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)))))
1232 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1233 const SimplifyQuery &Q) {
1234 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1237 /// Given operands for an LShr, see if we can fold the result.
1238 /// If not, this returns null.
1239 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1240 const SimplifyQuery &Q, unsigned MaxRecurse) {
1241 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1245 // (X << A) >> A -> X
1247 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1253 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1254 const SimplifyQuery &Q) {
1255 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1258 /// Given operands for an AShr, see if we can fold the result.
1259 /// If not, this returns null.
1260 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1261 const SimplifyQuery &Q, unsigned MaxRecurse) {
1262 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1266 // all ones >>a X -> -1
1267 // Do not return Op0 because it may contain undef elements if it's a vector.
1268 if (match(Op0, m_AllOnes()))
1269 return Constant::getAllOnesValue(Op0->getType());
1271 // (X << A) >> A -> X
1273 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1276 // Arithmetic shifting an all-sign-bit value is a no-op.
1277 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1278 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1284 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1285 const SimplifyQuery &Q) {
1286 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1289 /// Commuted variants are assumed to be handled by calling this function again
1290 /// with the parameters swapped.
1291 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1292 ICmpInst *UnsignedICmp, bool IsAnd) {
1295 ICmpInst::Predicate EqPred;
1296 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1297 !ICmpInst::isEquality(EqPred))
1300 ICmpInst::Predicate UnsignedPred;
1301 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1302 ICmpInst::isUnsigned(UnsignedPred))
1304 else if (match(UnsignedICmp,
1305 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1306 ICmpInst::isUnsigned(UnsignedPred))
1307 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1311 // X < Y && Y != 0 --> X < Y
1312 // X < Y || Y != 0 --> Y != 0
1313 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1314 return IsAnd ? UnsignedICmp : ZeroICmp;
1316 // X >= Y || Y != 0 --> true
1317 // X >= Y || Y == 0 --> X >= Y
1318 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1319 if (EqPred == ICmpInst::ICMP_NE)
1320 return getTrue(UnsignedICmp->getType());
1321 return UnsignedICmp;
1324 // X < Y && Y == 0 --> false
1325 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1327 return getFalse(UnsignedICmp->getType());
1332 /// Commuted variants are assumed to be handled by calling this function again
1333 /// with the parameters swapped.
1334 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1335 ICmpInst::Predicate Pred0, Pred1;
1337 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1338 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1341 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1342 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1343 // can eliminate Op1 from this 'and'.
1344 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1347 // Check for any combination of predicates that are guaranteed to be disjoint.
1348 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1349 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1350 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1351 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1352 return getFalse(Op0->getType());
1357 /// Commuted variants are assumed to be handled by calling this function again
1358 /// with the parameters swapped.
1359 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1360 ICmpInst::Predicate Pred0, Pred1;
1362 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1363 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1366 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1367 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1368 // can eliminate Op0 from this 'or'.
1369 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1372 // Check for any combination of predicates that cover the entire range of
1374 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1375 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1376 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1377 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1378 return getTrue(Op0->getType());
1383 /// Test if a pair of compares with a shared operand and 2 constants has an
1384 /// empty set intersection, full set union, or if one compare is a superset of
1386 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1388 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1389 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1392 const APInt *C0, *C1;
1393 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1394 !match(Cmp1->getOperand(1), m_APInt(C1)))
1397 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1398 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1400 // For and-of-compares, check if the intersection is empty:
1401 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1402 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1403 return getFalse(Cmp0->getType());
1405 // For or-of-compares, check if the union is full:
1406 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1407 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1408 return getTrue(Cmp0->getType());
1410 // Is one range a superset of the other?
1411 // If this is and-of-compares, take the smaller set:
1412 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1413 // If this is or-of-compares, take the larger set:
1414 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1415 if (Range0.contains(Range1))
1416 return IsAnd ? Cmp1 : Cmp0;
1417 if (Range1.contains(Range0))
1418 return IsAnd ? Cmp0 : Cmp1;
1423 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1425 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1426 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1427 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1430 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1433 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1434 Value *X = Cmp0->getOperand(0);
1435 Value *Y = Cmp1->getOperand(0);
1437 // If one of the compares is a masked version of a (not) null check, then
1438 // that compare implies the other, so we eliminate the other. Optionally, look
1439 // through a pointer-to-int cast to match a null check of a pointer type.
1441 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1442 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1443 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1444 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1445 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1446 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1449 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1450 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1451 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1452 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1453 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1454 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1460 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1461 // (icmp (add V, C0), C1) & (icmp V, C0)
1462 ICmpInst::Predicate Pred0, Pred1;
1463 const APInt *C0, *C1;
1465 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1468 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1471 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1472 if (AddInst->getOperand(1) != Op1->getOperand(1))
1475 Type *ITy = Op0->getType();
1476 bool isNSW = AddInst->hasNoSignedWrap();
1477 bool isNUW = AddInst->hasNoUnsignedWrap();
1479 const APInt Delta = *C1 - *C0;
1480 if (C0->isStrictlyPositive()) {
1482 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1483 return getFalse(ITy);
1484 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1485 return getFalse(ITy);
1488 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1489 return getFalse(ITy);
1490 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1491 return getFalse(ITy);
1494 if (C0->getBoolValue() && isNUW) {
1496 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1497 return getFalse(ITy);
1499 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1500 return getFalse(ITy);
1506 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1507 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1509 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1512 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1514 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1517 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1520 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1523 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1525 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1531 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1532 // (icmp (add V, C0), C1) | (icmp V, C0)
1533 ICmpInst::Predicate Pred0, Pred1;
1534 const APInt *C0, *C1;
1536 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1539 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1542 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1543 if (AddInst->getOperand(1) != Op1->getOperand(1))
1546 Type *ITy = Op0->getType();
1547 bool isNSW = AddInst->hasNoSignedWrap();
1548 bool isNUW = AddInst->hasNoUnsignedWrap();
1550 const APInt Delta = *C1 - *C0;
1551 if (C0->isStrictlyPositive()) {
1553 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1554 return getTrue(ITy);
1555 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1556 return getTrue(ITy);
1559 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1560 return getTrue(ITy);
1561 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1562 return getTrue(ITy);
1565 if (C0->getBoolValue() && isNUW) {
1567 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1568 return getTrue(ITy);
1570 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1571 return getTrue(ITy);
1577 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1578 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1580 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1583 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1585 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1588 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1591 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1594 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1596 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1602 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1603 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1604 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1605 if (LHS0->getType() != RHS0->getType())
1608 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1609 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1610 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1611 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1612 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1613 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1614 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1615 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1616 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1617 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1618 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1619 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1620 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1623 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1624 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1625 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1626 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1627 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1628 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1629 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1630 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1631 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1632 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1639 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1640 // Look through casts of the 'and' operands to find compares.
1641 auto *Cast0 = dyn_cast<CastInst>(Op0);
1642 auto *Cast1 = dyn_cast<CastInst>(Op1);
1643 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1644 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1645 Op0 = Cast0->getOperand(0);
1646 Op1 = Cast1->getOperand(0);
1650 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1651 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1653 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1654 simplifyOrOfICmps(ICmp0, ICmp1);
1656 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1657 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1659 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1666 // If we looked through casts, we can only handle a constant simplification
1667 // because we are not allowed to create a cast instruction here.
1668 if (auto *C = dyn_cast<Constant>(V))
1669 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1674 /// Given operands for an And, see if we can fold the result.
1675 /// If not, this returns null.
1676 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1677 unsigned MaxRecurse) {
1678 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1682 if (match(Op1, m_Undef()))
1683 return Constant::getNullValue(Op0->getType());
1690 if (match(Op1, m_Zero()))
1691 return Constant::getNullValue(Op0->getType());
1694 if (match(Op1, m_AllOnes()))
1697 // A & ~A = ~A & A = 0
1698 if (match(Op0, m_Not(m_Specific(Op1))) ||
1699 match(Op1, m_Not(m_Specific(Op0))))
1700 return Constant::getNullValue(Op0->getType());
1703 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1707 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1710 // A mask that only clears known zeros of a shifted value is a no-op.
1714 if (match(Op1, m_APInt(Mask))) {
1715 // If all bits in the inverted and shifted mask are clear:
1716 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1717 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1718 (~(*Mask)).lshr(*ShAmt).isNullValue())
1721 // If all bits in the inverted and shifted mask are clear:
1722 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1723 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1724 (~(*Mask)).shl(*ShAmt).isNullValue())
1728 // A & (-A) = A if A is a power of two or zero.
1729 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1730 match(Op1, m_Neg(m_Specific(Op0)))) {
1731 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1734 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1739 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1742 // Try some generic simplifications for associative operations.
1743 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1747 // And distributes over Or. Try some generic simplifications based on this.
1748 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1752 // And distributes over Xor. Try some generic simplifications based on this.
1753 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1757 // If the operation is with the result of a select instruction, check whether
1758 // operating on either branch of the select always yields the same value.
1759 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1760 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1764 // If the operation is with the result of a phi instruction, check whether
1765 // operating on all incoming values of the phi always yields the same value.
1766 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1767 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1774 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1775 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1778 /// Given operands for an Or, see if we can fold the result.
1779 /// If not, this returns null.
1780 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1781 unsigned MaxRecurse) {
1782 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1787 // Do not return Op1 because it may contain undef elements if it's a vector.
1788 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1789 return Constant::getAllOnesValue(Op0->getType());
1793 if (Op0 == Op1 || match(Op1, m_Zero()))
1796 // A | ~A = ~A | A = -1
1797 if (match(Op0, m_Not(m_Specific(Op1))) ||
1798 match(Op1, m_Not(m_Specific(Op0))))
1799 return Constant::getAllOnesValue(Op0->getType());
1802 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1806 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1809 // ~(A & ?) | A = -1
1810 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1811 return Constant::getAllOnesValue(Op1->getType());
1813 // A | ~(A & ?) = -1
1814 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1815 return Constant::getAllOnesValue(Op0->getType());
1818 // (A & ~B) | (A ^ B) -> (A ^ B)
1819 // (~B & A) | (A ^ B) -> (A ^ B)
1820 // (A & ~B) | (B ^ A) -> (B ^ A)
1821 // (~B & A) | (B ^ A) -> (B ^ A)
1822 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1823 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1824 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1827 // Commute the 'or' operands.
1828 // (A ^ B) | (A & ~B) -> (A ^ B)
1829 // (A ^ B) | (~B & A) -> (A ^ B)
1830 // (B ^ A) | (A & ~B) -> (B ^ A)
1831 // (B ^ A) | (~B & A) -> (B ^ A)
1832 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1833 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1834 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1837 // (A & B) | (~A ^ B) -> (~A ^ B)
1838 // (B & A) | (~A ^ B) -> (~A ^ B)
1839 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1840 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1841 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1842 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1843 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1846 // (~A ^ B) | (A & B) -> (~A ^ B)
1847 // (~A ^ B) | (B & A) -> (~A ^ B)
1848 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1849 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1850 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1851 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1852 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1855 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1858 // Try some generic simplifications for associative operations.
1859 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1863 // Or distributes over And. Try some generic simplifications based on this.
1864 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1868 // If the operation is with the result of a select instruction, check whether
1869 // operating on either branch of the select always yields the same value.
1870 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1871 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1875 // (A & C1)|(B & C2)
1876 const APInt *C1, *C2;
1877 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1878 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1880 // (A & C1)|(B & C2)
1881 // If we have: ((V + N) & C1) | (V & C2)
1882 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1883 // replace with V+N.
1885 if (C2->isMask() && // C2 == 0+1+
1886 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1887 // Add commutes, try both ways.
1888 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1891 // Or commutes, try both ways.
1893 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1894 // Add commutes, try both ways.
1895 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1901 // If the operation is with the result of a phi instruction, check whether
1902 // operating on all incoming values of the phi always yields the same value.
1903 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1904 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1910 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1911 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1914 /// Given operands for a Xor, see if we can fold the result.
1915 /// If not, this returns null.
1916 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1917 unsigned MaxRecurse) {
1918 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1921 // A ^ undef -> undef
1922 if (match(Op1, m_Undef()))
1926 if (match(Op1, m_Zero()))
1931 return Constant::getNullValue(Op0->getType());
1933 // A ^ ~A = ~A ^ A = -1
1934 if (match(Op0, m_Not(m_Specific(Op1))) ||
1935 match(Op1, m_Not(m_Specific(Op0))))
1936 return Constant::getAllOnesValue(Op0->getType());
1938 // Try some generic simplifications for associative operations.
1939 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1943 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1944 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1945 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1946 // only if B and C are equal. If B and C are equal then (since we assume
1947 // that operands have already been simplified) "select(cond, B, C)" should
1948 // have been simplified to the common value of B and C already. Analysing
1949 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1950 // for threading over phi nodes.
1955 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1956 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1960 static Type *GetCompareTy(Value *Op) {
1961 return CmpInst::makeCmpResultType(Op->getType());
1964 /// Rummage around inside V looking for something equivalent to the comparison
1965 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1966 /// Helper function for analyzing max/min idioms.
1967 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1968 Value *LHS, Value *RHS) {
1969 SelectInst *SI = dyn_cast<SelectInst>(V);
1972 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1975 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1976 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1978 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1979 LHS == CmpRHS && RHS == CmpLHS)
1984 // A significant optimization not implemented here is assuming that alloca
1985 // addresses are not equal to incoming argument values. They don't *alias*,
1986 // as we say, but that doesn't mean they aren't equal, so we take a
1987 // conservative approach.
1989 // This is inspired in part by C++11 5.10p1:
1990 // "Two pointers of the same type compare equal if and only if they are both
1991 // null, both point to the same function, or both represent the same
1994 // This is pretty permissive.
1996 // It's also partly due to C11 6.5.9p6:
1997 // "Two pointers compare equal if and only if both are null pointers, both are
1998 // pointers to the same object (including a pointer to an object and a
1999 // subobject at its beginning) or function, both are pointers to one past the
2000 // last element of the same array object, or one is a pointer to one past the
2001 // end of one array object and the other is a pointer to the start of a
2002 // different array object that happens to immediately follow the first array
2003 // object in the address space.)
2005 // C11's version is more restrictive, however there's no reason why an argument
2006 // couldn't be a one-past-the-end value for a stack object in the caller and be
2007 // equal to the beginning of a stack object in the callee.
2009 // If the C and C++ standards are ever made sufficiently restrictive in this
2010 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2011 // this optimization.
2013 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2014 const DominatorTree *DT, CmpInst::Predicate Pred,
2015 AssumptionCache *AC, const Instruction *CxtI,
2016 Value *LHS, Value *RHS) {
2017 // First, skip past any trivial no-ops.
2018 LHS = LHS->stripPointerCasts();
2019 RHS = RHS->stripPointerCasts();
2021 // A non-null pointer is not equal to a null pointer.
2022 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2023 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2024 return ConstantInt::get(GetCompareTy(LHS),
2025 !CmpInst::isTrueWhenEqual(Pred));
2027 // We can only fold certain predicates on pointer comparisons.
2032 // Equality comaprisons are easy to fold.
2033 case CmpInst::ICMP_EQ:
2034 case CmpInst::ICMP_NE:
2037 // We can only handle unsigned relational comparisons because 'inbounds' on
2038 // a GEP only protects against unsigned wrapping.
2039 case CmpInst::ICMP_UGT:
2040 case CmpInst::ICMP_UGE:
2041 case CmpInst::ICMP_ULT:
2042 case CmpInst::ICMP_ULE:
2043 // However, we have to switch them to their signed variants to handle
2044 // negative indices from the base pointer.
2045 Pred = ICmpInst::getSignedPredicate(Pred);
2049 // Strip off any constant offsets so that we can reason about them.
2050 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2051 // here and compare base addresses like AliasAnalysis does, however there are
2052 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2053 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2054 // doesn't need to guarantee pointer inequality when it says NoAlias.
2055 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2056 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2058 // If LHS and RHS are related via constant offsets to the same base
2059 // value, we can replace it with an icmp which just compares the offsets.
2061 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2063 // Various optimizations for (in)equality comparisons.
2064 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2065 // Different non-empty allocations that exist at the same time have
2066 // different addresses (if the program can tell). Global variables always
2067 // exist, so they always exist during the lifetime of each other and all
2068 // allocas. Two different allocas usually have different addresses...
2070 // However, if there's an @llvm.stackrestore dynamically in between two
2071 // allocas, they may have the same address. It's tempting to reduce the
2072 // scope of the problem by only looking at *static* allocas here. That would
2073 // cover the majority of allocas while significantly reducing the likelihood
2074 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2075 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2076 // an entry block. Also, if we have a block that's not attached to a
2077 // function, we can't tell if it's "static" under the current definition.
2078 // Theoretically, this problem could be fixed by creating a new kind of
2079 // instruction kind specifically for static allocas. Such a new instruction
2080 // could be required to be at the top of the entry block, thus preventing it
2081 // from being subject to a @llvm.stackrestore. Instcombine could even
2082 // convert regular allocas into these special allocas. It'd be nifty.
2083 // However, until then, this problem remains open.
2085 // So, we'll assume that two non-empty allocas have different addresses
2088 // With all that, if the offsets are within the bounds of their allocations
2089 // (and not one-past-the-end! so we can't use inbounds!), and their
2090 // allocations aren't the same, the pointers are not equal.
2092 // Note that it's not necessary to check for LHS being a global variable
2093 // address, due to canonicalization and constant folding.
2094 if (isa<AllocaInst>(LHS) &&
2095 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2096 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2097 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2098 uint64_t LHSSize, RHSSize;
2099 if (LHSOffsetCI && RHSOffsetCI &&
2100 getObjectSize(LHS, LHSSize, DL, TLI) &&
2101 getObjectSize(RHS, RHSSize, DL, TLI)) {
2102 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2103 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2104 if (!LHSOffsetValue.isNegative() &&
2105 !RHSOffsetValue.isNegative() &&
2106 LHSOffsetValue.ult(LHSSize) &&
2107 RHSOffsetValue.ult(RHSSize)) {
2108 return ConstantInt::get(GetCompareTy(LHS),
2109 !CmpInst::isTrueWhenEqual(Pred));
2113 // Repeat the above check but this time without depending on DataLayout
2114 // or being able to compute a precise size.
2115 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2116 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2117 LHSOffset->isNullValue() &&
2118 RHSOffset->isNullValue())
2119 return ConstantInt::get(GetCompareTy(LHS),
2120 !CmpInst::isTrueWhenEqual(Pred));
2123 // Even if an non-inbounds GEP occurs along the path we can still optimize
2124 // equality comparisons concerning the result. We avoid walking the whole
2125 // chain again by starting where the last calls to
2126 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2127 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2128 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2130 return ConstantExpr::getICmp(Pred,
2131 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2132 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2134 // If one side of the equality comparison must come from a noalias call
2135 // (meaning a system memory allocation function), and the other side must
2136 // come from a pointer that cannot overlap with dynamically-allocated
2137 // memory within the lifetime of the current function (allocas, byval
2138 // arguments, globals), then determine the comparison result here.
2139 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2140 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2141 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2143 // Is the set of underlying objects all noalias calls?
2144 auto IsNAC = [](ArrayRef<Value *> Objects) {
2145 return all_of(Objects, isNoAliasCall);
2148 // Is the set of underlying objects all things which must be disjoint from
2149 // noalias calls. For allocas, we consider only static ones (dynamic
2150 // allocas might be transformed into calls to malloc not simultaneously
2151 // live with the compared-to allocation). For globals, we exclude symbols
2152 // that might be resolve lazily to symbols in another dynamically-loaded
2153 // library (and, thus, could be malloc'ed by the implementation).
2154 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2155 return all_of(Objects, [](Value *V) {
2156 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2157 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2158 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2159 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2160 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2161 !GV->isThreadLocal();
2162 if (const Argument *A = dyn_cast<Argument>(V))
2163 return A->hasByValAttr();
2168 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2169 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2170 return ConstantInt::get(GetCompareTy(LHS),
2171 !CmpInst::isTrueWhenEqual(Pred));
2173 // Fold comparisons for non-escaping pointer even if the allocation call
2174 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2175 // dynamic allocation call could be either of the operands.
2176 Value *MI = nullptr;
2177 if (isAllocLikeFn(LHS, TLI) &&
2178 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2180 else if (isAllocLikeFn(RHS, TLI) &&
2181 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2183 // FIXME: We should also fold the compare when the pointer escapes, but the
2184 // compare dominates the pointer escape
2185 if (MI && !PointerMayBeCaptured(MI, true, true))
2186 return ConstantInt::get(GetCompareTy(LHS),
2187 CmpInst::isFalseWhenEqual(Pred));
2194 /// Fold an icmp when its operands have i1 scalar type.
2195 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2196 Value *RHS, const SimplifyQuery &Q) {
2197 Type *ITy = GetCompareTy(LHS); // The return type.
2198 Type *OpTy = LHS->getType(); // The operand type.
2199 if (!OpTy->isIntOrIntVectorTy(1))
2202 // A boolean compared to true/false can be simplified in 14 out of the 20
2203 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2204 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2205 if (match(RHS, m_Zero())) {
2207 case CmpInst::ICMP_NE: // X != 0 -> X
2208 case CmpInst::ICMP_UGT: // X >u 0 -> X
2209 case CmpInst::ICMP_SLT: // X <s 0 -> X
2212 case CmpInst::ICMP_ULT: // X <u 0 -> false
2213 case CmpInst::ICMP_SGT: // X >s 0 -> false
2214 return getFalse(ITy);
2216 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2217 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2218 return getTrue(ITy);
2222 } else if (match(RHS, m_One())) {
2224 case CmpInst::ICMP_EQ: // X == 1 -> X
2225 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2226 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2229 case CmpInst::ICMP_UGT: // X >u 1 -> false
2230 case CmpInst::ICMP_SLT: // X <s -1 -> false
2231 return getFalse(ITy);
2233 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2234 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2235 return getTrue(ITy);
2244 case ICmpInst::ICMP_UGE:
2245 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2246 return getTrue(ITy);
2248 case ICmpInst::ICMP_SGE:
2249 /// For signed comparison, the values for an i1 are 0 and -1
2250 /// respectively. This maps into a truth table of:
2251 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2252 /// 0 | 0 | 1 (0 >= 0) | 1
2253 /// 0 | 1 | 1 (0 >= -1) | 1
2254 /// 1 | 0 | 0 (-1 >= 0) | 0
2255 /// 1 | 1 | 1 (-1 >= -1) | 1
2256 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2257 return getTrue(ITy);
2259 case ICmpInst::ICMP_ULE:
2260 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2261 return getTrue(ITy);
2268 /// Try hard to fold icmp with zero RHS because this is a common case.
2269 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2270 Value *RHS, const SimplifyQuery &Q) {
2271 if (!match(RHS, m_Zero()))
2274 Type *ITy = GetCompareTy(LHS); // The return type.
2277 llvm_unreachable("Unknown ICmp predicate!");
2278 case ICmpInst::ICMP_ULT:
2279 return getFalse(ITy);
2280 case ICmpInst::ICMP_UGE:
2281 return getTrue(ITy);
2282 case ICmpInst::ICMP_EQ:
2283 case ICmpInst::ICMP_ULE:
2284 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2285 return getFalse(ITy);
2287 case ICmpInst::ICMP_NE:
2288 case ICmpInst::ICMP_UGT:
2289 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2290 return getTrue(ITy);
2292 case ICmpInst::ICMP_SLT: {
2293 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2294 if (LHSKnown.isNegative())
2295 return getTrue(ITy);
2296 if (LHSKnown.isNonNegative())
2297 return getFalse(ITy);
2300 case ICmpInst::ICMP_SLE: {
2301 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2302 if (LHSKnown.isNegative())
2303 return getTrue(ITy);
2304 if (LHSKnown.isNonNegative() &&
2305 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2306 return getFalse(ITy);
2309 case ICmpInst::ICMP_SGE: {
2310 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2311 if (LHSKnown.isNegative())
2312 return getFalse(ITy);
2313 if (LHSKnown.isNonNegative())
2314 return getTrue(ITy);
2317 case ICmpInst::ICMP_SGT: {
2318 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2319 if (LHSKnown.isNegative())
2320 return getFalse(ITy);
2321 if (LHSKnown.isNonNegative() &&
2322 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2323 return getTrue(ITy);
2331 /// Many binary operators with a constant operand have an easy-to-compute
2332 /// range of outputs. This can be used to fold a comparison to always true or
2334 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2335 unsigned Width = Lower.getBitWidth();
2337 switch (BO.getOpcode()) {
2338 case Instruction::Add:
2339 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2340 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2341 if (BO.hasNoUnsignedWrap()) {
2342 // 'add nuw x, C' produces [C, UINT_MAX].
2344 } else if (BO.hasNoSignedWrap()) {
2345 if (C->isNegative()) {
2346 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2347 Lower = APInt::getSignedMinValue(Width);
2348 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2350 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2351 Lower = APInt::getSignedMinValue(Width) + *C;
2352 Upper = APInt::getSignedMaxValue(Width) + 1;
2358 case Instruction::And:
2359 if (match(BO.getOperand(1), m_APInt(C)))
2360 // 'and x, C' produces [0, C].
2364 case Instruction::Or:
2365 if (match(BO.getOperand(1), m_APInt(C)))
2366 // 'or x, C' produces [C, UINT_MAX].
2370 case Instruction::AShr:
2371 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2372 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2373 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2374 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2375 } else if (match(BO.getOperand(0), m_APInt(C))) {
2376 unsigned ShiftAmount = Width - 1;
2377 if (!C->isNullValue() && BO.isExact())
2378 ShiftAmount = C->countTrailingZeros();
2379 if (C->isNegative()) {
2380 // 'ashr C, x' produces [C, C >> (Width-1)]
2382 Upper = C->ashr(ShiftAmount) + 1;
2384 // 'ashr C, x' produces [C >> (Width-1), C]
2385 Lower = C->ashr(ShiftAmount);
2391 case Instruction::LShr:
2392 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2393 // 'lshr x, C' produces [0, UINT_MAX >> C].
2394 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2395 } else if (match(BO.getOperand(0), m_APInt(C))) {
2396 // 'lshr C, x' produces [C >> (Width-1), C].
2397 unsigned ShiftAmount = Width - 1;
2398 if (!C->isNullValue() && BO.isExact())
2399 ShiftAmount = C->countTrailingZeros();
2400 Lower = C->lshr(ShiftAmount);
2405 case Instruction::Shl:
2406 if (match(BO.getOperand(0), m_APInt(C))) {
2407 if (BO.hasNoUnsignedWrap()) {
2408 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2410 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2411 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2412 if (C->isNegative()) {
2413 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2414 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2415 Lower = C->shl(ShiftAmount);
2418 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2419 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2421 Upper = C->shl(ShiftAmount) + 1;
2427 case Instruction::SDiv:
2428 if (match(BO.getOperand(1), m_APInt(C))) {
2429 APInt IntMin = APInt::getSignedMinValue(Width);
2430 APInt IntMax = APInt::getSignedMaxValue(Width);
2431 if (C->isAllOnesValue()) {
2432 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2433 // where C != -1 and C != 0 and C != 1
2436 } else if (C->countLeadingZeros() < Width - 1) {
2437 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2438 // where C != -1 and C != 0 and C != 1
2439 Lower = IntMin.sdiv(*C);
2440 Upper = IntMax.sdiv(*C);
2441 if (Lower.sgt(Upper))
2442 std::swap(Lower, Upper);
2444 assert(Upper != Lower && "Upper part of range has wrapped!");
2446 } else if (match(BO.getOperand(0), m_APInt(C))) {
2447 if (C->isMinSignedValue()) {
2448 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2450 Upper = Lower.lshr(1) + 1;
2452 // 'sdiv C, x' produces [-|C|, |C|].
2453 Upper = C->abs() + 1;
2454 Lower = (-Upper) + 1;
2459 case Instruction::UDiv:
2460 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2461 // 'udiv x, C' produces [0, UINT_MAX / C].
2462 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2463 } else if (match(BO.getOperand(0), m_APInt(C))) {
2464 // 'udiv C, x' produces [0, C].
2469 case Instruction::SRem:
2470 if (match(BO.getOperand(1), m_APInt(C))) {
2471 // 'srem x, C' produces (-|C|, |C|).
2473 Lower = (-Upper) + 1;
2477 case Instruction::URem:
2478 if (match(BO.getOperand(1), m_APInt(C)))
2479 // 'urem x, C' produces [0, C).
2488 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2490 Type *ITy = GetCompareTy(RHS); // The return type.
2493 // Sign-bit checks can be optimized to true/false after unsigned
2494 // floating-point casts:
2495 // icmp slt (bitcast (uitofp X)), 0 --> false
2496 // icmp sgt (bitcast (uitofp X)), -1 --> true
2497 if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2498 if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2499 return ConstantInt::getFalse(ITy);
2500 if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2501 return ConstantInt::getTrue(ITy);
2505 if (!match(RHS, m_APInt(C)))
2508 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2509 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2510 if (RHS_CR.isEmptySet())
2511 return ConstantInt::getFalse(ITy);
2512 if (RHS_CR.isFullSet())
2513 return ConstantInt::getTrue(ITy);
2515 // Find the range of possible values for binary operators.
2516 unsigned Width = C->getBitWidth();
2517 APInt Lower = APInt(Width, 0);
2518 APInt Upper = APInt(Width, 0);
2519 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2520 setLimitsForBinOp(*BO, Lower, Upper);
2522 ConstantRange LHS_CR =
2523 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2525 if (auto *I = dyn_cast<Instruction>(LHS))
2526 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2527 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2529 if (!LHS_CR.isFullSet()) {
2530 if (RHS_CR.contains(LHS_CR))
2531 return ConstantInt::getTrue(ITy);
2532 if (RHS_CR.inverse().contains(LHS_CR))
2533 return ConstantInt::getFalse(ITy);
2539 /// TODO: A large part of this logic is duplicated in InstCombine's
2540 /// foldICmpBinOp(). We should be able to share that and avoid the code
2542 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2543 Value *RHS, const SimplifyQuery &Q,
2544 unsigned MaxRecurse) {
2545 Type *ITy = GetCompareTy(LHS); // The return type.
2547 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2548 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2549 if (MaxRecurse && (LBO || RBO)) {
2550 // Analyze the case when either LHS or RHS is an add instruction.
2551 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2552 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2553 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2554 if (LBO && LBO->getOpcode() == Instruction::Add) {
2555 A = LBO->getOperand(0);
2556 B = LBO->getOperand(1);
2558 ICmpInst::isEquality(Pred) ||
2559 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2560 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2562 if (RBO && RBO->getOpcode() == Instruction::Add) {
2563 C = RBO->getOperand(0);
2564 D = RBO->getOperand(1);
2566 ICmpInst::isEquality(Pred) ||
2567 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2568 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2571 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2572 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2573 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2574 Constant::getNullValue(RHS->getType()), Q,
2578 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2579 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2581 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2582 C == LHS ? D : C, Q, MaxRecurse - 1))
2585 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2586 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2588 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2591 // C + B == C + D -> B == D
2594 } else if (A == D) {
2595 // D + B == C + D -> B == C
2598 } else if (B == C) {
2599 // A + C == C + D -> A == D
2604 // A + D == C + D -> A == C
2608 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2615 // icmp pred (or X, Y), X
2616 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2617 if (Pred == ICmpInst::ICMP_ULT)
2618 return getFalse(ITy);
2619 if (Pred == ICmpInst::ICMP_UGE)
2620 return getTrue(ITy);
2622 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2623 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2624 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2625 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2626 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2627 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2628 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2631 // icmp pred X, (or X, Y)
2632 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2633 if (Pred == ICmpInst::ICMP_ULE)
2634 return getTrue(ITy);
2635 if (Pred == ICmpInst::ICMP_UGT)
2636 return getFalse(ITy);
2638 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2639 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2640 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2641 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2642 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2643 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2644 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2649 // icmp pred (and X, Y), X
2650 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2651 if (Pred == ICmpInst::ICMP_UGT)
2652 return getFalse(ITy);
2653 if (Pred == ICmpInst::ICMP_ULE)
2654 return getTrue(ITy);
2656 // icmp pred X, (and X, Y)
2657 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2658 if (Pred == ICmpInst::ICMP_UGE)
2659 return getTrue(ITy);
2660 if (Pred == ICmpInst::ICMP_ULT)
2661 return getFalse(ITy);
2664 // 0 - (zext X) pred C
2665 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2666 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2667 if (RHSC->getValue().isStrictlyPositive()) {
2668 if (Pred == ICmpInst::ICMP_SLT)
2669 return ConstantInt::getTrue(RHSC->getContext());
2670 if (Pred == ICmpInst::ICMP_SGE)
2671 return ConstantInt::getFalse(RHSC->getContext());
2672 if (Pred == ICmpInst::ICMP_EQ)
2673 return ConstantInt::getFalse(RHSC->getContext());
2674 if (Pred == ICmpInst::ICMP_NE)
2675 return ConstantInt::getTrue(RHSC->getContext());
2677 if (RHSC->getValue().isNonNegative()) {
2678 if (Pred == ICmpInst::ICMP_SLE)
2679 return ConstantInt::getTrue(RHSC->getContext());
2680 if (Pred == ICmpInst::ICMP_SGT)
2681 return ConstantInt::getFalse(RHSC->getContext());
2686 // icmp pred (urem X, Y), Y
2687 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2691 case ICmpInst::ICMP_SGT:
2692 case ICmpInst::ICMP_SGE: {
2693 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2694 if (!Known.isNonNegative())
2698 case ICmpInst::ICMP_EQ:
2699 case ICmpInst::ICMP_UGT:
2700 case ICmpInst::ICMP_UGE:
2701 return getFalse(ITy);
2702 case ICmpInst::ICMP_SLT:
2703 case ICmpInst::ICMP_SLE: {
2704 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2705 if (!Known.isNonNegative())
2709 case ICmpInst::ICMP_NE:
2710 case ICmpInst::ICMP_ULT:
2711 case ICmpInst::ICMP_ULE:
2712 return getTrue(ITy);
2716 // icmp pred X, (urem Y, X)
2717 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2721 case ICmpInst::ICMP_SGT:
2722 case ICmpInst::ICMP_SGE: {
2723 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2724 if (!Known.isNonNegative())
2728 case ICmpInst::ICMP_NE:
2729 case ICmpInst::ICMP_UGT:
2730 case ICmpInst::ICMP_UGE:
2731 return getTrue(ITy);
2732 case ICmpInst::ICMP_SLT:
2733 case ICmpInst::ICMP_SLE: {
2734 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2735 if (!Known.isNonNegative())
2739 case ICmpInst::ICMP_EQ:
2740 case ICmpInst::ICMP_ULT:
2741 case ICmpInst::ICMP_ULE:
2742 return getFalse(ITy);
2748 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2749 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2750 // icmp pred (X op Y), X
2751 if (Pred == ICmpInst::ICMP_UGT)
2752 return getFalse(ITy);
2753 if (Pred == ICmpInst::ICMP_ULE)
2754 return getTrue(ITy);
2759 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2760 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2761 // icmp pred X, (X op Y)
2762 if (Pred == ICmpInst::ICMP_ULT)
2763 return getFalse(ITy);
2764 if (Pred == ICmpInst::ICMP_UGE)
2765 return getTrue(ITy);
2772 // where CI2 is a power of 2 and CI isn't
2773 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2774 const APInt *CI2Val, *CIVal = &CI->getValue();
2775 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2776 CI2Val->isPowerOf2()) {
2777 if (!CIVal->isPowerOf2()) {
2778 // CI2 << X can equal zero in some circumstances,
2779 // this simplification is unsafe if CI is zero.
2781 // We know it is safe if:
2782 // - The shift is nsw, we can't shift out the one bit.
2783 // - The shift is nuw, we can't shift out the one bit.
2786 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2787 CI2Val->isOneValue() || !CI->isZero()) {
2788 if (Pred == ICmpInst::ICMP_EQ)
2789 return ConstantInt::getFalse(RHS->getContext());
2790 if (Pred == ICmpInst::ICMP_NE)
2791 return ConstantInt::getTrue(RHS->getContext());
2794 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2795 if (Pred == ICmpInst::ICMP_UGT)
2796 return ConstantInt::getFalse(RHS->getContext());
2797 if (Pred == ICmpInst::ICMP_ULE)
2798 return ConstantInt::getTrue(RHS->getContext());
2803 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2804 LBO->getOperand(1) == RBO->getOperand(1)) {
2805 switch (LBO->getOpcode()) {
2808 case Instruction::UDiv:
2809 case Instruction::LShr:
2810 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2812 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2813 RBO->getOperand(0), Q, MaxRecurse - 1))
2816 case Instruction::SDiv:
2817 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2819 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2820 RBO->getOperand(0), Q, MaxRecurse - 1))
2823 case Instruction::AShr:
2824 if (!LBO->isExact() || !RBO->isExact())
2826 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2827 RBO->getOperand(0), Q, MaxRecurse - 1))
2830 case Instruction::Shl: {
2831 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2832 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2835 if (!NSW && ICmpInst::isSigned(Pred))
2837 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2838 RBO->getOperand(0), Q, MaxRecurse - 1))
2847 /// Simplify integer comparisons where at least one operand of the compare
2848 /// matches an integer min/max idiom.
2849 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2850 Value *RHS, const SimplifyQuery &Q,
2851 unsigned MaxRecurse) {
2852 Type *ITy = GetCompareTy(LHS); // The return type.
2854 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2855 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2857 // Signed variants on "max(a,b)>=a -> true".
2858 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2860 std::swap(A, B); // smax(A, B) pred A.
2861 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2862 // We analyze this as smax(A, B) pred A.
2864 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2865 (A == LHS || B == LHS)) {
2867 std::swap(A, B); // A pred smax(A, B).
2868 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2869 // We analyze this as smax(A, B) swapped-pred A.
2870 P = CmpInst::getSwappedPredicate(Pred);
2871 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2872 (A == RHS || B == RHS)) {
2874 std::swap(A, B); // smin(A, B) pred A.
2875 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2876 // We analyze this as smax(-A, -B) swapped-pred -A.
2877 // Note that we do not need to actually form -A or -B thanks to EqP.
2878 P = CmpInst::getSwappedPredicate(Pred);
2879 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2880 (A == LHS || B == LHS)) {
2882 std::swap(A, B); // A pred smin(A, B).
2883 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2884 // We analyze this as smax(-A, -B) pred -A.
2885 // Note that we do not need to actually form -A or -B thanks to EqP.
2888 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2889 // Cases correspond to "max(A, B) p A".
2893 case CmpInst::ICMP_EQ:
2894 case CmpInst::ICMP_SLE:
2895 // Equivalent to "A EqP B". This may be the same as the condition tested
2896 // in the max/min; if so, we can just return that.
2897 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2899 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2901 // Otherwise, see if "A EqP B" simplifies.
2903 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2906 case CmpInst::ICMP_NE:
2907 case CmpInst::ICMP_SGT: {
2908 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2909 // Equivalent to "A InvEqP B". This may be the same as the condition
2910 // tested in the max/min; if so, we can just return that.
2911 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2913 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2915 // Otherwise, see if "A InvEqP B" simplifies.
2917 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2921 case CmpInst::ICMP_SGE:
2923 return getTrue(ITy);
2924 case CmpInst::ICMP_SLT:
2926 return getFalse(ITy);
2930 // Unsigned variants on "max(a,b)>=a -> true".
2931 P = CmpInst::BAD_ICMP_PREDICATE;
2932 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2934 std::swap(A, B); // umax(A, B) pred A.
2935 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2936 // We analyze this as umax(A, B) pred A.
2938 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2939 (A == LHS || B == LHS)) {
2941 std::swap(A, B); // A pred umax(A, B).
2942 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2943 // We analyze this as umax(A, B) swapped-pred A.
2944 P = CmpInst::getSwappedPredicate(Pred);
2945 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2946 (A == RHS || B == RHS)) {
2948 std::swap(A, B); // umin(A, B) pred A.
2949 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2950 // We analyze this as umax(-A, -B) swapped-pred -A.
2951 // Note that we do not need to actually form -A or -B thanks to EqP.
2952 P = CmpInst::getSwappedPredicate(Pred);
2953 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2954 (A == LHS || B == LHS)) {
2956 std::swap(A, B); // A pred umin(A, B).
2957 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2958 // We analyze this as umax(-A, -B) pred -A.
2959 // Note that we do not need to actually form -A or -B thanks to EqP.
2962 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2963 // Cases correspond to "max(A, B) p A".
2967 case CmpInst::ICMP_EQ:
2968 case CmpInst::ICMP_ULE:
2969 // Equivalent to "A EqP B". This may be the same as the condition tested
2970 // in the max/min; if so, we can just return that.
2971 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2973 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2975 // Otherwise, see if "A EqP B" simplifies.
2977 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2980 case CmpInst::ICMP_NE:
2981 case CmpInst::ICMP_UGT: {
2982 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2983 // Equivalent to "A InvEqP B". This may be the same as the condition
2984 // tested in the max/min; if so, we can just return that.
2985 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2987 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2989 // Otherwise, see if "A InvEqP B" simplifies.
2991 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2995 case CmpInst::ICMP_UGE:
2997 return getTrue(ITy);
2998 case CmpInst::ICMP_ULT:
3000 return getFalse(ITy);
3004 // Variants on "max(x,y) >= min(x,z)".
3006 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3007 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3008 (A == C || A == D || B == C || B == D)) {
3009 // max(x, ?) pred min(x, ?).
3010 if (Pred == CmpInst::ICMP_SGE)
3012 return getTrue(ITy);
3013 if (Pred == CmpInst::ICMP_SLT)
3015 return getFalse(ITy);
3016 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3017 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3018 (A == C || A == D || B == C || B == D)) {
3019 // min(x, ?) pred max(x, ?).
3020 if (Pred == CmpInst::ICMP_SLE)
3022 return getTrue(ITy);
3023 if (Pred == CmpInst::ICMP_SGT)
3025 return getFalse(ITy);
3026 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3027 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3028 (A == C || A == D || B == C || B == D)) {
3029 // max(x, ?) pred min(x, ?).
3030 if (Pred == CmpInst::ICMP_UGE)
3032 return getTrue(ITy);
3033 if (Pred == CmpInst::ICMP_ULT)
3035 return getFalse(ITy);
3036 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3037 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3038 (A == C || A == D || B == C || B == D)) {
3039 // min(x, ?) pred max(x, ?).
3040 if (Pred == CmpInst::ICMP_ULE)
3042 return getTrue(ITy);
3043 if (Pred == CmpInst::ICMP_UGT)
3045 return getFalse(ITy);
3051 /// Given operands for an ICmpInst, see if we can fold the result.
3052 /// If not, this returns null.
3053 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3054 const SimplifyQuery &Q, unsigned MaxRecurse) {
3055 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3056 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3058 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3059 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3060 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3062 // If we have a constant, make sure it is on the RHS.
3063 std::swap(LHS, RHS);
3064 Pred = CmpInst::getSwappedPredicate(Pred);
3067 Type *ITy = GetCompareTy(LHS); // The return type.
3069 // icmp X, X -> true/false
3070 // icmp X, undef -> true/false because undef could be X.
3071 if (LHS == RHS || isa<UndefValue>(RHS))
3072 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3074 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3077 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3080 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3083 // If both operands have range metadata, use the metadata
3084 // to simplify the comparison.
3085 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3086 auto RHS_Instr = cast<Instruction>(RHS);
3087 auto LHS_Instr = cast<Instruction>(LHS);
3089 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3090 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3091 auto RHS_CR = getConstantRangeFromMetadata(
3092 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3093 auto LHS_CR = getConstantRangeFromMetadata(
3094 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3096 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3097 if (Satisfied_CR.contains(LHS_CR))
3098 return ConstantInt::getTrue(RHS->getContext());
3100 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3101 CmpInst::getInversePredicate(Pred), RHS_CR);
3102 if (InversedSatisfied_CR.contains(LHS_CR))
3103 return ConstantInt::getFalse(RHS->getContext());
3107 // Compare of cast, for example (zext X) != 0 -> X != 0
3108 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3109 Instruction *LI = cast<CastInst>(LHS);
3110 Value *SrcOp = LI->getOperand(0);
3111 Type *SrcTy = SrcOp->getType();
3112 Type *DstTy = LI->getType();
3114 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3115 // if the integer type is the same size as the pointer type.
3116 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3117 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3118 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3119 // Transfer the cast to the constant.
3120 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3121 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3124 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3125 if (RI->getOperand(0)->getType() == SrcTy)
3126 // Compare without the cast.
3127 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3133 if (isa<ZExtInst>(LHS)) {
3134 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3136 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3137 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3138 // Compare X and Y. Note that signed predicates become unsigned.
3139 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3140 SrcOp, RI->getOperand(0), Q,
3144 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3145 // too. If not, then try to deduce the result of the comparison.
3146 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3147 // Compute the constant that would happen if we truncated to SrcTy then
3148 // reextended to DstTy.
3149 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3150 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3152 // If the re-extended constant didn't change then this is effectively
3153 // also a case of comparing two zero-extended values.
3154 if (RExt == CI && MaxRecurse)
3155 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3156 SrcOp, Trunc, Q, MaxRecurse-1))
3159 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3160 // there. Use this to work out the result of the comparison.
3163 default: llvm_unreachable("Unknown ICmp predicate!");
3165 case ICmpInst::ICMP_EQ:
3166 case ICmpInst::ICMP_UGT:
3167 case ICmpInst::ICMP_UGE:
3168 return ConstantInt::getFalse(CI->getContext());
3170 case ICmpInst::ICMP_NE:
3171 case ICmpInst::ICMP_ULT:
3172 case ICmpInst::ICMP_ULE:
3173 return ConstantInt::getTrue(CI->getContext());
3175 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3176 // is non-negative then LHS <s RHS.
3177 case ICmpInst::ICMP_SGT:
3178 case ICmpInst::ICMP_SGE:
3179 return CI->getValue().isNegative() ?
3180 ConstantInt::getTrue(CI->getContext()) :
3181 ConstantInt::getFalse(CI->getContext());
3183 case ICmpInst::ICMP_SLT:
3184 case ICmpInst::ICMP_SLE:
3185 return CI->getValue().isNegative() ?
3186 ConstantInt::getFalse(CI->getContext()) :
3187 ConstantInt::getTrue(CI->getContext());
3193 if (isa<SExtInst>(LHS)) {
3194 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3196 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3197 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3198 // Compare X and Y. Note that the predicate does not change.
3199 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3203 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3204 // too. If not, then try to deduce the result of the comparison.
3205 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3206 // Compute the constant that would happen if we truncated to SrcTy then
3207 // reextended to DstTy.
3208 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3209 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3211 // If the re-extended constant didn't change then this is effectively
3212 // also a case of comparing two sign-extended values.
3213 if (RExt == CI && MaxRecurse)
3214 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3217 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3218 // bits there. Use this to work out the result of the comparison.
3221 default: llvm_unreachable("Unknown ICmp predicate!");
3222 case ICmpInst::ICMP_EQ:
3223 return ConstantInt::getFalse(CI->getContext());
3224 case ICmpInst::ICMP_NE:
3225 return ConstantInt::getTrue(CI->getContext());
3227 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3229 case ICmpInst::ICMP_SGT:
3230 case ICmpInst::ICMP_SGE:
3231 return CI->getValue().isNegative() ?
3232 ConstantInt::getTrue(CI->getContext()) :
3233 ConstantInt::getFalse(CI->getContext());
3234 case ICmpInst::ICMP_SLT:
3235 case ICmpInst::ICMP_SLE:
3236 return CI->getValue().isNegative() ?
3237 ConstantInt::getFalse(CI->getContext()) :
3238 ConstantInt::getTrue(CI->getContext());
3240 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3242 case ICmpInst::ICMP_UGT:
3243 case ICmpInst::ICMP_UGE:
3244 // Comparison is true iff the LHS <s 0.
3246 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3247 Constant::getNullValue(SrcTy),
3251 case ICmpInst::ICMP_ULT:
3252 case ICmpInst::ICMP_ULE:
3253 // Comparison is true iff the LHS >=s 0.
3255 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3256 Constant::getNullValue(SrcTy),
3266 // icmp eq|ne X, Y -> false|true if X != Y
3267 if (ICmpInst::isEquality(Pred) &&
3268 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3269 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3272 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3275 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3278 // Simplify comparisons of related pointers using a powerful, recursive
3279 // GEP-walk when we have target data available..
3280 if (LHS->getType()->isPointerTy())
3281 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3284 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3285 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3286 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3287 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3288 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3289 Q.DL.getTypeSizeInBits(CRHS->getType()))
3290 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3291 CLHS->getPointerOperand(),
3292 CRHS->getPointerOperand()))
3295 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3296 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3297 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3298 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3299 (ICmpInst::isEquality(Pred) ||
3300 (GLHS->isInBounds() && GRHS->isInBounds() &&
3301 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3302 // The bases are equal and the indices are constant. Build a constant
3303 // expression GEP with the same indices and a null base pointer to see
3304 // what constant folding can make out of it.
3305 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3306 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3307 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3308 GLHS->getSourceElementType(), Null, IndicesLHS);
3310 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3311 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3312 GLHS->getSourceElementType(), Null, IndicesRHS);
3313 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3318 // If the comparison is with the result of a select instruction, check whether
3319 // comparing with either branch of the select always yields the same value.
3320 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3321 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3324 // If the comparison is with the result of a phi instruction, check whether
3325 // doing the compare with each incoming phi value yields a common result.
3326 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3327 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3333 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3334 const SimplifyQuery &Q) {
3335 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3338 /// Given operands for an FCmpInst, see if we can fold the result.
3339 /// If not, this returns null.
3340 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3341 FastMathFlags FMF, const SimplifyQuery &Q,
3342 unsigned MaxRecurse) {
3343 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3344 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3346 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3347 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3348 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3350 // If we have a constant, make sure it is on the RHS.
3351 std::swap(LHS, RHS);
3352 Pred = CmpInst::getSwappedPredicate(Pred);
3355 // Fold trivial predicates.
3356 Type *RetTy = GetCompareTy(LHS);
3357 if (Pred == FCmpInst::FCMP_FALSE)
3358 return getFalse(RetTy);
3359 if (Pred == FCmpInst::FCMP_TRUE)
3360 return getTrue(RetTy);
3362 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3364 if (Pred == FCmpInst::FCMP_UNO)
3365 return getFalse(RetTy);
3366 if (Pred == FCmpInst::FCMP_ORD)
3367 return getTrue(RetTy);
3370 // NaN is unordered; NaN is not ordered.
3371 assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&
3372 "Comparison must be either ordered or unordered");
3373 if (match(RHS, m_NaN()))
3374 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3376 // fcmp pred x, undef and fcmp pred undef, x
3377 // fold to true if unordered, false if ordered
3378 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3379 // Choosing NaN for the undef will always make unordered comparison succeed
3380 // and ordered comparison fail.
3381 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3384 // fcmp x,x -> true/false. Not all compares are foldable.
3386 if (CmpInst::isTrueWhenEqual(Pred))
3387 return getTrue(RetTy);
3388 if (CmpInst::isFalseWhenEqual(Pred))
3389 return getFalse(RetTy);
3392 // Handle fcmp with constant RHS.
3394 if (match(RHS, m_APFloat(C))) {
3395 // Check whether the constant is an infinity.
3396 if (C->isInfinity()) {
3397 if (C->isNegative()) {
3399 case FCmpInst::FCMP_OLT:
3400 // No value is ordered and less than negative infinity.
3401 return getFalse(RetTy);
3402 case FCmpInst::FCMP_UGE:
3403 // All values are unordered with or at least negative infinity.
3404 return getTrue(RetTy);
3410 case FCmpInst::FCMP_OGT:
3411 // No value is ordered and greater than infinity.
3412 return getFalse(RetTy);
3413 case FCmpInst::FCMP_ULE:
3414 // All values are unordered with and at most infinity.
3415 return getTrue(RetTy);
3423 case FCmpInst::FCMP_UGE:
3424 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3425 return getTrue(RetTy);
3427 case FCmpInst::FCMP_OLT:
3429 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3430 return getFalse(RetTy);
3435 } else if (C->isNegative()) {
3436 assert(!C->isNaN() && "Unexpected NaN constant!");
3437 // TODO: We can catch more cases by using a range check rather than
3438 // relying on CannotBeOrderedLessThanZero.
3440 case FCmpInst::FCMP_UGE:
3441 case FCmpInst::FCMP_UGT:
3442 case FCmpInst::FCMP_UNE:
3443 // (X >= 0) implies (X > C) when (C < 0)
3444 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3445 return getTrue(RetTy);
3447 case FCmpInst::FCMP_OEQ:
3448 case FCmpInst::FCMP_OLE:
3449 case FCmpInst::FCMP_OLT:
3450 // (X >= 0) implies !(X < C) when (C < 0)
3451 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3452 return getFalse(RetTy);
3460 // If the comparison is with the result of a select instruction, check whether
3461 // comparing with either branch of the select always yields the same value.
3462 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3463 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3466 // If the comparison is with the result of a phi instruction, check whether
3467 // doing the compare with each incoming phi value yields a common result.
3468 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3469 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3475 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3476 FastMathFlags FMF, const SimplifyQuery &Q) {
3477 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3480 /// See if V simplifies when its operand Op is replaced with RepOp.
3481 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3482 const SimplifyQuery &Q,
3483 unsigned MaxRecurse) {
3484 // Trivial replacement.
3488 // We cannot replace a constant, and shouldn't even try.
3489 if (isa<Constant>(Op))
3492 auto *I = dyn_cast<Instruction>(V);
3496 // If this is a binary operator, try to simplify it with the replaced op.
3497 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3499 // %cmp = icmp eq i32 %x, 2147483647
3500 // %add = add nsw i32 %x, 1
3501 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3503 // We can't replace %sel with %add unless we strip away the flags.
3504 if (isa<OverflowingBinaryOperator>(B))
3505 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3507 if (isa<PossiblyExactOperator>(B))
3512 if (B->getOperand(0) == Op)
3513 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3515 if (B->getOperand(1) == Op)
3516 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3521 // Same for CmpInsts.
3522 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3524 if (C->getOperand(0) == Op)
3525 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3527 if (C->getOperand(1) == Op)
3528 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3534 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3536 SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3537 transform(GEP->operands(), NewOps.begin(),
3538 [&](Value *V) { return V == Op ? RepOp : V; });
3539 return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3544 // TODO: We could hand off more cases to instsimplify here.
3546 // If all operands are constant after substituting Op for RepOp then we can
3547 // constant fold the instruction.
3548 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3549 // Build a list of all constant operands.
3550 SmallVector<Constant *, 8> ConstOps;
3551 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3552 if (I->getOperand(i) == Op)
3553 ConstOps.push_back(CRepOp);
3554 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3555 ConstOps.push_back(COp);
3560 // All operands were constants, fold it.
3561 if (ConstOps.size() == I->getNumOperands()) {
3562 if (CmpInst *C = dyn_cast<CmpInst>(I))
3563 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3564 ConstOps[1], Q.DL, Q.TLI);
3566 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3567 if (!LI->isVolatile())
3568 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3570 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3577 /// Try to simplify a select instruction when its condition operand is an
3578 /// integer comparison where one operand of the compare is a constant.
3579 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3580 const APInt *Y, bool TrueWhenUnset) {
3583 // (X & Y) == 0 ? X & ~Y : X --> X
3584 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3585 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3587 return TrueWhenUnset ? FalseVal : TrueVal;
3589 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3590 // (X & Y) != 0 ? X : X & ~Y --> X
3591 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3593 return TrueWhenUnset ? FalseVal : TrueVal;
3595 if (Y->isPowerOf2()) {
3596 // (X & Y) == 0 ? X | Y : X --> X | Y
3597 // (X & Y) != 0 ? X | Y : X --> X
3598 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3600 return TrueWhenUnset ? TrueVal : FalseVal;
3602 // (X & Y) == 0 ? X : X | Y --> X
3603 // (X & Y) != 0 ? X : X | Y --> X | Y
3604 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3606 return TrueWhenUnset ? TrueVal : FalseVal;
3612 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3614 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3615 ICmpInst::Predicate Pred,
3616 Value *TrueVal, Value *FalseVal) {
3619 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3622 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3623 Pred == ICmpInst::ICMP_EQ);
3626 /// Try to simplify a select instruction when its condition operand is an
3627 /// integer comparison.
3628 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3629 Value *FalseVal, const SimplifyQuery &Q,
3630 unsigned MaxRecurse) {
3631 ICmpInst::Predicate Pred;
3632 Value *CmpLHS, *CmpRHS;
3633 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3636 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3639 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3640 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3641 Pred == ICmpInst::ICMP_EQ))
3645 // Check for other compares that behave like bit test.
3646 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3650 // If we have an equality comparison, then we know the value in one of the
3651 // arms of the select. See if substituting this value into the arm and
3652 // simplifying the result yields the same value as the other arm.
3653 if (Pred == ICmpInst::ICMP_EQ) {
3654 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3656 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3659 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3661 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3664 } else if (Pred == ICmpInst::ICMP_NE) {
3665 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3667 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3670 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3672 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3680 /// Given operands for a SelectInst, see if we can fold the result.
3681 /// If not, this returns null.
3682 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3683 const SimplifyQuery &Q, unsigned MaxRecurse) {
3684 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3685 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3686 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3687 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3689 // select undef, X, Y -> X or Y
3690 if (isa<UndefValue>(CondC))
3691 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3693 // TODO: Vector constants with undef elements don't simplify.
3695 // select true, X, Y -> X
3696 if (CondC->isAllOnesValue())
3698 // select false, X, Y -> Y
3699 if (CondC->isNullValue())
3703 // select ?, X, X -> X
3704 if (TrueVal == FalseVal)
3707 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3709 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3713 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3719 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3720 const SimplifyQuery &Q) {
3721 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3724 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3725 /// If not, this returns null.
3726 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3727 const SimplifyQuery &Q, unsigned) {
3728 // The type of the GEP pointer operand.
3730 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3732 // getelementptr P -> P.
3733 if (Ops.size() == 1)
3736 // Compute the (pointer) type returned by the GEP instruction.
3737 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3738 Type *GEPTy = PointerType::get(LastType, AS);
3739 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3740 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3741 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3742 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3744 if (isa<UndefValue>(Ops[0]))
3745 return UndefValue::get(GEPTy);
3747 if (Ops.size() == 2) {
3748 // getelementptr P, 0 -> P.
3749 if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3753 if (Ty->isSized()) {
3756 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3757 // getelementptr P, N -> P if P points to a type of zero size.
3758 if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3761 // The following transforms are only safe if the ptrtoint cast
3762 // doesn't truncate the pointers.
3763 if (Ops[1]->getType()->getScalarSizeInBits() ==
3764 Q.DL.getIndexSizeInBits(AS)) {
3765 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3766 if (match(P, m_Zero()))
3767 return Constant::getNullValue(GEPTy);
3769 if (match(P, m_PtrToInt(m_Value(Temp))))
3770 if (Temp->getType() == GEPTy)
3775 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3776 if (TyAllocSize == 1 &&
3777 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3778 if (Value *R = PtrToIntOrZero(P))
3781 // getelementptr V, (ashr (sub P, V), C) -> Q
3782 // if P points to a type of size 1 << C.
3784 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3785 m_ConstantInt(C))) &&
3786 TyAllocSize == 1ULL << C)
3787 if (Value *R = PtrToIntOrZero(P))
3790 // getelementptr V, (sdiv (sub P, V), C) -> Q
3791 // if P points to a type of size C.
3793 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3794 m_SpecificInt(TyAllocSize))))
3795 if (Value *R = PtrToIntOrZero(P))
3801 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3802 all_of(Ops.slice(1).drop_back(1),
3803 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3805 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3806 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3807 APInt BasePtrOffset(IdxWidth, 0);
3808 Value *StrippedBasePtr =
3809 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3812 // gep (gep V, C), (sub 0, V) -> C
3813 if (match(Ops.back(),
3814 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3815 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3816 return ConstantExpr::getIntToPtr(CI, GEPTy);
3818 // gep (gep V, C), (xor V, -1) -> C-1
3819 if (match(Ops.back(),
3820 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3821 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3822 return ConstantExpr::getIntToPtr(CI, GEPTy);
3827 // Check to see if this is constant foldable.
3828 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3831 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3833 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3838 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3839 const SimplifyQuery &Q) {
3840 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3843 /// Given operands for an InsertValueInst, see if we can fold the result.
3844 /// If not, this returns null.
3845 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3846 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3848 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3849 if (Constant *CVal = dyn_cast<Constant>(Val))
3850 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3852 // insertvalue x, undef, n -> x
3853 if (match(Val, m_Undef()))
3856 // insertvalue x, (extractvalue y, n), n
3857 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3858 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3859 EV->getIndices() == Idxs) {
3860 // insertvalue undef, (extractvalue y, n), n -> y
3861 if (match(Agg, m_Undef()))
3862 return EV->getAggregateOperand();
3864 // insertvalue y, (extractvalue y, n), n -> y
3865 if (Agg == EV->getAggregateOperand())
3872 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3873 ArrayRef<unsigned> Idxs,
3874 const SimplifyQuery &Q) {
3875 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3878 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3879 const SimplifyQuery &Q) {
3880 // Try to constant fold.
3881 auto *VecC = dyn_cast<Constant>(Vec);
3882 auto *ValC = dyn_cast<Constant>(Val);
3883 auto *IdxC = dyn_cast<Constant>(Idx);
3884 if (VecC && ValC && IdxC)
3885 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3887 // Fold into undef if index is out of bounds.
3888 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3889 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3890 if (CI->uge(NumElements))
3891 return UndefValue::get(Vec->getType());
3894 // If index is undef, it might be out of bounds (see above case)
3895 if (isa<UndefValue>(Idx))
3896 return UndefValue::get(Vec->getType());
3901 /// Given operands for an ExtractValueInst, see if we can fold the result.
3902 /// If not, this returns null.
3903 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3904 const SimplifyQuery &, unsigned) {
3905 if (auto *CAgg = dyn_cast<Constant>(Agg))
3906 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3908 // extractvalue x, (insertvalue y, elt, n), n -> elt
3909 unsigned NumIdxs = Idxs.size();
3910 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3911 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3912 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3913 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3914 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3915 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3916 Idxs.slice(0, NumCommonIdxs)) {
3917 if (NumIdxs == NumInsertValueIdxs)
3918 return IVI->getInsertedValueOperand();
3926 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3927 const SimplifyQuery &Q) {
3928 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3931 /// Given operands for an ExtractElementInst, see if we can fold the result.
3932 /// If not, this returns null.
3933 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3935 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3936 if (auto *CIdx = dyn_cast<Constant>(Idx))
3937 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3939 // The index is not relevant if our vector is a splat.
3940 if (auto *Splat = CVec->getSplatValue())
3943 if (isa<UndefValue>(Vec))
3944 return UndefValue::get(Vec->getType()->getVectorElementType());
3947 // If extracting a specified index from the vector, see if we can recursively
3948 // find a previously computed scalar that was inserted into the vector.
3949 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
3950 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
3951 // definitely out of bounds, thus undefined result
3952 return UndefValue::get(Vec->getType()->getVectorElementType());
3953 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3957 // An undef extract index can be arbitrarily chosen to be an out-of-range
3958 // index value, which would result in the instruction being undef.
3959 if (isa<UndefValue>(Idx))
3960 return UndefValue::get(Vec->getType()->getVectorElementType());
3965 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3966 const SimplifyQuery &Q) {
3967 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3970 /// See if we can fold the given phi. If not, returns null.
3971 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3972 // If all of the PHI's incoming values are the same then replace the PHI node
3973 // with the common value.
3974 Value *CommonValue = nullptr;
3975 bool HasUndefInput = false;
3976 for (Value *Incoming : PN->incoming_values()) {
3977 // If the incoming value is the phi node itself, it can safely be skipped.
3978 if (Incoming == PN) continue;
3979 if (isa<UndefValue>(Incoming)) {
3980 // Remember that we saw an undef value, but otherwise ignore them.
3981 HasUndefInput = true;
3984 if (CommonValue && Incoming != CommonValue)
3985 return nullptr; // Not the same, bail out.
3986 CommonValue = Incoming;
3989 // If CommonValue is null then all of the incoming values were either undef or
3990 // equal to the phi node itself.
3992 return UndefValue::get(PN->getType());
3994 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3995 // instruction, we cannot return X as the result of the PHI node unless it
3996 // dominates the PHI block.
3998 return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4003 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4004 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4005 if (auto *C = dyn_cast<Constant>(Op))
4006 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4008 if (auto *CI = dyn_cast<CastInst>(Op)) {
4009 auto *Src = CI->getOperand(0);
4010 Type *SrcTy = Src->getType();
4011 Type *MidTy = CI->getType();
4013 if (Src->getType() == Ty) {
4014 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4015 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4017 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4019 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4021 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4022 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4023 SrcIntPtrTy, MidIntPtrTy,
4024 DstIntPtrTy) == Instruction::BitCast)
4030 if (CastOpc == Instruction::BitCast)
4031 if (Op->getType() == Ty)
4037 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4038 const SimplifyQuery &Q) {
4039 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4042 /// For the given destination element of a shuffle, peek through shuffles to
4043 /// match a root vector source operand that contains that element in the same
4044 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4045 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4046 int MaskVal, Value *RootVec,
4047 unsigned MaxRecurse) {
4051 // Bail out if any mask value is undefined. That kind of shuffle may be
4052 // simplified further based on demanded bits or other folds.
4056 // The mask value chooses which source operand we need to look at next.
4057 int InVecNumElts = Op0->getType()->getVectorNumElements();
4058 int RootElt = MaskVal;
4059 Value *SourceOp = Op0;
4060 if (MaskVal >= InVecNumElts) {
4061 RootElt = MaskVal - InVecNumElts;
4065 // If the source operand is a shuffle itself, look through it to find the
4066 // matching root vector.
4067 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4068 return foldIdentityShuffles(
4069 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4070 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4073 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4076 // The source operand is not a shuffle. Initialize the root vector value for
4077 // this shuffle if that has not been done yet.
4081 // Give up as soon as a source operand does not match the existing root value.
4082 if (RootVec != SourceOp)
4085 // The element must be coming from the same lane in the source vector
4086 // (although it may have crossed lanes in intermediate shuffles).
4087 if (RootElt != DestElt)
4093 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4094 Type *RetTy, const SimplifyQuery &Q,
4095 unsigned MaxRecurse) {
4096 if (isa<UndefValue>(Mask))
4097 return UndefValue::get(RetTy);
4099 Type *InVecTy = Op0->getType();
4100 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4101 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4103 SmallVector<int, 32> Indices;
4104 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4105 assert(MaskNumElts == Indices.size() &&
4106 "Size of Indices not same as number of mask elements?");
4108 // Canonicalization: If mask does not select elements from an input vector,
4109 // replace that input vector with undef.
4110 bool MaskSelects0 = false, MaskSelects1 = false;
4111 for (unsigned i = 0; i != MaskNumElts; ++i) {
4112 if (Indices[i] == -1)
4114 if ((unsigned)Indices[i] < InVecNumElts)
4115 MaskSelects0 = true;
4117 MaskSelects1 = true;
4120 Op0 = UndefValue::get(InVecTy);
4122 Op1 = UndefValue::get(InVecTy);
4124 auto *Op0Const = dyn_cast<Constant>(Op0);
4125 auto *Op1Const = dyn_cast<Constant>(Op1);
4127 // If all operands are constant, constant fold the shuffle.
4128 if (Op0Const && Op1Const)
4129 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4131 // Canonicalization: if only one input vector is constant, it shall be the
4133 if (Op0Const && !Op1Const) {
4134 std::swap(Op0, Op1);
4135 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4138 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4139 // value type is same as the input vectors' type.
4140 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4141 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4142 OpShuf->getMask()->getSplatValue())
4145 // Don't fold a shuffle with undef mask elements. This may get folded in a
4146 // better way using demanded bits or other analysis.
4147 // TODO: Should we allow this?
4148 if (find(Indices, -1) != Indices.end())
4151 // Check if every element of this shuffle can be mapped back to the
4152 // corresponding element of a single root vector. If so, we don't need this
4153 // shuffle. This handles simple identity shuffles as well as chains of
4154 // shuffles that may widen/narrow and/or move elements across lanes and back.
4155 Value *RootVec = nullptr;
4156 for (unsigned i = 0; i != MaskNumElts; ++i) {
4157 // Note that recursion is limited for each vector element, so if any element
4158 // exceeds the limit, this will fail to simplify.
4160 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4162 // We can't replace a widening/narrowing shuffle with one of its operands.
4163 if (!RootVec || RootVec->getType() != RetTy)
4169 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4170 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4171 Type *RetTy, const SimplifyQuery &Q) {
4172 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4175 static Constant *propagateNaN(Constant *In) {
4176 // If the input is a vector with undef elements, just return a default NaN.
4178 return ConstantFP::getNaN(In->getType());
4180 // Propagate the existing NaN constant when possible.
4181 // TODO: Should we quiet a signaling NaN?
4185 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4186 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4187 return ConstantFP::getNaN(Op0->getType());
4189 if (match(Op0, m_NaN()))
4190 return propagateNaN(cast<Constant>(Op0));
4191 if (match(Op1, m_NaN()))
4192 return propagateNaN(cast<Constant>(Op1));
4197 /// Given operands for an FAdd, see if we can fold the result. If not, this
4199 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4200 const SimplifyQuery &Q, unsigned MaxRecurse) {
4201 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4204 if (Constant *C = simplifyFPBinop(Op0, Op1))
4208 if (match(Op1, m_NegZeroFP()))
4211 // fadd X, 0 ==> X, when we know X is not -0
4212 if (match(Op1, m_PosZeroFP()) &&
4213 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4216 // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4217 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4218 // Negative zeros are allowed because we always end up with positive zero:
4219 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4220 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4221 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4222 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4223 if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4224 match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4225 return ConstantFP::getNullValue(Op0->getType());
4230 /// Given operands for an FSub, see if we can fold the result. If not, this
4232 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4233 const SimplifyQuery &Q, unsigned MaxRecurse) {
4234 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4237 if (Constant *C = simplifyFPBinop(Op0, Op1))
4241 if (match(Op1, m_PosZeroFP()))
4244 // fsub X, -0 ==> X, when we know X is not -0
4245 if (match(Op1, m_NegZeroFP()) &&
4246 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4249 // fsub -0.0, (fsub -0.0, X) ==> X
4251 if (match(Op0, m_NegZeroFP()) &&
4252 match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4255 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4256 if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4257 match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4260 // fsub nnan x, x ==> 0.0
4261 if (FMF.noNaNs() && Op0 == Op1)
4262 return Constant::getNullValue(Op0->getType());
4267 /// Given the operands for an FMul, see if we can fold the result
4268 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4269 const SimplifyQuery &Q, unsigned MaxRecurse) {
4270 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4273 if (Constant *C = simplifyFPBinop(Op0, Op1))
4276 // fmul X, 1.0 ==> X
4277 if (match(Op1, m_FPOne()))
4280 // fmul nnan nsz X, 0 ==> 0
4281 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4282 return ConstantFP::getNullValue(Op0->getType());
4284 // sqrt(X) * sqrt(X) --> X, if we can:
4285 // 1. Remove the intermediate rounding (reassociate).
4286 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4287 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4289 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4290 FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4296 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4297 const SimplifyQuery &Q) {
4298 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4302 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4303 const SimplifyQuery &Q) {
4304 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4307 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4308 const SimplifyQuery &Q) {
4309 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4312 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4313 const SimplifyQuery &Q, unsigned) {
4314 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4317 if (Constant *C = simplifyFPBinop(Op0, Op1))
4321 if (match(Op1, m_FPOne()))
4325 // Requires that NaNs are off (X could be zero) and signed zeroes are
4326 // ignored (X could be positive or negative, so the output sign is unknown).
4327 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4328 return ConstantFP::getNullValue(Op0->getType());
4331 // X / X -> 1.0 is legal when NaNs are ignored.
4332 // We can ignore infinities because INF/INF is NaN.
4334 return ConstantFP::get(Op0->getType(), 1.0);
4336 // (X * Y) / Y --> X if we can reassociate to the above form.
4338 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4341 // -X / X -> -1.0 and
4342 // X / -X -> -1.0 are legal when NaNs are ignored.
4343 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4344 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4345 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4346 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4347 BinaryOperator::getFNegArgument(Op1) == Op0))
4348 return ConstantFP::get(Op0->getType(), -1.0);
4354 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4355 const SimplifyQuery &Q) {
4356 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4359 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4360 const SimplifyQuery &Q, unsigned) {
4361 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4364 if (Constant *C = simplifyFPBinop(Op0, Op1))
4367 // Unlike fdiv, the result of frem always matches the sign of the dividend.
4368 // The constant match may include undef elements in a vector, so return a full
4369 // zero constant as the result.
4372 if (match(Op0, m_PosZeroFP()))
4373 return ConstantFP::getNullValue(Op0->getType());
4375 if (match(Op0, m_NegZeroFP()))
4376 return ConstantFP::getNegativeZero(Op0->getType());
4382 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4383 const SimplifyQuery &Q) {
4384 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4387 //=== Helper functions for higher up the class hierarchy.
4389 /// Given operands for a BinaryOperator, see if we can fold the result.
4390 /// If not, this returns null.
4391 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4392 const SimplifyQuery &Q, unsigned MaxRecurse) {
4394 case Instruction::Add:
4395 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4396 case Instruction::Sub:
4397 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4398 case Instruction::Mul:
4399 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4400 case Instruction::SDiv:
4401 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4402 case Instruction::UDiv:
4403 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4404 case Instruction::SRem:
4405 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4406 case Instruction::URem:
4407 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4408 case Instruction::Shl:
4409 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4410 case Instruction::LShr:
4411 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4412 case Instruction::AShr:
4413 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4414 case Instruction::And:
4415 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4416 case Instruction::Or:
4417 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4418 case Instruction::Xor:
4419 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4420 case Instruction::FAdd:
4421 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4422 case Instruction::FSub:
4423 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4424 case Instruction::FMul:
4425 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4426 case Instruction::FDiv:
4427 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4428 case Instruction::FRem:
4429 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4431 llvm_unreachable("Unexpected opcode");
4435 /// Given operands for a BinaryOperator, see if we can fold the result.
4436 /// If not, this returns null.
4437 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4438 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4439 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4440 const FastMathFlags &FMF, const SimplifyQuery &Q,
4441 unsigned MaxRecurse) {
4443 case Instruction::FAdd:
4444 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4445 case Instruction::FSub:
4446 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4447 case Instruction::FMul:
4448 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4449 case Instruction::FDiv:
4450 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4452 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4456 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4457 const SimplifyQuery &Q) {
4458 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4461 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4462 FastMathFlags FMF, const SimplifyQuery &Q) {
4463 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4466 /// Given operands for a CmpInst, see if we can fold the result.
4467 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4468 const SimplifyQuery &Q, unsigned MaxRecurse) {
4469 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4470 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4471 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4474 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4475 const SimplifyQuery &Q) {
4476 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4479 static bool IsIdempotent(Intrinsic::ID ID) {
4481 default: return false;
4483 // Unary idempotent: f(f(x)) = f(x)
4484 case Intrinsic::fabs:
4485 case Intrinsic::floor:
4486 case Intrinsic::ceil:
4487 case Intrinsic::trunc:
4488 case Intrinsic::rint:
4489 case Intrinsic::nearbyint:
4490 case Intrinsic::round:
4491 case Intrinsic::canonicalize:
4496 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4497 const DataLayout &DL) {
4498 GlobalValue *PtrSym;
4500 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4503 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4504 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4505 Type *Int32PtrTy = Int32Ty->getPointerTo();
4506 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4508 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4509 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4512 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4513 if (OffsetInt % 4 != 0)
4516 Constant *C = ConstantExpr::getGetElementPtr(
4517 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4518 ConstantInt::get(Int64Ty, OffsetInt / 4));
4519 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4523 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4527 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4528 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4533 if (LoadedCE->getOpcode() != Instruction::Sub)
4536 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4537 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4539 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4541 Constant *LoadedRHS = LoadedCE->getOperand(1);
4542 GlobalValue *LoadedRHSSym;
4543 APInt LoadedRHSOffset;
4544 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4546 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4549 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4552 static bool maskIsAllZeroOrUndef(Value *Mask) {
4553 auto *ConstMask = dyn_cast<Constant>(Mask);
4556 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4558 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4560 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4561 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4568 template <typename IterTy>
4569 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4570 const SimplifyQuery &Q, unsigned MaxRecurse) {
4571 Intrinsic::ID IID = F->getIntrinsicID();
4572 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4575 if (NumOperands == 1) {
4576 // Perform idempotent optimizations
4577 if (IsIdempotent(IID)) {
4578 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4579 if (II->getIntrinsicID() == IID)
4584 Value *IIOperand = *ArgBegin;
4587 case Intrinsic::fabs: {
4588 if (SignBitMustBeZero(IIOperand, Q.TLI))
4592 case Intrinsic::bswap: {
4593 // bswap(bswap(x)) -> x
4594 if (match(IIOperand, m_BSwap(m_Value(X))))
4598 case Intrinsic::bitreverse: {
4599 // bitreverse(bitreverse(x)) -> x
4600 if (match(IIOperand, m_BitReverse(m_Value(X))))
4604 case Intrinsic::exp: {
4606 if (Q.CxtI->hasAllowReassoc() &&
4607 match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X))))
4611 case Intrinsic::exp2: {
4612 // exp2(log2(x)) -> x
4613 if (Q.CxtI->hasAllowReassoc() &&
4614 match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X))))
4618 case Intrinsic::log: {
4620 if (Q.CxtI->hasAllowReassoc() &&
4621 match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X))))
4625 case Intrinsic::log2: {
4626 // log2(exp2(x)) -> x
4627 if (Q.CxtI->hasAllowReassoc() &&
4628 match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) {
4639 if (NumOperands == 2) {
4640 Value *LHS = *ArgBegin;
4641 Value *RHS = *(ArgBegin + 1);
4642 Type *ReturnType = F->getReturnType();
4645 case Intrinsic::usub_with_overflow:
4646 case Intrinsic::ssub_with_overflow: {
4647 // X - X -> { 0, false }
4649 return Constant::getNullValue(ReturnType);
4651 // X - undef -> undef
4652 // undef - X -> undef
4653 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4654 return UndefValue::get(ReturnType);
4658 case Intrinsic::uadd_with_overflow:
4659 case Intrinsic::sadd_with_overflow: {
4660 // X + undef -> undef
4661 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4662 return UndefValue::get(ReturnType);
4666 case Intrinsic::umul_with_overflow:
4667 case Intrinsic::smul_with_overflow: {
4668 // 0 * X -> { 0, false }
4669 // X * 0 -> { 0, false }
4670 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4671 return Constant::getNullValue(ReturnType);
4673 // undef * X -> { 0, false }
4674 // X * undef -> { 0, false }
4675 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4676 return Constant::getNullValue(ReturnType);
4680 case Intrinsic::load_relative: {
4681 Constant *C0 = dyn_cast<Constant>(LHS);
4682 Constant *C1 = dyn_cast<Constant>(RHS);
4684 return SimplifyRelativeLoad(C0, C1, Q.DL);
4687 case Intrinsic::powi:
4688 if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) {
4689 // powi(x, 0) -> 1.0
4690 if (Power->isZero())
4691 return ConstantFP::get(LHS->getType(), 1.0);
4702 // Simplify calls to llvm.masked.load.*
4704 case Intrinsic::masked_load: {
4705 Value *MaskArg = ArgBegin[2];
4706 Value *PassthruArg = ArgBegin[3];
4707 // If the mask is all zeros or undef, the "passthru" argument is the result.
4708 if (maskIsAllZeroOrUndef(MaskArg))
4717 template <typename IterTy>
4718 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4719 IterTy ArgEnd, const SimplifyQuery &Q,
4720 unsigned MaxRecurse) {
4721 Type *Ty = V->getType();
4722 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4723 Ty = PTy->getElementType();
4724 FunctionType *FTy = cast<FunctionType>(Ty);
4726 // call undef -> undef
4727 // call null -> undef
4728 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4729 return UndefValue::get(FTy->getReturnType());
4731 Function *F = dyn_cast<Function>(V);
4735 if (F->isIntrinsic())
4736 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4739 if (!canConstantFoldCallTo(CS, F))
4742 SmallVector<Constant *, 4> ConstantArgs;
4743 ConstantArgs.reserve(ArgEnd - ArgBegin);
4744 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4745 Constant *C = dyn_cast<Constant>(*I);
4748 ConstantArgs.push_back(C);
4751 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4754 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4755 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4756 const SimplifyQuery &Q) {
4757 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4760 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4761 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4762 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4765 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4766 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4767 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4771 /// See if we can compute a simplified version of this instruction.
4772 /// If not, this returns null.
4774 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4775 OptimizationRemarkEmitter *ORE) {
4776 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4779 switch (I->getOpcode()) {
4781 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4783 case Instruction::FAdd:
4784 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4785 I->getFastMathFlags(), Q);
4787 case Instruction::Add:
4788 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4789 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4790 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4792 case Instruction::FSub:
4793 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4794 I->getFastMathFlags(), Q);
4796 case Instruction::Sub:
4797 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4798 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4799 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4801 case Instruction::FMul:
4802 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4803 I->getFastMathFlags(), Q);
4805 case Instruction::Mul:
4806 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4808 case Instruction::SDiv:
4809 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4811 case Instruction::UDiv:
4812 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4814 case Instruction::FDiv:
4815 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4816 I->getFastMathFlags(), Q);
4818 case Instruction::SRem:
4819 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4821 case Instruction::URem:
4822 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4824 case Instruction::FRem:
4825 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4826 I->getFastMathFlags(), Q);
4828 case Instruction::Shl:
4829 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4830 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4831 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4833 case Instruction::LShr:
4834 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4835 cast<BinaryOperator>(I)->isExact(), Q);
4837 case Instruction::AShr:
4838 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4839 cast<BinaryOperator>(I)->isExact(), Q);
4841 case Instruction::And:
4842 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4844 case Instruction::Or:
4845 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4847 case Instruction::Xor:
4848 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4850 case Instruction::ICmp:
4851 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4852 I->getOperand(0), I->getOperand(1), Q);
4854 case Instruction::FCmp:
4856 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4857 I->getOperand(1), I->getFastMathFlags(), Q);
4859 case Instruction::Select:
4860 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4861 I->getOperand(2), Q);
4863 case Instruction::GetElementPtr: {
4864 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4865 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4869 case Instruction::InsertValue: {
4870 InsertValueInst *IV = cast<InsertValueInst>(I);
4871 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4872 IV->getInsertedValueOperand(),
4873 IV->getIndices(), Q);
4876 case Instruction::InsertElement: {
4877 auto *IE = cast<InsertElementInst>(I);
4878 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4879 IE->getOperand(2), Q);
4882 case Instruction::ExtractValue: {
4883 auto *EVI = cast<ExtractValueInst>(I);
4884 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4885 EVI->getIndices(), Q);
4888 case Instruction::ExtractElement: {
4889 auto *EEI = cast<ExtractElementInst>(I);
4890 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4891 EEI->getIndexOperand(), Q);
4894 case Instruction::ShuffleVector: {
4895 auto *SVI = cast<ShuffleVectorInst>(I);
4896 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4897 SVI->getMask(), SVI->getType(), Q);
4900 case Instruction::PHI:
4901 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4903 case Instruction::Call: {
4904 CallSite CS(cast<CallInst>(I));
4905 Result = SimplifyCall(CS, Q);
4908 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4909 #include "llvm/IR/Instruction.def"
4910 #undef HANDLE_CAST_INST
4912 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4914 case Instruction::Alloca:
4915 // No simplifications for Alloca and it can't be constant folded.
4920 // In general, it is possible for computeKnownBits to determine all bits in a
4921 // value even when the operands are not all constants.
4922 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4923 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4924 if (Known.isConstant())
4925 Result = ConstantInt::get(I->getType(), Known.getConstant());
4928 /// If called on unreachable code, the above logic may report that the
4929 /// instruction simplified to itself. Make life easier for users by
4930 /// detecting that case here, returning a safe value instead.
4931 return Result == I ? UndefValue::get(I->getType()) : Result;
4934 /// Implementation of recursive simplification through an instruction's
4937 /// This is the common implementation of the recursive simplification routines.
4938 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4939 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4940 /// instructions to process and attempt to simplify it using
4941 /// InstructionSimplify.
4943 /// This routine returns 'true' only when *it* simplifies something. The passed
4944 /// in simplified value does not count toward this.
4945 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4946 const TargetLibraryInfo *TLI,
4947 const DominatorTree *DT,
4948 AssumptionCache *AC) {
4949 bool Simplified = false;
4950 SmallSetVector<Instruction *, 8> Worklist;
4951 const DataLayout &DL = I->getModule()->getDataLayout();
4953 // If we have an explicit value to collapse to, do that round of the
4954 // simplification loop by hand initially.
4956 for (User *U : I->users())
4958 Worklist.insert(cast<Instruction>(U));
4960 // Replace the instruction with its simplified value.
4961 I->replaceAllUsesWith(SimpleV);
4963 // Gracefully handle edge cases where the instruction is not wired into any
4965 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4966 !I->mayHaveSideEffects())
4967 I->eraseFromParent();
4972 // Note that we must test the size on each iteration, the worklist can grow.
4973 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4976 // See if this instruction simplifies.
4977 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4983 // Stash away all the uses of the old instruction so we can check them for
4984 // recursive simplifications after a RAUW. This is cheaper than checking all
4985 // uses of To on the recursive step in most cases.
4986 for (User *U : I->users())
4987 Worklist.insert(cast<Instruction>(U));
4989 // Replace the instruction with its simplified value.
4990 I->replaceAllUsesWith(SimpleV);
4992 // Gracefully handle edge cases where the instruction is not wired into any
4994 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4995 !I->mayHaveSideEffects())
4996 I->eraseFromParent();
5001 bool llvm::recursivelySimplifyInstruction(Instruction *I,
5002 const TargetLibraryInfo *TLI,
5003 const DominatorTree *DT,
5004 AssumptionCache *AC) {
5005 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
5008 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
5009 const TargetLibraryInfo *TLI,
5010 const DominatorTree *DT,
5011 AssumptionCache *AC) {
5012 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5013 assert(SimpleV && "Must provide a simplified value.");
5014 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
5018 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
5019 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
5020 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
5021 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
5022 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
5023 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5024 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5025 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5028 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5029 const DataLayout &DL) {
5030 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5033 template <class T, class... TArgs>
5034 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5036 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5037 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5038 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5039 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5041 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,