if (LHS == RHS || isa<UndefValue>(RHS))
3014 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3016 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3019 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3022 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3025 // If both operands have range metadata, use the metadata
3026 // to simplify the comparison.
3027 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3028 auto RHS_Instr = cast<Instruction>(RHS);
3029 auto LHS_Instr = cast<Instruction>(LHS);
3031 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3032 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3033 auto RHS_CR = getConstantRangeFromMetadata(
3034 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3035 auto LHS_CR = getConstantRangeFromMetadata(
3036 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3038 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3039 if (Satisfied_CR.contains(LHS_CR))
3040 return ConstantInt::getTrue(RHS->getContext());
3042 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3043 CmpInst::getInversePredicate(Pred), RHS_CR);
3044 if (InversedSatisfied_CR.contains(LHS_CR))
3045 return ConstantInt::getFalse(RHS->getContext());
3049 // Compare of cast, for example (zext X) != 0 -> X != 0
3050 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3051 Instruction *LI = cast<CastInst>(LHS);
3052 Value *SrcOp = LI->getOperand(0);
3053 Type *SrcTy = SrcOp->getType();
3054 Type *DstTy = LI->getType();
3056 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3057 // if the integer type is the same size as the pointer type.
3058 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3059 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3060 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3061 // Transfer the cast to the constant.
3062 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3063 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3066 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3067 if (RI->getOperand(0)->getType() == SrcTy)
3068 // Compare without the cast.
3069 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3075 if (isa<ZExtInst>(LHS)) {
3076 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3078 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3079 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3080 // Compare X and Y. Note that signed predicates become unsigned.
3081 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3082 SrcOp, RI->getOperand(0), Q,
3086 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3087 // too. If not, then try to deduce the result of the comparison.
3088 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3089 // Compute the constant that would happen if we truncated to SrcTy then
3090 // reextended to DstTy.
3091 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3092 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3094 // If the re-extended constant didn't change then this is effectively
3095 // also a case of comparing two zero-extended values.
3096 if (RExt == CI && MaxRecurse)
3097 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3098 SrcOp, Trunc, Q, MaxRecurse-1))
3101 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3102 // there. Use this to work out the result of the comparison.
3105 default: llvm_unreachable("Unknown ICmp predicate!");
3107 case ICmpInst::ICMP_EQ:
3108 case ICmpInst::ICMP_UGT:
3109 case ICmpInst::ICMP_UGE:
3110 return ConstantInt::getFalse(CI->getContext());
3112 case ICmpInst::ICMP_NE:
3113 case ICmpInst::ICMP_ULT:
3114 case ICmpInst::ICMP_ULE:
3115 return ConstantInt::getTrue(CI->getContext());
3117 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3118 // is non-negative then LHS <s RHS.
3119 case ICmpInst::ICMP_SGT:
3120 case ICmpInst::ICMP_SGE:
3121 return CI->getValue().isNegative() ?
3122 ConstantInt::getTrue(CI->getContext()) :
3123 ConstantInt::getFalse(CI->getContext());
3125 case ICmpInst::ICMP_SLT:
3126 case ICmpInst::ICMP_SLE:
3127 return CI->getValue().isNegative() ?
3128 ConstantInt::getFalse(CI->getContext()) :
3129 ConstantInt::getTrue(CI->getContext());
3135 if (isa<SExtInst>(LHS)) {
3136 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3138 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3139 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3140 // Compare X and Y. Note that the predicate does not change.
3141 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3145 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3146 // too. If not, then try to deduce the result of the comparison.
3147 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3148 // Compute the constant that would happen if we truncated to SrcTy then
3149 // reextended to DstTy.
3150 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3151 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3153 // If the re-extended constant didn't change then this is effectively
3154 // also a case of comparing two sign-extended values.
3155 if (RExt == CI && MaxRecurse)
3156 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3159 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3160 // bits there. Use this to work out the result of the comparison.
3163 default: llvm_unreachable("Unknown ICmp predicate!");
3164 case ICmpInst::ICMP_EQ:
3165 return ConstantInt::getFalse(CI->getContext());
3166 case ICmpInst::ICMP_NE:
3167 return ConstantInt::getTrue(CI->getContext());
3169 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3171 case ICmpInst::ICMP_SGT:
3172 case ICmpInst::ICMP_SGE:
3173 return CI->getValue().isNegative() ?
3174 ConstantInt::getTrue(CI->getContext()) :
3175 ConstantInt::getFalse(CI->getContext());
3176 case ICmpInst::ICMP_SLT:
3177 case ICmpInst::ICMP_SLE:
3178 return CI->getValue().isNegative() ?
3179 ConstantInt::getFalse(CI->getContext()) :
3180 ConstantInt::getTrue(CI->getContext());
3182 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3184 case ICmpInst::ICMP_UGT:
3185 case ICmpInst::ICMP_UGE:
3186 // Comparison is true iff the LHS <s 0.
3188 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3189 Constant::getNullValue(SrcTy),
3193 case ICmpInst::ICMP_ULT:
3194 case ICmpInst::ICMP_ULE:
3195 // Comparison is true iff the LHS >=s 0.
3197 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3198 Constant::getNullValue(SrcTy),
3208 // icmp eq|ne X, Y -> false|true if X != Y
3209 if (ICmpInst::isEquality(Pred) &&
3210 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3211 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3214 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3217 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3220 // Simplify comparisons of related pointers using a powerful, recursive
3221 // GEP-walk when we have target data available..
3222 if (LHS->getType()->isPointerTy())
3223 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3226 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3227 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3228 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3229 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3230 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3231 Q.DL.getTypeSizeInBits(CRHS->getType()))
3232 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3233 CLHS->getPointerOperand(),
3234 CRHS->getPointerOperand()))
3237 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3238 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3239 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3240 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3241 (ICmpInst::isEquality(Pred) ||
3242 (GLHS->isInBounds() && GRHS->isInBounds() &&
3243 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3244 // The bases are equal and the indices are constant. Build a constant
3245 // expression GEP with the same indices and a null base pointer to see
3246 // what constant folding can make out of it.
3247 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3248 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3249 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3250 GLHS->getSourceElementType(), Null, IndicesLHS);
3252 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3253 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3254 GLHS->getSourceElementType(), Null, IndicesRHS);
3255 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3260 // If the comparison is with the result of a select instruction, check whether
3261 // comparing with either branch of the select always yields the same value.
3262 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3263 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3266 // If the comparison is with the result of a phi instruction, check whether
3267 // doing the compare with each incoming phi value yields a common result.
3268 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3269 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3275 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3276 const SimplifyQuery &Q) {
3277 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3280 /// Given operands for an FCmpInst, see if we can fold the result.
3281 /// If not, this returns null.
3282 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3283 FastMathFlags FMF, const SimplifyQuery &Q,
3284 unsigned MaxRecurse) {
3285 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3286 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3288 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3289 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3290 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3292 // If we have a constant, make sure it is on the RHS.
3293 std::swap(LHS, RHS);
3294 Pred = CmpInst::getSwappedPredicate(Pred);
3297 // Fold trivial predicates.
3298 Type *RetTy = GetCompareTy(LHS);
3299 if (Pred == FCmpInst::FCMP_FALSE)
3300 return getFalse(RetTy);
3301 if (Pred == FCmpInst::FCMP_TRUE)
3302 return getTrue(RetTy);
3304 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3306 if (Pred == FCmpInst::FCMP_UNO)
3307 return getFalse(RetTy);
3308 if (Pred == FCmpInst::FCMP_ORD)
3309 return getTrue(RetTy);
3312 // fcmp pred x, undef and fcmp pred undef, x
3313 // fold to true if unordered, false if ordered
3314 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3315 // Choosing NaN for the undef will always make unordered comparison succeed
3316 // and ordered comparison fail.
3317 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3320 // fcmp x,x -> true/false. Not all compares are foldable.
3322 if (CmpInst::isTrueWhenEqual(Pred))
3323 return getTrue(RetTy);
3324 if (CmpInst::isFalseWhenEqual(Pred))
3325 return getFalse(RetTy);
3328 // Handle fcmp with constant RHS.
3330 if (match(RHS, m_APFloat(C))) {
3331 // If the constant is a nan, see if we can fold the comparison based on it.
3333 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3334 return getFalse(RetTy);
3335 assert(FCmpInst::isUnordered(Pred) &&
3336 "Comparison must be either ordered or unordered!");
3337 // True if unordered.
3338 return getTrue(RetTy);
3340 // Check whether the constant is an infinity.
3341 if (C->isInfinity()) {
3342 if (C->isNegative()) {
3344 case FCmpInst::FCMP_OLT:
3345 // No value is ordered and less than negative infinity.
3346 return getFalse(RetTy);
3347 case FCmpInst::FCMP_UGE:
3348 // All values are unordered with or at least negative infinity.
3349 return getTrue(RetTy);
3355 case FCmpInst::FCMP_OGT:
3356 // No value is ordered and greater than infinity.
3357 return getFalse(RetTy);
3358 case FCmpInst::FCMP_ULE:
3359 // All values are unordered with and at most infinity.
3360 return getTrue(RetTy);
3368 case FCmpInst::FCMP_UGE:
3369 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3370 return getTrue(RetTy);
3372 case FCmpInst::FCMP_OLT:
3374 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3375 return getFalse(RetTy);
3380 } else if (C->isNegative()) {
3381 assert(!C->isNaN() && "Unexpected NaN constant!");
3382 // TODO: We can catch more cases by using a range check rather than
3383 // relying on CannotBeOrderedLessThanZero.
3385 case FCmpInst::FCMP_UGE:
3386 case FCmpInst::FCMP_UGT:
3387 case FCmpInst::FCMP_UNE:
3388 // (X >= 0) implies (X > C) when (C < 0)
3389 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3390 return getTrue(RetTy);
3392 case FCmpInst::FCMP_OEQ:
3393 case FCmpInst::FCMP_OLE:
3394 case FCmpInst::FCMP_OLT:
3395 // (X >= 0) implies !(X < C) when (C < 0)
3396 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3397 return getFalse(RetTy);
3405 // If the comparison is with the result of a select instruction, check whether
3406 // comparing with either branch of the select always yields the same value.
3407 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3408 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3411 // If the comparison is with the result of a phi instruction, check whether
3412 // doing the compare with each incoming phi value yields a common result.
3413 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3414 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3420 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3421 FastMathFlags FMF, const SimplifyQuery &Q) {
3422 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3425 /// See if V simplifies when its operand Op is replaced with RepOp.
3426 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3427 const SimplifyQuery &Q,
3428 unsigned MaxRecurse) {
3429 // Trivial replacement.
3433 // We cannot replace a constant, and shouldn't even try.
3434 if (isa<Constant>(Op))
3437 auto *I = dyn_cast<Instruction>(V);
3441 // If this is a binary operator, try to simplify it with the replaced op.
3442 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3444 // %cmp = icmp eq i32 %x, 2147483647
3445 // %add = add nsw i32 %x, 1
3446 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3448 // We can't replace %sel with %add unless we strip away the flags.
3449 if (isa<OverflowingBinaryOperator>(B))
3450 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3452 if (isa<PossiblyExactOperator>(B))
3457 if (B->getOperand(0) == Op)
3458 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3460 if (B->getOperand(1) == Op)
3461 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3466 // Same for CmpInsts.
3467 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3469 if (C->getOperand(0) == Op)
3470 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3472 if (C->getOperand(1) == Op)
3473 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3478 // TODO: We could hand off more cases to instsimplify here.
3480 // If all operands are constant after substituting Op for RepOp then we can
3481 // constant fold the instruction.
3482 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3483 // Build a list of all constant operands.
3484 SmallVector<Constant *, 8> ConstOps;
3485 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3486 if (I->getOperand(i) == Op)
3487 ConstOps.push_back(CRepOp);
3488 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3489 ConstOps.push_back(COp);
3494 // All operands were constants, fold it.
3495 if (ConstOps.size() == I->getNumOperands()) {
3496 if (CmpInst *C = dyn_cast<CmpInst>(I))
3497 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3498 ConstOps[1], Q.DL, Q.TLI);
3500 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3501 if (!LI->isVolatile())
3502 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3504 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3511 /// Try to simplify a select instruction when its condition operand is an
3512 /// integer comparison where one operand of the compare is a constant.
3513 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3514 const APInt *Y, bool TrueWhenUnset) {
3517 // (X & Y) == 0 ? X & ~Y : X --> X
3518 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3519 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3521 return TrueWhenUnset ? FalseVal : TrueVal;
3523 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3524 // (X & Y) != 0 ? X : X & ~Y --> X
3525 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3527 return TrueWhenUnset ? FalseVal : TrueVal;
3529 if (Y->isPowerOf2()) {
3530 // (X & Y) == 0 ? X | Y : X --> X | Y
3531 // (X & Y) != 0 ? X | Y : X --> X
3532 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3534 return TrueWhenUnset ? TrueVal : FalseVal;
3536 // (X & Y) == 0 ? X : X | Y --> X
3537 // (X & Y) != 0 ? X : X | Y --> X | Y
3538 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3540 return TrueWhenUnset ? TrueVal : FalseVal;
3546 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3548 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3549 ICmpInst::Predicate Pred,
3550 Value *TrueVal, Value *FalseVal) {
3553 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3556 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3557 Pred == ICmpInst::ICMP_EQ);
3560 /// Try to simplify a select instruction when its condition operand is an
3561 /// integer comparison.
3562 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3563 Value *FalseVal, const SimplifyQuery &Q,
3564 unsigned MaxRecurse) {
3565 ICmpInst::Predicate Pred;
3566 Value *CmpLHS, *CmpRHS;
3567 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3570 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3573 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3574 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3575 Pred == ICmpInst::ICMP_EQ))
3579 // Check for other compares that behave like bit test.
3580 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3584 if (CondVal->hasOneUse()) {
3586 if (match(CmpRHS, m_APInt(C))) {
3587 // X < MIN ? T : F --> F
3588 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3590 // X < MIN ? T : F --> F
3591 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3593 // X > MAX ? T : F --> F
3594 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3596 // X > MAX ? T : F --> F
3597 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3602 // If we have an equality comparison, then we know the value in one of the
3603 // arms of the select. See if substituting this value into the arm and
3604 // simplifying the result yields the same value as the other arm.
3605 if (Pred == ICmpInst::ICMP_EQ) {
3606 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3608 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3611 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3613 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3616 } else if (Pred == ICmpInst::ICMP_NE) {
3617 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3619 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3622 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3624 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3632 /// Given operands for a SelectInst, see if we can fold the result.
3633 /// If not, this returns null.
3634 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3635 Value *FalseVal, const SimplifyQuery &Q,
3636 unsigned MaxRecurse) {
3637 // select true, X, Y -> X
3638 // select false, X, Y -> Y
3639 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3640 if (Constant *CT = dyn_cast<Constant>(TrueVal))
3641 if (Constant *CF = dyn_cast<Constant>(FalseVal))
3642 return ConstantFoldSelectInstruction(CB, CT, CF);
3643 if (CB->isAllOnesValue())
3645 if (CB->isNullValue())
3649 // select C, X, X -> X
3650 if (TrueVal == FalseVal)
3653 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3654 if (isa<Constant>(FalseVal))
3658 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3660 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3664 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3670 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3671 const SimplifyQuery &Q) {
3672 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3675 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3676 /// If not, this returns null.
3677 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3678 const SimplifyQuery &Q, unsigned) {
3679 // The type of the GEP pointer operand.
3681 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3683 // getelementptr P -> P.
3684 if (Ops.size() == 1)
3687 // Compute the (pointer) type returned by the GEP instruction.
3688 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3689 Type *GEPTy = PointerType::get(LastType, AS);
3690 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3691 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3692 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3693 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3695 if (isa<UndefValue>(Ops[0]))
3696 return UndefValue::get(GEPTy);
3698 if (Ops.size() == 2) {
3699 // getelementptr P, 0 -> P.
3700 if (match(Ops[1], m_Zero()))
3704 if (Ty->isSized()) {
3707 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3708 // getelementptr P, N -> P if P points to a type of zero size.
3709 if (TyAllocSize == 0)
3712 // The following transforms are only safe if the ptrtoint cast
3713 // doesn't truncate the pointers.
3714 if (Ops[1]->getType()->getScalarSizeInBits() ==
3715 Q.DL.getPointerSizeInBits(AS)) {
3716 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3717 if (match(P, m_Zero()))
3718 return Constant::getNullValue(GEPTy);
3720 if (match(P, m_PtrToInt(m_Value(Temp))))
3721 if (Temp->getType() == GEPTy)
3726 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3727 if (TyAllocSize == 1 &&
3728 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3729 if (Value *R = PtrToIntOrZero(P))
3732 // getelementptr V, (ashr (sub P, V), C) -> Q
3733 // if P points to a type of size 1 << C.
3735 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3736 m_ConstantInt(C))) &&
3737 TyAllocSize == 1ULL << C)
3738 if (Value *R = PtrToIntOrZero(P))
3741 // getelementptr V, (sdiv (sub P, V), C) -> Q
3742 // if P points to a type of size C.
3744 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3745 m_SpecificInt(TyAllocSize))))
3746 if (Value *R = PtrToIntOrZero(P))
3752 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3753 all_of(Ops.slice(1).drop_back(1),
3754 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3756 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3757 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3758 APInt BasePtrOffset(PtrWidth, 0);
3759 Value *StrippedBasePtr =
3760 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3763 // gep (gep V, C), (sub 0, V) -> C
3764 if (match(Ops.back(),
3765 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3766 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3767 return ConstantExpr::getIntToPtr(CI, GEPTy);
3769 // gep (gep V, C), (xor V, -1) -> C-1
3770 if (match(Ops.back(),
3771 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3772 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3773 return ConstantExpr::getIntToPtr(CI, GEPTy);
3778 // Check to see if this is constant foldable.
3779 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3782 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3784 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3789 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3790 const SimplifyQuery &Q) {
3791 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3794 /// Given operands for an InsertValueInst, see if we can fold the result.
3795 /// If not, this returns null.
3796 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3797 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3799 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3800 if (Constant *CVal = dyn_cast<Constant>(Val))
3801 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3803 // insertvalue x, undef, n -> x
3804 if (match(Val, m_Undef()))
3807 // insertvalue x, (extractvalue y, n), n
3808 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3809 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3810 EV->getIndices() == Idxs) {
3811 // insertvalue undef, (extractvalue y, n), n -> y
3812 if (match(Agg, m_Undef()))
3813 return EV->getAggregateOperand();
3815 // insertvalue y, (extractvalue y, n), n -> y
3816 if (Agg == EV->getAggregateOperand())
3823 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3824 ArrayRef<unsigned> Idxs,
3825 const SimplifyQuery &Q) {
3826 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3829 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3830 const SimplifyQuery &Q) {
3831 // Try to constant fold.
3832 auto *VecC = dyn_cast<Constant>(Vec);
3833 auto *ValC = dyn_cast<Constant>(Val);
3834 auto *IdxC = dyn_cast<Constant>(Idx);
3835 if (VecC && ValC && IdxC)
3836 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3838 // Fold into undef if index is out of bounds.
3839 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3840 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3842 if (CI->uge(NumElements))
3843 return UndefValue::get(Vec->getType());
3846 // TODO: We should also fold if index is iteslf an undef.
3851 /// Given operands for an ExtractValueInst, see if we can fold the result.
3852 /// If not, this returns null.
3853 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3854 const SimplifyQuery &, unsigned) {
3855 if (auto *CAgg = dyn_cast<Constant>(Agg))
3856 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3858 // extractvalue x, (insertvalue y, elt, n), n -> elt
3859 unsigned NumIdxs = Idxs.size();
3860 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3861 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3862 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3863 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3864 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3865 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3866 Idxs.slice(0, NumCommonIdxs)) {
3867 if (NumIdxs == NumInsertValueIdxs)
3868 return IVI->getInsertedValueOperand();
3876 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3877 const SimplifyQuery &Q) {
3878 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3881 /// Given operands for an ExtractElementInst, see if we can fold the result.
3882 /// If not, this returns null.
3883 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3885 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3886 if (auto *CIdx = dyn_cast<Constant>(Idx))
3887 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3889 // The index is not relevant if our vector is a splat.
3890 if (auto *Splat = CVec->getSplatValue())
3893 if (isa<UndefValue>(Vec))
3894 return UndefValue::get(Vec->getType()->getVectorElementType());
3897 // If extracting a specified index from the vector, see if we can recursively
3898 // find a previously computed scalar that was inserted into the vector.
3899 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3900 if (IdxC->getValue().ule(Vec->getType()->getVectorNumElements()))
3901 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3904 // An undef extract index can be arbitrarily chosen to be an out-of-range
3905 // index value, which would result in the instruction being undef.
3906 if (isa<UndefValue>(Idx))
3907 return UndefValue::get(Vec->getType()->getVectorElementType());
3912 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3913 const SimplifyQuery &Q) {
3914 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3917 /// See if we can fold the given phi. If not, returns null.
3918 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3919 // If all of the PHI's incoming values are the same then replace the PHI node
3920 // with the common value.
3921 Value *CommonValue = nullptr;
3922 bool HasUndefInput = false;
3923 for (Value *Incoming : PN->incoming_values()) {
3924 // If the incoming value is the phi node itself, it can safely be skipped.
3925 if (Incoming == PN) continue;
3926 if (isa<UndefValue>(Incoming)) {
3927 // Remember that we saw an undef value, but otherwise ignore them.
3928 HasUndefInput = true;
3931 if (CommonValue && Incoming != CommonValue)
3932 return nullptr; // Not the same, bail out.
3933 CommonValue = Incoming;
3936 // If CommonValue is null then all of the incoming values were either undef or
3937 // equal to the phi node itself.
3939 return UndefValue::get(PN->getType());
3941 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3942 // instruction, we cannot return X as the result of the PHI node unless it
3943 // dominates the PHI block.
3945 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3950 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
3951 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
3952 if (auto *C = dyn_cast<Constant>(Op))
3953 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
3955 if (auto *CI = dyn_cast<CastInst>(Op)) {
3956 auto *Src = CI->getOperand(0);
3957 Type *SrcTy = Src->getType();
3958 Type *MidTy = CI->getType();
3960 if (Src->getType() == Ty) {
3961 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
3962 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
3964 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
3966 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
3968 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
3969 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
3970 SrcIntPtrTy, MidIntPtrTy,
3971 DstIntPtrTy) == Instruction::BitCast)
3977 if (CastOpc == Instruction::BitCast)
3978 if (Op->getType() == Ty)
3984 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
3985 const SimplifyQuery &Q) {
3986 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
3989 /// For the given destination element of a shuffle, peek through shuffles to
3990 /// match a root vector source operand that contains that element in the same
3991 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
3992 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
3993 int MaskVal, Value *RootVec,
3994 unsigned MaxRecurse) {
3998 // Bail out if any mask value is undefined. That kind of shuffle may be
3999 // simplified further based on demanded bits or other folds.
4003 // The mask value chooses which source operand we need to look at next.
4004 int InVecNumElts = Op0->getType()->getVectorNumElements();
4005 int RootElt = MaskVal;
4006 Value *SourceOp = Op0;
4007 if (MaskVal >= InVecNumElts) {
4008 RootElt = MaskVal - InVecNumElts;
4012 // If the source operand is a shuffle itself, look through it to find the
4013 // matching root vector.
4014 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4015 return foldIdentityShuffles(
4016 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4017 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4020 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4023 // The source operand is not a shuffle. Initialize the root vector value for
4024 // this shuffle if that has not been done yet.
4028 // Give up as soon as a source operand does not match the existing root value.
4029 if (RootVec != SourceOp)
4032 // The element must be coming from the same lane in the source vector
4033 // (although it may have crossed lanes in intermediate shuffles).
4034 if (RootElt != DestElt)
4040 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4041 Type *RetTy, const SimplifyQuery &Q,
4042 unsigned MaxRecurse) {
4043 if (isa<UndefValue>(Mask))
4044 return UndefValue::get(RetTy);
4046 Type *InVecTy = Op0->getType();
4047 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4048 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4050 SmallVector<int, 32> Indices;
4051 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4052 assert(MaskNumElts == Indices.size() &&
4053 "Size of Indices not same as number of mask elements?");
4055 // Canonicalization: If mask does not select elements from an input vector,
4056 // replace that input vector with undef.
4057 bool MaskSelects0 = false, MaskSelects1 = false;
4058 for (unsigned i = 0; i != MaskNumElts; ++i) {
4059 if (Indices[i] == -1)
4061 if ((unsigned)Indices[i] < InVecNumElts)
4062 MaskSelects0 = true;
4064 MaskSelects1 = true;
4067 Op0 = UndefValue::get(InVecTy);
4069 Op1 = UndefValue::get(InVecTy);
4071 auto *Op0Const = dyn_cast<Constant>(Op0);
4072 auto *Op1Const = dyn_cast<Constant>(Op1);
4074 // If all operands are constant, constant fold the shuffle.
4075 if (Op0Const && Op1Const)
4076 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4078 // Canonicalization: if only one input vector is constant, it shall be the
4080 if (Op0Const && !Op1Const) {
4081 std::swap(Op0, Op1);
4082 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4085 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4086 // value type is same as the input vectors' type.
4087 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4088 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4089 OpShuf->getMask()->getSplatValue())
4092 // Don't fold a shuffle with undef mask elements. This may get folded in a
4093 // better way using demanded bits or other analysis.
4094 // TODO: Should we allow this?
4095 if (find(Indices, -1) != Indices.end())
4098 // Check if every element of this shuffle can be mapped back to the
4099 // corresponding element of a single root vector. If so, we don't need this
4100 // shuffle. This handles simple identity shuffles as well as chains of
4101 // shuffles that may widen/narrow and/or move elements across lanes and back.
4102 Value *RootVec = nullptr;
4103 for (unsigned i = 0; i != MaskNumElts; ++i) {
4104 // Note that recursion is limited for each vector element, so if any element
4105 // exceeds the limit, this will fail to simplify.
4107 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4109 // We can't replace a widening/narrowing shuffle with one of its operands.
4110 if (!RootVec || RootVec->getType() != RetTy)
4116 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4117 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4118 Type *RetTy, const SimplifyQuery &Q) {
4119 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4122 /// Given operands for an FAdd, see if we can fold the result. If not, this
4124 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4125 const SimplifyQuery &Q, unsigned MaxRecurse) {
4126 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4130 if (match(Op1, m_NegZero()))
4133 // fadd X, 0 ==> X, when we know X is not -0
4134 if (match(Op1, m_Zero()) &&
4135 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4138 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
4139 // where nnan and ninf have to occur at least once somewhere in this
4141 Value *SubOp = nullptr;
4142 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
4144 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
4147 Instruction *FSub = cast<Instruction>(SubOp);
4148 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
4149 (FMF.noInfs() || FSub->hasNoInfs()))
4150 return Constant::getNullValue(Op0->getType());
4156 /// Given operands for an FSub, see if we can fold the result. If not, this
4158 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4159 const SimplifyQuery &Q, unsigned MaxRecurse) {
4160 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4164 if (match(Op1, m_Zero()))
4167 // fsub X, -0 ==> X, when we know X is not -0
4168 if (match(Op1, m_NegZero()) &&
4169 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4172 // fsub -0.0, (fsub -0.0, X) ==> X
4174 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
4177 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4178 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
4179 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
4182 // fsub nnan x, x ==> 0.0
4183 if (FMF.noNaNs() && Op0 == Op1)
4184 return Constant::getNullValue(Op0->getType());
4189 /// Given the operands for an FMul, see if we can fold the result
4190 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4191 const SimplifyQuery &Q, unsigned MaxRecurse) {
4192 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4195 // fmul X, 1.0 ==> X
4196 if (match(Op1, m_FPOne()))
4199 // fmul nnan nsz X, 0 ==> 0
4200 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
4206 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4207 const SimplifyQuery &Q) {
4208 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4212 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4213 const SimplifyQuery &Q) {
4214 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4217 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4218 const SimplifyQuery &Q) {
4219 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4222 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4223 const SimplifyQuery &Q, unsigned) {
4224 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4227 // undef / X -> undef (the undef could be a snan).
4228 if (match(Op0, m_Undef()))
4231 // X / undef -> undef
4232 if (match(Op1, m_Undef()))
4236 if (match(Op1, m_FPOne()))
4240 // Requires that NaNs are off (X could be zero) and signed zeroes are
4241 // ignored (X could be positive or negative, so the output sign is unknown).
4242 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4246 // X / X -> 1.0 is legal when NaNs are ignored.
4248 return ConstantFP::get(Op0->getType(), 1.0);
4250 // -X / X -> -1.0 and
4251 // X / -X -> -1.0 are legal when NaNs are ignored.
4252 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4253 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4254 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4255 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4256 BinaryOperator::getFNegArgument(Op1) == Op0))
4257 return ConstantFP::get(Op0->getType(), -1.0);
4263 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4264 const SimplifyQuery &Q) {
4265 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4268 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4269 const SimplifyQuery &Q, unsigned) {
4270 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4273 // undef % X -> undef (the undef could be a snan).
4274 if (match(Op0, m_Undef()))
4277 // X % undef -> undef
4278 if (match(Op1, m_Undef()))
4282 // Requires that NaNs are off (X could be zero) and signed zeroes are
4283 // ignored (X could be positive or negative, so the output sign is unknown).
4284 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4290 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4291 const SimplifyQuery &Q) {
4292 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4295 //=== Helper functions for higher up the class hierarchy.
4297 /// Given operands for a BinaryOperator, see if we can fold the result.
4298 /// If not, this returns null.
4299 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4300 const SimplifyQuery &Q, unsigned MaxRecurse) {
4302 case Instruction::Add:
4303 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4304 case Instruction::Sub:
4305 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4306 case Instruction::Mul:
4307 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4308 case Instruction::SDiv:
4309 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4310 case Instruction::UDiv:
4311 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4312 case Instruction::SRem:
4313 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4314 case Instruction::URem:
4315 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4316 case Instruction::Shl:
4317 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4318 case Instruction::LShr:
4319 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4320 case Instruction::AShr:
4321 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4322 case Instruction::And:
4323 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4324 case Instruction::Or:
4325 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4326 case Instruction::Xor:
4327 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4328 case Instruction::FAdd:
4329 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4330 case Instruction::FSub:
4331 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4332 case Instruction::FMul:
4333 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4334 case Instruction::FDiv:
4335 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4336 case Instruction::FRem:
4337 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4339 llvm_unreachable("Unexpected opcode");
4343 /// Given operands for a BinaryOperator, see if we can fold the result.
4344 /// If not, this returns null.
4345 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4346 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4347 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4348 const FastMathFlags &FMF, const SimplifyQuery &Q,
4349 unsigned MaxRecurse) {
4351 case Instruction::FAdd:
4352 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4353 case Instruction::FSub:
4354 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4355 case Instruction::FMul:
4356 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4357 case Instruction::FDiv:
4358 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4360 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4364 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4365 const SimplifyQuery &Q) {
4366 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4369 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4370 FastMathFlags FMF, const SimplifyQuery &Q) {
4371 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4374 /// Given operands for a CmpInst, see if we can fold the result.
4375 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4376 const SimplifyQuery &Q, unsigned MaxRecurse) {
4377 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4378 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4379 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4382 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4383 const SimplifyQuery &Q) {
4384 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4387 static bool IsIdempotent(Intrinsic::ID ID) {
4389 default: return false;
4391 // Unary idempotent: f(f(x)) = f(x)
4392 case Intrinsic::fabs:
4393 case Intrinsic::floor:
4394 case Intrinsic::ceil:
4395 case Intrinsic::trunc:
4396 case Intrinsic::rint:
4397 case Intrinsic::nearbyint:
4398 case Intrinsic::round:
4399 case Intrinsic::canonicalize:
4404 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4405 const DataLayout &DL) {
4406 GlobalValue *PtrSym;
4408 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4411 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4412 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4413 Type *Int32PtrTy = Int32Ty->getPointerTo();
4414 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4416 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4417 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4420 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4421 if (OffsetInt % 4 != 0)
4424 Constant *C = ConstantExpr::getGetElementPtr(
4425 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4426 ConstantInt::get(Int64Ty, OffsetInt / 4));
4427 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4431 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4435 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4436 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4441 if (LoadedCE->getOpcode() != Instruction::Sub)
4444 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4445 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4447 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4449 Constant *LoadedRHS = LoadedCE->getOperand(1);
4450 GlobalValue *LoadedRHSSym;
4451 APInt LoadedRHSOffset;
4452 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4454 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4457 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4460 static bool maskIsAllZeroOrUndef(Value *Mask) {
4461 auto *ConstMask = dyn_cast<Constant>(Mask);
4464 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4466 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4468 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4469 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4476 template <typename IterTy>
4477 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4478 const SimplifyQuery &Q, unsigned MaxRecurse) {
4479 Intrinsic::ID IID = F->getIntrinsicID();
4480 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4483 if (NumOperands == 1) {
4484 // Perform idempotent optimizations
4485 if (IsIdempotent(IID)) {
4486 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4487 if (II->getIntrinsicID() == IID)
4493 case Intrinsic::fabs: {
4494 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4504 if (NumOperands == 2) {
4505 Value *LHS = *ArgBegin;
4506 Value *RHS = *(ArgBegin + 1);
4507 Type *ReturnType = F->getReturnType();
4510 case Intrinsic::usub_with_overflow:
4511 case Intrinsic::ssub_with_overflow: {
4512 // X - X -> { 0, false }
4514 return Constant::getNullValue(ReturnType);
4516 // X - undef -> undef
4517 // undef - X -> undef
4518 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4519 return UndefValue::get(ReturnType);
4523 case Intrinsic::uadd_with_overflow:
4524 case Intrinsic::sadd_with_overflow: {
4525 // X + undef -> undef
4526 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4527 return UndefValue::get(ReturnType);
4531 case Intrinsic::umul_with_overflow:
4532 case Intrinsic::smul_with_overflow: {
4533 // 0 * X -> { 0, false }
4534 // X * 0 -> { 0, false }
4535 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4536 return Constant::getNullValue(ReturnType);
4538 // undef * X -> { 0, false }
4539 // X * undef -> { 0, false }
4540 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4541 return Constant::getNullValue(ReturnType);
4545 case Intrinsic::load_relative: {
4546 Constant *C0 = dyn_cast<Constant>(LHS);
4547 Constant *C1 = dyn_cast<Constant>(RHS);
4549 return SimplifyRelativeLoad(C0, C1, Q.DL);
4557 // Simplify calls to llvm.masked.load.*
4559 case Intrinsic::masked_load: {
4560 Value *MaskArg = ArgBegin[2];
4561 Value *PassthruArg = ArgBegin[3];
4562 // If the mask is all zeros or undef, the "passthru" argument is the result.
4563 if (maskIsAllZeroOrUndef(MaskArg))
4572 template <typename IterTy>
4573 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4574 IterTy ArgEnd, const SimplifyQuery &Q,
4575 unsigned MaxRecurse) {
4576 Type *Ty = V->getType();
4577 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4578 Ty = PTy->getElementType();
4579 FunctionType *FTy = cast<FunctionType>(Ty);
4581 // call undef -> undef
4582 // call null -> undef
4583 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4584 return UndefValue::get(FTy->getReturnType());
4586 Function *F = dyn_cast<Function>(V);
4590 if (F->isIntrinsic())
4591 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4594 if (!canConstantFoldCallTo(CS, F))
4597 SmallVector<Constant *, 4> ConstantArgs;
4598 ConstantArgs.reserve(ArgEnd - ArgBegin);
4599 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4600 Constant *C = dyn_cast<Constant>(*I);
4603 ConstantArgs.push_back(C);
4606 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4609 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4610 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4611 const SimplifyQuery &Q) {
4612 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4615 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4616 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4617 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4620 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4621 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4622 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4626 /// See if we can compute a simplified version of this instruction.
4627 /// If not, this returns null.
4629 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4630 OptimizationRemarkEmitter *ORE) {
4631 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4634 switch (I->getOpcode()) {
4636 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4638 case Instruction::FAdd:
4639 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4640 I->getFastMathFlags(), Q);
4642 case Instruction::Add:
4643 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4644 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4645 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4647 case Instruction::FSub:
4648 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4649 I->getFastMathFlags(), Q);
4651 case Instruction::Sub:
4652 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4653 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4654 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4656 case Instruction::FMul:
4657 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4658 I->getFastMathFlags(), Q);
4660 case Instruction::Mul:
4661 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4663 case Instruction::SDiv:
4664 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4666 case Instruction::UDiv:
4667 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4669 case Instruction::FDiv:
4670 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4671 I->getFastMathFlags(), Q);
4673 case Instruction::SRem:
4674 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4676 case Instruction::URem:
4677 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4679 case Instruction::FRem:
4680 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4681 I->getFastMathFlags(), Q);
4683 case Instruction::Shl:
4684 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4685 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4686 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4688 case Instruction::LShr:
4689 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4690 cast<BinaryOperator>(I)->isExact(), Q);
4692 case Instruction::AShr:
4693 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4694 cast<BinaryOperator>(I)->isExact(), Q);
4696 case Instruction::And:
4697 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4699 case Instruction::Or:
4700 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4702 case Instruction::Xor:
4703 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4705 case Instruction::ICmp:
4706 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4707 I->getOperand(0), I->getOperand(1), Q);
4709 case Instruction::FCmp:
4711 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4712 I->getOperand(1), I->getFastMathFlags(), Q);
4714 case Instruction::Select:
4715 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4716 I->getOperand(2), Q);
4718 case Instruction::GetElementPtr: {
4719 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4720 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4724 case Instruction::InsertValue: {
4725 InsertValueInst *IV = cast<InsertValueInst>(I);
4726 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4727 IV->getInsertedValueOperand(),
4728 IV->getIndices(), Q);
4731 case Instruction::InsertElement: {
4732 auto *IE = cast<InsertElementInst>(I);
4733 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4734 IE->getOperand(2), Q);
4737 case Instruction::ExtractValue: {
4738 auto *EVI = cast<ExtractValueInst>(I);
4739 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4740 EVI->getIndices(), Q);
4743 case Instruction::ExtractElement: {
4744 auto *EEI = cast<ExtractElementInst>(I);
4745 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4746 EEI->getIndexOperand(), Q);
4749 case Instruction::ShuffleVector: {
4750 auto *SVI = cast<ShuffleVectorInst>(I);
4751 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4752 SVI->getMask(), SVI->getType(), Q);
4755 case Instruction::PHI:
4756 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4758 case Instruction::Call: {
4759 CallSite CS(cast<CallInst>(I));
4760 Result = SimplifyCall(CS, Q);
4763 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4764 #include "llvm/IR/Instruction.def"
4765 #undef HANDLE_CAST_INST
4767 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4769 case Instruction::Alloca:
4770 // No simplifications for Alloca and it can't be constant folded.
4775 // In general, it is possible for computeKnownBits to determine all bits in a
4776 // value even when the operands are not all constants.
4777 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4778 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4779 if (Known.isConstant())
4780 Result = ConstantInt::get(I->getType(), Known.getConstant());
4783 /// If called on unreachable code, the above logic may report that the
4784 /// instruction simplified to itself. Make life easier for users by
4785 /// detecting that case here, returning a safe value instead.
4786 return Result == I ? UndefValue::get(I->getType()) : Result;
4789 /// \brief Implementation of recursive simplification through an instruction's
4792 /// This is the common implementation of the recursive simplification routines.
4793 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4794 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4795 /// instructions to process and attempt to simplify it using
4796 /// InstructionSimplify.
4798 /// This routine returns 'true' only when *it* simplifies something. The passed
4799 /// in simplified value does not count toward this.
4800 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4801 const TargetLibraryInfo *TLI,
4802 const DominatorTree *DT,
4803 AssumptionCache *AC) {
4804 bool Simplified = false;
4805 SmallSetVector<Instruction *, 8> Worklist;
4806 const DataLayout &DL = I->getModule()->getDataLayout();
4808 // If we have an explicit value to collapse to, do that round of the
4809 // simplification loop by hand initially.
4811 for (User *U : I->users())
4813 Worklist.insert(cast<Instruction>(U));
4815 // Replace the instruction with its simplified value.
4816 I->replaceAllUsesWith(SimpleV);
4818 // Gracefully handle edge cases where the instruction is not wired into any
4820 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4821 !I->mayHaveSideEffects())
4822 I->eraseFromParent();
4827 // Note that we must test the size on each iteration, the worklist can grow.
4828 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4831 // See if this instruction simplifies.
4832 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4838 // Stash away all the uses of the old instruction so we can check them for
4839 // recursive simplifications after a RAUW. This is cheaper than checking all
4840 // uses of To on the recursive step in most cases.
4841 for (User *U : I->users())
4842 Worklist.insert(cast<Instruction>(U));
4844 // Replace the instruction with its simplified value.
4845 I->replaceAllUsesWith(SimpleV);
4847 // Gracefully handle edge cases where the instruction is not wired into any
4849 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4850 !I->mayHaveSideEffects())
4851 I->eraseFromParent();
4856 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4857 const TargetLibraryInfo *TLI,
4858 const DominatorTree *DT,
4859 AssumptionCache *AC) {
4860 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4863 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4864 const TargetLibraryInfo *TLI,
4865 const DominatorTree *DT,
4866 AssumptionCache *AC) {
4867 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4868 assert(SimpleV && "Must provide a simplified value.");
4869 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4873 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4874 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4875 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4876 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4877 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4878 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4879 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4880 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4883 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4884 const DataLayout &DL) {
4885 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4888 template <class T, class... TArgs>
4889 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4891 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4892 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4893 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4894 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4896 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,