1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements routines for folding instructions into simpler forms
11 // that do not require creating new instructions. This does constant folding
12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15 // simplified: This is usually true and assuming it simplifies the logic (if
16 // they have not been simplified then results are correct but maybe suboptimal).
18 //===----------------------------------------------------------------------===//
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/CmpInstAnalysis.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/Analysis/VectorUtils.h"
32 #include "llvm/IR/ConstantRange.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/GetElementPtrTypeIterator.h"
36 #include "llvm/IR/GlobalAlias.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Support/KnownBits.h"
43 using namespace llvm::PatternMatch;
45 #define DEBUG_TYPE "instsimplify"
47 enum { RecursionLimit = 3 };
49 STATISTIC(NumExpand, "Number of expansions");
50 STATISTIC(NumReassoc, "Number of reassociations");
52 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
53 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
55 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
56 const SimplifyQuery &, unsigned);
57 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
59 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
60 const SimplifyQuery &Q, unsigned MaxRecurse);
61 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
62 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 static Value *SimplifyCastInst(unsigned, Value *, Type *,
64 const SimplifyQuery &, unsigned);
66 /// For a boolean type or a vector of boolean type, return false or a vector
67 /// with every element false.
68 static Constant *getFalse(Type *Ty) {
69 return ConstantInt::getFalse(Ty);
72 /// For a boolean type or a vector of boolean type, return true or a vector
73 /// with every element true.
74 static Constant *getTrue(Type *Ty) {
75 return ConstantInt::getTrue(Ty);
78 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
79 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
81 CmpInst *Cmp = dyn_cast<CmpInst>(V);
84 CmpInst::Predicate CPred = Cmp->getPredicate();
85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
86 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
92 /// Does the given value dominate the specified phi node?
93 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
94 Instruction *I = dyn_cast<Instruction>(V);
96 // Arguments and constants dominate all instructions.
99 // If we are processing instructions (and/or basic blocks) that have not been
100 // fully added to a function, the parent nodes may still be null. Simply
101 // return the conservative answer in these cases.
102 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
105 // If we have a DominatorTree then do a precise test.
107 return DT->dominates(I, P);
109 // Otherwise, if the instruction is in the entry block and is not an invoke,
110 // then it obviously dominates all phi nodes.
111 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
118 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
119 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
120 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
121 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
122 /// Returns the simplified value, or null if no simplification was performed.
123 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
124 Instruction::BinaryOps OpcodeToExpand,
125 const SimplifyQuery &Q, unsigned MaxRecurse) {
126 // Recursion is always used, so bail out at once if we already hit the limit.
130 // Check whether the expression has the form "(A op' B) op C".
131 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
132 if (Op0->getOpcode() == OpcodeToExpand) {
133 // It does! Try turning it into "(A op C) op' (B op C)".
134 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
135 // Do "A op C" and "B op C" both simplify?
136 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
137 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
138 // They do! Return "L op' R" if it simplifies or is already available.
139 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
140 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
141 && L == B && R == A)) {
145 // Otherwise return "L op' R" if it simplifies.
146 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
153 // Check whether the expression has the form "A op (B op' C)".
154 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
155 if (Op1->getOpcode() == OpcodeToExpand) {
156 // It does! Try turning it into "(A op B) op' (A op C)".
157 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
158 // Do "A op B" and "A op C" both simplify?
159 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
160 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
161 // They do! Return "L op' R" if it simplifies or is already available.
162 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
163 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
164 && L == C && R == B)) {
168 // Otherwise return "L op' R" if it simplifies.
169 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
179 /// Generic simplifications for associative binary operations.
180 /// Returns the simpler value, or null if none was found.
181 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
182 Value *LHS, Value *RHS,
183 const SimplifyQuery &Q,
184 unsigned MaxRecurse) {
185 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
187 // Recursion is always used, so bail out at once if we already hit the limit.
191 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
192 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
194 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
195 if (Op0 && Op0->getOpcode() == Opcode) {
196 Value *A = Op0->getOperand(0);
197 Value *B = Op0->getOperand(1);
200 // Does "B op C" simplify?
201 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
202 // It does! Return "A op V" if it simplifies or is already available.
203 // If V equals B then "A op V" is just the LHS.
204 if (V == B) return LHS;
205 // Otherwise return "A op V" if it simplifies.
206 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
213 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
214 if (Op1 && Op1->getOpcode() == Opcode) {
216 Value *B = Op1->getOperand(0);
217 Value *C = Op1->getOperand(1);
219 // Does "A op B" simplify?
220 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
221 // It does! Return "V op C" if it simplifies or is already available.
222 // If V equals B then "V op C" is just the RHS.
223 if (V == B) return RHS;
224 // Otherwise return "V op C" if it simplifies.
225 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
232 // The remaining transforms require commutativity as well as associativity.
233 if (!Instruction::isCommutative(Opcode))
236 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
237 if (Op0 && Op0->getOpcode() == Opcode) {
238 Value *A = Op0->getOperand(0);
239 Value *B = Op0->getOperand(1);
242 // Does "C op A" simplify?
243 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
244 // It does! Return "V op B" if it simplifies or is already available.
245 // If V equals A then "V op B" is just the LHS.
246 if (V == A) return LHS;
247 // Otherwise return "V op B" if it simplifies.
248 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
255 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
256 if (Op1 && Op1->getOpcode() == Opcode) {
258 Value *B = Op1->getOperand(0);
259 Value *C = Op1->getOperand(1);
261 // Does "C op A" simplify?
262 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
263 // It does! Return "B op V" if it simplifies or is already available.
264 // If V equals C then "B op V" is just the RHS.
265 if (V == C) return RHS;
266 // Otherwise return "B op V" if it simplifies.
267 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
277 /// In the case of a binary operation with a select instruction as an operand,
278 /// try to simplify the binop by seeing whether evaluating it on both branches
279 /// of the select results in the same value. Returns the common value if so,
280 /// otherwise returns null.
281 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
282 Value *RHS, const SimplifyQuery &Q,
283 unsigned MaxRecurse) {
284 // Recursion is always used, so bail out at once if we already hit the limit.
289 if (isa<SelectInst>(LHS)) {
290 SI = cast<SelectInst>(LHS);
292 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
293 SI = cast<SelectInst>(RHS);
296 // Evaluate the BinOp on the true and false branches of the select.
300 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
301 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
303 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
304 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
307 // If they simplified to the same value, then return the common value.
308 // If they both failed to simplify then return null.
312 // If one branch simplified to undef, return the other one.
313 if (TV && isa<UndefValue>(TV))
315 if (FV && isa<UndefValue>(FV))
318 // If applying the operation did not change the true and false select values,
319 // then the result of the binop is the select itself.
320 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
323 // If one branch simplified and the other did not, and the simplified
324 // value is equal to the unsimplified one, return the simplified value.
325 // For example, select (cond, X, X & Z) & Z -> X & Z.
326 if ((FV && !TV) || (TV && !FV)) {
327 // Check that the simplified value has the form "X op Y" where "op" is the
328 // same as the original operation.
329 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
330 if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
331 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
332 // We already know that "op" is the same as for the simplified value. See
333 // if the operands match too. If so, return the simplified value.
334 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
335 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
336 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
337 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
338 Simplified->getOperand(1) == UnsimplifiedRHS)
340 if (Simplified->isCommutative() &&
341 Simplified->getOperand(1) == UnsimplifiedLHS &&
342 Simplified->getOperand(0) == UnsimplifiedRHS)
350 /// In the case of a comparison with a select instruction, try to simplify the
351 /// comparison by seeing whether both branches of the select result in the same
352 /// value. Returns the common value if so, otherwise returns null.
353 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
354 Value *RHS, const SimplifyQuery &Q,
355 unsigned MaxRecurse) {
356 // Recursion is always used, so bail out at once if we already hit the limit.
360 // Make sure the select is on the LHS.
361 if (!isa<SelectInst>(LHS)) {
363 Pred = CmpInst::getSwappedPredicate(Pred);
365 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
366 SelectInst *SI = cast<SelectInst>(LHS);
367 Value *Cond = SI->getCondition();
368 Value *TV = SI->getTrueValue();
369 Value *FV = SI->getFalseValue();
371 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
372 // Does "cmp TV, RHS" simplify?
373 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
375 // It not only simplified, it simplified to the select condition. Replace
377 TCmp = getTrue(Cond->getType());
379 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
380 // condition then we can replace it with 'true'. Otherwise give up.
381 if (!isSameCompare(Cond, Pred, TV, RHS))
383 TCmp = getTrue(Cond->getType());
386 // Does "cmp FV, RHS" simplify?
387 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
389 // It not only simplified, it simplified to the select condition. Replace
391 FCmp = getFalse(Cond->getType());
393 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
394 // condition then we can replace it with 'false'. Otherwise give up.
395 if (!isSameCompare(Cond, Pred, FV, RHS))
397 FCmp = getFalse(Cond->getType());
400 // If both sides simplified to the same value, then use it as the result of
401 // the original comparison.
405 // The remaining cases only make sense if the select condition has the same
406 // type as the result of the comparison, so bail out if this is not so.
407 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
409 // If the false value simplified to false, then the result of the compare
410 // is equal to "Cond && TCmp". This also catches the case when the false
411 // value simplified to false and the true value to true, returning "Cond".
412 if (match(FCmp, m_Zero()))
413 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
415 // If the true value simplified to true, then the result of the compare
416 // is equal to "Cond || FCmp".
417 if (match(TCmp, m_One()))
418 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
420 // Finally, if the false value simplified to true and the true value to
421 // false, then the result of the compare is equal to "!Cond".
422 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
424 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
431 /// In the case of a binary operation with an operand that is a PHI instruction,
432 /// try to simplify the binop by seeing whether evaluating it on the incoming
433 /// phi values yields the same result for every value. If so returns the common
434 /// value, otherwise returns null.
435 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
436 Value *RHS, const SimplifyQuery &Q,
437 unsigned MaxRecurse) {
438 // Recursion is always used, so bail out at once if we already hit the limit.
443 if (isa<PHINode>(LHS)) {
444 PI = cast<PHINode>(LHS);
445 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
446 if (!ValueDominatesPHI(RHS, PI, Q.DT))
449 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
450 PI = cast<PHINode>(RHS);
451 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
452 if (!ValueDominatesPHI(LHS, PI, Q.DT))
456 // Evaluate the BinOp on the incoming phi values.
457 Value *CommonValue = nullptr;
458 for (Value *Incoming : PI->incoming_values()) {
459 // If the incoming value is the phi node itself, it can safely be skipped.
460 if (Incoming == PI) continue;
461 Value *V = PI == LHS ?
462 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
463 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
464 // If the operation failed to simplify, or simplified to a different value
465 // to previously, then give up.
466 if (!V || (CommonValue && V != CommonValue))
474 /// In the case of a comparison with a PHI instruction, try to simplify the
475 /// comparison by seeing whether comparing with all of the incoming phi values
476 /// yields the same result every time. If so returns the common result,
477 /// otherwise returns null.
478 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
479 const SimplifyQuery &Q, unsigned MaxRecurse) {
480 // Recursion is always used, so bail out at once if we already hit the limit.
484 // Make sure the phi is on the LHS.
485 if (!isa<PHINode>(LHS)) {
487 Pred = CmpInst::getSwappedPredicate(Pred);
489 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
490 PHINode *PI = cast<PHINode>(LHS);
492 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
493 if (!ValueDominatesPHI(RHS, PI, Q.DT))
496 // Evaluate the BinOp on the incoming phi values.
497 Value *CommonValue = nullptr;
498 for (Value *Incoming : PI->incoming_values()) {
499 // If the incoming value is the phi node itself, it can safely be skipped.
500 if (Incoming == PI) continue;
501 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
502 // If the operation failed to simplify, or simplified to a different value
503 // to previously, then give up.
504 if (!V || (CommonValue && V != CommonValue))
512 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
513 Value *&Op0, Value *&Op1,
514 const SimplifyQuery &Q) {
515 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
516 if (auto *CRHS = dyn_cast<Constant>(Op1))
517 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
519 // Canonicalize the constant to the RHS if this is a commutative operation.
520 if (Instruction::isCommutative(Opcode))
526 /// Given operands for an Add, see if we can fold the result.
527 /// If not, this returns null.
528 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
529 const SimplifyQuery &Q, unsigned MaxRecurse) {
530 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
533 // X + undef -> undef
534 if (match(Op1, m_Undef()))
538 if (match(Op1, m_Zero()))
545 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
546 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
549 // X + ~X -> -1 since ~X = -X-1
550 Type *Ty = Op0->getType();
551 if (match(Op0, m_Not(m_Specific(Op1))) ||
552 match(Op1, m_Not(m_Specific(Op0))))
553 return Constant::getAllOnesValue(Ty);
555 // add nsw/nuw (xor Y, signmask), signmask --> Y
556 // The no-wrapping add guarantees that the top bit will be set by the add.
557 // Therefore, the xor must be clearing the already set sign bit of Y.
558 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
559 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
563 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
564 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
567 // Try some generic simplifications for associative operations.
568 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
572 // Threading Add over selects and phi nodes is pointless, so don't bother.
573 // Threading over the select in "A + select(cond, B, C)" means evaluating
574 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
575 // only if B and C are equal. If B and C are equal then (since we assume
576 // that operands have already been simplified) "select(cond, B, C)" should
577 // have been simplified to the common value of B and C already. Analysing
578 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
579 // for threading over phi nodes.
584 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
585 const SimplifyQuery &Query) {
586 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
589 /// \brief Compute the base pointer and cumulative constant offsets for V.
591 /// This strips all constant offsets off of V, leaving it the base pointer, and
592 /// accumulates the total constant offset applied in the returned constant. It
593 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
594 /// no constant offsets applied.
596 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
597 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
599 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
600 bool AllowNonInbounds = false) {
601 assert(V->getType()->isPtrOrPtrVectorTy());
603 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
604 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
606 // Even though we don't look through PHI nodes, we could be called on an
607 // instruction in an unreachable block, which may be on a cycle.
608 SmallPtrSet<Value *, 4> Visited;
611 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
612 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
613 !GEP->accumulateConstantOffset(DL, Offset))
615 V = GEP->getPointerOperand();
616 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
617 V = cast<Operator>(V)->getOperand(0);
618 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
619 if (GA->isInterposable())
621 V = GA->getAliasee();
623 if (auto CS = CallSite(V))
624 if (Value *RV = CS.getReturnedArgOperand()) {
630 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
631 } while (Visited.insert(V).second);
633 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
634 if (V->getType()->isVectorTy())
635 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
640 /// \brief Compute the constant difference between two pointer values.
641 /// If the difference is not a constant, returns zero.
642 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
644 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
645 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
647 // If LHS and RHS are not related via constant offsets to the same base
648 // value, there is nothing we can do here.
652 // Otherwise, the difference of LHS - RHS can be computed as:
654 // = (LHSOffset + Base) - (RHSOffset + Base)
655 // = LHSOffset - RHSOffset
656 return ConstantExpr::getSub(LHSOffset, RHSOffset);
659 /// Given operands for a Sub, see if we can fold the result.
660 /// If not, this returns null.
661 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
662 const SimplifyQuery &Q, unsigned MaxRecurse) {
663 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
666 // X - undef -> undef
667 // undef - X -> undef
668 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
669 return UndefValue::get(Op0->getType());
672 if (match(Op1, m_Zero()))
677 return Constant::getNullValue(Op0->getType());
679 // Is this a negation?
680 if (match(Op0, m_Zero())) {
681 // 0 - X -> 0 if the sub is NUW.
685 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
686 if (Known.Zero.isMaxSignedValue()) {
687 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
688 // Op1 must be 0 because negating the minimum signed value is undefined.
692 // 0 - X -> X if X is 0 or the minimum signed value.
697 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
698 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
699 Value *X = nullptr, *Y = nullptr, *Z = Op1;
700 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
701 // See if "V === Y - Z" simplifies.
702 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
703 // It does! Now see if "X + V" simplifies.
704 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
705 // It does, we successfully reassociated!
709 // See if "V === X - Z" simplifies.
710 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
711 // It does! Now see if "Y + V" simplifies.
712 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
713 // It does, we successfully reassociated!
719 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
720 // For example, X - (X + 1) -> -1
722 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
723 // See if "V === X - Y" simplifies.
724 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
725 // It does! Now see if "V - Z" simplifies.
726 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
727 // It does, we successfully reassociated!
731 // See if "V === X - Z" simplifies.
732 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
733 // It does! Now see if "V - Y" simplifies.
734 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
735 // It does, we successfully reassociated!
741 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
742 // For example, X - (X - Y) -> Y.
744 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
745 // See if "V === Z - X" simplifies.
746 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
747 // It does! Now see if "V + Y" simplifies.
748 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
749 // It does, we successfully reassociated!
754 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
755 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
756 match(Op1, m_Trunc(m_Value(Y))))
757 if (X->getType() == Y->getType())
758 // See if "V === X - Y" simplifies.
759 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
760 // It does! Now see if "trunc V" simplifies.
761 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
763 // It does, return the simplified "trunc V".
766 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
767 if (match(Op0, m_PtrToInt(m_Value(X))) &&
768 match(Op1, m_PtrToInt(m_Value(Y))))
769 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
770 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
773 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
774 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
777 // Threading Sub over selects and phi nodes is pointless, so don't bother.
778 // Threading over the select in "A - select(cond, B, C)" means evaluating
779 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
780 // only if B and C are equal. If B and C are equal then (since we assume
781 // that operands have already been simplified) "select(cond, B, C)" should
782 // have been simplified to the common value of B and C already. Analysing
783 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
784 // for threading over phi nodes.
789 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
790 const SimplifyQuery &Q) {
791 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
794 /// Given operands for a Mul, see if we can fold the result.
795 /// If not, this returns null.
796 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
797 unsigned MaxRecurse) {
798 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
802 if (match(Op1, m_Undef()))
803 return Constant::getNullValue(Op0->getType());
806 if (match(Op1, m_Zero()))
810 if (match(Op1, m_One()))
813 // (X / Y) * Y -> X if the division is exact.
815 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
816 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
820 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
821 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
824 // Try some generic simplifications for associative operations.
825 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
829 // Mul distributes over Add. Try some generic simplifications based on this.
830 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
834 // If the operation is with the result of a select instruction, check whether
835 // operating on either branch of the select always yields the same value.
836 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
837 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
841 // If the operation is with the result of a phi instruction, check whether
842 // operating on all incoming values of the phi always yields the same value.
843 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
844 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
851 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
852 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
855 /// Check for common or similar folds of integer division or integer remainder.
856 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
857 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
858 Type *Ty = Op0->getType();
860 // X / undef -> undef
861 // X % undef -> undef
862 if (match(Op1, m_Undef()))
867 // We don't need to preserve faults!
868 if (match(Op1, m_Zero()))
869 return UndefValue::get(Ty);
871 // If any element of a constant divisor vector is zero or undef, the whole op
873 auto *Op1C = dyn_cast<Constant>(Op1);
874 if (Op1C && Ty->isVectorTy()) {
875 unsigned NumElts = Ty->getVectorNumElements();
876 for (unsigned i = 0; i != NumElts; ++i) {
877 Constant *Elt = Op1C->getAggregateElement(i);
878 if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
879 return UndefValue::get(Ty);
885 if (match(Op0, m_Undef()))
886 return Constant::getNullValue(Ty);
890 if (match(Op0, m_Zero()))
896 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
900 // If this is a boolean op (single-bit element type), we can't have
901 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
902 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1))
903 return IsDiv ? Op0 : Constant::getNullValue(Ty);
908 /// Given a predicate and two operands, return true if the comparison is true.
909 /// This is a helper for div/rem simplification where we return some other value
910 /// when we can prove a relationship between the operands.
911 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
912 const SimplifyQuery &Q, unsigned MaxRecurse) {
913 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
914 Constant *C = dyn_cast_or_null<Constant>(V);
915 return (C && C->isAllOnesValue());
918 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
919 /// to simplify X % Y to X.
920 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
921 unsigned MaxRecurse, bool IsSigned) {
922 // Recursion is always used, so bail out at once if we already hit the limit.
929 // We require that 1 operand is a simple constant. That could be extended to
930 // 2 variables if we computed the sign bit for each.
932 // Make sure that a constant is not the minimum signed value because taking
933 // the abs() of that is undefined.
934 Type *Ty = X->getType();
936 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
937 // Is the variable divisor magnitude always greater than the constant
938 // dividend magnitude?
939 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
940 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
941 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
942 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
943 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
946 if (match(Y, m_APInt(C))) {
947 // Special-case: we can't take the abs() of a minimum signed value. If
948 // that's the divisor, then all we have to do is prove that the dividend
949 // is also not the minimum signed value.
950 if (C->isMinSignedValue())
951 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
953 // Is the variable dividend magnitude always less than the constant
954 // divisor magnitude?
955 // |X| < |C| --> X > -abs(C) and X < abs(C)
956 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
957 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
958 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
959 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
965 // IsSigned == false.
966 // Is the dividend unsigned less than the divisor?
967 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
970 /// These are simplifications common to SDiv and UDiv.
971 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
972 const SimplifyQuery &Q, unsigned MaxRecurse) {
973 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
976 if (Value *V = simplifyDivRem(Op0, Op1, true))
979 bool IsSigned = Opcode == Instruction::SDiv;
981 // (X * Y) / Y -> X if the multiplication does not overflow.
983 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
984 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
985 // If the Mul does not overflow, then we are good to go.
986 if ((IsSigned && Mul->hasNoSignedWrap()) ||
987 (!IsSigned && Mul->hasNoUnsignedWrap()))
989 // If X has the form X = A / Y, then X * Y cannot overflow.
990 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
991 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
995 // (X rem Y) / Y -> 0
996 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
997 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
998 return Constant::getNullValue(Op0->getType());
1000 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1001 ConstantInt *C1, *C2;
1002 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1003 match(Op1, m_ConstantInt(C2))) {
1005 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1007 return Constant::getNullValue(Op0->getType());
1010 // If the operation is with the result of a select instruction, check whether
1011 // operating on either branch of the select always yields the same value.
1012 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1013 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1016 // If the operation is with the result of a phi instruction, check whether
1017 // operating on all incoming values of the phi always yields the same value.
1018 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1019 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1022 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1023 return Constant::getNullValue(Op0->getType());
1028 /// These are simplifications common to SRem and URem.
1029 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1030 const SimplifyQuery &Q, unsigned MaxRecurse) {
1031 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1034 if (Value *V = simplifyDivRem(Op0, Op1, false))
1037 // (X % Y) % Y -> X % Y
1038 if ((Opcode == Instruction::SRem &&
1039 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1040 (Opcode == Instruction::URem &&
1041 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1044 // (X << Y) % X -> 0
1045 if ((Opcode == Instruction::SRem &&
1046 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1047 (Opcode == Instruction::URem &&
1048 match(Op0, m_NUWShl(m_Specific(Op1), m_Value()))))
1049 return Constant::getNullValue(Op0->getType());
1051 // If the operation is with the result of a select instruction, check whether
1052 // operating on either branch of the select always yields the same value.
1053 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1054 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1057 // If the operation is with the result of a phi instruction, check whether
1058 // operating on all incoming values of the phi always yields the same value.
1059 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1060 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1063 // If X / Y == 0, then X % Y == X.
1064 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1070 /// Given operands for an SDiv, see if we can fold the result.
1071 /// If not, this returns null.
1072 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1073 unsigned MaxRecurse) {
1074 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1077 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1078 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1081 /// Given operands for a UDiv, see if we can fold the result.
1082 /// If not, this returns null.
1083 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1084 unsigned MaxRecurse) {
1085 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1088 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1089 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1092 /// Given operands for an SRem, see if we can fold the result.
1093 /// If not, this returns null.
1094 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1095 unsigned MaxRecurse) {
1096 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1099 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1100 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1103 /// Given operands for a URem, see if we can fold the result.
1104 /// If not, this returns null.
1105 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1106 unsigned MaxRecurse) {
1107 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1110 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1111 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1114 /// Returns true if a shift by \c Amount always yields undef.
1115 static bool isUndefShift(Value *Amount) {
1116 Constant *C = dyn_cast<Constant>(Amount);
1120 // X shift by undef -> undef because it may shift by the bitwidth.
1121 if (isa<UndefValue>(C))
1124 // Shifting by the bitwidth or more is undefined.
1125 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1126 if (CI->getValue().getLimitedValue() >=
1127 CI->getType()->getScalarSizeInBits())
1130 // If all lanes of a vector shift are undefined the whole shift is.
1131 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1132 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1133 if (!isUndefShift(C->getAggregateElement(I)))
1141 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1142 /// If not, this returns null.
1143 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1144 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1145 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1148 // 0 shift by X -> 0
1149 if (match(Op0, m_Zero()))
1152 // X shift by 0 -> X
1153 if (match(Op1, m_Zero()))
1156 // Fold undefined shifts.
1157 if (isUndefShift(Op1))
1158 return UndefValue::get(Op0->getType());
1160 // If the operation is with the result of a select instruction, check whether
1161 // operating on either branch of the select always yields the same value.
1162 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1163 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1166 // If the operation is with the result of a phi instruction, check whether
1167 // operating on all incoming values of the phi always yields the same value.
1168 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1169 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1172 // If any bits in the shift amount make that value greater than or equal to
1173 // the number of bits in the type, the shift is undefined.
1174 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1175 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1176 return UndefValue::get(Op0->getType());
1178 // If all valid bits in the shift amount are known zero, the first operand is
1180 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1181 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1187 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1188 /// fold the result. If not, this returns null.
1189 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1190 Value *Op1, bool isExact, const SimplifyQuery &Q,
1191 unsigned MaxRecurse) {
1192 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1197 return Constant::getNullValue(Op0->getType());
1200 // undef >> X -> undef (if it's exact)
1201 if (match(Op0, m_Undef()))
1202 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1204 // The low bit cannot be shifted out of an exact shift if it is set.
1206 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1207 if (Op0Known.One[0])
1214 /// Given operands for an Shl, see if we can fold the result.
1215 /// If not, this returns null.
1216 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1217 const SimplifyQuery &Q, unsigned MaxRecurse) {
1218 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1222 // undef << X -> undef if (if it's NSW/NUW)
1223 if (match(Op0, m_Undef()))
1224 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1226 // (X >> A) << A -> X
1228 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1233 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1234 const SimplifyQuery &Q) {
1235 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1238 /// Given operands for an LShr, see if we can fold the result.
1239 /// If not, this returns null.
1240 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1241 const SimplifyQuery &Q, unsigned MaxRecurse) {
1242 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1246 // (X << A) >> A -> X
1248 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1254 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1255 const SimplifyQuery &Q) {
1256 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1259 /// Given operands for an AShr, see if we can fold the result.
1260 /// If not, this returns null.
1261 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1262 const SimplifyQuery &Q, unsigned MaxRecurse) {
1263 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1267 // all ones >>a X -> -1
1268 // Do not return Op0 because it may contain undef elements if it's a vector.
1269 if (match(Op0, m_AllOnes()))
1270 return Constant::getAllOnesValue(Op0->getType());
1272 // (X << A) >> A -> X
1274 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1277 // Arithmetic shifting an all-sign-bit value is a no-op.
1278 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1279 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1285 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1286 const SimplifyQuery &Q) {
1287 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1290 /// Commuted variants are assumed to be handled by calling this function again
1291 /// with the parameters swapped.
1292 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1293 ICmpInst *UnsignedICmp, bool IsAnd) {
1296 ICmpInst::Predicate EqPred;
1297 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1298 !ICmpInst::isEquality(EqPred))
1301 ICmpInst::Predicate UnsignedPred;
1302 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1303 ICmpInst::isUnsigned(UnsignedPred))
1305 else if (match(UnsignedICmp,
1306 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1307 ICmpInst::isUnsigned(UnsignedPred))
1308 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1312 // X < Y && Y != 0 --> X < Y
1313 // X < Y || Y != 0 --> Y != 0
1314 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1315 return IsAnd ? UnsignedICmp : ZeroICmp;
1317 // X >= Y || Y != 0 --> true
1318 // X >= Y || Y == 0 --> X >= Y
1319 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1320 if (EqPred == ICmpInst::ICMP_NE)
1321 return getTrue(UnsignedICmp->getType());
1322 return UnsignedICmp;
1325 // X < Y && Y == 0 --> false
1326 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1328 return getFalse(UnsignedICmp->getType());
1333 /// Commuted variants are assumed to be handled by calling this function again
1334 /// with the parameters swapped.
1335 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1336 ICmpInst::Predicate Pred0, Pred1;
1338 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1339 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1342 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1343 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1344 // can eliminate Op1 from this 'and'.
1345 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1348 // Check for any combination of predicates that are guaranteed to be disjoint.
1349 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1350 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1351 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1352 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1353 return getFalse(Op0->getType());
1358 /// Commuted variants are assumed to be handled by calling this function again
1359 /// with the parameters swapped.
1360 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1361 ICmpInst::Predicate Pred0, Pred1;
1363 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1364 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1367 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1368 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1369 // can eliminate Op0 from this 'or'.
1370 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1373 // Check for any combination of predicates that cover the entire range of
1375 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1376 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1377 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1378 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1379 return getTrue(Op0->getType());
1384 /// Test if a pair of compares with a shared operand and 2 constants has an
1385 /// empty set intersection, full set union, or if one compare is a superset of
1387 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1389 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1390 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1393 const APInt *C0, *C1;
1394 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1395 !match(Cmp1->getOperand(1), m_APInt(C1)))
1398 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1399 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1401 // For and-of-compares, check if the intersection is empty:
1402 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1403 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1404 return getFalse(Cmp0->getType());
1406 // For or-of-compares, check if the union is full:
1407 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1408 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1409 return getTrue(Cmp0->getType());
1411 // Is one range a superset of the other?
1412 // If this is and-of-compares, take the smaller set:
1413 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1414 // If this is or-of-compares, take the larger set:
1415 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1416 if (Range0.contains(Range1))
1417 return IsAnd ? Cmp1 : Cmp0;
1418 if (Range1.contains(Range0))
1419 return IsAnd ? Cmp0 : Cmp1;
1424 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1426 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1427 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1428 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1431 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1434 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1435 Value *X = Cmp0->getOperand(0);
1436 Value *Y = Cmp1->getOperand(0);
1438 // If one of the compares is a masked version of a (not) null check, then
1439 // that compare implies the other, so we eliminate the other. Optionally, look
1440 // through a pointer-to-int cast to match a null check of a pointer type.
1442 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1443 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1444 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1445 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1446 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1447 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1450 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1451 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1452 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1453 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1454 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1455 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1461 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1462 // (icmp (add V, C0), C1) & (icmp V, C0)
1463 ICmpInst::Predicate Pred0, Pred1;
1464 const APInt *C0, *C1;
1466 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1469 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1472 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1473 if (AddInst->getOperand(1) != Op1->getOperand(1))
1476 Type *ITy = Op0->getType();
1477 bool isNSW = AddInst->hasNoSignedWrap();
1478 bool isNUW = AddInst->hasNoUnsignedWrap();
1480 const APInt Delta = *C1 - *C0;
1481 if (C0->isStrictlyPositive()) {
1483 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1484 return getFalse(ITy);
1485 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1486 return getFalse(ITy);
1489 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1490 return getFalse(ITy);
1491 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1492 return getFalse(ITy);
1495 if (C0->getBoolValue() && isNUW) {
1497 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1498 return getFalse(ITy);
1500 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1501 return getFalse(ITy);
1507 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1508 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1510 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1513 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1515 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1518 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1521 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1524 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1526 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1532 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1533 // (icmp (add V, C0), C1) | (icmp V, C0)
1534 ICmpInst::Predicate Pred0, Pred1;
1535 const APInt *C0, *C1;
1537 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1540 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1543 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1544 if (AddInst->getOperand(1) != Op1->getOperand(1))
1547 Type *ITy = Op0->getType();
1548 bool isNSW = AddInst->hasNoSignedWrap();
1549 bool isNUW = AddInst->hasNoUnsignedWrap();
1551 const APInt Delta = *C1 - *C0;
1552 if (C0->isStrictlyPositive()) {
1554 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1555 return getTrue(ITy);
1556 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1557 return getTrue(ITy);
1560 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1561 return getTrue(ITy);
1562 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1563 return getTrue(ITy);
1566 if (C0->getBoolValue() && isNUW) {
1568 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1569 return getTrue(ITy);
1571 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1572 return getTrue(ITy);
1578 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1579 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1581 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1584 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1586 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1589 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1592 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1595 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1597 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1603 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1604 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1605 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1606 if (LHS0->getType() != RHS0->getType())
1609 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1610 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1611 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1612 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1613 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1614 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1615 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1616 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1617 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1618 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1619 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1620 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1621 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1624 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1625 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1626 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1627 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1628 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1629 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1630 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1631 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1632 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1633 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1640 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1641 // Look through casts of the 'and' operands to find compares.
1642 auto *Cast0 = dyn_cast<CastInst>(Op0);
1643 auto *Cast1 = dyn_cast<CastInst>(Op1);
1644 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1645 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1646 Op0 = Cast0->getOperand(0);
1647 Op1 = Cast1->getOperand(0);
1651 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1652 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1654 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1655 simplifyOrOfICmps(ICmp0, ICmp1);
1657 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1658 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1660 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1667 // If we looked through casts, we can only handle a constant simplification
1668 // because we are not allowed to create a cast instruction here.
1669 if (auto *C = dyn_cast<Constant>(V))
1670 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1675 /// Given operands for an And, see if we can fold the result.
1676 /// If not, this returns null.
1677 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1678 unsigned MaxRecurse) {
1679 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1683 if (match(Op1, m_Undef()))
1684 return Constant::getNullValue(Op0->getType());
1691 if (match(Op1, m_Zero()))
1695 if (match(Op1, m_AllOnes()))
1698 // A & ~A = ~A & A = 0
1699 if (match(Op0, m_Not(m_Specific(Op1))) ||
1700 match(Op1, m_Not(m_Specific(Op0))))
1701 return Constant::getNullValue(Op0->getType());
1704 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1708 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1711 // A mask that only clears known zeros of a shifted value is a no-op.
1715 if (match(Op1, m_APInt(Mask))) {
1716 // If all bits in the inverted and shifted mask are clear:
1717 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1718 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1719 (~(*Mask)).lshr(*ShAmt).isNullValue())
1722 // If all bits in the inverted and shifted mask are clear:
1723 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1724 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1725 (~(*Mask)).shl(*ShAmt).isNullValue())
1729 // A & (-A) = A if A is a power of two or zero.
1730 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1731 match(Op1, m_Neg(m_Specific(Op0)))) {
1732 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1735 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1740 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1743 // Try some generic simplifications for associative operations.
1744 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1748 // And distributes over Or. Try some generic simplifications based on this.
1749 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1753 // And distributes over Xor. Try some generic simplifications based on this.
1754 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1758 // If the operation is with the result of a select instruction, check whether
1759 // operating on either branch of the select always yields the same value.
1760 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1761 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1765 // If the operation is with the result of a phi instruction, check whether
1766 // operating on all incoming values of the phi always yields the same value.
1767 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1768 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1775 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1776 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1779 /// Given operands for an Or, see if we can fold the result.
1780 /// If not, this returns null.
1781 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1782 unsigned MaxRecurse) {
1783 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1788 // Do not return Op1 because it may contain undef elements if it's a vector.
1789 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1790 return Constant::getAllOnesValue(Op0->getType());
1794 if (Op0 == Op1 || match(Op1, m_Zero()))
1797 // A | ~A = ~A | A = -1
1798 if (match(Op0, m_Not(m_Specific(Op1))) ||
1799 match(Op1, m_Not(m_Specific(Op0))))
1800 return Constant::getAllOnesValue(Op0->getType());
1803 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1807 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1810 // ~(A & ?) | A = -1
1811 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1812 return Constant::getAllOnesValue(Op1->getType());
1814 // A | ~(A & ?) = -1
1815 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1816 return Constant::getAllOnesValue(Op0->getType());
1819 // (A & ~B) | (A ^ B) -> (A ^ B)
1820 // (~B & A) | (A ^ B) -> (A ^ B)
1821 // (A & ~B) | (B ^ A) -> (B ^ A)
1822 // (~B & A) | (B ^ A) -> (B ^ A)
1823 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1824 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1825 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1828 // Commute the 'or' operands.
1829 // (A ^ B) | (A & ~B) -> (A ^ B)
1830 // (A ^ B) | (~B & A) -> (A ^ B)
1831 // (B ^ A) | (A & ~B) -> (B ^ A)
1832 // (B ^ A) | (~B & A) -> (B ^ A)
1833 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1834 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1835 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1838 // (A & B) | (~A ^ B) -> (~A ^ B)
1839 // (B & A) | (~A ^ B) -> (~A ^ B)
1840 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1841 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1842 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1843 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1844 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1847 // (~A ^ B) | (A & B) -> (~A ^ B)
1848 // (~A ^ B) | (B & A) -> (~A ^ B)
1849 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1850 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1851 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1852 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1853 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1856 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1859 // Try some generic simplifications for associative operations.
1860 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1864 // Or distributes over And. Try some generic simplifications based on this.
1865 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1869 // If the operation is with the result of a select instruction, check whether
1870 // operating on either branch of the select always yields the same value.
1871 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1872 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1876 // (A & C1)|(B & C2)
1877 const APInt *C1, *C2;
1878 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1879 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1881 // (A & C1)|(B & C2)
1882 // If we have: ((V + N) & C1) | (V & C2)
1883 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1884 // replace with V+N.
1886 if (C2->isMask() && // C2 == 0+1+
1887 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1888 // Add commutes, try both ways.
1889 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1892 // Or commutes, try both ways.
1894 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1895 // Add commutes, try both ways.
1896 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1902 // If the operation is with the result of a phi instruction, check whether
1903 // operating on all incoming values of the phi always yields the same value.
1904 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1905 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1911 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1912 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1915 /// Given operands for a Xor, see if we can fold the result.
1916 /// If not, this returns null.
1917 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1918 unsigned MaxRecurse) {
1919 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1922 // A ^ undef -> undef
1923 if (match(Op1, m_Undef()))
1927 if (match(Op1, m_Zero()))
1932 return Constant::getNullValue(Op0->getType());
1934 // A ^ ~A = ~A ^ A = -1
1935 if (match(Op0, m_Not(m_Specific(Op1))) ||
1936 match(Op1, m_Not(m_Specific(Op0))))
1937 return Constant::getAllOnesValue(Op0->getType());
1939 // Try some generic simplifications for associative operations.
1940 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1944 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1945 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1946 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1947 // only if B and C are equal. If B and C are equal then (since we assume
1948 // that operands have already been simplified) "select(cond, B, C)" should
1949 // have been simplified to the common value of B and C already. Analysing
1950 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1951 // for threading over phi nodes.
1956 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1957 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1961 static Type *GetCompareTy(Value *Op) {
1962 return CmpInst::makeCmpResultType(Op->getType());
1965 /// Rummage around inside V looking for something equivalent to the comparison
1966 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
1967 /// Helper function for analyzing max/min idioms.
1968 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1969 Value *LHS, Value *RHS) {
1970 SelectInst *SI = dyn_cast<SelectInst>(V);
1973 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1976 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1977 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1979 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1980 LHS == CmpRHS && RHS == CmpLHS)
1985 // A significant optimization not implemented here is assuming that alloca
1986 // addresses are not equal to incoming argument values. They don't *alias*,
1987 // as we say, but that doesn't mean they aren't equal, so we take a
1988 // conservative approach.
1990 // This is inspired in part by C++11 5.10p1:
1991 // "Two pointers of the same type compare equal if and only if they are both
1992 // null, both point to the same function, or both represent the same
1995 // This is pretty permissive.
1997 // It's also partly due to C11 6.5.9p6:
1998 // "Two pointers compare equal if and only if both are null pointers, both are
1999 // pointers to the same object (including a pointer to an object and a
2000 // subobject at its beginning) or function, both are pointers to one past the
2001 // last element of the same array object, or one is a pointer to one past the
2002 // end of one array object and the other is a pointer to the start of a
2003 // different array object that happens to immediately follow the first array
2004 // object in the address space.)
2006 // C11's version is more restrictive, however there's no reason why an argument
2007 // couldn't be a one-past-the-end value for a stack object in the caller and be
2008 // equal to the beginning of a stack object in the callee.
2010 // If the C and C++ standards are ever made sufficiently restrictive in this
2011 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2012 // this optimization.
2014 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2015 const DominatorTree *DT, CmpInst::Predicate Pred,
2016 AssumptionCache *AC, const Instruction *CxtI,
2017 Value *LHS, Value *RHS) {
2018 // First, skip past any trivial no-ops.
2019 LHS = LHS->stripPointerCasts();
2020 RHS = RHS->stripPointerCasts();
2022 // A non-null pointer is not equal to a null pointer.
2023 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2024 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2025 return ConstantInt::get(GetCompareTy(LHS),
2026 !CmpInst::isTrueWhenEqual(Pred));
2028 // We can only fold certain predicates on pointer comparisons.
2033 // Equality comaprisons are easy to fold.
2034 case CmpInst::ICMP_EQ:
2035 case CmpInst::ICMP_NE:
2038 // We can only handle unsigned relational comparisons because 'inbounds' on
2039 // a GEP only protects against unsigned wrapping.
2040 case CmpInst::ICMP_UGT:
2041 case CmpInst::ICMP_UGE:
2042 case CmpInst::ICMP_ULT:
2043 case CmpInst::ICMP_ULE:
2044 // However, we have to switch them to their signed variants to handle
2045 // negative indices from the base pointer.
2046 Pred = ICmpInst::getSignedPredicate(Pred);
2050 // Strip off any constant offsets so that we can reason about them.
2051 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2052 // here and compare base addresses like AliasAnalysis does, however there are
2053 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2054 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2055 // doesn't need to guarantee pointer inequality when it says NoAlias.
2056 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2057 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2059 // If LHS and RHS are related via constant offsets to the same base
2060 // value, we can replace it with an icmp which just compares the offsets.
2062 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2064 // Various optimizations for (in)equality comparisons.
2065 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2066 // Different non-empty allocations that exist at the same time have
2067 // different addresses (if the program can tell). Global variables always
2068 // exist, so they always exist during the lifetime of each other and all
2069 // allocas. Two different allocas usually have different addresses...
2071 // However, if there's an @llvm.stackrestore dynamically in between two
2072 // allocas, they may have the same address. It's tempting to reduce the
2073 // scope of the problem by only looking at *static* allocas here. That would
2074 // cover the majority of allocas while significantly reducing the likelihood
2075 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2076 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2077 // an entry block. Also, if we have a block that's not attached to a
2078 // function, we can't tell if it's "static" under the current definition.
2079 // Theoretically, this problem could be fixed by creating a new kind of
2080 // instruction kind specifically for static allocas. Such a new instruction
2081 // could be required to be at the top of the entry block, thus preventing it
2082 // from being subject to a @llvm.stackrestore. Instcombine could even
2083 // convert regular allocas into these special allocas. It'd be nifty.
2084 // However, until then, this problem remains open.
2086 // So, we'll assume that two non-empty allocas have different addresses
2089 // With all that, if the offsets are within the bounds of their allocations
2090 // (and not one-past-the-end! so we can't use inbounds!), and their
2091 // allocations aren't the same, the pointers are not equal.
2093 // Note that it's not necessary to check for LHS being a global variable
2094 // address, due to canonicalization and constant folding.
2095 if (isa<AllocaInst>(LHS) &&
2096 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2097 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2098 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2099 uint64_t LHSSize, RHSSize;
2100 if (LHSOffsetCI && RHSOffsetCI &&
2101 getObjectSize(LHS, LHSSize, DL, TLI) &&
2102 getObjectSize(RHS, RHSSize, DL, TLI)) {
2103 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2104 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2105 if (!LHSOffsetValue.isNegative() &&
2106 !RHSOffsetValue.isNegative() &&
2107 LHSOffsetValue.ult(LHSSize) &&
2108 RHSOffsetValue.ult(RHSSize)) {
2109 return ConstantInt::get(GetCompareTy(LHS),
2110 !CmpInst::isTrueWhenEqual(Pred));
2114 // Repeat the above check but this time without depending on DataLayout
2115 // or being able to compute a precise size.
2116 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2117 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2118 LHSOffset->isNullValue() &&
2119 RHSOffset->isNullValue())
2120 return ConstantInt::get(GetCompareTy(LHS),
2121 !CmpInst::isTrueWhenEqual(Pred));
2124 // Even if an non-inbounds GEP occurs along the path we can still optimize
2125 // equality comparisons concerning the result. We avoid walking the whole
2126 // chain again by starting where the last calls to
2127 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2128 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2129 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2131 return ConstantExpr::getICmp(Pred,
2132 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2133 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2135 // If one side of the equality comparison must come from a noalias call
2136 // (meaning a system memory allocation function), and the other side must
2137 // come from a pointer that cannot overlap with dynamically-allocated
2138 // memory within the lifetime of the current function (allocas, byval
2139 // arguments, globals), then determine the comparison result here.
2140 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2141 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2142 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2144 // Is the set of underlying objects all noalias calls?
2145 auto IsNAC = [](ArrayRef<Value *> Objects) {
2146 return all_of(Objects, isNoAliasCall);
2149 // Is the set of underlying objects all things which must be disjoint from
2150 // noalias calls. For allocas, we consider only static ones (dynamic
2151 // allocas might be transformed into calls to malloc not simultaneously
2152 // live with the compared-to allocation). For globals, we exclude symbols
2153 // that might be resolve lazily to symbols in another dynamically-loaded
2154 // library (and, thus, could be malloc'ed by the implementation).
2155 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2156 return all_of(Objects, [](Value *V) {
2157 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2158 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2159 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2160 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2161 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2162 !GV->isThreadLocal();
2163 if (const Argument *A = dyn_cast<Argument>(V))
2164 return A->hasByValAttr();
2169 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2170 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2171 return ConstantInt::get(GetCompareTy(LHS),
2172 !CmpInst::isTrueWhenEqual(Pred));
2174 // Fold comparisons for non-escaping pointer even if the allocation call
2175 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2176 // dynamic allocation call could be either of the operands.
2177 Value *MI = nullptr;
2178 if (isAllocLikeFn(LHS, TLI) &&
2179 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2181 else if (isAllocLikeFn(RHS, TLI) &&
2182 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2184 // FIXME: We should also fold the compare when the pointer escapes, but the
2185 // compare dominates the pointer escape
2186 if (MI && !PointerMayBeCaptured(MI, true, true))
2187 return ConstantInt::get(GetCompareTy(LHS),
2188 CmpInst::isFalseWhenEqual(Pred));
2195 /// Fold an icmp when its operands have i1 scalar type.
2196 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2197 Value *RHS, const SimplifyQuery &Q) {
2198 Type *ITy = GetCompareTy(LHS); // The return type.
2199 Type *OpTy = LHS->getType(); // The operand type.
2200 if (!OpTy->isIntOrIntVectorTy(1))
2203 // A boolean compared to true/false can be simplified in 14 out of the 20
2204 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2205 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2206 if (match(RHS, m_Zero())) {
2208 case CmpInst::ICMP_NE: // X != 0 -> X
2209 case CmpInst::ICMP_UGT: // X >u 0 -> X
2210 case CmpInst::ICMP_SLT: // X <s 0 -> X
2213 case CmpInst::ICMP_ULT: // X <u 0 -> false
2214 case CmpInst::ICMP_SGT: // X >s 0 -> false
2215 return getFalse(ITy);
2217 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2218 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2219 return getTrue(ITy);
2223 } else if (match(RHS, m_One())) {
2225 case CmpInst::ICMP_EQ: // X == 1 -> X
2226 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2227 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2230 case CmpInst::ICMP_UGT: // X >u 1 -> false
2231 case CmpInst::ICMP_SLT: // X <s -1 -> false
2232 return getFalse(ITy);
2234 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2235 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2236 return getTrue(ITy);
2245 case ICmpInst::ICMP_UGE:
2246 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2247 return getTrue(ITy);
2249 case ICmpInst::ICMP_SGE:
2250 /// For signed comparison, the values for an i1 are 0 and -1
2251 /// respectively. This maps into a truth table of:
2252 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2253 /// 0 | 0 | 1 (0 >= 0) | 1
2254 /// 0 | 1 | 1 (0 >= -1) | 1
2255 /// 1 | 0 | 0 (-1 >= 0) | 0
2256 /// 1 | 1 | 1 (-1 >= -1) | 1
2257 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2258 return getTrue(ITy);
2260 case ICmpInst::ICMP_ULE:
2261 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2262 return getTrue(ITy);
2269 /// Try hard to fold icmp with zero RHS because this is a common case.
2270 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2271 Value *RHS, const SimplifyQuery &Q) {
2272 if (!match(RHS, m_Zero()))
2275 Type *ITy = GetCompareTy(LHS); // The return type.
2278 llvm_unreachable("Unknown ICmp predicate!");
2279 case ICmpInst::ICMP_ULT:
2280 return getFalse(ITy);
2281 case ICmpInst::ICMP_UGE:
2282 return getTrue(ITy);
2283 case ICmpInst::ICMP_EQ:
2284 case ICmpInst::ICMP_ULE:
2285 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2286 return getFalse(ITy);
2288 case ICmpInst::ICMP_NE:
2289 case ICmpInst::ICMP_UGT:
2290 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2291 return getTrue(ITy);
2293 case ICmpInst::ICMP_SLT: {
2294 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2295 if (LHSKnown.isNegative())
2296 return getTrue(ITy);
2297 if (LHSKnown.isNonNegative())
2298 return getFalse(ITy);
2301 case ICmpInst::ICMP_SLE: {
2302 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2303 if (LHSKnown.isNegative())
2304 return getTrue(ITy);
2305 if (LHSKnown.isNonNegative() &&
2306 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2307 return getFalse(ITy);
2310 case ICmpInst::ICMP_SGE: {
2311 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2312 if (LHSKnown.isNegative())
2313 return getFalse(ITy);
2314 if (LHSKnown.isNonNegative())
2315 return getTrue(ITy);
2318 case ICmpInst::ICMP_SGT: {
2319 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2320 if (LHSKnown.isNegative())
2321 return getFalse(ITy);
2322 if (LHSKnown.isNonNegative() &&
2323 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2324 return getTrue(ITy);
2332 /// Many binary operators with a constant operand have an easy-to-compute
2333 /// range of outputs. This can be used to fold a comparison to always true or
2335 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2336 unsigned Width = Lower.getBitWidth();
2338 switch (BO.getOpcode()) {
2339 case Instruction::Add:
2340 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2341 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2342 if (BO.hasNoUnsignedWrap()) {
2343 // 'add nuw x, C' produces [C, UINT_MAX].
2345 } else if (BO.hasNoSignedWrap()) {
2346 if (C->isNegative()) {
2347 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2348 Lower = APInt::getSignedMinValue(Width);
2349 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2351 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2352 Lower = APInt::getSignedMinValue(Width) + *C;
2353 Upper = APInt::getSignedMaxValue(Width) + 1;
2359 case Instruction::And:
2360 if (match(BO.getOperand(1), m_APInt(C)))
2361 // 'and x, C' produces [0, C].
2365 case Instruction::Or:
2366 if (match(BO.getOperand(1), m_APInt(C)))
2367 // 'or x, C' produces [C, UINT_MAX].
2371 case Instruction::AShr:
2372 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2373 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2374 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2375 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2376 } else if (match(BO.getOperand(0), m_APInt(C))) {
2377 unsigned ShiftAmount = Width - 1;
2378 if (!C->isNullValue() && BO.isExact())
2379 ShiftAmount = C->countTrailingZeros();
2380 if (C->isNegative()) {
2381 // 'ashr C, x' produces [C, C >> (Width-1)]
2383 Upper = C->ashr(ShiftAmount) + 1;
2385 // 'ashr C, x' produces [C >> (Width-1), C]
2386 Lower = C->ashr(ShiftAmount);
2392 case Instruction::LShr:
2393 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2394 // 'lshr x, C' produces [0, UINT_MAX >> C].
2395 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2396 } else if (match(BO.getOperand(0), m_APInt(C))) {
2397 // 'lshr C, x' produces [C >> (Width-1), C].
2398 unsigned ShiftAmount = Width - 1;
2399 if (!C->isNullValue() && BO.isExact())
2400 ShiftAmount = C->countTrailingZeros();
2401 Lower = C->lshr(ShiftAmount);
2406 case Instruction::Shl:
2407 if (match(BO.getOperand(0), m_APInt(C))) {
2408 if (BO.hasNoUnsignedWrap()) {
2409 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2411 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2412 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2413 if (C->isNegative()) {
2414 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2415 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2416 Lower = C->shl(ShiftAmount);
2419 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2420 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2422 Upper = C->shl(ShiftAmount) + 1;
2428 case Instruction::SDiv:
2429 if (match(BO.getOperand(1), m_APInt(C))) {
2430 APInt IntMin = APInt::getSignedMinValue(Width);
2431 APInt IntMax = APInt::getSignedMaxValue(Width);
2432 if (C->isAllOnesValue()) {
2433 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2434 // where C != -1 and C != 0 and C != 1
2437 } else if (C->countLeadingZeros() < Width - 1) {
2438 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2439 // where C != -1 and C != 0 and C != 1
2440 Lower = IntMin.sdiv(*C);
2441 Upper = IntMax.sdiv(*C);
2442 if (Lower.sgt(Upper))
2443 std::swap(Lower, Upper);
2445 assert(Upper != Lower && "Upper part of range has wrapped!");
2447 } else if (match(BO.getOperand(0), m_APInt(C))) {
2448 if (C->isMinSignedValue()) {
2449 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2451 Upper = Lower.lshr(1) + 1;
2453 // 'sdiv C, x' produces [-|C|, |C|].
2454 Upper = C->abs() + 1;
2455 Lower = (-Upper) + 1;
2460 case Instruction::UDiv:
2461 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2462 // 'udiv x, C' produces [0, UINT_MAX / C].
2463 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2464 } else if (match(BO.getOperand(0), m_APInt(C))) {
2465 // 'udiv C, x' produces [0, C].
2470 case Instruction::SRem:
2471 if (match(BO.getOperand(1), m_APInt(C))) {
2472 // 'srem x, C' produces (-|C|, |C|).
2474 Lower = (-Upper) + 1;
2478 case Instruction::URem:
2479 if (match(BO.getOperand(1), m_APInt(C)))
2480 // 'urem x, C' produces [0, C).
2489 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2492 if (!match(RHS, m_APInt(C)))
2495 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2496 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2497 if (RHS_CR.isEmptySet())
2498 return ConstantInt::getFalse(GetCompareTy(RHS));
2499 if (RHS_CR.isFullSet())
2500 return ConstantInt::getTrue(GetCompareTy(RHS));
2502 // Find the range of possible values for binary operators.
2503 unsigned Width = C->getBitWidth();
2504 APInt Lower = APInt(Width, 0);
2505 APInt Upper = APInt(Width, 0);
2506 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2507 setLimitsForBinOp(*BO, Lower, Upper);
2509 ConstantRange LHS_CR =
2510 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2512 if (auto *I = dyn_cast<Instruction>(LHS))
2513 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2514 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2516 if (!LHS_CR.isFullSet()) {
2517 if (RHS_CR.contains(LHS_CR))
2518 return ConstantInt::getTrue(GetCompareTy(RHS));
2519 if (RHS_CR.inverse().contains(LHS_CR))
2520 return ConstantInt::getFalse(GetCompareTy(RHS));
2526 /// TODO: A large part of this logic is duplicated in InstCombine's
2527 /// foldICmpBinOp(). We should be able to share that and avoid the code
2529 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2530 Value *RHS, const SimplifyQuery &Q,
2531 unsigned MaxRecurse) {
2532 Type *ITy = GetCompareTy(LHS); // The return type.
2534 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2535 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2536 if (MaxRecurse && (LBO || RBO)) {
2537 // Analyze the case when either LHS or RHS is an add instruction.
2538 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2539 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2540 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2541 if (LBO && LBO->getOpcode() == Instruction::Add) {
2542 A = LBO->getOperand(0);
2543 B = LBO->getOperand(1);
2545 ICmpInst::isEquality(Pred) ||
2546 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2547 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2549 if (RBO && RBO->getOpcode() == Instruction::Add) {
2550 C = RBO->getOperand(0);
2551 D = RBO->getOperand(1);
2553 ICmpInst::isEquality(Pred) ||
2554 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2555 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2558 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2559 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2560 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2561 Constant::getNullValue(RHS->getType()), Q,
2565 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2566 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2568 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2569 C == LHS ? D : C, Q, MaxRecurse - 1))
2572 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2573 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2575 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2578 // C + B == C + D -> B == D
2581 } else if (A == D) {
2582 // D + B == C + D -> B == C
2585 } else if (B == C) {
2586 // A + C == C + D -> A == D
2591 // A + D == C + D -> A == C
2595 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2602 // icmp pred (or X, Y), X
2603 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2604 if (Pred == ICmpInst::ICMP_ULT)
2605 return getFalse(ITy);
2606 if (Pred == ICmpInst::ICMP_UGE)
2607 return getTrue(ITy);
2609 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2610 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2611 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2612 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2613 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2614 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2615 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2618 // icmp pred X, (or X, Y)
2619 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2620 if (Pred == ICmpInst::ICMP_ULE)
2621 return getTrue(ITy);
2622 if (Pred == ICmpInst::ICMP_UGT)
2623 return getFalse(ITy);
2625 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2626 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2627 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2628 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2629 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2630 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2631 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2636 // icmp pred (and X, Y), X
2637 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2638 if (Pred == ICmpInst::ICMP_UGT)
2639 return getFalse(ITy);
2640 if (Pred == ICmpInst::ICMP_ULE)
2641 return getTrue(ITy);
2643 // icmp pred X, (and X, Y)
2644 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2645 if (Pred == ICmpInst::ICMP_UGE)
2646 return getTrue(ITy);
2647 if (Pred == ICmpInst::ICMP_ULT)
2648 return getFalse(ITy);
2651 // 0 - (zext X) pred C
2652 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2653 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2654 if (RHSC->getValue().isStrictlyPositive()) {
2655 if (Pred == ICmpInst::ICMP_SLT)
2656 return ConstantInt::getTrue(RHSC->getContext());
2657 if (Pred == ICmpInst::ICMP_SGE)
2658 return ConstantInt::getFalse(RHSC->getContext());
2659 if (Pred == ICmpInst::ICMP_EQ)
2660 return ConstantInt::getFalse(RHSC->getContext());
2661 if (Pred == ICmpInst::ICMP_NE)
2662 return ConstantInt::getTrue(RHSC->getContext());
2664 if (RHSC->getValue().isNonNegative()) {
2665 if (Pred == ICmpInst::ICMP_SLE)
2666 return ConstantInt::getTrue(RHSC->getContext());
2667 if (Pred == ICmpInst::ICMP_SGT)
2668 return ConstantInt::getFalse(RHSC->getContext());
2673 // icmp pred (urem X, Y), Y
2674 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2678 case ICmpInst::ICMP_SGT:
2679 case ICmpInst::ICMP_SGE: {
2680 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2681 if (!Known.isNonNegative())
2685 case ICmpInst::ICMP_EQ:
2686 case ICmpInst::ICMP_UGT:
2687 case ICmpInst::ICMP_UGE:
2688 return getFalse(ITy);
2689 case ICmpInst::ICMP_SLT:
2690 case ICmpInst::ICMP_SLE: {
2691 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2692 if (!Known.isNonNegative())
2696 case ICmpInst::ICMP_NE:
2697 case ICmpInst::ICMP_ULT:
2698 case ICmpInst::ICMP_ULE:
2699 return getTrue(ITy);
2703 // icmp pred X, (urem Y, X)
2704 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2708 case ICmpInst::ICMP_SGT:
2709 case ICmpInst::ICMP_SGE: {
2710 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2711 if (!Known.isNonNegative())
2715 case ICmpInst::ICMP_NE:
2716 case ICmpInst::ICMP_UGT:
2717 case ICmpInst::ICMP_UGE:
2718 return getTrue(ITy);
2719 case ICmpInst::ICMP_SLT:
2720 case ICmpInst::ICMP_SLE: {
2721 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2722 if (!Known.isNonNegative())
2726 case ICmpInst::ICMP_EQ:
2727 case ICmpInst::ICMP_ULT:
2728 case ICmpInst::ICMP_ULE:
2729 return getFalse(ITy);
2735 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2736 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2737 // icmp pred (X op Y), X
2738 if (Pred == ICmpInst::ICMP_UGT)
2739 return getFalse(ITy);
2740 if (Pred == ICmpInst::ICMP_ULE)
2741 return getTrue(ITy);
2746 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2747 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2748 // icmp pred X, (X op Y)
2749 if (Pred == ICmpInst::ICMP_ULT)
2750 return getFalse(ITy);
2751 if (Pred == ICmpInst::ICMP_UGE)
2752 return getTrue(ITy);
2759 // where CI2 is a power of 2 and CI isn't
2760 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2761 const APInt *CI2Val, *CIVal = &CI->getValue();
2762 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2763 CI2Val->isPowerOf2()) {
2764 if (!CIVal->isPowerOf2()) {
2765 // CI2 << X can equal zero in some circumstances,
2766 // this simplification is unsafe if CI is zero.
2768 // We know it is safe if:
2769 // - The shift is nsw, we can't shift out the one bit.
2770 // - The shift is nuw, we can't shift out the one bit.
2773 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2774 CI2Val->isOneValue() || !CI->isZero()) {
2775 if (Pred == ICmpInst::ICMP_EQ)
2776 return ConstantInt::getFalse(RHS->getContext());
2777 if (Pred == ICmpInst::ICMP_NE)
2778 return ConstantInt::getTrue(RHS->getContext());
2781 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2782 if (Pred == ICmpInst::ICMP_UGT)
2783 return ConstantInt::getFalse(RHS->getContext());
2784 if (Pred == ICmpInst::ICMP_ULE)
2785 return ConstantInt::getTrue(RHS->getContext());
2790 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2791 LBO->getOperand(1) == RBO->getOperand(1)) {
2792 switch (LBO->getOpcode()) {
2795 case Instruction::UDiv:
2796 case Instruction::LShr:
2797 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2799 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2800 RBO->getOperand(0), Q, MaxRecurse - 1))
2803 case Instruction::SDiv:
2804 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2806 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2807 RBO->getOperand(0), Q, MaxRecurse - 1))
2810 case Instruction::AShr:
2811 if (!LBO->isExact() || !RBO->isExact())
2813 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2814 RBO->getOperand(0), Q, MaxRecurse - 1))
2817 case Instruction::Shl: {
2818 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2819 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2822 if (!NSW && ICmpInst::isSigned(Pred))
2824 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2825 RBO->getOperand(0), Q, MaxRecurse - 1))
2834 /// Simplify integer comparisons where at least one operand of the compare
2835 /// matches an integer min/max idiom.
2836 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2837 Value *RHS, const SimplifyQuery &Q,
2838 unsigned MaxRecurse) {
2839 Type *ITy = GetCompareTy(LHS); // The return type.
2841 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2842 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2844 // Signed variants on "max(a,b)>=a -> true".
2845 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2847 std::swap(A, B); // smax(A, B) pred A.
2848 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2849 // We analyze this as smax(A, B) pred A.
2851 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2852 (A == LHS || B == LHS)) {
2854 std::swap(A, B); // A pred smax(A, B).
2855 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2856 // We analyze this as smax(A, B) swapped-pred A.
2857 P = CmpInst::getSwappedPredicate(Pred);
2858 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2859 (A == RHS || B == RHS)) {
2861 std::swap(A, B); // smin(A, B) pred A.
2862 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2863 // We analyze this as smax(-A, -B) swapped-pred -A.
2864 // Note that we do not need to actually form -A or -B thanks to EqP.
2865 P = CmpInst::getSwappedPredicate(Pred);
2866 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2867 (A == LHS || B == LHS)) {
2869 std::swap(A, B); // A pred smin(A, B).
2870 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2871 // We analyze this as smax(-A, -B) pred -A.
2872 // Note that we do not need to actually form -A or -B thanks to EqP.
2875 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2876 // Cases correspond to "max(A, B) p A".
2880 case CmpInst::ICMP_EQ:
2881 case CmpInst::ICMP_SLE:
2882 // Equivalent to "A EqP B". This may be the same as the condition tested
2883 // in the max/min; if so, we can just return that.
2884 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2886 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2888 // Otherwise, see if "A EqP B" simplifies.
2890 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2893 case CmpInst::ICMP_NE:
2894 case CmpInst::ICMP_SGT: {
2895 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2896 // Equivalent to "A InvEqP B". This may be the same as the condition
2897 // tested in the max/min; if so, we can just return that.
2898 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2900 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2902 // Otherwise, see if "A InvEqP B" simplifies.
2904 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2908 case CmpInst::ICMP_SGE:
2910 return getTrue(ITy);
2911 case CmpInst::ICMP_SLT:
2913 return getFalse(ITy);
2917 // Unsigned variants on "max(a,b)>=a -> true".
2918 P = CmpInst::BAD_ICMP_PREDICATE;
2919 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2921 std::swap(A, B); // umax(A, B) pred A.
2922 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2923 // We analyze this as umax(A, B) pred A.
2925 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2926 (A == LHS || B == LHS)) {
2928 std::swap(A, B); // A pred umax(A, B).
2929 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2930 // We analyze this as umax(A, B) swapped-pred A.
2931 P = CmpInst::getSwappedPredicate(Pred);
2932 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2933 (A == RHS || B == RHS)) {
2935 std::swap(A, B); // umin(A, B) pred A.
2936 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2937 // We analyze this as umax(-A, -B) swapped-pred -A.
2938 // Note that we do not need to actually form -A or -B thanks to EqP.
2939 P = CmpInst::getSwappedPredicate(Pred);
2940 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2941 (A == LHS || B == LHS)) {
2943 std::swap(A, B); // A pred umin(A, B).
2944 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2945 // We analyze this as umax(-A, -B) pred -A.
2946 // Note that we do not need to actually form -A or -B thanks to EqP.
2949 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2950 // Cases correspond to "max(A, B) p A".
2954 case CmpInst::ICMP_EQ:
2955 case CmpInst::ICMP_ULE:
2956 // Equivalent to "A EqP B". This may be the same as the condition tested
2957 // in the max/min; if so, we can just return that.
2958 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2960 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2962 // Otherwise, see if "A EqP B" simplifies.
2964 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2967 case CmpInst::ICMP_NE:
2968 case CmpInst::ICMP_UGT: {
2969 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2970 // Equivalent to "A InvEqP B". This may be the same as the condition
2971 // tested in the max/min; if so, we can just return that.
2972 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2974 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2976 // Otherwise, see if "A InvEqP B" simplifies.
2978 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2982 case CmpInst::ICMP_UGE:
2984 return getTrue(ITy);
2985 case CmpInst::ICMP_ULT:
2987 return getFalse(ITy);
2991 // Variants on "max(x,y) >= min(x,z)".
2993 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2994 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2995 (A == C || A == D || B == C || B == D)) {
2996 // max(x, ?) pred min(x, ?).
2997 if (Pred == CmpInst::ICMP_SGE)
2999 return getTrue(ITy);
3000 if (Pred == CmpInst::ICMP_SLT)
3002 return getFalse(ITy);
3003 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3004 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3005 (A == C || A == D || B == C || B == D)) {
3006 // min(x, ?) pred max(x, ?).
3007 if (Pred == CmpInst::ICMP_SLE)
3009 return getTrue(ITy);
3010 if (Pred == CmpInst::ICMP_SGT)
3012 return getFalse(ITy);
3013 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3014 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3015 (A == C || A == D || B == C || B == D)) {
3016 // max(x, ?) pred min(x, ?).
3017 if (Pred == CmpInst::ICMP_UGE)
3019 return getTrue(ITy);
3020 if (Pred == CmpInst::ICMP_ULT)
3022 return getFalse(ITy);
3023 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3024 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3025 (A == C || A == D || B == C || B == D)) {
3026 // min(x, ?) pred max(x, ?).
3027 if (Pred == CmpInst::ICMP_ULE)
3029 return getTrue(ITy);
3030 if (Pred == CmpInst::ICMP_UGT)
3032 return getFalse(ITy);
3038 /// Given operands for an ICmpInst, see if we can fold the result.
3039 /// If not, this returns null.
3040 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3041 const SimplifyQuery &Q, unsigned MaxRecurse) {
3042 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3043 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3045 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3046 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3047 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3049 // If we have a constant, make sure it is on the RHS.
3050 std::swap(LHS, RHS);
3051 Pred = CmpInst::getSwappedPredicate(Pred);
3054 Type *ITy = GetCompareTy(LHS); // The return type.
3056 // icmp X, X -> true/false
3057 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3058 // because X could be 0.
3059 if (LHS == RHS || isa<UndefValue>(RHS))
3060 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3062 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3065 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3068 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3071 // If both operands have range metadata, use the metadata
3072 // to simplify the comparison.
3073 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3074 auto RHS_Instr = cast<Instruction>(RHS);
3075 auto LHS_Instr = cast<Instruction>(LHS);
3077 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3078 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3079 auto RHS_CR = getConstantRangeFromMetadata(
3080 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3081 auto LHS_CR = getConstantRangeFromMetadata(
3082 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3084 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3085 if (Satisfied_CR.contains(LHS_CR))
3086 return ConstantInt::getTrue(RHS->getContext());
3088 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3089 CmpInst::getInversePredicate(Pred), RHS_CR);
3090 if (InversedSatisfied_CR.contains(LHS_CR))
3091 return ConstantInt::getFalse(RHS->getContext());
3095 // Compare of cast, for example (zext X) != 0 -> X != 0
3096 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3097 Instruction *LI = cast<CastInst>(LHS);
3098 Value *SrcOp = LI->getOperand(0);
3099 Type *SrcTy = SrcOp->getType();
3100 Type *DstTy = LI->getType();
3102 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3103 // if the integer type is the same size as the pointer type.
3104 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3105 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3106 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3107 // Transfer the cast to the constant.
3108 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3109 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3112 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3113 if (RI->getOperand(0)->getType() == SrcTy)
3114 // Compare without the cast.
3115 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3121 if (isa<ZExtInst>(LHS)) {
3122 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3124 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3125 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3126 // Compare X and Y. Note that signed predicates become unsigned.
3127 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3128 SrcOp, RI->getOperand(0), Q,
3132 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3133 // too. If not, then try to deduce the result of the comparison.
3134 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3135 // Compute the constant that would happen if we truncated to SrcTy then
3136 // reextended to DstTy.
3137 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3138 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3140 // If the re-extended constant didn't change then this is effectively
3141 // also a case of comparing two zero-extended values.
3142 if (RExt == CI && MaxRecurse)
3143 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3144 SrcOp, Trunc, Q, MaxRecurse-1))
3147 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3148 // there. Use this to work out the result of the comparison.
3151 default: llvm_unreachable("Unknown ICmp predicate!");
3153 case ICmpInst::ICMP_EQ:
3154 case ICmpInst::ICMP_UGT:
3155 case ICmpInst::ICMP_UGE:
3156 return ConstantInt::getFalse(CI->getContext());
3158 case ICmpInst::ICMP_NE:
3159 case ICmpInst::ICMP_ULT:
3160 case ICmpInst::ICMP_ULE:
3161 return ConstantInt::getTrue(CI->getContext());
3163 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3164 // is non-negative then LHS <s RHS.
3165 case ICmpInst::ICMP_SGT:
3166 case ICmpInst::ICMP_SGE:
3167 return CI->getValue().isNegative() ?
3168 ConstantInt::getTrue(CI->getContext()) :
3169 ConstantInt::getFalse(CI->getContext());
3171 case ICmpInst::ICMP_SLT:
3172 case ICmpInst::ICMP_SLE:
3173 return CI->getValue().isNegative() ?
3174 ConstantInt::getFalse(CI->getContext()) :
3175 ConstantInt::getTrue(CI->getContext());
3181 if (isa<SExtInst>(LHS)) {
3182 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3184 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3185 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3186 // Compare X and Y. Note that the predicate does not change.
3187 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3191 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3192 // too. If not, then try to deduce the result of the comparison.
3193 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3194 // Compute the constant that would happen if we truncated to SrcTy then
3195 // reextended to DstTy.
3196 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3197 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3199 // If the re-extended constant didn't change then this is effectively
3200 // also a case of comparing two sign-extended values.
3201 if (RExt == CI && MaxRecurse)
3202 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3205 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3206 // bits there. Use this to work out the result of the comparison.
3209 default: llvm_unreachable("Unknown ICmp predicate!");
3210 case ICmpInst::ICMP_EQ:
3211 return ConstantInt::getFalse(CI->getContext());
3212 case ICmpInst::ICMP_NE:
3213 return ConstantInt::getTrue(CI->getContext());
3215 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3217 case ICmpInst::ICMP_SGT:
3218 case ICmpInst::ICMP_SGE:
3219 return CI->getValue().isNegative() ?
3220 ConstantInt::getTrue(CI->getContext()) :
3221 ConstantInt::getFalse(CI->getContext());
3222 case ICmpInst::ICMP_SLT:
3223 case ICmpInst::ICMP_SLE:
3224 return CI->getValue().isNegative() ?
3225 ConstantInt::getFalse(CI->getContext()) :
3226 ConstantInt::getTrue(CI->getContext());
3228 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3230 case ICmpInst::ICMP_UGT:
3231 case ICmpInst::ICMP_UGE:
3232 // Comparison is true iff the LHS <s 0.
3234 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3235 Constant::getNullValue(SrcTy),
3239 case ICmpInst::ICMP_ULT:
3240 case ICmpInst::ICMP_ULE:
3241 // Comparison is true iff the LHS >=s 0.
3243 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3244 Constant::getNullValue(SrcTy),
3254 // icmp eq|ne X, Y -> false|true if X != Y
3255 if (ICmpInst::isEquality(Pred) &&
3256 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3257 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3260 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3263 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3266 // Simplify comparisons of related pointers using a powerful, recursive
3267 // GEP-walk when we have target data available..
3268 if (LHS->getType()->isPointerTy())
3269 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3272 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3273 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3274 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3275 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3276 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3277 Q.DL.getTypeSizeInBits(CRHS->getType()))
3278 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3279 CLHS->getPointerOperand(),
3280 CRHS->getPointerOperand()))
3283 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3284 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3285 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3286 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3287 (ICmpInst::isEquality(Pred) ||
3288 (GLHS->isInBounds() && GRHS->isInBounds() &&
3289 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3290 // The bases are equal and the indices are constant. Build a constant
3291 // expression GEP with the same indices and a null base pointer to see
3292 // what constant folding can make out of it.
3293 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3294 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3295 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3296 GLHS->getSourceElementType(), Null, IndicesLHS);
3298 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3299 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3300 GLHS->getSourceElementType(), Null, IndicesRHS);
3301 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3306 // If the comparison is with the result of a select instruction, check whether
3307 // comparing with either branch of the select always yields the same value.
3308 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3309 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3312 // If the comparison is with the result of a phi instruction, check whether
3313 // doing the compare with each incoming phi value yields a common result.
3314 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3315 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3321 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3322 const SimplifyQuery &Q) {
3323 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3326 /// Given operands for an FCmpInst, see if we can fold the result.
3327 /// If not, this returns null.
3328 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3329 FastMathFlags FMF, const SimplifyQuery &Q,
3330 unsigned MaxRecurse) {
3331 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3332 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3334 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3335 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3336 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3338 // If we have a constant, make sure it is on the RHS.
3339 std::swap(LHS, RHS);
3340 Pred = CmpInst::getSwappedPredicate(Pred);
3343 // Fold trivial predicates.
3344 Type *RetTy = GetCompareTy(LHS);
3345 if (Pred == FCmpInst::FCMP_FALSE)
3346 return getFalse(RetTy);
3347 if (Pred == FCmpInst::FCMP_TRUE)
3348 return getTrue(RetTy);
3350 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3352 if (Pred == FCmpInst::FCMP_UNO)
3353 return getFalse(RetTy);
3354 if (Pred == FCmpInst::FCMP_ORD)
3355 return getTrue(RetTy);
3358 // fcmp pred x, undef and fcmp pred undef, x
3359 // fold to true if unordered, false if ordered
3360 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3361 // Choosing NaN for the undef will always make unordered comparison succeed
3362 // and ordered comparison fail.
3363 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3366 // fcmp x,x -> true/false. Not all compares are foldable.
3368 if (CmpInst::isTrueWhenEqual(Pred))
3369 return getTrue(RetTy);
3370 if (CmpInst::isFalseWhenEqual(Pred))
3371 return getFalse(RetTy);
3374 // Handle fcmp with constant RHS.
3376 if (match(RHS, m_APFloat(C))) {
3377 // If the constant is a nan, see if we can fold the comparison based on it.
3379 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3380 return getFalse(RetTy);
3381 assert(FCmpInst::isUnordered(Pred) &&
3382 "Comparison must be either ordered or unordered!");
3383 // True if unordered.
3384 return getTrue(RetTy);
3386 // Check whether the constant is an infinity.
3387 if (C->isInfinity()) {
3388 if (C->isNegative()) {
3390 case FCmpInst::FCMP_OLT:
3391 // No value is ordered and less than negative infinity.
3392 return getFalse(RetTy);
3393 case FCmpInst::FCMP_UGE:
3394 // All values are unordered with or at least negative infinity.
3395 return getTrue(RetTy);
3401 case FCmpInst::FCMP_OGT:
3402 // No value is ordered and greater than infinity.
3403 return getFalse(RetTy);
3404 case FCmpInst::FCMP_ULE:
3405 // All values are unordered with and at most infinity.
3406 return getTrue(RetTy);
3414 case FCmpInst::FCMP_UGE:
3415 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3416 return getTrue(RetTy);
3418 case FCmpInst::FCMP_OLT:
3420 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3421 return getFalse(RetTy);
3426 } else if (C->isNegative()) {
3427 assert(!C->isNaN() && "Unexpected NaN constant!");
3428 // TODO: We can catch more cases by using a range check rather than
3429 // relying on CannotBeOrderedLessThanZero.
3431 case FCmpInst::FCMP_UGE:
3432 case FCmpInst::FCMP_UGT:
3433 case FCmpInst::FCMP_UNE:
3434 // (X >= 0) implies (X > C) when (C < 0)
3435 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3436 return getTrue(RetTy);
3438 case FCmpInst::FCMP_OEQ:
3439 case FCmpInst::FCMP_OLE:
3440 case FCmpInst::FCMP_OLT:
3441 // (X >= 0) implies !(X < C) when (C < 0)
3442 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3443 return getFalse(RetTy);
3451 // If the comparison is with the result of a select instruction, check whether
3452 // comparing with either branch of the select always yields the same value.
3453 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3454 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3457 // If the comparison is with the result of a phi instruction, check whether
3458 // doing the compare with each incoming phi value yields a common result.
3459 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3460 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3466 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3467 FastMathFlags FMF, const SimplifyQuery &Q) {
3468 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3471 /// See if V simplifies when its operand Op is replaced with RepOp.
3472 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3473 const SimplifyQuery &Q,
3474 unsigned MaxRecurse) {
3475 // Trivial replacement.
3479 // We cannot replace a constant, and shouldn't even try.
3480 if (isa<Constant>(Op))
3483 auto *I = dyn_cast<Instruction>(V);
3487 // If this is a binary operator, try to simplify it with the replaced op.
3488 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3490 // %cmp = icmp eq i32 %x, 2147483647
3491 // %add = add nsw i32 %x, 1
3492 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3494 // We can't replace %sel with %add unless we strip away the flags.
3495 if (isa<OverflowingBinaryOperator>(B))
3496 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3498 if (isa<PossiblyExactOperator>(B))
3503 if (B->getOperand(0) == Op)
3504 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3506 if (B->getOperand(1) == Op)
3507 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3512 // Same for CmpInsts.
3513 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3515 if (C->getOperand(0) == Op)
3516 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3518 if (C->getOperand(1) == Op)
3519 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3524 // TODO: We could hand off more cases to instsimplify here.
3526 // If all operands are constant after substituting Op for RepOp then we can
3527 // constant fold the instruction.
3528 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3529 // Build a list of all constant operands.
3530 SmallVector<Constant *, 8> ConstOps;
3531 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3532 if (I->getOperand(i) == Op)
3533 ConstOps.push_back(CRepOp);
3534 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3535 ConstOps.push_back(COp);
3540 // All operands were constants, fold it.
3541 if (ConstOps.size() == I->getNumOperands()) {
3542 if (CmpInst *C = dyn_cast<CmpInst>(I))
3543 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3544 ConstOps[1], Q.DL, Q.TLI);
3546 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3547 if (!LI->isVolatile())
3548 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3550 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3557 /// Try to simplify a select instruction when its condition operand is an
3558 /// integer comparison where one operand of the compare is a constant.
3559 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3560 const APInt *Y, bool TrueWhenUnset) {
3563 // (X & Y) == 0 ? X & ~Y : X --> X
3564 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3565 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3567 return TrueWhenUnset ? FalseVal : TrueVal;
3569 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3570 // (X & Y) != 0 ? X : X & ~Y --> X
3571 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3573 return TrueWhenUnset ? FalseVal : TrueVal;
3575 if (Y->isPowerOf2()) {
3576 // (X & Y) == 0 ? X | Y : X --> X | Y
3577 // (X & Y) != 0 ? X | Y : X --> X
3578 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3580 return TrueWhenUnset ? TrueVal : FalseVal;
3582 // (X & Y) == 0 ? X : X | Y --> X
3583 // (X & Y) != 0 ? X : X | Y --> X | Y
3584 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3586 return TrueWhenUnset ? TrueVal : FalseVal;
3592 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3594 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3595 ICmpInst::Predicate Pred,
3596 Value *TrueVal, Value *FalseVal) {
3599 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3602 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3603 Pred == ICmpInst::ICMP_EQ);
3606 /// Try to simplify a select instruction when its condition operand is an
3607 /// integer comparison.
3608 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3609 Value *FalseVal, const SimplifyQuery &Q,
3610 unsigned MaxRecurse) {
3611 ICmpInst::Predicate Pred;
3612 Value *CmpLHS, *CmpRHS;
3613 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3616 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3619 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3620 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3621 Pred == ICmpInst::ICMP_EQ))
3625 // Check for other compares that behave like bit test.
3626 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3630 if (CondVal->hasOneUse()) {
3632 if (match(CmpRHS, m_APInt(C))) {
3633 // X < MIN ? T : F --> F
3634 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3636 // X < MIN ? T : F --> F
3637 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3639 // X > MAX ? T : F --> F
3640 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3642 // X > MAX ? T : F --> F
3643 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3648 // If we have an equality comparison, then we know the value in one of the
3649 // arms of the select. See if substituting this value into the arm and
3650 // simplifying the result yields the same value as the other arm.
3651 if (Pred == ICmpInst::ICMP_EQ) {
3652 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3654 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3657 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3659 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3662 } else if (Pred == ICmpInst::ICMP_NE) {
3663 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3665 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3668 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3670 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3678 /// Given operands for a SelectInst, see if we can fold the result.
3679 /// If not, this returns null.
3680 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3681 const SimplifyQuery &Q, unsigned MaxRecurse) {
3682 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3683 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3684 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3685 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3687 // select undef, X, Y -> X or Y
3688 if (isa<UndefValue>(CondC))
3689 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3691 // TODO: Vector constants with undef elements don't simplify.
3693 // select true, X, Y -> X
3694 if (CondC->isAllOnesValue())
3696 // select false, X, Y -> Y
3697 if (CondC->isNullValue())
3701 // select ?, X, X -> X
3702 if (TrueVal == FalseVal)
3705 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3707 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3711 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3717 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3718 const SimplifyQuery &Q) {
3719 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3722 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3723 /// If not, this returns null.
3724 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3725 const SimplifyQuery &Q, unsigned) {
3726 // The type of the GEP pointer operand.
3728 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3730 // getelementptr P -> P.
3731 if (Ops.size() == 1)
3734 // Compute the (pointer) type returned by the GEP instruction.
3735 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3736 Type *GEPTy = PointerType::get(LastType, AS);
3737 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3738 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3739 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3740 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3742 if (isa<UndefValue>(Ops[0]))
3743 return UndefValue::get(GEPTy);
3745 if (Ops.size() == 2) {
3746 // getelementptr P, 0 -> P.
3747 if (match(Ops[1], m_Zero()))
3751 if (Ty->isSized()) {
3754 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3755 // getelementptr P, N -> P if P points to a type of zero size.
3756 if (TyAllocSize == 0)
3759 // The following transforms are only safe if the ptrtoint cast
3760 // doesn't truncate the pointers.
3761 if (Ops[1]->getType()->getScalarSizeInBits() ==
3762 Q.DL.getIndexSizeInBits(AS)) {
3763 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3764 if (match(P, m_Zero()))
3765 return Constant::getNullValue(GEPTy);
3767 if (match(P, m_PtrToInt(m_Value(Temp))))
3768 if (Temp->getType() == GEPTy)
3773 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3774 if (TyAllocSize == 1 &&
3775 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3776 if (Value *R = PtrToIntOrZero(P))
3779 // getelementptr V, (ashr (sub P, V), C) -> Q
3780 // if P points to a type of size 1 << C.
3782 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3783 m_ConstantInt(C))) &&
3784 TyAllocSize == 1ULL << C)
3785 if (Value *R = PtrToIntOrZero(P))
3788 // getelementptr V, (sdiv (sub P, V), C) -> Q
3789 // if P points to a type of size C.
3791 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3792 m_SpecificInt(TyAllocSize))))
3793 if (Value *R = PtrToIntOrZero(P))
3799 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3800 all_of(Ops.slice(1).drop_back(1),
3801 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3803 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3804 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3805 APInt BasePtrOffset(IdxWidth, 0);
3806 Value *StrippedBasePtr =
3807 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3810 // gep (gep V, C), (sub 0, V) -> C
3811 if (match(Ops.back(),
3812 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3813 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3814 return ConstantExpr::getIntToPtr(CI, GEPTy);
3816 // gep (gep V, C), (xor V, -1) -> C-1
3817 if (match(Ops.back(),
3818 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3819 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3820 return ConstantExpr::getIntToPtr(CI, GEPTy);
3825 // Check to see if this is constant foldable.
3826 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3829 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3831 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3836 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3837 const SimplifyQuery &Q) {
3838 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3841 /// Given operands for an InsertValueInst, see if we can fold the result.
3842 /// If not, this returns null.
3843 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3844 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3846 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3847 if (Constant *CVal = dyn_cast<Constant>(Val))
3848 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3850 // insertvalue x, undef, n -> x
3851 if (match(Val, m_Undef()))
3854 // insertvalue x, (extractvalue y, n), n
3855 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3856 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3857 EV->getIndices() == Idxs) {
3858 // insertvalue undef, (extractvalue y, n), n -> y
3859 if (match(Agg, m_Undef()))
3860 return EV->getAggregateOperand();
3862 // insertvalue y, (extractvalue y, n), n -> y
3863 if (Agg == EV->getAggregateOperand())
3870 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3871 ArrayRef<unsigned> Idxs,
3872 const SimplifyQuery &Q) {
3873 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3876 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3877 const SimplifyQuery &Q) {
3878 // Try to constant fold.
3879 auto *VecC = dyn_cast<Constant>(Vec);
3880 auto *ValC = dyn_cast<Constant>(Val);
3881 auto *IdxC = dyn_cast<Constant>(Idx);
3882 if (VecC && ValC && IdxC)
3883 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3885 // Fold into undef if index is out of bounds.
3886 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3887 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3888 if (CI->uge(NumElements))
3889 return UndefValue::get(Vec->getType());
3892 // If index is undef, it might be out of bounds (see above case)
3893 if (isa<UndefValue>(Idx))
3894 return UndefValue::get(Vec->getType());
3899 /// Given operands for an ExtractValueInst, see if we can fold the result.
3900 /// If not, this returns null.
3901 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3902 const SimplifyQuery &, unsigned) {
3903 if (auto *CAgg = dyn_cast<Constant>(Agg))
3904 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3906 // extractvalue x, (insertvalue y, elt, n), n -> elt
3907 unsigned NumIdxs = Idxs.size();
3908 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3909 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3910 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3911 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3912 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3913 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3914 Idxs.slice(0, NumCommonIdxs)) {
3915 if (NumIdxs == NumInsertValueIdxs)
3916 return IVI->getInsertedValueOperand();
3924 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3925 const SimplifyQuery &Q) {
3926 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3929 /// Given operands for an ExtractElementInst, see if we can fold the result.
3930 /// If not, this returns null.
3931 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3933 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3934 if (auto *CIdx = dyn_cast<Constant>(Idx))
3935 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3937 // The index is not relevant if our vector is a splat.
3938 if (auto *Splat = CVec->getSplatValue())
3941 if (isa<UndefValue>(Vec))
3942 return UndefValue::get(Vec->getType()->getVectorElementType());
3945 // If extracting a specified index from the vector, see if we can recursively
3946 // find a previously computed scalar that was inserted into the vector.
3947 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
3948 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
3949 // definitely out of bounds, thus undefined result
3950 return UndefValue::get(Vec->getType()->getVectorElementType());
3951 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3955 // An undef extract index can be arbitrarily chosen to be an out-of-range
3956 // index value, which would result in the instruction being undef.
3957 if (isa<UndefValue>(Idx))
3958 return UndefValue::get(Vec->getType()->getVectorElementType());
3963 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3964 const SimplifyQuery &Q) {
3965 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3968 /// See if we can fold the given phi. If not, returns null.
3969 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3970 // If all of the PHI's incoming values are the same then replace the PHI node
3971 // with the common value.
3972 Value *CommonValue = nullptr;
3973 bool HasUndefInput = false;
3974 for (Value *Incoming : PN->incoming_values()) {
3975 // If the incoming value is the phi node itself, it can safely be skipped.
3976 if (Incoming == PN) continue;
3977 if (isa<UndefValue>(Incoming)) {
3978 // Remember that we saw an undef value, but otherwise ignore them.
3979 HasUndefInput = true;
3982 if (CommonValue && Incoming != CommonValue)
3983 return nullptr; // Not the same, bail out.
3984 CommonValue = Incoming;
3987 // If CommonValue is null then all of the incoming values were either undef or
3988 // equal to the phi node itself.
3990 return UndefValue::get(PN->getType());
3992 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3993 // instruction, we cannot return X as the result of the PHI node unless it
3994 // dominates the PHI block.
3996 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4001 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4002 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4003 if (auto *C = dyn_cast<Constant>(Op))
4004 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4006 if (auto *CI = dyn_cast<CastInst>(Op)) {
4007 auto *Src = CI->getOperand(0);
4008 Type *SrcTy = Src->getType();
4009 Type *MidTy = CI->getType();
4011 if (Src->getType() == Ty) {
4012 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4013 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4015 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4017 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4019 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4020 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4021 SrcIntPtrTy, MidIntPtrTy,
4022 DstIntPtrTy) == Instruction::BitCast)
4028 if (CastOpc == Instruction::BitCast)
4029 if (Op->getType() == Ty)
4035 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4036 const SimplifyQuery &Q) {
4037 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4040 /// For the given destination element of a shuffle, peek through shuffles to
4041 /// match a root vector source operand that contains that element in the same
4042 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4043 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4044 int MaskVal, Value *RootVec,
4045 unsigned MaxRecurse) {
4049 // Bail out if any mask value is undefined. That kind of shuffle may be
4050 // simplified further based on demanded bits or other folds.
4054 // The mask value chooses which source operand we need to look at next.
4055 int InVecNumElts = Op0->getType()->getVectorNumElements();
4056 int RootElt = MaskVal;
4057 Value *SourceOp = Op0;
4058 if (MaskVal >= InVecNumElts) {
4059 RootElt = MaskVal - InVecNumElts;
4063 // If the source operand is a shuffle itself, look through it to find the
4064 // matching root vector.
4065 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4066 return foldIdentityShuffles(
4067 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4068 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4071 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4074 // The source operand is not a shuffle. Initialize the root vector value for
4075 // this shuffle if that has not been done yet.
4079 // Give up as soon as a source operand does not match the existing root value.
4080 if (RootVec != SourceOp)
4083 // The element must be coming from the same lane in the source vector
4084 // (although it may have crossed lanes in intermediate shuffles).
4085 if (RootElt != DestElt)
4091 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4092 Type *RetTy, const SimplifyQuery &Q,
4093 unsigned MaxRecurse) {
4094 if (isa<UndefValue>(Mask))
4095 return UndefValue::get(RetTy);
4097 Type *InVecTy = Op0->getType();
4098 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4099 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4101 SmallVector<int, 32> Indices;
4102 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4103 assert(MaskNumElts == Indices.size() &&
4104 "Size of Indices not same as number of mask elements?");
4106 // Canonicalization: If mask does not select elements from an input vector,
4107 // replace that input vector with undef.
4108 bool MaskSelects0 = false, MaskSelects1 = false;
4109 for (unsigned i = 0; i != MaskNumElts; ++i) {
4110 if (Indices[i] == -1)
4112 if ((unsigned)Indices[i] < InVecNumElts)
4113 MaskSelects0 = true;
4115 MaskSelects1 = true;
4118 Op0 = UndefValue::get(InVecTy);
4120 Op1 = UndefValue::get(InVecTy);
4122 auto *Op0Const = dyn_cast<Constant>(Op0);
4123 auto *Op1Const = dyn_cast<Constant>(Op1);
4125 // If all operands are constant, constant fold the shuffle.
4126 if (Op0Const && Op1Const)
4127 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4129 // Canonicalization: if only one input vector is constant, it shall be the
4131 if (Op0Const && !Op1Const) {
4132 std::swap(Op0, Op1);
4133 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4136 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4137 // value type is same as the input vectors' type.
4138 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4139 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4140 OpShuf->getMask()->getSplatValue())
4143 // Don't fold a shuffle with undef mask elements. This may get folded in a
4144 // better way using demanded bits or other analysis.
4145 // TODO: Should we allow this?
4146 if (find(Indices, -1) != Indices.end())
4149 // Check if every element of this shuffle can be mapped back to the
4150 // corresponding element of a single root vector. If so, we don't need this
4151 // shuffle. This handles simple identity shuffles as well as chains of
4152 // shuffles that may widen/narrow and/or move elements across lanes and back.
4153 Value *RootVec = nullptr;
4154 for (unsigned i = 0; i != MaskNumElts; ++i) {
4155 // Note that recursion is limited for each vector element, so if any element
4156 // exceeds the limit, this will fail to simplify.
4158 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4160 // We can't replace a widening/narrowing shuffle with one of its operands.
4161 if (!RootVec || RootVec->getType() != RetTy)
4167 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4168 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4169 Type *RetTy, const SimplifyQuery &Q) {
4170 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4173 /// Given operands for an FAdd, see if we can fold the result. If not, this
4175 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4176 const SimplifyQuery &Q, unsigned MaxRecurse) {
4177 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4181 if (match(Op1, m_NegZero()))
4184 // fadd X, 0 ==> X, when we know X is not -0
4185 if (match(Op1, m_Zero()) &&
4186 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4189 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
4190 // where nnan and ninf have to occur at least once somewhere in this
4192 Value *SubOp = nullptr;
4193 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
4195 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
4198 Instruction *FSub = cast<Instruction>(SubOp);
4199 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
4200 (FMF.noInfs() || FSub->hasNoInfs()))
4201 return Constant::getNullValue(Op0->getType());
4207 /// Given operands for an FSub, see if we can fold the result. If not, this
4209 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4210 const SimplifyQuery &Q, unsigned MaxRecurse) {
4211 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4215 if (match(Op1, m_Zero()))
4218 // fsub X, -0 ==> X, when we know X is not -0
4219 if (match(Op1, m_NegZero()) &&
4220 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4223 // fsub -0.0, (fsub -0.0, X) ==> X
4225 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
4228 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4229 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
4230 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
4233 // fsub nnan x, x ==> 0.0
4234 if (FMF.noNaNs() && Op0 == Op1)
4235 return Constant::getNullValue(Op0->getType());
4240 /// Given the operands for an FMul, see if we can fold the result
4241 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4242 const SimplifyQuery &Q, unsigned MaxRecurse) {
4243 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4246 // fmul X, 1.0 ==> X
4247 if (match(Op1, m_FPOne()))
4250 // fmul nnan nsz X, 0 ==> 0
4251 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
4254 // sqrt(X) * sqrt(X) --> X
4256 if (FMF.isFast() && Op0 == Op1 &&
4257 match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))))
4263 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4264 const SimplifyQuery &Q) {
4265 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4269 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4270 const SimplifyQuery &Q) {
4271 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4274 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4275 const SimplifyQuery &Q) {
4276 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4279 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4280 const SimplifyQuery &Q, unsigned) {
4281 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4284 // undef / X -> undef (the undef could be a snan).
4285 if (match(Op0, m_Undef()))
4288 // X / undef -> undef
4289 if (match(Op1, m_Undef()))
4293 if (match(Op1, m_FPOne()))
4297 // Requires that NaNs are off (X could be zero) and signed zeroes are
4298 // ignored (X could be positive or negative, so the output sign is unknown).
4299 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4303 // X / X -> 1.0 is legal when NaNs are ignored.
4304 // We can ignore infinities because INF/INF is NaN.
4306 return ConstantFP::get(Op0->getType(), 1.0);
4308 // (X * Y) / Y --> X if we can reassociate to the above form.
4310 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4313 // -X / X -> -1.0 and
4314 // X / -X -> -1.0 are legal when NaNs are ignored.
4315 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4316 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4317 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4318 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4319 BinaryOperator::getFNegArgument(Op1) == Op0))
4320 return ConstantFP::get(Op0->getType(), -1.0);
4326 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4327 const SimplifyQuery &Q) {
4328 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4331 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4332 const SimplifyQuery &Q, unsigned) {
4333 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4336 // undef % X -> undef (the undef could be a snan).
4337 if (match(Op0, m_Undef()))
4340 // X % undef -> undef
4341 if (match(Op1, m_Undef()))
4345 // Requires that NaNs are off (X could be zero) and signed zeroes are
4346 // ignored (X could be positive or negative, so the output sign is unknown).
4347 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4353 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4354 const SimplifyQuery &Q) {
4355 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4358 //=== Helper functions for higher up the class hierarchy.
4360 /// Given operands for a BinaryOperator, see if we can fold the result.
4361 /// If not, this returns null.
4362 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4363 const SimplifyQuery &Q, unsigned MaxRecurse) {
4365 case Instruction::Add:
4366 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4367 case Instruction::Sub:
4368 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4369 case Instruction::Mul:
4370 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4371 case Instruction::SDiv:
4372 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4373 case Instruction::UDiv:
4374 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4375 case Instruction::SRem:
4376 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4377 case Instruction::URem:
4378 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4379 case Instruction::Shl:
4380 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4381 case Instruction::LShr:
4382 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4383 case Instruction::AShr:
4384 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4385 case Instruction::And:
4386 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4387 case Instruction::Or:
4388 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4389 case Instruction::Xor:
4390 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4391 case Instruction::FAdd:
4392 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4393 case Instruction::FSub:
4394 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4395 case Instruction::FMul:
4396 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4397 case Instruction::FDiv:
4398 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4399 case Instruction::FRem:
4400 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4402 llvm_unreachable("Unexpected opcode");
4406 /// Given operands for a BinaryOperator, see if we can fold the result.
4407 /// If not, this returns null.
4408 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4409 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4410 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4411 const FastMathFlags &FMF, const SimplifyQuery &Q,
4412 unsigned MaxRecurse) {
4414 case Instruction::FAdd:
4415 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4416 case Instruction::FSub:
4417 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4418 case Instruction::FMul:
4419 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4420 case Instruction::FDiv:
4421 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4423 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4427 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4428 const SimplifyQuery &Q) {
4429 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4432 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4433 FastMathFlags FMF, const SimplifyQuery &Q) {
4434 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4437 /// Given operands for a CmpInst, see if we can fold the result.
4438 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4439 const SimplifyQuery &Q, unsigned MaxRecurse) {
4440 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4441 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4442 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4445 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4446 const SimplifyQuery &Q) {
4447 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4450 static bool IsIdempotent(Intrinsic::ID ID) {
4452 default: return false;
4454 // Unary idempotent: f(f(x)) = f(x)
4455 case Intrinsic::fabs:
4456 case Intrinsic::floor:
4457 case Intrinsic::ceil:
4458 case Intrinsic::trunc:
4459 case Intrinsic::rint:
4460 case Intrinsic::nearbyint:
4461 case Intrinsic::round:
4462 case Intrinsic::canonicalize:
4467 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4468 const DataLayout &DL) {
4469 GlobalValue *PtrSym;
4471 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4474 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4475 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4476 Type *Int32PtrTy = Int32Ty->getPointerTo();
4477 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4479 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4480 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4483 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4484 if (OffsetInt % 4 != 0)
4487 Constant *C = ConstantExpr::getGetElementPtr(
4488 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4489 ConstantInt::get(Int64Ty, OffsetInt / 4));
4490 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4494 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4498 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4499 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4504 if (LoadedCE->getOpcode() != Instruction::Sub)
4507 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4508 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4510 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4512 Constant *LoadedRHS = LoadedCE->getOperand(1);
4513 GlobalValue *LoadedRHSSym;
4514 APInt LoadedRHSOffset;
4515 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4517 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4520 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4523 static bool maskIsAllZeroOrUndef(Value *Mask) {
4524 auto *ConstMask = dyn_cast<Constant>(Mask);
4527 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4529 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4531 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4532 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4539 template <typename IterTy>
4540 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4541 const SimplifyQuery &Q, unsigned MaxRecurse) {
4542 Intrinsic::ID IID = F->getIntrinsicID();
4543 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4546 if (NumOperands == 1) {
4547 // Perform idempotent optimizations
4548 if (IsIdempotent(IID)) {
4549 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4550 if (II->getIntrinsicID() == IID)
4555 Value *IIOperand = *ArgBegin;
4558 case Intrinsic::fabs: {
4559 if (SignBitMustBeZero(IIOperand, Q.TLI))
4563 case Intrinsic::bswap: {
4564 // bswap(bswap(x)) -> x
4565 if (match(IIOperand, m_BSwap(m_Value(X))))
4569 case Intrinsic::bitreverse: {
4570 // bitreverse(bitreverse(x)) -> x
4571 if (match(IIOperand, m_BitReverse(m_Value(X))))
4575 case Intrinsic::exp: {
4577 if (Q.CxtI->hasAllowReassoc() &&
4578 match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X))))
4582 case Intrinsic::exp2: {
4583 // exp2(log2(x)) -> x
4584 if (Q.CxtI->hasAllowReassoc() &&
4585 match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X))))
4589 case Intrinsic::log: {
4591 if (Q.CxtI->hasAllowReassoc() &&
4592 match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X))))
4596 case Intrinsic::log2: {
4597 // log2(exp2(x)) -> x
4598 if (Q.CxtI->hasAllowReassoc() &&
4599 match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) {
4610 if (NumOperands == 2) {
4611 Value *LHS = *ArgBegin;
4612 Value *RHS = *(ArgBegin + 1);
4613 Type *ReturnType = F->getReturnType();
4616 case Intrinsic::usub_with_overflow:
4617 case Intrinsic::ssub_with_overflow: {
4618 // X - X -> { 0, false }
4620 return Constant::getNullValue(ReturnType);
4622 // X - undef -> undef
4623 // undef - X -> undef
4624 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4625 return UndefValue::get(ReturnType);
4629 case Intrinsic::uadd_with_overflow:
4630 case Intrinsic::sadd_with_overflow: {
4631 // X + undef -> undef
4632 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4633 return UndefValue::get(ReturnType);
4637 case Intrinsic::umul_with_overflow:
4638 case Intrinsic::smul_with_overflow: {
4639 // 0 * X -> { 0, false }
4640 // X * 0 -> { 0, false }
4641 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4642 return Constant::getNullValue(ReturnType);
4644 // undef * X -> { 0, false }
4645 // X * undef -> { 0, false }
4646 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4647 return Constant::getNullValue(ReturnType);
4651 case Intrinsic::load_relative: {
4652 Constant *C0 = dyn_cast<Constant>(LHS);
4653 Constant *C1 = dyn_cast<Constant>(RHS);
4655 return SimplifyRelativeLoad(C0, C1, Q.DL);
4658 case Intrinsic::powi:
4659 if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) {
4660 // powi(x, 0) -> 1.0
4661 if (Power->isZero())
4662 return ConstantFP::get(LHS->getType(), 1.0);
4673 // Simplify calls to llvm.masked.load.*
4675 case Intrinsic::masked_load: {
4676 Value *MaskArg = ArgBegin[2];
4677 Value *PassthruArg = ArgBegin[3];
4678 // If the mask is all zeros or undef, the "passthru" argument is the result.
4679 if (maskIsAllZeroOrUndef(MaskArg))
4688 template <typename IterTy>
4689 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4690 IterTy ArgEnd, const SimplifyQuery &Q,
4691 unsigned MaxRecurse) {
4692 Type *Ty = V->getType();
4693 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4694 Ty = PTy->getElementType();
4695 FunctionType *FTy = cast<FunctionType>(Ty);
4697 // call undef -> undef
4698 // call null -> undef
4699 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4700 return UndefValue::get(FTy->getReturnType());
4702 Function *F = dyn_cast<Function>(V);
4706 if (F->isIntrinsic())
4707 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4710 if (!canConstantFoldCallTo(CS, F))
4713 SmallVector<Constant *, 4> ConstantArgs;
4714 ConstantArgs.reserve(ArgEnd - ArgBegin);
4715 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4716 Constant *C = dyn_cast<Constant>(*I);
4719 ConstantArgs.push_back(C);
4722 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4725 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4726 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4727 const SimplifyQuery &Q) {
4728 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4731 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4732 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4733 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4736 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4737 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4738 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4742 /// See if we can compute a simplified version of this instruction.
4743 /// If not, this returns null.
4745 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4746 OptimizationRemarkEmitter *ORE) {
4747 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4750 switch (I->getOpcode()) {
4752 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4754 case Instruction::FAdd:
4755 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4756 I->getFastMathFlags(), Q);
4758 case Instruction::Add:
4759 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4760 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4761 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4763 case Instruction::FSub:
4764 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4765 I->getFastMathFlags(), Q);
4767 case Instruction::Sub:
4768 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4769 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4770 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4772 case Instruction::FMul:
4773 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4774 I->getFastMathFlags(), Q);
4776 case Instruction::Mul:
4777 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4779 case Instruction::SDiv:
4780 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4782 case Instruction::UDiv:
4783 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4785 case Instruction::FDiv:
4786 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4787 I->getFastMathFlags(), Q);
4789 case Instruction::SRem:
4790 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4792 case Instruction::URem:
4793 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4795 case Instruction::FRem:
4796 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4797 I->getFastMathFlags(), Q);
4799 case Instruction::Shl:
4800 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4801 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4802 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4804 case Instruction::LShr:
4805 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4806 cast<BinaryOperator>(I)->isExact(), Q);
4808 case Instruction::AShr:
4809 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4810 cast<BinaryOperator>(I)->isExact(), Q);
4812 case Instruction::And:
4813 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4815 case Instruction::Or:
4816 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4818 case Instruction::Xor:
4819 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4821 case Instruction::ICmp:
4822 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4823 I->getOperand(0), I->getOperand(1), Q);
4825 case Instruction::FCmp:
4827 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4828 I->getOperand(1), I->getFastMathFlags(), Q);
4830 case Instruction::Select:
4831 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4832 I->getOperand(2), Q);
4834 case Instruction::GetElementPtr: {
4835 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4836 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4840 case Instruction::InsertValue: {
4841 InsertValueInst *IV = cast<InsertValueInst>(I);
4842 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4843 IV->getInsertedValueOperand(),
4844 IV->getIndices(), Q);
4847 case Instruction::InsertElement: {
4848 auto *IE = cast<InsertElementInst>(I);
4849 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4850 IE->getOperand(2), Q);
4853 case Instruction::ExtractValue: {
4854 auto *EVI = cast<ExtractValueInst>(I);
4855 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4856 EVI->getIndices(), Q);
4859 case Instruction::ExtractElement: {
4860 auto *EEI = cast<ExtractElementInst>(I);
4861 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4862 EEI->getIndexOperand(), Q);
4865 case Instruction::ShuffleVector: {
4866 auto *SVI = cast<ShuffleVectorInst>(I);
4867 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4868 SVI->getMask(), SVI->getType(), Q);
4871 case Instruction::PHI:
4872 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4874 case Instruction::Call: {
4875 CallSite CS(cast<CallInst>(I));
4876 Result = SimplifyCall(CS, Q);
4879 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4880 #include "llvm/IR/Instruction.def"
4881 #undef HANDLE_CAST_INST
4883 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4885 case Instruction::Alloca:
4886 // No simplifications for Alloca and it can't be constant folded.
4891 // In general, it is possible for computeKnownBits to determine all bits in a
4892 // value even when the operands are not all constants.
4893 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4894 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4895 if (Known.isConstant())
4896 Result = ConstantInt::get(I->getType(), Known.getConstant());
4899 /// If called on unreachable code, the above logic may report that the
4900 /// instruction simplified to itself. Make life easier for users by
4901 /// detecting that case here, returning a safe value instead.
4902 return Result == I ? UndefValue::get(I->getType()) : Result;
4905 /// \brief Implementation of recursive simplification through an instruction's
4908 /// This is the common implementation of the recursive simplification routines.
4909 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4910 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4911 /// instructions to process and attempt to simplify it using
4912 /// InstructionSimplify.
4914 /// This routine returns 'true' only when *it* simplifies something. The passed
4915 /// in simplified value does not count toward this.
4916 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4917 const TargetLibraryInfo *TLI,
4918 const DominatorTree *DT,
4919 AssumptionCache *AC) {
4920 bool Simplified = false;
4921 SmallSetVector<Instruction *, 8> Worklist;
4922 const DataLayout &DL = I->getModule()->getDataLayout();
4924 // If we have an explicit value to collapse to, do that round of the
4925 // simplification loop by hand initially.
4927 for (User *U : I->users())
4929 Worklist.insert(cast<Instruction>(U));
4931 // Replace the instruction with its simplified value.
4932 I->replaceAllUsesWith(SimpleV);
4934 // Gracefully handle edge cases where the instruction is not wired into any
4936 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4937 !I->mayHaveSideEffects())
4938 I->eraseFromParent();
4943 // Note that we must test the size on each iteration, the worklist can grow.
4944 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4947 // See if this instruction simplifies.
4948 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4954 // Stash away all the uses of the old instruction so we can check them for
4955 // recursive simplifications after a RAUW. This is cheaper than checking all
4956 // uses of To on the recursive step in most cases.
4957 for (User *U : I->users())
4958 Worklist.insert(cast<Instruction>(U));
4960 // Replace the instruction with its simplified value.
4961 I->replaceAllUsesWith(SimpleV);
4963 // Gracefully handle edge cases where the instruction is not wired into any
4965 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4966 !I->mayHaveSideEffects())
4967 I->eraseFromParent();
4972 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4973 const TargetLibraryInfo *TLI,
4974 const DominatorTree *DT,
4975 AssumptionCache *AC) {
4976 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4979 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4980 const TargetLibraryInfo *TLI,
4981 const DominatorTree *DT,
4982 AssumptionCache *AC) {
4983 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4984 assert(SimpleV && "Must provide a simplified value.");
4985 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4989 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4990 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4991 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4992 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4993 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4994 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4995 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4996 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4999 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5000 const DataLayout &DL) {
5001 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5004 template <class T, class... TArgs>
5005 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5007 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5008 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5009 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5010 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5012 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,