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()))
543 // If two operands are negative, return 0.
544 if (isKnownNegation(Op0, Op1))
545 return Constant::getNullValue(Op0->getType());
551 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
552 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
555 // X + ~X -> -1 since ~X = -X-1
556 Type *Ty = Op0->getType();
557 if (match(Op0, m_Not(m_Specific(Op1))) ||
558 match(Op1, m_Not(m_Specific(Op0))))
559 return Constant::getAllOnesValue(Ty);
561 // add nsw/nuw (xor Y, signmask), signmask --> Y
562 // The no-wrapping add guarantees that the top bit will be set by the add.
563 // Therefore, the xor must be clearing the already set sign bit of Y.
564 if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
565 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
568 // add nuw %x, -1 -> -1, because %x can only be 0.
569 if (IsNUW && match(Op1, m_AllOnes()))
570 return Op1; // Which is -1.
573 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
574 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
577 // Try some generic simplifications for associative operations.
578 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
582 // Threading Add over selects and phi nodes is pointless, so don't bother.
583 // Threading over the select in "A + select(cond, B, C)" means evaluating
584 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
585 // only if B and C are equal. If B and C are equal then (since we assume
586 // that operands have already been simplified) "select(cond, B, C)" should
587 // have been simplified to the common value of B and C already. Analysing
588 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
589 // for threading over phi nodes.
594 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
595 const SimplifyQuery &Query) {
596 return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
599 /// Compute the base pointer and cumulative constant offsets for V.
601 /// This strips all constant offsets off of V, leaving it the base pointer, and
602 /// accumulates the total constant offset applied in the returned constant. It
603 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
604 /// no constant offsets applied.
606 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
607 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
609 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
610 bool AllowNonInbounds = false) {
611 assert(V->getType()->isPtrOrPtrVectorTy());
613 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
614 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
616 // Even though we don't look through PHI nodes, we could be called on an
617 // instruction in an unreachable block, which may be on a cycle.
618 SmallPtrSet<Value *, 4> Visited;
621 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
622 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
623 !GEP->accumulateConstantOffset(DL, Offset))
625 V = GEP->getPointerOperand();
626 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
627 V = cast<Operator>(V)->getOperand(0);
628 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
629 if (GA->isInterposable())
631 V = GA->getAliasee();
633 if (auto CS = CallSite(V))
634 if (Value *RV = CS.getReturnedArgOperand()) {
640 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
641 } while (Visited.insert(V).second);
643 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
644 if (V->getType()->isVectorTy())
645 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
650 /// Compute the constant difference between two pointer values.
651 /// If the difference is not a constant, returns zero.
652 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
654 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
655 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
657 // If LHS and RHS are not related via constant offsets to the same base
658 // value, there is nothing we can do here.
662 // Otherwise, the difference of LHS - RHS can be computed as:
664 // = (LHSOffset + Base) - (RHSOffset + Base)
665 // = LHSOffset - RHSOffset
666 return ConstantExpr::getSub(LHSOffset, RHSOffset);
669 /// Given operands for a Sub, see if we can fold the result.
670 /// If not, this returns null.
671 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
672 const SimplifyQuery &Q, unsigned MaxRecurse) {
673 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
676 // X - undef -> undef
677 // undef - X -> undef
678 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
679 return UndefValue::get(Op0->getType());
682 if (match(Op1, m_Zero()))
687 return Constant::getNullValue(Op0->getType());
689 // Is this a negation?
690 if (match(Op0, m_Zero())) {
691 // 0 - X -> 0 if the sub is NUW.
693 return Constant::getNullValue(Op0->getType());
695 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
696 if (Known.Zero.isMaxSignedValue()) {
697 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
698 // Op1 must be 0 because negating the minimum signed value is undefined.
700 return Constant::getNullValue(Op0->getType());
702 // 0 - X -> X if X is 0 or the minimum signed value.
707 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
708 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
709 Value *X = nullptr, *Y = nullptr, *Z = Op1;
710 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
711 // See if "V === Y - Z" simplifies.
712 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
713 // It does! Now see if "X + V" simplifies.
714 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
715 // It does, we successfully reassociated!
719 // See if "V === X - Z" simplifies.
720 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
721 // It does! Now see if "Y + V" simplifies.
722 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
723 // It does, we successfully reassociated!
729 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
730 // For example, X - (X + 1) -> -1
732 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
733 // See if "V === X - Y" simplifies.
734 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
735 // It does! Now see if "V - Z" simplifies.
736 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
737 // It does, we successfully reassociated!
741 // See if "V === X - Z" simplifies.
742 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
743 // It does! Now see if "V - Y" simplifies.
744 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
745 // It does, we successfully reassociated!
751 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
752 // For example, X - (X - Y) -> Y.
754 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
755 // See if "V === Z - X" simplifies.
756 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
757 // It does! Now see if "V + Y" simplifies.
758 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
759 // It does, we successfully reassociated!
764 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
765 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
766 match(Op1, m_Trunc(m_Value(Y))))
767 if (X->getType() == Y->getType())
768 // See if "V === X - Y" simplifies.
769 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
770 // It does! Now see if "trunc V" simplifies.
771 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
773 // It does, return the simplified "trunc V".
776 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
777 if (match(Op0, m_PtrToInt(m_Value(X))) &&
778 match(Op1, m_PtrToInt(m_Value(Y))))
779 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
780 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
783 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
784 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
787 // Threading Sub over selects and phi nodes is pointless, so don't bother.
788 // Threading over the select in "A - select(cond, B, C)" means evaluating
789 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
790 // only if B and C are equal. If B and C are equal then (since we assume
791 // that operands have already been simplified) "select(cond, B, C)" should
792 // have been simplified to the common value of B and C already. Analysing
793 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
794 // for threading over phi nodes.
799 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
800 const SimplifyQuery &Q) {
801 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
804 /// Given operands for a Mul, see if we can fold the result.
805 /// If not, this returns null.
806 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
807 unsigned MaxRecurse) {
808 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
813 if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
814 return Constant::getNullValue(Op0->getType());
817 if (match(Op1, m_One()))
820 // (X / Y) * Y -> X if the division is exact.
822 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
823 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
827 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
828 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
831 // Try some generic simplifications for associative operations.
832 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
836 // Mul distributes over Add. Try some generic simplifications based on this.
837 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
841 // If the operation is with the result of a select instruction, check whether
842 // operating on either branch of the select always yields the same value.
843 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
844 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
848 // If the operation is with the result of a phi instruction, check whether
849 // operating on all incoming values of the phi always yields the same value.
850 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
851 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
858 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
859 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
862 /// Check for common or similar folds of integer division or integer remainder.
863 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
864 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
865 Type *Ty = Op0->getType();
867 // X / undef -> undef
868 // X % undef -> undef
869 if (match(Op1, m_Undef()))
874 // We don't need to preserve faults!
875 if (match(Op1, m_Zero()))
876 return UndefValue::get(Ty);
878 // If any element of a constant divisor vector is zero or undef, the whole op
880 auto *Op1C = dyn_cast<Constant>(Op1);
881 if (Op1C && Ty->isVectorTy()) {
882 unsigned NumElts = Ty->getVectorNumElements();
883 for (unsigned i = 0; i != NumElts; ++i) {
884 Constant *Elt = Op1C->getAggregateElement(i);
885 if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
886 return UndefValue::get(Ty);
892 if (match(Op0, m_Undef()))
893 return Constant::getNullValue(Ty);
897 if (match(Op0, m_Zero()))
898 return Constant::getNullValue(Op0->getType());
903 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
907 // If this is a boolean op (single-bit element type), we can't have
908 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
909 // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
911 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
912 (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
913 return IsDiv ? Op0 : Constant::getNullValue(Ty);
918 /// Given a predicate and two operands, return true if the comparison is true.
919 /// This is a helper for div/rem simplification where we return some other value
920 /// when we can prove a relationship between the operands.
921 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
922 const SimplifyQuery &Q, unsigned MaxRecurse) {
923 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
924 Constant *C = dyn_cast_or_null<Constant>(V);
925 return (C && C->isAllOnesValue());
928 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
929 /// to simplify X % Y to X.
930 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
931 unsigned MaxRecurse, bool IsSigned) {
932 // Recursion is always used, so bail out at once if we already hit the limit.
939 // We require that 1 operand is a simple constant. That could be extended to
940 // 2 variables if we computed the sign bit for each.
942 // Make sure that a constant is not the minimum signed value because taking
943 // the abs() of that is undefined.
944 Type *Ty = X->getType();
946 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
947 // Is the variable divisor magnitude always greater than the constant
948 // dividend magnitude?
949 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
950 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
951 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
952 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
953 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
956 if (match(Y, m_APInt(C))) {
957 // Special-case: we can't take the abs() of a minimum signed value. If
958 // that's the divisor, then all we have to do is prove that the dividend
959 // is also not the minimum signed value.
960 if (C->isMinSignedValue())
961 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
963 // Is the variable dividend magnitude always less than the constant
964 // divisor magnitude?
965 // |X| < |C| --> X > -abs(C) and X < abs(C)
966 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
967 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
968 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
969 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
975 // IsSigned == false.
976 // Is the dividend unsigned less than the divisor?
977 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
980 /// These are simplifications common to SDiv and UDiv.
981 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
982 const SimplifyQuery &Q, unsigned MaxRecurse) {
983 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
986 if (Value *V = simplifyDivRem(Op0, Op1, true))
989 bool IsSigned = Opcode == Instruction::SDiv;
991 // (X * Y) / Y -> X if the multiplication does not overflow.
993 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
994 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
995 // If the Mul does not overflow, then we are good to go.
996 if ((IsSigned && Mul->hasNoSignedWrap()) ||
997 (!IsSigned && Mul->hasNoUnsignedWrap()))
999 // If X has the form X = A / Y, then X * Y cannot overflow.
1000 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1001 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1005 // (X rem Y) / Y -> 0
1006 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1007 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1008 return Constant::getNullValue(Op0->getType());
1010 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1011 ConstantInt *C1, *C2;
1012 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1013 match(Op1, m_ConstantInt(C2))) {
1015 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1017 return Constant::getNullValue(Op0->getType());
1020 // If the operation is with the result of a select instruction, check whether
1021 // operating on either branch of the select always yields the same value.
1022 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1023 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1026 // If the operation is with the result of a phi instruction, check whether
1027 // operating on all incoming values of the phi always yields the same value.
1028 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1029 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1032 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1033 return Constant::getNullValue(Op0->getType());
1038 /// These are simplifications common to SRem and URem.
1039 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1040 const SimplifyQuery &Q, unsigned MaxRecurse) {
1041 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1044 if (Value *V = simplifyDivRem(Op0, Op1, false))
1047 // (X % Y) % Y -> X % Y
1048 if ((Opcode == Instruction::SRem &&
1049 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1050 (Opcode == Instruction::URem &&
1051 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1054 // (X << Y) % X -> 0
1055 if ((Opcode == Instruction::SRem &&
1056 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1057 (Opcode == Instruction::URem &&
1058 match(Op0, m_NUWShl(m_Specific(Op1), m_Value()))))
1059 return Constant::getNullValue(Op0->getType());
1061 // If the operation is with the result of a select instruction, check whether
1062 // operating on either branch of the select always yields the same value.
1063 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1064 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1067 // If the operation is with the result of a phi instruction, check whether
1068 // operating on all incoming values of the phi always yields the same value.
1069 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1070 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1073 // If X / Y == 0, then X % Y == X.
1074 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1080 /// Given operands for an SDiv, see if we can fold the result.
1081 /// If not, this returns null.
1082 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1083 unsigned MaxRecurse) {
1084 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1087 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1088 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1091 /// Given operands for a UDiv, see if we can fold the result.
1092 /// If not, this returns null.
1093 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1094 unsigned MaxRecurse) {
1095 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1098 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1099 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1102 /// Given operands for an SRem, see if we can fold the result.
1103 /// If not, this returns null.
1104 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1105 unsigned MaxRecurse) {
1106 // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1107 // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1109 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1110 return ConstantInt::getNullValue(Op0->getType());
1112 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1115 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1116 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1119 /// Given operands for a URem, see if we can fold the result.
1120 /// If not, this returns null.
1121 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1122 unsigned MaxRecurse) {
1123 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1126 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1127 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1130 /// Returns true if a shift by \c Amount always yields undef.
1131 static bool isUndefShift(Value *Amount) {
1132 Constant *C = dyn_cast<Constant>(Amount);
1136 // X shift by undef -> undef because it may shift by the bitwidth.
1137 if (isa<UndefValue>(C))
1140 // Shifting by the bitwidth or more is undefined.
1141 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1142 if (CI->getValue().getLimitedValue() >=
1143 CI->getType()->getScalarSizeInBits())
1146 // If all lanes of a vector shift are undefined the whole shift is.
1147 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1148 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1149 if (!isUndefShift(C->getAggregateElement(I)))
1157 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1158 /// If not, this returns null.
1159 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1160 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1161 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1164 // 0 shift by X -> 0
1165 if (match(Op0, m_Zero()))
1166 return Constant::getNullValue(Op0->getType());
1168 // X shift by 0 -> X
1169 // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1172 if (match(Op1, m_Zero()) ||
1173 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1176 // Fold undefined shifts.
1177 if (isUndefShift(Op1))
1178 return UndefValue::get(Op0->getType());
1180 // If the operation is with the result of a select instruction, check whether
1181 // operating on either branch of the select always yields the same value.
1182 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1183 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1186 // If the operation is with the result of a phi instruction, check whether
1187 // operating on all incoming values of the phi always yields the same value.
1188 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1189 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1192 // If any bits in the shift amount make that value greater than or equal to
1193 // the number of bits in the type, the shift is undefined.
1194 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1195 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1196 return UndefValue::get(Op0->getType());
1198 // If all valid bits in the shift amount are known zero, the first operand is
1200 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1201 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1207 /// Given operands for an Shl, LShr or AShr, see if we can
1208 /// fold the result. If not, this returns null.
1209 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1210 Value *Op1, bool isExact, const SimplifyQuery &Q,
1211 unsigned MaxRecurse) {
1212 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1217 return Constant::getNullValue(Op0->getType());
1220 // undef >> X -> undef (if it's exact)
1221 if (match(Op0, m_Undef()))
1222 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1224 // The low bit cannot be shifted out of an exact shift if it is set.
1226 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1227 if (Op0Known.One[0])
1234 /// Given operands for an Shl, see if we can fold the result.
1235 /// If not, this returns null.
1236 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1237 const SimplifyQuery &Q, unsigned MaxRecurse) {
1238 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1242 // undef << X -> undef if (if it's NSW/NUW)
1243 if (match(Op0, m_Undef()))
1244 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1246 // (X >> A) << A -> X
1248 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1251 // shl nuw i8 C, %x -> C iff C has sign bit set.
1252 if (isNUW && match(Op0, m_Negative()))
1254 // NOTE: could use computeKnownBits() / LazyValueInfo,
1255 // but the cost-benefit analysis suggests it isn't worth it.
1260 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1261 const SimplifyQuery &Q) {
1262 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1265 /// Given operands for an LShr, see if we can fold the result.
1266 /// If not, this returns null.
1267 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1268 const SimplifyQuery &Q, unsigned MaxRecurse) {
1269 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1273 // (X << A) >> A -> X
1275 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1281 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1282 const SimplifyQuery &Q) {
1283 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1286 /// Given operands for an AShr, see if we can fold the result.
1287 /// If not, this returns null.
1288 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1289 const SimplifyQuery &Q, unsigned MaxRecurse) {
1290 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1294 // all ones >>a X -> -1
1295 // Do not return Op0 because it may contain undef elements if it's a vector.
1296 if (match(Op0, m_AllOnes()))
1297 return Constant::getAllOnesValue(Op0->getType());
1299 // (X << A) >> A -> X
1301 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1304 // Arithmetic shifting an all-sign-bit value is a no-op.
1305 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1306 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1312 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1313 const SimplifyQuery &Q) {
1314 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1317 /// Commuted variants are assumed to be handled by calling this function again
1318 /// with the parameters swapped.
1319 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1320 ICmpInst *UnsignedICmp, bool IsAnd) {
1323 ICmpInst::Predicate EqPred;
1324 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1325 !ICmpInst::isEquality(EqPred))
1328 ICmpInst::Predicate UnsignedPred;
1329 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1330 ICmpInst::isUnsigned(UnsignedPred))
1332 else if (match(UnsignedICmp,
1333 m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1334 ICmpInst::isUnsigned(UnsignedPred))
1335 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1339 // X < Y && Y != 0 --> X < Y
1340 // X < Y || Y != 0 --> Y != 0
1341 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1342 return IsAnd ? UnsignedICmp : ZeroICmp;
1344 // X >= Y || Y != 0 --> true
1345 // X >= Y || Y == 0 --> X >= Y
1346 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1347 if (EqPred == ICmpInst::ICMP_NE)
1348 return getTrue(UnsignedICmp->getType());
1349 return UnsignedICmp;
1352 // X < Y && Y == 0 --> false
1353 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1355 return getFalse(UnsignedICmp->getType());
1360 /// Commuted variants are assumed to be handled by calling this function again
1361 /// with the parameters swapped.
1362 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1363 ICmpInst::Predicate Pred0, Pred1;
1365 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1366 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1369 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1370 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1371 // can eliminate Op1 from this 'and'.
1372 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1375 // Check for any combination of predicates that are guaranteed to be disjoint.
1376 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1377 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1378 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1379 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1380 return getFalse(Op0->getType());
1385 /// Commuted variants are assumed to be handled by calling this function again
1386 /// with the parameters swapped.
1387 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1388 ICmpInst::Predicate Pred0, Pred1;
1390 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1391 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1394 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1395 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1396 // can eliminate Op0 from this 'or'.
1397 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1400 // Check for any combination of predicates that cover the entire range of
1402 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1403 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1404 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1405 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1406 return getTrue(Op0->getType());
1411 /// Test if a pair of compares with a shared operand and 2 constants has an
1412 /// empty set intersection, full set union, or if one compare is a superset of
1414 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1416 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1417 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1420 const APInt *C0, *C1;
1421 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1422 !match(Cmp1->getOperand(1), m_APInt(C1)))
1425 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1426 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1428 // For and-of-compares, check if the intersection is empty:
1429 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1430 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1431 return getFalse(Cmp0->getType());
1433 // For or-of-compares, check if the union is full:
1434 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1435 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1436 return getTrue(Cmp0->getType());
1438 // Is one range a superset of the other?
1439 // If this is and-of-compares, take the smaller set:
1440 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1441 // If this is or-of-compares, take the larger set:
1442 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1443 if (Range0.contains(Range1))
1444 return IsAnd ? Cmp1 : Cmp0;
1445 if (Range1.contains(Range0))
1446 return IsAnd ? Cmp0 : Cmp1;
1451 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1453 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1454 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1455 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1458 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1461 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1462 Value *X = Cmp0->getOperand(0);
1463 Value *Y = Cmp1->getOperand(0);
1465 // If one of the compares is a masked version of a (not) null check, then
1466 // that compare implies the other, so we eliminate the other. Optionally, look
1467 // through a pointer-to-int cast to match a null check of a pointer type.
1469 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1470 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1471 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1472 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1473 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1474 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1477 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1478 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1479 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1480 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1481 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1482 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1488 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1489 // (icmp (add V, C0), C1) & (icmp V, C0)
1490 ICmpInst::Predicate Pred0, Pred1;
1491 const APInt *C0, *C1;
1493 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1496 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1499 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1500 if (AddInst->getOperand(1) != Op1->getOperand(1))
1503 Type *ITy = Op0->getType();
1504 bool isNSW = AddInst->hasNoSignedWrap();
1505 bool isNUW = AddInst->hasNoUnsignedWrap();
1507 const APInt Delta = *C1 - *C0;
1508 if (C0->isStrictlyPositive()) {
1510 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1511 return getFalse(ITy);
1512 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1513 return getFalse(ITy);
1516 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1517 return getFalse(ITy);
1518 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1519 return getFalse(ITy);
1522 if (C0->getBoolValue() && isNUW) {
1524 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1525 return getFalse(ITy);
1527 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1528 return getFalse(ITy);
1534 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1535 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1537 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1540 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1542 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1545 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1548 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1551 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1553 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1559 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1560 // (icmp (add V, C0), C1) | (icmp V, C0)
1561 ICmpInst::Predicate Pred0, Pred1;
1562 const APInt *C0, *C1;
1564 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1567 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1570 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1571 if (AddInst->getOperand(1) != Op1->getOperand(1))
1574 Type *ITy = Op0->getType();
1575 bool isNSW = AddInst->hasNoSignedWrap();
1576 bool isNUW = AddInst->hasNoUnsignedWrap();
1578 const APInt Delta = *C1 - *C0;
1579 if (C0->isStrictlyPositive()) {
1581 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1582 return getTrue(ITy);
1583 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1584 return getTrue(ITy);
1587 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1588 return getTrue(ITy);
1589 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1590 return getTrue(ITy);
1593 if (C0->getBoolValue() && isNUW) {
1595 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1596 return getTrue(ITy);
1598 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1599 return getTrue(ITy);
1605 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1606 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1608 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1611 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1613 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1616 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1619 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1622 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1624 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1630 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1631 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1632 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1633 if (LHS0->getType() != RHS0->getType())
1636 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1637 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1638 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1639 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1640 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1641 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1642 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1643 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1644 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1645 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1646 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1647 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1648 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1651 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1652 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1653 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1654 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1655 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1656 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1657 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1658 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1659 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1660 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1667 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1668 // Look through casts of the 'and' operands to find compares.
1669 auto *Cast0 = dyn_cast<CastInst>(Op0);
1670 auto *Cast1 = dyn_cast<CastInst>(Op1);
1671 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1672 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1673 Op0 = Cast0->getOperand(0);
1674 Op1 = Cast1->getOperand(0);
1678 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1679 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1681 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1682 simplifyOrOfICmps(ICmp0, ICmp1);
1684 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1685 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1687 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1694 // If we looked through casts, we can only handle a constant simplification
1695 // because we are not allowed to create a cast instruction here.
1696 if (auto *C = dyn_cast<Constant>(V))
1697 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1702 /// Given operands for an And, see if we can fold the result.
1703 /// If not, this returns null.
1704 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1705 unsigned MaxRecurse) {
1706 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1710 if (match(Op1, m_Undef()))
1711 return Constant::getNullValue(Op0->getType());
1718 if (match(Op1, m_Zero()))
1719 return Constant::getNullValue(Op0->getType());
1722 if (match(Op1, m_AllOnes()))
1725 // A & ~A = ~A & A = 0
1726 if (match(Op0, m_Not(m_Specific(Op1))) ||
1727 match(Op1, m_Not(m_Specific(Op0))))
1728 return Constant::getNullValue(Op0->getType());
1731 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1735 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1738 // A mask that only clears known zeros of a shifted value is a no-op.
1742 if (match(Op1, m_APInt(Mask))) {
1743 // If all bits in the inverted and shifted mask are clear:
1744 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1745 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1746 (~(*Mask)).lshr(*ShAmt).isNullValue())
1749 // If all bits in the inverted and shifted mask are clear:
1750 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1751 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1752 (~(*Mask)).shl(*ShAmt).isNullValue())
1756 // A & (-A) = A if A is a power of two or zero.
1757 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1758 match(Op1, m_Neg(m_Specific(Op0)))) {
1759 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1762 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1767 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1770 // Try some generic simplifications for associative operations.
1771 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1775 // And distributes over Or. Try some generic simplifications based on this.
1776 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1780 // And distributes over Xor. Try some generic simplifications based on this.
1781 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1785 // If the operation is with the result of a select instruction, check whether
1786 // operating on either branch of the select always yields the same value.
1787 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1788 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1792 // If the operation is with the result of a phi instruction, check whether
1793 // operating on all incoming values of the phi always yields the same value.
1794 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1795 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1802 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1803 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1806 /// Given operands for an Or, see if we can fold the result.
1807 /// If not, this returns null.
1808 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1809 unsigned MaxRecurse) {
1810 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1815 // Do not return Op1 because it may contain undef elements if it's a vector.
1816 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1817 return Constant::getAllOnesValue(Op0->getType());
1821 if (Op0 == Op1 || match(Op1, m_Zero()))
1824 // A | ~A = ~A | A = -1
1825 if (match(Op0, m_Not(m_Specific(Op1))) ||
1826 match(Op1, m_Not(m_Specific(Op0))))
1827 return Constant::getAllOnesValue(Op0->getType());
1830 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1834 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1837 // ~(A & ?) | A = -1
1838 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1839 return Constant::getAllOnesValue(Op1->getType());
1841 // A | ~(A & ?) = -1
1842 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1843 return Constant::getAllOnesValue(Op0->getType());
1846 // (A & ~B) | (A ^ B) -> (A ^ B)
1847 // (~B & A) | (A ^ B) -> (A ^ B)
1848 // (A & ~B) | (B ^ A) -> (B ^ A)
1849 // (~B & A) | (B ^ A) -> (B ^ A)
1850 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1851 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1852 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1855 // Commute the 'or' operands.
1856 // (A ^ B) | (A & ~B) -> (A ^ B)
1857 // (A ^ B) | (~B & A) -> (A ^ B)
1858 // (B ^ A) | (A & ~B) -> (B ^ A)
1859 // (B ^ A) | (~B & A) -> (B ^ A)
1860 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1861 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1862 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1865 // (A & B) | (~A ^ B) -> (~A ^ B)
1866 // (B & A) | (~A ^ B) -> (~A ^ B)
1867 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1868 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1869 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1870 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1871 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1874 // (~A ^ B) | (A & B) -> (~A ^ B)
1875 // (~A ^ B) | (B & A) -> (~A ^ B)
1876 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1877 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1878 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1879 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1880 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1883 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1886 // Try some generic simplifications for associative operations.
1887 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1891 // Or distributes over And. Try some generic simplifications based on this.
1892 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1896 // If the operation is with the result of a select instruction, check whether
1897 // operating on either branch of the select always yields the same value.
1898 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1899 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1903 // (A & C1)|(B & C2)
1904 const APInt *C1, *C2;
1905 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1906 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1908 // (A & C1)|(B & C2)
1909 // If we have: ((V + N) & C1) | (V & C2)
1910 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1911 // replace with V+N.
1913 if (C2->isMask() && // C2 == 0+1+
1914 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1915 // Add commutes, try both ways.
1916 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1919 // Or commutes, try both ways.
1921 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1922 // Add commutes, try both ways.
1923 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1929 // If the operation is with the result of a phi instruction, check whether
1930 // operating on all incoming values of the phi always yields the same value.
1931 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1932 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1938 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1939 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1942 /// Given operands for a Xor, see if we can fold the result.
1943 /// If not, this returns null.
1944 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1945 unsigned MaxRecurse) {
1946 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1949 // A ^ undef -> undef
1950 if (match(Op1, m_Undef()))
1954 if (match(Op1, m_Zero()))
1959 return Constant::getNullValue(Op0->getType());
1961 // A ^ ~A = ~A ^ A = -1
1962 if (match(Op0, m_Not(m_Specific(Op1))) ||
1963 match(Op1, m_Not(m_Specific(Op0))))
1964 return Constant::getAllOnesValue(Op0->getType());
1966 // Try some generic simplifications for associative operations.
1967 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1971 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1972 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1973 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1974 // only if B and C are equal. If B and C are equal then (since we assume
1975 // that operands have already been simplified) "select(cond, B, C)" should
1976 // have been simplified to the common value of B and C already. Analysing
1977 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1978 // for threading over phi nodes.
1983 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1984 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1988 static Type *GetCompareTy(Value *Op) {
1989 return CmpInst::makeCmpResultType(Op->getType());
1992 /// Rummage around inside V looking for something equivalent to the comparison
1993 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1994 /// Helper function for analyzing max/min idioms.
1995 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1996 Value *LHS, Value *RHS) {
1997 SelectInst *SI = dyn_cast<SelectInst>(V);
2000 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2003 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2004 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2006 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2007 LHS == CmpRHS && RHS == CmpLHS)
2012 // A significant optimization not implemented here is assuming that alloca
2013 // addresses are not equal to incoming argument values. They don't *alias*,
2014 // as we say, but that doesn't mean they aren't equal, so we take a
2015 // conservative approach.
2017 // This is inspired in part by C++11 5.10p1:
2018 // "Two pointers of the same type compare equal if and only if they are both
2019 // null, both point to the same function, or both represent the same
2022 // This is pretty permissive.
2024 // It's also partly due to C11 6.5.9p6:
2025 // "Two pointers compare equal if and only if both are null pointers, both are
2026 // pointers to the same object (including a pointer to an object and a
2027 // subobject at its beginning) or function, both are pointers to one past the
2028 // last element of the same array object, or one is a pointer to one past the
2029 // end of one array object and the other is a pointer to the start of a
2030 // different array object that happens to immediately follow the first array
2031 // object in the address space.)
2033 // C11's version is more restrictive, however there's no reason why an argument
2034 // couldn't be a one-past-the-end value for a stack object in the caller and be
2035 // equal to the beginning of a stack object in the callee.
2037 // If the C and C++ standards are ever made sufficiently restrictive in this
2038 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2039 // this optimization.
2041 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2042 const DominatorTree *DT, CmpInst::Predicate Pred,
2043 AssumptionCache *AC, const Instruction *CxtI,
2044 Value *LHS, Value *RHS) {
2045 // First, skip past any trivial no-ops.
2046 LHS = LHS->stripPointerCasts();
2047 RHS = RHS->stripPointerCasts();
2049 // A non-null pointer is not equal to a null pointer.
2050 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2051 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2052 return ConstantInt::get(GetCompareTy(LHS),
2053 !CmpInst::isTrueWhenEqual(Pred));
2055 // We can only fold certain predicates on pointer comparisons.
2060 // Equality comaprisons are easy to fold.
2061 case CmpInst::ICMP_EQ:
2062 case CmpInst::ICMP_NE:
2065 // We can only handle unsigned relational comparisons because 'inbounds' on
2066 // a GEP only protects against unsigned wrapping.
2067 case CmpInst::ICMP_UGT:
2068 case CmpInst::ICMP_UGE:
2069 case CmpInst::ICMP_ULT:
2070 case CmpInst::ICMP_ULE:
2071 // However, we have to switch them to their signed variants to handle
2072 // negative indices from the base pointer.
2073 Pred = ICmpInst::getSignedPredicate(Pred);
2077 // Strip off any constant offsets so that we can reason about them.
2078 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2079 // here and compare base addresses like AliasAnalysis does, however there are
2080 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2081 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2082 // doesn't need to guarantee pointer inequality when it says NoAlias.
2083 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2084 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2086 // If LHS and RHS are related via constant offsets to the same base
2087 // value, we can replace it with an icmp which just compares the offsets.
2089 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2091 // Various optimizations for (in)equality comparisons.
2092 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2093 // Different non-empty allocations that exist at the same time have
2094 // different addresses (if the program can tell). Global variables always
2095 // exist, so they always exist during the lifetime of each other and all
2096 // allocas. Two different allocas usually have different addresses...
2098 // However, if there's an @llvm.stackrestore dynamically in between two
2099 // allocas, they may have the same address. It's tempting to reduce the
2100 // scope of the problem by only looking at *static* allocas here. That would
2101 // cover the majority of allocas while significantly reducing the likelihood
2102 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2103 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2104 // an entry block. Also, if we have a block that's not attached to a
2105 // function, we can't tell if it's "static" under the current definition.
2106 // Theoretically, this problem could be fixed by creating a new kind of
2107 // instruction kind specifically for static allocas. Such a new instruction
2108 // could be required to be at the top of the entry block, thus preventing it
2109 // from being subject to a @llvm.stackrestore. Instcombine could even
2110 // convert regular allocas into these special allocas. It'd be nifty.
2111 // However, until then, this problem remains open.
2113 // So, we'll assume that two non-empty allocas have different addresses
2116 // With all that, if the offsets are within the bounds of their allocations
2117 // (and not one-past-the-end! so we can't use inbounds!), and their
2118 // allocations aren't the same, the pointers are not equal.
2120 // Note that it's not necessary to check for LHS being a global variable
2121 // address, due to canonicalization and constant folding.
2122 if (isa<AllocaInst>(LHS) &&
2123 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2124 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2125 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2126 uint64_t LHSSize, RHSSize;
2127 ObjectSizeOpts Opts;
2128 Opts.NullIsUnknownSize =
2129 NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2130 if (LHSOffsetCI && RHSOffsetCI &&
2131 getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2132 getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2133 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2134 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2135 if (!LHSOffsetValue.isNegative() &&
2136 !RHSOffsetValue.isNegative() &&
2137 LHSOffsetValue.ult(LHSSize) &&
2138 RHSOffsetValue.ult(RHSSize)) {
2139 return ConstantInt::get(GetCompareTy(LHS),
2140 !CmpInst::isTrueWhenEqual(Pred));
2144 // Repeat the above check but this time without depending on DataLayout
2145 // or being able to compute a precise size.
2146 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2147 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2148 LHSOffset->isNullValue() &&
2149 RHSOffset->isNullValue())
2150 return ConstantInt::get(GetCompareTy(LHS),
2151 !CmpInst::isTrueWhenEqual(Pred));
2154 // Even if an non-inbounds GEP occurs along the path we can still optimize
2155 // equality comparisons concerning the result. We avoid walking the whole
2156 // chain again by starting where the last calls to
2157 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2158 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2159 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2161 return ConstantExpr::getICmp(Pred,
2162 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2163 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2165 // If one side of the equality comparison must come from a noalias call
2166 // (meaning a system memory allocation function), and the other side must
2167 // come from a pointer that cannot overlap with dynamically-allocated
2168 // memory within the lifetime of the current function (allocas, byval
2169 // arguments, globals), then determine the comparison result here.
2170 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2171 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2172 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2174 // Is the set of underlying objects all noalias calls?
2175 auto IsNAC = [](ArrayRef<Value *> Objects) {
2176 return all_of(Objects, isNoAliasCall);
2179 // Is the set of underlying objects all things which must be disjoint from
2180 // noalias calls. For allocas, we consider only static ones (dynamic
2181 // allocas might be transformed into calls to malloc not simultaneously
2182 // live with the compared-to allocation). For globals, we exclude symbols
2183 // that might be resolve lazily to symbols in another dynamically-loaded
2184 // library (and, thus, could be malloc'ed by the implementation).
2185 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2186 return all_of(Objects, [](Value *V) {
2187 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2188 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2189 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2190 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2191 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2192 !GV->isThreadLocal();
2193 if (const Argument *A = dyn_cast<Argument>(V))
2194 return A->hasByValAttr();
2199 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2200 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2201 return ConstantInt::get(GetCompareTy(LHS),
2202 !CmpInst::isTrueWhenEqual(Pred));
2204 // Fold comparisons for non-escaping pointer even if the allocation call
2205 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2206 // dynamic allocation call could be either of the operands.
2207 Value *MI = nullptr;
2208 if (isAllocLikeFn(LHS, TLI) &&
2209 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2211 else if (isAllocLikeFn(RHS, TLI) &&
2212 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2214 // FIXME: We should also fold the compare when the pointer escapes, but the
2215 // compare dominates the pointer escape
2216 if (MI && !PointerMayBeCaptured(MI, true, true))
2217 return ConstantInt::get(GetCompareTy(LHS),
2218 CmpInst::isFalseWhenEqual(Pred));
2225 /// Fold an icmp when its operands have i1 scalar type.
2226 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2227 Value *RHS, const SimplifyQuery &Q) {
2228 Type *ITy = GetCompareTy(LHS); // The return type.
2229 Type *OpTy = LHS->getType(); // The operand type.
2230 if (!OpTy->isIntOrIntVectorTy(1))
2233 // A boolean compared to true/false can be simplified in 14 out of the 20
2234 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2235 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2236 if (match(RHS, m_Zero())) {
2238 case CmpInst::ICMP_NE: // X != 0 -> X
2239 case CmpInst::ICMP_UGT: // X >u 0 -> X
2240 case CmpInst::ICMP_SLT: // X <s 0 -> X
2243 case CmpInst::ICMP_ULT: // X <u 0 -> false
2244 case CmpInst::ICMP_SGT: // X >s 0 -> false
2245 return getFalse(ITy);
2247 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2248 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2249 return getTrue(ITy);
2253 } else if (match(RHS, m_One())) {
2255 case CmpInst::ICMP_EQ: // X == 1 -> X
2256 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2257 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2260 case CmpInst::ICMP_UGT: // X >u 1 -> false
2261 case CmpInst::ICMP_SLT: // X <s -1 -> false
2262 return getFalse(ITy);
2264 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2265 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2266 return getTrue(ITy);
2275 case ICmpInst::ICMP_UGE:
2276 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2277 return getTrue(ITy);
2279 case ICmpInst::ICMP_SGE:
2280 /// For signed comparison, the values for an i1 are 0 and -1
2281 /// respectively. This maps into a truth table of:
2282 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2283 /// 0 | 0 | 1 (0 >= 0) | 1
2284 /// 0 | 1 | 1 (0 >= -1) | 1
2285 /// 1 | 0 | 0 (-1 >= 0) | 0
2286 /// 1 | 1 | 1 (-1 >= -1) | 1
2287 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2288 return getTrue(ITy);
2290 case ICmpInst::ICMP_ULE:
2291 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2292 return getTrue(ITy);
2299 /// Try hard to fold icmp with zero RHS because this is a common case.
2300 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2301 Value *RHS, const SimplifyQuery &Q) {
2302 if (!match(RHS, m_Zero()))
2305 Type *ITy = GetCompareTy(LHS); // The return type.
2308 llvm_unreachable("Unknown ICmp predicate!");
2309 case ICmpInst::ICMP_ULT:
2310 return getFalse(ITy);
2311 case ICmpInst::ICMP_UGE:
2312 return getTrue(ITy);
2313 case ICmpInst::ICMP_EQ:
2314 case ICmpInst::ICMP_ULE:
2315 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2316 return getFalse(ITy);
2318 case ICmpInst::ICMP_NE:
2319 case ICmpInst::ICMP_UGT:
2320 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2321 return getTrue(ITy);
2323 case ICmpInst::ICMP_SLT: {
2324 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2325 if (LHSKnown.isNegative())
2326 return getTrue(ITy);
2327 if (LHSKnown.isNonNegative())
2328 return getFalse(ITy);
2331 case ICmpInst::ICMP_SLE: {
2332 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2333 if (LHSKnown.isNegative())
2334 return getTrue(ITy);
2335 if (LHSKnown.isNonNegative() &&
2336 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2337 return getFalse(ITy);
2340 case ICmpInst::ICMP_SGE: {
2341 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2342 if (LHSKnown.isNegative())
2343 return getFalse(ITy);
2344 if (LHSKnown.isNonNegative())
2345 return getTrue(ITy);
2348 case ICmpInst::ICMP_SGT: {
2349 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2350 if (LHSKnown.isNegative())
2351 return getFalse(ITy);
2352 if (LHSKnown.isNonNegative() &&
2353 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2354 return getTrue(ITy);
2362 /// Many binary operators with a constant operand have an easy-to-compute
2363 /// range of outputs. This can be used to fold a comparison to always true or
2365 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2366 unsigned Width = Lower.getBitWidth();
2368 switch (BO.getOpcode()) {
2369 case Instruction::Add:
2370 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2371 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2372 if (BO.hasNoUnsignedWrap()) {
2373 // 'add nuw x, C' produces [C, UINT_MAX].
2375 } else if (BO.hasNoSignedWrap()) {
2376 if (C->isNegative()) {
2377 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2378 Lower = APInt::getSignedMinValue(Width);
2379 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2381 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2382 Lower = APInt::getSignedMinValue(Width) + *C;
2383 Upper = APInt::getSignedMaxValue(Width) + 1;
2389 case Instruction::And:
2390 if (match(BO.getOperand(1), m_APInt(C)))
2391 // 'and x, C' produces [0, C].
2395 case Instruction::Or:
2396 if (match(BO.getOperand(1), m_APInt(C)))
2397 // 'or x, C' produces [C, UINT_MAX].
2401 case Instruction::AShr:
2402 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2403 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2404 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2405 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2406 } else if (match(BO.getOperand(0), m_APInt(C))) {
2407 unsigned ShiftAmount = Width - 1;
2408 if (!C->isNullValue() && BO.isExact())
2409 ShiftAmount = C->countTrailingZeros();
2410 if (C->isNegative()) {
2411 // 'ashr C, x' produces [C, C >> (Width-1)]
2413 Upper = C->ashr(ShiftAmount) + 1;
2415 // 'ashr C, x' produces [C >> (Width-1), C]
2416 Lower = C->ashr(ShiftAmount);
2422 case Instruction::LShr:
2423 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2424 // 'lshr x, C' produces [0, UINT_MAX >> C].
2425 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2426 } else if (match(BO.getOperand(0), m_APInt(C))) {
2427 // 'lshr C, x' produces [C >> (Width-1), C].
2428 unsigned ShiftAmount = Width - 1;
2429 if (!C->isNullValue() && BO.isExact())
2430 ShiftAmount = C->countTrailingZeros();
2431 Lower = C->lshr(ShiftAmount);
2436 case Instruction::Shl:
2437 if (match(BO.getOperand(0), m_APInt(C))) {
2438 if (BO.hasNoUnsignedWrap()) {
2439 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2441 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2442 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2443 if (C->isNegative()) {
2444 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2445 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2446 Lower = C->shl(ShiftAmount);
2449 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2450 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2452 Upper = C->shl(ShiftAmount) + 1;
2458 case Instruction::SDiv:
2459 if (match(BO.getOperand(1), m_APInt(C))) {
2460 APInt IntMin = APInt::getSignedMinValue(Width);
2461 APInt IntMax = APInt::getSignedMaxValue(Width);
2462 if (C->isAllOnesValue()) {
2463 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2464 // where C != -1 and C != 0 and C != 1
2467 } else if (C->countLeadingZeros() < Width - 1) {
2468 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2469 // where C != -1 and C != 0 and C != 1
2470 Lower = IntMin.sdiv(*C);
2471 Upper = IntMax.sdiv(*C);
2472 if (Lower.sgt(Upper))
2473 std::swap(Lower, Upper);
2475 assert(Upper != Lower && "Upper part of range has wrapped!");
2477 } else if (match(BO.getOperand(0), m_APInt(C))) {
2478 if (C->isMinSignedValue()) {
2479 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2481 Upper = Lower.lshr(1) + 1;
2483 // 'sdiv C, x' produces [-|C|, |C|].
2484 Upper = C->abs() + 1;
2485 Lower = (-Upper) + 1;
2490 case Instruction::UDiv:
2491 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2492 // 'udiv x, C' produces [0, UINT_MAX / C].
2493 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2494 } else if (match(BO.getOperand(0), m_APInt(C))) {
2495 // 'udiv C, x' produces [0, C].
2500 case Instruction::SRem:
2501 if (match(BO.getOperand(1), m_APInt(C))) {
2502 // 'srem x, C' produces (-|C|, |C|).
2504 Lower = (-Upper) + 1;
2508 case Instruction::URem:
2509 if (match(BO.getOperand(1), m_APInt(C)))
2510 // 'urem x, C' produces [0, C).
2519 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2521 Type *ITy = GetCompareTy(RHS); // The return type.
2524 // Sign-bit checks can be optimized to true/false after unsigned
2525 // floating-point casts:
2526 // icmp slt (bitcast (uitofp X)), 0 --> false
2527 // icmp sgt (bitcast (uitofp X)), -1 --> true
2528 if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2529 if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2530 return ConstantInt::getFalse(ITy);
2531 if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2532 return ConstantInt::getTrue(ITy);
2536 if (!match(RHS, m_APInt(C)))
2539 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2540 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2541 if (RHS_CR.isEmptySet())
2542 return ConstantInt::getFalse(ITy);
2543 if (RHS_CR.isFullSet())
2544 return ConstantInt::getTrue(ITy);
2546 // Find the range of possible values for binary operators.
2547 unsigned Width = C->getBitWidth();
2548 APInt Lower = APInt(Width, 0);
2549 APInt Upper = APInt(Width, 0);
2550 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2551 setLimitsForBinOp(*BO, Lower, Upper);
2553 ConstantRange LHS_CR =
2554 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2556 if (auto *I = dyn_cast<Instruction>(LHS))
2557 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2558 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2560 if (!LHS_CR.isFullSet()) {
2561 if (RHS_CR.contains(LHS_CR))
2562 return ConstantInt::getTrue(ITy);
2563 if (RHS_CR.inverse().contains(LHS_CR))
2564 return ConstantInt::getFalse(ITy);
2570 /// TODO: A large part of this logic is duplicated in InstCombine's
2571 /// foldICmpBinOp(). We should be able to share that and avoid the code
2573 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2574 Value *RHS, const SimplifyQuery &Q,
2575 unsigned MaxRecurse) {
2576 Type *ITy = GetCompareTy(LHS); // The return type.
2578 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2579 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2580 if (MaxRecurse && (LBO || RBO)) {
2581 // Analyze the case when either LHS or RHS is an add instruction.
2582 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2583 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2584 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2585 if (LBO && LBO->getOpcode() == Instruction::Add) {
2586 A = LBO->getOperand(0);
2587 B = LBO->getOperand(1);
2589 ICmpInst::isEquality(Pred) ||
2590 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2591 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2593 if (RBO && RBO->getOpcode() == Instruction::Add) {
2594 C = RBO->getOperand(0);
2595 D = RBO->getOperand(1);
2597 ICmpInst::isEquality(Pred) ||
2598 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2599 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2602 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2603 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2604 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2605 Constant::getNullValue(RHS->getType()), Q,
2609 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2610 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2612 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2613 C == LHS ? D : C, Q, MaxRecurse - 1))
2616 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2617 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2619 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2622 // C + B == C + D -> B == D
2625 } else if (A == D) {
2626 // D + B == C + D -> B == C
2629 } else if (B == C) {
2630 // A + C == C + D -> A == D
2635 // A + D == C + D -> A == C
2639 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2646 // icmp pred (or X, Y), X
2647 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2648 if (Pred == ICmpInst::ICMP_ULT)
2649 return getFalse(ITy);
2650 if (Pred == ICmpInst::ICMP_UGE)
2651 return getTrue(ITy);
2653 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2654 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2655 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2656 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2657 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2658 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2659 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2662 // icmp pred X, (or X, Y)
2663 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2664 if (Pred == ICmpInst::ICMP_ULE)
2665 return getTrue(ITy);
2666 if (Pred == ICmpInst::ICMP_UGT)
2667 return getFalse(ITy);
2669 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2670 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2671 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2672 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2673 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2674 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2675 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2680 // icmp pred (and X, Y), X
2681 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2682 if (Pred == ICmpInst::ICMP_UGT)
2683 return getFalse(ITy);
2684 if (Pred == ICmpInst::ICMP_ULE)
2685 return getTrue(ITy);
2687 // icmp pred X, (and X, Y)
2688 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2689 if (Pred == ICmpInst::ICMP_UGE)
2690 return getTrue(ITy);
2691 if (Pred == ICmpInst::ICMP_ULT)
2692 return getFalse(ITy);
2695 // 0 - (zext X) pred C
2696 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2697 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2698 if (RHSC->getValue().isStrictlyPositive()) {
2699 if (Pred == ICmpInst::ICMP_SLT)
2700 return ConstantInt::getTrue(RHSC->getContext());
2701 if (Pred == ICmpInst::ICMP_SGE)
2702 return ConstantInt::getFalse(RHSC->getContext());
2703 if (Pred == ICmpInst::ICMP_EQ)
2704 return ConstantInt::getFalse(RHSC->getContext());
2705 if (Pred == ICmpInst::ICMP_NE)
2706 return ConstantInt::getTrue(RHSC->getContext());
2708 if (RHSC->getValue().isNonNegative()) {
2709 if (Pred == ICmpInst::ICMP_SLE)
2710 return ConstantInt::getTrue(RHSC->getContext());
2711 if (Pred == ICmpInst::ICMP_SGT)
2712 return ConstantInt::getFalse(RHSC->getContext());
2717 // icmp pred (urem X, Y), Y
2718 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2722 case ICmpInst::ICMP_SGT:
2723 case ICmpInst::ICMP_SGE: {
2724 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2725 if (!Known.isNonNegative())
2729 case ICmpInst::ICMP_EQ:
2730 case ICmpInst::ICMP_UGT:
2731 case ICmpInst::ICMP_UGE:
2732 return getFalse(ITy);
2733 case ICmpInst::ICMP_SLT:
2734 case ICmpInst::ICMP_SLE: {
2735 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2736 if (!Known.isNonNegative())
2740 case ICmpInst::ICMP_NE:
2741 case ICmpInst::ICMP_ULT:
2742 case ICmpInst::ICMP_ULE:
2743 return getTrue(ITy);
2747 // icmp pred X, (urem Y, X)
2748 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2752 case ICmpInst::ICMP_SGT:
2753 case ICmpInst::ICMP_SGE: {
2754 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2755 if (!Known.isNonNegative())
2759 case ICmpInst::ICMP_NE:
2760 case ICmpInst::ICMP_UGT:
2761 case ICmpInst::ICMP_UGE:
2762 return getTrue(ITy);
2763 case ICmpInst::ICMP_SLT:
2764 case ICmpInst::ICMP_SLE: {
2765 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2766 if (!Known.isNonNegative())
2770 case ICmpInst::ICMP_EQ:
2771 case ICmpInst::ICMP_ULT:
2772 case ICmpInst::ICMP_ULE:
2773 return getFalse(ITy);
2779 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2780 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2781 // icmp pred (X op Y), X
2782 if (Pred == ICmpInst::ICMP_UGT)
2783 return getFalse(ITy);
2784 if (Pred == ICmpInst::ICMP_ULE)
2785 return getTrue(ITy);
2790 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2791 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2792 // icmp pred X, (X op Y)
2793 if (Pred == ICmpInst::ICMP_ULT)
2794 return getFalse(ITy);
2795 if (Pred == ICmpInst::ICMP_UGE)
2796 return getTrue(ITy);
2803 // where CI2 is a power of 2 and CI isn't
2804 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2805 const APInt *CI2Val, *CIVal = &CI->getValue();
2806 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2807 CI2Val->isPowerOf2()) {
2808 if (!CIVal->isPowerOf2()) {
2809 // CI2 << X can equal zero in some circumstances,
2810 // this simplification is unsafe if CI is zero.
2812 // We know it is safe if:
2813 // - The shift is nsw, we can't shift out the one bit.
2814 // - The shift is nuw, we can't shift out the one bit.
2817 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2818 CI2Val->isOneValue() || !CI->isZero()) {
2819 if (Pred == ICmpInst::ICMP_EQ)
2820 return ConstantInt::getFalse(RHS->getContext());
2821 if (Pred == ICmpInst::ICMP_NE)
2822 return ConstantInt::getTrue(RHS->getContext());
2825 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2826 if (Pred == ICmpInst::ICMP_UGT)
2827 return ConstantInt::getFalse(RHS->getContext());
2828 if (Pred == ICmpInst::ICMP_ULE)
2829 return ConstantInt::getTrue(RHS->getContext());
2834 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2835 LBO->getOperand(1) == RBO->getOperand(1)) {
2836 switch (LBO->getOpcode()) {
2839 case Instruction::UDiv:
2840 case Instruction::LShr:
2841 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2843 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2844 RBO->getOperand(0), Q, MaxRecurse - 1))
2847 case Instruction::SDiv:
2848 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2850 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2851 RBO->getOperand(0), Q, MaxRecurse - 1))
2854 case Instruction::AShr:
2855 if (!LBO->isExact() || !RBO->isExact())
2857 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2858 RBO->getOperand(0), Q, MaxRecurse - 1))
2861 case Instruction::Shl: {
2862 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2863 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2866 if (!NSW && ICmpInst::isSigned(Pred))
2868 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2869 RBO->getOperand(0), Q, MaxRecurse - 1))
2878 /// Simplify integer comparisons where at least one operand of the compare
2879 /// matches an integer min/max idiom.
2880 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2881 Value *RHS, const SimplifyQuery &Q,
2882 unsigned MaxRecurse) {
2883 Type *ITy = GetCompareTy(LHS); // The return type.
2885 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2886 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2888 // Signed variants on "max(a,b)>=a -> true".
2889 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2891 std::swap(A, B); // smax(A, B) pred A.
2892 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2893 // We analyze this as smax(A, B) pred A.
2895 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2896 (A == LHS || B == LHS)) {
2898 std::swap(A, B); // A pred smax(A, B).
2899 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2900 // We analyze this as smax(A, B) swapped-pred A.
2901 P = CmpInst::getSwappedPredicate(Pred);
2902 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2903 (A == RHS || B == RHS)) {
2905 std::swap(A, B); // smin(A, B) pred A.
2906 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2907 // We analyze this as smax(-A, -B) swapped-pred -A.
2908 // Note that we do not need to actually form -A or -B thanks to EqP.
2909 P = CmpInst::getSwappedPredicate(Pred);
2910 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2911 (A == LHS || B == LHS)) {
2913 std::swap(A, B); // A pred smin(A, B).
2914 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2915 // We analyze this as smax(-A, -B) pred -A.
2916 // Note that we do not need to actually form -A or -B thanks to EqP.
2919 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2920 // Cases correspond to "max(A, B) p A".
2924 case CmpInst::ICMP_EQ:
2925 case CmpInst::ICMP_SLE:
2926 // Equivalent to "A EqP B". This may be the same as the condition tested
2927 // in the max/min; if so, we can just return that.
2928 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2930 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2932 // Otherwise, see if "A EqP B" simplifies.
2934 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2937 case CmpInst::ICMP_NE:
2938 case CmpInst::ICMP_SGT: {
2939 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2940 // Equivalent to "A InvEqP B". This may be the same as the condition
2941 // tested in the max/min; if so, we can just return that.
2942 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2944 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2946 // Otherwise, see if "A InvEqP B" simplifies.
2948 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2952 case CmpInst::ICMP_SGE:
2954 return getTrue(ITy);
2955 case CmpInst::ICMP_SLT:
2957 return getFalse(ITy);
2961 // Unsigned variants on "max(a,b)>=a -> true".
2962 P = CmpInst::BAD_ICMP_PREDICATE;
2963 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2965 std::swap(A, B); // umax(A, B) pred A.
2966 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2967 // We analyze this as umax(A, B) pred A.
2969 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2970 (A == LHS || B == LHS)) {
2972 std::swap(A, B); // A pred umax(A, B).
2973 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2974 // We analyze this as umax(A, B) swapped-pred A.
2975 P = CmpInst::getSwappedPredicate(Pred);
2976 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2977 (A == RHS || B == RHS)) {
2979 std::swap(A, B); // umin(A, B) pred A.
2980 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2981 // We analyze this as umax(-A, -B) swapped-pred -A.
2982 // Note that we do not need to actually form -A or -B thanks to EqP.
2983 P = CmpInst::getSwappedPredicate(Pred);
2984 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2985 (A == LHS || B == LHS)) {
2987 std::swap(A, B); // A pred umin(A, B).
2988 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2989 // We analyze this as umax(-A, -B) pred -A.
2990 // Note that we do not need to actually form -A or -B thanks to EqP.
2993 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2994 // Cases correspond to "max(A, B) p A".
2998 case CmpInst::ICMP_EQ:
2999 case CmpInst::ICMP_ULE:
3000 // Equivalent to "A EqP B". This may be the same as the condition tested
3001 // in the max/min; if so, we can just return that.
3002 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3004 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3006 // Otherwise, see if "A EqP B" simplifies.
3008 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3011 case CmpInst::ICMP_NE:
3012 case CmpInst::ICMP_UGT: {
3013 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3014 // Equivalent to "A InvEqP B". This may be the same as the condition
3015 // tested in the max/min; if so, we can just return that.
3016 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3018 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3020 // Otherwise, see if "A InvEqP B" simplifies.
3022 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3026 case CmpInst::ICMP_UGE:
3028 return getTrue(ITy);
3029 case CmpInst::ICMP_ULT:
3031 return getFalse(ITy);
3035 // Variants on "max(x,y) >= min(x,z)".
3037 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3038 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3039 (A == C || A == D || B == C || B == D)) {
3040 // max(x, ?) pred min(x, ?).
3041 if (Pred == CmpInst::ICMP_SGE)
3043 return getTrue(ITy);
3044 if (Pred == CmpInst::ICMP_SLT)
3046 return getFalse(ITy);
3047 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3048 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3049 (A == C || A == D || B == C || B == D)) {
3050 // min(x, ?) pred max(x, ?).
3051 if (Pred == CmpInst::ICMP_SLE)
3053 return getTrue(ITy);
3054 if (Pred == CmpInst::ICMP_SGT)
3056 return getFalse(ITy);
3057 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3058 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3059 (A == C || A == D || B == C || B == D)) {
3060 // max(x, ?) pred min(x, ?).
3061 if (Pred == CmpInst::ICMP_UGE)
3063 return getTrue(ITy);
3064 if (Pred == CmpInst::ICMP_ULT)
3066 return getFalse(ITy);
3067 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3068 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3069 (A == C || A == D || B == C || B == D)) {
3070 // min(x, ?) pred max(x, ?).
3071 if (Pred == CmpInst::ICMP_ULE)
3073 return getTrue(ITy);
3074 if (Pred == CmpInst::ICMP_UGT)
3076 return getFalse(ITy);
3082 /// Given operands for an ICmpInst, see if we can fold the result.
3083 /// If not, this returns null.
3084 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3085 const SimplifyQuery &Q, unsigned MaxRecurse) {
3086 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3087 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3089 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3090 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3091 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3093 // If we have a constant, make sure it is on the RHS.
3094 std::swap(LHS, RHS);
3095 Pred = CmpInst::getSwappedPredicate(Pred);
3098 Type *ITy = GetCompareTy(LHS); // The return type.
3100 // icmp X, X -> true/false
3101 // icmp X, undef -> true/false because undef could be X.
3102 if (LHS == RHS || isa<UndefValue>(RHS))
3103 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3105 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3108 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3111 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3114 // If both operands have range metadata, use the metadata
3115 // to simplify the comparison.
3116 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3117 auto RHS_Instr = cast<Instruction>(RHS);
3118 auto LHS_Instr = cast<Instruction>(LHS);
3120 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3121 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3122 auto RHS_CR = getConstantRangeFromMetadata(
3123 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3124 auto LHS_CR = getConstantRangeFromMetadata(
3125 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3127 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3128 if (Satisfied_CR.contains(LHS_CR))
3129 return ConstantInt::getTrue(RHS->getContext());
3131 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3132 CmpInst::getInversePredicate(Pred), RHS_CR);
3133 if (InversedSatisfied_CR.contains(LHS_CR))
3134 return ConstantInt::getFalse(RHS->getContext());
3138 // Compare of cast, for example (zext X) != 0 -> X != 0
3139 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3140 Instruction *LI = cast<CastInst>(LHS);
3141 Value *SrcOp = LI->getOperand(0);
3142 Type *SrcTy = SrcOp->getType();
3143 Type *DstTy = LI->getType();
3145 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3146 // if the integer type is the same size as the pointer type.
3147 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3148 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3149 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3150 // Transfer the cast to the constant.
3151 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3152 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3155 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3156 if (RI->getOperand(0)->getType() == SrcTy)
3157 // Compare without the cast.
3158 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3164 if (isa<ZExtInst>(LHS)) {
3165 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3167 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3168 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3169 // Compare X and Y. Note that signed predicates become unsigned.
3170 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3171 SrcOp, RI->getOperand(0), Q,
3175 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3176 // too. If not, then try to deduce the result of the comparison.
3177 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3178 // Compute the constant that would happen if we truncated to SrcTy then
3179 // reextended to DstTy.
3180 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3181 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3183 // If the re-extended constant didn't change then this is effectively
3184 // also a case of comparing two zero-extended values.
3185 if (RExt == CI && MaxRecurse)
3186 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3187 SrcOp, Trunc, Q, MaxRecurse-1))
3190 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3191 // there. Use this to work out the result of the comparison.
3194 default: llvm_unreachable("Unknown ICmp predicate!");
3196 case ICmpInst::ICMP_EQ:
3197 case ICmpInst::ICMP_UGT:
3198 case ICmpInst::ICMP_UGE:
3199 return ConstantInt::getFalse(CI->getContext());
3201 case ICmpInst::ICMP_NE:
3202 case ICmpInst::ICMP_ULT:
3203 case ICmpInst::ICMP_ULE:
3204 return ConstantInt::getTrue(CI->getContext());
3206 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3207 // is non-negative then LHS <s RHS.
3208 case ICmpInst::ICMP_SGT:
3209 case ICmpInst::ICMP_SGE:
3210 return CI->getValue().isNegative() ?
3211 ConstantInt::getTrue(CI->getContext()) :
3212 ConstantInt::getFalse(CI->getContext());
3214 case ICmpInst::ICMP_SLT:
3215 case ICmpInst::ICMP_SLE:
3216 return CI->getValue().isNegative() ?
3217 ConstantInt::getFalse(CI->getContext()) :
3218 ConstantInt::getTrue(CI->getContext());
3224 if (isa<SExtInst>(LHS)) {
3225 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3227 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3228 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3229 // Compare X and Y. Note that the predicate does not change.
3230 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3234 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3235 // too. If not, then try to deduce the result of the comparison.
3236 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3237 // Compute the constant that would happen if we truncated to SrcTy then
3238 // reextended to DstTy.
3239 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3240 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3242 // If the re-extended constant didn't change then this is effectively
3243 // also a case of comparing two sign-extended values.
3244 if (RExt == CI && MaxRecurse)
3245 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3248 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3249 // bits there. Use this to work out the result of the comparison.
3252 default: llvm_unreachable("Unknown ICmp predicate!");
3253 case ICmpInst::ICMP_EQ:
3254 return ConstantInt::getFalse(CI->getContext());
3255 case ICmpInst::ICMP_NE:
3256 return ConstantInt::getTrue(CI->getContext());
3258 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3260 case ICmpInst::ICMP_SGT:
3261 case ICmpInst::ICMP_SGE:
3262 return CI->getValue().isNegative() ?
3263 ConstantInt::getTrue(CI->getContext()) :
3264 ConstantInt::getFalse(CI->getContext());
3265 case ICmpInst::ICMP_SLT:
3266 case ICmpInst::ICMP_SLE:
3267 return CI->getValue().isNegative() ?
3268 ConstantInt::getFalse(CI->getContext()) :
3269 ConstantInt::getTrue(CI->getContext());
3271 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3273 case ICmpInst::ICMP_UGT:
3274 case ICmpInst::ICMP_UGE:
3275 // Comparison is true iff the LHS <s 0.
3277 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3278 Constant::getNullValue(SrcTy),
3282 case ICmpInst::ICMP_ULT:
3283 case ICmpInst::ICMP_ULE:
3284 // Comparison is true iff the LHS >=s 0.
3286 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3287 Constant::getNullValue(SrcTy),
3297 // icmp eq|ne X, Y -> false|true if X != Y
3298 if (ICmpInst::isEquality(Pred) &&
3299 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3300 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3303 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3306 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3309 // Simplify comparisons of related pointers using a powerful, recursive
3310 // GEP-walk when we have target data available..
3311 if (LHS->getType()->isPointerTy())
3312 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3315 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3316 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3317 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3318 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3319 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3320 Q.DL.getTypeSizeInBits(CRHS->getType()))
3321 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3322 CLHS->getPointerOperand(),
3323 CRHS->getPointerOperand()))
3326 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3327 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3328 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3329 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3330 (ICmpInst::isEquality(Pred) ||
3331 (GLHS->isInBounds() && GRHS->isInBounds() &&
3332 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3333 // The bases are equal and the indices are constant. Build a constant
3334 // expression GEP with the same indices and a null base pointer to see
3335 // what constant folding can make out of it.
3336 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3337 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3338 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3339 GLHS->getSourceElementType(), Null, IndicesLHS);
3341 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3342 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3343 GLHS->getSourceElementType(), Null, IndicesRHS);
3344 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3349 // If the comparison is with the result of a select instruction, check whether
3350 // comparing with either branch of the select always yields the same value.
3351 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3352 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3355 // If the comparison is with the result of a phi instruction, check whether
3356 // doing the compare with each incoming phi value yields a common result.
3357 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3358 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3364 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3365 const SimplifyQuery &Q) {
3366 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3369 /// Given operands for an FCmpInst, see if we can fold the result.
3370 /// If not, this returns null.
3371 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3372 FastMathFlags FMF, const SimplifyQuery &Q,
3373 unsigned MaxRecurse) {
3374 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3375 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3377 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3378 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3379 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3381 // If we have a constant, make sure it is on the RHS.
3382 std::swap(LHS, RHS);
3383 Pred = CmpInst::getSwappedPredicate(Pred);
3386 // Fold trivial predicates.
3387 Type *RetTy = GetCompareTy(LHS);
3388 if (Pred == FCmpInst::FCMP_FALSE)
3389 return getFalse(RetTy);
3390 if (Pred == FCmpInst::FCMP_TRUE)
3391 return getTrue(RetTy);
3393 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3395 if (Pred == FCmpInst::FCMP_UNO)
3396 return getFalse(RetTy);
3397 if (Pred == FCmpInst::FCMP_ORD)
3398 return getTrue(RetTy);
3401 // NaN is unordered; NaN is not ordered.
3402 assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&
3403 "Comparison must be either ordered or unordered");
3404 if (match(RHS, m_NaN()))
3405 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3407 // fcmp pred x, undef and fcmp pred undef, x
3408 // fold to true if unordered, false if ordered
3409 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3410 // Choosing NaN for the undef will always make unordered comparison succeed
3411 // and ordered comparison fail.
3412 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3415 // fcmp x,x -> true/false. Not all compares are foldable.
3417 if (CmpInst::isTrueWhenEqual(Pred))
3418 return getTrue(RetTy);
3419 if (CmpInst::isFalseWhenEqual(Pred))
3420 return getFalse(RetTy);
3423 // Handle fcmp with constant RHS.
3425 if (match(RHS, m_APFloat(C))) {
3426 // Check whether the constant is an infinity.
3427 if (C->isInfinity()) {
3428 if (C->isNegative()) {
3430 case FCmpInst::FCMP_OLT:
3431 // No value is ordered and less than negative infinity.
3432 return getFalse(RetTy);
3433 case FCmpInst::FCMP_UGE:
3434 // All values are unordered with or at least negative infinity.
3435 return getTrue(RetTy);
3441 case FCmpInst::FCMP_OGT:
3442 // No value is ordered and greater than infinity.
3443 return getFalse(RetTy);
3444 case FCmpInst::FCMP_ULE:
3445 // All values are unordered with and at most infinity.
3446 return getTrue(RetTy);
3454 case FCmpInst::FCMP_UGE:
3455 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3456 return getTrue(RetTy);
3458 case FCmpInst::FCMP_OLT:
3460 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3461 return getFalse(RetTy);
3466 } else if (C->isNegative()) {
3467 assert(!C->isNaN() && "Unexpected NaN constant!");
3468 // TODO: We can catch more cases by using a range check rather than
3469 // relying on CannotBeOrderedLessThanZero.
3471 case FCmpInst::FCMP_UGE:
3472 case FCmpInst::FCMP_UGT:
3473 case FCmpInst::FCMP_UNE:
3474 // (X >= 0) implies (X > C) when (C < 0)
3475 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3476 return getTrue(RetTy);
3478 case FCmpInst::FCMP_OEQ:
3479 case FCmpInst::FCMP_OLE:
3480 case FCmpInst::FCMP_OLT:
3481 // (X >= 0) implies !(X < C) when (C < 0)
3482 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3483 return getFalse(RetTy);
3491 // If the comparison is with the result of a select instruction, check whether
3492 // comparing with either branch of the select always yields the same value.
3493 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3494 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3497 // If the comparison is with the result of a phi instruction, check whether
3498 // doing the compare with each incoming phi value yields a common result.
3499 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3500 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3506 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3507 FastMathFlags FMF, const SimplifyQuery &Q) {
3508 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3511 /// See if V simplifies when its operand Op is replaced with RepOp.
3512 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3513 const SimplifyQuery &Q,
3514 unsigned MaxRecurse) {
3515 // Trivial replacement.
3519 // We cannot replace a constant, and shouldn't even try.
3520 if (isa<Constant>(Op))
3523 auto *I = dyn_cast<Instruction>(V);
3527 // If this is a binary operator, try to simplify it with the replaced op.
3528 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3530 // %cmp = icmp eq i32 %x, 2147483647
3531 // %add = add nsw i32 %x, 1
3532 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3534 // We can't replace %sel with %add unless we strip away the flags.
3535 if (isa<OverflowingBinaryOperator>(B))
3536 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3538 if (isa<PossiblyExactOperator>(B))
3543 if (B->getOperand(0) == Op)
3544 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3546 if (B->getOperand(1) == Op)
3547 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3552 // Same for CmpInsts.
3553 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3555 if (C->getOperand(0) == Op)
3556 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3558 if (C->getOperand(1) == Op)
3559 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3565 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3567 SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3568 transform(GEP->operands(), NewOps.begin(),
3569 [&](Value *V) { return V == Op ? RepOp : V; });
3570 return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3575 // TODO: We could hand off more cases to instsimplify here.
3577 // If all operands are constant after substituting Op for RepOp then we can
3578 // constant fold the instruction.
3579 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3580 // Build a list of all constant operands.
3581 SmallVector<Constant *, 8> ConstOps;
3582 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3583 if (I->getOperand(i) == Op)
3584 ConstOps.push_back(CRepOp);
3585 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3586 ConstOps.push_back(COp);
3591 // All operands were constants, fold it.
3592 if (ConstOps.size() == I->getNumOperands()) {
3593 if (CmpInst *C = dyn_cast<CmpInst>(I))
3594 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3595 ConstOps[1], Q.DL, Q.TLI);
3597 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3598 if (!LI->isVolatile())
3599 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3601 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3608 /// Try to simplify a select instruction when its condition operand is an
3609 /// integer comparison where one operand of the compare is a constant.
3610 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3611 const APInt *Y, bool TrueWhenUnset) {
3614 // (X & Y) == 0 ? X & ~Y : X --> X
3615 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3616 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3618 return TrueWhenUnset ? FalseVal : TrueVal;
3620 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3621 // (X & Y) != 0 ? X : X & ~Y --> X
3622 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3624 return TrueWhenUnset ? FalseVal : TrueVal;
3626 if (Y->isPowerOf2()) {
3627 // (X & Y) == 0 ? X | Y : X --> X | Y
3628 // (X & Y) != 0 ? X | Y : X --> X
3629 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3631 return TrueWhenUnset ? TrueVal : FalseVal;
3633 // (X & Y) == 0 ? X : X | Y --> X
3634 // (X & Y) != 0 ? X : X | Y --> X | Y
3635 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3637 return TrueWhenUnset ? TrueVal : FalseVal;
3643 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3645 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3646 ICmpInst::Predicate Pred,
3647 Value *TrueVal, Value *FalseVal) {
3650 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3653 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3654 Pred == ICmpInst::ICMP_EQ);
3657 /// Try to simplify a select instruction when its condition operand is an
3658 /// integer comparison.
3659 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3660 Value *FalseVal, const SimplifyQuery &Q,
3661 unsigned MaxRecurse) {
3662 ICmpInst::Predicate Pred;
3663 Value *CmpLHS, *CmpRHS;
3664 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3667 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3670 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3671 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3672 Pred == ICmpInst::ICMP_EQ))
3676 // Check for other compares that behave like bit test.
3677 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3681 // If we have an equality comparison, then we know the value in one of the
3682 // arms of the select. See if substituting this value into the arm and
3683 // simplifying the result yields the same value as the other arm.
3684 if (Pred == ICmpInst::ICMP_EQ) {
3685 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3687 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3690 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3692 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3695 } else if (Pred == ICmpInst::ICMP_NE) {
3696 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3698 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3701 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3703 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3711 /// Given operands for a SelectInst, see if we can fold the result.
3712 /// If not, this returns null.
3713 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3714 const SimplifyQuery &Q, unsigned MaxRecurse) {
3715 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3716 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3717 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3718 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3720 // select undef, X, Y -> X or Y
3721 if (isa<UndefValue>(CondC))
3722 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3724 // TODO: Vector constants with undef elements don't simplify.
3726 // select true, X, Y -> X
3727 if (CondC->isAllOnesValue())
3729 // select false, X, Y -> Y
3730 if (CondC->isNullValue())
3734 // select ?, X, X -> X
3735 if (TrueVal == FalseVal)
3738 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3740 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3744 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3750 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3751 const SimplifyQuery &Q) {
3752 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3755 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3756 /// If not, this returns null.
3757 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3758 const SimplifyQuery &Q, unsigned) {
3759 // The type of the GEP pointer operand.
3761 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3763 // getelementptr P -> P.
3764 if (Ops.size() == 1)
3767 // Compute the (pointer) type returned by the GEP instruction.
3768 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3769 Type *GEPTy = PointerType::get(LastType, AS);
3770 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3771 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3772 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3773 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3775 if (isa<UndefValue>(Ops[0]))
3776 return UndefValue::get(GEPTy);
3778 if (Ops.size() == 2) {
3779 // getelementptr P, 0 -> P.
3780 if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3784 if (Ty->isSized()) {
3787 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3788 // getelementptr P, N -> P if P points to a type of zero size.
3789 if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3792 // The following transforms are only safe if the ptrtoint cast
3793 // doesn't truncate the pointers.
3794 if (Ops[1]->getType()->getScalarSizeInBits() ==
3795 Q.DL.getIndexSizeInBits(AS)) {
3796 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3797 if (match(P, m_Zero()))
3798 return Constant::getNullValue(GEPTy);
3800 if (match(P, m_PtrToInt(m_Value(Temp))))
3801 if (Temp->getType() == GEPTy)
3806 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3807 if (TyAllocSize == 1 &&
3808 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3809 if (Value *R = PtrToIntOrZero(P))
3812 // getelementptr V, (ashr (sub P, V), C) -> Q
3813 // if P points to a type of size 1 << C.
3815 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3816 m_ConstantInt(C))) &&
3817 TyAllocSize == 1ULL << C)
3818 if (Value *R = PtrToIntOrZero(P))
3821 // getelementptr V, (sdiv (sub P, V), C) -> Q
3822 // if P points to a type of size C.
3824 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3825 m_SpecificInt(TyAllocSize))))
3826 if (Value *R = PtrToIntOrZero(P))
3832 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3833 all_of(Ops.slice(1).drop_back(1),
3834 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3836 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3837 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3838 APInt BasePtrOffset(IdxWidth, 0);
3839 Value *StrippedBasePtr =
3840 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3843 // gep (gep V, C), (sub 0, V) -> C
3844 if (match(Ops.back(),
3845 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3846 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3847 return ConstantExpr::getIntToPtr(CI, GEPTy);
3849 // gep (gep V, C), (xor V, -1) -> C-1
3850 if (match(Ops.back(),
3851 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3852 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3853 return ConstantExpr::getIntToPtr(CI, GEPTy);
3858 // Check to see if this is constant foldable.
3859 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3862 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3864 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3869 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3870 const SimplifyQuery &Q) {
3871 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3874 /// Given operands for an InsertValueInst, see if we can fold the result.
3875 /// If not, this returns null.
3876 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3877 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3879 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3880 if (Constant *CVal = dyn_cast<Constant>(Val))
3881 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3883 // insertvalue x, undef, n -> x
3884 if (match(Val, m_Undef()))
3887 // insertvalue x, (extractvalue y, n), n
3888 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3889 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3890 EV->getIndices() == Idxs) {
3891 // insertvalue undef, (extractvalue y, n), n -> y
3892 if (match(Agg, m_Undef()))
3893 return EV->getAggregateOperand();
3895 // insertvalue y, (extractvalue y, n), n -> y
3896 if (Agg == EV->getAggregateOperand())
3903 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3904 ArrayRef<unsigned> Idxs,
3905 const SimplifyQuery &Q) {
3906 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3909 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3910 const SimplifyQuery &Q) {
3911 // Try to constant fold.
3912 auto *VecC = dyn_cast<Constant>(Vec);
3913 auto *ValC = dyn_cast<Constant>(Val);
3914 auto *IdxC = dyn_cast<Constant>(Idx);
3915 if (VecC && ValC && IdxC)
3916 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3918 // Fold into undef if index is out of bounds.
3919 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3920 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3921 if (CI->uge(NumElements))
3922 return UndefValue::get(Vec->getType());
3925 // If index is undef, it might be out of bounds (see above case)
3926 if (isa<UndefValue>(Idx))
3927 return UndefValue::get(Vec->getType());
3932 /// Given operands for an ExtractValueInst, see if we can fold the result.
3933 /// If not, this returns null.
3934 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3935 const SimplifyQuery &, unsigned) {
3936 if (auto *CAgg = dyn_cast<Constant>(Agg))
3937 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3939 // extractvalue x, (insertvalue y, elt, n), n -> elt
3940 unsigned NumIdxs = Idxs.size();
3941 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3942 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3943 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3944 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3945 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3946 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3947 Idxs.slice(0, NumCommonIdxs)) {
3948 if (NumIdxs == NumInsertValueIdxs)
3949 return IVI->getInsertedValueOperand();
3957 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3958 const SimplifyQuery &Q) {
3959 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3962 /// Given operands for an ExtractElementInst, see if we can fold the result.
3963 /// If not, this returns null.
3964 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3966 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3967 if (auto *CIdx = dyn_cast<Constant>(Idx))
3968 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3970 // The index is not relevant if our vector is a splat.
3971 if (auto *Splat = CVec->getSplatValue())
3974 if (isa<UndefValue>(Vec))
3975 return UndefValue::get(Vec->getType()->getVectorElementType());
3978 // If extracting a specified index from the vector, see if we can recursively
3979 // find a previously computed scalar that was inserted into the vector.
3980 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
3981 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
3982 // definitely out of bounds, thus undefined result
3983 return UndefValue::get(Vec->getType()->getVectorElementType());
3984 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3988 // An undef extract index can be arbitrarily chosen to be an out-of-range
3989 // index value, which would result in the instruction being undef.
3990 if (isa<UndefValue>(Idx))
3991 return UndefValue::get(Vec->getType()->getVectorElementType());
3996 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3997 const SimplifyQuery &Q) {
3998 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
4001 /// See if we can fold the given phi. If not, returns null.
4002 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4003 // If all of the PHI's incoming values are the same then replace the PHI node
4004 // with the common value.
4005 Value *CommonValue = nullptr;
4006 bool HasUndefInput = false;
4007 for (Value *Incoming : PN->incoming_values()) {
4008 // If the incoming value is the phi node itself, it can safely be skipped.
4009 if (Incoming == PN) continue;
4010 if (isa<UndefValue>(Incoming)) {
4011 // Remember that we saw an undef value, but otherwise ignore them.
4012 HasUndefInput = true;
4015 if (CommonValue && Incoming != CommonValue)
4016 return nullptr; // Not the same, bail out.
4017 CommonValue = Incoming;
4020 // If CommonValue is null then all of the incoming values were either undef or
4021 // equal to the phi node itself.
4023 return UndefValue::get(PN->getType());
4025 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4026 // instruction, we cannot return X as the result of the PHI node unless it
4027 // dominates the PHI block.
4029 return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4034 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4035 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4036 if (auto *C = dyn_cast<Constant>(Op))
4037 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4039 if (auto *CI = dyn_cast<CastInst>(Op)) {
4040 auto *Src = CI->getOperand(0);
4041 Type *SrcTy = Src->getType();
4042 Type *MidTy = CI->getType();
4044 if (Src->getType() == Ty) {
4045 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4046 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4048 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4050 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4052 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4053 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4054 SrcIntPtrTy, MidIntPtrTy,
4055 DstIntPtrTy) == Instruction::BitCast)
4061 if (CastOpc == Instruction::BitCast)
4062 if (Op->getType() == Ty)
4068 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4069 const SimplifyQuery &Q) {
4070 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4073 /// For the given destination element of a shuffle, peek through shuffles to
4074 /// match a root vector source operand that contains that element in the same
4075 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4076 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4077 int MaskVal, Value *RootVec,
4078 unsigned MaxRecurse) {
4082 // Bail out if any mask value is undefined. That kind of shuffle may be
4083 // simplified further based on demanded bits or other folds.
4087 // The mask value chooses which source operand we need to look at next.
4088 int InVecNumElts = Op0->getType()->getVectorNumElements();
4089 int RootElt = MaskVal;
4090 Value *SourceOp = Op0;
4091 if (MaskVal >= InVecNumElts) {
4092 RootElt = MaskVal - InVecNumElts;
4096 // If the source operand is a shuffle itself, look through it to find the
4097 // matching root vector.
4098 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4099 return foldIdentityShuffles(
4100 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4101 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4104 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4107 // The source operand is not a shuffle. Initialize the root vector value for
4108 // this shuffle if that has not been done yet.
4112 // Give up as soon as a source operand does not match the existing root value.
4113 if (RootVec != SourceOp)
4116 // The element must be coming from the same lane in the source vector
4117 // (although it may have crossed lanes in intermediate shuffles).
4118 if (RootElt != DestElt)
4124 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4125 Type *RetTy, const SimplifyQuery &Q,
4126 unsigned MaxRecurse) {
4127 if (isa<UndefValue>(Mask))
4128 return UndefValue::get(RetTy);
4130 Type *InVecTy = Op0->getType();
4131 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4132 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4134 SmallVector<int, 32> Indices;
4135 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4136 assert(MaskNumElts == Indices.size() &&
4137 "Size of Indices not same as number of mask elements?");
4139 // Canonicalization: If mask does not select elements from an input vector,
4140 // replace that input vector with undef.
4141 bool MaskSelects0 = false, MaskSelects1 = false;
4142 for (unsigned i = 0; i != MaskNumElts; ++i) {
4143 if (Indices[i] == -1)
4145 if ((unsigned)Indices[i] < InVecNumElts)
4146 MaskSelects0 = true;
4148 MaskSelects1 = true;
4151 Op0 = UndefValue::get(InVecTy);
4153 Op1 = UndefValue::get(InVecTy);
4155 auto *Op0Const = dyn_cast<Constant>(Op0);
4156 auto *Op1Const = dyn_cast<Constant>(Op1);
4158 // If all operands are constant, constant fold the shuffle.
4159 if (Op0Const && Op1Const)
4160 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4162 // Canonicalization: if only one input vector is constant, it shall be the
4164 if (Op0Const && !Op1Const) {
4165 std::swap(Op0, Op1);
4166 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4169 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4170 // value type is same as the input vectors' type.
4171 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4172 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4173 OpShuf->getMask()->getSplatValue())
4176 // Don't fold a shuffle with undef mask elements. This may get folded in a
4177 // better way using demanded bits or other analysis.
4178 // TODO: Should we allow this?
4179 if (find(Indices, -1) != Indices.end())
4182 // Check if every element of this shuffle can be mapped back to the
4183 // corresponding element of a single root vector. If so, we don't need this
4184 // shuffle. This handles simple identity shuffles as well as chains of
4185 // shuffles that may widen/narrow and/or move elements across lanes and back.
4186 Value *RootVec = nullptr;
4187 for (unsigned i = 0; i != MaskNumElts; ++i) {
4188 // Note that recursion is limited for each vector element, so if any element
4189 // exceeds the limit, this will fail to simplify.
4191 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4193 // We can't replace a widening/narrowing shuffle with one of its operands.
4194 if (!RootVec || RootVec->getType() != RetTy)
4200 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4201 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4202 Type *RetTy, const SimplifyQuery &Q) {
4203 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4206 static Constant *propagateNaN(Constant *In) {
4207 // If the input is a vector with undef elements, just return a default NaN.
4209 return ConstantFP::getNaN(In->getType());
4211 // Propagate the existing NaN constant when possible.
4212 // TODO: Should we quiet a signaling NaN?
4216 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4217 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4218 return ConstantFP::getNaN(Op0->getType());
4220 if (match(Op0, m_NaN()))
4221 return propagateNaN(cast<Constant>(Op0));
4222 if (match(Op1, m_NaN()))
4223 return propagateNaN(cast<Constant>(Op1));
4228 /// Given operands for an FAdd, see if we can fold the result. If not, this
4230 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4231 const SimplifyQuery &Q, unsigned MaxRecurse) {
4232 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4235 if (Constant *C = simplifyFPBinop(Op0, Op1))
4239 if (match(Op1, m_NegZeroFP()))
4242 // fadd X, 0 ==> X, when we know X is not -0
4243 if (match(Op1, m_PosZeroFP()) &&
4244 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4247 // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4248 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4249 // Negative zeros are allowed because we always end up with positive zero:
4250 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4251 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4252 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4253 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4254 if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4255 match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4256 return ConstantFP::getNullValue(Op0->getType());
4261 /// Given operands for an FSub, see if we can fold the result. If not, this
4263 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4264 const SimplifyQuery &Q, unsigned MaxRecurse) {
4265 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4268 if (Constant *C = simplifyFPBinop(Op0, Op1))
4272 if (match(Op1, m_PosZeroFP()))
4275 // fsub X, -0 ==> X, when we know X is not -0
4276 if (match(Op1, m_NegZeroFP()) &&
4277 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4280 // fsub -0.0, (fsub -0.0, X) ==> X
4282 if (match(Op0, m_NegZeroFP()) &&
4283 match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4286 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4287 if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4288 match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4291 // fsub nnan x, x ==> 0.0
4292 if (FMF.noNaNs() && Op0 == Op1)
4293 return Constant::getNullValue(Op0->getType());
4298 /// Given the operands for an FMul, see if we can fold the result
4299 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4300 const SimplifyQuery &Q, unsigned MaxRecurse) {
4301 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4304 if (Constant *C = simplifyFPBinop(Op0, Op1))
4307 // fmul X, 1.0 ==> X
4308 if (match(Op1, m_FPOne()))
4311 // fmul nnan nsz X, 0 ==> 0
4312 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4313 return ConstantFP::getNullValue(Op0->getType());
4315 // sqrt(X) * sqrt(X) --> X, if we can:
4316 // 1. Remove the intermediate rounding (reassociate).
4317 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4318 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4320 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4321 FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4327 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4328 const SimplifyQuery &Q) {
4329 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4333 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4334 const SimplifyQuery &Q) {
4335 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4338 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4339 const SimplifyQuery &Q) {
4340 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4343 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4344 const SimplifyQuery &Q, unsigned) {
4345 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4348 if (Constant *C = simplifyFPBinop(Op0, Op1))
4352 if (match(Op1, m_FPOne()))
4356 // Requires that NaNs are off (X could be zero) and signed zeroes are
4357 // ignored (X could be positive or negative, so the output sign is unknown).
4358 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4359 return ConstantFP::getNullValue(Op0->getType());
4362 // X / X -> 1.0 is legal when NaNs are ignored.
4363 // We can ignore infinities because INF/INF is NaN.
4365 return ConstantFP::get(Op0->getType(), 1.0);
4367 // (X * Y) / Y --> X if we can reassociate to the above form.
4369 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4372 // -X / X -> -1.0 and
4373 // X / -X -> -1.0 are legal when NaNs are ignored.
4374 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4375 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4376 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4377 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4378 BinaryOperator::getFNegArgument(Op1) == Op0))
4379 return ConstantFP::get(Op0->getType(), -1.0);
4385 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4386 const SimplifyQuery &Q) {
4387 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4390 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4391 const SimplifyQuery &Q, unsigned) {
4392 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4395 if (Constant *C = simplifyFPBinop(Op0, Op1))
4398 // Unlike fdiv, the result of frem always matches the sign of the dividend.
4399 // The constant match may include undef elements in a vector, so return a full
4400 // zero constant as the result.
4403 if (match(Op0, m_PosZeroFP()))
4404 return ConstantFP::getNullValue(Op0->getType());
4406 if (match(Op0, m_NegZeroFP()))
4407 return ConstantFP::getNegativeZero(Op0->getType());
4413 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4414 const SimplifyQuery &Q) {
4415 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4418 //=== Helper functions for higher up the class hierarchy.
4420 /// Given operands for a BinaryOperator, see if we can fold the result.
4421 /// If not, this returns null.
4422 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4423 const SimplifyQuery &Q, unsigned MaxRecurse) {
4425 case Instruction::Add:
4426 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4427 case Instruction::Sub:
4428 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4429 case Instruction::Mul:
4430 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4431 case Instruction::SDiv:
4432 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4433 case Instruction::UDiv:
4434 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4435 case Instruction::SRem:
4436 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4437 case Instruction::URem:
4438 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4439 case Instruction::Shl:
4440 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4441 case Instruction::LShr:
4442 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4443 case Instruction::AShr:
4444 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4445 case Instruction::And:
4446 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4447 case Instruction::Or:
4448 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4449 case Instruction::Xor:
4450 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4451 case Instruction::FAdd:
4452 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4453 case Instruction::FSub:
4454 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4455 case Instruction::FMul:
4456 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4457 case Instruction::FDiv:
4458 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4459 case Instruction::FRem:
4460 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4462 llvm_unreachable("Unexpected opcode");
4466 /// Given operands for a BinaryOperator, see if we can fold the result.
4467 /// If not, this returns null.
4468 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4469 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4470 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4471 const FastMathFlags &FMF, const SimplifyQuery &Q,
4472 unsigned MaxRecurse) {
4474 case Instruction::FAdd:
4475 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4476 case Instruction::FSub:
4477 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4478 case Instruction::FMul:
4479 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4480 case Instruction::FDiv:
4481 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4483 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4487 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4488 const SimplifyQuery &Q) {
4489 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4492 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4493 FastMathFlags FMF, const SimplifyQuery &Q) {
4494 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4497 /// Given operands for a CmpInst, see if we can fold the result.
4498 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4499 const SimplifyQuery &Q, unsigned MaxRecurse) {
4500 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4501 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4502 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4505 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4506 const SimplifyQuery &Q) {
4507 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4510 static bool IsIdempotent(Intrinsic::ID ID) {
4512 default: return false;
4514 // Unary idempotent: f(f(x)) = f(x)
4515 case Intrinsic::fabs:
4516 case Intrinsic::floor:
4517 case Intrinsic::ceil:
4518 case Intrinsic::trunc:
4519 case Intrinsic::rint:
4520 case Intrinsic::nearbyint:
4521 case Intrinsic::round:
4522 case Intrinsic::canonicalize:
4527 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4528 const DataLayout &DL) {
4529 GlobalValue *PtrSym;
4531 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4534 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4535 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4536 Type *Int32PtrTy = Int32Ty->getPointerTo();
4537 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4539 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4540 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4543 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4544 if (OffsetInt % 4 != 0)
4547 Constant *C = ConstantExpr::getGetElementPtr(
4548 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4549 ConstantInt::get(Int64Ty, OffsetInt / 4));
4550 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4554 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4558 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4559 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4564 if (LoadedCE->getOpcode() != Instruction::Sub)
4567 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4568 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4570 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4572 Constant *LoadedRHS = LoadedCE->getOperand(1);
4573 GlobalValue *LoadedRHSSym;
4574 APInt LoadedRHSOffset;
4575 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4577 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4580 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4583 static bool maskIsAllZeroOrUndef(Value *Mask) {
4584 auto *ConstMask = dyn_cast<Constant>(Mask);
4587 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4589 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4591 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4592 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4599 template <typename IterTy>
4600 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4601 const SimplifyQuery &Q, unsigned MaxRecurse) {
4602 Intrinsic::ID IID = F->getIntrinsicID();
4603 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4606 if (NumOperands == 1) {
4607 // Perform idempotent optimizations
4608 if (IsIdempotent(IID)) {
4609 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4610 if (II->getIntrinsicID() == IID)
4615 Value *IIOperand = *ArgBegin;
4618 case Intrinsic::fabs: {
4619 if (SignBitMustBeZero(IIOperand, Q.TLI))
4623 case Intrinsic::bswap: {
4624 // bswap(bswap(x)) -> x
4625 if (match(IIOperand, m_BSwap(m_Value(X))))
4629 case Intrinsic::bitreverse: {
4630 // bitreverse(bitreverse(x)) -> x
4631 if (match(IIOperand, m_BitReverse(m_Value(X))))
4635 case Intrinsic::exp: {
4637 if (Q.CxtI->hasAllowReassoc() &&
4638 match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X))))
4642 case Intrinsic::exp2: {
4643 // exp2(log2(x)) -> x
4644 if (Q.CxtI->hasAllowReassoc() &&
4645 match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X))))
4649 case Intrinsic::log: {
4651 if (Q.CxtI->hasAllowReassoc() &&
4652 match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X))))
4656 case Intrinsic::log2: {
4657 // log2(exp2(x)) -> x
4658 if (Q.CxtI->hasAllowReassoc() &&
4659 match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) {
4670 if (NumOperands == 2) {
4671 Value *LHS = *ArgBegin;
4672 Value *RHS = *(ArgBegin + 1);
4673 Type *ReturnType = F->getReturnType();
4676 case Intrinsic::usub_with_overflow:
4677 case Intrinsic::ssub_with_overflow: {
4678 // X - X -> { 0, false }
4680 return Constant::getNullValue(ReturnType);
4682 // X - undef -> undef
4683 // undef - X -> undef
4684 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4685 return UndefValue::get(ReturnType);
4689 case Intrinsic::uadd_with_overflow:
4690 case Intrinsic::sadd_with_overflow: {
4691 // X + undef -> undef
4692 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4693 return UndefValue::get(ReturnType);
4697 case Intrinsic::umul_with_overflow:
4698 case Intrinsic::smul_with_overflow: {
4699 // 0 * X -> { 0, false }
4700 // X * 0 -> { 0, false }
4701 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4702 return Constant::getNullValue(ReturnType);
4704 // undef * X -> { 0, false }
4705 // X * undef -> { 0, false }
4706 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4707 return Constant::getNullValue(ReturnType);
4711 case Intrinsic::load_relative: {
4712 Constant *C0 = dyn_cast<Constant>(LHS);
4713 Constant *C1 = dyn_cast<Constant>(RHS);
4715 return SimplifyRelativeLoad(C0, C1, Q.DL);
4718 case Intrinsic::powi:
4719 if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) {
4720 // powi(x, 0) -> 1.0
4721 if (Power->isZero())
4722 return ConstantFP::get(LHS->getType(), 1.0);
4733 // Simplify calls to llvm.masked.load.*
4735 case Intrinsic::masked_load: {
4736 Value *MaskArg = ArgBegin[2];
4737 Value *PassthruArg = ArgBegin[3];
4738 // If the mask is all zeros or undef, the "passthru" argument is the result.
4739 if (maskIsAllZeroOrUndef(MaskArg))
4748 template <typename IterTy>
4749 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4750 IterTy ArgEnd, const SimplifyQuery &Q,
4751 unsigned MaxRecurse) {
4752 Type *Ty = V->getType();
4753 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4754 Ty = PTy->getElementType();
4755 FunctionType *FTy = cast<FunctionType>(Ty);
4757 // call undef -> undef
4758 // call null -> undef
4759 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4760 return UndefValue::get(FTy->getReturnType());
4762 Function *F = dyn_cast<Function>(V);
4766 if (F->isIntrinsic())
4767 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4770 if (!canConstantFoldCallTo(CS, F))
4773 SmallVector<Constant *, 4> ConstantArgs;
4774 ConstantArgs.reserve(ArgEnd - ArgBegin);
4775 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4776 Constant *C = dyn_cast<Constant>(*I);
4779 ConstantArgs.push_back(C);
4782 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4785 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4786 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4787 const SimplifyQuery &Q) {
4788 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4791 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4792 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4793 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4796 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4797 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4798 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4802 /// See if we can compute a simplified version of this instruction.
4803 /// If not, this returns null.
4805 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4806 OptimizationRemarkEmitter *ORE) {
4807 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4810 switch (I->getOpcode()) {
4812 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4814 case Instruction::FAdd:
4815 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4816 I->getFastMathFlags(), Q);
4818 case Instruction::Add:
4819 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4820 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4821 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4823 case Instruction::FSub:
4824 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4825 I->getFastMathFlags(), Q);
4827 case Instruction::Sub:
4828 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4829 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4830 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4832 case Instruction::FMul:
4833 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4834 I->getFastMathFlags(), Q);
4836 case Instruction::Mul:
4837 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4839 case Instruction::SDiv:
4840 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4842 case Instruction::UDiv:
4843 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4845 case Instruction::FDiv:
4846 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4847 I->getFastMathFlags(), Q);
4849 case Instruction::SRem:
4850 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4852 case Instruction::URem:
4853 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4855 case Instruction::FRem:
4856 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4857 I->getFastMathFlags(), Q);
4859 case Instruction::Shl:
4860 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4861 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4862 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4864 case Instruction::LShr:
4865 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4866 cast<BinaryOperator>(I)->isExact(), Q);
4868 case Instruction::AShr:
4869 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4870 cast<BinaryOperator>(I)->isExact(), Q);
4872 case Instruction::And:
4873 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4875 case Instruction::Or:
4876 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4878 case Instruction::Xor:
4879 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4881 case Instruction::ICmp:
4882 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4883 I->getOperand(0), I->getOperand(1), Q);
4885 case Instruction::FCmp:
4887 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4888 I->getOperand(1), I->getFastMathFlags(), Q);
4890 case Instruction::Select:
4891 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4892 I->getOperand(2), Q);
4894 case Instruction::GetElementPtr: {
4895 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4896 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4900 case Instruction::InsertValue: {
4901 InsertValueInst *IV = cast<InsertValueInst>(I);
4902 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4903 IV->getInsertedValueOperand(),
4904 IV->getIndices(), Q);
4907 case Instruction::InsertElement: {
4908 auto *IE = cast<InsertElementInst>(I);
4909 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4910 IE->getOperand(2), Q);
4913 case Instruction::ExtractValue: {
4914 auto *EVI = cast<ExtractValueInst>(I);
4915 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4916 EVI->getIndices(), Q);
4919 case Instruction::ExtractElement: {
4920 auto *EEI = cast<ExtractElementInst>(I);
4921 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4922 EEI->getIndexOperand(), Q);
4925 case Instruction::ShuffleVector: {
4926 auto *SVI = cast<ShuffleVectorInst>(I);
4927 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4928 SVI->getMask(), SVI->getType(), Q);
4931 case Instruction::PHI:
4932 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4934 case Instruction::Call: {
4935 CallSite CS(cast<CallInst>(I));
4936 Result = SimplifyCall(CS, Q);
4939 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4940 #include "llvm/IR/Instruction.def"
4941 #undef HANDLE_CAST_INST
4943 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4945 case Instruction::Alloca:
4946 // No simplifications for Alloca and it can't be constant folded.
4951 // In general, it is possible for computeKnownBits to determine all bits in a
4952 // value even when the operands are not all constants.
4953 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4954 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4955 if (Known.isConstant())
4956 Result = ConstantInt::get(I->getType(), Known.getConstant());
4959 /// If called on unreachable code, the above logic may report that the
4960 /// instruction simplified to itself. Make life easier for users by
4961 /// detecting that case here, returning a safe value instead.
4962 return Result == I ? UndefValue::get(I->getType()) : Result;
4965 /// Implementation of recursive simplification through an instruction's
4968 /// This is the common implementation of the recursive simplification routines.
4969 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4970 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4971 /// instructions to process and attempt to simplify it using
4972 /// InstructionSimplify.
4974 /// This routine returns 'true' only when *it* simplifies something. The passed
4975 /// in simplified value does not count toward this.
4976 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4977 const TargetLibraryInfo *TLI,
4978 const DominatorTree *DT,
4979 AssumptionCache *AC) {
4980 bool Simplified = false;
4981 SmallSetVector<Instruction *, 8> Worklist;
4982 const DataLayout &DL = I->getModule()->getDataLayout();
4984 // If we have an explicit value to collapse to, do that round of the
4985 // simplification loop by hand initially.
4987 for (User *U : I->users())
4989 Worklist.insert(cast<Instruction>(U));
4991 // Replace the instruction with its simplified value.
4992 I->replaceAllUsesWith(SimpleV);
4994 // Gracefully handle edge cases where the instruction is not wired into any
4996 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4997 !I->mayHaveSideEffects())
4998 I->eraseFromParent();
5003 // Note that we must test the size on each iteration, the worklist can grow.
5004 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
5007 // See if this instruction simplifies.
5008 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
5014 // Stash away all the uses of the old instruction so we can check them for
5015 // recursive simplifications after a RAUW. This is cheaper than checking all
5016 // uses of To on the recursive step in most cases.
5017 for (User *U : I->users())
5018 Worklist.insert(cast<Instruction>(U));
5020 // Replace the instruction with its simplified value.
5021 I->replaceAllUsesWith(SimpleV);
5023 // Gracefully handle edge cases where the instruction is not wired into any
5025 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
5026 !I->mayHaveSideEffects())
5027 I->eraseFromParent();
5032 bool llvm::recursivelySimplifyInstruction(Instruction *I,
5033 const TargetLibraryInfo *TLI,
5034 const DominatorTree *DT,
5035 AssumptionCache *AC) {
5036 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
5039 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
5040 const TargetLibraryInfo *TLI,
5041 const DominatorTree *DT,
5042 AssumptionCache *AC) {
5043 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5044 assert(SimpleV && "Must provide a simplified value.");
5045 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
5049 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
5050 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
5051 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
5052 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
5053 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
5054 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5055 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5056 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5059 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5060 const DataLayout &DL) {
5061 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5064 template <class T, class... TArgs>
5065 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5067 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5068 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5069 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5070 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5072 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,