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.
982 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
983 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
984 // If the Mul does not overflow, then we are good to go.
985 if ((IsSigned && Mul->hasNoSignedWrap()) ||
986 (!IsSigned && Mul->hasNoUnsignedWrap()))
988 // If X has the form X = A / Y, then X * Y cannot overflow.
989 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
990 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
994 // (X rem Y) / Y -> 0
995 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
996 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
997 return Constant::getNullValue(Op0->getType());
999 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1000 ConstantInt *C1, *C2;
1001 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1002 match(Op1, m_ConstantInt(C2))) {
1004 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1006 return Constant::getNullValue(Op0->getType());
1009 // If the operation is with the result of a select instruction, check whether
1010 // operating on either branch of the select always yields the same value.
1011 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1012 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1015 // If the operation is with the result of a phi instruction, check whether
1016 // operating on all incoming values of the phi always yields the same value.
1017 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1018 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1021 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1022 return Constant::getNullValue(Op0->getType());
1027 /// These are simplifications common to SRem and URem.
1028 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1029 const SimplifyQuery &Q, unsigned MaxRecurse) {
1030 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1033 if (Value *V = simplifyDivRem(Op0, Op1, false))
1036 // (X % Y) % Y -> X % Y
1037 if ((Opcode == Instruction::SRem &&
1038 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1039 (Opcode == Instruction::URem &&
1040 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1043 // If the operation is with the result of a select instruction, check whether
1044 // operating on either branch of the select always yields the same value.
1045 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1046 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1049 // If the operation is with the result of a phi instruction, check whether
1050 // operating on all incoming values of the phi always yields the same value.
1051 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1052 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1055 // If X / Y == 0, then X % Y == X.
1056 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1062 /// Given operands for an SDiv, see if we can fold the result.
1063 /// If not, this returns null.
1064 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1065 unsigned MaxRecurse) {
1066 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1069 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1070 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1073 /// Given operands for a UDiv, see if we can fold the result.
1074 /// If not, this returns null.
1075 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1076 unsigned MaxRecurse) {
1077 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1080 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1081 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1084 /// Given operands for an SRem, see if we can fold the result.
1085 /// If not, this returns null.
1086 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1087 unsigned MaxRecurse) {
1088 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1091 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1092 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1095 /// Given operands for a URem, see if we can fold the result.
1096 /// If not, this returns null.
1097 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1098 unsigned MaxRecurse) {
1099 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1102 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1103 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1106 /// Returns true if a shift by \c Amount always yields undef.
1107 static bool isUndefShift(Value *Amount) {
1108 Constant *C = dyn_cast<Constant>(Amount);
1112 // X shift by undef -> undef because it may shift by the bitwidth.
1113 if (isa<UndefValue>(C))
1116 // Shifting by the bitwidth or more is undefined.
1117 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1118 if (CI->getValue().getLimitedValue() >=
1119 CI->getType()->getScalarSizeInBits())
1122 // If all lanes of a vector shift are undefined the whole shift is.
1123 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1124 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1125 if (!isUndefShift(C->getAggregateElement(I)))
1133 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1134 /// If not, this returns null.
1135 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1136 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1137 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1140 // 0 shift by X -> 0
1141 if (match(Op0, m_Zero()))
1144 // X shift by 0 -> X
1145 if (match(Op1, m_Zero()))
1148 // Fold undefined shifts.
1149 if (isUndefShift(Op1))
1150 return UndefValue::get(Op0->getType());
1152 // If the operation is with the result of a select instruction, check whether
1153 // operating on either branch of the select always yields the same value.
1154 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1155 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1158 // If the operation is with the result of a phi instruction, check whether
1159 // operating on all incoming values of the phi always yields the same value.
1160 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1161 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1164 // If any bits in the shift amount make that value greater than or equal to
1165 // the number of bits in the type, the shift is undefined.
1166 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1167 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1168 return UndefValue::get(Op0->getType());
1170 // If all valid bits in the shift amount are known zero, the first operand is
1172 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1173 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1179 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1180 /// fold the result. If not, this returns null.
1181 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1182 Value *Op1, bool isExact, const SimplifyQuery &Q,
1183 unsigned MaxRecurse) {
1184 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1189 return Constant::getNullValue(Op0->getType());
1192 // undef >> X -> undef (if it's exact)
1193 if (match(Op0, m_Undef()))
1194 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1196 // The low bit cannot be shifted out of an exact shift if it is set.
1198 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1199 if (Op0Known.One[0])
1206 /// Given operands for an Shl, see if we can fold the result.
1207 /// If not, this returns null.
1208 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1209 const SimplifyQuery &Q, unsigned MaxRecurse) {
1210 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1214 // undef << X -> undef if (if it's NSW/NUW)
1215 if (match(Op0, m_Undef()))
1216 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1218 // (X >> A) << A -> X
1220 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1225 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1226 const SimplifyQuery &Q) {
1227 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1230 /// Given operands for an LShr, see if we can fold the result.
1231 /// If not, this returns null.
1232 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1233 const SimplifyQuery &Q, unsigned MaxRecurse) {
1234 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1238 // (X << A) >> A -> X
1240 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1246 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1247 const SimplifyQuery &Q) {
1248 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1251 /// Given operands for an AShr, see if we can fold the result.
1252 /// If not, this returns null.
1253 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1254 const SimplifyQuery &Q, unsigned MaxRecurse) {
1255 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1259 // all ones >>a X -> all ones
1260 if (match(Op0, m_AllOnes()))
1263 // (X << A) >> A -> X
1265 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1268 // Arithmetic shifting an all-sign-bit value is a no-op.
1269 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1270 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1276 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1277 const SimplifyQuery &Q) {
1278 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1281 /// Commuted variants are assumed to be handled by calling this function again
1282 /// with the parameters swapped.
1283 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1284 ICmpInst *UnsignedICmp, bool IsAnd) {
1287 ICmpInst::Predicate EqPred;
1288 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1289 !ICmpInst::isEquality(EqPred))
1292 ICmpInst::Predicate UnsignedPred;
1293 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1294 ICmpInst::isUnsigned(UnsignedPred))
1296 else if (match(UnsignedICmp,
1297 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1298 ICmpInst::isUnsigned(UnsignedPred))
1299 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1303 // X < Y && Y != 0 --> X < Y
1304 // X < Y || Y != 0 --> Y != 0
1305 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1306 return IsAnd ? UnsignedICmp : ZeroICmp;
1308 // X >= Y || Y != 0 --> true
1309 // X >= Y || Y == 0 --> X >= Y
1310 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1311 if (EqPred == ICmpInst::ICMP_NE)
1312 return getTrue(UnsignedICmp->getType());
1313 return UnsignedICmp;
1316 // X < Y && Y == 0 --> false
1317 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1319 return getFalse(UnsignedICmp->getType());
1324 /// Commuted variants are assumed to be handled by calling this function again
1325 /// with the parameters swapped.
1326 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1327 ICmpInst::Predicate Pred0, Pred1;
1329 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1330 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1333 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1334 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1335 // can eliminate Op1 from this 'and'.
1336 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1339 // Check for any combination of predicates that are guaranteed to be disjoint.
1340 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1341 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1342 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1343 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1344 return getFalse(Op0->getType());
1349 /// Commuted variants are assumed to be handled by calling this function again
1350 /// with the parameters swapped.
1351 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1352 ICmpInst::Predicate Pred0, Pred1;
1354 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1355 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1358 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1359 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1360 // can eliminate Op0 from this 'or'.
1361 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1364 // Check for any combination of predicates that cover the entire range of
1366 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1367 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1368 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1369 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1370 return getTrue(Op0->getType());
1375 /// Test if a pair of compares with a shared operand and 2 constants has an
1376 /// empty set intersection, full set union, or if one compare is a superset of
1378 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1380 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1381 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1384 const APInt *C0, *C1;
1385 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1386 !match(Cmp1->getOperand(1), m_APInt(C1)))
1389 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1390 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1392 // For and-of-compares, check if the intersection is empty:
1393 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1394 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1395 return getFalse(Cmp0->getType());
1397 // For or-of-compares, check if the union is full:
1398 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1399 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1400 return getTrue(Cmp0->getType());
1402 // Is one range a superset of the other?
1403 // If this is and-of-compares, take the smaller set:
1404 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1405 // If this is or-of-compares, take the larger set:
1406 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1407 if (Range0.contains(Range1))
1408 return IsAnd ? Cmp1 : Cmp0;
1409 if (Range1.contains(Range0))
1410 return IsAnd ? Cmp0 : Cmp1;
1415 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1417 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1418 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1419 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1422 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1425 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1426 Value *X = Cmp0->getOperand(0);
1427 Value *Y = Cmp1->getOperand(0);
1429 // If one of the compares is a masked version of a (not) null check, then
1430 // that compare implies the other, so we eliminate the other. Optionally, look
1431 // through a pointer-to-int cast to match a null check of a pointer type.
1433 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1434 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1435 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1436 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1437 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1438 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1441 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1442 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1443 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1444 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1445 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1446 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1452 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1453 // (icmp (add V, C0), C1) & (icmp V, C0)
1454 ICmpInst::Predicate Pred0, Pred1;
1455 const APInt *C0, *C1;
1457 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1460 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1463 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1464 if (AddInst->getOperand(1) != Op1->getOperand(1))
1467 Type *ITy = Op0->getType();
1468 bool isNSW = AddInst->hasNoSignedWrap();
1469 bool isNUW = AddInst->hasNoUnsignedWrap();
1471 const APInt Delta = *C1 - *C0;
1472 if (C0->isStrictlyPositive()) {
1474 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1475 return getFalse(ITy);
1476 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1477 return getFalse(ITy);
1480 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1481 return getFalse(ITy);
1482 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1483 return getFalse(ITy);
1486 if (C0->getBoolValue() && isNUW) {
1488 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1489 return getFalse(ITy);
1491 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1492 return getFalse(ITy);
1498 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1499 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1501 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1504 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1506 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1509 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1512 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1515 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1517 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1523 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1524 // (icmp (add V, C0), C1) | (icmp V, C0)
1525 ICmpInst::Predicate Pred0, Pred1;
1526 const APInt *C0, *C1;
1528 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1531 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1534 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1535 if (AddInst->getOperand(1) != Op1->getOperand(1))
1538 Type *ITy = Op0->getType();
1539 bool isNSW = AddInst->hasNoSignedWrap();
1540 bool isNUW = AddInst->hasNoUnsignedWrap();
1542 const APInt Delta = *C1 - *C0;
1543 if (C0->isStrictlyPositive()) {
1545 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1546 return getTrue(ITy);
1547 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1548 return getTrue(ITy);
1551 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1552 return getTrue(ITy);
1553 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1554 return getTrue(ITy);
1557 if (C0->getBoolValue() && isNUW) {
1559 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1560 return getTrue(ITy);
1562 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1563 return getTrue(ITy);
1569 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1570 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1572 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1575 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1577 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1580 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1583 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1586 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1588 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1594 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1595 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1596 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1597 if (LHS0->getType() != RHS0->getType())
1600 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1601 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1602 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1603 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1604 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1605 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1606 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1607 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1608 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1609 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1610 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1611 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1612 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1615 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1616 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1617 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1618 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1619 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1620 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1621 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1622 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1623 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1624 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1631 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1632 // Look through casts of the 'and' operands to find compares.
1633 auto *Cast0 = dyn_cast<CastInst>(Op0);
1634 auto *Cast1 = dyn_cast<CastInst>(Op1);
1635 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1636 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1637 Op0 = Cast0->getOperand(0);
1638 Op1 = Cast1->getOperand(0);
1642 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1643 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1645 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1646 simplifyOrOfICmps(ICmp0, ICmp1);
1648 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1649 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1651 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1658 // If we looked through casts, we can only handle a constant simplification
1659 // because we are not allowed to create a cast instruction here.
1660 if (auto *C = dyn_cast<Constant>(V))
1661 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1666 /// Given operands for an And, see if we can fold the result.
1667 /// If not, this returns null.
1668 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1669 unsigned MaxRecurse) {
1670 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1674 if (match(Op1, m_Undef()))
1675 return Constant::getNullValue(Op0->getType());
1682 if (match(Op1, m_Zero()))
1686 if (match(Op1, m_AllOnes()))
1689 // A & ~A = ~A & A = 0
1690 if (match(Op0, m_Not(m_Specific(Op1))) ||
1691 match(Op1, m_Not(m_Specific(Op0))))
1692 return Constant::getNullValue(Op0->getType());
1695 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1699 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1702 // A mask that only clears known zeros of a shifted value is a no-op.
1706 if (match(Op1, m_APInt(Mask))) {
1707 // If all bits in the inverted and shifted mask are clear:
1708 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1709 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1710 (~(*Mask)).lshr(*ShAmt).isNullValue())
1713 // If all bits in the inverted and shifted mask are clear:
1714 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1715 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1716 (~(*Mask)).shl(*ShAmt).isNullValue())
1720 // A & (-A) = A if A is a power of two or zero.
1721 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1722 match(Op1, m_Neg(m_Specific(Op0)))) {
1723 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1726 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1731 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1734 // Try some generic simplifications for associative operations.
1735 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1739 // And distributes over Or. Try some generic simplifications based on this.
1740 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1744 // And distributes over Xor. Try some generic simplifications based on this.
1745 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1749 // If the operation is with the result of a select instruction, check whether
1750 // operating on either branch of the select always yields the same value.
1751 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1752 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1756 // If the operation is with the result of a phi instruction, check whether
1757 // operating on all incoming values of the phi always yields the same value.
1758 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1759 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1766 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1767 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1770 /// Given operands for an Or, see if we can fold the result.
1771 /// If not, this returns null.
1772 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1773 unsigned MaxRecurse) {
1774 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1778 if (match(Op1, m_Undef()))
1779 return Constant::getAllOnesValue(Op0->getType());
1786 if (match(Op1, m_Zero()))
1790 if (match(Op1, m_AllOnes()))
1793 // A | ~A = ~A | A = -1
1794 if (match(Op0, m_Not(m_Specific(Op1))) ||
1795 match(Op1, m_Not(m_Specific(Op0))))
1796 return Constant::getAllOnesValue(Op0->getType());
1799 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1803 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1806 // ~(A & ?) | A = -1
1807 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1808 return Constant::getAllOnesValue(Op1->getType());
1810 // A | ~(A & ?) = -1
1811 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1812 return Constant::getAllOnesValue(Op0->getType());
1815 // (A & ~B) | (A ^ B) -> (A ^ B)
1816 // (~B & A) | (A ^ B) -> (A ^ B)
1817 // (A & ~B) | (B ^ A) -> (B ^ A)
1818 // (~B & A) | (B ^ A) -> (B ^ A)
1819 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1820 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1821 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1824 // Commute the 'or' operands.
1825 // (A ^ B) | (A & ~B) -> (A ^ B)
1826 // (A ^ B) | (~B & A) -> (A ^ B)
1827 // (B ^ A) | (A & ~B) -> (B ^ A)
1828 // (B ^ A) | (~B & A) -> (B ^ A)
1829 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1830 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1831 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1834 // (A & B) | (~A ^ B) -> (~A ^ B)
1835 // (B & A) | (~A ^ B) -> (~A ^ B)
1836 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1837 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1838 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1839 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1840 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1843 // (~A ^ B) | (A & B) -> (~A ^ B)
1844 // (~A ^ B) | (B & A) -> (~A ^ B)
1845 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1846 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1847 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1848 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1849 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1852 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1855 // Try some generic simplifications for associative operations.
1856 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1860 // Or distributes over And. Try some generic simplifications based on this.
1861 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1865 // If the operation is with the result of a select instruction, check whether
1866 // operating on either branch of the select always yields the same value.
1867 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1868 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1872 // (A & C1)|(B & C2)
1873 const APInt *C1, *C2;
1874 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1875 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1877 // (A & C1)|(B & C2)
1878 // If we have: ((V + N) & C1) | (V & C2)
1879 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1880 // replace with V+N.
1882 if (C2->isMask() && // C2 == 0+1+
1883 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1884 // Add commutes, try both ways.
1885 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1888 // Or commutes, try both ways.
1890 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1891 // Add commutes, try both ways.
1892 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1898 // If the operation is with the result of a phi instruction, check whether
1899 // operating on all incoming values of the phi always yields the same value.
1900 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1901 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1907 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1908 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1911 /// Given operands for a Xor, see if we can fold the result.
1912 /// If not, this returns null.
1913 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1914 unsigned MaxRecurse) {
1915 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1918 // A ^ undef -> undef
1919 if (match(Op1, m_Undef()))
1923 if (match(Op1, m_Zero()))
1928 return Constant::getNullValue(Op0->getType());
1930 // A ^ ~A = ~A ^ A = -1
1931 if (match(Op0, m_Not(m_Specific(Op1))) ||
1932 match(Op1, m_Not(m_Specific(Op0))))
1933 return Constant::getAllOnesValue(Op0->getType());
1935 // Try some generic simplifications for associative operations.
1936 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1940 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1941 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1942 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1943 // only if B and C are equal. If B and C are equal then (since we assume
1944 // that operands have already been simplified) "select(cond, B, C)" should
1945 // have been simplified to the common value of B and C already. Analysing
1946 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1947 // for threading over phi nodes.
1952 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1953 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1957 static Type *GetCompareTy(Value *Op) {
1958 return CmpInst::makeCmpResultType(Op->getType());
1961 /// Rummage around inside V looking for something equivalent to the comparison
1962 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1963 /// Helper function for analyzing max/min idioms.
1964 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1965 Value *LHS, Value *RHS) {
1966 SelectInst *SI = dyn_cast<SelectInst>(V);
1969 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1972 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1973 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1975 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1976 LHS == CmpRHS && RHS == CmpLHS)
1981 // A significant optimization not implemented here is assuming that alloca
1982 // addresses are not equal to incoming argument values. They don't *alias*,
1983 // as we say, but that doesn't mean they aren't equal, so we take a
1984 // conservative approach.
1986 // This is inspired in part by C++11 5.10p1:
1987 // "Two pointers of the same type compare equal if and only if they are both
1988 // null, both point to the same function, or both represent the same
1991 // This is pretty permissive.
1993 // It's also partly due to C11 6.5.9p6:
1994 // "Two pointers compare equal if and only if both are null pointers, both are
1995 // pointers to the same object (including a pointer to an object and a
1996 // subobject at its beginning) or function, both are pointers to one past the
1997 // last element of the same array object, or one is a pointer to one past the
1998 // end of one array object and the other is a pointer to the start of a
1999 // different array object that happens to immediately follow the first array
2000 // object in the address space.)
2002 // C11's version is more restrictive, however there's no reason why an argument
2003 // couldn't be a one-past-the-end value for a stack object in the caller and be
2004 // equal to the beginning of a stack object in the callee.
2006 // If the C and C++ standards are ever made sufficiently restrictive in this
2007 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2008 // this optimization.
2010 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2011 const DominatorTree *DT, CmpInst::Predicate Pred,
2012 AssumptionCache *AC, const Instruction *CxtI,
2013 Value *LHS, Value *RHS) {
2014 // First, skip past any trivial no-ops.
2015 LHS = LHS->stripPointerCasts();
2016 RHS = RHS->stripPointerCasts();
2018 // A non-null pointer is not equal to a null pointer.
2019 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2020 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2021 return ConstantInt::get(GetCompareTy(LHS),
2022 !CmpInst::isTrueWhenEqual(Pred));
2024 // We can only fold certain predicates on pointer comparisons.
2029 // Equality comaprisons are easy to fold.
2030 case CmpInst::ICMP_EQ:
2031 case CmpInst::ICMP_NE:
2034 // We can only handle unsigned relational comparisons because 'inbounds' on
2035 // a GEP only protects against unsigned wrapping.
2036 case CmpInst::ICMP_UGT:
2037 case CmpInst::ICMP_UGE:
2038 case CmpInst::ICMP_ULT:
2039 case CmpInst::ICMP_ULE:
2040 // However, we have to switch them to their signed variants to handle
2041 // negative indices from the base pointer.
2042 Pred = ICmpInst::getSignedPredicate(Pred);
2046 // Strip off any constant offsets so that we can reason about them.
2047 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2048 // here and compare base addresses like AliasAnalysis does, however there are
2049 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2050 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2051 // doesn't need to guarantee pointer inequality when it says NoAlias.
2052 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2053 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2055 // If LHS and RHS are related via constant offsets to the same base
2056 // value, we can replace it with an icmp which just compares the offsets.
2058 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2060 // Various optimizations for (in)equality comparisons.
2061 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2062 // Different non-empty allocations that exist at the same time have
2063 // different addresses (if the program can tell). Global variables always
2064 // exist, so they always exist during the lifetime of each other and all
2065 // allocas. Two different allocas usually have different addresses...
2067 // However, if there's an @llvm.stackrestore dynamically in between two
2068 // allocas, they may have the same address. It's tempting to reduce the
2069 // scope of the problem by only looking at *static* allocas here. That would
2070 // cover the majority of allocas while significantly reducing the likelihood
2071 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2072 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2073 // an entry block. Also, if we have a block that's not attached to a
2074 // function, we can't tell if it's "static" under the current definition.
2075 // Theoretically, this problem could be fixed by creating a new kind of
2076 // instruction kind specifically for static allocas. Such a new instruction
2077 // could be required to be at the top of the entry block, thus preventing it
2078 // from being subject to a @llvm.stackrestore. Instcombine could even
2079 // convert regular allocas into these special allocas. It'd be nifty.
2080 // However, until then, this problem remains open.
2082 // So, we'll assume that two non-empty allocas have different addresses
2085 // With all that, if the offsets are within the bounds of their allocations
2086 // (and not one-past-the-end! so we can't use inbounds!), and their
2087 // allocations aren't the same, the pointers are not equal.
2089 // Note that it's not necessary to check for LHS being a global variable
2090 // address, due to canonicalization and constant folding.
2091 if (isa<AllocaInst>(LHS) &&
2092 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2093 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2094 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2095 uint64_t LHSSize, RHSSize;
2096 if (LHSOffsetCI && RHSOffsetCI &&
2097 getObjectSize(LHS, LHSSize, DL, TLI) &&
2098 getObjectSize(RHS, RHSSize, DL, TLI)) {
2099 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2100 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2101 if (!LHSOffsetValue.isNegative() &&
2102 !RHSOffsetValue.isNegative() &&
2103 LHSOffsetValue.ult(LHSSize) &&
2104 RHSOffsetValue.ult(RHSSize)) {
2105 return ConstantInt::get(GetCompareTy(LHS),
2106 !CmpInst::isTrueWhenEqual(Pred));
2110 // Repeat the above check but this time without depending on DataLayout
2111 // or being able to compute a precise size.
2112 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2113 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2114 LHSOffset->isNullValue() &&
2115 RHSOffset->isNullValue())
2116 return ConstantInt::get(GetCompareTy(LHS),
2117 !CmpInst::isTrueWhenEqual(Pred));
2120 // Even if an non-inbounds GEP occurs along the path we can still optimize
2121 // equality comparisons concerning the result. We avoid walking the whole
2122 // chain again by starting where the last calls to
2123 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2124 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2125 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2127 return ConstantExpr::getICmp(Pred,
2128 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2129 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2131 // If one side of the equality comparison must come from a noalias call
2132 // (meaning a system memory allocation function), and the other side must
2133 // come from a pointer that cannot overlap with dynamically-allocated
2134 // memory within the lifetime of the current function (allocas, byval
2135 // arguments, globals), then determine the comparison result here.
2136 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2137 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2138 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2140 // Is the set of underlying objects all noalias calls?
2141 auto IsNAC = [](ArrayRef<Value *> Objects) {
2142 return all_of(Objects, isNoAliasCall);
2145 // Is the set of underlying objects all things which must be disjoint from
2146 // noalias calls. For allocas, we consider only static ones (dynamic
2147 // allocas might be transformed into calls to malloc not simultaneously
2148 // live with the compared-to allocation). For globals, we exclude symbols
2149 // that might be resolve lazily to symbols in another dynamically-loaded
2150 // library (and, thus, could be malloc'ed by the implementation).
2151 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2152 return all_of(Objects, [](Value *V) {
2153 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2154 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2155 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2156 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2157 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2158 !GV->isThreadLocal();
2159 if (const Argument *A = dyn_cast<Argument>(V))
2160 return A->hasByValAttr();
2165 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2166 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2167 return ConstantInt::get(GetCompareTy(LHS),
2168 !CmpInst::isTrueWhenEqual(Pred));
2170 // Fold comparisons for non-escaping pointer even if the allocation call
2171 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2172 // dynamic allocation call could be either of the operands.
2173 Value *MI = nullptr;
2174 if (isAllocLikeFn(LHS, TLI) &&
2175 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2177 else if (isAllocLikeFn(RHS, TLI) &&
2178 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2180 // FIXME: We should also fold the compare when the pointer escapes, but the
2181 // compare dominates the pointer escape
2182 if (MI && !PointerMayBeCaptured(MI, true, true))
2183 return ConstantInt::get(GetCompareTy(LHS),
2184 CmpInst::isFalseWhenEqual(Pred));
2191 /// Fold an icmp when its operands have i1 scalar type.
2192 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2193 Value *RHS, const SimplifyQuery &Q) {
2194 Type *ITy = GetCompareTy(LHS); // The return type.
2195 Type *OpTy = LHS->getType(); // The operand type.
2196 if (!OpTy->isIntOrIntVectorTy(1))
2199 // A boolean compared to true/false can be simplified in 14 out of the 20
2200 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2201 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2202 if (match(RHS, m_Zero())) {
2204 case CmpInst::ICMP_NE: // X != 0 -> X
2205 case CmpInst::ICMP_UGT: // X >u 0 -> X
2206 case CmpInst::ICMP_SLT: // X <s 0 -> X
2209 case CmpInst::ICMP_ULT: // X <u 0 -> false
2210 case CmpInst::ICMP_SGT: // X >s 0 -> false
2211 return getFalse(ITy);
2213 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2214 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2215 return getTrue(ITy);
2219 } else if (match(RHS, m_One())) {
2221 case CmpInst::ICMP_EQ: // X == 1 -> X
2222 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2223 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2226 case CmpInst::ICMP_UGT: // X >u 1 -> false
2227 case CmpInst::ICMP_SLT: // X <s -1 -> false
2228 return getFalse(ITy);
2230 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2231 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2232 return getTrue(ITy);
2241 case ICmpInst::ICMP_UGE:
2242 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2243 return getTrue(ITy);
2245 case ICmpInst::ICMP_SGE:
2246 /// For signed comparison, the values for an i1 are 0 and -1
2247 /// respectively. This maps into a truth table of:
2248 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2249 /// 0 | 0 | 1 (0 >= 0) | 1
2250 /// 0 | 1 | 1 (0 >= -1) | 1
2251 /// 1 | 0 | 0 (-1 >= 0) | 0
2252 /// 1 | 1 | 1 (-1 >= -1) | 1
2253 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2254 return getTrue(ITy);
2256 case ICmpInst::ICMP_ULE:
2257 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2258 return getTrue(ITy);
2265 /// Try hard to fold icmp with zero RHS because this is a common case.
2266 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2267 Value *RHS, const SimplifyQuery &Q) {
2268 if (!match(RHS, m_Zero()))
2271 Type *ITy = GetCompareTy(LHS); // The return type.
2274 llvm_unreachable("Unknown ICmp predicate!");
2275 case ICmpInst::ICMP_ULT:
2276 return getFalse(ITy);
2277 case ICmpInst::ICMP_UGE:
2278 return getTrue(ITy);
2279 case ICmpInst::ICMP_EQ:
2280 case ICmpInst::ICMP_ULE:
2281 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2282 return getFalse(ITy);
2284 case ICmpInst::ICMP_NE:
2285 case ICmpInst::ICMP_UGT:
2286 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2287 return getTrue(ITy);
2289 case ICmpInst::ICMP_SLT: {
2290 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2291 if (LHSKnown.isNegative())
2292 return getTrue(ITy);
2293 if (LHSKnown.isNonNegative())
2294 return getFalse(ITy);
2297 case ICmpInst::ICMP_SLE: {
2298 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2299 if (LHSKnown.isNegative())
2300 return getTrue(ITy);
2301 if (LHSKnown.isNonNegative() &&
2302 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2303 return getFalse(ITy);
2306 case ICmpInst::ICMP_SGE: {
2307 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2308 if (LHSKnown.isNegative())
2309 return getFalse(ITy);
2310 if (LHSKnown.isNonNegative())
2311 return getTrue(ITy);
2314 case ICmpInst::ICMP_SGT: {
2315 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2316 if (LHSKnown.isNegative())
2317 return getFalse(ITy);
2318 if (LHSKnown.isNonNegative() &&
2319 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2320 return getTrue(ITy);
2328 /// Many binary operators with a constant operand have an easy-to-compute
2329 /// range of outputs. This can be used to fold a comparison to always true or
2331 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2332 unsigned Width = Lower.getBitWidth();
2334 switch (BO.getOpcode()) {
2335 case Instruction::Add:
2336 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2337 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2338 if (BO.hasNoUnsignedWrap()) {
2339 // 'add nuw x, C' produces [C, UINT_MAX].
2341 } else if (BO.hasNoSignedWrap()) {
2342 if (C->isNegative()) {
2343 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2344 Lower = APInt::getSignedMinValue(Width);
2345 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2347 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2348 Lower = APInt::getSignedMinValue(Width) + *C;
2349 Upper = APInt::getSignedMaxValue(Width) + 1;
2355 case Instruction::And:
2356 if (match(BO.getOperand(1), m_APInt(C)))
2357 // 'and x, C' produces [0, C].
2361 case Instruction::Or:
2362 if (match(BO.getOperand(1), m_APInt(C)))
2363 // 'or x, C' produces [C, UINT_MAX].
2367 case Instruction::AShr:
2368 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2369 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2370 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2371 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2372 } else if (match(BO.getOperand(0), m_APInt(C))) {
2373 unsigned ShiftAmount = Width - 1;
2374 if (!C->isNullValue() && BO.isExact())
2375 ShiftAmount = C->countTrailingZeros();
2376 if (C->isNegative()) {
2377 // 'ashr C, x' produces [C, C >> (Width-1)]
2379 Upper = C->ashr(ShiftAmount) + 1;
2381 // 'ashr C, x' produces [C >> (Width-1), C]
2382 Lower = C->ashr(ShiftAmount);
2388 case Instruction::LShr:
2389 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2390 // 'lshr x, C' produces [0, UINT_MAX >> C].
2391 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2392 } else if (match(BO.getOperand(0), m_APInt(C))) {
2393 // 'lshr C, x' produces [C >> (Width-1), C].
2394 unsigned ShiftAmount = Width - 1;
2395 if (!C->isNullValue() && BO.isExact())
2396 ShiftAmount = C->countTrailingZeros();
2397 Lower = C->lshr(ShiftAmount);
2402 case Instruction::Shl:
2403 if (match(BO.getOperand(0), m_APInt(C))) {
2404 if (BO.hasNoUnsignedWrap()) {
2405 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2407 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2408 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2409 if (C->isNegative()) {
2410 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2411 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2412 Lower = C->shl(ShiftAmount);
2415 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2416 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2418 Upper = C->shl(ShiftAmount) + 1;
2424 case Instruction::SDiv:
2425 if (match(BO.getOperand(1), m_APInt(C))) {
2426 APInt IntMin = APInt::getSignedMinValue(Width);
2427 APInt IntMax = APInt::getSignedMaxValue(Width);
2428 if (C->isAllOnesValue()) {
2429 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2430 // where C != -1 and C != 0 and C != 1
2433 } else if (C->countLeadingZeros() < Width - 1) {
2434 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2435 // where C != -1 and C != 0 and C != 1
2436 Lower = IntMin.sdiv(*C);
2437 Upper = IntMax.sdiv(*C);
2438 if (Lower.sgt(Upper))
2439 std::swap(Lower, Upper);
2441 assert(Upper != Lower && "Upper part of range has wrapped!");
2443 } else if (match(BO.getOperand(0), m_APInt(C))) {
2444 if (C->isMinSignedValue()) {
2445 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2447 Upper = Lower.lshr(1) + 1;
2449 // 'sdiv C, x' produces [-|C|, |C|].
2450 Upper = C->abs() + 1;
2451 Lower = (-Upper) + 1;
2456 case Instruction::UDiv:
2457 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2458 // 'udiv x, C' produces [0, UINT_MAX / C].
2459 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2460 } else if (match(BO.getOperand(0), m_APInt(C))) {
2461 // 'udiv C, x' produces [0, C].
2466 case Instruction::SRem:
2467 if (match(BO.getOperand(1), m_APInt(C))) {
2468 // 'srem x, C' produces (-|C|, |C|).
2470 Lower = (-Upper) + 1;
2474 case Instruction::URem:
2475 if (match(BO.getOperand(1), m_APInt(C)))
2476 // 'urem x, C' produces [0, C).
2485 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2488 if (!match(RHS, m_APInt(C)))
2491 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2492 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2493 if (RHS_CR.isEmptySet())
2494 return ConstantInt::getFalse(GetCompareTy(RHS));
2495 if (RHS_CR.isFullSet())
2496 return ConstantInt::getTrue(GetCompareTy(RHS));
2498 // Find the range of possible values for binary operators.
2499 unsigned Width = C->getBitWidth();
2500 APInt Lower = APInt(Width, 0);
2501 APInt Upper = APInt(Width, 0);
2502 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2503 setLimitsForBinOp(*BO, Lower, Upper);
2505 ConstantRange LHS_CR =
2506 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2508 if (auto *I = dyn_cast<Instruction>(LHS))
2509 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2510 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2512 if (!LHS_CR.isFullSet()) {
2513 if (RHS_CR.contains(LHS_CR))
2514 return ConstantInt::getTrue(GetCompareTy(RHS));
2515 if (RHS_CR.inverse().contains(LHS_CR))
2516 return ConstantInt::getFalse(GetCompareTy(RHS));
2522 /// TODO: A large part of this logic is duplicated in InstCombine's
2523 /// foldICmpBinOp(). We should be able to share that and avoid the code
2525 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2526 Value *RHS, const SimplifyQuery &Q,
2527 unsigned MaxRecurse) {
2528 Type *ITy = GetCompareTy(LHS); // The return type.
2530 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2531 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2532 if (MaxRecurse && (LBO || RBO)) {
2533 // Analyze the case when either LHS or RHS is an add instruction.
2534 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2535 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2536 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2537 if (LBO && LBO->getOpcode() == Instruction::Add) {
2538 A = LBO->getOperand(0);
2539 B = LBO->getOperand(1);
2541 ICmpInst::isEquality(Pred) ||
2542 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2543 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2545 if (RBO && RBO->getOpcode() == Instruction::Add) {
2546 C = RBO->getOperand(0);
2547 D = RBO->getOperand(1);
2549 ICmpInst::isEquality(Pred) ||
2550 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2551 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2554 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2555 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2556 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2557 Constant::getNullValue(RHS->getType()), Q,
2561 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2562 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2564 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2565 C == LHS ? D : C, Q, MaxRecurse - 1))
2568 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2569 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2571 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2574 // C + B == C + D -> B == D
2577 } else if (A == D) {
2578 // D + B == C + D -> B == C
2581 } else if (B == C) {
2582 // A + C == C + D -> A == D
2587 // A + D == C + D -> A == C
2591 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2598 // icmp pred (or X, Y), X
2599 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2600 if (Pred == ICmpInst::ICMP_ULT)
2601 return getFalse(ITy);
2602 if (Pred == ICmpInst::ICMP_UGE)
2603 return getTrue(ITy);
2605 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2606 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2607 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2608 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2609 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2610 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2611 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2614 // icmp pred X, (or X, Y)
2615 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2616 if (Pred == ICmpInst::ICMP_ULE)
2617 return getTrue(ITy);
2618 if (Pred == ICmpInst::ICMP_UGT)
2619 return getFalse(ITy);
2621 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2622 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2623 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2624 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2625 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2626 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2627 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2632 // icmp pred (and X, Y), X
2633 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2634 if (Pred == ICmpInst::ICMP_UGT)
2635 return getFalse(ITy);
2636 if (Pred == ICmpInst::ICMP_ULE)
2637 return getTrue(ITy);
2639 // icmp pred X, (and X, Y)
2640 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2641 if (Pred == ICmpInst::ICMP_UGE)
2642 return getTrue(ITy);
2643 if (Pred == ICmpInst::ICMP_ULT)
2644 return getFalse(ITy);
2647 // 0 - (zext X) pred C
2648 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2649 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2650 if (RHSC->getValue().isStrictlyPositive()) {
2651 if (Pred == ICmpInst::ICMP_SLT)
2652 return ConstantInt::getTrue(RHSC->getContext());
2653 if (Pred == ICmpInst::ICMP_SGE)
2654 return ConstantInt::getFalse(RHSC->getContext());
2655 if (Pred == ICmpInst::ICMP_EQ)
2656 return ConstantInt::getFalse(RHSC->getContext());
2657 if (Pred == ICmpInst::ICMP_NE)
2658 return ConstantInt::getTrue(RHSC->getContext());
2660 if (RHSC->getValue().isNonNegative()) {
2661 if (Pred == ICmpInst::ICMP_SLE)
2662 return ConstantInt::getTrue(RHSC->getContext());
2663 if (Pred == ICmpInst::ICMP_SGT)
2664 return ConstantInt::getFalse(RHSC->getContext());
2669 // icmp pred (urem X, Y), Y
2670 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2674 case ICmpInst::ICMP_SGT:
2675 case ICmpInst::ICMP_SGE: {
2676 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2677 if (!Known.isNonNegative())
2681 case ICmpInst::ICMP_EQ:
2682 case ICmpInst::ICMP_UGT:
2683 case ICmpInst::ICMP_UGE:
2684 return getFalse(ITy);
2685 case ICmpInst::ICMP_SLT:
2686 case ICmpInst::ICMP_SLE: {
2687 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2688 if (!Known.isNonNegative())
2692 case ICmpInst::ICMP_NE:
2693 case ICmpInst::ICMP_ULT:
2694 case ICmpInst::ICMP_ULE:
2695 return getTrue(ITy);
2699 // icmp pred X, (urem Y, X)
2700 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2704 case ICmpInst::ICMP_SGT:
2705 case ICmpInst::ICMP_SGE: {
2706 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2707 if (!Known.isNonNegative())
2711 case ICmpInst::ICMP_NE:
2712 case ICmpInst::ICMP_UGT:
2713 case ICmpInst::ICMP_UGE:
2714 return getTrue(ITy);
2715 case ICmpInst::ICMP_SLT:
2716 case ICmpInst::ICMP_SLE: {
2717 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2718 if (!Known.isNonNegative())
2722 case ICmpInst::ICMP_EQ:
2723 case ICmpInst::ICMP_ULT:
2724 case ICmpInst::ICMP_ULE:
2725 return getFalse(ITy);
2731 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2732 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2733 // icmp pred (X op Y), X
2734 if (Pred == ICmpInst::ICMP_UGT)
2735 return getFalse(ITy);
2736 if (Pred == ICmpInst::ICMP_ULE)
2737 return getTrue(ITy);
2742 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2743 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2744 // icmp pred X, (X op Y)
2745 if (Pred == ICmpInst::ICMP_ULT)
2746 return getFalse(ITy);
2747 if (Pred == ICmpInst::ICMP_UGE)
2748 return getTrue(ITy);
2755 // where CI2 is a power of 2 and CI isn't
2756 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2757 const APInt *CI2Val, *CIVal = &CI->getValue();
2758 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2759 CI2Val->isPowerOf2()) {
2760 if (!CIVal->isPowerOf2()) {
2761 // CI2 << X can equal zero in some circumstances,
2762 // this simplification is unsafe if CI is zero.
2764 // We know it is safe if:
2765 // - The shift is nsw, we can't shift out the one bit.
2766 // - The shift is nuw, we can't shift out the one bit.
2769 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2770 CI2Val->isOneValue() || !CI->isZero()) {
2771 if (Pred == ICmpInst::ICMP_EQ)
2772 return ConstantInt::getFalse(RHS->getContext());
2773 if (Pred == ICmpInst::ICMP_NE)
2774 return ConstantInt::getTrue(RHS->getContext());
2777 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2778 if (Pred == ICmpInst::ICMP_UGT)
2779 return ConstantInt::getFalse(RHS->getContext());
2780 if (Pred == ICmpInst::ICMP_ULE)
2781 return ConstantInt::getTrue(RHS->getContext());
2786 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2787 LBO->getOperand(1) == RBO->getOperand(1)) {
2788 switch (LBO->getOpcode()) {
2791 case Instruction::UDiv:
2792 case Instruction::LShr:
2793 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2795 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2796 RBO->getOperand(0), Q, MaxRecurse - 1))
2799 case Instruction::SDiv:
2800 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2802 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2803 RBO->getOperand(0), Q, MaxRecurse - 1))
2806 case Instruction::AShr:
2807 if (!LBO->isExact() || !RBO->isExact())
2809 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2810 RBO->getOperand(0), Q, MaxRecurse - 1))
2813 case Instruction::Shl: {
2814 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2815 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2818 if (!NSW && ICmpInst::isSigned(Pred))
2820 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2821 RBO->getOperand(0), Q, MaxRecurse - 1))
2830 /// Simplify integer comparisons where at least one operand of the compare
2831 /// matches an integer min/max idiom.
2832 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2833 Value *RHS, const SimplifyQuery &Q,
2834 unsigned MaxRecurse) {
2835 Type *ITy = GetCompareTy(LHS); // The return type.
2837 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2838 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2840 // Signed variants on "max(a,b)>=a -> true".
2841 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2843 std::swap(A, B); // smax(A, B) pred A.
2844 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2845 // We analyze this as smax(A, B) pred A.
2847 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2848 (A == LHS || B == LHS)) {
2850 std::swap(A, B); // A pred smax(A, B).
2851 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2852 // We analyze this as smax(A, B) swapped-pred A.
2853 P = CmpInst::getSwappedPredicate(Pred);
2854 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2855 (A == RHS || B == RHS)) {
2857 std::swap(A, B); // smin(A, B) pred A.
2858 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2859 // We analyze this as smax(-A, -B) swapped-pred -A.
2860 // Note that we do not need to actually form -A or -B thanks to EqP.
2861 P = CmpInst::getSwappedPredicate(Pred);
2862 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2863 (A == LHS || B == LHS)) {
2865 std::swap(A, B); // A pred smin(A, B).
2866 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2867 // We analyze this as smax(-A, -B) pred -A.
2868 // Note that we do not need to actually form -A or -B thanks to EqP.
2871 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2872 // Cases correspond to "max(A, B) p A".
2876 case CmpInst::ICMP_EQ:
2877 case CmpInst::ICMP_SLE:
2878 // Equivalent to "A EqP B". This may be the same as the condition tested
2879 // in the max/min; if so, we can just return that.
2880 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2882 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2884 // Otherwise, see if "A EqP B" simplifies.
2886 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2889 case CmpInst::ICMP_NE:
2890 case CmpInst::ICMP_SGT: {
2891 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2892 // Equivalent to "A InvEqP B". This may be the same as the condition
2893 // tested in the max/min; if so, we can just return that.
2894 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2896 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2898 // Otherwise, see if "A InvEqP B" simplifies.
2900 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2904 case CmpInst::ICMP_SGE:
2906 return getTrue(ITy);
2907 case CmpInst::ICMP_SLT:
2909 return getFalse(ITy);
2913 // Unsigned variants on "max(a,b)>=a -> true".
2914 P = CmpInst::BAD_ICMP_PREDICATE;
2915 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2917 std::swap(A, B); // umax(A, B) pred A.
2918 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2919 // We analyze this as umax(A, B) pred A.
2921 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2922 (A == LHS || B == LHS)) {
2924 std::swap(A, B); // A pred umax(A, B).
2925 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2926 // We analyze this as umax(A, B) swapped-pred A.
2927 P = CmpInst::getSwappedPredicate(Pred);
2928 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2929 (A == RHS || B == RHS)) {
2931 std::swap(A, B); // umin(A, B) pred A.
2932 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2933 // We analyze this as umax(-A, -B) swapped-pred -A.
2934 // Note that we do not need to actually form -A or -B thanks to EqP.
2935 P = CmpInst::getSwappedPredicate(Pred);
2936 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2937 (A == LHS || B == LHS)) {
2939 std::swap(A, B); // A pred umin(A, B).
2940 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2941 // We analyze this as umax(-A, -B) pred -A.
2942 // Note that we do not need to actually form -A or -B thanks to EqP.
2945 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2946 // Cases correspond to "max(A, B) p A".
2950 case CmpInst::ICMP_EQ:
2951 case CmpInst::ICMP_ULE:
2952 // Equivalent to "A EqP B". This may be the same as the condition tested
2953 // in the max/min; if so, we can just return that.
2954 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2956 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2958 // Otherwise, see if "A EqP B" simplifies.
2960 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2963 case CmpInst::ICMP_NE:
2964 case CmpInst::ICMP_UGT: {
2965 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2966 // Equivalent to "A InvEqP B". This may be the same as the condition
2967 // tested in the max/min; if so, we can just return that.
2968 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2970 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2972 // Otherwise, see if "A InvEqP B" simplifies.
2974 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2978 case CmpInst::ICMP_UGE:
2980 return getTrue(ITy);
2981 case CmpInst::ICMP_ULT:
2983 return getFalse(ITy);
2987 // Variants on "max(x,y) >= min(x,z)".
2989 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2990 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2991 (A == C || A == D || B == C || B == D)) {
2992 // max(x, ?) pred min(x, ?).
2993 if (Pred == CmpInst::ICMP_SGE)
2995 return getTrue(ITy);
2996 if (Pred == CmpInst::ICMP_SLT)
2998 return getFalse(ITy);
2999 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3000 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3001 (A == C || A == D || B == C || B == D)) {
3002 // min(x, ?) pred max(x, ?).
3003 if (Pred == CmpInst::ICMP_SLE)
3005 return getTrue(ITy);
3006 if (Pred == CmpInst::ICMP_SGT)
3008 return getFalse(ITy);
3009 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3010 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3011 (A == C || A == D || B == C || B == D)) {
3012 // max(x, ?) pred min(x, ?).
3013 if (Pred == CmpInst::ICMP_UGE)
3015 return getTrue(ITy);
3016 if (Pred == CmpInst::ICMP_ULT)
3018 return getFalse(ITy);
3019 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3020 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3021 (A == C || A == D || B == C || B == D)) {
3022 // min(x, ?) pred max(x, ?).
3023 if (Pred == CmpInst::ICMP_ULE)
3025 return getTrue(ITy);
3026 if (Pred == CmpInst::ICMP_UGT)
3028 return getFalse(ITy);
3034 /// Given operands for an ICmpInst, see if we can fold the result.
3035 /// If not, this returns null.
3036 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3037 const SimplifyQuery &Q, unsigned MaxRecurse) {
3038 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3039 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3041 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3042 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3043 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3045 // If we have a constant, make sure it is on the RHS.
3046 std::swap(LHS, RHS);
3047 Pred = CmpInst::getSwappedPredicate(Pred);
3050 Type *ITy = GetCompareTy(LHS); // The return type.
3052 // icmp X, X -> true/false
3053 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3054 // because X could be 0.
3055 if (LHS == RHS || isa<UndefValue>(RHS))
3056 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3058 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3061 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3064 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3067 // If both operands have range metadata, use the metadata
3068 // to simplify the comparison.
3069 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3070 auto RHS_Instr = cast<Instruction>(RHS);
3071 auto LHS_Instr = cast<Instruction>(LHS);
3073 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3074 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3075 auto RHS_CR = getConstantRangeFromMetadata(
3076 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3077 auto LHS_CR = getConstantRangeFromMetadata(
3078 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3080 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3081 if (Satisfied_CR.contains(LHS_CR))
3082 return ConstantInt::getTrue(RHS->getContext());
3084 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3085 CmpInst::getInversePredicate(Pred), RHS_CR);
3086 if (InversedSatisfied_CR.contains(LHS_CR))
3087 return ConstantInt::getFalse(RHS->getContext());
3091 // Compare of cast, for example (zext X) != 0 -> X != 0
3092 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3093 Instruction *LI = cast<CastInst>(LHS);
3094 Value *SrcOp = LI->getOperand(0);
3095 Type *SrcTy = SrcOp->getType();
3096 Type *DstTy = LI->getType();
3098 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3099 // if the integer type is the same size as the pointer type.
3100 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3101 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3102 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3103 // Transfer the cast to the constant.
3104 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3105 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3108 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3109 if (RI->getOperand(0)->getType() == SrcTy)
3110 // Compare without the cast.
3111 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3117 if (isa<ZExtInst>(LHS)) {
3118 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3120 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3121 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3122 // Compare X and Y. Note that signed predicates become unsigned.
3123 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3124 SrcOp, RI->getOperand(0), Q,
3128 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3129 // too. If not, then try to deduce the result of the comparison.
3130 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3131 // Compute the constant that would happen if we truncated to SrcTy then
3132 // reextended to DstTy.
3133 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3134 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3136 // If the re-extended constant didn't change then this is effectively
3137 // also a case of comparing two zero-extended values.
3138 if (RExt == CI && MaxRecurse)
3139 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3140 SrcOp, Trunc, Q, MaxRecurse-1))
3143 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3144 // there. Use this to work out the result of the comparison.
3147 default: llvm_unreachable("Unknown ICmp predicate!");
3149 case ICmpInst::ICMP_EQ:
3150 case ICmpInst::ICMP_UGT:
3151 case ICmpInst::ICMP_UGE:
3152 return ConstantInt::getFalse(CI->getContext());
3154 case ICmpInst::ICMP_NE:
3155 case ICmpInst::ICMP_ULT:
3156 case ICmpInst::ICMP_ULE:
3157 return ConstantInt::getTrue(CI->getContext());
3159 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3160 // is non-negative then LHS <s RHS.
3161 case ICmpInst::ICMP_SGT:
3162 case ICmpInst::ICMP_SGE:
3163 return CI->getValue().isNegative() ?
3164 ConstantInt::getTrue(CI->getContext()) :
3165 ConstantInt::getFalse(CI->getContext());
3167 case ICmpInst::ICMP_SLT:
3168 case ICmpInst::ICMP_SLE:
3169 return CI->getValue().isNegative() ?
3170 ConstantInt::getFalse(CI->getContext()) :
3171 ConstantInt::getTrue(CI->getContext());
3177 if (isa<SExtInst>(LHS)) {
3178 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3180 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3181 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3182 // Compare X and Y. Note that the predicate does not change.
3183 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3187 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3188 // too. If not, then try to deduce the result of the comparison.
3189 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3190 // Compute the constant that would happen if we truncated to SrcTy then
3191 // reextended to DstTy.
3192 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3193 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3195 // If the re-extended constant didn't change then this is effectively
3196 // also a case of comparing two sign-extended values.
3197 if (RExt == CI && MaxRecurse)
3198 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3201 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3202 // bits there. Use this to work out the result of the comparison.
3205 default: llvm_unreachable("Unknown ICmp predicate!");
3206 case ICmpInst::ICMP_EQ:
3207 return ConstantInt::getFalse(CI->getContext());
3208 case ICmpInst::ICMP_NE:
3209 return ConstantInt::getTrue(CI->getContext());
3211 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3213 case ICmpInst::ICMP_SGT:
3214 case ICmpInst::ICMP_SGE:
3215 return CI->getValue().isNegative() ?
3216 ConstantInt::getTrue(CI->getContext()) :
3217 ConstantInt::getFalse(CI->getContext());
3218 case ICmpInst::ICMP_SLT:
3219 case ICmpInst::ICMP_SLE:
3220 return CI->getValue().isNegative() ?
3221 ConstantInt::getFalse(CI->getContext()) :
3222 ConstantInt::getTrue(CI->getContext());
3224 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3226 case ICmpInst::ICMP_UGT:
3227 case ICmpInst::ICMP_UGE:
3228 // Comparison is true iff the LHS <s 0.
3230 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3231 Constant::getNullValue(SrcTy),
3235 case ICmpInst::ICMP_ULT:
3236 case ICmpInst::ICMP_ULE:
3237 // Comparison is true iff the LHS >=s 0.
3239 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3240 Constant::getNullValue(SrcTy),
3250 // icmp eq|ne X, Y -> false|true if X != Y
3251 if (ICmpInst::isEquality(Pred) &&
3252 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3253 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3256 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3259 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3262 // Simplify comparisons of related pointers using a powerful, recursive
3263 // GEP-walk when we have target data available..
3264 if (LHS->getType()->isPointerTy())
3265 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3268 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3269 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3270 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3271 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3272 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3273 Q.DL.getTypeSizeInBits(CRHS->getType()))
3274 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3275 CLHS->getPointerOperand(),
3276 CRHS->getPointerOperand()))
3279 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3280 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3281 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3282 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3283 (ICmpInst::isEquality(Pred) ||
3284 (GLHS->isInBounds() && GRHS->isInBounds() &&
3285 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3286 // The bases are equal and the indices are constant. Build a constant
3287 // expression GEP with the same indices and a null base pointer to see
3288 // what constant folding can make out of it.
3289 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3290 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3291 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3292 GLHS->getSourceElementType(), Null, IndicesLHS);
3294 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3295 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3296 GLHS->getSourceElementType(), Null, IndicesRHS);
3297 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3302 // If the comparison is with the result of a select instruction, check whether
3303 // comparing with either branch of the select always yields the same value.
3304 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3305 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3308 // If the comparison is with the result of a phi instruction, check whether
3309 // doing the compare with each incoming phi value yields a common result.
3310 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3311 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3317 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3318 const SimplifyQuery &Q) {
3319 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3322 /// Given operands for an FCmpInst, see if we can fold the result.
3323 /// If not, this returns null.
3324 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3325 FastMathFlags FMF, const SimplifyQuery &Q,
3326 unsigned MaxRecurse) {
3327 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3328 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3330 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3331 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3332 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3334 // If we have a constant, make sure it is on the RHS.
3335 std::swap(LHS, RHS);
3336 Pred = CmpInst::getSwappedPredicate(Pred);
3339 // Fold trivial predicates.
3340 Type *RetTy = GetCompareTy(LHS);
3341 if (Pred == FCmpInst::FCMP_FALSE)
3342 return getFalse(RetTy);
3343 if (Pred == FCmpInst::FCMP_TRUE)
3344 return getTrue(RetTy);
3346 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3348 if (Pred == FCmpInst::FCMP_UNO)
3349 return getFalse(RetTy);
3350 if (Pred == FCmpInst::FCMP_ORD)
3351 return getTrue(RetTy);
3354 // fcmp pred x, undef and fcmp pred undef, x
3355 // fold to true if unordered, false if ordered
3356 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3357 // Choosing NaN for the undef will always make unordered comparison succeed
3358 // and ordered comparison fail.
3359 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3362 // fcmp x,x -> true/false. Not all compares are foldable.
3364 if (CmpInst::isTrueWhenEqual(Pred))
3365 return getTrue(RetTy);
3366 if (CmpInst::isFalseWhenEqual(Pred))
3367 return getFalse(RetTy);
3370 // Handle fcmp with constant RHS.
3372 if (match(RHS, m_APFloat(C))) {
3373 // If the constant is a nan, see if we can fold the comparison based on it.
3375 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3376 return getFalse(RetTy);
3377 assert(FCmpInst::isUnordered(Pred) &&
3378 "Comparison must be either ordered or unordered!");
3379 // True if unordered.
3380 return getTrue(RetTy);
3382 // Check whether the constant is an infinity.
3383 if (C->isInfinity()) {
3384 if (C->isNegative()) {
3386 case FCmpInst::FCMP_OLT:
3387 // No value is ordered and less than negative infinity.
3388 return getFalse(RetTy);
3389 case FCmpInst::FCMP_UGE:
3390 // All values are unordered with or at least negative infinity.
3391 return getTrue(RetTy);
3397 case FCmpInst::FCMP_OGT:
3398 // No value is ordered and greater than infinity.
3399 return getFalse(RetTy);
3400 case FCmpInst::FCMP_ULE:
3401 // All values are unordered with and at most infinity.
3402 return getTrue(RetTy);
3410 case FCmpInst::FCMP_UGE:
3411 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3412 return getTrue(RetTy);
3414 case FCmpInst::FCMP_OLT:
3416 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3417 return getFalse(RetTy);
3422 } else if (C->isNegative()) {
3423 assert(!C->isNaN() && "Unexpected NaN constant!");
3424 // TODO: We can catch more cases by using a range check rather than
3425 // relying on CannotBeOrderedLessThanZero.
3427 case FCmpInst::FCMP_UGE:
3428 case FCmpInst::FCMP_UGT:
3429 case FCmpInst::FCMP_UNE:
3430 // (X >= 0) implies (X > C) when (C < 0)
3431 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3432 return getTrue(RetTy);
3434 case FCmpInst::FCMP_OEQ:
3435 case FCmpInst::FCMP_OLE:
3436 case FCmpInst::FCMP_OLT:
3437 // (X >= 0) implies !(X < C) when (C < 0)
3438 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3439 return getFalse(RetTy);
3447 // If the comparison is with the result of a select instruction, check whether
3448 // comparing with either branch of the select always yields the same value.
3449 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3450 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3453 // If the comparison is with the result of a phi instruction, check whether
3454 // doing the compare with each incoming phi value yields a common result.
3455 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3456 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3462 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3463 FastMathFlags FMF, const SimplifyQuery &Q) {
3464 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3467 /// See if V simplifies when its operand Op is replaced with RepOp.
3468 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3469 const SimplifyQuery &Q,
3470 unsigned MaxRecurse) {
3471 // Trivial replacement.
3475 // We cannot replace a constant, and shouldn't even try.
3476 if (isa<Constant>(Op))
3479 auto *I = dyn_cast<Instruction>(V);
3483 // If this is a binary operator, try to simplify it with the replaced op.
3484 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3486 // %cmp = icmp eq i32 %x, 2147483647
3487 // %add = add nsw i32 %x, 1
3488 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3490 // We can't replace %sel with %add unless we strip away the flags.
3491 if (isa<OverflowingBinaryOperator>(B))
3492 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3494 if (isa<PossiblyExactOperator>(B))
3499 if (B->getOperand(0) == Op)
3500 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3502 if (B->getOperand(1) == Op)
3503 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3508 // Same for CmpInsts.
3509 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3511 if (C->getOperand(0) == Op)
3512 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3514 if (C->getOperand(1) == Op)
3515 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3520 // TODO: We could hand off more cases to instsimplify here.
3522 // If all operands are constant after substituting Op for RepOp then we can
3523 // constant fold the instruction.
3524 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3525 // Build a list of all constant operands.
3526 SmallVector<Constant *, 8> ConstOps;
3527 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3528 if (I->getOperand(i) == Op)
3529 ConstOps.push_back(CRepOp);
3530 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3531 ConstOps.push_back(COp);
3536 // All operands were constants, fold it.
3537 if (ConstOps.size() == I->getNumOperands()) {
3538 if (CmpInst *C = dyn_cast<CmpInst>(I))
3539 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3540 ConstOps[1], Q.DL, Q.TLI);
3542 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3543 if (!LI->isVolatile())
3544 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3546 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3553 /// Try to simplify a select instruction when its condition operand is an
3554 /// integer comparison where one operand of the compare is a constant.
3555 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3556 const APInt *Y, bool TrueWhenUnset) {
3559 // (X & Y) == 0 ? X & ~Y : X --> X
3560 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3561 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3563 return TrueWhenUnset ? FalseVal : TrueVal;
3565 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3566 // (X & Y) != 0 ? X : X & ~Y --> X
3567 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3569 return TrueWhenUnset ? FalseVal : TrueVal;
3571 if (Y->isPowerOf2()) {
3572 // (X & Y) == 0 ? X | Y : X --> X | Y
3573 // (X & Y) != 0 ? X | Y : X --> X
3574 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3576 return TrueWhenUnset ? TrueVal : FalseVal;
3578 // (X & Y) == 0 ? X : X | Y --> X
3579 // (X & Y) != 0 ? X : X | Y --> X | Y
3580 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3582 return TrueWhenUnset ? TrueVal : FalseVal;
3588 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3590 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3591 ICmpInst::Predicate Pred,
3592 Value *TrueVal, Value *FalseVal) {
3595 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3598 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3599 Pred == ICmpInst::ICMP_EQ);
3602 /// Try to simplify a select instruction when its condition operand is an
3603 /// integer comparison.
3604 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3605 Value *FalseVal, const SimplifyQuery &Q,
3606 unsigned MaxRecurse) {
3607 ICmpInst::Predicate Pred;
3608 Value *CmpLHS, *CmpRHS;
3609 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3612 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3615 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3616 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3617 Pred == ICmpInst::ICMP_EQ))
3621 // Check for other compares that behave like bit test.
3622 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3626 if (CondVal->hasOneUse()) {
3628 if (match(CmpRHS, m_APInt(C))) {
3629 // X < MIN ? T : F --> F
3630 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3632 // X < MIN ? T : F --> F
3633 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3635 // X > MAX ? T : F --> F
3636 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3638 // X > MAX ? T : F --> F
3639 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3644 // If we have an equality comparison, then we know the value in one of the
3645 // arms of the select. See if substituting this value into the arm and
3646 // simplifying the result yields the same value as the other arm.
3647 if (Pred == ICmpInst::ICMP_EQ) {
3648 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3650 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3653 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3655 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3658 } else if (Pred == ICmpInst::ICMP_NE) {
3659 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3661 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3664 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3666 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3674 /// Given operands for a SelectInst, see if we can fold the result.
3675 /// If not, this returns null.
3676 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3677 Value *FalseVal, const SimplifyQuery &Q,
3678 unsigned MaxRecurse) {
3679 // select true, X, Y -> X
3680 // select false, X, Y -> Y
3681 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3682 if (Constant *CT = dyn_cast<Constant>(TrueVal))
3683 if (Constant *CF = dyn_cast<Constant>(FalseVal))
3684 return ConstantFoldSelectInstruction(CB, CT, CF);
3685 if (CB->isAllOnesValue())
3687 if (CB->isNullValue())
3691 // select C, X, X -> X
3692 if (TrueVal == FalseVal)
3695 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3696 if (isa<Constant>(FalseVal))
3700 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3702 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3706 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3712 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3713 const SimplifyQuery &Q) {
3714 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3717 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3718 /// If not, this returns null.
3719 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3720 const SimplifyQuery &Q, unsigned) {
3721 // The type of the GEP pointer operand.
3723 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3725 // getelementptr P -> P.
3726 if (Ops.size() == 1)
3729 // Compute the (pointer) type returned by the GEP instruction.
3730 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3731 Type *GEPTy = PointerType::get(LastType, AS);
3732 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3733 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3734 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3735 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3737 if (isa<UndefValue>(Ops[0]))
3738 return UndefValue::get(GEPTy);
3740 if (Ops.size() == 2) {
3741 // getelementptr P, 0 -> P.
3742 if (match(Ops[1], m_Zero()))
3746 if (Ty->isSized()) {
3749 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3750 // getelementptr P, N -> P if P points to a type of zero size.
3751 if (TyAllocSize == 0)
3754 // The following transforms are only safe if the ptrtoint cast
3755 // doesn't truncate the pointers.
3756 if (Ops[1]->getType()->getScalarSizeInBits() ==
3757 Q.DL.getPointerSizeInBits(AS)) {
3758 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3759 if (match(P, m_Zero()))
3760 return Constant::getNullValue(GEPTy);
3762 if (match(P, m_PtrToInt(m_Value(Temp))))
3763 if (Temp->getType() == GEPTy)
3768 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3769 if (TyAllocSize == 1 &&
3770 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3771 if (Value *R = PtrToIntOrZero(P))
3774 // getelementptr V, (ashr (sub P, V), C) -> Q
3775 // if P points to a type of size 1 << C.
3777 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3778 m_ConstantInt(C))) &&
3779 TyAllocSize == 1ULL << C)
3780 if (Value *R = PtrToIntOrZero(P))
3783 // getelementptr V, (sdiv (sub P, V), C) -> Q
3784 // if P points to a type of size C.
3786 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3787 m_SpecificInt(TyAllocSize))))
3788 if (Value *R = PtrToIntOrZero(P))
3794 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3795 all_of(Ops.slice(1).drop_back(1),
3796 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3798 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3799 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3800 APInt BasePtrOffset(PtrWidth, 0);
3801 Value *StrippedBasePtr =
3802 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3805 // gep (gep V, C), (sub 0, V) -> C
3806 if (match(Ops.back(),
3807 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3808 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3809 return ConstantExpr::getIntToPtr(CI, GEPTy);
3811 // gep (gep V, C), (xor V, -1) -> C-1
3812 if (match(Ops.back(),
3813 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3814 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3815 return ConstantExpr::getIntToPtr(CI, GEPTy);
3820 // Check to see if this is constant foldable.
3821 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3824 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3826 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3831 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3832 const SimplifyQuery &Q) {
3833 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3836 /// Given operands for an InsertValueInst, see if we can fold the result.
3837 /// If not, this returns null.
3838 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3839 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3841 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3842 if (Constant *CVal = dyn_cast<Constant>(Val))
3843 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3845 // insertvalue x, undef, n -> x
3846 if (match(Val, m_Undef()))
3849 // insertvalue x, (extractvalue y, n), n
3850 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3851 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3852 EV->getIndices() == Idxs) {
3853 // insertvalue undef, (extractvalue y, n), n -> y
3854 if (match(Agg, m_Undef()))
3855 return EV->getAggregateOperand();
3857 // insertvalue y, (extractvalue y, n), n -> y
3858 if (Agg == EV->getAggregateOperand())
3865 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3866 ArrayRef<unsigned> Idxs,
3867 const SimplifyQuery &Q) {
3868 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3871 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3872 const SimplifyQuery &Q) {
3873 // Try to constant fold.
3874 auto *VecC = dyn_cast<Constant>(Vec);
3875 auto *ValC = dyn_cast<Constant>(Val);
3876 auto *IdxC = dyn_cast<Constant>(Idx);
3877 if (VecC && ValC && IdxC)
3878 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3880 // Fold into undef if index is out of bounds.
3881 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3882 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3883 if (CI->uge(NumElements))
3884 return UndefValue::get(Vec->getType());
3887 // If index is undef, it might be out of bounds (see above case)
3888 if (isa<UndefValue>(Idx))
3889 return UndefValue::get(Vec->getType());
3894 /// Given operands for an ExtractValueInst, see if we can fold the result.
3895 /// If not, this returns null.
3896 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3897 const SimplifyQuery &, unsigned) {
3898 if (auto *CAgg = dyn_cast<Constant>(Agg))
3899 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3901 // extractvalue x, (insertvalue y, elt, n), n -> elt
3902 unsigned NumIdxs = Idxs.size();
3903 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3904 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3905 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3906 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3907 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3908 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3909 Idxs.slice(0, NumCommonIdxs)) {
3910 if (NumIdxs == NumInsertValueIdxs)
3911 return IVI->getInsertedValueOperand();
3919 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3920 const SimplifyQuery &Q) {
3921 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3924 /// Given operands for an ExtractElementInst, see if we can fold the result.
3925 /// If not, this returns null.
3926 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3928 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3929 if (auto *CIdx = dyn_cast<Constant>(Idx))
3930 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3932 // The index is not relevant if our vector is a splat.
3933 if (auto *Splat = CVec->getSplatValue())
3936 if (isa<UndefValue>(Vec))
3937 return UndefValue::get(Vec->getType()->getVectorElementType());
3940 // If extracting a specified index from the vector, see if we can recursively
3941 // find a previously computed scalar that was inserted into the vector.
3942 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
3943 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
3944 // definitely out of bounds, thus undefined result
3945 return UndefValue::get(Vec->getType()->getVectorElementType());
3946 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3950 // An undef extract index can be arbitrarily chosen to be an out-of-range
3951 // index value, which would result in the instruction being undef.
3952 if (isa<UndefValue>(Idx))
3953 return UndefValue::get(Vec->getType()->getVectorElementType());
3958 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3959 const SimplifyQuery &Q) {
3960 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3963 /// See if we can fold the given phi. If not, returns null.
3964 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3965 // If all of the PHI's incoming values are the same then replace the PHI node
3966 // with the common value.
3967 Value *CommonValue = nullptr;
3968 bool HasUndefInput = false;
3969 for (Value *Incoming : PN->incoming_values()) {
3970 // If the incoming value is the phi node itself, it can safely be skipped.
3971 if (Incoming == PN) continue;
3972 if (isa<UndefValue>(Incoming)) {
3973 // Remember that we saw an undef value, but otherwise ignore them.
3974 HasUndefInput = true;
3977 if (CommonValue && Incoming != CommonValue)
3978 return nullptr; // Not the same, bail out.
3979 CommonValue = Incoming;
3982 // If CommonValue is null then all of the incoming values were either undef or
3983 // equal to the phi node itself.
3985 return UndefValue::get(PN->getType());
3987 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3988 // instruction, we cannot return X as the result of the PHI node unless it
3989 // dominates the PHI block.
3991 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3996 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
3997 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
3998 if (auto *C = dyn_cast<Constant>(Op))
3999 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4001 if (auto *CI = dyn_cast<CastInst>(Op)) {
4002 auto *Src = CI->getOperand(0);
4003 Type *SrcTy = Src->getType();
4004 Type *MidTy = CI->getType();
4006 if (Src->getType() == Ty) {
4007 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4008 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4010 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4012 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4014 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4015 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4016 SrcIntPtrTy, MidIntPtrTy,
4017 DstIntPtrTy) == Instruction::BitCast)
4023 if (CastOpc == Instruction::BitCast)
4024 if (Op->getType() == Ty)
4030 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4031 const SimplifyQuery &Q) {
4032 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4035 /// For the given destination element of a shuffle, peek through shuffles to
4036 /// match a root vector source operand that contains that element in the same
4037 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4038 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4039 int MaskVal, Value *RootVec,
4040 unsigned MaxRecurse) {
4044 // Bail out if any mask value is undefined. That kind of shuffle may be
4045 // simplified further based on demanded bits or other folds.
4049 // The mask value chooses which source operand we need to look at next.
4050 int InVecNumElts = Op0->getType()->getVectorNumElements();
4051 int RootElt = MaskVal;
4052 Value *SourceOp = Op0;
4053 if (MaskVal >= InVecNumElts) {
4054 RootElt = MaskVal - InVecNumElts;
4058 // If the source operand is a shuffle itself, look through it to find the
4059 // matching root vector.
4060 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4061 return foldIdentityShuffles(
4062 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4063 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4066 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4069 // The source operand is not a shuffle. Initialize the root vector value for
4070 // this shuffle if that has not been done yet.
4074 // Give up as soon as a source operand does not match the existing root value.
4075 if (RootVec != SourceOp)
4078 // The element must be coming from the same lane in the source vector
4079 // (although it may have crossed lanes in intermediate shuffles).
4080 if (RootElt != DestElt)
4086 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4087 Type *RetTy, const SimplifyQuery &Q,
4088 unsigned MaxRecurse) {
4089 if (isa<UndefValue>(Mask))
4090 return UndefValue::get(RetTy);
4092 Type *InVecTy = Op0->getType();
4093 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4094 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4096 SmallVector<int, 32> Indices;
4097 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4098 assert(MaskNumElts == Indices.size() &&
4099 "Size of Indices not same as number of mask elements?");
4101 // Canonicalization: If mask does not select elements from an input vector,
4102 // replace that input vector with undef.
4103 bool MaskSelects0 = false, MaskSelects1 = false;
4104 for (unsigned i = 0; i != MaskNumElts; ++i) {
4105 if (Indices[i] == -1)
4107 if ((unsigned)Indices[i] < InVecNumElts)
4108 MaskSelects0 = true;
4110 MaskSelects1 = true;
4113 Op0 = UndefValue::get(InVecTy);
4115 Op1 = UndefValue::get(InVecTy);
4117 auto *Op0Const = dyn_cast<Constant>(Op0);
4118 auto *Op1Const = dyn_cast<Constant>(Op1);
4120 // If all operands are constant, constant fold the shuffle.
4121 if (Op0Const && Op1Const)
4122 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4124 // Canonicalization: if only one input vector is constant, it shall be the
4126 if (Op0Const && !Op1Const) {
4127 std::swap(Op0, Op1);
4128 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4131 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4132 // value type is same as the input vectors' type.
4133 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4134 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4135 OpShuf->getMask()->getSplatValue())
4138 // Don't fold a shuffle with undef mask elements. This may get folded in a
4139 // better way using demanded bits or other analysis.
4140 // TODO: Should we allow this?
4141 if (find(Indices, -1) != Indices.end())
4144 // Check if every element of this shuffle can be mapped back to the
4145 // corresponding element of a single root vector. If so, we don't need this
4146 // shuffle. This handles simple identity shuffles as well as chains of
4147 // shuffles that may widen/narrow and/or move elements across lanes and back.
4148 Value *RootVec = nullptr;
4149 for (unsigned i = 0; i != MaskNumElts; ++i) {
4150 // Note that recursion is limited for each vector element, so if any element
4151 // exceeds the limit, this will fail to simplify.
4153 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4155 // We can't replace a widening/narrowing shuffle with one of its operands.
4156 if (!RootVec || RootVec->getType() != RetTy)
4162 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4163 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4164 Type *RetTy, const SimplifyQuery &Q) {
4165 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4168 /// Given operands for an FAdd, see if we can fold the result. If not, this
4170 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4171 const SimplifyQuery &Q, unsigned MaxRecurse) {
4172 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4176 if (match(Op1, m_NegZero()))
4179 // fadd X, 0 ==> X, when we know X is not -0
4180 if (match(Op1, m_Zero()) &&
4181 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4184 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
4185 // where nnan and ninf have to occur at least once somewhere in this
4187 Value *SubOp = nullptr;
4188 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
4190 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
4193 Instruction *FSub = cast<Instruction>(SubOp);
4194 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
4195 (FMF.noInfs() || FSub->hasNoInfs()))
4196 return Constant::getNullValue(Op0->getType());
4202 /// Given operands for an FSub, see if we can fold the result. If not, this
4204 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4205 const SimplifyQuery &Q, unsigned MaxRecurse) {
4206 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4210 if (match(Op1, m_Zero()))
4213 // fsub X, -0 ==> X, when we know X is not -0
4214 if (match(Op1, m_NegZero()) &&
4215 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4218 // fsub -0.0, (fsub -0.0, X) ==> X
4220 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
4223 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4224 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
4225 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
4228 // fsub nnan x, x ==> 0.0
4229 if (FMF.noNaNs() && Op0 == Op1)
4230 return Constant::getNullValue(Op0->getType());
4235 /// Given the operands for an FMul, see if we can fold the result
4236 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4237 const SimplifyQuery &Q, unsigned MaxRecurse) {
4238 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4241 // fmul X, 1.0 ==> X
4242 if (match(Op1, m_FPOne()))
4245 // fmul nnan nsz X, 0 ==> 0
4246 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
4252 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4253 const SimplifyQuery &Q) {
4254 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4258 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4259 const SimplifyQuery &Q) {
4260 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4263 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4264 const SimplifyQuery &Q) {
4265 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4268 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4269 const SimplifyQuery &Q, unsigned) {
4270 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, 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 if (match(Op1, m_FPOne()))
4286 // Requires that NaNs are off (X could be zero) and signed zeroes are
4287 // ignored (X could be positive or negative, so the output sign is unknown).
4288 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4292 // X / X -> 1.0 is legal when NaNs are ignored.
4294 return ConstantFP::get(Op0->getType(), 1.0);
4296 // -X / X -> -1.0 and
4297 // X / -X -> -1.0 are legal when NaNs are ignored.
4298 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4299 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4300 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4301 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4302 BinaryOperator::getFNegArgument(Op1) == Op0))
4303 return ConstantFP::get(Op0->getType(), -1.0);
4309 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4310 const SimplifyQuery &Q) {
4311 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4314 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4315 const SimplifyQuery &Q, unsigned) {
4316 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4319 // undef % X -> undef (the undef could be a snan).
4320 if (match(Op0, m_Undef()))
4323 // X % undef -> undef
4324 if (match(Op1, m_Undef()))
4328 // Requires that NaNs are off (X could be zero) and signed zeroes are
4329 // ignored (X could be positive or negative, so the output sign is unknown).
4330 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4336 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4337 const SimplifyQuery &Q) {
4338 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4341 //=== Helper functions for higher up the class hierarchy.
4343 /// Given operands for a BinaryOperator, see if we can fold the result.
4344 /// If not, this returns null.
4345 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4346 const SimplifyQuery &Q, unsigned MaxRecurse) {
4348 case Instruction::Add:
4349 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4350 case Instruction::Sub:
4351 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4352 case Instruction::Mul:
4353 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4354 case Instruction::SDiv:
4355 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4356 case Instruction::UDiv:
4357 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4358 case Instruction::SRem:
4359 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4360 case Instruction::URem:
4361 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4362 case Instruction::Shl:
4363 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4364 case Instruction::LShr:
4365 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4366 case Instruction::AShr:
4367 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4368 case Instruction::And:
4369 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4370 case Instruction::Or:
4371 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4372 case Instruction::Xor:
4373 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4374 case Instruction::FAdd:
4375 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4376 case Instruction::FSub:
4377 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4378 case Instruction::FMul:
4379 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4380 case Instruction::FDiv:
4381 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4382 case Instruction::FRem:
4383 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4385 llvm_unreachable("Unexpected opcode");
4389 /// Given operands for a BinaryOperator, see if we can fold the result.
4390 /// If not, this returns null.
4391 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4392 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4393 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4394 const FastMathFlags &FMF, const SimplifyQuery &Q,
4395 unsigned MaxRecurse) {
4397 case Instruction::FAdd:
4398 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4399 case Instruction::FSub:
4400 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4401 case Instruction::FMul:
4402 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4403 case Instruction::FDiv:
4404 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4406 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4410 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4411 const SimplifyQuery &Q) {
4412 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4415 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4416 FastMathFlags FMF, const SimplifyQuery &Q) {
4417 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4420 /// Given operands for a CmpInst, see if we can fold the result.
4421 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4422 const SimplifyQuery &Q, unsigned MaxRecurse) {
4423 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4424 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4425 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4428 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4429 const SimplifyQuery &Q) {
4430 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4433 static bool IsIdempotent(Intrinsic::ID ID) {
4435 default: return false;
4437 // Unary idempotent: f(f(x)) = f(x)
4438 case Intrinsic::fabs:
4439 case Intrinsic::floor:
4440 case Intrinsic::ceil:
4441 case Intrinsic::trunc:
4442 case Intrinsic::rint:
4443 case Intrinsic::nearbyint:
4444 case Intrinsic::round:
4445 case Intrinsic::canonicalize:
4450 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4451 const DataLayout &DL) {
4452 GlobalValue *PtrSym;
4454 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4457 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4458 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4459 Type *Int32PtrTy = Int32Ty->getPointerTo();
4460 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4462 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4463 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4466 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4467 if (OffsetInt % 4 != 0)
4470 Constant *C = ConstantExpr::getGetElementPtr(
4471 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4472 ConstantInt::get(Int64Ty, OffsetInt / 4));
4473 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4477 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4481 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4482 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4487 if (LoadedCE->getOpcode() != Instruction::Sub)
4490 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4491 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4493 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4495 Constant *LoadedRHS = LoadedCE->getOperand(1);
4496 GlobalValue *LoadedRHSSym;
4497 APInt LoadedRHSOffset;
4498 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4500 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4503 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4506 static bool maskIsAllZeroOrUndef(Value *Mask) {
4507 auto *ConstMask = dyn_cast<Constant>(Mask);
4510 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4512 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4514 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4515 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4522 template <typename IterTy>
4523 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4524 const SimplifyQuery &Q, unsigned MaxRecurse) {
4525 Intrinsic::ID IID = F->getIntrinsicID();
4526 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4529 if (NumOperands == 1) {
4530 // Perform idempotent optimizations
4531 if (IsIdempotent(IID)) {
4532 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4533 if (II->getIntrinsicID() == IID)
4538 Value *IIOperand = *ArgBegin;
4541 case Intrinsic::fabs: {
4542 if (SignBitMustBeZero(IIOperand, Q.TLI))
4546 case Intrinsic::bswap: {
4547 // bswap(bswap(x)) -> x
4548 if (match(IIOperand, m_BSwap(m_Value(X))))
4552 case Intrinsic::bitreverse: {
4553 // bitreverse(bitreverse(x)) -> x
4554 if (match(IIOperand, m_BitReverse(m_Value(X))))
4558 case Intrinsic::exp: {
4560 if (Q.CxtI->isFast() &&
4561 match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X))))
4565 case Intrinsic::exp2: {
4566 // exp2(log2(x)) -> x
4567 if (Q.CxtI->isFast() &&
4568 match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X))))
4572 case Intrinsic::log: {
4574 if (Q.CxtI->isFast() &&
4575 match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X))))
4579 case Intrinsic::log2: {
4580 // log2(exp2(x)) -> x
4581 if (Q.CxtI->isFast() &&
4582 match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) {
4593 if (NumOperands == 2) {
4594 Value *LHS = *ArgBegin;
4595 Value *RHS = *(ArgBegin + 1);
4596 Type *ReturnType = F->getReturnType();
4599 case Intrinsic::usub_with_overflow:
4600 case Intrinsic::ssub_with_overflow: {
4601 // X - X -> { 0, false }
4603 return Constant::getNullValue(ReturnType);
4605 // X - undef -> undef
4606 // undef - X -> undef
4607 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4608 return UndefValue::get(ReturnType);
4612 case Intrinsic::uadd_with_overflow:
4613 case Intrinsic::sadd_with_overflow: {
4614 // X + undef -> undef
4615 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4616 return UndefValue::get(ReturnType);
4620 case Intrinsic::umul_with_overflow:
4621 case Intrinsic::smul_with_overflow: {
4622 // 0 * X -> { 0, false }
4623 // X * 0 -> { 0, false }
4624 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4625 return Constant::getNullValue(ReturnType);
4627 // undef * X -> { 0, false }
4628 // X * undef -> { 0, false }
4629 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4630 return Constant::getNullValue(ReturnType);
4634 case Intrinsic::load_relative: {
4635 Constant *C0 = dyn_cast<Constant>(LHS);
4636 Constant *C1 = dyn_cast<Constant>(RHS);
4638 return SimplifyRelativeLoad(C0, C1, Q.DL);
4641 case Intrinsic::powi:
4642 if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) {
4643 // powi(x, 0) -> 1.0
4644 if (Power->isZero())
4645 return ConstantFP::get(LHS->getType(), 1.0);
4656 // Simplify calls to llvm.masked.load.*
4658 case Intrinsic::masked_load: {
4659 Value *MaskArg = ArgBegin[2];
4660 Value *PassthruArg = ArgBegin[3];
4661 // If the mask is all zeros or undef, the "passthru" argument is the result.
4662 if (maskIsAllZeroOrUndef(MaskArg))
4671 template <typename IterTy>
4672 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4673 IterTy ArgEnd, const SimplifyQuery &Q,
4674 unsigned MaxRecurse) {
4675 Type *Ty = V->getType();
4676 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4677 Ty = PTy->getElementType();
4678 FunctionType *FTy = cast<FunctionType>(Ty);
4680 // call undef -> undef
4681 // call null -> undef
4682 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4683 return UndefValue::get(FTy->getReturnType());
4685 Function *F = dyn_cast<Function>(V);
4689 if (F->isIntrinsic())
4690 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4693 if (!canConstantFoldCallTo(CS, F))
4696 SmallVector<Constant *, 4> ConstantArgs;
4697 ConstantArgs.reserve(ArgEnd - ArgBegin);
4698 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4699 Constant *C = dyn_cast<Constant>(*I);
4702 ConstantArgs.push_back(C);
4705 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4708 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4709 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4710 const SimplifyQuery &Q) {
4711 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4714 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4715 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4716 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4719 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4720 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4721 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4725 /// See if we can compute a simplified version of this instruction.
4726 /// If not, this returns null.
4728 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4729 OptimizationRemarkEmitter *ORE) {
4730 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4733 switch (I->getOpcode()) {
4735 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4737 case Instruction::FAdd:
4738 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4739 I->getFastMathFlags(), Q);
4741 case Instruction::Add:
4742 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4743 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4744 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4746 case Instruction::FSub:
4747 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4748 I->getFastMathFlags(), Q);
4750 case Instruction::Sub:
4751 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4752 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4753 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4755 case Instruction::FMul:
4756 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4757 I->getFastMathFlags(), Q);
4759 case Instruction::Mul:
4760 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4762 case Instruction::SDiv:
4763 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4765 case Instruction::UDiv:
4766 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4768 case Instruction::FDiv:
4769 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4770 I->getFastMathFlags(), Q);
4772 case Instruction::SRem:
4773 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4775 case Instruction::URem:
4776 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4778 case Instruction::FRem:
4779 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4780 I->getFastMathFlags(), Q);
4782 case Instruction::Shl:
4783 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4784 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4785 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4787 case Instruction::LShr:
4788 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4789 cast<BinaryOperator>(I)->isExact(), Q);
4791 case Instruction::AShr:
4792 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4793 cast<BinaryOperator>(I)->isExact(), Q);
4795 case Instruction::And:
4796 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4798 case Instruction::Or:
4799 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4801 case Instruction::Xor:
4802 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4804 case Instruction::ICmp:
4805 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4806 I->getOperand(0), I->getOperand(1), Q);
4808 case Instruction::FCmp:
4810 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4811 I->getOperand(1), I->getFastMathFlags(), Q);
4813 case Instruction::Select:
4814 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4815 I->getOperand(2), Q);
4817 case Instruction::GetElementPtr: {
4818 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4819 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4823 case Instruction::InsertValue: {
4824 InsertValueInst *IV = cast<InsertValueInst>(I);
4825 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4826 IV->getInsertedValueOperand(),
4827 IV->getIndices(), Q);
4830 case Instruction::InsertElement: {
4831 auto *IE = cast<InsertElementInst>(I);
4832 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4833 IE->getOperand(2), Q);
4836 case Instruction::ExtractValue: {
4837 auto *EVI = cast<ExtractValueInst>(I);
4838 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4839 EVI->getIndices(), Q);
4842 case Instruction::ExtractElement: {
4843 auto *EEI = cast<ExtractElementInst>(I);
4844 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4845 EEI->getIndexOperand(), Q);
4848 case Instruction::ShuffleVector: {
4849 auto *SVI = cast<ShuffleVectorInst>(I);
4850 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4851 SVI->getMask(), SVI->getType(), Q);
4854 case Instruction::PHI:
4855 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4857 case Instruction::Call: {
4858 CallSite CS(cast<CallInst>(I));
4859 Result = SimplifyCall(CS, Q);
4862 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4863 #include "llvm/IR/Instruction.def"
4864 #undef HANDLE_CAST_INST
4866 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4868 case Instruction::Alloca:
4869 // No simplifications for Alloca and it can't be constant folded.
4874 // In general, it is possible for computeKnownBits to determine all bits in a
4875 // value even when the operands are not all constants.
4876 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4877 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4878 if (Known.isConstant())
4879 Result = ConstantInt::get(I->getType(), Known.getConstant());
4882 /// If called on unreachable code, the above logic may report that the
4883 /// instruction simplified to itself. Make life easier for users by
4884 /// detecting that case here, returning a safe value instead.
4885 return Result == I ? UndefValue::get(I->getType()) : Result;
4888 /// \brief Implementation of recursive simplification through an instruction's
4891 /// This is the common implementation of the recursive simplification routines.
4892 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4893 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4894 /// instructions to process and attempt to simplify it using
4895 /// InstructionSimplify.
4897 /// This routine returns 'true' only when *it* simplifies something. The passed
4898 /// in simplified value does not count toward this.
4899 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4900 const TargetLibraryInfo *TLI,
4901 const DominatorTree *DT,
4902 AssumptionCache *AC) {
4903 bool Simplified = false;
4904 SmallSetVector<Instruction *, 8> Worklist;
4905 const DataLayout &DL = I->getModule()->getDataLayout();
4907 // If we have an explicit value to collapse to, do that round of the
4908 // simplification loop by hand initially.
4910 for (User *U : I->users())
4912 Worklist.insert(cast<Instruction>(U));
4914 // Replace the instruction with its simplified value.
4915 I->replaceAllUsesWith(SimpleV);
4917 // Gracefully handle edge cases where the instruction is not wired into any
4919 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4920 !I->mayHaveSideEffects())
4921 I->eraseFromParent();
4926 // Note that we must test the size on each iteration, the worklist can grow.
4927 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4930 // See if this instruction simplifies.
4931 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4937 // Stash away all the uses of the old instruction so we can check them for
4938 // recursive simplifications after a RAUW. This is cheaper than checking all
4939 // uses of To on the recursive step in most cases.
4940 for (User *U : I->users())
4941 Worklist.insert(cast<Instruction>(U));
4943 // Replace the instruction with its simplified value.
4944 I->replaceAllUsesWith(SimpleV);
4946 // Gracefully handle edge cases where the instruction is not wired into any
4948 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4949 !I->mayHaveSideEffects())
4950 I->eraseFromParent();
4955 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4956 const TargetLibraryInfo *TLI,
4957 const DominatorTree *DT,
4958 AssumptionCache *AC) {
4959 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4962 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4963 const TargetLibraryInfo *TLI,
4964 const DominatorTree *DT,
4965 AssumptionCache *AC) {
4966 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4967 assert(SimpleV && "Must provide a simplified value.");
4968 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4972 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4973 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4974 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4975 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4976 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4977 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4978 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4979 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4982 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4983 const DataLayout &DL) {
4984 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4987 template <class T, class... TArgs>
4988 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4990 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4991 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4992 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4993 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4995 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,