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())))
564 // add nuw %x, -1 -> -1, because %x can only be 0.
565 if (IsNUW && match(Op1, m_AllOnes()))
566 return Op1; // Which is -1.
569 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
570 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
573 // Try some generic simplifications for associative operations.
574 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
578 // Threading Add over selects and phi nodes is pointless, so don't bother.
579 // Threading over the select in "A + select(cond, B, C)" means evaluating
580 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
581 // only if B and C are equal. If B and C are equal then (since we assume
582 // that operands have already been simplified) "select(cond, B, C)" should
583 // have been simplified to the common value of B and C already. Analysing
584 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
585 // for threading over phi nodes.
590 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
591 const SimplifyQuery &Query) {
592 return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
595 /// Compute the base pointer and cumulative constant offsets for V.
597 /// This strips all constant offsets off of V, leaving it the base pointer, and
598 /// accumulates the total constant offset applied in the returned constant. It
599 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
600 /// no constant offsets applied.
602 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
603 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
605 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
606 bool AllowNonInbounds = false) {
607 assert(V->getType()->isPtrOrPtrVectorTy());
609 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
610 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
612 // Even though we don't look through PHI nodes, we could be called on an
613 // instruction in an unreachable block, which may be on a cycle.
614 SmallPtrSet<Value *, 4> Visited;
617 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
618 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
619 !GEP->accumulateConstantOffset(DL, Offset))
621 V = GEP->getPointerOperand();
622 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
623 V = cast<Operator>(V)->getOperand(0);
624 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
625 if (GA->isInterposable())
627 V = GA->getAliasee();
629 if (auto CS = CallSite(V))
630 if (Value *RV = CS.getReturnedArgOperand()) {
636 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
637 } while (Visited.insert(V).second);
639 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
640 if (V->getType()->isVectorTy())
641 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
646 /// Compute the constant difference between two pointer values.
647 /// If the difference is not a constant, returns zero.
648 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
650 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
651 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
653 // If LHS and RHS are not related via constant offsets to the same base
654 // value, there is nothing we can do here.
658 // Otherwise, the difference of LHS - RHS can be computed as:
660 // = (LHSOffset + Base) - (RHSOffset + Base)
661 // = LHSOffset - RHSOffset
662 return ConstantExpr::getSub(LHSOffset, RHSOffset);
665 /// Given operands for a Sub, see if we can fold the result.
666 /// If not, this returns null.
667 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
668 const SimplifyQuery &Q, unsigned MaxRecurse) {
669 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
672 // X - undef -> undef
673 // undef - X -> undef
674 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
675 return UndefValue::get(Op0->getType());
678 if (match(Op1, m_Zero()))
683 return Constant::getNullValue(Op0->getType());
685 // Is this a negation?
686 if (match(Op0, m_Zero())) {
687 // 0 - X -> 0 if the sub is NUW.
689 return Constant::getNullValue(Op0->getType());
691 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
692 if (Known.Zero.isMaxSignedValue()) {
693 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
694 // Op1 must be 0 because negating the minimum signed value is undefined.
696 return Constant::getNullValue(Op0->getType());
698 // 0 - X -> X if X is 0 or the minimum signed value.
703 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
704 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
705 Value *X = nullptr, *Y = nullptr, *Z = Op1;
706 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
707 // See if "V === Y - Z" simplifies.
708 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
709 // It does! Now see if "X + V" simplifies.
710 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
711 // It does, we successfully reassociated!
715 // See if "V === X - Z" simplifies.
716 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
717 // It does! Now see if "Y + V" simplifies.
718 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
719 // It does, we successfully reassociated!
725 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
726 // For example, X - (X + 1) -> -1
728 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
729 // See if "V === X - Y" simplifies.
730 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
731 // It does! Now see if "V - Z" simplifies.
732 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
733 // It does, we successfully reassociated!
737 // See if "V === X - Z" simplifies.
738 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
739 // It does! Now see if "V - Y" simplifies.
740 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
741 // It does, we successfully reassociated!
747 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
748 // For example, X - (X - Y) -> Y.
750 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
751 // See if "V === Z - X" simplifies.
752 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
753 // It does! Now see if "V + Y" simplifies.
754 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
755 // It does, we successfully reassociated!
760 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
761 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
762 match(Op1, m_Trunc(m_Value(Y))))
763 if (X->getType() == Y->getType())
764 // See if "V === X - Y" simplifies.
765 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
766 // It does! Now see if "trunc V" simplifies.
767 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
769 // It does, return the simplified "trunc V".
772 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
773 if (match(Op0, m_PtrToInt(m_Value(X))) &&
774 match(Op1, m_PtrToInt(m_Value(Y))))
775 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
776 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
779 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
780 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
783 // Threading Sub over selects and phi nodes is pointless, so don't bother.
784 // Threading over the select in "A - select(cond, B, C)" means evaluating
785 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
786 // only if B and C are equal. If B and C are equal then (since we assume
787 // that operands have already been simplified) "select(cond, B, C)" should
788 // have been simplified to the common value of B and C already. Analysing
789 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
790 // for threading over phi nodes.
795 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
796 const SimplifyQuery &Q) {
797 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
800 /// Given operands for a Mul, see if we can fold the result.
801 /// If not, this returns null.
802 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
803 unsigned MaxRecurse) {
804 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
809 if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
810 return Constant::getNullValue(Op0->getType());
813 if (match(Op1, m_One()))
816 // (X / Y) * Y -> X if the division is exact.
818 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
819 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
823 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
824 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
827 // Try some generic simplifications for associative operations.
828 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
832 // Mul distributes over Add. Try some generic simplifications based on this.
833 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
837 // If the operation is with the result of a select instruction, check whether
838 // operating on either branch of the select always yields the same value.
839 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
840 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
844 // If the operation is with the result of a phi instruction, check whether
845 // operating on all incoming values of the phi always yields the same value.
846 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
847 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
854 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
855 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
858 /// Check for common or similar folds of integer division or integer remainder.
859 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
860 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
861 Type *Ty = Op0->getType();
863 // X / undef -> undef
864 // X % undef -> undef
865 if (match(Op1, m_Undef()))
870 // We don't need to preserve faults!
871 if (match(Op1, m_Zero()))
872 return UndefValue::get(Ty);
874 // If any element of a constant divisor vector is zero or undef, the whole op
876 auto *Op1C = dyn_cast<Constant>(Op1);
877 if (Op1C && Ty->isVectorTy()) {
878 unsigned NumElts = Ty->getVectorNumElements();
879 for (unsigned i = 0; i != NumElts; ++i) {
880 Constant *Elt = Op1C->getAggregateElement(i);
881 if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
882 return UndefValue::get(Ty);
888 if (match(Op0, m_Undef()))
889 return Constant::getNullValue(Ty);
893 if (match(Op0, m_Zero()))
894 return Constant::getNullValue(Op0->getType());
899 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
903 // If this is a boolean op (single-bit element type), we can't have
904 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
905 // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
907 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
908 (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
909 return IsDiv ? Op0 : Constant::getNullValue(Ty);
914 /// Given a predicate and two operands, return true if the comparison is true.
915 /// This is a helper for div/rem simplification where we return some other value
916 /// when we can prove a relationship between the operands.
917 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
918 const SimplifyQuery &Q, unsigned MaxRecurse) {
919 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
920 Constant *C = dyn_cast_or_null<Constant>(V);
921 return (C && C->isAllOnesValue());
924 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
925 /// to simplify X % Y to X.
926 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
927 unsigned MaxRecurse, bool IsSigned) {
928 // Recursion is always used, so bail out at once if we already hit the limit.
935 // We require that 1 operand is a simple constant. That could be extended to
936 // 2 variables if we computed the sign bit for each.
938 // Make sure that a constant is not the minimum signed value because taking
939 // the abs() of that is undefined.
940 Type *Ty = X->getType();
942 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
943 // Is the variable divisor magnitude always greater than the constant
944 // dividend magnitude?
945 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
946 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
947 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
948 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
949 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
952 if (match(Y, m_APInt(C))) {
953 // Special-case: we can't take the abs() of a minimum signed value. If
954 // that's the divisor, then all we have to do is prove that the dividend
955 // is also not the minimum signed value.
956 if (C->isMinSignedValue())
957 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
959 // Is the variable dividend magnitude always less than the constant
960 // divisor magnitude?
961 // |X| < |C| --> X > -abs(C) and X < abs(C)
962 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
963 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
964 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
965 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
971 // IsSigned == false.
972 // Is the dividend unsigned less than the divisor?
973 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
976 /// These are simplifications common to SDiv and UDiv.
977 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
978 const SimplifyQuery &Q, unsigned MaxRecurse) {
979 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
982 if (Value *V = simplifyDivRem(Op0, Op1, true))
985 bool IsSigned = Opcode == Instruction::SDiv;
987 // (X * Y) / Y -> X if the multiplication does not overflow.
989 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
990 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
991 // If the Mul does not overflow, then we are good to go.
992 if ((IsSigned && Mul->hasNoSignedWrap()) ||
993 (!IsSigned && Mul->hasNoUnsignedWrap()))
995 // If X has the form X = A / Y, then X * Y cannot overflow.
996 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
997 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1001 // (X rem Y) / Y -> 0
1002 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1003 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1004 return Constant::getNullValue(Op0->getType());
1006 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1007 ConstantInt *C1, *C2;
1008 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1009 match(Op1, m_ConstantInt(C2))) {
1011 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1013 return Constant::getNullValue(Op0->getType());
1016 // If the operation is with the result of a select instruction, check whether
1017 // operating on either branch of the select always yields the same value.
1018 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1019 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1022 // If the operation is with the result of a phi instruction, check whether
1023 // operating on all incoming values of the phi always yields the same value.
1024 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1025 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1028 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1029 return Constant::getNullValue(Op0->getType());
1034 /// These are simplifications common to SRem and URem.
1035 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1036 const SimplifyQuery &Q, unsigned MaxRecurse) {
1037 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1040 if (Value *V = simplifyDivRem(Op0, Op1, false))
1043 // (X % Y) % Y -> X % Y
1044 if ((Opcode == Instruction::SRem &&
1045 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1046 (Opcode == Instruction::URem &&
1047 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1050 // (X << Y) % X -> 0
1051 if ((Opcode == Instruction::SRem &&
1052 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1053 (Opcode == Instruction::URem &&
1054 match(Op0, m_NUWShl(m_Specific(Op1), m_Value()))))
1055 return Constant::getNullValue(Op0->getType());
1057 // If the operation is with the result of a select instruction, check whether
1058 // operating on either branch of the select always yields the same value.
1059 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1060 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1063 // If the operation is with the result of a phi instruction, check whether
1064 // operating on all incoming values of the phi always yields the same value.
1065 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1066 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1069 // If X / Y == 0, then X % Y == X.
1070 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1076 /// Given operands for an SDiv, see if we can fold the result.
1077 /// If not, this returns null.
1078 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1079 unsigned MaxRecurse) {
1080 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1083 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1084 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1087 /// Given operands for a UDiv, see if we can fold the result.
1088 /// If not, this returns null.
1089 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1090 unsigned MaxRecurse) {
1091 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1094 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1095 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1098 /// Given operands for an SRem, see if we can fold the result.
1099 /// If not, this returns null.
1100 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1101 unsigned MaxRecurse) {
1102 // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1103 // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1105 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1106 return ConstantInt::getNullValue(Op0->getType());
1108 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1111 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1112 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1115 /// Given operands for a URem, see if we can fold the result.
1116 /// If not, this returns null.
1117 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1118 unsigned MaxRecurse) {
1119 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1122 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1123 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1126 /// Returns true if a shift by \c Amount always yields undef.
1127 static bool isUndefShift(Value *Amount) {
1128 Constant *C = dyn_cast<Constant>(Amount);
1132 // X shift by undef -> undef because it may shift by the bitwidth.
1133 if (isa<UndefValue>(C))
1136 // Shifting by the bitwidth or more is undefined.
1137 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1138 if (CI->getValue().getLimitedValue() >=
1139 CI->getType()->getScalarSizeInBits())
1142 // If all lanes of a vector shift are undefined the whole shift is.
1143 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1144 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1145 if (!isUndefShift(C->getAggregateElement(I)))
1153 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1154 /// If not, this returns null.
1155 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1156 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1157 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1160 // 0 shift by X -> 0
1161 if (match(Op0, m_Zero()))
1162 return Constant::getNullValue(Op0->getType());
1164 // X shift by 0 -> X
1165 // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1168 if (match(Op1, m_Zero()) ||
1169 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1172 // Fold undefined shifts.
1173 if (isUndefShift(Op1))
1174 return UndefValue::get(Op0->getType());
1176 // If the operation is with the result of a select instruction, check whether
1177 // operating on either branch of the select always yields the same value.
1178 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1179 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1182 // If the operation is with the result of a phi instruction, check whether
1183 // operating on all incoming values of the phi always yields the same value.
1184 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1185 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1188 // If any bits in the shift amount make that value greater than or equal to
1189 // the number of bits in the type, the shift is undefined.
1190 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1191 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1192 return UndefValue::get(Op0->getType());
1194 // If all valid bits in the shift amount are known zero, the first operand is
1196 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1197 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1203 /// Given operands for an Shl, LShr or AShr, see if we can
1204 /// fold the result. If not, this returns null.
1205 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1206 Value *Op1, bool isExact, const SimplifyQuery &Q,
1207 unsigned MaxRecurse) {
1208 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1213 return Constant::getNullValue(Op0->getType());
1216 // undef >> X -> undef (if it's exact)
1217 if (match(Op0, m_Undef()))
1218 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1220 // The low bit cannot be shifted out of an exact shift if it is set.
1222 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1223 if (Op0Known.One[0])
1230 /// Given operands for an Shl, see if we can fold the result.
1231 /// If not, this returns null.
1232 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1233 const SimplifyQuery &Q, unsigned MaxRecurse) {
1234 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1238 // undef << X -> undef if (if it's NSW/NUW)
1239 if (match(Op0, m_Undef()))
1240 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1242 // (X >> A) << A -> X
1244 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1247 // shl nuw i8 C, %x -> C iff C has sign bit set.
1248 if (isNUW && match(Op0, m_Negative()))
1250 // NOTE: could use computeKnownBits() / LazyValueInfo,
1251 // but the cost-benefit analysis suggests it isn't worth it.
1256 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1257 const SimplifyQuery &Q) {
1258 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1261 /// Given operands for an LShr, see if we can fold the result.
1262 /// If not, this returns null.
1263 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1264 const SimplifyQuery &Q, unsigned MaxRecurse) {
1265 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1269 // (X << A) >> A -> X
1271 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1277 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1278 const SimplifyQuery &Q) {
1279 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1282 /// Given operands for an AShr, see if we can fold the result.
1283 /// If not, this returns null.
1284 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1285 const SimplifyQuery &Q, unsigned MaxRecurse) {
1286 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1290 // all ones >>a X -> -1
1291 // Do not return Op0 because it may contain undef elements if it's a vector.
1292 if (match(Op0, m_AllOnes()))
1293 return Constant::getAllOnesValue(Op0->getType());
1295 // (X << A) >> A -> X
1297 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1300 // Arithmetic shifting an all-sign-bit value is a no-op.
1301 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1302 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1308 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1309 const SimplifyQuery &Q) {
1310 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1313 /// Commuted variants are assumed to be handled by calling this function again
1314 /// with the parameters swapped.
1315 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1316 ICmpInst *UnsignedICmp, bool IsAnd) {
1319 ICmpInst::Predicate EqPred;
1320 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1321 !ICmpInst::isEquality(EqPred))
1324 ICmpInst::Predicate UnsignedPred;
1325 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1326 ICmpInst::isUnsigned(UnsignedPred))
1328 else if (match(UnsignedICmp,
1329 m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1330 ICmpInst::isUnsigned(UnsignedPred))
1331 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1335 // X < Y && Y != 0 --> X < Y
1336 // X < Y || Y != 0 --> Y != 0
1337 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1338 return IsAnd ? UnsignedICmp : ZeroICmp;
1340 // X >= Y || Y != 0 --> true
1341 // X >= Y || Y == 0 --> X >= Y
1342 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1343 if (EqPred == ICmpInst::ICMP_NE)
1344 return getTrue(UnsignedICmp->getType());
1345 return UnsignedICmp;
1348 // X < Y && Y == 0 --> false
1349 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1351 return getFalse(UnsignedICmp->getType());
1356 /// Commuted variants are assumed to be handled by calling this function again
1357 /// with the parameters swapped.
1358 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1359 ICmpInst::Predicate Pred0, Pred1;
1361 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1362 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1365 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1366 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1367 // can eliminate Op1 from this 'and'.
1368 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1371 // Check for any combination of predicates that are guaranteed to be disjoint.
1372 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1373 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1374 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1375 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1376 return getFalse(Op0->getType());
1381 /// Commuted variants are assumed to be handled by calling this function again
1382 /// with the parameters swapped.
1383 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1384 ICmpInst::Predicate Pred0, Pred1;
1386 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1387 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1390 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1391 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1392 // can eliminate Op0 from this 'or'.
1393 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1396 // Check for any combination of predicates that cover the entire range of
1398 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1399 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1400 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1401 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1402 return getTrue(Op0->getType());
1407 /// Test if a pair of compares with a shared operand and 2 constants has an
1408 /// empty set intersection, full set union, or if one compare is a superset of
1410 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1412 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1413 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1416 const APInt *C0, *C1;
1417 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1418 !match(Cmp1->getOperand(1), m_APInt(C1)))
1421 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1422 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1424 // For and-of-compares, check if the intersection is empty:
1425 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1426 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1427 return getFalse(Cmp0->getType());
1429 // For or-of-compares, check if the union is full:
1430 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1431 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1432 return getTrue(Cmp0->getType());
1434 // Is one range a superset of the other?
1435 // If this is and-of-compares, take the smaller set:
1436 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1437 // If this is or-of-compares, take the larger set:
1438 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1439 if (Range0.contains(Range1))
1440 return IsAnd ? Cmp1 : Cmp0;
1441 if (Range1.contains(Range0))
1442 return IsAnd ? Cmp0 : Cmp1;
1447 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1449 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1450 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1451 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1454 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1457 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1458 Value *X = Cmp0->getOperand(0);
1459 Value *Y = Cmp1->getOperand(0);
1461 // If one of the compares is a masked version of a (not) null check, then
1462 // that compare implies the other, so we eliminate the other. Optionally, look
1463 // through a pointer-to-int cast to match a null check of a pointer type.
1465 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1466 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1467 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1468 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1469 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1470 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1473 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1474 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1475 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1476 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1477 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1478 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1484 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1485 // (icmp (add V, C0), C1) & (icmp V, C0)
1486 ICmpInst::Predicate Pred0, Pred1;
1487 const APInt *C0, *C1;
1489 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1492 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1495 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1496 if (AddInst->getOperand(1) != Op1->getOperand(1))
1499 Type *ITy = Op0->getType();
1500 bool isNSW = AddInst->hasNoSignedWrap();
1501 bool isNUW = AddInst->hasNoUnsignedWrap();
1503 const APInt Delta = *C1 - *C0;
1504 if (C0->isStrictlyPositive()) {
1506 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1507 return getFalse(ITy);
1508 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1509 return getFalse(ITy);
1512 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1513 return getFalse(ITy);
1514 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1515 return getFalse(ITy);
1518 if (C0->getBoolValue() && isNUW) {
1520 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1521 return getFalse(ITy);
1523 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1524 return getFalse(ITy);
1530 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1531 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1533 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1536 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1538 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1541 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1544 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1547 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1549 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1555 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1556 // (icmp (add V, C0), C1) | (icmp V, C0)
1557 ICmpInst::Predicate Pred0, Pred1;
1558 const APInt *C0, *C1;
1560 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1563 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1566 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1567 if (AddInst->getOperand(1) != Op1->getOperand(1))
1570 Type *ITy = Op0->getType();
1571 bool isNSW = AddInst->hasNoSignedWrap();
1572 bool isNUW = AddInst->hasNoUnsignedWrap();
1574 const APInt Delta = *C1 - *C0;
1575 if (C0->isStrictlyPositive()) {
1577 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1578 return getTrue(ITy);
1579 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1580 return getTrue(ITy);
1583 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1584 return getTrue(ITy);
1585 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1586 return getTrue(ITy);
1589 if (C0->getBoolValue() && isNUW) {
1591 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1592 return getTrue(ITy);
1594 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1595 return getTrue(ITy);
1601 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1602 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1604 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1607 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1609 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1612 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1615 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1618 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1620 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1626 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1627 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1628 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1629 if (LHS0->getType() != RHS0->getType())
1632 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1633 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1634 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1635 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1636 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1637 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1638 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1639 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1640 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1641 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1642 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1643 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1644 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1647 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1648 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1649 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1650 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1651 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1652 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1653 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1654 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1655 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1656 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1663 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1664 // Look through casts of the 'and' operands to find compares.
1665 auto *Cast0 = dyn_cast<CastInst>(Op0);
1666 auto *Cast1 = dyn_cast<CastInst>(Op1);
1667 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1668 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1669 Op0 = Cast0->getOperand(0);
1670 Op1 = Cast1->getOperand(0);
1674 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1675 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1677 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1678 simplifyOrOfICmps(ICmp0, ICmp1);
1680 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1681 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1683 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1690 // If we looked through casts, we can only handle a constant simplification
1691 // because we are not allowed to create a cast instruction here.
1692 if (auto *C = dyn_cast<Constant>(V))
1693 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1698 /// Given operands for an And, see if we can fold the result.
1699 /// If not, this returns null.
1700 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1701 unsigned MaxRecurse) {
1702 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1706 if (match(Op1, m_Undef()))
1707 return Constant::getNullValue(Op0->getType());
1714 if (match(Op1, m_Zero()))
1715 return Constant::getNullValue(Op0->getType());
1718 if (match(Op1, m_AllOnes()))
1721 // A & ~A = ~A & A = 0
1722 if (match(Op0, m_Not(m_Specific(Op1))) ||
1723 match(Op1, m_Not(m_Specific(Op0))))
1724 return Constant::getNullValue(Op0->getType());
1727 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1731 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1734 // A mask that only clears known zeros of a shifted value is a no-op.
1738 if (match(Op1, m_APInt(Mask))) {
1739 // If all bits in the inverted and shifted mask are clear:
1740 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1741 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1742 (~(*Mask)).lshr(*ShAmt).isNullValue())
1745 // If all bits in the inverted and shifted mask are clear:
1746 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1747 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1748 (~(*Mask)).shl(*ShAmt).isNullValue())
1752 // A & (-A) = A if A is a power of two or zero.
1753 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1754 match(Op1, m_Neg(m_Specific(Op0)))) {
1755 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1758 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1763 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1766 // Try some generic simplifications for associative operations.
1767 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1771 // And distributes over Or. Try some generic simplifications based on this.
1772 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1776 // And distributes over Xor. Try some generic simplifications based on this.
1777 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1781 // If the operation is with the result of a select instruction, check whether
1782 // operating on either branch of the select always yields the same value.
1783 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1784 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1788 // If the operation is with the result of a phi instruction, check whether
1789 // operating on all incoming values of the phi always yields the same value.
1790 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1791 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1798 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1799 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1802 /// Given operands for an Or, see if we can fold the result.
1803 /// If not, this returns null.
1804 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1805 unsigned MaxRecurse) {
1806 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1811 // Do not return Op1 because it may contain undef elements if it's a vector.
1812 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1813 return Constant::getAllOnesValue(Op0->getType());
1817 if (Op0 == Op1 || match(Op1, m_Zero()))
1820 // A | ~A = ~A | A = -1
1821 if (match(Op0, m_Not(m_Specific(Op1))) ||
1822 match(Op1, m_Not(m_Specific(Op0))))
1823 return Constant::getAllOnesValue(Op0->getType());
1826 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1830 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1833 // ~(A & ?) | A = -1
1834 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1835 return Constant::getAllOnesValue(Op1->getType());
1837 // A | ~(A & ?) = -1
1838 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1839 return Constant::getAllOnesValue(Op0->getType());
1842 // (A & ~B) | (A ^ B) -> (A ^ B)
1843 // (~B & A) | (A ^ B) -> (A ^ B)
1844 // (A & ~B) | (B ^ A) -> (B ^ A)
1845 // (~B & A) | (B ^ A) -> (B ^ A)
1846 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1847 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1848 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1851 // Commute the 'or' operands.
1852 // (A ^ B) | (A & ~B) -> (A ^ B)
1853 // (A ^ B) | (~B & A) -> (A ^ B)
1854 // (B ^ A) | (A & ~B) -> (B ^ A)
1855 // (B ^ A) | (~B & A) -> (B ^ A)
1856 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1857 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1858 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1861 // (A & B) | (~A ^ B) -> (~A ^ B)
1862 // (B & A) | (~A ^ B) -> (~A ^ B)
1863 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1864 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1865 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1866 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1867 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1870 // (~A ^ B) | (A & B) -> (~A ^ B)
1871 // (~A ^ B) | (B & A) -> (~A ^ B)
1872 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1873 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1874 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1875 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1876 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1879 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1882 // Try some generic simplifications for associative operations.
1883 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1887 // Or distributes over And. Try some generic simplifications based on this.
1888 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1892 // If the operation is with the result of a select instruction, check whether
1893 // operating on either branch of the select always yields the same value.
1894 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1895 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1899 // (A & C1)|(B & C2)
1900 const APInt *C1, *C2;
1901 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1902 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1904 // (A & C1)|(B & C2)
1905 // If we have: ((V + N) & C1) | (V & C2)
1906 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1907 // replace with V+N.
1909 if (C2->isMask() && // C2 == 0+1+
1910 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1911 // Add commutes, try both ways.
1912 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1915 // Or commutes, try both ways.
1917 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1918 // Add commutes, try both ways.
1919 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1925 // If the operation is with the result of a phi instruction, check whether
1926 // operating on all incoming values of the phi always yields the same value.
1927 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1928 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1934 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1935 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1938 /// Given operands for a Xor, see if we can fold the result.
1939 /// If not, this returns null.
1940 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1941 unsigned MaxRecurse) {
1942 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1945 // A ^ undef -> undef
1946 if (match(Op1, m_Undef()))
1950 if (match(Op1, m_Zero()))
1955 return Constant::getNullValue(Op0->getType());
1957 // A ^ ~A = ~A ^ A = -1
1958 if (match(Op0, m_Not(m_Specific(Op1))) ||
1959 match(Op1, m_Not(m_Specific(Op0))))
1960 return Constant::getAllOnesValue(Op0->getType());
1962 // Try some generic simplifications for associative operations.
1963 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1967 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1968 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1969 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1970 // only if B and C are equal. If B and C are equal then (since we assume
1971 // that operands have already been simplified) "select(cond, B, C)" should
1972 // have been simplified to the common value of B and C already. Analysing
1973 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1974 // for threading over phi nodes.
1979 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1980 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1984 static Type *GetCompareTy(Value *Op) {
1985 return CmpInst::makeCmpResultType(Op->getType());
1988 /// Rummage around inside V looking for something equivalent to the comparison
1989 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1990 /// Helper function for analyzing max/min idioms.
1991 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1992 Value *LHS, Value *RHS) {
1993 SelectInst *SI = dyn_cast<SelectInst>(V);
1996 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1999 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2000 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2002 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2003 LHS == CmpRHS && RHS == CmpLHS)
2008 // A significant optimization not implemented here is assuming that alloca
2009 // addresses are not equal to incoming argument values. They don't *alias*,
2010 // as we say, but that doesn't mean they aren't equal, so we take a
2011 // conservative approach.
2013 // This is inspired in part by C++11 5.10p1:
2014 // "Two pointers of the same type compare equal if and only if they are both
2015 // null, both point to the same function, or both represent the same
2018 // This is pretty permissive.
2020 // It's also partly due to C11 6.5.9p6:
2021 // "Two pointers compare equal if and only if both are null pointers, both are
2022 // pointers to the same object (including a pointer to an object and a
2023 // subobject at its beginning) or function, both are pointers to one past the
2024 // last element of the same array object, or one is a pointer to one past the
2025 // end of one array object and the other is a pointer to the start of a
2026 // different array object that happens to immediately follow the first array
2027 // object in the address space.)
2029 // C11's version is more restrictive, however there's no reason why an argument
2030 // couldn't be a one-past-the-end value for a stack object in the caller and be
2031 // equal to the beginning of a stack object in the callee.
2033 // If the C and C++ standards are ever made sufficiently restrictive in this
2034 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2035 // this optimization.
2037 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2038 const DominatorTree *DT, CmpInst::Predicate Pred,
2039 AssumptionCache *AC, const Instruction *CxtI,
2040 Value *LHS, Value *RHS) {
2041 // First, skip past any trivial no-ops.
2042 LHS = LHS->stripPointerCasts();
2043 RHS = RHS->stripPointerCasts();
2045 // A non-null pointer is not equal to a null pointer.
2046 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2047 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2048 return ConstantInt::get(GetCompareTy(LHS),
2049 !CmpInst::isTrueWhenEqual(Pred));
2051 // We can only fold certain predicates on pointer comparisons.
2056 // Equality comaprisons are easy to fold.
2057 case CmpInst::ICMP_EQ:
2058 case CmpInst::ICMP_NE:
2061 // We can only handle unsigned relational comparisons because 'inbounds' on
2062 // a GEP only protects against unsigned wrapping.
2063 case CmpInst::ICMP_UGT:
2064 case CmpInst::ICMP_UGE:
2065 case CmpInst::ICMP_ULT:
2066 case CmpInst::ICMP_ULE:
2067 // However, we have to switch them to their signed variants to handle
2068 // negative indices from the base pointer.
2069 Pred = ICmpInst::getSignedPredicate(Pred);
2073 // Strip off any constant offsets so that we can reason about them.
2074 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2075 // here and compare base addresses like AliasAnalysis does, however there are
2076 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2077 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2078 // doesn't need to guarantee pointer inequality when it says NoAlias.
2079 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2080 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2082 // If LHS and RHS are related via constant offsets to the same base
2083 // value, we can replace it with an icmp which just compares the offsets.
2085 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2087 // Various optimizations for (in)equality comparisons.
2088 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2089 // Different non-empty allocations that exist at the same time have
2090 // different addresses (if the program can tell). Global variables always
2091 // exist, so they always exist during the lifetime of each other and all
2092 // allocas. Two different allocas usually have different addresses...
2094 // However, if there's an @llvm.stackrestore dynamically in between two
2095 // allocas, they may have the same address. It's tempting to reduce the
2096 // scope of the problem by only looking at *static* allocas here. That would
2097 // cover the majority of allocas while significantly reducing the likelihood
2098 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2099 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2100 // an entry block. Also, if we have a block that's not attached to a
2101 // function, we can't tell if it's "static" under the current definition.
2102 // Theoretically, this problem could be fixed by creating a new kind of
2103 // instruction kind specifically for static allocas. Such a new instruction
2104 // could be required to be at the top of the entry block, thus preventing it
2105 // from being subject to a @llvm.stackrestore. Instcombine could even
2106 // convert regular allocas into these special allocas. It'd be nifty.
2107 // However, until then, this problem remains open.
2109 // So, we'll assume that two non-empty allocas have different addresses
2112 // With all that, if the offsets are within the bounds of their allocations
2113 // (and not one-past-the-end! so we can't use inbounds!), and their
2114 // allocations aren't the same, the pointers are not equal.
2116 // Note that it's not necessary to check for LHS being a global variable
2117 // address, due to canonicalization and constant folding.
2118 if (isa<AllocaInst>(LHS) &&
2119 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2120 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2121 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2122 uint64_t LHSSize, RHSSize;
2123 if (LHSOffsetCI && RHSOffsetCI &&
2124 getObjectSize(LHS, LHSSize, DL, TLI) &&
2125 getObjectSize(RHS, RHSSize, DL, TLI)) {
2126 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2127 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2128 if (!LHSOffsetValue.isNegative() &&
2129 !RHSOffsetValue.isNegative() &&
2130 LHSOffsetValue.ult(LHSSize) &&
2131 RHSOffsetValue.ult(RHSSize)) {
2132 return ConstantInt::get(GetCompareTy(LHS),
2133 !CmpInst::isTrueWhenEqual(Pred));
2137 // Repeat the above check but this time without depending on DataLayout
2138 // or being able to compute a precise size.
2139 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2140 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2141 LHSOffset->isNullValue() &&
2142 RHSOffset->isNullValue())
2143 return ConstantInt::get(GetCompareTy(LHS),
2144 !CmpInst::isTrueWhenEqual(Pred));
2147 // Even if an non-inbounds GEP occurs along the path we can still optimize
2148 // equality comparisons concerning the result. We avoid walking the whole
2149 // chain again by starting where the last calls to
2150 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2151 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2152 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2154 return ConstantExpr::getICmp(Pred,
2155 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2156 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2158 // If one side of the equality comparison must come from a noalias call
2159 // (meaning a system memory allocation function), and the other side must
2160 // come from a pointer that cannot overlap with dynamically-allocated
2161 // memory within the lifetime of the current function (allocas, byval
2162 // arguments, globals), then determine the comparison result here.
2163 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2164 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2165 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2167 // Is the set of underlying objects all noalias calls?
2168 auto IsNAC = [](ArrayRef<Value *> Objects) {
2169 return all_of(Objects, isNoAliasCall);
2172 // Is the set of underlying objects all things which must be disjoint from
2173 // noalias calls. For allocas, we consider only static ones (dynamic
2174 // allocas might be transformed into calls to malloc not simultaneously
2175 // live with the compared-to allocation). For globals, we exclude symbols
2176 // that might be resolve lazily to symbols in another dynamically-loaded
2177 // library (and, thus, could be malloc'ed by the implementation).
2178 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2179 return all_of(Objects, [](Value *V) {
2180 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2181 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2182 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2183 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2184 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2185 !GV->isThreadLocal();
2186 if (const Argument *A = dyn_cast<Argument>(V))
2187 return A->hasByValAttr();
2192 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2193 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2194 return ConstantInt::get(GetCompareTy(LHS),
2195 !CmpInst::isTrueWhenEqual(Pred));
2197 // Fold comparisons for non-escaping pointer even if the allocation call
2198 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2199 // dynamic allocation call could be either of the operands.
2200 Value *MI = nullptr;
2201 if (isAllocLikeFn(LHS, TLI) &&
2202 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2204 else if (isAllocLikeFn(RHS, TLI) &&
2205 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2207 // FIXME: We should also fold the compare when the pointer escapes, but the
2208 // compare dominates the pointer escape
2209 if (MI && !PointerMayBeCaptured(MI, true, true))
2210 return ConstantInt::get(GetCompareTy(LHS),
2211 CmpInst::isFalseWhenEqual(Pred));
2218 /// Fold an icmp when its operands have i1 scalar type.
2219 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2220 Value *RHS, const SimplifyQuery &Q) {
2221 Type *ITy = GetCompareTy(LHS); // The return type.
2222 Type *OpTy = LHS->getType(); // The operand type.
2223 if (!OpTy->isIntOrIntVectorTy(1))
2226 // A boolean compared to true/false can be simplified in 14 out of the 20
2227 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2228 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2229 if (match(RHS, m_Zero())) {
2231 case CmpInst::ICMP_NE: // X != 0 -> X
2232 case CmpInst::ICMP_UGT: // X >u 0 -> X
2233 case CmpInst::ICMP_SLT: // X <s 0 -> X
2236 case CmpInst::ICMP_ULT: // X <u 0 -> false
2237 case CmpInst::ICMP_SGT: // X >s 0 -> false
2238 return getFalse(ITy);
2240 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2241 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2242 return getTrue(ITy);
2246 } else if (match(RHS, m_One())) {
2248 case CmpInst::ICMP_EQ: // X == 1 -> X
2249 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2250 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2253 case CmpInst::ICMP_UGT: // X >u 1 -> false
2254 case CmpInst::ICMP_SLT: // X <s -1 -> false
2255 return getFalse(ITy);
2257 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2258 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2259 return getTrue(ITy);
2268 case ICmpInst::ICMP_UGE:
2269 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2270 return getTrue(ITy);
2272 case ICmpInst::ICMP_SGE:
2273 /// For signed comparison, the values for an i1 are 0 and -1
2274 /// respectively. This maps into a truth table of:
2275 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2276 /// 0 | 0 | 1 (0 >= 0) | 1
2277 /// 0 | 1 | 1 (0 >= -1) | 1
2278 /// 1 | 0 | 0 (-1 >= 0) | 0
2279 /// 1 | 1 | 1 (-1 >= -1) | 1
2280 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2281 return getTrue(ITy);
2283 case ICmpInst::ICMP_ULE:
2284 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2285 return getTrue(ITy);
2292 /// Try hard to fold icmp with zero RHS because this is a common case.
2293 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2294 Value *RHS, const SimplifyQuery &Q) {
2295 if (!match(RHS, m_Zero()))
2298 Type *ITy = GetCompareTy(LHS); // The return type.
2301 llvm_unreachable("Unknown ICmp predicate!");
2302 case ICmpInst::ICMP_ULT:
2303 return getFalse(ITy);
2304 case ICmpInst::ICMP_UGE:
2305 return getTrue(ITy);
2306 case ICmpInst::ICMP_EQ:
2307 case ICmpInst::ICMP_ULE:
2308 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2309 return getFalse(ITy);
2311 case ICmpInst::ICMP_NE:
2312 case ICmpInst::ICMP_UGT:
2313 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2314 return getTrue(ITy);
2316 case ICmpInst::ICMP_SLT: {
2317 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2318 if (LHSKnown.isNegative())
2319 return getTrue(ITy);
2320 if (LHSKnown.isNonNegative())
2321 return getFalse(ITy);
2324 case ICmpInst::ICMP_SLE: {
2325 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2326 if (LHSKnown.isNegative())
2327 return getTrue(ITy);
2328 if (LHSKnown.isNonNegative() &&
2329 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2330 return getFalse(ITy);
2333 case ICmpInst::ICMP_SGE: {
2334 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2335 if (LHSKnown.isNegative())
2336 return getFalse(ITy);
2337 if (LHSKnown.isNonNegative())
2338 return getTrue(ITy);
2341 case ICmpInst::ICMP_SGT: {
2342 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2343 if (LHSKnown.isNegative())
2344 return getFalse(ITy);
2345 if (LHSKnown.isNonNegative() &&
2346 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2347 return getTrue(ITy);
2355 /// Many binary operators with a constant operand have an easy-to-compute
2356 /// range of outputs. This can be used to fold a comparison to always true or
2358 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2359 unsigned Width = Lower.getBitWidth();
2361 switch (BO.getOpcode()) {
2362 case Instruction::Add:
2363 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2364 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2365 if (BO.hasNoUnsignedWrap()) {
2366 // 'add nuw x, C' produces [C, UINT_MAX].
2368 } else if (BO.hasNoSignedWrap()) {
2369 if (C->isNegative()) {
2370 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2371 Lower = APInt::getSignedMinValue(Width);
2372 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2374 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2375 Lower = APInt::getSignedMinValue(Width) + *C;
2376 Upper = APInt::getSignedMaxValue(Width) + 1;
2382 case Instruction::And:
2383 if (match(BO.getOperand(1), m_APInt(C)))
2384 // 'and x, C' produces [0, C].
2388 case Instruction::Or:
2389 if (match(BO.getOperand(1), m_APInt(C)))
2390 // 'or x, C' produces [C, UINT_MAX].
2394 case Instruction::AShr:
2395 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2396 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2397 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2398 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2399 } else if (match(BO.getOperand(0), m_APInt(C))) {
2400 unsigned ShiftAmount = Width - 1;
2401 if (!C->isNullValue() && BO.isExact())
2402 ShiftAmount = C->countTrailingZeros();
2403 if (C->isNegative()) {
2404 // 'ashr C, x' produces [C, C >> (Width-1)]
2406 Upper = C->ashr(ShiftAmount) + 1;
2408 // 'ashr C, x' produces [C >> (Width-1), C]
2409 Lower = C->ashr(ShiftAmount);
2415 case Instruction::LShr:
2416 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2417 // 'lshr x, C' produces [0, UINT_MAX >> C].
2418 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2419 } else if (match(BO.getOperand(0), m_APInt(C))) {
2420 // 'lshr C, x' produces [C >> (Width-1), C].
2421 unsigned ShiftAmount = Width - 1;
2422 if (!C->isNullValue() && BO.isExact())
2423 ShiftAmount = C->countTrailingZeros();
2424 Lower = C->lshr(ShiftAmount);
2429 case Instruction::Shl:
2430 if (match(BO.getOperand(0), m_APInt(C))) {
2431 if (BO.hasNoUnsignedWrap()) {
2432 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2434 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2435 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2436 if (C->isNegative()) {
2437 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2438 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2439 Lower = C->shl(ShiftAmount);
2442 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2443 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2445 Upper = C->shl(ShiftAmount) + 1;
2451 case Instruction::SDiv:
2452 if (match(BO.getOperand(1), m_APInt(C))) {
2453 APInt IntMin = APInt::getSignedMinValue(Width);
2454 APInt IntMax = APInt::getSignedMaxValue(Width);
2455 if (C->isAllOnesValue()) {
2456 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2457 // where C != -1 and C != 0 and C != 1
2460 } else if (C->countLeadingZeros() < Width - 1) {
2461 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2462 // where C != -1 and C != 0 and C != 1
2463 Lower = IntMin.sdiv(*C);
2464 Upper = IntMax.sdiv(*C);
2465 if (Lower.sgt(Upper))
2466 std::swap(Lower, Upper);
2468 assert(Upper != Lower && "Upper part of range has wrapped!");
2470 } else if (match(BO.getOperand(0), m_APInt(C))) {
2471 if (C->isMinSignedValue()) {
2472 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2474 Upper = Lower.lshr(1) + 1;
2476 // 'sdiv C, x' produces [-|C|, |C|].
2477 Upper = C->abs() + 1;
2478 Lower = (-Upper) + 1;
2483 case Instruction::UDiv:
2484 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2485 // 'udiv x, C' produces [0, UINT_MAX / C].
2486 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2487 } else if (match(BO.getOperand(0), m_APInt(C))) {
2488 // 'udiv C, x' produces [0, C].
2493 case Instruction::SRem:
2494 if (match(BO.getOperand(1), m_APInt(C))) {
2495 // 'srem x, C' produces (-|C|, |C|).
2497 Lower = (-Upper) + 1;
2501 case Instruction::URem:
2502 if (match(BO.getOperand(1), m_APInt(C)))
2503 // 'urem x, C' produces [0, C).
2512 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2514 Type *ITy = GetCompareTy(RHS); // The return type.
2517 // Sign-bit checks can be optimized to true/false after unsigned
2518 // floating-point casts:
2519 // icmp slt (bitcast (uitofp X)), 0 --> false
2520 // icmp sgt (bitcast (uitofp X)), -1 --> true
2521 if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2522 if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2523 return ConstantInt::getFalse(ITy);
2524 if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2525 return ConstantInt::getTrue(ITy);
2529 if (!match(RHS, m_APInt(C)))
2532 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2533 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2534 if (RHS_CR.isEmptySet())
2535 return ConstantInt::getFalse(ITy);
2536 if (RHS_CR.isFullSet())
2537 return ConstantInt::getTrue(ITy);
2539 // Find the range of possible values for binary operators.
2540 unsigned Width = C->getBitWidth();
2541 APInt Lower = APInt(Width, 0);
2542 APInt Upper = APInt(Width, 0);
2543 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2544 setLimitsForBinOp(*BO, Lower, Upper);
2546 ConstantRange LHS_CR =
2547 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2549 if (auto *I = dyn_cast<Instruction>(LHS))
2550 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2551 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2553 if (!LHS_CR.isFullSet()) {
2554 if (RHS_CR.contains(LHS_CR))
2555 return ConstantInt::getTrue(ITy);
2556 if (RHS_CR.inverse().contains(LHS_CR))
2557 return ConstantInt::getFalse(ITy);
2563 /// TODO: A large part of this logic is duplicated in InstCombine's
2564 /// foldICmpBinOp(). We should be able to share that and avoid the code
2566 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2567 Value *RHS, const SimplifyQuery &Q,
2568 unsigned MaxRecurse) {
2569 Type *ITy = GetCompareTy(LHS); // The return type.
2571 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2572 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2573 if (MaxRecurse && (LBO || RBO)) {
2574 // Analyze the case when either LHS or RHS is an add instruction.
2575 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2576 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2577 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2578 if (LBO && LBO->getOpcode() == Instruction::Add) {
2579 A = LBO->getOperand(0);
2580 B = LBO->getOperand(1);
2582 ICmpInst::isEquality(Pred) ||
2583 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2584 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2586 if (RBO && RBO->getOpcode() == Instruction::Add) {
2587 C = RBO->getOperand(0);
2588 D = RBO->getOperand(1);
2590 ICmpInst::isEquality(Pred) ||
2591 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2592 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2595 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2596 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2597 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2598 Constant::getNullValue(RHS->getType()), Q,
2602 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2603 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2605 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2606 C == LHS ? D : C, Q, MaxRecurse - 1))
2609 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2610 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2612 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2615 // C + B == C + D -> B == D
2618 } else if (A == D) {
2619 // D + B == C + D -> B == C
2622 } else if (B == C) {
2623 // A + C == C + D -> A == D
2628 // A + D == C + D -> A == C
2632 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2639 // icmp pred (or X, Y), X
2640 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2641 if (Pred == ICmpInst::ICMP_ULT)
2642 return getFalse(ITy);
2643 if (Pred == ICmpInst::ICMP_UGE)
2644 return getTrue(ITy);
2646 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2647 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2648 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2649 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2650 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2651 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2652 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2655 // icmp pred X, (or X, Y)
2656 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2657 if (Pred == ICmpInst::ICMP_ULE)
2658 return getTrue(ITy);
2659 if (Pred == ICmpInst::ICMP_UGT)
2660 return getFalse(ITy);
2662 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2663 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2664 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2665 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2666 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2667 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2668 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2673 // icmp pred (and X, Y), X
2674 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2675 if (Pred == ICmpInst::ICMP_UGT)
2676 return getFalse(ITy);
2677 if (Pred == ICmpInst::ICMP_ULE)
2678 return getTrue(ITy);
2680 // icmp pred X, (and X, Y)
2681 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2682 if (Pred == ICmpInst::ICMP_UGE)
2683 return getTrue(ITy);
2684 if (Pred == ICmpInst::ICMP_ULT)
2685 return getFalse(ITy);
2688 // 0 - (zext X) pred C
2689 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2690 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2691 if (RHSC->getValue().isStrictlyPositive()) {
2692 if (Pred == ICmpInst::ICMP_SLT)
2693 return ConstantInt::getTrue(RHSC->getContext());
2694 if (Pred == ICmpInst::ICMP_SGE)
2695 return ConstantInt::getFalse(RHSC->getContext());
2696 if (Pred == ICmpInst::ICMP_EQ)
2697 return ConstantInt::getFalse(RHSC->getContext());
2698 if (Pred == ICmpInst::ICMP_NE)
2699 return ConstantInt::getTrue(RHSC->getContext());
2701 if (RHSC->getValue().isNonNegative()) {
2702 if (Pred == ICmpInst::ICMP_SLE)
2703 return ConstantInt::getTrue(RHSC->getContext());
2704 if (Pred == ICmpInst::ICMP_SGT)
2705 return ConstantInt::getFalse(RHSC->getContext());
2710 // icmp pred (urem X, Y), Y
2711 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2715 case ICmpInst::ICMP_SGT:
2716 case ICmpInst::ICMP_SGE: {
2717 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2718 if (!Known.isNonNegative())
2722 case ICmpInst::ICMP_EQ:
2723 case ICmpInst::ICMP_UGT:
2724 case ICmpInst::ICMP_UGE:
2725 return getFalse(ITy);
2726 case ICmpInst::ICMP_SLT:
2727 case ICmpInst::ICMP_SLE: {
2728 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2729 if (!Known.isNonNegative())
2733 case ICmpInst::ICMP_NE:
2734 case ICmpInst::ICMP_ULT:
2735 case ICmpInst::ICMP_ULE:
2736 return getTrue(ITy);
2740 // icmp pred X, (urem Y, X)
2741 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2745 case ICmpInst::ICMP_SGT:
2746 case ICmpInst::ICMP_SGE: {
2747 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2748 if (!Known.isNonNegative())
2752 case ICmpInst::ICMP_NE:
2753 case ICmpInst::ICMP_UGT:
2754 case ICmpInst::ICMP_UGE:
2755 return getTrue(ITy);
2756 case ICmpInst::ICMP_SLT:
2757 case ICmpInst::ICMP_SLE: {
2758 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2759 if (!Known.isNonNegative())
2763 case ICmpInst::ICMP_EQ:
2764 case ICmpInst::ICMP_ULT:
2765 case ICmpInst::ICMP_ULE:
2766 return getFalse(ITy);
2772 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2773 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2774 // icmp pred (X op Y), X
2775 if (Pred == ICmpInst::ICMP_UGT)
2776 return getFalse(ITy);
2777 if (Pred == ICmpInst::ICMP_ULE)
2778 return getTrue(ITy);
2783 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2784 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2785 // icmp pred X, (X op Y)
2786 if (Pred == ICmpInst::ICMP_ULT)
2787 return getFalse(ITy);
2788 if (Pred == ICmpInst::ICMP_UGE)
2789 return getTrue(ITy);
2796 // where CI2 is a power of 2 and CI isn't
2797 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2798 const APInt *CI2Val, *CIVal = &CI->getValue();
2799 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2800 CI2Val->isPowerOf2()) {
2801 if (!CIVal->isPowerOf2()) {
2802 // CI2 << X can equal zero in some circumstances,
2803 // this simplification is unsafe if CI is zero.
2805 // We know it is safe if:
2806 // - The shift is nsw, we can't shift out the one bit.
2807 // - The shift is nuw, we can't shift out the one bit.
2810 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2811 CI2Val->isOneValue() || !CI->isZero()) {
2812 if (Pred == ICmpInst::ICMP_EQ)
2813 return ConstantInt::getFalse(RHS->getContext());
2814 if (Pred == ICmpInst::ICMP_NE)
2815 return ConstantInt::getTrue(RHS->getContext());
2818 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2819 if (Pred == ICmpInst::ICMP_UGT)
2820 return ConstantInt::getFalse(RHS->getContext());
2821 if (Pred == ICmpInst::ICMP_ULE)
2822 return ConstantInt::getTrue(RHS->getContext());
2827 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2828 LBO->getOperand(1) == RBO->getOperand(1)) {
2829 switch (LBO->getOpcode()) {
2832 case Instruction::UDiv:
2833 case Instruction::LShr:
2834 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2836 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2837 RBO->getOperand(0), Q, MaxRecurse - 1))
2840 case Instruction::SDiv:
2841 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2843 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2844 RBO->getOperand(0), Q, MaxRecurse - 1))
2847 case Instruction::AShr:
2848 if (!LBO->isExact() || !RBO->isExact())
2850 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2851 RBO->getOperand(0), Q, MaxRecurse - 1))
2854 case Instruction::Shl: {
2855 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2856 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2859 if (!NSW && ICmpInst::isSigned(Pred))
2861 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2862 RBO->getOperand(0), Q, MaxRecurse - 1))
2871 /// Simplify integer comparisons where at least one operand of the compare
2872 /// matches an integer min/max idiom.
2873 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2874 Value *RHS, const SimplifyQuery &Q,
2875 unsigned MaxRecurse) {
2876 Type *ITy = GetCompareTy(LHS); // The return type.
2878 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2879 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2881 // Signed variants on "max(a,b)>=a -> true".
2882 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2884 std::swap(A, B); // smax(A, B) pred A.
2885 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2886 // We analyze this as smax(A, B) pred A.
2888 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2889 (A == LHS || B == LHS)) {
2891 std::swap(A, B); // A pred smax(A, B).
2892 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2893 // We analyze this as smax(A, B) swapped-pred A.
2894 P = CmpInst::getSwappedPredicate(Pred);
2895 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2896 (A == RHS || B == RHS)) {
2898 std::swap(A, B); // smin(A, B) pred A.
2899 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2900 // We analyze this as smax(-A, -B) swapped-pred -A.
2901 // Note that we do not need to actually form -A or -B thanks to EqP.
2902 P = CmpInst::getSwappedPredicate(Pred);
2903 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2904 (A == LHS || B == LHS)) {
2906 std::swap(A, B); // A pred smin(A, B).
2907 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2908 // We analyze this as smax(-A, -B) pred -A.
2909 // Note that we do not need to actually form -A or -B thanks to EqP.
2912 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2913 // Cases correspond to "max(A, B) p A".
2917 case CmpInst::ICMP_EQ:
2918 case CmpInst::ICMP_SLE:
2919 // Equivalent to "A EqP B". This may be the same as the condition tested
2920 // in the max/min; if so, we can just return that.
2921 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2923 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2925 // Otherwise, see if "A EqP B" simplifies.
2927 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2930 case CmpInst::ICMP_NE:
2931 case CmpInst::ICMP_SGT: {
2932 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2933 // Equivalent to "A InvEqP B". This may be the same as the condition
2934 // tested in the max/min; if so, we can just return that.
2935 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2937 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2939 // Otherwise, see if "A InvEqP B" simplifies.
2941 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2945 case CmpInst::ICMP_SGE:
2947 return getTrue(ITy);
2948 case CmpInst::ICMP_SLT:
2950 return getFalse(ITy);
2954 // Unsigned variants on "max(a,b)>=a -> true".
2955 P = CmpInst::BAD_ICMP_PREDICATE;
2956 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2958 std::swap(A, B); // umax(A, B) pred A.
2959 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2960 // We analyze this as umax(A, B) pred A.
2962 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2963 (A == LHS || B == LHS)) {
2965 std::swap(A, B); // A pred umax(A, B).
2966 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2967 // We analyze this as umax(A, B) swapped-pred A.
2968 P = CmpInst::getSwappedPredicate(Pred);
2969 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2970 (A == RHS || B == RHS)) {
2972 std::swap(A, B); // umin(A, B) pred A.
2973 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2974 // We analyze this as umax(-A, -B) swapped-pred -A.
2975 // Note that we do not need to actually form -A or -B thanks to EqP.
2976 P = CmpInst::getSwappedPredicate(Pred);
2977 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2978 (A == LHS || B == LHS)) {
2980 std::swap(A, B); // A pred umin(A, B).
2981 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2982 // We analyze this as umax(-A, -B) pred -A.
2983 // Note that we do not need to actually form -A or -B thanks to EqP.
2986 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2987 // Cases correspond to "max(A, B) p A".
2991 case CmpInst::ICMP_EQ:
2992 case CmpInst::ICMP_ULE:
2993 // Equivalent to "A EqP B". This may be the same as the condition tested
2994 // in the max/min; if so, we can just return that.
2995 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2997 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2999 // Otherwise, see if "A EqP B" simplifies.
3001 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3004 case CmpInst::ICMP_NE:
3005 case CmpInst::ICMP_UGT: {
3006 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3007 // Equivalent to "A InvEqP B". This may be the same as the condition
3008 // tested in the max/min; if so, we can just return that.
3009 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3011 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3013 // Otherwise, see if "A InvEqP B" simplifies.
3015 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3019 case CmpInst::ICMP_UGE:
3021 return getTrue(ITy);
3022 case CmpInst::ICMP_ULT:
3024 return getFalse(ITy);
3028 // Variants on "max(x,y) >= min(x,z)".
3030 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3031 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3032 (A == C || A == D || B == C || B == D)) {
3033 // max(x, ?) pred min(x, ?).
3034 if (Pred == CmpInst::ICMP_SGE)
3036 return getTrue(ITy);
3037 if (Pred == CmpInst::ICMP_SLT)
3039 return getFalse(ITy);
3040 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3041 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3042 (A == C || A == D || B == C || B == D)) {
3043 // min(x, ?) pred max(x, ?).
3044 if (Pred == CmpInst::ICMP_SLE)
3046 return getTrue(ITy);
3047 if (Pred == CmpInst::ICMP_SGT)
3049 return getFalse(ITy);
3050 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3051 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3052 (A == C || A == D || B == C || B == D)) {
3053 // max(x, ?) pred min(x, ?).
3054 if (Pred == CmpInst::ICMP_UGE)
3056 return getTrue(ITy);
3057 if (Pred == CmpInst::ICMP_ULT)
3059 return getFalse(ITy);
3060 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3061 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3062 (A == C || A == D || B == C || B == D)) {
3063 // min(x, ?) pred max(x, ?).
3064 if (Pred == CmpInst::ICMP_ULE)
3066 return getTrue(ITy);
3067 if (Pred == CmpInst::ICMP_UGT)
3069 return getFalse(ITy);
3075 /// Given operands for an ICmpInst, see if we can fold the result.
3076 /// If not, this returns null.
3077 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3078 const SimplifyQuery &Q, unsigned MaxRecurse) {
3079 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3080 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3082 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3083 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3084 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3086 // If we have a constant, make sure it is on the RHS.
3087 std::swap(LHS, RHS);
3088 Pred = CmpInst::getSwappedPredicate(Pred);
3091 Type *ITy = GetCompareTy(LHS); // The return type.
3093 // icmp X, X -> true/false
3094 // icmp X, undef -> true/false because undef could be X.
3095 if (LHS == RHS || isa<UndefValue>(RHS))
3096 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3098 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3101 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3104 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3107 // If both operands have range metadata, use the metadata
3108 // to simplify the comparison.
3109 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3110 auto RHS_Instr = cast<Instruction>(RHS);
3111 auto LHS_Instr = cast<Instruction>(LHS);
3113 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3114 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3115 auto RHS_CR = getConstantRangeFromMetadata(
3116 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3117 auto LHS_CR = getConstantRangeFromMetadata(
3118 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3120 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3121 if (Satisfied_CR.contains(LHS_CR))
3122 return ConstantInt::getTrue(RHS->getContext());
3124 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3125 CmpInst::getInversePredicate(Pred), RHS_CR);
3126 if (InversedSatisfied_CR.contains(LHS_CR))
3127 return ConstantInt::getFalse(RHS->getContext());
3131 // Compare of cast, for example (zext X) != 0 -> X != 0
3132 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3133 Instruction *LI = cast<CastInst>(LHS);
3134 Value *SrcOp = LI->getOperand(0);
3135 Type *SrcTy = SrcOp->getType();
3136 Type *DstTy = LI->getType();
3138 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3139 // if the integer type is the same size as the pointer type.
3140 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3141 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3142 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3143 // Transfer the cast to the constant.
3144 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3145 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3148 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3149 if (RI->getOperand(0)->getType() == SrcTy)
3150 // Compare without the cast.
3151 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3157 if (isa<ZExtInst>(LHS)) {
3158 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3160 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3161 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3162 // Compare X and Y. Note that signed predicates become unsigned.
3163 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3164 SrcOp, RI->getOperand(0), Q,
3168 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3169 // too. If not, then try to deduce the result of the comparison.
3170 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3171 // Compute the constant that would happen if we truncated to SrcTy then
3172 // reextended to DstTy.
3173 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3174 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3176 // If the re-extended constant didn't change then this is effectively
3177 // also a case of comparing two zero-extended values.
3178 if (RExt == CI && MaxRecurse)
3179 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3180 SrcOp, Trunc, Q, MaxRecurse-1))
3183 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3184 // there. Use this to work out the result of the comparison.
3187 default: llvm_unreachable("Unknown ICmp predicate!");
3189 case ICmpInst::ICMP_EQ:
3190 case ICmpInst::ICMP_UGT:
3191 case ICmpInst::ICMP_UGE:
3192 return ConstantInt::getFalse(CI->getContext());
3194 case ICmpInst::ICMP_NE:
3195 case ICmpInst::ICMP_ULT:
3196 case ICmpInst::ICMP_ULE:
3197 return ConstantInt::getTrue(CI->getContext());
3199 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3200 // is non-negative then LHS <s RHS.
3201 case ICmpInst::ICMP_SGT:
3202 case ICmpInst::ICMP_SGE:
3203 return CI->getValue().isNegative() ?
3204 ConstantInt::getTrue(CI->getContext()) :
3205 ConstantInt::getFalse(CI->getContext());
3207 case ICmpInst::ICMP_SLT:
3208 case ICmpInst::ICMP_SLE:
3209 return CI->getValue().isNegative() ?
3210 ConstantInt::getFalse(CI->getContext()) :
3211 ConstantInt::getTrue(CI->getContext());
3217 if (isa<SExtInst>(LHS)) {
3218 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3220 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3221 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3222 // Compare X and Y. Note that the predicate does not change.
3223 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3227 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3228 // too. If not, then try to deduce the result of the comparison.
3229 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3230 // Compute the constant that would happen if we truncated to SrcTy then
3231 // reextended to DstTy.
3232 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3233 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3235 // If the re-extended constant didn't change then this is effectively
3236 // also a case of comparing two sign-extended values.
3237 if (RExt == CI && MaxRecurse)
3238 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3241 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3242 // bits there. Use this to work out the result of the comparison.
3245 default: llvm_unreachable("Unknown ICmp predicate!");
3246 case ICmpInst::ICMP_EQ:
3247 return ConstantInt::getFalse(CI->getContext());
3248 case ICmpInst::ICMP_NE:
3249 return ConstantInt::getTrue(CI->getContext());
3251 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3253 case ICmpInst::ICMP_SGT:
3254 case ICmpInst::ICMP_SGE:
3255 return CI->getValue().isNegative() ?
3256 ConstantInt::getTrue(CI->getContext()) :
3257 ConstantInt::getFalse(CI->getContext());
3258 case ICmpInst::ICMP_SLT:
3259 case ICmpInst::ICMP_SLE:
3260 return CI->getValue().isNegative() ?
3261 ConstantInt::getFalse(CI->getContext()) :
3262 ConstantInt::getTrue(CI->getContext());
3264 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3266 case ICmpInst::ICMP_UGT:
3267 case ICmpInst::ICMP_UGE:
3268 // Comparison is true iff the LHS <s 0.
3270 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3271 Constant::getNullValue(SrcTy),
3275 case ICmpInst::ICMP_ULT:
3276 case ICmpInst::ICMP_ULE:
3277 // Comparison is true iff the LHS >=s 0.
3279 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3280 Constant::getNullValue(SrcTy),
3290 // icmp eq|ne X, Y -> false|true if X != Y
3291 if (ICmpInst::isEquality(Pred) &&
3292 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3293 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3296 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3299 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3302 // Simplify comparisons of related pointers using a powerful, recursive
3303 // GEP-walk when we have target data available..
3304 if (LHS->getType()->isPointerTy())
3305 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3308 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3309 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3310 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3311 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3312 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3313 Q.DL.getTypeSizeInBits(CRHS->getType()))
3314 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3315 CLHS->getPointerOperand(),
3316 CRHS->getPointerOperand()))
3319 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3320 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3321 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3322 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3323 (ICmpInst::isEquality(Pred) ||
3324 (GLHS->isInBounds() && GRHS->isInBounds() &&
3325 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3326 // The bases are equal and the indices are constant. Build a constant
3327 // expression GEP with the same indices and a null base pointer to see
3328 // what constant folding can make out of it.
3329 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3330 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3331 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3332 GLHS->getSourceElementType(), Null, IndicesLHS);
3334 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3335 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3336 GLHS->getSourceElementType(), Null, IndicesRHS);
3337 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3342 // If the comparison is with the result of a select instruction, check whether
3343 // comparing with either branch of the select always yields the same value.
3344 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3345 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3348 // If the comparison is with the result of a phi instruction, check whether
3349 // doing the compare with each incoming phi value yields a common result.
3350 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3351 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3357 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3358 const SimplifyQuery &Q) {
3359 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3362 /// Given operands for an FCmpInst, see if we can fold the result.
3363 /// If not, this returns null.
3364 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3365 FastMathFlags FMF, const SimplifyQuery &Q,
3366 unsigned MaxRecurse) {
3367 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3368 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3370 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3371 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3372 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3374 // If we have a constant, make sure it is on the RHS.
3375 std::swap(LHS, RHS);
3376 Pred = CmpInst::getSwappedPredicate(Pred);
3379 // Fold trivial predicates.
3380 Type *RetTy = GetCompareTy(LHS);
3381 if (Pred == FCmpInst::FCMP_FALSE)
3382 return getFalse(RetTy);
3383 if (Pred == FCmpInst::FCMP_TRUE)
3384 return getTrue(RetTy);
3386 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3388 if (Pred == FCmpInst::FCMP_UNO)
3389 return getFalse(RetTy);
3390 if (Pred == FCmpInst::FCMP_ORD)
3391 return getTrue(RetTy);
3394 // NaN is unordered; NaN is not ordered.
3395 assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&
3396 "Comparison must be either ordered or unordered");
3397 if (match(RHS, m_NaN()))
3398 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3400 // fcmp pred x, undef and fcmp pred undef, x
3401 // fold to true if unordered, false if ordered
3402 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3403 // Choosing NaN for the undef will always make unordered comparison succeed
3404 // and ordered comparison fail.
3405 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3408 // fcmp x,x -> true/false. Not all compares are foldable.
3410 if (CmpInst::isTrueWhenEqual(Pred))
3411 return getTrue(RetTy);
3412 if (CmpInst::isFalseWhenEqual(Pred))
3413 return getFalse(RetTy);
3416 // Handle fcmp with constant RHS.
3418 if (match(RHS, m_APFloat(C))) {
3419 // Check whether the constant is an infinity.
3420 if (C->isInfinity()) {
3421 if (C->isNegative()) {
3423 case FCmpInst::FCMP_OLT:
3424 // No value is ordered and less than negative infinity.
3425 return getFalse(RetTy);
3426 case FCmpInst::FCMP_UGE:
3427 // All values are unordered with or at least negative infinity.
3428 return getTrue(RetTy);
3434 case FCmpInst::FCMP_OGT:
3435 // No value is ordered and greater than infinity.
3436 return getFalse(RetTy);
3437 case FCmpInst::FCMP_ULE:
3438 // All values are unordered with and at most infinity.
3439 return getTrue(RetTy);
3447 case FCmpInst::FCMP_UGE:
3448 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3449 return getTrue(RetTy);
3451 case FCmpInst::FCMP_OLT:
3453 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3454 return getFalse(RetTy);
3459 } else if (C->isNegative()) {
3460 assert(!C->isNaN() && "Unexpected NaN constant!");
3461 // TODO: We can catch more cases by using a range check rather than
3462 // relying on CannotBeOrderedLessThanZero.
3464 case FCmpInst::FCMP_UGE:
3465 case FCmpInst::FCMP_UGT:
3466 case FCmpInst::FCMP_UNE:
3467 // (X >= 0) implies (X > C) when (C < 0)
3468 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3469 return getTrue(RetTy);
3471 case FCmpInst::FCMP_OEQ:
3472 case FCmpInst::FCMP_OLE:
3473 case FCmpInst::FCMP_OLT:
3474 // (X >= 0) implies !(X < C) when (C < 0)
3475 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3476 return getFalse(RetTy);
3484 // If the comparison is with the result of a select instruction, check whether
3485 // comparing with either branch of the select always yields the same value.
3486 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3487 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3490 // If the comparison is with the result of a phi instruction, check whether
3491 // doing the compare with each incoming phi value yields a common result.
3492 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3493 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3499 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3500 FastMathFlags FMF, const SimplifyQuery &Q) {
3501 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3504 /// See if V simplifies when its operand Op is replaced with RepOp.
3505 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3506 const SimplifyQuery &Q,
3507 unsigned MaxRecurse) {
3508 // Trivial replacement.
3512 // We cannot replace a constant, and shouldn't even try.
3513 if (isa<Constant>(Op))
3516 auto *I = dyn_cast<Instruction>(V);
3520 // If this is a binary operator, try to simplify it with the replaced op.
3521 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3523 // %cmp = icmp eq i32 %x, 2147483647
3524 // %add = add nsw i32 %x, 1
3525 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3527 // We can't replace %sel with %add unless we strip away the flags.
3528 if (isa<OverflowingBinaryOperator>(B))
3529 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3531 if (isa<PossiblyExactOperator>(B))
3536 if (B->getOperand(0) == Op)
3537 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3539 if (B->getOperand(1) == Op)
3540 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3545 // Same for CmpInsts.
3546 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3548 if (C->getOperand(0) == Op)
3549 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3551 if (C->getOperand(1) == Op)
3552 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3558 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3560 SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3561 transform(GEP->operands(), NewOps.begin(),
3562 [&](Value *V) { return V == Op ? RepOp : V; });
3563 return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3568 // TODO: We could hand off more cases to instsimplify here.
3570 // If all operands are constant after substituting Op for RepOp then we can
3571 // constant fold the instruction.
3572 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3573 // Build a list of all constant operands.
3574 SmallVector<Constant *, 8> ConstOps;
3575 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3576 if (I->getOperand(i) == Op)
3577 ConstOps.push_back(CRepOp);
3578 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3579 ConstOps.push_back(COp);
3584 // All operands were constants, fold it.
3585 if (ConstOps.size() == I->getNumOperands()) {
3586 if (CmpInst *C = dyn_cast<CmpInst>(I))
3587 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3588 ConstOps[1], Q.DL, Q.TLI);
3590 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3591 if (!LI->isVolatile())
3592 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3594 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3601 /// Try to simplify a select instruction when its condition operand is an
3602 /// integer comparison where one operand of the compare is a constant.
3603 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3604 const APInt *Y, bool TrueWhenUnset) {
3607 // (X & Y) == 0 ? X & ~Y : X --> X
3608 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3609 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3611 return TrueWhenUnset ? FalseVal : TrueVal;
3613 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3614 // (X & Y) != 0 ? X : X & ~Y --> X
3615 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3617 return TrueWhenUnset ? FalseVal : TrueVal;
3619 if (Y->isPowerOf2()) {
3620 // (X & Y) == 0 ? X | Y : X --> X | Y
3621 // (X & Y) != 0 ? X | Y : X --> X
3622 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3624 return TrueWhenUnset ? TrueVal : FalseVal;
3626 // (X & Y) == 0 ? X : X | Y --> X
3627 // (X & Y) != 0 ? X : X | Y --> X | Y
3628 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3630 return TrueWhenUnset ? TrueVal : FalseVal;
3636 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3638 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3639 ICmpInst::Predicate Pred,
3640 Value *TrueVal, Value *FalseVal) {
3643 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3646 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3647 Pred == ICmpInst::ICMP_EQ);
3650 /// Try to simplify a select instruction when its condition operand is an
3651 /// integer comparison.
3652 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3653 Value *FalseVal, const SimplifyQuery &Q,
3654 unsigned MaxRecurse) {
3655 ICmpInst::Predicate Pred;
3656 Value *CmpLHS, *CmpRHS;
3657 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3660 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3663 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3664 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3665 Pred == ICmpInst::ICMP_EQ))
3669 // Check for other compares that behave like bit test.
3670 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3674 // If we have an equality comparison, then we know the value in one of the
3675 // arms of the select. See if substituting this value into the arm and
3676 // simplifying the result yields the same value as the other arm.
3677 if (Pred == ICmpInst::ICMP_EQ) {
3678 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3680 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3683 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3685 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3688 } else if (Pred == ICmpInst::ICMP_NE) {
3689 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3691 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3694 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3696 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3704 /// Given operands for a SelectInst, see if we can fold the result.
3705 /// If not, this returns null.
3706 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3707 const SimplifyQuery &Q, unsigned MaxRecurse) {
3708 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3709 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3710 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3711 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3713 // select undef, X, Y -> X or Y
3714 if (isa<UndefValue>(CondC))
3715 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3717 // TODO: Vector constants with undef elements don't simplify.
3719 // select true, X, Y -> X
3720 if (CondC->isAllOnesValue())
3722 // select false, X, Y -> Y
3723 if (CondC->isNullValue())
3727 // select ?, X, X -> X
3728 if (TrueVal == FalseVal)
3731 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3733 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3737 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3743 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3744 const SimplifyQuery &Q) {
3745 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3748 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3749 /// If not, this returns null.
3750 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3751 const SimplifyQuery &Q, unsigned) {
3752 // The type of the GEP pointer operand.
3754 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3756 // getelementptr P -> P.
3757 if (Ops.size() == 1)
3760 // Compute the (pointer) type returned by the GEP instruction.
3761 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3762 Type *GEPTy = PointerType::get(LastType, AS);
3763 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3764 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3765 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3766 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3768 if (isa<UndefValue>(Ops[0]))
3769 return UndefValue::get(GEPTy);
3771 if (Ops.size() == 2) {
3772 // getelementptr P, 0 -> P.
3773 if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3777 if (Ty->isSized()) {
3780 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3781 // getelementptr P, N -> P if P points to a type of zero size.
3782 if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3785 // The following transforms are only safe if the ptrtoint cast
3786 // doesn't truncate the pointers.
3787 if (Ops[1]->getType()->getScalarSizeInBits() ==
3788 Q.DL.getIndexSizeInBits(AS)) {
3789 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3790 if (match(P, m_Zero()))
3791 return Constant::getNullValue(GEPTy);
3793 if (match(P, m_PtrToInt(m_Value(Temp))))
3794 if (Temp->getType() == GEPTy)
3799 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3800 if (TyAllocSize == 1 &&
3801 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3802 if (Value *R = PtrToIntOrZero(P))
3805 // getelementptr V, (ashr (sub P, V), C) -> Q
3806 // if P points to a type of size 1 << C.
3808 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3809 m_ConstantInt(C))) &&
3810 TyAllocSize == 1ULL << C)
3811 if (Value *R = PtrToIntOrZero(P))
3814 // getelementptr V, (sdiv (sub P, V), C) -> Q
3815 // if P points to a type of size C.
3817 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3818 m_SpecificInt(TyAllocSize))))
3819 if (Value *R = PtrToIntOrZero(P))
3825 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3826 all_of(Ops.slice(1).drop_back(1),
3827 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3829 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3830 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3831 APInt BasePtrOffset(IdxWidth, 0);
3832 Value *StrippedBasePtr =
3833 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3836 // gep (gep V, C), (sub 0, V) -> C
3837 if (match(Ops.back(),
3838 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3839 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3840 return ConstantExpr::getIntToPtr(CI, GEPTy);
3842 // gep (gep V, C), (xor V, -1) -> C-1
3843 if (match(Ops.back(),
3844 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3845 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3846 return ConstantExpr::getIntToPtr(CI, GEPTy);
3851 // Check to see if this is constant foldable.
3852 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3855 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3857 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3862 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3863 const SimplifyQuery &Q) {
3864 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3867 /// Given operands for an InsertValueInst, see if we can fold the result.
3868 /// If not, this returns null.
3869 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3870 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3872 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3873 if (Constant *CVal = dyn_cast<Constant>(Val))
3874 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3876 // insertvalue x, undef, n -> x
3877 if (match(Val, m_Undef()))
3880 // insertvalue x, (extractvalue y, n), n
3881 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3882 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3883 EV->getIndices() == Idxs) {
3884 // insertvalue undef, (extractvalue y, n), n -> y
3885 if (match(Agg, m_Undef()))
3886 return EV->getAggregateOperand();
3888 // insertvalue y, (extractvalue y, n), n -> y
3889 if (Agg == EV->getAggregateOperand())
3896 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3897 ArrayRef<unsigned> Idxs,
3898 const SimplifyQuery &Q) {
3899 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3902 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3903 const SimplifyQuery &Q) {
3904 // Try to constant fold.
3905 auto *VecC = dyn_cast<Constant>(Vec);
3906 auto *ValC = dyn_cast<Constant>(Val);
3907 auto *IdxC = dyn_cast<Constant>(Idx);
3908 if (VecC && ValC && IdxC)
3909 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3911 // Fold into undef if index is out of bounds.
3912 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3913 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3914 if (CI->uge(NumElements))
3915 return UndefValue::get(Vec->getType());
3918 // If index is undef, it might be out of bounds (see above case)
3919 if (isa<UndefValue>(Idx))
3920 return UndefValue::get(Vec->getType());
3925 /// Given operands for an ExtractValueInst, see if we can fold the result.
3926 /// If not, this returns null.
3927 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3928 const SimplifyQuery &, unsigned) {
3929 if (auto *CAgg = dyn_cast<Constant>(Agg))
3930 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3932 // extractvalue x, (insertvalue y, elt, n), n -> elt
3933 unsigned NumIdxs = Idxs.size();
3934 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3935 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3936 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3937 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3938 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3939 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3940 Idxs.slice(0, NumCommonIdxs)) {
3941 if (NumIdxs == NumInsertValueIdxs)
3942 return IVI->getInsertedValueOperand();
3950 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3951 const SimplifyQuery &Q) {
3952 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3955 /// Given operands for an ExtractElementInst, see if we can fold the result.
3956 /// If not, this returns null.
3957 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3959 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3960 if (auto *CIdx = dyn_cast<Constant>(Idx))
3961 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3963 // The index is not relevant if our vector is a splat.
3964 if (auto *Splat = CVec->getSplatValue())
3967 if (isa<UndefValue>(Vec))
3968 return UndefValue::get(Vec->getType()->getVectorElementType());
3971 // If extracting a specified index from the vector, see if we can recursively
3972 // find a previously computed scalar that was inserted into the vector.
3973 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
3974 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
3975 // definitely out of bounds, thus undefined result
3976 return UndefValue::get(Vec->getType()->getVectorElementType());
3977 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3981 // An undef extract index can be arbitrarily chosen to be an out-of-range
3982 // index value, which would result in the instruction being undef.
3983 if (isa<UndefValue>(Idx))
3984 return UndefValue::get(Vec->getType()->getVectorElementType());
3989 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3990 const SimplifyQuery &Q) {
3991 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3994 /// See if we can fold the given phi. If not, returns null.
3995 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3996 // If all of the PHI's incoming values are the same then replace the PHI node
3997 // with the common value.
3998 Value *CommonValue = nullptr;
3999 bool HasUndefInput = false;
4000 for (Value *Incoming : PN->incoming_values()) {
4001 // If the incoming value is the phi node itself, it can safely be skipped.
4002 if (Incoming == PN) continue;
4003 if (isa<UndefValue>(Incoming)) {
4004 // Remember that we saw an undef value, but otherwise ignore them.
4005 HasUndefInput = true;
4008 if (CommonValue && Incoming != CommonValue)
4009 return nullptr; // Not the same, bail out.
4010 CommonValue = Incoming;
4013 // If CommonValue is null then all of the incoming values were either undef or
4014 // equal to the phi node itself.
4016 return UndefValue::get(PN->getType());
4018 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4019 // instruction, we cannot return X as the result of the PHI node unless it
4020 // dominates the PHI block.
4022 return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4027 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4028 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4029 if (auto *C = dyn_cast<Constant>(Op))
4030 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4032 if (auto *CI = dyn_cast<CastInst>(Op)) {
4033 auto *Src = CI->getOperand(0);
4034 Type *SrcTy = Src->getType();
4035 Type *MidTy = CI->getType();
4037 if (Src->getType() == Ty) {
4038 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4039 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4041 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4043 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4045 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4046 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4047 SrcIntPtrTy, MidIntPtrTy,
4048 DstIntPtrTy) == Instruction::BitCast)
4054 if (CastOpc == Instruction::BitCast)
4055 if (Op->getType() == Ty)
4061 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4062 const SimplifyQuery &Q) {
4063 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4066 /// For the given destination element of a shuffle, peek through shuffles to
4067 /// match a root vector source operand that contains that element in the same
4068 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4069 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4070 int MaskVal, Value *RootVec,
4071 unsigned MaxRecurse) {
4075 // Bail out if any mask value is undefined. That kind of shuffle may be
4076 // simplified further based on demanded bits or other folds.
4080 // The mask value chooses which source operand we need to look at next.
4081 int InVecNumElts = Op0->getType()->getVectorNumElements();
4082 int RootElt = MaskVal;
4083 Value *SourceOp = Op0;
4084 if (MaskVal >= InVecNumElts) {
4085 RootElt = MaskVal - InVecNumElts;
4089 // If the source operand is a shuffle itself, look through it to find the
4090 // matching root vector.
4091 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4092 return foldIdentityShuffles(
4093 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4094 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4097 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4100 // The source operand is not a shuffle. Initialize the root vector value for
4101 // this shuffle if that has not been done yet.
4105 // Give up as soon as a source operand does not match the existing root value.
4106 if (RootVec != SourceOp)
4109 // The element must be coming from the same lane in the source vector
4110 // (although it may have crossed lanes in intermediate shuffles).
4111 if (RootElt != DestElt)
4117 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4118 Type *RetTy, const SimplifyQuery &Q,
4119 unsigned MaxRecurse) {
4120 if (isa<UndefValue>(Mask))
4121 return UndefValue::get(RetTy);
4123 Type *InVecTy = Op0->getType();
4124 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4125 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4127 SmallVector<int, 32> Indices;
4128 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4129 assert(MaskNumElts == Indices.size() &&
4130 "Size of Indices not same as number of mask elements?");
4132 // Canonicalization: If mask does not select elements from an input vector,
4133 // replace that input vector with undef.
4134 bool MaskSelects0 = false, MaskSelects1 = false;
4135 for (unsigned i = 0; i != MaskNumElts; ++i) {
4136 if (Indices[i] == -1)
4138 if ((unsigned)Indices[i] < InVecNumElts)
4139 MaskSelects0 = true;
4141 MaskSelects1 = true;
4144 Op0 = UndefValue::get(InVecTy);
4146 Op1 = UndefValue::get(InVecTy);
4148 auto *Op0Const = dyn_cast<Constant>(Op0);
4149 auto *Op1Const = dyn_cast<Constant>(Op1);
4151 // If all operands are constant, constant fold the shuffle.
4152 if (Op0Const && Op1Const)
4153 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4155 // Canonicalization: if only one input vector is constant, it shall be the
4157 if (Op0Const && !Op1Const) {
4158 std::swap(Op0, Op1);
4159 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4162 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4163 // value type is same as the input vectors' type.
4164 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4165 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4166 OpShuf->getMask()->getSplatValue())
4169 // Don't fold a shuffle with undef mask elements. This may get folded in a
4170 // better way using demanded bits or other analysis.
4171 // TODO: Should we allow this?
4172 if (find(Indices, -1) != Indices.end())
4175 // Check if every element of this shuffle can be mapped back to the
4176 // corresponding element of a single root vector. If so, we don't need this
4177 // shuffle. This handles simple identity shuffles as well as chains of
4178 // shuffles that may widen/narrow and/or move elements across lanes and back.
4179 Value *RootVec = nullptr;
4180 for (unsigned i = 0; i != MaskNumElts; ++i) {
4181 // Note that recursion is limited for each vector element, so if any element
4182 // exceeds the limit, this will fail to simplify.
4184 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4186 // We can't replace a widening/narrowing shuffle with one of its operands.
4187 if (!RootVec || RootVec->getType() != RetTy)
4193 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4194 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4195 Type *RetTy, const SimplifyQuery &Q) {
4196 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4199 static Constant *propagateNaN(Constant *In) {
4200 // If the input is a vector with undef elements, just return a default NaN.
4202 return ConstantFP::getNaN(In->getType());
4204 // Propagate the existing NaN constant when possible.
4205 // TODO: Should we quiet a signaling NaN?
4209 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4210 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4211 return ConstantFP::getNaN(Op0->getType());
4213 if (match(Op0, m_NaN()))
4214 return propagateNaN(cast<Constant>(Op0));
4215 if (match(Op1, m_NaN()))
4216 return propagateNaN(cast<Constant>(Op1));
4221 /// Given operands for an FAdd, see if we can fold the result. If not, this
4223 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4224 const SimplifyQuery &Q, unsigned MaxRecurse) {
4225 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4228 if (Constant *C = simplifyFPBinop(Op0, Op1))
4232 if (match(Op1, m_NegZeroFP()))
4235 // fadd X, 0 ==> X, when we know X is not -0
4236 if (match(Op1, m_PosZeroFP()) &&
4237 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4240 // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4241 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4242 // Negative zeros are allowed because we always end up with positive zero:
4243 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4244 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4245 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4246 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4247 if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4248 match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4249 return ConstantFP::getNullValue(Op0->getType());
4254 /// Given operands for an FSub, see if we can fold the result. If not, this
4256 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4257 const SimplifyQuery &Q, unsigned MaxRecurse) {
4258 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4261 if (Constant *C = simplifyFPBinop(Op0, Op1))
4265 if (match(Op1, m_PosZeroFP()))
4268 // fsub X, -0 ==> X, when we know X is not -0
4269 if (match(Op1, m_NegZeroFP()) &&
4270 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4273 // fsub -0.0, (fsub -0.0, X) ==> X
4275 if (match(Op0, m_NegZeroFP()) &&
4276 match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4279 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4280 if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4281 match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4284 // fsub nnan x, x ==> 0.0
4285 if (FMF.noNaNs() && Op0 == Op1)
4286 return Constant::getNullValue(Op0->getType());
4291 /// Given the operands for an FMul, see if we can fold the result
4292 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4293 const SimplifyQuery &Q, unsigned MaxRecurse) {
4294 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4297 if (Constant *C = simplifyFPBinop(Op0, Op1))
4300 // fmul X, 1.0 ==> X
4301 if (match(Op1, m_FPOne()))
4304 // fmul nnan nsz X, 0 ==> 0
4305 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4306 return ConstantFP::getNullValue(Op0->getType());
4308 // sqrt(X) * sqrt(X) --> X, if we can:
4309 // 1. Remove the intermediate rounding (reassociate).
4310 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4311 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4313 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4314 FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4320 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4321 const SimplifyQuery &Q) {
4322 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4326 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4327 const SimplifyQuery &Q) {
4328 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4331 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4332 const SimplifyQuery &Q) {
4333 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4336 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4337 const SimplifyQuery &Q, unsigned) {
4338 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4341 if (Constant *C = simplifyFPBinop(Op0, Op1))
4345 if (match(Op1, m_FPOne()))
4349 // Requires that NaNs are off (X could be zero) and signed zeroes are
4350 // ignored (X could be positive or negative, so the output sign is unknown).
4351 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4352 return ConstantFP::getNullValue(Op0->getType());
4355 // X / X -> 1.0 is legal when NaNs are ignored.
4356 // We can ignore infinities because INF/INF is NaN.
4358 return ConstantFP::get(Op0->getType(), 1.0);
4360 // (X * Y) / Y --> X if we can reassociate to the above form.
4362 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4365 // -X / X -> -1.0 and
4366 // X / -X -> -1.0 are legal when NaNs are ignored.
4367 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4368 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4369 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4370 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4371 BinaryOperator::getFNegArgument(Op1) == Op0))
4372 return ConstantFP::get(Op0->getType(), -1.0);
4378 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4379 const SimplifyQuery &Q) {
4380 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4383 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4384 const SimplifyQuery &Q, unsigned) {
4385 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4388 if (Constant *C = simplifyFPBinop(Op0, Op1))
4391 // Unlike fdiv, the result of frem always matches the sign of the dividend.
4392 // The constant match may include undef elements in a vector, so return a full
4393 // zero constant as the result.
4396 if (match(Op0, m_PosZeroFP()))
4397 return ConstantFP::getNullValue(Op0->getType());
4399 if (match(Op0, m_NegZeroFP()))
4400 return ConstantFP::getNegativeZero(Op0->getType());
4406 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4407 const SimplifyQuery &Q) {
4408 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4411 //=== Helper functions for higher up the class hierarchy.
4413 /// Given operands for a BinaryOperator, see if we can fold the result.
4414 /// If not, this returns null.
4415 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4416 const SimplifyQuery &Q, unsigned MaxRecurse) {
4418 case Instruction::Add:
4419 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4420 case Instruction::Sub:
4421 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4422 case Instruction::Mul:
4423 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4424 case Instruction::SDiv:
4425 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4426 case Instruction::UDiv:
4427 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4428 case Instruction::SRem:
4429 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4430 case Instruction::URem:
4431 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4432 case Instruction::Shl:
4433 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4434 case Instruction::LShr:
4435 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4436 case Instruction::AShr:
4437 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4438 case Instruction::And:
4439 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4440 case Instruction::Or:
4441 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4442 case Instruction::Xor:
4443 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4444 case Instruction::FAdd:
4445 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4446 case Instruction::FSub:
4447 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4448 case Instruction::FMul:
4449 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4450 case Instruction::FDiv:
4451 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4452 case Instruction::FRem:
4453 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4455 llvm_unreachable("Unexpected opcode");
4459 /// Given operands for a BinaryOperator, see if we can fold the result.
4460 /// If not, this returns null.
4461 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4462 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4463 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4464 const FastMathFlags &FMF, const SimplifyQuery &Q,
4465 unsigned MaxRecurse) {
4467 case Instruction::FAdd:
4468 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4469 case Instruction::FSub:
4470 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4471 case Instruction::FMul:
4472 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4473 case Instruction::FDiv:
4474 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4476 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4480 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4481 const SimplifyQuery &Q) {
4482 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4485 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4486 FastMathFlags FMF, const SimplifyQuery &Q) {
4487 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4490 /// Given operands for a CmpInst, see if we can fold the result.
4491 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4492 const SimplifyQuery &Q, unsigned MaxRecurse) {
4493 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4494 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4495 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4498 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4499 const SimplifyQuery &Q) {
4500 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4503 static bool IsIdempotent(Intrinsic::ID ID) {
4505 default: return false;
4507 // Unary idempotent: f(f(x)) = f(x)
4508 case Intrinsic::fabs:
4509 case Intrinsic::floor:
4510 case Intrinsic::ceil:
4511 case Intrinsic::trunc:
4512 case Intrinsic::rint:
4513 case Intrinsic::nearbyint:
4514 case Intrinsic::round:
4515 case Intrinsic::canonicalize:
4520 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4521 const DataLayout &DL) {
4522 GlobalValue *PtrSym;
4524 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4527 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4528 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4529 Type *Int32PtrTy = Int32Ty->getPointerTo();
4530 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4532 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4533 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4536 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4537 if (OffsetInt % 4 != 0)
4540 Constant *C = ConstantExpr::getGetElementPtr(
4541 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4542 ConstantInt::get(Int64Ty, OffsetInt / 4));
4543 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4547 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4551 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4552 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4557 if (LoadedCE->getOpcode() != Instruction::Sub)
4560 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4561 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4563 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4565 Constant *LoadedRHS = LoadedCE->getOperand(1);
4566 GlobalValue *LoadedRHSSym;
4567 APInt LoadedRHSOffset;
4568 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4570 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4573 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4576 static bool maskIsAllZeroOrUndef(Value *Mask) {
4577 auto *ConstMask = dyn_cast<Constant>(Mask);
4580 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4582 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4584 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4585 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4592 template <typename IterTy>
4593 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4594 const SimplifyQuery &Q, unsigned MaxRecurse) {
4595 Intrinsic::ID IID = F->getIntrinsicID();
4596 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4599 if (NumOperands == 1) {
4600 // Perform idempotent optimizations
4601 if (IsIdempotent(IID)) {
4602 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4603 if (II->getIntrinsicID() == IID)
4608 Value *IIOperand = *ArgBegin;
4611 case Intrinsic::fabs: {
4612 if (SignBitMustBeZero(IIOperand, Q.TLI))
4616 case Intrinsic::bswap: {
4617 // bswap(bswap(x)) -> x
4618 if (match(IIOperand, m_BSwap(m_Value(X))))
4622 case Intrinsic::bitreverse: {
4623 // bitreverse(bitreverse(x)) -> x
4624 if (match(IIOperand, m_BitReverse(m_Value(X))))
4628 case Intrinsic::exp: {
4630 if (Q.CxtI->hasAllowReassoc() &&
4631 match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X))))
4635 case Intrinsic::exp2: {
4636 // exp2(log2(x)) -> x
4637 if (Q.CxtI->hasAllowReassoc() &&
4638 match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X))))
4642 case Intrinsic::log: {
4644 if (Q.CxtI->hasAllowReassoc() &&
4645 match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X))))
4649 case Intrinsic::log2: {
4650 // log2(exp2(x)) -> x
4651 if (Q.CxtI->hasAllowReassoc() &&
4652 match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) {
4663 if (NumOperands == 2) {
4664 Value *LHS = *ArgBegin;
4665 Value *RHS = *(ArgBegin + 1);
4666 Type *ReturnType = F->getReturnType();
4669 case Intrinsic::usub_with_overflow:
4670 case Intrinsic::ssub_with_overflow: {
4671 // X - X -> { 0, false }
4673 return Constant::getNullValue(ReturnType);
4675 // X - undef -> undef
4676 // undef - X -> undef
4677 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4678 return UndefValue::get(ReturnType);
4682 case Intrinsic::uadd_with_overflow:
4683 case Intrinsic::sadd_with_overflow: {
4684 // X + undef -> undef
4685 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4686 return UndefValue::get(ReturnType);
4690 case Intrinsic::umul_with_overflow:
4691 case Intrinsic::smul_with_overflow: {
4692 // 0 * X -> { 0, false }
4693 // X * 0 -> { 0, false }
4694 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4695 return Constant::getNullValue(ReturnType);
4697 // undef * X -> { 0, false }
4698 // X * undef -> { 0, false }
4699 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4700 return Constant::getNullValue(ReturnType);
4704 case Intrinsic::load_relative: {
4705 Constant *C0 = dyn_cast<Constant>(LHS);
4706 Constant *C1 = dyn_cast<Constant>(RHS);
4708 return SimplifyRelativeLoad(C0, C1, Q.DL);
4711 case Intrinsic::powi:
4712 if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) {
4713 // powi(x, 0) -> 1.0
4714 if (Power->isZero())
4715 return ConstantFP::get(LHS->getType(), 1.0);
4726 // Simplify calls to llvm.masked.load.*
4728 case Intrinsic::masked_load: {
4729 Value *MaskArg = ArgBegin[2];
4730 Value *PassthruArg = ArgBegin[3];
4731 // If the mask is all zeros or undef, the "passthru" argument is the result.
4732 if (maskIsAllZeroOrUndef(MaskArg))
4741 template <typename IterTy>
4742 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4743 IterTy ArgEnd, const SimplifyQuery &Q,
4744 unsigned MaxRecurse) {
4745 Type *Ty = V->getType();
4746 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4747 Ty = PTy->getElementType();
4748 FunctionType *FTy = cast<FunctionType>(Ty);
4750 // call undef -> undef
4751 // call null -> undef
4752 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4753 return UndefValue::get(FTy->getReturnType());
4755 Function *F = dyn_cast<Function>(V);
4759 if (F->isIntrinsic())
4760 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4763 if (!canConstantFoldCallTo(CS, F))
4766 SmallVector<Constant *, 4> ConstantArgs;
4767 ConstantArgs.reserve(ArgEnd - ArgBegin);
4768 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4769 Constant *C = dyn_cast<Constant>(*I);
4772 ConstantArgs.push_back(C);
4775 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4778 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4779 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4780 const SimplifyQuery &Q) {
4781 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4784 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4785 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4786 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4789 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4790 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4791 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4795 /// See if we can compute a simplified version of this instruction.
4796 /// If not, this returns null.
4798 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4799 OptimizationRemarkEmitter *ORE) {
4800 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4803 switch (I->getOpcode()) {
4805 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4807 case Instruction::FAdd:
4808 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4809 I->getFastMathFlags(), Q);
4811 case Instruction::Add:
4812 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4813 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4814 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4816 case Instruction::FSub:
4817 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4818 I->getFastMathFlags(), Q);
4820 case Instruction::Sub:
4821 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4822 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4823 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4825 case Instruction::FMul:
4826 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4827 I->getFastMathFlags(), Q);
4829 case Instruction::Mul:
4830 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4832 case Instruction::SDiv:
4833 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4835 case Instruction::UDiv:
4836 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4838 case Instruction::FDiv:
4839 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4840 I->getFastMathFlags(), Q);
4842 case Instruction::SRem:
4843 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4845 case Instruction::URem:
4846 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4848 case Instruction::FRem:
4849 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4850 I->getFastMathFlags(), Q);
4852 case Instruction::Shl:
4853 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4854 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4855 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4857 case Instruction::LShr:
4858 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4859 cast<BinaryOperator>(I)->isExact(), Q);
4861 case Instruction::AShr:
4862 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4863 cast<BinaryOperator>(I)->isExact(), Q);
4865 case Instruction::And:
4866 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4868 case Instruction::Or:
4869 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4871 case Instruction::Xor:
4872 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4874 case Instruction::ICmp:
4875 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4876 I->getOperand(0), I->getOperand(1), Q);
4878 case Instruction::FCmp:
4880 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4881 I->getOperand(1), I->getFastMathFlags(), Q);
4883 case Instruction::Select:
4884 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4885 I->getOperand(2), Q);
4887 case Instruction::GetElementPtr: {
4888 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4889 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4893 case Instruction::InsertValue: {
4894 InsertValueInst *IV = cast<InsertValueInst>(I);
4895 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4896 IV->getInsertedValueOperand(),
4897 IV->getIndices(), Q);
4900 case Instruction::InsertElement: {
4901 auto *IE = cast<InsertElementInst>(I);
4902 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4903 IE->getOperand(2), Q);
4906 case Instruction::ExtractValue: {
4907 auto *EVI = cast<ExtractValueInst>(I);
4908 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4909 EVI->getIndices(), Q);
4912 case Instruction::ExtractElement: {
4913 auto *EEI = cast<ExtractElementInst>(I);
4914 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4915 EEI->getIndexOperand(), Q);
4918 case Instruction::ShuffleVector: {
4919 auto *SVI = cast<ShuffleVectorInst>(I);
4920 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4921 SVI->getMask(), SVI->getType(), Q);
4924 case Instruction::PHI:
4925 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4927 case Instruction::Call: {
4928 CallSite CS(cast<CallInst>(I));
4929 Result = SimplifyCall(CS, Q);
4932 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4933 #include "llvm/IR/Instruction.def"
4934 #undef HANDLE_CAST_INST
4936 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4938 case Instruction::Alloca:
4939 // No simplifications for Alloca and it can't be constant folded.
4944 // In general, it is possible for computeKnownBits to determine all bits in a
4945 // value even when the operands are not all constants.
4946 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4947 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4948 if (Known.isConstant())
4949 Result = ConstantInt::get(I->getType(), Known.getConstant());
4952 /// If called on unreachable code, the above logic may report that the
4953 /// instruction simplified to itself. Make life easier for users by
4954 /// detecting that case here, returning a safe value instead.
4955 return Result == I ? UndefValue::get(I->getType()) : Result;
4958 /// Implementation of recursive simplification through an instruction's
4961 /// This is the common implementation of the recursive simplification routines.
4962 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4963 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4964 /// instructions to process and attempt to simplify it using
4965 /// InstructionSimplify.
4967 /// This routine returns 'true' only when *it* simplifies something. The passed
4968 /// in simplified value does not count toward this.
4969 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4970 const TargetLibraryInfo *TLI,
4971 const DominatorTree *DT,
4972 AssumptionCache *AC) {
4973 bool Simplified = false;
4974 SmallSetVector<Instruction *, 8> Worklist;
4975 const DataLayout &DL = I->getModule()->getDataLayout();
4977 // If we have an explicit value to collapse to, do that round of the
4978 // simplification loop by hand initially.
4980 for (User *U : I->users())
4982 Worklist.insert(cast<Instruction>(U));
4984 // Replace the instruction with its simplified value.
4985 I->replaceAllUsesWith(SimpleV);
4987 // Gracefully handle edge cases where the instruction is not wired into any
4989 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4990 !I->mayHaveSideEffects())
4991 I->eraseFromParent();
4996 // Note that we must test the size on each iteration, the worklist can grow.
4997 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
5000 // See if this instruction simplifies.
5001 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
5007 // Stash away all the uses of the old instruction so we can check them for
5008 // recursive simplifications after a RAUW. This is cheaper than checking all
5009 // uses of To on the recursive step in most cases.
5010 for (User *U : I->users())
5011 Worklist.insert(cast<Instruction>(U));
5013 // Replace the instruction with its simplified value.
5014 I->replaceAllUsesWith(SimpleV);
5016 // Gracefully handle edge cases where the instruction is not wired into any
5018 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
5019 !I->mayHaveSideEffects())
5020 I->eraseFromParent();
5025 bool llvm::recursivelySimplifyInstruction(Instruction *I,
5026 const TargetLibraryInfo *TLI,
5027 const DominatorTree *DT,
5028 AssumptionCache *AC) {
5029 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
5032 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
5033 const TargetLibraryInfo *TLI,
5034 const DominatorTree *DT,
5035 AssumptionCache *AC) {
5036 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5037 assert(SimpleV && "Must provide a simplified value.");
5038 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
5042 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
5043 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
5044 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
5045 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
5046 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
5047 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5048 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5049 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5052 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5053 const DataLayout &DL) {
5054 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5057 template <class T, class... TArgs>
5058 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5060 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5061 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5062 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5063 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5065 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,