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);
65 static Value *SimplifyGEPInst(Type *, ArrayRef<Value *>, const SimplifyQuery &,
68 /// For a boolean type or a vector of boolean type, return false or a vector
69 /// with every element false.
70 static Constant *getFalse(Type *Ty) {
71 return ConstantInt::getFalse(Ty);
74 /// For a boolean type or a vector of boolean type, return true or a vector
75 /// with every element true.
76 static Constant *getTrue(Type *Ty) {
77 return ConstantInt::getTrue(Ty);
80 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
81 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
83 CmpInst *Cmp = dyn_cast<CmpInst>(V);
86 CmpInst::Predicate CPred = Cmp->getPredicate();
87 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
88 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
90 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
94 /// Does the given value dominate the specified phi node?
95 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
96 Instruction *I = dyn_cast<Instruction>(V);
98 // Arguments and constants dominate all instructions.
101 // If we are processing instructions (and/or basic blocks) that have not been
102 // fully added to a function, the parent nodes may still be null. Simply
103 // return the conservative answer in these cases.
104 if (!I->getParent() || !P->getParent() || !I->getFunction())
107 // If we have a DominatorTree then do a precise test.
109 return DT->dominates(I, P);
111 // Otherwise, if the instruction is in the entry block and is not an invoke,
112 // then it obviously dominates all phi nodes.
113 if (I->getParent() == &I->getFunction()->getEntryBlock() &&
120 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
121 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
122 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
123 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
124 /// Returns the simplified value, or null if no simplification was performed.
125 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
126 Instruction::BinaryOps OpcodeToExpand,
127 const SimplifyQuery &Q, unsigned MaxRecurse) {
128 // Recursion is always used, so bail out at once if we already hit the limit.
132 // Check whether the expression has the form "(A op' B) op C".
133 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
134 if (Op0->getOpcode() == OpcodeToExpand) {
135 // It does! Try turning it into "(A op C) op' (B op C)".
136 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
137 // Do "A op C" and "B op C" both simplify?
138 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
139 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
140 // They do! Return "L op' R" if it simplifies or is already available.
141 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
142 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
143 && L == B && R == A)) {
147 // Otherwise return "L op' R" if it simplifies.
148 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
155 // Check whether the expression has the form "A op (B op' C)".
156 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
157 if (Op1->getOpcode() == OpcodeToExpand) {
158 // It does! Try turning it into "(A op B) op' (A op C)".
159 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
160 // Do "A op B" and "A op C" both simplify?
161 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
162 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
163 // They do! Return "L op' R" if it simplifies or is already available.
164 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
165 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
166 && L == C && R == B)) {
170 // Otherwise return "L op' R" if it simplifies.
171 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
181 /// Generic simplifications for associative binary operations.
182 /// Returns the simpler value, or null if none was found.
183 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
184 Value *LHS, Value *RHS,
185 const SimplifyQuery &Q,
186 unsigned MaxRecurse) {
187 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
189 // Recursion is always used, so bail out at once if we already hit the limit.
193 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
194 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
196 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
197 if (Op0 && Op0->getOpcode() == Opcode) {
198 Value *A = Op0->getOperand(0);
199 Value *B = Op0->getOperand(1);
202 // Does "B op C" simplify?
203 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
204 // It does! Return "A op V" if it simplifies or is already available.
205 // If V equals B then "A op V" is just the LHS.
206 if (V == B) return LHS;
207 // Otherwise return "A op V" if it simplifies.
208 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
215 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
216 if (Op1 && Op1->getOpcode() == Opcode) {
218 Value *B = Op1->getOperand(0);
219 Value *C = Op1->getOperand(1);
221 // Does "A op B" simplify?
222 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
223 // It does! Return "V op C" if it simplifies or is already available.
224 // If V equals B then "V op C" is just the RHS.
225 if (V == B) return RHS;
226 // Otherwise return "V op C" if it simplifies.
227 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
234 // The remaining transforms require commutativity as well as associativity.
235 if (!Instruction::isCommutative(Opcode))
238 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
239 if (Op0 && Op0->getOpcode() == Opcode) {
240 Value *A = Op0->getOperand(0);
241 Value *B = Op0->getOperand(1);
244 // Does "C op A" simplify?
245 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
246 // It does! Return "V op B" if it simplifies or is already available.
247 // If V equals A then "V op B" is just the LHS.
248 if (V == A) return LHS;
249 // Otherwise return "V op B" if it simplifies.
250 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
257 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
258 if (Op1 && Op1->getOpcode() == Opcode) {
260 Value *B = Op1->getOperand(0);
261 Value *C = Op1->getOperand(1);
263 // Does "C op A" simplify?
264 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
265 // It does! Return "B op V" if it simplifies or is already available.
266 // If V equals C then "B op V" is just the RHS.
267 if (V == C) return RHS;
268 // Otherwise return "B op V" if it simplifies.
269 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
279 /// In the case of a binary operation with a select instruction as an operand,
280 /// try to simplify the binop by seeing whether evaluating it on both branches
281 /// of the select results in the same value. Returns the common value if so,
282 /// otherwise returns null.
283 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
284 Value *RHS, const SimplifyQuery &Q,
285 unsigned MaxRecurse) {
286 // Recursion is always used, so bail out at once if we already hit the limit.
291 if (isa<SelectInst>(LHS)) {
292 SI = cast<SelectInst>(LHS);
294 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
295 SI = cast<SelectInst>(RHS);
298 // Evaluate the BinOp on the true and false branches of the select.
302 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
303 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
305 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
306 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
309 // If they simplified to the same value, then return the common value.
310 // If they both failed to simplify then return null.
314 // If one branch simplified to undef, return the other one.
315 if (TV && isa<UndefValue>(TV))
317 if (FV && isa<UndefValue>(FV))
320 // If applying the operation did not change the true and false select values,
321 // then the result of the binop is the select itself.
322 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
325 // If one branch simplified and the other did not, and the simplified
326 // value is equal to the unsimplified one, return the simplified value.
327 // For example, select (cond, X, X & Z) & Z -> X & Z.
328 if ((FV && !TV) || (TV && !FV)) {
329 // Check that the simplified value has the form "X op Y" where "op" is the
330 // same as the original operation.
331 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
332 if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
333 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
334 // We already know that "op" is the same as for the simplified value. See
335 // if the operands match too. If so, return the simplified value.
336 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
337 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
338 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
339 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
340 Simplified->getOperand(1) == UnsimplifiedRHS)
342 if (Simplified->isCommutative() &&
343 Simplified->getOperand(1) == UnsimplifiedLHS &&
344 Simplified->getOperand(0) == UnsimplifiedRHS)
352 /// In the case of a comparison with a select instruction, try to simplify the
353 /// comparison by seeing whether both branches of the select result in the same
354 /// value. Returns the common value if so, otherwise returns null.
355 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
356 Value *RHS, const SimplifyQuery &Q,
357 unsigned MaxRecurse) {
358 // Recursion is always used, so bail out at once if we already hit the limit.
362 // Make sure the select is on the LHS.
363 if (!isa<SelectInst>(LHS)) {
365 Pred = CmpInst::getSwappedPredicate(Pred);
367 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
368 SelectInst *SI = cast<SelectInst>(LHS);
369 Value *Cond = SI->getCondition();
370 Value *TV = SI->getTrueValue();
371 Value *FV = SI->getFalseValue();
373 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
374 // Does "cmp TV, RHS" simplify?
375 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
377 // It not only simplified, it simplified to the select condition. Replace
379 TCmp = getTrue(Cond->getType());
381 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
382 // condition then we can replace it with 'true'. Otherwise give up.
383 if (!isSameCompare(Cond, Pred, TV, RHS))
385 TCmp = getTrue(Cond->getType());
388 // Does "cmp FV, RHS" simplify?
389 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
391 // It not only simplified, it simplified to the select condition. Replace
393 FCmp = getFalse(Cond->getType());
395 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
396 // condition then we can replace it with 'false'. Otherwise give up.
397 if (!isSameCompare(Cond, Pred, FV, RHS))
399 FCmp = getFalse(Cond->getType());
402 // If both sides simplified to the same value, then use it as the result of
403 // the original comparison.
407 // The remaining cases only make sense if the select condition has the same
408 // type as the result of the comparison, so bail out if this is not so.
409 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
411 // If the false value simplified to false, then the result of the compare
412 // is equal to "Cond && TCmp". This also catches the case when the false
413 // value simplified to false and the true value to true, returning "Cond".
414 if (match(FCmp, m_Zero()))
415 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
417 // If the true value simplified to true, then the result of the compare
418 // is equal to "Cond || FCmp".
419 if (match(TCmp, m_One()))
420 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
422 // Finally, if the false value simplified to true and the true value to
423 // false, then the result of the compare is equal to "!Cond".
424 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
426 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
433 /// In the case of a binary operation with an operand that is a PHI instruction,
434 /// try to simplify the binop by seeing whether evaluating it on the incoming
435 /// phi values yields the same result for every value. If so returns the common
436 /// value, otherwise returns null.
437 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
438 Value *RHS, const SimplifyQuery &Q,
439 unsigned MaxRecurse) {
440 // Recursion is always used, so bail out at once if we already hit the limit.
445 if (isa<PHINode>(LHS)) {
446 PI = cast<PHINode>(LHS);
447 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
448 if (!valueDominatesPHI(RHS, PI, Q.DT))
451 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
452 PI = cast<PHINode>(RHS);
453 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
454 if (!valueDominatesPHI(LHS, PI, Q.DT))
458 // Evaluate the BinOp on the incoming phi values.
459 Value *CommonValue = nullptr;
460 for (Value *Incoming : PI->incoming_values()) {
461 // If the incoming value is the phi node itself, it can safely be skipped.
462 if (Incoming == PI) continue;
463 Value *V = PI == LHS ?
464 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
465 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
466 // If the operation failed to simplify, or simplified to a different value
467 // to previously, then give up.
468 if (!V || (CommonValue && V != CommonValue))
476 /// In the case of a comparison with a PHI instruction, try to simplify the
477 /// comparison by seeing whether comparing with all of the incoming phi values
478 /// yields the same result every time. If so returns the common result,
479 /// otherwise returns null.
480 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
481 const SimplifyQuery &Q, unsigned MaxRecurse) {
482 // Recursion is always used, so bail out at once if we already hit the limit.
486 // Make sure the phi is on the LHS.
487 if (!isa<PHINode>(LHS)) {
489 Pred = CmpInst::getSwappedPredicate(Pred);
491 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
492 PHINode *PI = cast<PHINode>(LHS);
494 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
495 if (!valueDominatesPHI(RHS, PI, Q.DT))
498 // Evaluate the BinOp on the incoming phi values.
499 Value *CommonValue = nullptr;
500 for (Value *Incoming : PI->incoming_values()) {
501 // If the incoming value is the phi node itself, it can safely be skipped.
502 if (Incoming == PI) continue;
503 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
504 // If the operation failed to simplify, or simplified to a different value
505 // to previously, then give up.
506 if (!V || (CommonValue && V != CommonValue))
514 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
515 Value *&Op0, Value *&Op1,
516 const SimplifyQuery &Q) {
517 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
518 if (auto *CRHS = dyn_cast<Constant>(Op1))
519 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
521 // Canonicalize the constant to the RHS if this is a commutative operation.
522 if (Instruction::isCommutative(Opcode))
528 /// Given operands for an Add, see if we can fold the result.
529 /// If not, this returns null.
530 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
531 const SimplifyQuery &Q, unsigned MaxRecurse) {
532 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
535 // X + undef -> undef
536 if (match(Op1, m_Undef()))
540 if (match(Op1, m_Zero()))
547 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
548 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
551 // X + ~X -> -1 since ~X = -X-1
552 Type *Ty = Op0->getType();
553 if (match(Op0, m_Not(m_Specific(Op1))) ||
554 match(Op1, m_Not(m_Specific(Op0))))
555 return Constant::getAllOnesValue(Ty);
557 // add nsw/nuw (xor Y, signmask), signmask --> Y
558 // The no-wrapping add guarantees that the top bit will be set by the add.
559 // Therefore, the xor must be clearing the already set sign bit of Y.
560 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
561 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
565 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
566 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
569 // Try some generic simplifications for associative operations.
570 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
574 // Threading Add over selects and phi nodes is pointless, so don't bother.
575 // Threading over the select in "A + select(cond, B, C)" means evaluating
576 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
577 // only if B and C are equal. If B and C are equal then (since we assume
578 // that operands have already been simplified) "select(cond, B, C)" should
579 // have been simplified to the common value of B and C already. Analysing
580 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
581 // for threading over phi nodes.
586 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
587 const SimplifyQuery &Query) {
588 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
591 /// Compute the base pointer and cumulative constant offsets for V.
593 /// This strips all constant offsets off of V, leaving it the base pointer, and
594 /// accumulates the total constant offset applied in the returned constant. It
595 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
596 /// no constant offsets applied.
598 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
599 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
601 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
602 bool AllowNonInbounds = false) {
603 assert(V->getType()->isPtrOrPtrVectorTy());
605 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
606 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
608 // Even though we don't look through PHI nodes, we could be called on an
609 // instruction in an unreachable block, which may be on a cycle.
610 SmallPtrSet<Value *, 4> Visited;
613 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
614 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
615 !GEP->accumulateConstantOffset(DL, Offset))
617 V = GEP->getPointerOperand();
618 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
619 V = cast<Operator>(V)->getOperand(0);
620 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
621 if (GA->isInterposable())
623 V = GA->getAliasee();
625 if (auto CS = CallSite(V))
626 if (Value *RV = CS.getReturnedArgOperand()) {
632 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
633 } while (Visited.insert(V).second);
635 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
636 if (V->getType()->isVectorTy())
637 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
642 /// Compute the constant difference between two pointer values.
643 /// If the difference is not a constant, returns zero.
644 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
646 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
647 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
649 // If LHS and RHS are not related via constant offsets to the same base
650 // value, there is nothing we can do here.
654 // Otherwise, the difference of LHS - RHS can be computed as:
656 // = (LHSOffset + Base) - (RHSOffset + Base)
657 // = LHSOffset - RHSOffset
658 return ConstantExpr::getSub(LHSOffset, RHSOffset);
661 /// Given operands for a Sub, see if we can fold the result.
662 /// If not, this returns null.
663 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
664 const SimplifyQuery &Q, unsigned MaxRecurse) {
665 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
668 // X - undef -> undef
669 // undef - X -> undef
670 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
671 return UndefValue::get(Op0->getType());
674 if (match(Op1, m_Zero()))
679 return Constant::getNullValue(Op0->getType());
681 // Is this a negation?
682 if (match(Op0, m_Zero())) {
683 // 0 - X -> 0 if the sub is NUW.
685 return Constant::getNullValue(Op0->getType());
687 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
688 if (Known.Zero.isMaxSignedValue()) {
689 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
690 // Op1 must be 0 because negating the minimum signed value is undefined.
692 return Constant::getNullValue(Op0->getType());
694 // 0 - X -> X if X is 0 or the minimum signed value.
699 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
700 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
701 Value *X = nullptr, *Y = nullptr, *Z = Op1;
702 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
703 // See if "V === Y - Z" simplifies.
704 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
705 // It does! Now see if "X + V" simplifies.
706 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
707 // It does, we successfully reassociated!
711 // See if "V === X - Z" simplifies.
712 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
713 // It does! Now see if "Y + V" simplifies.
714 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
715 // It does, we successfully reassociated!
721 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
722 // For example, X - (X + 1) -> -1
724 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
725 // See if "V === X - Y" simplifies.
726 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
727 // It does! Now see if "V - Z" simplifies.
728 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
729 // It does, we successfully reassociated!
733 // See if "V === X - Z" simplifies.
734 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
735 // It does! Now see if "V - Y" simplifies.
736 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
737 // It does, we successfully reassociated!
743 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
744 // For example, X - (X - Y) -> Y.
746 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
747 // See if "V === Z - X" simplifies.
748 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
749 // It does! Now see if "V + Y" simplifies.
750 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
751 // It does, we successfully reassociated!
756 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
757 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
758 match(Op1, m_Trunc(m_Value(Y))))
759 if (X->getType() == Y->getType())
760 // See if "V === X - Y" simplifies.
761 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
762 // It does! Now see if "trunc V" simplifies.
763 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
765 // It does, return the simplified "trunc V".
768 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
769 if (match(Op0, m_PtrToInt(m_Value(X))) &&
770 match(Op1, m_PtrToInt(m_Value(Y))))
771 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
772 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
775 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
776 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
779 // Threading Sub over selects and phi nodes is pointless, so don't bother.
780 // Threading over the select in "A - select(cond, B, C)" means evaluating
781 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
782 // only if B and C are equal. If B and C are equal then (since we assume
783 // that operands have already been simplified) "select(cond, B, C)" should
784 // have been simplified to the common value of B and C already. Analysing
785 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
786 // for threading over phi nodes.
791 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
792 const SimplifyQuery &Q) {
793 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
796 /// Given operands for a Mul, see if we can fold the result.
797 /// If not, this returns null.
798 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
799 unsigned MaxRecurse) {
800 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
805 if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
806 return Constant::getNullValue(Op0->getType());
809 if (match(Op1, m_One()))
812 // (X / Y) * Y -> X if the division is exact.
814 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
815 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
819 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
820 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
823 // Try some generic simplifications for associative operations.
824 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
828 // Mul distributes over Add. Try some generic simplifications based on this.
829 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
833 // If the operation is with the result of a select instruction, check whether
834 // operating on either branch of the select always yields the same value.
835 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
836 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
840 // If the operation is with the result of a phi instruction, check whether
841 // operating on all incoming values of the phi always yields the same value.
842 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
843 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
850 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
851 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
854 /// Check for common or similar folds of integer division or integer remainder.
855 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
856 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
857 Type *Ty = Op0->getType();
859 // X / undef -> undef
860 // X % undef -> undef
861 if (match(Op1, m_Undef()))
866 // We don't need to preserve faults!
867 if (match(Op1, m_Zero()))
868 return UndefValue::get(Ty);
870 // If any element of a constant divisor vector is zero or undef, the whole op
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() || isa<UndefValue>(Elt)))
878 return UndefValue::get(Ty);
884 if (match(Op0, m_Undef()))
885 return Constant::getNullValue(Ty);
889 if (match(Op0, m_Zero()))
890 return Constant::getNullValue(Op0->getType());
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 // (X << Y) % X -> 0
1044 if ((Opcode == Instruction::SRem &&
1045 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1046 (Opcode == Instruction::URem &&
1047 match(Op0, m_NUWShl(m_Specific(Op1), m_Value()))))
1048 return Constant::getNullValue(Op0->getType());
1050 // If the operation is with the result of a select instruction, check whether
1051 // operating on either branch of the select always yields the same value.
1052 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1053 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1056 // If the operation is with the result of a phi instruction, check whether
1057 // operating on all incoming values of the phi always yields the same value.
1058 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1059 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1062 // If X / Y == 0, then X % Y == X.
1063 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1069 /// Given operands for an SDiv, see if we can fold the result.
1070 /// If not, this returns null.
1071 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1072 unsigned MaxRecurse) {
1073 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1076 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1077 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1080 /// Given operands for a UDiv, see if we can fold the result.
1081 /// If not, this returns null.
1082 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1083 unsigned MaxRecurse) {
1084 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1087 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1088 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1091 /// Given operands for an SRem, see if we can fold the result.
1092 /// If not, this returns null.
1093 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1094 unsigned MaxRecurse) {
1095 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1098 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1099 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1102 /// Given operands for a URem, see if we can fold the result.
1103 /// If not, this returns null.
1104 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1105 unsigned MaxRecurse) {
1106 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1109 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1110 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1113 /// Returns true if a shift by \c Amount always yields undef.
1114 static bool isUndefShift(Value *Amount) {
1115 Constant *C = dyn_cast<Constant>(Amount);
1119 // X shift by undef -> undef because it may shift by the bitwidth.
1120 if (isa<UndefValue>(C))
1123 // Shifting by the bitwidth or more is undefined.
1124 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1125 if (CI->getValue().getLimitedValue() >=
1126 CI->getType()->getScalarSizeInBits())
1129 // If all lanes of a vector shift are undefined the whole shift is.
1130 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1131 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1132 if (!isUndefShift(C->getAggregateElement(I)))
1140 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1141 /// If not, this returns null.
1142 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1143 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1144 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1147 // 0 shift by X -> 0
1148 if (match(Op0, m_Zero()))
1149 return Constant::getNullValue(Op0->getType());
1151 // X shift by 0 -> X
1152 if (match(Op1, m_Zero()))
1155 // Fold undefined shifts.
1156 if (isUndefShift(Op1))
1157 return UndefValue::get(Op0->getType());
1159 // If the operation is with the result of a select instruction, check whether
1160 // operating on either branch of the select always yields the same value.
1161 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1162 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1165 // If the operation is with the result of a phi instruction, check whether
1166 // operating on all incoming values of the phi always yields the same value.
1167 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1168 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1171 // If any bits in the shift amount make that value greater than or equal to
1172 // the number of bits in the type, the shift is undefined.
1173 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1174 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1175 return UndefValue::get(Op0->getType());
1177 // If all valid bits in the shift amount are known zero, the first operand is
1179 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1180 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1186 /// Given operands for an Shl, LShr or AShr, see if we can
1187 /// fold the result. If not, this returns null.
1188 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1189 Value *Op1, bool isExact, const SimplifyQuery &Q,
1190 unsigned MaxRecurse) {
1191 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1196 return Constant::getNullValue(Op0->getType());
1199 // undef >> X -> undef (if it's exact)
1200 if (match(Op0, m_Undef()))
1201 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1203 // The low bit cannot be shifted out of an exact shift if it is set.
1205 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1206 if (Op0Known.One[0])
1213 /// Given operands for an Shl, see if we can fold the result.
1214 /// If not, this returns null.
1215 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1216 const SimplifyQuery &Q, unsigned MaxRecurse) {
1217 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1221 // undef << X -> undef if (if it's NSW/NUW)
1222 if (match(Op0, m_Undef()))
1223 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1225 // (X >> A) << A -> X
1227 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1230 // shl nuw i8 C, %x -> C iff C has sign bit set.
1231 if (isNUW && match(Op0, m_Negative()))
1233 // NOTE: could use computeKnownBits() / LazyValueInfo,
1234 // but the cost-benefit analysis suggests it isn't worth it.
1239 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1240 const SimplifyQuery &Q) {
1241 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1244 /// Given operands for an LShr, see if we can fold the result.
1245 /// If not, this returns null.
1246 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1247 const SimplifyQuery &Q, unsigned MaxRecurse) {
1248 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1252 // (X << A) >> A -> X
1254 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1260 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1261 const SimplifyQuery &Q) {
1262 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1265 /// Given operands for an AShr, see if we can fold the result.
1266 /// If not, this returns null.
1267 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1268 const SimplifyQuery &Q, unsigned MaxRecurse) {
1269 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1273 // all ones >>a X -> -1
1274 // Do not return Op0 because it may contain undef elements if it's a vector.
1275 if (match(Op0, m_AllOnes()))
1276 return Constant::getAllOnesValue(Op0->getType());
1278 // (X << A) >> A -> X
1280 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1283 // Arithmetic shifting an all-sign-bit value is a no-op.
1284 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1285 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1291 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1292 const SimplifyQuery &Q) {
1293 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1296 /// Commuted variants are assumed to be handled by calling this function again
1297 /// with the parameters swapped.
1298 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1299 ICmpInst *UnsignedICmp, bool IsAnd) {
1302 ICmpInst::Predicate EqPred;
1303 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1304 !ICmpInst::isEquality(EqPred))
1307 ICmpInst::Predicate UnsignedPred;
1308 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1309 ICmpInst::isUnsigned(UnsignedPred))
1311 else if (match(UnsignedICmp,
1312 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1313 ICmpInst::isUnsigned(UnsignedPred))
1314 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1318 // X < Y && Y != 0 --> X < Y
1319 // X < Y || Y != 0 --> Y != 0
1320 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1321 return IsAnd ? UnsignedICmp : ZeroICmp;
1323 // X >= Y || Y != 0 --> true
1324 // X >= Y || Y == 0 --> X >= Y
1325 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1326 if (EqPred == ICmpInst::ICMP_NE)
1327 return getTrue(UnsignedICmp->getType());
1328 return UnsignedICmp;
1331 // X < Y && Y == 0 --> false
1332 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1334 return getFalse(UnsignedICmp->getType());
1339 /// Commuted variants are assumed to be handled by calling this function again
1340 /// with the parameters swapped.
1341 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1342 ICmpInst::Predicate Pred0, Pred1;
1344 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1345 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1348 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1349 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1350 // can eliminate Op1 from this 'and'.
1351 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1354 // Check for any combination of predicates that are guaranteed to be disjoint.
1355 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1356 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1357 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1358 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1359 return getFalse(Op0->getType());
1364 /// Commuted variants are assumed to be handled by calling this function again
1365 /// with the parameters swapped.
1366 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1367 ICmpInst::Predicate Pred0, Pred1;
1369 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1370 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1373 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1374 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1375 // can eliminate Op0 from this 'or'.
1376 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1379 // Check for any combination of predicates that cover the entire range of
1381 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1382 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1383 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1384 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1385 return getTrue(Op0->getType());
1390 /// Test if a pair of compares with a shared operand and 2 constants has an
1391 /// empty set intersection, full set union, or if one compare is a superset of
1393 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1395 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1396 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1399 const APInt *C0, *C1;
1400 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1401 !match(Cmp1->getOperand(1), m_APInt(C1)))
1404 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1405 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1407 // For and-of-compares, check if the intersection is empty:
1408 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1409 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1410 return getFalse(Cmp0->getType());
1412 // For or-of-compares, check if the union is full:
1413 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1414 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1415 return getTrue(Cmp0->getType());
1417 // Is one range a superset of the other?
1418 // If this is and-of-compares, take the smaller set:
1419 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1420 // If this is or-of-compares, take the larger set:
1421 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1422 if (Range0.contains(Range1))
1423 return IsAnd ? Cmp1 : Cmp0;
1424 if (Range1.contains(Range0))
1425 return IsAnd ? Cmp0 : Cmp1;
1430 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1432 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1433 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1434 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1437 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1440 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1441 Value *X = Cmp0->getOperand(0);
1442 Value *Y = Cmp1->getOperand(0);
1444 // If one of the compares is a masked version of a (not) null check, then
1445 // that compare implies the other, so we eliminate the other. Optionally, look
1446 // through a pointer-to-int cast to match a null check of a pointer type.
1448 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1449 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1450 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1451 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1452 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1453 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1456 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1457 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1458 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1459 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1460 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1461 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1467 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1468 // (icmp (add V, C0), C1) & (icmp V, C0)
1469 ICmpInst::Predicate Pred0, Pred1;
1470 const APInt *C0, *C1;
1472 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1475 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1478 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1479 if (AddInst->getOperand(1) != Op1->getOperand(1))
1482 Type *ITy = Op0->getType();
1483 bool isNSW = AddInst->hasNoSignedWrap();
1484 bool isNUW = AddInst->hasNoUnsignedWrap();
1486 const APInt Delta = *C1 - *C0;
1487 if (C0->isStrictlyPositive()) {
1489 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1490 return getFalse(ITy);
1491 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1492 return getFalse(ITy);
1495 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1496 return getFalse(ITy);
1497 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1498 return getFalse(ITy);
1501 if (C0->getBoolValue() && isNUW) {
1503 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1504 return getFalse(ITy);
1506 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1507 return getFalse(ITy);
1513 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1514 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1516 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1519 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1521 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1524 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1527 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1530 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1532 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1538 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1539 // (icmp (add V, C0), C1) | (icmp V, C0)
1540 ICmpInst::Predicate Pred0, Pred1;
1541 const APInt *C0, *C1;
1543 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1546 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1549 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1550 if (AddInst->getOperand(1) != Op1->getOperand(1))
1553 Type *ITy = Op0->getType();
1554 bool isNSW = AddInst->hasNoSignedWrap();
1555 bool isNUW = AddInst->hasNoUnsignedWrap();
1557 const APInt Delta = *C1 - *C0;
1558 if (C0->isStrictlyPositive()) {
1560 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1561 return getTrue(ITy);
1562 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1563 return getTrue(ITy);
1566 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1567 return getTrue(ITy);
1568 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1569 return getTrue(ITy);
1572 if (C0->getBoolValue() && isNUW) {
1574 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1575 return getTrue(ITy);
1577 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1578 return getTrue(ITy);
1584 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1585 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1587 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1590 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1592 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1595 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1598 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1601 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1603 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1609 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1610 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1611 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1612 if (LHS0->getType() != RHS0->getType())
1615 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1616 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1617 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1618 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1619 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1620 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1621 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1622 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1623 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1624 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1625 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1626 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1627 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1630 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1631 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1632 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1633 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1634 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1635 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1636 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1637 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1638 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1639 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1646 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1647 // Look through casts of the 'and' operands to find compares.
1648 auto *Cast0 = dyn_cast<CastInst>(Op0);
1649 auto *Cast1 = dyn_cast<CastInst>(Op1);
1650 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1651 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1652 Op0 = Cast0->getOperand(0);
1653 Op1 = Cast1->getOperand(0);
1657 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1658 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1660 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1661 simplifyOrOfICmps(ICmp0, ICmp1);
1663 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1664 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1666 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1673 // If we looked through casts, we can only handle a constant simplification
1674 // because we are not allowed to create a cast instruction here.
1675 if (auto *C = dyn_cast<Constant>(V))
1676 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1681 /// Given operands for an And, see if we can fold the result.
1682 /// If not, this returns null.
1683 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1684 unsigned MaxRecurse) {
1685 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1689 if (match(Op1, m_Undef()))
1690 return Constant::getNullValue(Op0->getType());
1697 if (match(Op1, m_Zero()))
1698 return Constant::getNullValue(Op0->getType());
1701 if (match(Op1, m_AllOnes()))
1704 // A & ~A = ~A & A = 0
1705 if (match(Op0, m_Not(m_Specific(Op1))) ||
1706 match(Op1, m_Not(m_Specific(Op0))))
1707 return Constant::getNullValue(Op0->getType());
1710 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1714 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1717 // A mask that only clears known zeros of a shifted value is a no-op.
1721 if (match(Op1, m_APInt(Mask))) {
1722 // If all bits in the inverted and shifted mask are clear:
1723 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1724 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1725 (~(*Mask)).lshr(*ShAmt).isNullValue())
1728 // If all bits in the inverted and shifted mask are clear:
1729 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1730 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1731 (~(*Mask)).shl(*ShAmt).isNullValue())
1735 // A & (-A) = A if A is a power of two or zero.
1736 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1737 match(Op1, m_Neg(m_Specific(Op0)))) {
1738 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1741 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1746 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1749 // Try some generic simplifications for associative operations.
1750 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1754 // And distributes over Or. Try some generic simplifications based on this.
1755 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1759 // And distributes over Xor. Try some generic simplifications based on this.
1760 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1764 // If the operation is with the result of a select instruction, check whether
1765 // operating on either branch of the select always yields the same value.
1766 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1767 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1771 // If the operation is with the result of a phi instruction, check whether
1772 // operating on all incoming values of the phi always yields the same value.
1773 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1774 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1781 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1782 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1785 /// Given operands for an Or, see if we can fold the result.
1786 /// If not, this returns null.
1787 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1788 unsigned MaxRecurse) {
1789 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1794 // Do not return Op1 because it may contain undef elements if it's a vector.
1795 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1796 return Constant::getAllOnesValue(Op0->getType());
1800 if (Op0 == Op1 || match(Op1, m_Zero()))
1803 // A | ~A = ~A | A = -1
1804 if (match(Op0, m_Not(m_Specific(Op1))) ||
1805 match(Op1, m_Not(m_Specific(Op0))))
1806 return Constant::getAllOnesValue(Op0->getType());
1809 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1813 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1816 // ~(A & ?) | A = -1
1817 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1818 return Constant::getAllOnesValue(Op1->getType());
1820 // A | ~(A & ?) = -1
1821 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1822 return Constant::getAllOnesValue(Op0->getType());
1825 // (A & ~B) | (A ^ B) -> (A ^ B)
1826 // (~B & A) | (A ^ B) -> (A ^ B)
1827 // (A & ~B) | (B ^ A) -> (B ^ A)
1828 // (~B & A) | (B ^ A) -> (B ^ A)
1829 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1830 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1831 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1834 // Commute the 'or' operands.
1835 // (A ^ B) | (A & ~B) -> (A ^ B)
1836 // (A ^ B) | (~B & A) -> (A ^ B)
1837 // (B ^ A) | (A & ~B) -> (B ^ A)
1838 // (B ^ A) | (~B & A) -> (B ^ A)
1839 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1840 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1841 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1844 // (A & B) | (~A ^ B) -> (~A ^ B)
1845 // (B & A) | (~A ^ B) -> (~A ^ B)
1846 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1847 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1848 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1849 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1850 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1853 // (~A ^ B) | (A & B) -> (~A ^ B)
1854 // (~A ^ B) | (B & A) -> (~A ^ B)
1855 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1856 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1857 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1858 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1859 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1862 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1865 // Try some generic simplifications for associative operations.
1866 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1870 // Or distributes over And. Try some generic simplifications based on this.
1871 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1875 // If the operation is with the result of a select instruction, check whether
1876 // operating on either branch of the select always yields the same value.
1877 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1878 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1882 // (A & C1)|(B & C2)
1883 const APInt *C1, *C2;
1884 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1885 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1887 // (A & C1)|(B & C2)
1888 // If we have: ((V + N) & C1) | (V & C2)
1889 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1890 // replace with V+N.
1892 if (C2->isMask() && // C2 == 0+1+
1893 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1894 // Add commutes, try both ways.
1895 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1898 // Or commutes, try both ways.
1900 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1901 // Add commutes, try both ways.
1902 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1908 // If the operation is with the result of a phi instruction, check whether
1909 // operating on all incoming values of the phi always yields the same value.
1910 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1911 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1917 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1918 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1921 /// Given operands for a Xor, see if we can fold the result.
1922 /// If not, this returns null.
1923 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1924 unsigned MaxRecurse) {
1925 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1928 // A ^ undef -> undef
1929 if (match(Op1, m_Undef()))
1933 if (match(Op1, m_Zero()))
1938 return Constant::getNullValue(Op0->getType());
1940 // A ^ ~A = ~A ^ A = -1
1941 if (match(Op0, m_Not(m_Specific(Op1))) ||
1942 match(Op1, m_Not(m_Specific(Op0))))
1943 return Constant::getAllOnesValue(Op0->getType());
1945 // Try some generic simplifications for associative operations.
1946 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1950 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1951 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1952 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1953 // only if B and C are equal. If B and C are equal then (since we assume
1954 // that operands have already been simplified) "select(cond, B, C)" should
1955 // have been simplified to the common value of B and C already. Analysing
1956 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1957 // for threading over phi nodes.
1962 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1963 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1967 static Type *GetCompareTy(Value *Op) {
1968 return CmpInst::makeCmpResultType(Op->getType());
1971 /// Rummage around inside V looking for something equivalent to the comparison
1972 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1973 /// Helper function for analyzing max/min idioms.
1974 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1975 Value *LHS, Value *RHS) {
1976 SelectInst *SI = dyn_cast<SelectInst>(V);
1979 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1982 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1983 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1985 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1986 LHS == CmpRHS && RHS == CmpLHS)
1991 // A significant optimization not implemented here is assuming that alloca
1992 // addresses are not equal to incoming argument values. They don't *alias*,
1993 // as we say, but that doesn't mean they aren't equal, so we take a
1994 // conservative approach.
1996 // This is inspired in part by C++11 5.10p1:
1997 // "Two pointers of the same type compare equal if and only if they are both
1998 // null, both point to the same function, or both represent the same
2001 // This is pretty permissive.
2003 // It's also partly due to C11 6.5.9p6:
2004 // "Two pointers compare equal if and only if both are null pointers, both are
2005 // pointers to the same object (including a pointer to an object and a
2006 // subobject at its beginning) or function, both are pointers to one past the
2007 // last element of the same array object, or one is a pointer to one past the
2008 // end of one array object and the other is a pointer to the start of a
2009 // different array object that happens to immediately follow the first array
2010 // object in the address space.)
2012 // C11's version is more restrictive, however there's no reason why an argument
2013 // couldn't be a one-past-the-end value for a stack object in the caller and be
2014 // equal to the beginning of a stack object in the callee.
2016 // If the C and C++ standards are ever made sufficiently restrictive in this
2017 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2018 // this optimization.
2020 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2021 const DominatorTree *DT, CmpInst::Predicate Pred,
2022 AssumptionCache *AC, const Instruction *CxtI,
2023 Value *LHS, Value *RHS) {
2024 // First, skip past any trivial no-ops.
2025 LHS = LHS->stripPointerCasts();
2026 RHS = RHS->stripPointerCasts();
2028 // A non-null pointer is not equal to a null pointer.
2029 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2030 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2031 return ConstantInt::get(GetCompareTy(LHS),
2032 !CmpInst::isTrueWhenEqual(Pred));
2034 // We can only fold certain predicates on pointer comparisons.
2039 // Equality comaprisons are easy to fold.
2040 case CmpInst::ICMP_EQ:
2041 case CmpInst::ICMP_NE:
2044 // We can only handle unsigned relational comparisons because 'inbounds' on
2045 // a GEP only protects against unsigned wrapping.
2046 case CmpInst::ICMP_UGT:
2047 case CmpInst::ICMP_UGE:
2048 case CmpInst::ICMP_ULT:
2049 case CmpInst::ICMP_ULE:
2050 // However, we have to switch them to their signed variants to handle
2051 // negative indices from the base pointer.
2052 Pred = ICmpInst::getSignedPredicate(Pred);
2056 // Strip off any constant offsets so that we can reason about them.
2057 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2058 // here and compare base addresses like AliasAnalysis does, however there are
2059 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2060 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2061 // doesn't need to guarantee pointer inequality when it says NoAlias.
2062 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2063 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2065 // If LHS and RHS are related via constant offsets to the same base
2066 // value, we can replace it with an icmp which just compares the offsets.
2068 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2070 // Various optimizations for (in)equality comparisons.
2071 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2072 // Different non-empty allocations that exist at the same time have
2073 // different addresses (if the program can tell). Global variables always
2074 // exist, so they always exist during the lifetime of each other and all
2075 // allocas. Two different allocas usually have different addresses...
2077 // However, if there's an @llvm.stackrestore dynamically in between two
2078 // allocas, they may have the same address. It's tempting to reduce the
2079 // scope of the problem by only looking at *static* allocas here. That would
2080 // cover the majority of allocas while significantly reducing the likelihood
2081 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2082 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2083 // an entry block. Also, if we have a block that's not attached to a
2084 // function, we can't tell if it's "static" under the current definition.
2085 // Theoretically, this problem could be fixed by creating a new kind of
2086 // instruction kind specifically for static allocas. Such a new instruction
2087 // could be required to be at the top of the entry block, thus preventing it
2088 // from being subject to a @llvm.stackrestore. Instcombine could even
2089 // convert regular allocas into these special allocas. It'd be nifty.
2090 // However, until then, this problem remains open.
2092 // So, we'll assume that two non-empty allocas have different addresses
2095 // With all that, if the offsets are within the bounds of their allocations
2096 // (and not one-past-the-end! so we can't use inbounds!), and their
2097 // allocations aren't the same, the pointers are not equal.
2099 // Note that it's not necessary to check for LHS being a global variable
2100 // address, due to canonicalization and constant folding.
2101 if (isa<AllocaInst>(LHS) &&
2102 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2103 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2104 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2105 uint64_t LHSSize, RHSSize;
2106 if (LHSOffsetCI && RHSOffsetCI &&
2107 getObjectSize(LHS, LHSSize, DL, TLI) &&
2108 getObjectSize(RHS, RHSSize, DL, TLI)) {
2109 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2110 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2111 if (!LHSOffsetValue.isNegative() &&
2112 !RHSOffsetValue.isNegative() &&
2113 LHSOffsetValue.ult(LHSSize) &&
2114 RHSOffsetValue.ult(RHSSize)) {
2115 return ConstantInt::get(GetCompareTy(LHS),
2116 !CmpInst::isTrueWhenEqual(Pred));
2120 // Repeat the above check but this time without depending on DataLayout
2121 // or being able to compute a precise size.
2122 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2123 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2124 LHSOffset->isNullValue() &&
2125 RHSOffset->isNullValue())
2126 return ConstantInt::get(GetCompareTy(LHS),
2127 !CmpInst::isTrueWhenEqual(Pred));
2130 // Even if an non-inbounds GEP occurs along the path we can still optimize
2131 // equality comparisons concerning the result. We avoid walking the whole
2132 // chain again by starting where the last calls to
2133 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2134 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2135 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2137 return ConstantExpr::getICmp(Pred,
2138 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2139 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2141 // If one side of the equality comparison must come from a noalias call
2142 // (meaning a system memory allocation function), and the other side must
2143 // come from a pointer that cannot overlap with dynamically-allocated
2144 // memory within the lifetime of the current function (allocas, byval
2145 // arguments, globals), then determine the comparison result here.
2146 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2147 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2148 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2150 // Is the set of underlying objects all noalias calls?
2151 auto IsNAC = [](ArrayRef<Value *> Objects) {
2152 return all_of(Objects, isNoAliasCall);
2155 // Is the set of underlying objects all things which must be disjoint from
2156 // noalias calls. For allocas, we consider only static ones (dynamic
2157 // allocas might be transformed into calls to malloc not simultaneously
2158 // live with the compared-to allocation). For globals, we exclude symbols
2159 // that might be resolve lazily to symbols in another dynamically-loaded
2160 // library (and, thus, could be malloc'ed by the implementation).
2161 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2162 return all_of(Objects, [](Value *V) {
2163 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2164 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2165 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2166 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2167 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2168 !GV->isThreadLocal();
2169 if (const Argument *A = dyn_cast<Argument>(V))
2170 return A->hasByValAttr();
2175 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2176 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2177 return ConstantInt::get(GetCompareTy(LHS),
2178 !CmpInst::isTrueWhenEqual(Pred));
2180 // Fold comparisons for non-escaping pointer even if the allocation call
2181 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2182 // dynamic allocation call could be either of the operands.
2183 Value *MI = nullptr;
2184 if (isAllocLikeFn(LHS, TLI) &&
2185 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2187 else if (isAllocLikeFn(RHS, TLI) &&
2188 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2190 // FIXME: We should also fold the compare when the pointer escapes, but the
2191 // compare dominates the pointer escape
2192 if (MI && !PointerMayBeCaptured(MI, true, true))
2193 return ConstantInt::get(GetCompareTy(LHS),
2194 CmpInst::isFalseWhenEqual(Pred));
2201 /// Fold an icmp when its operands have i1 scalar type.
2202 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2203 Value *RHS, const SimplifyQuery &Q) {
2204 Type *ITy = GetCompareTy(LHS); // The return type.
2205 Type *OpTy = LHS->getType(); // The operand type.
2206 if (!OpTy->isIntOrIntVectorTy(1))
2209 // A boolean compared to true/false can be simplified in 14 out of the 20
2210 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2211 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2212 if (match(RHS, m_Zero())) {
2214 case CmpInst::ICMP_NE: // X != 0 -> X
2215 case CmpInst::ICMP_UGT: // X >u 0 -> X
2216 case CmpInst::ICMP_SLT: // X <s 0 -> X
2219 case CmpInst::ICMP_ULT: // X <u 0 -> false
2220 case CmpInst::ICMP_SGT: // X >s 0 -> false
2221 return getFalse(ITy);
2223 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2224 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2225 return getTrue(ITy);
2229 } else if (match(RHS, m_One())) {
2231 case CmpInst::ICMP_EQ: // X == 1 -> X
2232 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2233 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2236 case CmpInst::ICMP_UGT: // X >u 1 -> false
2237 case CmpInst::ICMP_SLT: // X <s -1 -> false
2238 return getFalse(ITy);
2240 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2241 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2242 return getTrue(ITy);
2251 case ICmpInst::ICMP_UGE:
2252 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2253 return getTrue(ITy);
2255 case ICmpInst::ICMP_SGE:
2256 /// For signed comparison, the values for an i1 are 0 and -1
2257 /// respectively. This maps into a truth table of:
2258 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2259 /// 0 | 0 | 1 (0 >= 0) | 1
2260 /// 0 | 1 | 1 (0 >= -1) | 1
2261 /// 1 | 0 | 0 (-1 >= 0) | 0
2262 /// 1 | 1 | 1 (-1 >= -1) | 1
2263 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2264 return getTrue(ITy);
2266 case ICmpInst::ICMP_ULE:
2267 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2268 return getTrue(ITy);
2275 /// Try hard to fold icmp with zero RHS because this is a common case.
2276 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2277 Value *RHS, const SimplifyQuery &Q) {
2278 if (!match(RHS, m_Zero()))
2281 Type *ITy = GetCompareTy(LHS); // The return type.
2284 llvm_unreachable("Unknown ICmp predicate!");
2285 case ICmpInst::ICMP_ULT:
2286 return getFalse(ITy);
2287 case ICmpInst::ICMP_UGE:
2288 return getTrue(ITy);
2289 case ICmpInst::ICMP_EQ:
2290 case ICmpInst::ICMP_ULE:
2291 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2292 return getFalse(ITy);
2294 case ICmpInst::ICMP_NE:
2295 case ICmpInst::ICMP_UGT:
2296 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2297 return getTrue(ITy);
2299 case ICmpInst::ICMP_SLT: {
2300 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2301 if (LHSKnown.isNegative())
2302 return getTrue(ITy);
2303 if (LHSKnown.isNonNegative())
2304 return getFalse(ITy);
2307 case ICmpInst::ICMP_SLE: {
2308 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2309 if (LHSKnown.isNegative())
2310 return getTrue(ITy);
2311 if (LHSKnown.isNonNegative() &&
2312 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2313 return getFalse(ITy);
2316 case ICmpInst::ICMP_SGE: {
2317 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2318 if (LHSKnown.isNegative())
2319 return getFalse(ITy);
2320 if (LHSKnown.isNonNegative())
2321 return getTrue(ITy);
2324 case ICmpInst::ICMP_SGT: {
2325 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2326 if (LHSKnown.isNegative())
2327 return getFalse(ITy);
2328 if (LHSKnown.isNonNegative() &&
2329 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2330 return getTrue(ITy);
2338 /// Many binary operators with a constant operand have an easy-to-compute
2339 /// range of outputs. This can be used to fold a comparison to always true or
2341 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2342 unsigned Width = Lower.getBitWidth();
2344 switch (BO.getOpcode()) {
2345 case Instruction::Add:
2346 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2347 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2348 if (BO.hasNoUnsignedWrap()) {
2349 // 'add nuw x, C' produces [C, UINT_MAX].
2351 } else if (BO.hasNoSignedWrap()) {
2352 if (C->isNegative()) {
2353 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2354 Lower = APInt::getSignedMinValue(Width);
2355 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2357 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2358 Lower = APInt::getSignedMinValue(Width) + *C;
2359 Upper = APInt::getSignedMaxValue(Width) + 1;
2365 case Instruction::And:
2366 if (match(BO.getOperand(1), m_APInt(C)))
2367 // 'and x, C' produces [0, C].
2371 case Instruction::Or:
2372 if (match(BO.getOperand(1), m_APInt(C)))
2373 // 'or x, C' produces [C, UINT_MAX].
2377 case Instruction::AShr:
2378 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2379 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2380 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2381 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2382 } else if (match(BO.getOperand(0), m_APInt(C))) {
2383 unsigned ShiftAmount = Width - 1;
2384 if (!C->isNullValue() && BO.isExact())
2385 ShiftAmount = C->countTrailingZeros();
2386 if (C->isNegative()) {
2387 // 'ashr C, x' produces [C, C >> (Width-1)]
2389 Upper = C->ashr(ShiftAmount) + 1;
2391 // 'ashr C, x' produces [C >> (Width-1), C]
2392 Lower = C->ashr(ShiftAmount);
2398 case Instruction::LShr:
2399 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2400 // 'lshr x, C' produces [0, UINT_MAX >> C].
2401 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2402 } else if (match(BO.getOperand(0), m_APInt(C))) {
2403 // 'lshr C, x' produces [C >> (Width-1), C].
2404 unsigned ShiftAmount = Width - 1;
2405 if (!C->isNullValue() && BO.isExact())
2406 ShiftAmount = C->countTrailingZeros();
2407 Lower = C->lshr(ShiftAmount);
2412 case Instruction::Shl:
2413 if (match(BO.getOperand(0), m_APInt(C))) {
2414 if (BO.hasNoUnsignedWrap()) {
2415 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2417 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2418 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2419 if (C->isNegative()) {
2420 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2421 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2422 Lower = C->shl(ShiftAmount);
2425 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2426 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2428 Upper = C->shl(ShiftAmount) + 1;
2434 case Instruction::SDiv:
2435 if (match(BO.getOperand(1), m_APInt(C))) {
2436 APInt IntMin = APInt::getSignedMinValue(Width);
2437 APInt IntMax = APInt::getSignedMaxValue(Width);
2438 if (C->isAllOnesValue()) {
2439 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2440 // where C != -1 and C != 0 and C != 1
2443 } else if (C->countLeadingZeros() < Width - 1) {
2444 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2445 // where C != -1 and C != 0 and C != 1
2446 Lower = IntMin.sdiv(*C);
2447 Upper = IntMax.sdiv(*C);
2448 if (Lower.sgt(Upper))
2449 std::swap(Lower, Upper);
2451 assert(Upper != Lower && "Upper part of range has wrapped!");
2453 } else if (match(BO.getOperand(0), m_APInt(C))) {
2454 if (C->isMinSignedValue()) {
2455 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2457 Upper = Lower.lshr(1) + 1;
2459 // 'sdiv C, x' produces [-|C|, |C|].
2460 Upper = C->abs() + 1;
2461 Lower = (-Upper) + 1;
2466 case Instruction::UDiv:
2467 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2468 // 'udiv x, C' produces [0, UINT_MAX / C].
2469 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2470 } else if (match(BO.getOperand(0), m_APInt(C))) {
2471 // 'udiv C, x' produces [0, C].
2476 case Instruction::SRem:
2477 if (match(BO.getOperand(1), m_APInt(C))) {
2478 // 'srem x, C' produces (-|C|, |C|).
2480 Lower = (-Upper) + 1;
2484 case Instruction::URem:
2485 if (match(BO.getOperand(1), m_APInt(C)))
2486 // 'urem x, C' produces [0, C).
2495 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2497 Type *ITy = GetCompareTy(RHS); // The return type.
2500 // Sign-bit checks can be optimized to true/false after unsigned
2501 // floating-point casts:
2502 // icmp slt (bitcast (uitofp X)), 0 --> false
2503 // icmp sgt (bitcast (uitofp X)), -1 --> true
2504 if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2505 if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2506 return ConstantInt::getFalse(ITy);
2507 if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2508 return ConstantInt::getTrue(ITy);
2512 if (!match(RHS, m_APInt(C)))
2515 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2516 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2517 if (RHS_CR.isEmptySet())
2518 return ConstantInt::getFalse(ITy);
2519 if (RHS_CR.isFullSet())
2520 return ConstantInt::getTrue(ITy);
2522 // Find the range of possible values for binary operators.
2523 unsigned Width = C->getBitWidth();
2524 APInt Lower = APInt(Width, 0);
2525 APInt Upper = APInt(Width, 0);
2526 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2527 setLimitsForBinOp(*BO, Lower, Upper);
2529 ConstantRange LHS_CR =
2530 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2532 if (auto *I = dyn_cast<Instruction>(LHS))
2533 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2534 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2536 if (!LHS_CR.isFullSet()) {
2537 if (RHS_CR.contains(LHS_CR))
2538 return ConstantInt::getTrue(ITy);
2539 if (RHS_CR.inverse().contains(LHS_CR))
2540 return ConstantInt::getFalse(ITy);
2546 /// TODO: A large part of this logic is duplicated in InstCombine's
2547 /// foldICmpBinOp(). We should be able to share that and avoid the code
2549 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2550 Value *RHS, const SimplifyQuery &Q,
2551 unsigned MaxRecurse) {
2552 Type *ITy = GetCompareTy(LHS); // The return type.
2554 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2555 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2556 if (MaxRecurse && (LBO || RBO)) {
2557 // Analyze the case when either LHS or RHS is an add instruction.
2558 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2559 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2560 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2561 if (LBO && LBO->getOpcode() == Instruction::Add) {
2562 A = LBO->getOperand(0);
2563 B = LBO->getOperand(1);
2565 ICmpInst::isEquality(Pred) ||
2566 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2567 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2569 if (RBO && RBO->getOpcode() == Instruction::Add) {
2570 C = RBO->getOperand(0);
2571 D = RBO->getOperand(1);
2573 ICmpInst::isEquality(Pred) ||
2574 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2575 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2578 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2579 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2580 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2581 Constant::getNullValue(RHS->getType()), Q,
2585 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2586 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2588 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2589 C == LHS ? D : C, Q, MaxRecurse - 1))
2592 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2593 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2595 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2598 // C + B == C + D -> B == D
2601 } else if (A == D) {
2602 // D + B == C + D -> B == C
2605 } else if (B == C) {
2606 // A + C == C + D -> A == D
2611 // A + D == C + D -> A == C
2615 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2622 // icmp pred (or X, Y), X
2623 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2624 if (Pred == ICmpInst::ICMP_ULT)
2625 return getFalse(ITy);
2626 if (Pred == ICmpInst::ICMP_UGE)
2627 return getTrue(ITy);
2629 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2630 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2631 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2632 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2633 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2634 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2635 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2638 // icmp pred X, (or X, Y)
2639 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2640 if (Pred == ICmpInst::ICMP_ULE)
2641 return getTrue(ITy);
2642 if (Pred == ICmpInst::ICMP_UGT)
2643 return getFalse(ITy);
2645 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2646 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2647 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2648 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2649 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2650 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2651 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2656 // icmp pred (and X, Y), X
2657 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2658 if (Pred == ICmpInst::ICMP_UGT)
2659 return getFalse(ITy);
2660 if (Pred == ICmpInst::ICMP_ULE)
2661 return getTrue(ITy);
2663 // icmp pred X, (and X, Y)
2664 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2665 if (Pred == ICmpInst::ICMP_UGE)
2666 return getTrue(ITy);
2667 if (Pred == ICmpInst::ICMP_ULT)
2668 return getFalse(ITy);
2671 // 0 - (zext X) pred C
2672 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2673 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2674 if (RHSC->getValue().isStrictlyPositive()) {
2675 if (Pred == ICmpInst::ICMP_SLT)
2676 return ConstantInt::getTrue(RHSC->getContext());
2677 if (Pred == ICmpInst::ICMP_SGE)
2678 return ConstantInt::getFalse(RHSC->getContext());
2679 if (Pred == ICmpInst::ICMP_EQ)
2680 return ConstantInt::getFalse(RHSC->getContext());
2681 if (Pred == ICmpInst::ICMP_NE)
2682 return ConstantInt::getTrue(RHSC->getContext());
2684 if (RHSC->getValue().isNonNegative()) {
2685 if (Pred == ICmpInst::ICMP_SLE)
2686 return ConstantInt::getTrue(RHSC->getContext());
2687 if (Pred == ICmpInst::ICMP_SGT)
2688 return ConstantInt::getFalse(RHSC->getContext());
2693 // icmp pred (urem X, Y), Y
2694 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2698 case ICmpInst::ICMP_SGT:
2699 case ICmpInst::ICMP_SGE: {
2700 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2701 if (!Known.isNonNegative())
2705 case ICmpInst::ICMP_EQ:
2706 case ICmpInst::ICMP_UGT:
2707 case ICmpInst::ICMP_UGE:
2708 return getFalse(ITy);
2709 case ICmpInst::ICMP_SLT:
2710 case ICmpInst::ICMP_SLE: {
2711 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2712 if (!Known.isNonNegative())
2716 case ICmpInst::ICMP_NE:
2717 case ICmpInst::ICMP_ULT:
2718 case ICmpInst::ICMP_ULE:
2719 return getTrue(ITy);
2723 // icmp pred X, (urem Y, X)
2724 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2728 case ICmpInst::ICMP_SGT:
2729 case ICmpInst::ICMP_SGE: {
2730 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2731 if (!Known.isNonNegative())
2735 case ICmpInst::ICMP_NE:
2736 case ICmpInst::ICMP_UGT:
2737 case ICmpInst::ICMP_UGE:
2738 return getTrue(ITy);
2739 case ICmpInst::ICMP_SLT:
2740 case ICmpInst::ICMP_SLE: {
2741 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2742 if (!Known.isNonNegative())
2746 case ICmpInst::ICMP_EQ:
2747 case ICmpInst::ICMP_ULT:
2748 case ICmpInst::ICMP_ULE:
2749 return getFalse(ITy);
2755 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2756 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2757 // icmp pred (X op Y), X
2758 if (Pred == ICmpInst::ICMP_UGT)
2759 return getFalse(ITy);
2760 if (Pred == ICmpInst::ICMP_ULE)
2761 return getTrue(ITy);
2766 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2767 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2768 // icmp pred X, (X op Y)
2769 if (Pred == ICmpInst::ICMP_ULT)
2770 return getFalse(ITy);
2771 if (Pred == ICmpInst::ICMP_UGE)
2772 return getTrue(ITy);
2779 // where CI2 is a power of 2 and CI isn't
2780 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2781 const APInt *CI2Val, *CIVal = &CI->getValue();
2782 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2783 CI2Val->isPowerOf2()) {
2784 if (!CIVal->isPowerOf2()) {
2785 // CI2 << X can equal zero in some circumstances,
2786 // this simplification is unsafe if CI is zero.
2788 // We know it is safe if:
2789 // - The shift is nsw, we can't shift out the one bit.
2790 // - The shift is nuw, we can't shift out the one bit.
2793 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2794 CI2Val->isOneValue() || !CI->isZero()) {
2795 if (Pred == ICmpInst::ICMP_EQ)
2796 return ConstantInt::getFalse(RHS->getContext());
2797 if (Pred == ICmpInst::ICMP_NE)
2798 return ConstantInt::getTrue(RHS->getContext());
2801 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2802 if (Pred == ICmpInst::ICMP_UGT)
2803 return ConstantInt::getFalse(RHS->getContext());
2804 if (Pred == ICmpInst::ICMP_ULE)
2805 return ConstantInt::getTrue(RHS->getContext());
2810 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2811 LBO->getOperand(1) == RBO->getOperand(1)) {
2812 switch (LBO->getOpcode()) {
2815 case Instruction::UDiv:
2816 case Instruction::LShr:
2817 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2819 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2820 RBO->getOperand(0), Q, MaxRecurse - 1))
2823 case Instruction::SDiv:
2824 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2826 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2827 RBO->getOperand(0), Q, MaxRecurse - 1))
2830 case Instruction::AShr:
2831 if (!LBO->isExact() || !RBO->isExact())
2833 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2834 RBO->getOperand(0), Q, MaxRecurse - 1))
2837 case Instruction::Shl: {
2838 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2839 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2842 if (!NSW && ICmpInst::isSigned(Pred))
2844 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2845 RBO->getOperand(0), Q, MaxRecurse - 1))
2854 /// Simplify integer comparisons where at least one operand of the compare
2855 /// matches an integer min/max idiom.
2856 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2857 Value *RHS, const SimplifyQuery &Q,
2858 unsigned MaxRecurse) {
2859 Type *ITy = GetCompareTy(LHS); // The return type.
2861 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2862 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2864 // Signed variants on "max(a,b)>=a -> true".
2865 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2867 std::swap(A, B); // smax(A, B) pred A.
2868 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2869 // We analyze this as smax(A, B) pred A.
2871 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2872 (A == LHS || B == LHS)) {
2874 std::swap(A, B); // A pred smax(A, B).
2875 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2876 // We analyze this as smax(A, B) swapped-pred A.
2877 P = CmpInst::getSwappedPredicate(Pred);
2878 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2879 (A == RHS || B == RHS)) {
2881 std::swap(A, B); // smin(A, B) pred A.
2882 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2883 // We analyze this as smax(-A, -B) swapped-pred -A.
2884 // Note that we do not need to actually form -A or -B thanks to EqP.
2885 P = CmpInst::getSwappedPredicate(Pred);
2886 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2887 (A == LHS || B == LHS)) {
2889 std::swap(A, B); // A pred smin(A, B).
2890 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2891 // We analyze this as smax(-A, -B) pred -A.
2892 // Note that we do not need to actually form -A or -B thanks to EqP.
2895 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2896 // Cases correspond to "max(A, B) p A".
2900 case CmpInst::ICMP_EQ:
2901 case CmpInst::ICMP_SLE:
2902 // Equivalent to "A EqP B". This may be the same as the condition tested
2903 // in the max/min; if so, we can just return that.
2904 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2906 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2908 // Otherwise, see if "A EqP B" simplifies.
2910 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2913 case CmpInst::ICMP_NE:
2914 case CmpInst::ICMP_SGT: {
2915 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2916 // Equivalent to "A InvEqP B". This may be the same as the condition
2917 // tested in the max/min; if so, we can just return that.
2918 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2920 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2922 // Otherwise, see if "A InvEqP B" simplifies.
2924 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2928 case CmpInst::ICMP_SGE:
2930 return getTrue(ITy);
2931 case CmpInst::ICMP_SLT:
2933 return getFalse(ITy);
2937 // Unsigned variants on "max(a,b)>=a -> true".
2938 P = CmpInst::BAD_ICMP_PREDICATE;
2939 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2941 std::swap(A, B); // umax(A, B) pred A.
2942 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2943 // We analyze this as umax(A, B) pred A.
2945 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2946 (A == LHS || B == LHS)) {
2948 std::swap(A, B); // A pred umax(A, B).
2949 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2950 // We analyze this as umax(A, B) swapped-pred A.
2951 P = CmpInst::getSwappedPredicate(Pred);
2952 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2953 (A == RHS || B == RHS)) {
2955 std::swap(A, B); // umin(A, B) pred A.
2956 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2957 // We analyze this as umax(-A, -B) swapped-pred -A.
2958 // Note that we do not need to actually form -A or -B thanks to EqP.
2959 P = CmpInst::getSwappedPredicate(Pred);
2960 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2961 (A == LHS || B == LHS)) {
2963 std::swap(A, B); // A pred umin(A, B).
2964 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2965 // We analyze this as umax(-A, -B) pred -A.
2966 // Note that we do not need to actually form -A or -B thanks to EqP.
2969 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2970 // Cases correspond to "max(A, B) p A".
2974 case CmpInst::ICMP_EQ:
2975 case CmpInst::ICMP_ULE:
2976 // Equivalent to "A EqP B". This may be the same as the condition tested
2977 // in the max/min; if so, we can just return that.
2978 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2980 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2982 // Otherwise, see if "A EqP B" simplifies.
2984 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2987 case CmpInst::ICMP_NE:
2988 case CmpInst::ICMP_UGT: {
2989 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2990 // Equivalent to "A InvEqP B". This may be the same as the condition
2991 // tested in the max/min; if so, we can just return that.
2992 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2994 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2996 // Otherwise, see if "A InvEqP B" simplifies.
2998 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3002 case CmpInst::ICMP_UGE:
3004 return getTrue(ITy);
3005 case CmpInst::ICMP_ULT:
3007 return getFalse(ITy);
3011 // Variants on "max(x,y) >= min(x,z)".
3013 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3014 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3015 (A == C || A == D || B == C || B == D)) {
3016 // max(x, ?) pred min(x, ?).
3017 if (Pred == CmpInst::ICMP_SGE)
3019 return getTrue(ITy);
3020 if (Pred == CmpInst::ICMP_SLT)
3022 return getFalse(ITy);
3023 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3024 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3025 (A == C || A == D || B == C || B == D)) {
3026 // min(x, ?) pred max(x, ?).
3027 if (Pred == CmpInst::ICMP_SLE)
3029 return getTrue(ITy);
3030 if (Pred == CmpInst::ICMP_SGT)
3032 return getFalse(ITy);
3033 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3034 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3035 (A == C || A == D || B == C || B == D)) {
3036 // max(x, ?) pred min(x, ?).
3037 if (Pred == CmpInst::ICMP_UGE)
3039 return getTrue(ITy);
3040 if (Pred == CmpInst::ICMP_ULT)
3042 return getFalse(ITy);
3043 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3044 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3045 (A == C || A == D || B == C || B == D)) {
3046 // min(x, ?) pred max(x, ?).
3047 if (Pred == CmpInst::ICMP_ULE)
3049 return getTrue(ITy);
3050 if (Pred == CmpInst::ICMP_UGT)
3052 return getFalse(ITy);
3058 /// Given operands for an ICmpInst, see if we can fold the result.
3059 /// If not, this returns null.
3060 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3061 const SimplifyQuery &Q, unsigned MaxRecurse) {
3062 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3063 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3065 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3066 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3067 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3069 // If we have a constant, make sure it is on the RHS.
3070 std::swap(LHS, RHS);
3071 Pred = CmpInst::getSwappedPredicate(Pred);
3074 Type *ITy = GetCompareTy(LHS); // The return type.
3076 // icmp X, X -> true/false
3077 // icmp X, undef -> true/false because undef could be X.
3078 if (LHS == RHS || isa<UndefValue>(RHS))
3079 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3081 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3084 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3087 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3090 // If both operands have range metadata, use the metadata
3091 // to simplify the comparison.
3092 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3093 auto RHS_Instr = cast<Instruction>(RHS);
3094 auto LHS_Instr = cast<Instruction>(LHS);
3096 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3097 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3098 auto RHS_CR = getConstantRangeFromMetadata(
3099 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3100 auto LHS_CR = getConstantRangeFromMetadata(
3101 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3103 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3104 if (Satisfied_CR.contains(LHS_CR))
3105 return ConstantInt::getTrue(RHS->getContext());
3107 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3108 CmpInst::getInversePredicate(Pred), RHS_CR);
3109 if (InversedSatisfied_CR.contains(LHS_CR))
3110 return ConstantInt::getFalse(RHS->getContext());
3114 // Compare of cast, for example (zext X) != 0 -> X != 0
3115 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3116 Instruction *LI = cast<CastInst>(LHS);
3117 Value *SrcOp = LI->getOperand(0);
3118 Type *SrcTy = SrcOp->getType();
3119 Type *DstTy = LI->getType();
3121 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3122 // if the integer type is the same size as the pointer type.
3123 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3124 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3125 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3126 // Transfer the cast to the constant.
3127 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3128 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3131 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3132 if (RI->getOperand(0)->getType() == SrcTy)
3133 // Compare without the cast.
3134 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3140 if (isa<ZExtInst>(LHS)) {
3141 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3143 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3144 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3145 // Compare X and Y. Note that signed predicates become unsigned.
3146 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3147 SrcOp, RI->getOperand(0), Q,
3151 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3152 // too. If not, then try to deduce the result of the comparison.
3153 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3154 // Compute the constant that would happen if we truncated to SrcTy then
3155 // reextended to DstTy.
3156 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3157 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3159 // If the re-extended constant didn't change then this is effectively
3160 // also a case of comparing two zero-extended values.
3161 if (RExt == CI && MaxRecurse)
3162 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3163 SrcOp, Trunc, Q, MaxRecurse-1))
3166 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3167 // there. Use this to work out the result of the comparison.
3170 default: llvm_unreachable("Unknown ICmp predicate!");
3172 case ICmpInst::ICMP_EQ:
3173 case ICmpInst::ICMP_UGT:
3174 case ICmpInst::ICMP_UGE:
3175 return ConstantInt::getFalse(CI->getContext());
3177 case ICmpInst::ICMP_NE:
3178 case ICmpInst::ICMP_ULT:
3179 case ICmpInst::ICMP_ULE:
3180 return ConstantInt::getTrue(CI->getContext());
3182 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3183 // is non-negative then LHS <s RHS.
3184 case ICmpInst::ICMP_SGT:
3185 case ICmpInst::ICMP_SGE:
3186 return CI->getValue().isNegative() ?
3187 ConstantInt::getTrue(CI->getContext()) :
3188 ConstantInt::getFalse(CI->getContext());
3190 case ICmpInst::ICMP_SLT:
3191 case ICmpInst::ICMP_SLE:
3192 return CI->getValue().isNegative() ?
3193 ConstantInt::getFalse(CI->getContext()) :
3194 ConstantInt::getTrue(CI->getContext());
3200 if (isa<SExtInst>(LHS)) {
3201 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3203 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3204 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3205 // Compare X and Y. Note that the predicate does not change.
3206 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3210 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3211 // too. If not, then try to deduce the result of the comparison.
3212 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3213 // Compute the constant that would happen if we truncated to SrcTy then
3214 // reextended to DstTy.
3215 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3216 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3218 // If the re-extended constant didn't change then this is effectively
3219 // also a case of comparing two sign-extended values.
3220 if (RExt == CI && MaxRecurse)
3221 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3224 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3225 // bits there. Use this to work out the result of the comparison.
3228 default: llvm_unreachable("Unknown ICmp predicate!");
3229 case ICmpInst::ICMP_EQ:
3230 return ConstantInt::getFalse(CI->getContext());
3231 case ICmpInst::ICMP_NE:
3232 return ConstantInt::getTrue(CI->getContext());
3234 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3236 case ICmpInst::ICMP_SGT:
3237 case ICmpInst::ICMP_SGE:
3238 return CI->getValue().isNegative() ?
3239 ConstantInt::getTrue(CI->getContext()) :
3240 ConstantInt::getFalse(CI->getContext());
3241 case ICmpInst::ICMP_SLT:
3242 case ICmpInst::ICMP_SLE:
3243 return CI->getValue().isNegative() ?
3244 ConstantInt::getFalse(CI->getContext()) :
3245 ConstantInt::getTrue(CI->getContext());
3247 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3249 case ICmpInst::ICMP_UGT:
3250 case ICmpInst::ICMP_UGE:
3251 // Comparison is true iff the LHS <s 0.
3253 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3254 Constant::getNullValue(SrcTy),
3258 case ICmpInst::ICMP_ULT:
3259 case ICmpInst::ICMP_ULE:
3260 // Comparison is true iff the LHS >=s 0.
3262 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3263 Constant::getNullValue(SrcTy),
3273 // icmp eq|ne X, Y -> false|true if X != Y
3274 if (ICmpInst::isEquality(Pred) &&
3275 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3276 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3279 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3282 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3285 // Simplify comparisons of related pointers using a powerful, recursive
3286 // GEP-walk when we have target data available..
3287 if (LHS->getType()->isPointerTy())
3288 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3291 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3292 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3293 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3294 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3295 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3296 Q.DL.getTypeSizeInBits(CRHS->getType()))
3297 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3298 CLHS->getPointerOperand(),
3299 CRHS->getPointerOperand()))
3302 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3303 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3304 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3305 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3306 (ICmpInst::isEquality(Pred) ||
3307 (GLHS->isInBounds() && GRHS->isInBounds() &&
3308 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3309 // The bases are equal and the indices are constant. Build a constant
3310 // expression GEP with the same indices and a null base pointer to see
3311 // what constant folding can make out of it.
3312 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3313 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3314 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3315 GLHS->getSourceElementType(), Null, IndicesLHS);
3317 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3318 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3319 GLHS->getSourceElementType(), Null, IndicesRHS);
3320 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3325 // If the comparison is with the result of a select instruction, check whether
3326 // comparing with either branch of the select always yields the same value.
3327 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3328 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3331 // If the comparison is with the result of a phi instruction, check whether
3332 // doing the compare with each incoming phi value yields a common result.
3333 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3334 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3340 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3341 const SimplifyQuery &Q) {
3342 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3345 /// Given operands for an FCmpInst, see if we can fold the result.
3346 /// If not, this returns null.
3347 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3348 FastMathFlags FMF, const SimplifyQuery &Q,
3349 unsigned MaxRecurse) {
3350 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3351 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3353 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3354 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3355 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3357 // If we have a constant, make sure it is on the RHS.
3358 std::swap(LHS, RHS);
3359 Pred = CmpInst::getSwappedPredicate(Pred);
3362 // Fold trivial predicates.
3363 Type *RetTy = GetCompareTy(LHS);
3364 if (Pred == FCmpInst::FCMP_FALSE)
3365 return getFalse(RetTy);
3366 if (Pred == FCmpInst::FCMP_TRUE)
3367 return getTrue(RetTy);
3369 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3371 if (Pred == FCmpInst::FCMP_UNO)
3372 return getFalse(RetTy);
3373 if (Pred == FCmpInst::FCMP_ORD)
3374 return getTrue(RetTy);
3377 // NaN is unordered; NaN is not ordered.
3378 assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&
3379 "Comparison must be either ordered or unordered");
3380 if (match(RHS, m_NaN()))
3381 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3383 // fcmp pred x, undef and fcmp pred undef, x
3384 // fold to true if unordered, false if ordered
3385 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3386 // Choosing NaN for the undef will always make unordered comparison succeed
3387 // and ordered comparison fail.
3388 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3391 // fcmp x,x -> true/false. Not all compares are foldable.
3393 if (CmpInst::isTrueWhenEqual(Pred))
3394 return getTrue(RetTy);
3395 if (CmpInst::isFalseWhenEqual(Pred))
3396 return getFalse(RetTy);
3399 // Handle fcmp with constant RHS.
3401 if (match(RHS, m_APFloat(C))) {
3402 // Check whether the constant is an infinity.
3403 if (C->isInfinity()) {
3404 if (C->isNegative()) {
3406 case FCmpInst::FCMP_OLT:
3407 // No value is ordered and less than negative infinity.
3408 return getFalse(RetTy);
3409 case FCmpInst::FCMP_UGE:
3410 // All values are unordered with or at least negative infinity.
3411 return getTrue(RetTy);
3417 case FCmpInst::FCMP_OGT:
3418 // No value is ordered and greater than infinity.
3419 return getFalse(RetTy);
3420 case FCmpInst::FCMP_ULE:
3421 // All values are unordered with and at most infinity.
3422 return getTrue(RetTy);
3430 case FCmpInst::FCMP_UGE:
3431 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3432 return getTrue(RetTy);
3434 case FCmpInst::FCMP_OLT:
3436 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3437 return getFalse(RetTy);
3442 } else if (C->isNegative()) {
3443 assert(!C->isNaN() && "Unexpected NaN constant!");
3444 // TODO: We can catch more cases by using a range check rather than
3445 // relying on CannotBeOrderedLessThanZero.
3447 case FCmpInst::FCMP_UGE:
3448 case FCmpInst::FCMP_UGT:
3449 case FCmpInst::FCMP_UNE:
3450 // (X >= 0) implies (X > C) when (C < 0)
3451 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3452 return getTrue(RetTy);
3454 case FCmpInst::FCMP_OEQ:
3455 case FCmpInst::FCMP_OLE:
3456 case FCmpInst::FCMP_OLT:
3457 // (X >= 0) implies !(X < C) when (C < 0)
3458 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3459 return getFalse(RetTy);
3467 // If the comparison is with the result of a select instruction, check whether
3468 // comparing with either branch of the select always yields the same value.
3469 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3470 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3473 // If the comparison is with the result of a phi instruction, check whether
3474 // doing the compare with each incoming phi value yields a common result.
3475 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3476 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3482 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3483 FastMathFlags FMF, const SimplifyQuery &Q) {
3484 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3487 /// See if V simplifies when its operand Op is replaced with RepOp.
3488 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3489 const SimplifyQuery &Q,
3490 unsigned MaxRecurse) {
3491 // Trivial replacement.
3495 // We cannot replace a constant, and shouldn't even try.
3496 if (isa<Constant>(Op))
3499 auto *I = dyn_cast<Instruction>(V);
3503 // If this is a binary operator, try to simplify it with the replaced op.
3504 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3506 // %cmp = icmp eq i32 %x, 2147483647
3507 // %add = add nsw i32 %x, 1
3508 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3510 // We can't replace %sel with %add unless we strip away the flags.
3511 if (isa<OverflowingBinaryOperator>(B))
3512 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3514 if (isa<PossiblyExactOperator>(B))
3519 if (B->getOperand(0) == Op)
3520 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3522 if (B->getOperand(1) == Op)
3523 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3528 // Same for CmpInsts.
3529 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3531 if (C->getOperand(0) == Op)
3532 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3534 if (C->getOperand(1) == Op)
3535 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3541 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3543 SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3544 transform(GEP->operands(), NewOps.begin(),
3545 [&](Value *V) { return V == Op ? RepOp : V; });
3546 return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3551 // TODO: We could hand off more cases to instsimplify here.
3553 // If all operands are constant after substituting Op for RepOp then we can
3554 // constant fold the instruction.
3555 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3556 // Build a list of all constant operands.
3557 SmallVector<Constant *, 8> ConstOps;
3558 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3559 if (I->getOperand(i) == Op)
3560 ConstOps.push_back(CRepOp);
3561 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3562 ConstOps.push_back(COp);
3567 // All operands were constants, fold it.
3568 if (ConstOps.size() == I->getNumOperands()) {
3569 if (CmpInst *C = dyn_cast<CmpInst>(I))
3570 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3571 ConstOps[1], Q.DL, Q.TLI);
3573 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3574 if (!LI->isVolatile())
3575 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3577 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3584 /// Try to simplify a select instruction when its condition operand is an
3585 /// integer comparison where one operand of the compare is a constant.
3586 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3587 const APInt *Y, bool TrueWhenUnset) {
3590 // (X & Y) == 0 ? X & ~Y : X --> X
3591 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3592 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3594 return TrueWhenUnset ? FalseVal : TrueVal;
3596 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3597 // (X & Y) != 0 ? X : X & ~Y --> X
3598 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3600 return TrueWhenUnset ? FalseVal : TrueVal;
3602 if (Y->isPowerOf2()) {
3603 // (X & Y) == 0 ? X | Y : X --> X | Y
3604 // (X & Y) != 0 ? X | Y : X --> X
3605 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3607 return TrueWhenUnset ? TrueVal : FalseVal;
3609 // (X & Y) == 0 ? X : X | Y --> X
3610 // (X & Y) != 0 ? X : X | Y --> X | Y
3611 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3613 return TrueWhenUnset ? TrueVal : FalseVal;
3619 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3621 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3622 ICmpInst::Predicate Pred,
3623 Value *TrueVal, Value *FalseVal) {
3626 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3629 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3630 Pred == ICmpInst::ICMP_EQ);
3633 /// Try to simplify a select instruction when its condition operand is an
3634 /// integer comparison.
3635 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3636 Value *FalseVal, const SimplifyQuery &Q,
3637 unsigned MaxRecurse) {
3638 ICmpInst::Predicate Pred;
3639 Value *CmpLHS, *CmpRHS;
3640 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3643 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3646 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3647 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3648 Pred == ICmpInst::ICMP_EQ))
3652 // Check for other compares that behave like bit test.
3653 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3657 // If we have an equality comparison, then we know the value in one of the
3658 // arms of the select. See if substituting this value into the arm and
3659 // simplifying the result yields the same value as the other arm.
3660 if (Pred == ICmpInst::ICMP_EQ) {
3661 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3663 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3666 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3668 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3671 } else if (Pred == ICmpInst::ICMP_NE) {
3672 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3674 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3677 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3679 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3687 /// Given operands for a SelectInst, see if we can fold the result.
3688 /// If not, this returns null.
3689 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3690 const SimplifyQuery &Q, unsigned MaxRecurse) {
3691 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3692 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3693 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3694 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3696 // select undef, X, Y -> X or Y
3697 if (isa<UndefValue>(CondC))
3698 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3700 // TODO: Vector constants with undef elements don't simplify.
3702 // select true, X, Y -> X
3703 if (CondC->isAllOnesValue())
3705 // select false, X, Y -> Y
3706 if (CondC->isNullValue())
3710 // select ?, X, X -> X
3711 if (TrueVal == FalseVal)
3714 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3716 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3720 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3726 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3727 const SimplifyQuery &Q) {
3728 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3731 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3732 /// If not, this returns null.
3733 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3734 const SimplifyQuery &Q, unsigned) {
3735 // The type of the GEP pointer operand.
3737 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3739 // getelementptr P -> P.
3740 if (Ops.size() == 1)
3743 // Compute the (pointer) type returned by the GEP instruction.
3744 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3745 Type *GEPTy = PointerType::get(LastType, AS);
3746 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3747 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3748 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3749 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3751 if (isa<UndefValue>(Ops[0]))
3752 return UndefValue::get(GEPTy);
3754 if (Ops.size() == 2) {
3755 // getelementptr P, 0 -> P.
3756 if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3760 if (Ty->isSized()) {
3763 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3764 // getelementptr P, N -> P if P points to a type of zero size.
3765 if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3768 // The following transforms are only safe if the ptrtoint cast
3769 // doesn't truncate the pointers.
3770 if (Ops[1]->getType()->getScalarSizeInBits() ==
3771 Q.DL.getIndexSizeInBits(AS)) {
3772 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3773 if (match(P, m_Zero()))
3774 return Constant::getNullValue(GEPTy);
3776 if (match(P, m_PtrToInt(m_Value(Temp))))
3777 if (Temp->getType() == GEPTy)
3782 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3783 if (TyAllocSize == 1 &&
3784 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3785 if (Value *R = PtrToIntOrZero(P))
3788 // getelementptr V, (ashr (sub P, V), C) -> Q
3789 // if P points to a type of size 1 << C.
3791 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3792 m_ConstantInt(C))) &&
3793 TyAllocSize == 1ULL << C)
3794 if (Value *R = PtrToIntOrZero(P))
3797 // getelementptr V, (sdiv (sub P, V), C) -> Q
3798 // if P points to a type of size C.
3800 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3801 m_SpecificInt(TyAllocSize))))
3802 if (Value *R = PtrToIntOrZero(P))
3808 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3809 all_of(Ops.slice(1).drop_back(1),
3810 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3812 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3813 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3814 APInt BasePtrOffset(IdxWidth, 0);
3815 Value *StrippedBasePtr =
3816 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3819 // gep (gep V, C), (sub 0, V) -> C
3820 if (match(Ops.back(),
3821 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3822 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3823 return ConstantExpr::getIntToPtr(CI, GEPTy);
3825 // gep (gep V, C), (xor V, -1) -> C-1
3826 if (match(Ops.back(),
3827 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3828 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3829 return ConstantExpr::getIntToPtr(CI, GEPTy);
3834 // Check to see if this is constant foldable.
3835 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3838 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3840 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3845 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3846 const SimplifyQuery &Q) {
3847 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3850 /// Given operands for an InsertValueInst, see if we can fold the result.
3851 /// If not, this returns null.
3852 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3853 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3855 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3856 if (Constant *CVal = dyn_cast<Constant>(Val))
3857 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3859 // insertvalue x, undef, n -> x
3860 if (match(Val, m_Undef()))
3863 // insertvalue x, (extractvalue y, n), n
3864 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3865 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3866 EV->getIndices() == Idxs) {
3867 // insertvalue undef, (extractvalue y, n), n -> y
3868 if (match(Agg, m_Undef()))
3869 return EV->getAggregateOperand();
3871 // insertvalue y, (extractvalue y, n), n -> y
3872 if (Agg == EV->getAggregateOperand())
3879 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3880 ArrayRef<unsigned> Idxs,
3881 const SimplifyQuery &Q) {
3882 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3885 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3886 const SimplifyQuery &Q) {
3887 // Try to constant fold.
3888 auto *VecC = dyn_cast<Constant>(Vec);
3889 auto *ValC = dyn_cast<Constant>(Val);
3890 auto *IdxC = dyn_cast<Constant>(Idx);
3891 if (VecC && ValC && IdxC)
3892 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3894 // Fold into undef if index is out of bounds.
3895 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3896 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3897 if (CI->uge(NumElements))
3898 return UndefValue::get(Vec->getType());
3901 // If index is undef, it might be out of bounds (see above case)
3902 if (isa<UndefValue>(Idx))
3903 return UndefValue::get(Vec->getType());
3908 /// Given operands for an ExtractValueInst, see if we can fold the result.
3909 /// If not, this returns null.
3910 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3911 const SimplifyQuery &, unsigned) {
3912 if (auto *CAgg = dyn_cast<Constant>(Agg))
3913 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3915 // extractvalue x, (insertvalue y, elt, n), n -> elt
3916 unsigned NumIdxs = Idxs.size();
3917 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3918 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3919 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3920 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3921 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3922 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3923 Idxs.slice(0, NumCommonIdxs)) {
3924 if (NumIdxs == NumInsertValueIdxs)
3925 return IVI->getInsertedValueOperand();
3933 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3934 const SimplifyQuery &Q) {
3935 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3938 /// Given operands for an ExtractElementInst, see if we can fold the result.
3939 /// If not, this returns null.
3940 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3942 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3943 if (auto *CIdx = dyn_cast<Constant>(Idx))
3944 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3946 // The index is not relevant if our vector is a splat.
3947 if (auto *Splat = CVec->getSplatValue())
3950 if (isa<UndefValue>(Vec))
3951 return UndefValue::get(Vec->getType()->getVectorElementType());
3954 // If extracting a specified index from the vector, see if we can recursively
3955 // find a previously computed scalar that was inserted into the vector.
3956 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
3957 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
3958 // definitely out of bounds, thus undefined result
3959 return UndefValue::get(Vec->getType()->getVectorElementType());
3960 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3964 // An undef extract index can be arbitrarily chosen to be an out-of-range
3965 // index value, which would result in the instruction being undef.
3966 if (isa<UndefValue>(Idx))
3967 return UndefValue::get(Vec->getType()->getVectorElementType());
3972 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3973 const SimplifyQuery &Q) {
3974 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3977 /// See if we can fold the given phi. If not, returns null.
3978 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3979 // If all of the PHI's incoming values are the same then replace the PHI node
3980 // with the common value.
3981 Value *CommonValue = nullptr;
3982 bool HasUndefInput = false;
3983 for (Value *Incoming : PN->incoming_values()) {
3984 // If the incoming value is the phi node itself, it can safely be skipped.
3985 if (Incoming == PN) continue;
3986 if (isa<UndefValue>(Incoming)) {
3987 // Remember that we saw an undef value, but otherwise ignore them.
3988 HasUndefInput = true;
3991 if (CommonValue && Incoming != CommonValue)
3992 return nullptr; // Not the same, bail out.
3993 CommonValue = Incoming;
3996 // If CommonValue is null then all of the incoming values were either undef or
3997 // equal to the phi node itself.
3999 return UndefValue::get(PN->getType());
4001 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4002 // instruction, we cannot return X as the result of the PHI node unless it
4003 // dominates the PHI block.
4005 return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4010 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4011 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4012 if (auto *C = dyn_cast<Constant>(Op))
4013 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4015 if (auto *CI = dyn_cast<CastInst>(Op)) {
4016 auto *Src = CI->getOperand(0);
4017 Type *SrcTy = Src->getType();
4018 Type *MidTy = CI->getType();
4020 if (Src->getType() == Ty) {
4021 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4022 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4024 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4026 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4028 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4029 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4030 SrcIntPtrTy, MidIntPtrTy,
4031 DstIntPtrTy) == Instruction::BitCast)
4037 if (CastOpc == Instruction::BitCast)
4038 if (Op->getType() == Ty)
4044 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4045 const SimplifyQuery &Q) {
4046 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4049 /// For the given destination element of a shuffle, peek through shuffles to
4050 /// match a root vector source operand that contains that element in the same
4051 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4052 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4053 int MaskVal, Value *RootVec,
4054 unsigned MaxRecurse) {
4058 // Bail out if any mask value is undefined. That kind of shuffle may be
4059 // simplified further based on demanded bits or other folds.
4063 // The mask value chooses which source operand we need to look at next.
4064 int InVecNumElts = Op0->getType()->getVectorNumElements();
4065 int RootElt = MaskVal;
4066 Value *SourceOp = Op0;
4067 if (MaskVal >= InVecNumElts) {
4068 RootElt = MaskVal - InVecNumElts;
4072 // If the source operand is a shuffle itself, look through it to find the
4073 // matching root vector.
4074 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4075 return foldIdentityShuffles(
4076 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4077 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4080 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4083 // The source operand is not a shuffle. Initialize the root vector value for
4084 // this shuffle if that has not been done yet.
4088 // Give up as soon as a source operand does not match the existing root value.
4089 if (RootVec != SourceOp)
4092 // The element must be coming from the same lane in the source vector
4093 // (although it may have crossed lanes in intermediate shuffles).
4094 if (RootElt != DestElt)
4100 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4101 Type *RetTy, const SimplifyQuery &Q,
4102 unsigned MaxRecurse) {
4103 if (isa<UndefValue>(Mask))
4104 return UndefValue::get(RetTy);
4106 Type *InVecTy = Op0->getType();
4107 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4108 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4110 SmallVector<int, 32> Indices;
4111 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4112 assert(MaskNumElts == Indices.size() &&
4113 "Size of Indices not same as number of mask elements?");
4115 // Canonicalization: If mask does not select elements from an input vector,
4116 // replace that input vector with undef.
4117 bool MaskSelects0 = false, MaskSelects1 = false;
4118 for (unsigned i = 0; i != MaskNumElts; ++i) {
4119 if (Indices[i] == -1)
4121 if ((unsigned)Indices[i] < InVecNumElts)
4122 MaskSelects0 = true;
4124 MaskSelects1 = true;
4127 Op0 = UndefValue::get(InVecTy);
4129 Op1 = UndefValue::get(InVecTy);
4131 auto *Op0Const = dyn_cast<Constant>(Op0);
4132 auto *Op1Const = dyn_cast<Constant>(Op1);
4134 // If all operands are constant, constant fold the shuffle.
4135 if (Op0Const && Op1Const)
4136 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4138 // Canonicalization: if only one input vector is constant, it shall be the
4140 if (Op0Const && !Op1Const) {
4141 std::swap(Op0, Op1);
4142 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4145 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4146 // value type is same as the input vectors' type.
4147 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4148 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4149 OpShuf->getMask()->getSplatValue())
4152 // Don't fold a shuffle with undef mask elements. This may get folded in a
4153 // better way using demanded bits or other analysis.
4154 // TODO: Should we allow this?
4155 if (find(Indices, -1) != Indices.end())
4158 // Check if every element of this shuffle can be mapped back to the
4159 // corresponding element of a single root vector. If so, we don't need this
4160 // shuffle. This handles simple identity shuffles as well as chains of
4161 // shuffles that may widen/narrow and/or move elements across lanes and back.
4162 Value *RootVec = nullptr;
4163 for (unsigned i = 0; i != MaskNumElts; ++i) {
4164 // Note that recursion is limited for each vector element, so if any element
4165 // exceeds the limit, this will fail to simplify.
4167 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4169 // We can't replace a widening/narrowing shuffle with one of its operands.
4170 if (!RootVec || RootVec->getType() != RetTy)
4176 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4177 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4178 Type *RetTy, const SimplifyQuery &Q) {
4179 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4182 static Constant *propagateNaN(Constant *In) {
4183 // If the input is a vector with undef elements, just return a default NaN.
4185 return ConstantFP::getNaN(In->getType());
4187 // Propagate the existing NaN constant when possible.
4188 // TODO: Should we quiet a signaling NaN?
4192 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4193 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4194 return ConstantFP::getNaN(Op0->getType());
4196 if (match(Op0, m_NaN()))
4197 return propagateNaN(cast<Constant>(Op0));
4198 if (match(Op1, m_NaN()))
4199 return propagateNaN(cast<Constant>(Op1));
4204 /// Given operands for an FAdd, see if we can fold the result. If not, this
4206 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4207 const SimplifyQuery &Q, unsigned MaxRecurse) {
4208 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4211 if (Constant *C = simplifyFPBinop(Op0, Op1))
4215 if (match(Op1, m_NegZeroFP()))
4218 // fadd X, 0 ==> X, when we know X is not -0
4219 if (match(Op1, m_PosZeroFP()) &&
4220 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4223 // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4224 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4225 // Negative zeros are allowed because we always end up with positive zero:
4226 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4227 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4228 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4229 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4230 if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4231 match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4232 return ConstantFP::getNullValue(Op0->getType());
4237 /// Given operands for an FSub, see if we can fold the result. If not, this
4239 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4240 const SimplifyQuery &Q, unsigned MaxRecurse) {
4241 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4244 if (Constant *C = simplifyFPBinop(Op0, Op1))
4248 if (match(Op1, m_PosZeroFP()))
4251 // fsub X, -0 ==> X, when we know X is not -0
4252 if (match(Op1, m_NegZeroFP()) &&
4253 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4256 // fsub -0.0, (fsub -0.0, X) ==> X
4258 if (match(Op0, m_NegZeroFP()) &&
4259 match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4262 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4263 if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4264 match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4267 // fsub nnan x, x ==> 0.0
4268 if (FMF.noNaNs() && Op0 == Op1)
4269 return Constant::getNullValue(Op0->getType());
4274 /// Given the operands for an FMul, see if we can fold the result
4275 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4276 const SimplifyQuery &Q, unsigned MaxRecurse) {
4277 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4280 if (Constant *C = simplifyFPBinop(Op0, Op1))
4283 // fmul X, 1.0 ==> X
4284 if (match(Op1, m_FPOne()))
4287 // fmul nnan nsz X, 0 ==> 0
4288 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4289 return ConstantFP::getNullValue(Op0->getType());
4291 // sqrt(X) * sqrt(X) --> X, if we can:
4292 // 1. Remove the intermediate rounding (reassociate).
4293 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4294 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4296 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4297 FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4303 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4304 const SimplifyQuery &Q) {
4305 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4309 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4310 const SimplifyQuery &Q) {
4311 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4314 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4315 const SimplifyQuery &Q) {
4316 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4319 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4320 const SimplifyQuery &Q, unsigned) {
4321 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4324 if (Constant *C = simplifyFPBinop(Op0, Op1))
4328 if (match(Op1, m_FPOne()))
4332 // Requires that NaNs are off (X could be zero) and signed zeroes are
4333 // ignored (X could be positive or negative, so the output sign is unknown).
4334 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4335 return ConstantFP::getNullValue(Op0->getType());
4338 // X / X -> 1.0 is legal when NaNs are ignored.
4339 // We can ignore infinities because INF/INF is NaN.
4341 return ConstantFP::get(Op0->getType(), 1.0);
4343 // (X * Y) / Y --> X if we can reassociate to the above form.
4345 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4348 // -X / X -> -1.0 and
4349 // X / -X -> -1.0 are legal when NaNs are ignored.
4350 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4351 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4352 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4353 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4354 BinaryOperator::getFNegArgument(Op1) == Op0))
4355 return ConstantFP::get(Op0->getType(), -1.0);
4361 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4362 const SimplifyQuery &Q) {
4363 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4366 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4367 const SimplifyQuery &Q, unsigned) {
4368 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4371 if (Constant *C = simplifyFPBinop(Op0, Op1))
4374 // Unlike fdiv, the result of frem always matches the sign of the dividend.
4375 // The constant match may include undef elements in a vector, so return a full
4376 // zero constant as the result.
4379 if (match(Op0, m_PosZeroFP()))
4380 return ConstantFP::getNullValue(Op0->getType());
4382 if (match(Op0, m_NegZeroFP()))
4383 return ConstantFP::getNegativeZero(Op0->getType());
4389 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4390 const SimplifyQuery &Q) {
4391 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4394 //=== Helper functions for higher up the class hierarchy.
4396 /// Given operands for a BinaryOperator, see if we can fold the result.
4397 /// If not, this returns null.
4398 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4399 const SimplifyQuery &Q, unsigned MaxRecurse) {
4401 case Instruction::Add:
4402 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4403 case Instruction::Sub:
4404 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4405 case Instruction::Mul:
4406 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4407 case Instruction::SDiv:
4408 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4409 case Instruction::UDiv:
4410 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4411 case Instruction::SRem:
4412 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4413 case Instruction::URem:
4414 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4415 case Instruction::Shl:
4416 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4417 case Instruction::LShr:
4418 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4419 case Instruction::AShr:
4420 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4421 case Instruction::And:
4422 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4423 case Instruction::Or:
4424 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4425 case Instruction::Xor:
4426 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4427 case Instruction::FAdd:
4428 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4429 case Instruction::FSub:
4430 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4431 case Instruction::FMul:
4432 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4433 case Instruction::FDiv:
4434 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4435 case Instruction::FRem:
4436 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4438 llvm_unreachable("Unexpected opcode");
4442 /// Given operands for a BinaryOperator, see if we can fold the result.
4443 /// If not, this returns null.
4444 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4445 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4446 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4447 const FastMathFlags &FMF, const SimplifyQuery &Q,
4448 unsigned MaxRecurse) {
4450 case Instruction::FAdd:
4451 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4452 case Instruction::FSub:
4453 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4454 case Instruction::FMul:
4455 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4456 case Instruction::FDiv:
4457 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4459 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4463 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4464 const SimplifyQuery &Q) {
4465 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4468 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4469 FastMathFlags FMF, const SimplifyQuery &Q) {
4470 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4473 /// Given operands for a CmpInst, see if we can fold the result.
4474 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4475 const SimplifyQuery &Q, unsigned MaxRecurse) {
4476 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4477 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4478 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4481 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4482 const SimplifyQuery &Q) {
4483 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4486 static bool IsIdempotent(Intrinsic::ID ID) {
4488 default: return false;
4490 // Unary idempotent: f(f(x)) = f(x)
4491 case Intrinsic::fabs:
4492 case Intrinsic::floor:
4493 case Intrinsic::ceil:
4494 case Intrinsic::trunc:
4495 case Intrinsic::rint:
4496 case Intrinsic::nearbyint:
4497 case Intrinsic::round:
4498 case Intrinsic::canonicalize:
4503 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4504 const DataLayout &DL) {
4505 GlobalValue *PtrSym;
4507 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4510 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4511 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4512 Type *Int32PtrTy = Int32Ty->getPointerTo();
4513 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4515 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4516 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4519 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4520 if (OffsetInt % 4 != 0)
4523 Constant *C = ConstantExpr::getGetElementPtr(
4524 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4525 ConstantInt::get(Int64Ty, OffsetInt / 4));
4526 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4530 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4534 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4535 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4540 if (LoadedCE->getOpcode() != Instruction::Sub)
4543 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4544 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4546 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4548 Constant *LoadedRHS = LoadedCE->getOperand(1);
4549 GlobalValue *LoadedRHSSym;
4550 APInt LoadedRHSOffset;
4551 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4553 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4556 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4559 static bool maskIsAllZeroOrUndef(Value *Mask) {
4560 auto *ConstMask = dyn_cast<Constant>(Mask);
4563 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4565 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4567 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4568 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4575 template <typename IterTy>
4576 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4577 const SimplifyQuery &Q, unsigned MaxRecurse) {
4578 Intrinsic::ID IID = F->getIntrinsicID();
4579 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4582 if (NumOperands == 1) {
4583 // Perform idempotent optimizations
4584 if (IsIdempotent(IID)) {
4585 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4586 if (II->getIntrinsicID() == IID)
4591 Value *IIOperand = *ArgBegin;
4594 case Intrinsic::fabs: {
4595 if (SignBitMustBeZero(IIOperand, Q.TLI))
4599 case Intrinsic::bswap: {
4600 // bswap(bswap(x)) -> x
4601 if (match(IIOperand, m_BSwap(m_Value(X))))
4605 case Intrinsic::bitreverse: {
4606 // bitreverse(bitreverse(x)) -> x
4607 if (match(IIOperand, m_BitReverse(m_Value(X))))
4611 case Intrinsic::exp: {
4613 if (Q.CxtI->hasAllowReassoc() &&
4614 match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X))))
4618 case Intrinsic::exp2: {
4619 // exp2(log2(x)) -> x
4620 if (Q.CxtI->hasAllowReassoc() &&
4621 match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X))))
4625 case Intrinsic::log: {
4627 if (Q.CxtI->hasAllowReassoc() &&
4628 match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X))))
4632 case Intrinsic::log2: {
4633 // log2(exp2(x)) -> x
4634 if (Q.CxtI->hasAllowReassoc() &&
4635 match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) {
4646 if (NumOperands == 2) {
4647 Value *LHS = *ArgBegin;
4648 Value *RHS = *(ArgBegin + 1);
4649 Type *ReturnType = F->getReturnType();
4652 case Intrinsic::usub_with_overflow:
4653 case Intrinsic::ssub_with_overflow: {
4654 // X - X -> { 0, false }
4656 return Constant::getNullValue(ReturnType);
4658 // X - undef -> undef
4659 // undef - X -> undef
4660 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4661 return UndefValue::get(ReturnType);
4665 case Intrinsic::uadd_with_overflow:
4666 case Intrinsic::sadd_with_overflow: {
4667 // X + undef -> undef
4668 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4669 return UndefValue::get(ReturnType);
4673 case Intrinsic::umul_with_overflow:
4674 case Intrinsic::smul_with_overflow: {
4675 // 0 * X -> { 0, false }
4676 // X * 0 -> { 0, false }
4677 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4678 return Constant::getNullValue(ReturnType);
4680 // undef * X -> { 0, false }
4681 // X * undef -> { 0, false }
4682 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4683 return Constant::getNullValue(ReturnType);
4687 case Intrinsic::load_relative: {
4688 Constant *C0 = dyn_cast<Constant>(LHS);
4689 Constant *C1 = dyn_cast<Constant>(RHS);
4691 return SimplifyRelativeLoad(C0, C1, Q.DL);
4694 case Intrinsic::powi:
4695 if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) {
4696 // powi(x, 0) -> 1.0
4697 if (Power->isZero())
4698 return ConstantFP::get(LHS->getType(), 1.0);
4709 // Simplify calls to llvm.masked.load.*
4711 case Intrinsic::masked_load: {
4712 Value *MaskArg = ArgBegin[2];
4713 Value *PassthruArg = ArgBegin[3];
4714 // If the mask is all zeros or undef, the "passthru" argument is the result.
4715 if (maskIsAllZeroOrUndef(MaskArg))
4724 template <typename IterTy>
4725 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4726 IterTy ArgEnd, const SimplifyQuery &Q,
4727 unsigned MaxRecurse) {
4728 Type *Ty = V->getType();
4729 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4730 Ty = PTy->getElementType();
4731 FunctionType *FTy = cast<FunctionType>(Ty);
4733 // call undef -> undef
4734 // call null -> undef
4735 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4736 return UndefValue::get(FTy->getReturnType());
4738 Function *F = dyn_cast<Function>(V);
4742 if (F->isIntrinsic())
4743 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4746 if (!canConstantFoldCallTo(CS, F))
4749 SmallVector<Constant *, 4> ConstantArgs;
4750 ConstantArgs.reserve(ArgEnd - ArgBegin);
4751 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4752 Constant *C = dyn_cast<Constant>(*I);
4755 ConstantArgs.push_back(C);
4758 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4761 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4762 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4763 const SimplifyQuery &Q) {
4764 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4767 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4768 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4769 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4772 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4773 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4774 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4778 /// See if we can compute a simplified version of this instruction.
4779 /// If not, this returns null.
4781 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4782 OptimizationRemarkEmitter *ORE) {
4783 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4786 switch (I->getOpcode()) {
4788 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4790 case Instruction::FAdd:
4791 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4792 I->getFastMathFlags(), Q);
4794 case Instruction::Add:
4795 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4796 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4797 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4799 case Instruction::FSub:
4800 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4801 I->getFastMathFlags(), Q);
4803 case Instruction::Sub:
4804 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4805 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4806 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4808 case Instruction::FMul:
4809 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4810 I->getFastMathFlags(), Q);
4812 case Instruction::Mul:
4813 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4815 case Instruction::SDiv:
4816 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4818 case Instruction::UDiv:
4819 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4821 case Instruction::FDiv:
4822 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4823 I->getFastMathFlags(), Q);
4825 case Instruction::SRem:
4826 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4828 case Instruction::URem:
4829 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4831 case Instruction::FRem:
4832 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4833 I->getFastMathFlags(), Q);
4835 case Instruction::Shl:
4836 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4837 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4838 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4840 case Instruction::LShr:
4841 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4842 cast<BinaryOperator>(I)->isExact(), Q);
4844 case Instruction::AShr:
4845 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4846 cast<BinaryOperator>(I)->isExact(), Q);
4848 case Instruction::And:
4849 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4851 case Instruction::Or:
4852 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4854 case Instruction::Xor:
4855 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4857 case Instruction::ICmp:
4858 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4859 I->getOperand(0), I->getOperand(1), Q);
4861 case Instruction::FCmp:
4863 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4864 I->getOperand(1), I->getFastMathFlags(), Q);
4866 case Instruction::Select:
4867 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4868 I->getOperand(2), Q);
4870 case Instruction::GetElementPtr: {
4871 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4872 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4876 case Instruction::InsertValue: {
4877 InsertValueInst *IV = cast<InsertValueInst>(I);
4878 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4879 IV->getInsertedValueOperand(),
4880 IV->getIndices(), Q);
4883 case Instruction::InsertElement: {
4884 auto *IE = cast<InsertElementInst>(I);
4885 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4886 IE->getOperand(2), Q);
4889 case Instruction::ExtractValue: {
4890 auto *EVI = cast<ExtractValueInst>(I);
4891 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4892 EVI->getIndices(), Q);
4895 case Instruction::ExtractElement: {
4896 auto *EEI = cast<ExtractElementInst>(I);
4897 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4898 EEI->getIndexOperand(), Q);
4901 case Instruction::ShuffleVector: {
4902 auto *SVI = cast<ShuffleVectorInst>(I);
4903 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4904 SVI->getMask(), SVI->getType(), Q);
4907 case Instruction::PHI:
4908 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4910 case Instruction::Call: {
4911 CallSite CS(cast<CallInst>(I));
4912 Result = SimplifyCall(CS, Q);
4915 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4916 #include "llvm/IR/Instruction.def"
4917 #undef HANDLE_CAST_INST
4919 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4921 case Instruction::Alloca:
4922 // No simplifications for Alloca and it can't be constant folded.
4927 // In general, it is possible for computeKnownBits to determine all bits in a
4928 // value even when the operands are not all constants.
4929 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4930 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4931 if (Known.isConstant())
4932 Result = ConstantInt::get(I->getType(), Known.getConstant());
4935 /// If called on unreachable code, the above logic may report that the
4936 /// instruction simplified to itself. Make life easier for users by
4937 /// detecting that case here, returning a safe value instead.
4938 return Result == I ? UndefValue::get(I->getType()) : Result;
4941 /// Implementation of recursive simplification through an instruction's
4944 /// This is the common implementation of the recursive simplification routines.
4945 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4946 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4947 /// instructions to process and attempt to simplify it using
4948 /// InstructionSimplify.
4950 /// This routine returns 'true' only when *it* simplifies something. The passed
4951 /// in simplified value does not count toward this.
4952 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4953 const TargetLibraryInfo *TLI,
4954 const DominatorTree *DT,
4955 AssumptionCache *AC) {
4956 bool Simplified = false;
4957 SmallSetVector<Instruction *, 8> Worklist;
4958 const DataLayout &DL = I->getModule()->getDataLayout();
4960 // If we have an explicit value to collapse to, do that round of the
4961 // simplification loop by hand initially.
4963 for (User *U : I->users())
4965 Worklist.insert(cast<Instruction>(U));
4967 // Replace the instruction with its simplified value.
4968 I->replaceAllUsesWith(SimpleV);
4970 // Gracefully handle edge cases where the instruction is not wired into any
4972 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4973 !I->mayHaveSideEffects())
4974 I->eraseFromParent();
4979 // Note that we must test the size on each iteration, the worklist can grow.
4980 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4983 // See if this instruction simplifies.
4984 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4990 // Stash away all the uses of the old instruction so we can check them for
4991 // recursive simplifications after a RAUW. This is cheaper than checking all
4992 // uses of To on the recursive step in most cases.
4993 for (User *U : I->users())
4994 Worklist.insert(cast<Instruction>(U));
4996 // Replace the instruction with its simplified value.
4997 I->replaceAllUsesWith(SimpleV);
4999 // Gracefully handle edge cases where the instruction is not wired into any
5001 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
5002 !I->mayHaveSideEffects())
5003 I->eraseFromParent();
5008 bool llvm::recursivelySimplifyInstruction(Instruction *I,
5009 const TargetLibraryInfo *TLI,
5010 const DominatorTree *DT,
5011 AssumptionCache *AC) {
5012 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
5015 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
5016 const TargetLibraryInfo *TLI,
5017 const DominatorTree *DT,
5018 AssumptionCache *AC) {
5019 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5020 assert(SimpleV && "Must provide a simplified value.");
5021 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
5025 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
5026 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
5027 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
5028 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
5029 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
5030 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5031 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5032 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5035 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5036 const DataLayout &DL) {
5037 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5040 template <class T, class... TArgs>
5041 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5043 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5044 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5045 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5046 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5048 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,