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/ConstantFolding.h"
27 #include "llvm/Analysis/LoopAnalysisManager.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/OptimizationDiagnosticInfo.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() == 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()->getScalarType()->isIntegerTy(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()->getScalarType()->isIntegerTy(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 an FAdd, see if we can fold the result. If not, this
796 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
797 const SimplifyQuery &Q, unsigned MaxRecurse) {
798 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
802 if (match(Op1, m_NegZero()))
805 // fadd X, 0 ==> X, when we know X is not -0
806 if (match(Op1, m_Zero()) &&
807 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
810 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
811 // where nnan and ninf have to occur at least once somewhere in this
813 Value *SubOp = nullptr;
814 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
816 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
819 Instruction *FSub = cast<Instruction>(SubOp);
820 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
821 (FMF.noInfs() || FSub->hasNoInfs()))
822 return Constant::getNullValue(Op0->getType());
828 /// Given operands for an FSub, see if we can fold the result. If not, this
830 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
831 const SimplifyQuery &Q, unsigned MaxRecurse) {
832 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
836 if (match(Op1, m_Zero()))
839 // fsub X, -0 ==> X, when we know X is not -0
840 if (match(Op1, m_NegZero()) &&
841 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
844 // fsub -0.0, (fsub -0.0, X) ==> X
846 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
849 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
850 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
851 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
854 // fsub nnan x, x ==> 0.0
855 if (FMF.noNaNs() && Op0 == Op1)
856 return Constant::getNullValue(Op0->getType());
861 /// Given the operands for an FMul, see if we can fold the result
862 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
863 const SimplifyQuery &Q, unsigned MaxRecurse) {
864 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
868 if (match(Op1, m_FPOne()))
871 // fmul nnan nsz X, 0 ==> 0
872 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
878 /// Given operands for a Mul, see if we can fold the result.
879 /// If not, this returns null.
880 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
881 unsigned MaxRecurse) {
882 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
886 if (match(Op1, m_Undef()))
887 return Constant::getNullValue(Op0->getType());
890 if (match(Op1, m_Zero()))
894 if (match(Op1, m_One()))
897 // (X / Y) * Y -> X if the division is exact.
899 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
900 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
904 if (MaxRecurse && Op0->getType()->getScalarType()->isIntegerTy(1))
905 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
908 // Try some generic simplifications for associative operations.
909 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
913 // Mul distributes over Add. Try some generic simplifications based on this.
914 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
918 // If the operation is with the result of a select instruction, check whether
919 // operating on either branch of the select always yields the same value.
920 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
921 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
925 // If the operation is with the result of a phi instruction, check whether
926 // operating on all incoming values of the phi always yields the same value.
927 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
928 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
935 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
936 const SimplifyQuery &Q) {
937 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
941 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
942 const SimplifyQuery &Q) {
943 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
946 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
947 const SimplifyQuery &Q) {
948 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
951 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
952 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
955 /// Check for common or similar folds of integer division or integer remainder.
956 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
957 Type *Ty = Op0->getType();
959 // X / undef -> undef
960 // X % undef -> undef
961 if (match(Op1, m_Undef()))
966 // We don't need to preserve faults!
967 if (match(Op1, m_Zero()))
968 return UndefValue::get(Ty);
970 // If any element of a constant divisor vector is zero, the whole op is undef.
971 auto *Op1C = dyn_cast<Constant>(Op1);
972 if (Op1C && Ty->isVectorTy()) {
973 unsigned NumElts = Ty->getVectorNumElements();
974 for (unsigned i = 0; i != NumElts; ++i) {
975 Constant *Elt = Op1C->getAggregateElement(i);
976 if (Elt && Elt->isNullValue())
977 return UndefValue::get(Ty);
983 if (match(Op0, m_Undef()))
984 return Constant::getNullValue(Ty);
988 if (match(Op0, m_Zero()))
994 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
998 // If this is a boolean op (single-bit element type), we can't have
999 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
1000 if (match(Op1, m_One()) || Ty->getScalarType()->isIntegerTy(1))
1001 return IsDiv ? Op0 : Constant::getNullValue(Ty);
1006 /// Given operands for an SDiv or UDiv, see if we can fold the result.
1007 /// If not, this returns null.
1008 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1009 const SimplifyQuery &Q, unsigned MaxRecurse) {
1010 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1013 if (Value *V = simplifyDivRem(Op0, Op1, true))
1016 bool isSigned = Opcode == Instruction::SDiv;
1018 // (X * Y) / Y -> X if the multiplication does not overflow.
1019 Value *X = nullptr, *Y = nullptr;
1020 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1021 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1022 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1023 // If the Mul knows it does not overflow, then we are good to go.
1024 if ((isSigned && Mul->hasNoSignedWrap()) ||
1025 (!isSigned && Mul->hasNoUnsignedWrap()))
1027 // If X has the form X = A / Y then X * Y cannot overflow.
1028 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1029 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1033 // (X rem Y) / Y -> 0
1034 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1035 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1036 return Constant::getNullValue(Op0->getType());
1038 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1039 ConstantInt *C1, *C2;
1040 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1041 match(Op1, m_ConstantInt(C2))) {
1043 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1045 return Constant::getNullValue(Op0->getType());
1048 // If the operation is with the result of a select instruction, check whether
1049 // operating on either branch of the select always yields the same value.
1050 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1051 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1054 // If the operation is with the result of a phi instruction, check whether
1055 // operating on all incoming values of the phi always yields the same value.
1056 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1057 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1063 /// Given operands for an SDiv, see if we can fold the result.
1064 /// If not, this returns null.
1065 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1066 unsigned MaxRecurse) {
1067 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1073 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1074 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1077 /// Given operands for a UDiv, see if we can fold the result.
1078 /// If not, this returns null.
1079 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1080 unsigned MaxRecurse) {
1081 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1084 // udiv %V, C -> 0 if %V < C
1086 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1087 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1088 if (C->isAllOnesValue()) {
1089 return Constant::getNullValue(Op0->getType());
1097 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1098 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1101 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1102 const SimplifyQuery &Q, unsigned) {
1103 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
1106 // undef / X -> undef (the undef could be a snan).
1107 if (match(Op0, m_Undef()))
1110 // X / undef -> undef
1111 if (match(Op1, m_Undef()))
1115 if (match(Op1, m_FPOne()))
1119 // Requires that NaNs are off (X could be zero) and signed zeroes are
1120 // ignored (X could be positive or negative, so the output sign is unknown).
1121 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1125 // X / X -> 1.0 is legal when NaNs are ignored.
1127 return ConstantFP::get(Op0->getType(), 1.0);
1129 // -X / X -> -1.0 and
1130 // X / -X -> -1.0 are legal when NaNs are ignored.
1131 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1132 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1133 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1134 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1135 BinaryOperator::getFNegArgument(Op1) == Op0))
1136 return ConstantFP::get(Op0->getType(), -1.0);
1142 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1143 const SimplifyQuery &Q) {
1144 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
1147 /// Given operands for an SRem or URem, see if we can fold the result.
1148 /// If not, this returns null.
1149 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1150 const SimplifyQuery &Q, unsigned MaxRecurse) {
1151 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1154 if (Value *V = simplifyDivRem(Op0, Op1, false))
1157 // (X % Y) % Y -> X % Y
1158 if ((Opcode == Instruction::SRem &&
1159 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1160 (Opcode == Instruction::URem &&
1161 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1164 // If the operation is with the result of a select instruction, check whether
1165 // operating on either branch of the select always yields the same value.
1166 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1167 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1170 // If the operation is with the result of a phi instruction, check whether
1171 // operating on all incoming values of the phi always yields the same value.
1172 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1173 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1179 /// Given operands for an SRem, see if we can fold the result.
1180 /// If not, this returns null.
1181 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1182 unsigned MaxRecurse) {
1183 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1189 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1190 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1193 /// Given operands for a URem, see if we can fold the result.
1194 /// If not, this returns null.
1195 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1196 unsigned MaxRecurse) {
1197 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1200 // urem %V, C -> %V if %V < C
1202 if (Constant *C = dyn_cast_or_null<Constant>(SimplifyICmpInst(
1203 ICmpInst::ICMP_ULT, Op0, Op1, Q, MaxRecurse - 1))) {
1204 if (C->isAllOnesValue()) {
1213 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1214 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1217 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1218 const SimplifyQuery &Q, unsigned) {
1219 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
1222 // undef % X -> undef (the undef could be a snan).
1223 if (match(Op0, m_Undef()))
1226 // X % undef -> undef
1227 if (match(Op1, m_Undef()))
1231 // Requires that NaNs are off (X could be zero) and signed zeroes are
1232 // ignored (X could be positive or negative, so the output sign is unknown).
1233 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1239 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1240 const SimplifyQuery &Q) {
1241 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
1244 /// Returns true if a shift by \c Amount always yields undef.
1245 static bool isUndefShift(Value *Amount) {
1246 Constant *C = dyn_cast<Constant>(Amount);
1250 // X shift by undef -> undef because it may shift by the bitwidth.
1251 if (isa<UndefValue>(C))
1254 // Shifting by the bitwidth or more is undefined.
1255 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1256 if (CI->getValue().getLimitedValue() >=
1257 CI->getType()->getScalarSizeInBits())
1260 // If all lanes of a vector shift are undefined the whole shift is.
1261 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1262 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1263 if (!isUndefShift(C->getAggregateElement(I)))
1271 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1272 /// If not, this returns null.
1273 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1274 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1275 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1278 // 0 shift by X -> 0
1279 if (match(Op0, m_Zero()))
1282 // X shift by 0 -> X
1283 if (match(Op1, m_Zero()))
1286 // Fold undefined shifts.
1287 if (isUndefShift(Op1))
1288 return UndefValue::get(Op0->getType());
1290 // If the operation is with the result of a select instruction, check whether
1291 // operating on either branch of the select always yields the same value.
1292 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1293 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1296 // If the operation is with the result of a phi instruction, check whether
1297 // operating on all incoming values of the phi always yields the same value.
1298 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1299 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1302 // If any bits in the shift amount make that value greater than or equal to
1303 // the number of bits in the type, the shift is undefined.
1304 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1305 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1306 return UndefValue::get(Op0->getType());
1308 // If all valid bits in the shift amount are known zero, the first operand is
1310 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1311 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1317 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1318 /// fold the result. If not, this returns null.
1319 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1320 Value *Op1, bool isExact, const SimplifyQuery &Q,
1321 unsigned MaxRecurse) {
1322 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1327 return Constant::getNullValue(Op0->getType());
1330 // undef >> X -> undef (if it's exact)
1331 if (match(Op0, m_Undef()))
1332 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1334 // The low bit cannot be shifted out of an exact shift if it is set.
1336 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1337 if (Op0Known.One[0])
1344 /// Given operands for an Shl, see if we can fold the result.
1345 /// If not, this returns null.
1346 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1347 const SimplifyQuery &Q, unsigned MaxRecurse) {
1348 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1352 // undef << X -> undef if (if it's NSW/NUW)
1353 if (match(Op0, m_Undef()))
1354 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1356 // (X >> A) << A -> X
1358 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1363 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1364 const SimplifyQuery &Q) {
1365 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1368 /// Given operands for an LShr, see if we can fold the result.
1369 /// If not, this returns null.
1370 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1371 const SimplifyQuery &Q, unsigned MaxRecurse) {
1372 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1376 // (X << A) >> A -> X
1378 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1384 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1385 const SimplifyQuery &Q) {
1386 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1389 /// Given operands for an AShr, see if we can fold the result.
1390 /// If not, this returns null.
1391 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1392 const SimplifyQuery &Q, unsigned MaxRecurse) {
1393 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1397 // all ones >>a X -> all ones
1398 if (match(Op0, m_AllOnes()))
1401 // (X << A) >> A -> X
1403 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1406 // Arithmetic shifting an all-sign-bit value is a no-op.
1407 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1408 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1414 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1415 const SimplifyQuery &Q) {
1416 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1419 /// Commuted variants are assumed to be handled by calling this function again
1420 /// with the parameters swapped.
1421 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1422 ICmpInst *UnsignedICmp, bool IsAnd) {
1425 ICmpInst::Predicate EqPred;
1426 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1427 !ICmpInst::isEquality(EqPred))
1430 ICmpInst::Predicate UnsignedPred;
1431 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1432 ICmpInst::isUnsigned(UnsignedPred))
1434 else if (match(UnsignedICmp,
1435 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1436 ICmpInst::isUnsigned(UnsignedPred))
1437 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1441 // X < Y && Y != 0 --> X < Y
1442 // X < Y || Y != 0 --> Y != 0
1443 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1444 return IsAnd ? UnsignedICmp : ZeroICmp;
1446 // X >= Y || Y != 0 --> true
1447 // X >= Y || Y == 0 --> X >= Y
1448 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1449 if (EqPred == ICmpInst::ICMP_NE)
1450 return getTrue(UnsignedICmp->getType());
1451 return UnsignedICmp;
1454 // X < Y && Y == 0 --> false
1455 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1457 return getFalse(UnsignedICmp->getType());
1462 /// Commuted variants are assumed to be handled by calling this function again
1463 /// with the parameters swapped.
1464 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1465 ICmpInst::Predicate Pred0, Pred1;
1467 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1468 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1471 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1472 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1473 // can eliminate Op1 from this 'and'.
1474 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1477 // Check for any combination of predicates that are guaranteed to be disjoint.
1478 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1479 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1480 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1481 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1482 return getFalse(Op0->getType());
1487 /// Commuted variants are assumed to be handled by calling this function again
1488 /// with the parameters swapped.
1489 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1490 ICmpInst::Predicate Pred0, Pred1;
1492 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1493 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1496 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1497 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1498 // can eliminate Op0 from this 'or'.
1499 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1502 // Check for any combination of predicates that cover the entire range of
1504 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1505 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1506 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1507 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1508 return getTrue(Op0->getType());
1513 /// Test if a pair of compares with a shared operand and 2 constants has an
1514 /// empty set intersection, full set union, or if one compare is a superset of
1516 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1518 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1519 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1522 const APInt *C0, *C1;
1523 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1524 !match(Cmp1->getOperand(1), m_APInt(C1)))
1527 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1528 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1530 // For and-of-compares, check if the intersection is empty:
1531 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1532 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1533 return getFalse(Cmp0->getType());
1535 // For or-of-compares, check if the union is full:
1536 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1537 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1538 return getTrue(Cmp0->getType());
1540 // Is one range a superset of the other?
1541 // If this is and-of-compares, take the smaller set:
1542 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1543 // If this is or-of-compares, take the larger set:
1544 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1545 if (Range0.contains(Range1))
1546 return IsAnd ? Cmp1 : Cmp0;
1547 if (Range1.contains(Range0))
1548 return IsAnd ? Cmp0 : Cmp1;
1553 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1554 // (icmp (add V, C0), C1) & (icmp V, C0)
1555 ICmpInst::Predicate Pred0, Pred1;
1556 const APInt *C0, *C1;
1558 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1561 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1564 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1565 if (AddInst->getOperand(1) != Op1->getOperand(1))
1568 Type *ITy = Op0->getType();
1569 bool isNSW = AddInst->hasNoSignedWrap();
1570 bool isNUW = AddInst->hasNoUnsignedWrap();
1572 const APInt Delta = *C1 - *C0;
1573 if (C0->isStrictlyPositive()) {
1575 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1576 return getFalse(ITy);
1577 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1578 return getFalse(ITy);
1581 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1582 return getFalse(ITy);
1583 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1584 return getFalse(ITy);
1587 if (C0->getBoolValue() && isNUW) {
1589 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1590 return getFalse(ITy);
1592 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1593 return getFalse(ITy);
1599 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1600 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1602 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1605 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1607 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1610 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1613 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1615 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1621 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1622 // (icmp (add V, C0), C1) | (icmp V, C0)
1623 ICmpInst::Predicate Pred0, Pred1;
1624 const APInt *C0, *C1;
1626 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1629 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1632 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1633 if (AddInst->getOperand(1) != Op1->getOperand(1))
1636 Type *ITy = Op0->getType();
1637 bool isNSW = AddInst->hasNoSignedWrap();
1638 bool isNUW = AddInst->hasNoUnsignedWrap();
1640 const APInt Delta = *C1 - *C0;
1641 if (C0->isStrictlyPositive()) {
1643 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1644 return getTrue(ITy);
1645 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1646 return getTrue(ITy);
1649 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1650 return getTrue(ITy);
1651 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1652 return getTrue(ITy);
1655 if (C0->getBoolValue() && isNUW) {
1657 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1658 return getTrue(ITy);
1660 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1661 return getTrue(ITy);
1667 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1668 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1670 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1673 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1675 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1678 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1681 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1683 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1689 static Value *simplifyAndOrOfICmps(Value *Op0, Value *Op1, bool IsAnd) {
1690 // Look through casts of the 'and' operands to find compares.
1691 auto *Cast0 = dyn_cast<CastInst>(Op0);
1692 auto *Cast1 = dyn_cast<CastInst>(Op1);
1693 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1694 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1695 Op0 = Cast0->getOperand(0);
1696 Op1 = Cast1->getOperand(0);
1699 auto *Cmp0 = dyn_cast<ICmpInst>(Op0);
1700 auto *Cmp1 = dyn_cast<ICmpInst>(Op1);
1705 IsAnd ? simplifyAndOfICmps(Cmp0, Cmp1) : simplifyOrOfICmps(Cmp0, Cmp1);
1711 // If we looked through casts, we can only handle a constant simplification
1712 // because we are not allowed to create a cast instruction here.
1713 if (auto *C = dyn_cast<Constant>(V))
1714 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1719 /// Given operands for an And, see if we can fold the result.
1720 /// If not, this returns null.
1721 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1722 unsigned MaxRecurse) {
1723 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1727 if (match(Op1, m_Undef()))
1728 return Constant::getNullValue(Op0->getType());
1735 if (match(Op1, m_Zero()))
1739 if (match(Op1, m_AllOnes()))
1742 // A & ~A = ~A & A = 0
1743 if (match(Op0, m_Not(m_Specific(Op1))) ||
1744 match(Op1, m_Not(m_Specific(Op0))))
1745 return Constant::getNullValue(Op0->getType());
1748 Value *A = nullptr, *B = nullptr;
1749 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1750 (A == Op1 || B == Op1))
1754 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1755 (A == Op0 || B == Op0))
1758 // A mask that only clears known zeros of a shifted value is a no-op.
1762 if (match(Op1, m_APInt(Mask))) {
1763 // If all bits in the inverted and shifted mask are clear:
1764 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1765 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1766 (~(*Mask)).lshr(*ShAmt).isNullValue())
1769 // If all bits in the inverted and shifted mask are clear:
1770 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1771 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1772 (~(*Mask)).shl(*ShAmt).isNullValue())
1776 // A & (-A) = A if A is a power of two or zero.
1777 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1778 match(Op1, m_Neg(m_Specific(Op0)))) {
1779 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1782 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1787 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, true))
1790 // Try some generic simplifications for associative operations.
1791 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1795 // And distributes over Or. Try some generic simplifications based on this.
1796 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1800 // And distributes over Xor. Try some generic simplifications based on this.
1801 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1805 // If the operation is with the result of a select instruction, check whether
1806 // operating on either branch of the select always yields the same value.
1807 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1808 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1812 // If the operation is with the result of a phi instruction, check whether
1813 // operating on all incoming values of the phi always yields the same value.
1814 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1815 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1822 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1823 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1826 /// Given operands for an Or, see if we can fold the result.
1827 /// If not, this returns null.
1828 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1829 unsigned MaxRecurse) {
1830 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1834 if (match(Op1, m_Undef()))
1835 return Constant::getAllOnesValue(Op0->getType());
1842 if (match(Op1, m_Zero()))
1846 if (match(Op1, m_AllOnes()))
1849 // A | ~A = ~A | A = -1
1850 if (match(Op0, m_Not(m_Specific(Op1))) ||
1851 match(Op1, m_Not(m_Specific(Op0))))
1852 return Constant::getAllOnesValue(Op0->getType());
1855 Value *A = nullptr, *B = nullptr;
1856 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1857 (A == Op1 || B == Op1))
1861 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1862 (A == Op0 || B == Op0))
1865 // ~(A & ?) | A = -1
1866 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1867 (A == Op1 || B == Op1))
1868 return Constant::getAllOnesValue(Op1->getType());
1870 // A | ~(A & ?) = -1
1871 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1872 (A == Op0 || B == Op0))
1873 return Constant::getAllOnesValue(Op0->getType());
1875 // (A & ~B) | (A ^ B) -> (A ^ B)
1876 // (~B & A) | (A ^ B) -> (A ^ B)
1877 // (A & ~B) | (B ^ A) -> (B ^ A)
1878 // (~B & A) | (B ^ A) -> (B ^ A)
1879 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1880 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1881 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1884 // Commute the 'or' operands.
1885 // (A ^ B) | (A & ~B) -> (A ^ B)
1886 // (A ^ B) | (~B & A) -> (A ^ B)
1887 // (B ^ A) | (A & ~B) -> (B ^ A)
1888 // (B ^ A) | (~B & A) -> (B ^ A)
1889 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1890 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1891 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1894 // (A & B) | (~A ^ B) -> (~A ^ B)
1895 // (B & A) | (~A ^ B) -> (~A ^ B)
1896 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1897 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1898 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1899 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1900 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1903 // (~A ^ B) | (A & B) -> (~A ^ B)
1904 // (~A ^ B) | (B & A) -> (~A ^ B)
1905 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1906 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1907 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1908 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1909 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1912 if (Value *V = simplifyAndOrOfICmps(Op0, Op1, false))
1915 // Try some generic simplifications for associative operations.
1916 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1920 // Or distributes over And. Try some generic simplifications based on this.
1921 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1925 // If the operation is with the result of a select instruction, check whether
1926 // operating on either branch of the select always yields the same value.
1927 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1928 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1932 // (A & C1)|(B & C2)
1933 const APInt *C1, *C2;
1934 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1935 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1937 // (A & C1)|(B & C2)
1938 // If we have: ((V + N) & C1) | (V & C2)
1939 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1940 // replace with V+N.
1942 if (C2->isMask() && // C2 == 0+1+
1943 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1944 // Add commutes, try both ways.
1945 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1948 // Or commutes, try both ways.
1950 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1951 // Add commutes, try both ways.
1952 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1958 // If the operation is with the result of a phi instruction, check whether
1959 // operating on all incoming values of the phi always yields the same value.
1960 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1961 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1967 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1968 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1971 /// Given operands for a Xor, see if we can fold the result.
1972 /// If not, this returns null.
1973 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1974 unsigned MaxRecurse) {
1975 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1978 // A ^ undef -> undef
1979 if (match(Op1, m_Undef()))
1983 if (match(Op1, m_Zero()))
1988 return Constant::getNullValue(Op0->getType());
1990 // A ^ ~A = ~A ^ A = -1
1991 if (match(Op0, m_Not(m_Specific(Op1))) ||
1992 match(Op1, m_Not(m_Specific(Op0))))
1993 return Constant::getAllOnesValue(Op0->getType());
1995 // Try some generic simplifications for associative operations.
1996 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2000 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2001 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2002 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2003 // only if B and C are equal. If B and C are equal then (since we assume
2004 // that operands have already been simplified) "select(cond, B, C)" should
2005 // have been simplified to the common value of B and C already. Analysing
2006 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2007 // for threading over phi nodes.
2012 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2013 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2017 static Type *GetCompareTy(Value *Op) {
2018 return CmpInst::makeCmpResultType(Op->getType());
2021 /// Rummage around inside V looking for something equivalent to the comparison
2022 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2023 /// Helper function for analyzing max/min idioms.
2024 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2025 Value *LHS, Value *RHS) {
2026 SelectInst *SI = dyn_cast<SelectInst>(V);
2029 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2032 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2033 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2035 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2036 LHS == CmpRHS && RHS == CmpLHS)
2041 // A significant optimization not implemented here is assuming that alloca
2042 // addresses are not equal to incoming argument values. They don't *alias*,
2043 // as we say, but that doesn't mean they aren't equal, so we take a
2044 // conservative approach.
2046 // This is inspired in part by C++11 5.10p1:
2047 // "Two pointers of the same type compare equal if and only if they are both
2048 // null, both point to the same function, or both represent the same
2051 // This is pretty permissive.
2053 // It's also partly due to C11 6.5.9p6:
2054 // "Two pointers compare equal if and only if both are null pointers, both are
2055 // pointers to the same object (including a pointer to an object and a
2056 // subobject at its beginning) or function, both are pointers to one past the
2057 // last element of the same array object, or one is a pointer to one past the
2058 // end of one array object and the other is a pointer to the start of a
2059 // different array object that happens to immediately follow the first array
2060 // object in the address space.)
2062 // C11's version is more restrictive, however there's no reason why an argument
2063 // couldn't be a one-past-the-end value for a stack object in the caller and be
2064 // equal to the beginning of a stack object in the callee.
2066 // If the C and C++ standards are ever made sufficiently restrictive in this
2067 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2068 // this optimization.
2070 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2071 const DominatorTree *DT, CmpInst::Predicate Pred,
2072 const Instruction *CxtI, Value *LHS, Value *RHS) {
2073 // First, skip past any trivial no-ops.
2074 LHS = LHS->stripPointerCasts();
2075 RHS = RHS->stripPointerCasts();
2077 // A non-null pointer is not equal to a null pointer.
2078 if (llvm::isKnownNonNull(LHS) && isa<ConstantPointerNull>(RHS) &&
2079 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2080 return ConstantInt::get(GetCompareTy(LHS),
2081 !CmpInst::isTrueWhenEqual(Pred));
2083 // We can only fold certain predicates on pointer comparisons.
2088 // Equality comaprisons are easy to fold.
2089 case CmpInst::ICMP_EQ:
2090 case CmpInst::ICMP_NE:
2093 // We can only handle unsigned relational comparisons because 'inbounds' on
2094 // a GEP only protects against unsigned wrapping.
2095 case CmpInst::ICMP_UGT:
2096 case CmpInst::ICMP_UGE:
2097 case CmpInst::ICMP_ULT:
2098 case CmpInst::ICMP_ULE:
2099 // However, we have to switch them to their signed variants to handle
2100 // negative indices from the base pointer.
2101 Pred = ICmpInst::getSignedPredicate(Pred);
2105 // Strip off any constant offsets so that we can reason about them.
2106 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2107 // here and compare base addresses like AliasAnalysis does, however there are
2108 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2109 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2110 // doesn't need to guarantee pointer inequality when it says NoAlias.
2111 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2112 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2114 // If LHS and RHS are related via constant offsets to the same base
2115 // value, we can replace it with an icmp which just compares the offsets.
2117 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2119 // Various optimizations for (in)equality comparisons.
2120 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2121 // Different non-empty allocations that exist at the same time have
2122 // different addresses (if the program can tell). Global variables always
2123 // exist, so they always exist during the lifetime of each other and all
2124 // allocas. Two different allocas usually have different addresses...
2126 // However, if there's an @llvm.stackrestore dynamically in between two
2127 // allocas, they may have the same address. It's tempting to reduce the
2128 // scope of the problem by only looking at *static* allocas here. That would
2129 // cover the majority of allocas while significantly reducing the likelihood
2130 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2131 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2132 // an entry block. Also, if we have a block that's not attached to a
2133 // function, we can't tell if it's "static" under the current definition.
2134 // Theoretically, this problem could be fixed by creating a new kind of
2135 // instruction kind specifically for static allocas. Such a new instruction
2136 // could be required to be at the top of the entry block, thus preventing it
2137 // from being subject to a @llvm.stackrestore. Instcombine could even
2138 // convert regular allocas into these special allocas. It'd be nifty.
2139 // However, until then, this problem remains open.
2141 // So, we'll assume that two non-empty allocas have different addresses
2144 // With all that, if the offsets are within the bounds of their allocations
2145 // (and not one-past-the-end! so we can't use inbounds!), and their
2146 // allocations aren't the same, the pointers are not equal.
2148 // Note that it's not necessary to check for LHS being a global variable
2149 // address, due to canonicalization and constant folding.
2150 if (isa<AllocaInst>(LHS) &&
2151 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2152 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2153 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2154 uint64_t LHSSize, RHSSize;
2155 if (LHSOffsetCI && RHSOffsetCI &&
2156 getObjectSize(LHS, LHSSize, DL, TLI) &&
2157 getObjectSize(RHS, RHSSize, DL, TLI)) {
2158 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2159 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2160 if (!LHSOffsetValue.isNegative() &&
2161 !RHSOffsetValue.isNegative() &&
2162 LHSOffsetValue.ult(LHSSize) &&
2163 RHSOffsetValue.ult(RHSSize)) {
2164 return ConstantInt::get(GetCompareTy(LHS),
2165 !CmpInst::isTrueWhenEqual(Pred));
2169 // Repeat the above check but this time without depending on DataLayout
2170 // or being able to compute a precise size.
2171 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2172 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2173 LHSOffset->isNullValue() &&
2174 RHSOffset->isNullValue())
2175 return ConstantInt::get(GetCompareTy(LHS),
2176 !CmpInst::isTrueWhenEqual(Pred));
2179 // Even if an non-inbounds GEP occurs along the path we can still optimize
2180 // equality comparisons concerning the result. We avoid walking the whole
2181 // chain again by starting where the last calls to
2182 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2183 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2184 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2186 return ConstantExpr::getICmp(Pred,
2187 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2188 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2190 // If one side of the equality comparison must come from a noalias call
2191 // (meaning a system memory allocation function), and the other side must
2192 // come from a pointer that cannot overlap with dynamically-allocated
2193 // memory within the lifetime of the current function (allocas, byval
2194 // arguments, globals), then determine the comparison result here.
2195 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2196 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2197 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2199 // Is the set of underlying objects all noalias calls?
2200 auto IsNAC = [](ArrayRef<Value *> Objects) {
2201 return all_of(Objects, isNoAliasCall);
2204 // Is the set of underlying objects all things which must be disjoint from
2205 // noalias calls. For allocas, we consider only static ones (dynamic
2206 // allocas might be transformed into calls to malloc not simultaneously
2207 // live with the compared-to allocation). For globals, we exclude symbols
2208 // that might be resolve lazily to symbols in another dynamically-loaded
2209 // library (and, thus, could be malloc'ed by the implementation).
2210 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2211 return all_of(Objects, [](Value *V) {
2212 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2213 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2214 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2215 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2216 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2217 !GV->isThreadLocal();
2218 if (const Argument *A = dyn_cast<Argument>(V))
2219 return A->hasByValAttr();
2224 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2225 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2226 return ConstantInt::get(GetCompareTy(LHS),
2227 !CmpInst::isTrueWhenEqual(Pred));
2229 // Fold comparisons for non-escaping pointer even if the allocation call
2230 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2231 // dynamic allocation call could be either of the operands.
2232 Value *MI = nullptr;
2233 if (isAllocLikeFn(LHS, TLI) && llvm::isKnownNonNullAt(RHS, CxtI, DT))
2235 else if (isAllocLikeFn(RHS, TLI) && llvm::isKnownNonNullAt(LHS, CxtI, DT))
2237 // FIXME: We should also fold the compare when the pointer escapes, but the
2238 // compare dominates the pointer escape
2239 if (MI && !PointerMayBeCaptured(MI, true, true))
2240 return ConstantInt::get(GetCompareTy(LHS),
2241 CmpInst::isFalseWhenEqual(Pred));
2248 /// Fold an icmp when its operands have i1 scalar type.
2249 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2250 Value *RHS, const SimplifyQuery &Q) {
2251 Type *ITy = GetCompareTy(LHS); // The return type.
2252 Type *OpTy = LHS->getType(); // The operand type.
2253 if (!OpTy->getScalarType()->isIntegerTy(1))
2256 // A boolean compared to true/false can be simplified in 14 out of the 20
2257 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2258 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2259 if (match(RHS, m_Zero())) {
2261 case CmpInst::ICMP_NE: // X != 0 -> X
2262 case CmpInst::ICMP_UGT: // X >u 0 -> X
2263 case CmpInst::ICMP_SLT: // X <s 0 -> X
2266 case CmpInst::ICMP_ULT: // X <u 0 -> false
2267 case CmpInst::ICMP_SGT: // X >s 0 -> false
2268 return getFalse(ITy);
2270 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2271 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2272 return getTrue(ITy);
2276 } else if (match(RHS, m_One())) {
2278 case CmpInst::ICMP_EQ: // X == 1 -> X
2279 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2280 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2283 case CmpInst::ICMP_UGT: // X >u 1 -> false
2284 case CmpInst::ICMP_SLT: // X <s -1 -> false
2285 return getFalse(ITy);
2287 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2288 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2289 return getTrue(ITy);
2298 case ICmpInst::ICMP_UGE:
2299 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2300 return getTrue(ITy);
2302 case ICmpInst::ICMP_SGE:
2303 /// For signed comparison, the values for an i1 are 0 and -1
2304 /// respectively. This maps into a truth table of:
2305 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2306 /// 0 | 0 | 1 (0 >= 0) | 1
2307 /// 0 | 1 | 1 (0 >= -1) | 1
2308 /// 1 | 0 | 0 (-1 >= 0) | 0
2309 /// 1 | 1 | 1 (-1 >= -1) | 1
2310 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2311 return getTrue(ITy);
2313 case ICmpInst::ICMP_ULE:
2314 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2315 return getTrue(ITy);
2322 /// Try hard to fold icmp with zero RHS because this is a common case.
2323 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2324 Value *RHS, const SimplifyQuery &Q) {
2325 if (!match(RHS, m_Zero()))
2328 Type *ITy = GetCompareTy(LHS); // The return type.
2331 llvm_unreachable("Unknown ICmp predicate!");
2332 case ICmpInst::ICMP_ULT:
2333 return getFalse(ITy);
2334 case ICmpInst::ICMP_UGE:
2335 return getTrue(ITy);
2336 case ICmpInst::ICMP_EQ:
2337 case ICmpInst::ICMP_ULE:
2338 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2339 return getFalse(ITy);
2341 case ICmpInst::ICMP_NE:
2342 case ICmpInst::ICMP_UGT:
2343 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2344 return getTrue(ITy);
2346 case ICmpInst::ICMP_SLT: {
2347 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2348 if (LHSKnown.isNegative())
2349 return getTrue(ITy);
2350 if (LHSKnown.isNonNegative())
2351 return getFalse(ITy);
2354 case ICmpInst::ICMP_SLE: {
2355 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2356 if (LHSKnown.isNegative())
2357 return getTrue(ITy);
2358 if (LHSKnown.isNonNegative() &&
2359 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2360 return getFalse(ITy);
2363 case ICmpInst::ICMP_SGE: {
2364 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2365 if (LHSKnown.isNegative())
2366 return getFalse(ITy);
2367 if (LHSKnown.isNonNegative())
2368 return getTrue(ITy);
2371 case ICmpInst::ICMP_SGT: {
2372 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2373 if (LHSKnown.isNegative())
2374 return getFalse(ITy);
2375 if (LHSKnown.isNonNegative() &&
2376 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2377 return getTrue(ITy);
2385 /// Many binary operators with a constant operand have an easy-to-compute
2386 /// range of outputs. This can be used to fold a comparison to always true or
2388 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2389 unsigned Width = Lower.getBitWidth();
2391 switch (BO.getOpcode()) {
2392 case Instruction::Add:
2393 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2394 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2395 if (BO.hasNoUnsignedWrap()) {
2396 // 'add nuw x, C' produces [C, UINT_MAX].
2398 } else if (BO.hasNoSignedWrap()) {
2399 if (C->isNegative()) {
2400 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2401 Lower = APInt::getSignedMinValue(Width);
2402 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2404 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2405 Lower = APInt::getSignedMinValue(Width) + *C;
2406 Upper = APInt::getSignedMaxValue(Width) + 1;
2412 case Instruction::And:
2413 if (match(BO.getOperand(1), m_APInt(C)))
2414 // 'and x, C' produces [0, C].
2418 case Instruction::Or:
2419 if (match(BO.getOperand(1), m_APInt(C)))
2420 // 'or x, C' produces [C, UINT_MAX].
2424 case Instruction::AShr:
2425 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2426 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2427 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2428 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2429 } else if (match(BO.getOperand(0), m_APInt(C))) {
2430 unsigned ShiftAmount = Width - 1;
2431 if (!C->isNullValue() && BO.isExact())
2432 ShiftAmount = C->countTrailingZeros();
2433 if (C->isNegative()) {
2434 // 'ashr C, x' produces [C, C >> (Width-1)]
2436 Upper = C->ashr(ShiftAmount) + 1;
2438 // 'ashr C, x' produces [C >> (Width-1), C]
2439 Lower = C->ashr(ShiftAmount);
2445 case Instruction::LShr:
2446 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2447 // 'lshr x, C' produces [0, UINT_MAX >> C].
2448 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2449 } else if (match(BO.getOperand(0), m_APInt(C))) {
2450 // 'lshr C, x' produces [C >> (Width-1), C].
2451 unsigned ShiftAmount = Width - 1;
2452 if (!C->isNullValue() && BO.isExact())
2453 ShiftAmount = C->countTrailingZeros();
2454 Lower = C->lshr(ShiftAmount);
2459 case Instruction::Shl:
2460 if (match(BO.getOperand(0), m_APInt(C))) {
2461 if (BO.hasNoUnsignedWrap()) {
2462 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2464 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2465 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2466 if (C->isNegative()) {
2467 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2468 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2469 Lower = C->shl(ShiftAmount);
2472 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2473 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2475 Upper = C->shl(ShiftAmount) + 1;
2481 case Instruction::SDiv:
2482 if (match(BO.getOperand(1), m_APInt(C))) {
2483 APInt IntMin = APInt::getSignedMinValue(Width);
2484 APInt IntMax = APInt::getSignedMaxValue(Width);
2485 if (C->isAllOnesValue()) {
2486 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2487 // where C != -1 and C != 0 and C != 1
2490 } else if (C->countLeadingZeros() < Width - 1) {
2491 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2492 // where C != -1 and C != 0 and C != 1
2493 Lower = IntMin.sdiv(*C);
2494 Upper = IntMax.sdiv(*C);
2495 if (Lower.sgt(Upper))
2496 std::swap(Lower, Upper);
2498 assert(Upper != Lower && "Upper part of range has wrapped!");
2500 } else if (match(BO.getOperand(0), m_APInt(C))) {
2501 if (C->isMinSignedValue()) {
2502 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2504 Upper = Lower.lshr(1) + 1;
2506 // 'sdiv C, x' produces [-|C|, |C|].
2507 Upper = C->abs() + 1;
2508 Lower = (-Upper) + 1;
2513 case Instruction::UDiv:
2514 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2515 // 'udiv x, C' produces [0, UINT_MAX / C].
2516 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2517 } else if (match(BO.getOperand(0), m_APInt(C))) {
2518 // 'udiv C, x' produces [0, C].
2523 case Instruction::SRem:
2524 if (match(BO.getOperand(1), m_APInt(C))) {
2525 // 'srem x, C' produces (-|C|, |C|).
2527 Lower = (-Upper) + 1;
2531 case Instruction::URem:
2532 if (match(BO.getOperand(1), m_APInt(C)))
2533 // 'urem x, C' produces [0, C).
2542 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2545 if (!match(RHS, m_APInt(C)))
2548 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2549 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2550 if (RHS_CR.isEmptySet())
2551 return ConstantInt::getFalse(GetCompareTy(RHS));
2552 if (RHS_CR.isFullSet())
2553 return ConstantInt::getTrue(GetCompareTy(RHS));
2555 // Find the range of possible values for binary operators.
2556 unsigned Width = C->getBitWidth();
2557 APInt Lower = APInt(Width, 0);
2558 APInt Upper = APInt(Width, 0);
2559 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2560 setLimitsForBinOp(*BO, Lower, Upper);
2562 ConstantRange LHS_CR =
2563 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2565 if (auto *I = dyn_cast<Instruction>(LHS))
2566 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2567 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2569 if (!LHS_CR.isFullSet()) {
2570 if (RHS_CR.contains(LHS_CR))
2571 return ConstantInt::getTrue(GetCompareTy(RHS));
2572 if (RHS_CR.inverse().contains(LHS_CR))
2573 return ConstantInt::getFalse(GetCompareTy(RHS));
2579 /// TODO: A large part of this logic is duplicated in InstCombine's
2580 /// foldICmpBinOp(). We should be able to share that and avoid the code
2582 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2583 Value *RHS, const SimplifyQuery &Q,
2584 unsigned MaxRecurse) {
2585 Type *ITy = GetCompareTy(LHS); // The return type.
2587 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2588 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2589 if (MaxRecurse && (LBO || RBO)) {
2590 // Analyze the case when either LHS or RHS is an add instruction.
2591 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2592 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2593 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2594 if (LBO && LBO->getOpcode() == Instruction::Add) {
2595 A = LBO->getOperand(0);
2596 B = LBO->getOperand(1);
2598 ICmpInst::isEquality(Pred) ||
2599 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2600 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2602 if (RBO && RBO->getOpcode() == Instruction::Add) {
2603 C = RBO->getOperand(0);
2604 D = RBO->getOperand(1);
2606 ICmpInst::isEquality(Pred) ||
2607 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2608 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2611 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2612 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2613 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2614 Constant::getNullValue(RHS->getType()), Q,
2618 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2619 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2621 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2622 C == LHS ? D : C, Q, MaxRecurse - 1))
2625 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2626 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2628 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2631 // C + B == C + D -> B == D
2634 } else if (A == D) {
2635 // D + B == C + D -> B == C
2638 } else if (B == C) {
2639 // A + C == C + D -> A == D
2644 // A + D == C + D -> A == C
2648 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2655 // icmp pred (or X, Y), X
2656 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2657 if (Pred == ICmpInst::ICMP_ULT)
2658 return getFalse(ITy);
2659 if (Pred == ICmpInst::ICMP_UGE)
2660 return getTrue(ITy);
2662 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2663 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2664 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2665 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2666 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2667 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2668 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2671 // icmp pred X, (or X, Y)
2672 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2673 if (Pred == ICmpInst::ICMP_ULE)
2674 return getTrue(ITy);
2675 if (Pred == ICmpInst::ICMP_UGT)
2676 return getFalse(ITy);
2678 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2679 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2680 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2681 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2682 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2683 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2684 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2689 // icmp pred (and X, Y), X
2690 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2691 if (Pred == ICmpInst::ICMP_UGT)
2692 return getFalse(ITy);
2693 if (Pred == ICmpInst::ICMP_ULE)
2694 return getTrue(ITy);
2696 // icmp pred X, (and X, Y)
2697 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2698 if (Pred == ICmpInst::ICMP_UGE)
2699 return getTrue(ITy);
2700 if (Pred == ICmpInst::ICMP_ULT)
2701 return getFalse(ITy);
2704 // 0 - (zext X) pred C
2705 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2706 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2707 if (RHSC->getValue().isStrictlyPositive()) {
2708 if (Pred == ICmpInst::ICMP_SLT)
2709 return ConstantInt::getTrue(RHSC->getContext());
2710 if (Pred == ICmpInst::ICMP_SGE)
2711 return ConstantInt::getFalse(RHSC->getContext());
2712 if (Pred == ICmpInst::ICMP_EQ)
2713 return ConstantInt::getFalse(RHSC->getContext());
2714 if (Pred == ICmpInst::ICMP_NE)
2715 return ConstantInt::getTrue(RHSC->getContext());
2717 if (RHSC->getValue().isNonNegative()) {
2718 if (Pred == ICmpInst::ICMP_SLE)
2719 return ConstantInt::getTrue(RHSC->getContext());
2720 if (Pred == ICmpInst::ICMP_SGT)
2721 return ConstantInt::getFalse(RHSC->getContext());
2726 // icmp pred (urem X, Y), Y
2727 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2731 case ICmpInst::ICMP_SGT:
2732 case ICmpInst::ICMP_SGE: {
2733 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2734 if (!Known.isNonNegative())
2738 case ICmpInst::ICMP_EQ:
2739 case ICmpInst::ICMP_UGT:
2740 case ICmpInst::ICMP_UGE:
2741 return getFalse(ITy);
2742 case ICmpInst::ICMP_SLT:
2743 case ICmpInst::ICMP_SLE: {
2744 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2745 if (!Known.isNonNegative())
2749 case ICmpInst::ICMP_NE:
2750 case ICmpInst::ICMP_ULT:
2751 case ICmpInst::ICMP_ULE:
2752 return getTrue(ITy);
2756 // icmp pred X, (urem Y, X)
2757 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2761 case ICmpInst::ICMP_SGT:
2762 case ICmpInst::ICMP_SGE: {
2763 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2764 if (!Known.isNonNegative())
2768 case ICmpInst::ICMP_NE:
2769 case ICmpInst::ICMP_UGT:
2770 case ICmpInst::ICMP_UGE:
2771 return getTrue(ITy);
2772 case ICmpInst::ICMP_SLT:
2773 case ICmpInst::ICMP_SLE: {
2774 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2775 if (!Known.isNonNegative())
2779 case ICmpInst::ICMP_EQ:
2780 case ICmpInst::ICMP_ULT:
2781 case ICmpInst::ICMP_ULE:
2782 return getFalse(ITy);
2788 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2789 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2790 // icmp pred (X op Y), X
2791 if (Pred == ICmpInst::ICMP_UGT)
2792 return getFalse(ITy);
2793 if (Pred == ICmpInst::ICMP_ULE)
2794 return getTrue(ITy);
2799 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2800 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2801 // icmp pred X, (X op Y)
2802 if (Pred == ICmpInst::ICMP_ULT)
2803 return getFalse(ITy);
2804 if (Pred == ICmpInst::ICMP_UGE)
2805 return getTrue(ITy);
2812 // where CI2 is a power of 2 and CI isn't
2813 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2814 const APInt *CI2Val, *CIVal = &CI->getValue();
2815 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2816 CI2Val->isPowerOf2()) {
2817 if (!CIVal->isPowerOf2()) {
2818 // CI2 << X can equal zero in some circumstances,
2819 // this simplification is unsafe if CI is zero.
2821 // We know it is safe if:
2822 // - The shift is nsw, we can't shift out the one bit.
2823 // - The shift is nuw, we can't shift out the one bit.
2826 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2827 CI2Val->isOneValue() || !CI->isZero()) {
2828 if (Pred == ICmpInst::ICMP_EQ)
2829 return ConstantInt::getFalse(RHS->getContext());
2830 if (Pred == ICmpInst::ICMP_NE)
2831 return ConstantInt::getTrue(RHS->getContext());
2834 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2835 if (Pred == ICmpInst::ICMP_UGT)
2836 return ConstantInt::getFalse(RHS->getContext());
2837 if (Pred == ICmpInst::ICMP_ULE)
2838 return ConstantInt::getTrue(RHS->getContext());
2843 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2844 LBO->getOperand(1) == RBO->getOperand(1)) {
2845 switch (LBO->getOpcode()) {
2848 case Instruction::UDiv:
2849 case Instruction::LShr:
2850 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2852 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2853 RBO->getOperand(0), Q, MaxRecurse - 1))
2856 case Instruction::SDiv:
2857 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2859 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2860 RBO->getOperand(0), Q, MaxRecurse - 1))
2863 case Instruction::AShr:
2864 if (!LBO->isExact() || !RBO->isExact())
2866 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2867 RBO->getOperand(0), Q, MaxRecurse - 1))
2870 case Instruction::Shl: {
2871 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2872 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2875 if (!NSW && ICmpInst::isSigned(Pred))
2877 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2878 RBO->getOperand(0), Q, MaxRecurse - 1))
2887 /// Simplify integer comparisons where at least one operand of the compare
2888 /// matches an integer min/max idiom.
2889 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2890 Value *RHS, const SimplifyQuery &Q,
2891 unsigned MaxRecurse) {
2892 Type *ITy = GetCompareTy(LHS); // The return type.
2894 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2895 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2897 // Signed variants on "max(a,b)>=a -> true".
2898 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2900 std::swap(A, B); // smax(A, B) pred A.
2901 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2902 // We analyze this as smax(A, B) pred A.
2904 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2905 (A == LHS || B == LHS)) {
2907 std::swap(A, B); // A pred smax(A, B).
2908 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2909 // We analyze this as smax(A, B) swapped-pred A.
2910 P = CmpInst::getSwappedPredicate(Pred);
2911 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2912 (A == RHS || B == RHS)) {
2914 std::swap(A, B); // smin(A, B) pred A.
2915 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2916 // We analyze this as smax(-A, -B) swapped-pred -A.
2917 // Note that we do not need to actually form -A or -B thanks to EqP.
2918 P = CmpInst::getSwappedPredicate(Pred);
2919 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2920 (A == LHS || B == LHS)) {
2922 std::swap(A, B); // A pred smin(A, B).
2923 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2924 // We analyze this as smax(-A, -B) pred -A.
2925 // Note that we do not need to actually form -A or -B thanks to EqP.
2928 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2929 // Cases correspond to "max(A, B) p A".
2933 case CmpInst::ICMP_EQ:
2934 case CmpInst::ICMP_SLE:
2935 // Equivalent to "A EqP B". This may be the same as the condition tested
2936 // in the max/min; if so, we can just return that.
2937 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2939 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2941 // Otherwise, see if "A EqP B" simplifies.
2943 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2946 case CmpInst::ICMP_NE:
2947 case CmpInst::ICMP_SGT: {
2948 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2949 // Equivalent to "A InvEqP B". This may be the same as the condition
2950 // tested in the max/min; if so, we can just return that.
2951 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2953 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2955 // Otherwise, see if "A InvEqP B" simplifies.
2957 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2961 case CmpInst::ICMP_SGE:
2963 return getTrue(ITy);
2964 case CmpInst::ICMP_SLT:
2966 return getFalse(ITy);
2970 // Unsigned variants on "max(a,b)>=a -> true".
2971 P = CmpInst::BAD_ICMP_PREDICATE;
2972 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2974 std::swap(A, B); // umax(A, B) pred A.
2975 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2976 // We analyze this as umax(A, B) pred A.
2978 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2979 (A == LHS || B == LHS)) {
2981 std::swap(A, B); // A pred umax(A, B).
2982 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2983 // We analyze this as umax(A, B) swapped-pred A.
2984 P = CmpInst::getSwappedPredicate(Pred);
2985 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2986 (A == RHS || B == RHS)) {
2988 std::swap(A, B); // umin(A, B) pred A.
2989 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2990 // We analyze this as umax(-A, -B) swapped-pred -A.
2991 // Note that we do not need to actually form -A or -B thanks to EqP.
2992 P = CmpInst::getSwappedPredicate(Pred);
2993 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2994 (A == LHS || B == LHS)) {
2996 std::swap(A, B); // A pred umin(A, B).
2997 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2998 // We analyze this as umax(-A, -B) pred -A.
2999 // Note that we do not need to actually form -A or -B thanks to EqP.
3002 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3003 // Cases correspond to "max(A, B) p A".
3007 case CmpInst::ICMP_EQ:
3008 case CmpInst::ICMP_ULE:
3009 // Equivalent to "A EqP B". This may be the same as the condition tested
3010 // in the max/min; if so, we can just return that.
3011 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3013 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3015 // Otherwise, see if "A EqP B" simplifies.
3017 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3020 case CmpInst::ICMP_NE:
3021 case CmpInst::ICMP_UGT: {
3022 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3023 // Equivalent to "A InvEqP B". This may be the same as the condition
3024 // tested in the max/min; if so, we can just return that.
3025 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3027 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3029 // Otherwise, see if "A InvEqP B" simplifies.
3031 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3035 case CmpInst::ICMP_UGE:
3037 return getTrue(ITy);
3038 case CmpInst::ICMP_ULT:
3040 return getFalse(ITy);
3044 // Variants on "max(x,y) >= min(x,z)".
3046 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3047 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3048 (A == C || A == D || B == C || B == D)) {
3049 // max(x, ?) pred min(x, ?).
3050 if (Pred == CmpInst::ICMP_SGE)
3052 return getTrue(ITy);
3053 if (Pred == CmpInst::ICMP_SLT)
3055 return getFalse(ITy);
3056 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3057 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3058 (A == C || A == D || B == C || B == D)) {
3059 // min(x, ?) pred max(x, ?).
3060 if (Pred == CmpInst::ICMP_SLE)
3062 return getTrue(ITy);
3063 if (Pred == CmpInst::ICMP_SGT)
3065 return getFalse(ITy);
3066 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3067 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3068 (A == C || A == D || B == C || B == D)) {
3069 // max(x, ?) pred min(x, ?).
3070 if (Pred == CmpInst::ICMP_UGE)
3072 return getTrue(ITy);
3073 if (Pred == CmpInst::ICMP_ULT)
3075 return getFalse(ITy);
3076 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3077 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3078 (A == C || A == D || B == C || B == D)) {
3079 // min(x, ?) pred max(x, ?).
3080 if (Pred == CmpInst::ICMP_ULE)
3082 return getTrue(ITy);
3083 if (Pred == CmpInst::ICMP_UGT)
3085 return getFalse(ITy);
3091 /// Given operands for an ICmpInst, see if we can fold the result.
3092 /// If not, this returns null.
3093 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3094 const SimplifyQuery &Q, unsigned MaxRecurse) {
3095 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3096 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3098 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3099 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3100 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3102 // If we have a constant, make sure it is on the RHS.
3103 std::swap(LHS, RHS);
3104 Pred = CmpInst::getSwappedPredicate(Pred);
3107 Type *ITy = GetCompareTy(LHS); // The return type.
3109 // icmp X, X -> true/false
3110 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3111 // because X could be 0.
3112 if (LHS == RHS || isa<UndefValue>(RHS))
3113 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3115 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3118 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3121 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3124 // If both operands have range metadata, use the metadata
3125 // to simplify the comparison.
3126 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3127 auto RHS_Instr = cast<Instruction>(RHS);
3128 auto LHS_Instr = cast<Instruction>(LHS);
3130 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3131 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3132 auto RHS_CR = getConstantRangeFromMetadata(
3133 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3134 auto LHS_CR = getConstantRangeFromMetadata(
3135 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3137 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3138 if (Satisfied_CR.contains(LHS_CR))
3139 return ConstantInt::getTrue(RHS->getContext());
3141 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3142 CmpInst::getInversePredicate(Pred), RHS_CR);
3143 if (InversedSatisfied_CR.contains(LHS_CR))
3144 return ConstantInt::getFalse(RHS->getContext());
3148 // Compare of cast, for example (zext X) != 0 -> X != 0
3149 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3150 Instruction *LI = cast<CastInst>(LHS);
3151 Value *SrcOp = LI->getOperand(0);
3152 Type *SrcTy = SrcOp->getType();
3153 Type *DstTy = LI->getType();
3155 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3156 // if the integer type is the same size as the pointer type.
3157 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3158 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3159 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3160 // Transfer the cast to the constant.
3161 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3162 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3165 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3166 if (RI->getOperand(0)->getType() == SrcTy)
3167 // Compare without the cast.
3168 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3174 if (isa<ZExtInst>(LHS)) {
3175 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3177 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3178 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3179 // Compare X and Y. Note that signed predicates become unsigned.
3180 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3181 SrcOp, RI->getOperand(0), Q,
3185 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3186 // too. If not, then try to deduce the result of the comparison.
3187 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3188 // Compute the constant that would happen if we truncated to SrcTy then
3189 // reextended to DstTy.
3190 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3191 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3193 // If the re-extended constant didn't change then this is effectively
3194 // also a case of comparing two zero-extended values.
3195 if (RExt == CI && MaxRecurse)
3196 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3197 SrcOp, Trunc, Q, MaxRecurse-1))
3200 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3201 // there. Use this to work out the result of the comparison.
3204 default: llvm_unreachable("Unknown ICmp predicate!");
3206 case ICmpInst::ICMP_EQ:
3207 case ICmpInst::ICMP_UGT:
3208 case ICmpInst::ICMP_UGE:
3209 return ConstantInt::getFalse(CI->getContext());
3211 case ICmpInst::ICMP_NE:
3212 case ICmpInst::ICMP_ULT:
3213 case ICmpInst::ICMP_ULE:
3214 return ConstantInt::getTrue(CI->getContext());
3216 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3217 // is non-negative then LHS <s RHS.
3218 case ICmpInst::ICMP_SGT:
3219 case ICmpInst::ICMP_SGE:
3220 return CI->getValue().isNegative() ?
3221 ConstantInt::getTrue(CI->getContext()) :
3222 ConstantInt::getFalse(CI->getContext());
3224 case ICmpInst::ICMP_SLT:
3225 case ICmpInst::ICMP_SLE:
3226 return CI->getValue().isNegative() ?
3227 ConstantInt::getFalse(CI->getContext()) :
3228 ConstantInt::getTrue(CI->getContext());
3234 if (isa<SExtInst>(LHS)) {
3235 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3237 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3238 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3239 // Compare X and Y. Note that the predicate does not change.
3240 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3244 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3245 // too. If not, then try to deduce the result of the comparison.
3246 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3247 // Compute the constant that would happen if we truncated to SrcTy then
3248 // reextended to DstTy.
3249 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3250 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3252 // If the re-extended constant didn't change then this is effectively
3253 // also a case of comparing two sign-extended values.
3254 if (RExt == CI && MaxRecurse)
3255 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3258 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3259 // bits there. Use this to work out the result of the comparison.
3262 default: llvm_unreachable("Unknown ICmp predicate!");
3263 case ICmpInst::ICMP_EQ:
3264 return ConstantInt::getFalse(CI->getContext());
3265 case ICmpInst::ICMP_NE:
3266 return ConstantInt::getTrue(CI->getContext());
3268 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3270 case ICmpInst::ICMP_SGT:
3271 case ICmpInst::ICMP_SGE:
3272 return CI->getValue().isNegative() ?
3273 ConstantInt::getTrue(CI->getContext()) :
3274 ConstantInt::getFalse(CI->getContext());
3275 case ICmpInst::ICMP_SLT:
3276 case ICmpInst::ICMP_SLE:
3277 return CI->getValue().isNegative() ?
3278 ConstantInt::getFalse(CI->getContext()) :
3279 ConstantInt::getTrue(CI->getContext());
3281 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3283 case ICmpInst::ICMP_UGT:
3284 case ICmpInst::ICMP_UGE:
3285 // Comparison is true iff the LHS <s 0.
3287 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3288 Constant::getNullValue(SrcTy),
3292 case ICmpInst::ICMP_ULT:
3293 case ICmpInst::ICMP_ULE:
3294 // Comparison is true iff the LHS >=s 0.
3296 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3297 Constant::getNullValue(SrcTy),
3307 // icmp eq|ne X, Y -> false|true if X != Y
3308 if (ICmpInst::isEquality(Pred) &&
3309 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3310 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3313 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3316 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3319 // Simplify comparisons of related pointers using a powerful, recursive
3320 // GEP-walk when we have target data available..
3321 if (LHS->getType()->isPointerTy())
3322 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI, LHS, RHS))
3324 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3325 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3326 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3327 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3328 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3329 Q.DL.getTypeSizeInBits(CRHS->getType()))
3330 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.CxtI,
3331 CLHS->getPointerOperand(),
3332 CRHS->getPointerOperand()))
3335 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3336 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3337 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3338 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3339 (ICmpInst::isEquality(Pred) ||
3340 (GLHS->isInBounds() && GRHS->isInBounds() &&
3341 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3342 // The bases are equal and the indices are constant. Build a constant
3343 // expression GEP with the same indices and a null base pointer to see
3344 // what constant folding can make out of it.
3345 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3346 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3347 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3348 GLHS->getSourceElementType(), Null, IndicesLHS);
3350 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3351 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3352 GLHS->getSourceElementType(), Null, IndicesRHS);
3353 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3358 // If the comparison is with the result of a select instruction, check whether
3359 // comparing with either branch of the select always yields the same value.
3360 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3361 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3364 // If the comparison is with the result of a phi instruction, check whether
3365 // doing the compare with each incoming phi value yields a common result.
3366 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3367 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3373 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3374 const SimplifyQuery &Q) {
3375 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3378 /// Given operands for an FCmpInst, see if we can fold the result.
3379 /// If not, this returns null.
3380 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3381 FastMathFlags FMF, const SimplifyQuery &Q,
3382 unsigned MaxRecurse) {
3383 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3384 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3386 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3387 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3388 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3390 // If we have a constant, make sure it is on the RHS.
3391 std::swap(LHS, RHS);
3392 Pred = CmpInst::getSwappedPredicate(Pred);
3395 // Fold trivial predicates.
3396 Type *RetTy = GetCompareTy(LHS);
3397 if (Pred == FCmpInst::FCMP_FALSE)
3398 return getFalse(RetTy);
3399 if (Pred == FCmpInst::FCMP_TRUE)
3400 return getTrue(RetTy);
3402 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3404 if (Pred == FCmpInst::FCMP_UNO)
3405 return getFalse(RetTy);
3406 if (Pred == FCmpInst::FCMP_ORD)
3407 return getTrue(RetTy);
3410 // fcmp pred x, undef and fcmp pred undef, x
3411 // fold to true if unordered, false if ordered
3412 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3413 // Choosing NaN for the undef will always make unordered comparison succeed
3414 // and ordered comparison fail.
3415 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3418 // fcmp x,x -> true/false. Not all compares are foldable.
3420 if (CmpInst::isTrueWhenEqual(Pred))
3421 return getTrue(RetTy);
3422 if (CmpInst::isFalseWhenEqual(Pred))
3423 return getFalse(RetTy);
3426 // Handle fcmp with constant RHS
3427 const ConstantFP *CFP = nullptr;
3428 if (const auto *RHSC = dyn_cast<Constant>(RHS)) {
3429 if (RHS->getType()->isVectorTy())
3430 CFP = dyn_cast_or_null<ConstantFP>(RHSC->getSplatValue());
3432 CFP = dyn_cast<ConstantFP>(RHSC);
3435 // If the constant is a nan, see if we can fold the comparison based on it.
3436 if (CFP->getValueAPF().isNaN()) {
3437 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3438 return getFalse(RetTy);
3439 assert(FCmpInst::isUnordered(Pred) &&
3440 "Comparison must be either ordered or unordered!");
3441 // True if unordered.
3442 return getTrue(RetTy);
3444 // Check whether the constant is an infinity.
3445 if (CFP->getValueAPF().isInfinity()) {
3446 if (CFP->getValueAPF().isNegative()) {
3448 case FCmpInst::FCMP_OLT:
3449 // No value is ordered and less than negative infinity.
3450 return getFalse(RetTy);
3451 case FCmpInst::FCMP_UGE:
3452 // All values are unordered with or at least negative infinity.
3453 return getTrue(RetTy);
3459 case FCmpInst::FCMP_OGT:
3460 // No value is ordered and greater than infinity.
3461 return getFalse(RetTy);
3462 case FCmpInst::FCMP_ULE:
3463 // All values are unordered with and at most infinity.
3464 return getTrue(RetTy);
3470 if (CFP->getValueAPF().isZero()) {
3472 case FCmpInst::FCMP_UGE:
3473 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3474 return getTrue(RetTy);
3476 case FCmpInst::FCMP_OLT:
3478 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3479 return getFalse(RetTy);
3487 // If the comparison is with the result of a select instruction, check whether
3488 // comparing with either branch of the select always yields the same value.
3489 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3490 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3493 // If the comparison is with the result of a phi instruction, check whether
3494 // doing the compare with each incoming phi value yields a common result.
3495 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3496 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3502 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3503 FastMathFlags FMF, const SimplifyQuery &Q) {
3504 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3507 /// See if V simplifies when its operand Op is replaced with RepOp.
3508 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3509 const SimplifyQuery &Q,
3510 unsigned MaxRecurse) {
3511 // Trivial replacement.
3515 // We cannot replace a constant, and shouldn't even try.
3516 if (isa<Constant>(Op))
3519 auto *I = dyn_cast<Instruction>(V);
3523 // If this is a binary operator, try to simplify it with the replaced op.
3524 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3526 // %cmp = icmp eq i32 %x, 2147483647
3527 // %add = add nsw i32 %x, 1
3528 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3530 // We can't replace %sel with %add unless we strip away the flags.
3531 if (isa<OverflowingBinaryOperator>(B))
3532 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3534 if (isa<PossiblyExactOperator>(B))
3539 if (B->getOperand(0) == Op)
3540 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3542 if (B->getOperand(1) == Op)
3543 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3548 // Same for CmpInsts.
3549 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3551 if (C->getOperand(0) == Op)
3552 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3554 if (C->getOperand(1) == Op)
3555 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3560 // TODO: We could hand off more cases to instsimplify here.
3562 // If all operands are constant after substituting Op for RepOp then we can
3563 // constant fold the instruction.
3564 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3565 // Build a list of all constant operands.
3566 SmallVector<Constant *, 8> ConstOps;
3567 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3568 if (I->getOperand(i) == Op)
3569 ConstOps.push_back(CRepOp);
3570 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3571 ConstOps.push_back(COp);
3576 // All operands were constants, fold it.
3577 if (ConstOps.size() == I->getNumOperands()) {
3578 if (CmpInst *C = dyn_cast<CmpInst>(I))
3579 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3580 ConstOps[1], Q.DL, Q.TLI);
3582 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3583 if (!LI->isVolatile())
3584 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3586 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3593 /// Try to simplify a select instruction when its condition operand is an
3594 /// integer comparison where one operand of the compare is a constant.
3595 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3596 const APInt *Y, bool TrueWhenUnset) {
3599 // (X & Y) == 0 ? X & ~Y : X --> X
3600 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3601 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3603 return TrueWhenUnset ? FalseVal : TrueVal;
3605 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3606 // (X & Y) != 0 ? X : X & ~Y --> X
3607 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3609 return TrueWhenUnset ? FalseVal : TrueVal;
3611 if (Y->isPowerOf2()) {
3612 // (X & Y) == 0 ? X | Y : X --> X | Y
3613 // (X & Y) != 0 ? X | Y : X --> X
3614 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3616 return TrueWhenUnset ? TrueVal : FalseVal;
3618 // (X & Y) == 0 ? X : X | Y --> X
3619 // (X & Y) != 0 ? X : X | Y --> X | Y
3620 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3622 return TrueWhenUnset ? TrueVal : FalseVal;
3628 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3630 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *TrueVal,
3632 bool TrueWhenUnset) {
3633 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
3637 APInt MinSignedValue;
3639 if (match(CmpLHS, m_Trunc(m_Value(X))) && (X == TrueVal || X == FalseVal)) {
3640 // icmp slt (trunc X), 0 <--> icmp ne (and X, C), 0
3641 // icmp sgt (trunc X), -1 <--> icmp eq (and X, C), 0
3642 unsigned DestSize = CmpLHS->getType()->getScalarSizeInBits();
3643 MinSignedValue = APInt::getSignedMinValue(DestSize).zext(BitWidth);
3645 // icmp slt X, 0 <--> icmp ne (and X, C), 0
3646 // icmp sgt X, -1 <--> icmp eq (and X, C), 0
3648 MinSignedValue = APInt::getSignedMinValue(BitWidth);
3651 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, &MinSignedValue,
3658 /// Try to simplify a select instruction when its condition operand is an
3659 /// integer comparison.
3660 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3661 Value *FalseVal, const SimplifyQuery &Q,
3662 unsigned MaxRecurse) {
3663 ICmpInst::Predicate Pred;
3664 Value *CmpLHS, *CmpRHS;
3665 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3668 // FIXME: This code is nearly duplicated in InstCombine. Using/refactoring
3669 // decomposeBitTestICmp() might help.
3670 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3673 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3674 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3675 Pred == ICmpInst::ICMP_EQ))
3677 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3678 // Comparing signed-less-than 0 checks if the sign bit is set.
3679 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3682 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3683 // Comparing signed-greater-than -1 checks if the sign bit is not set.
3684 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, TrueVal, FalseVal,
3689 if (CondVal->hasOneUse()) {
3691 if (match(CmpRHS, m_APInt(C))) {
3692 // X < MIN ? T : F --> F
3693 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3695 // X < MIN ? T : F --> F
3696 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3698 // X > MAX ? T : F --> F
3699 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3701 // X > MAX ? T : F --> F
3702 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3707 // If we have an equality comparison, then we know the value in one of the
3708 // arms of the select. See if substituting this value into the arm and
3709 // simplifying the result yields the same value as the other arm.
3710 if (Pred == ICmpInst::ICMP_EQ) {
3711 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3713 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3716 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3718 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3721 } else if (Pred == ICmpInst::ICMP_NE) {
3722 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3724 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3727 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3729 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3737 /// Given operands for a SelectInst, see if we can fold the result.
3738 /// If not, this returns null.
3739 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3740 Value *FalseVal, const SimplifyQuery &Q,
3741 unsigned MaxRecurse) {
3742 // select true, X, Y -> X
3743 // select false, X, Y -> Y
3744 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3745 if (CB->isAllOnesValue())
3747 if (CB->isNullValue())
3751 // select C, X, X -> X
3752 if (TrueVal == FalseVal)
3755 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3756 if (isa<Constant>(FalseVal))
3760 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3762 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3766 simplifySelectWithICmpCond(CondVal, TrueVal, FalseVal, Q, MaxRecurse))
3772 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3773 const SimplifyQuery &Q) {
3774 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3777 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3778 /// If not, this returns null.
3779 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3780 const SimplifyQuery &Q, unsigned) {
3781 // The type of the GEP pointer operand.
3783 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3785 // getelementptr P -> P.
3786 if (Ops.size() == 1)
3789 // Compute the (pointer) type returned by the GEP instruction.
3790 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3791 Type *GEPTy = PointerType::get(LastType, AS);
3792 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3793 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3794 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3795 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3797 if (isa<UndefValue>(Ops[0]))
3798 return UndefValue::get(GEPTy);
3800 if (Ops.size() == 2) {
3801 // getelementptr P, 0 -> P.
3802 if (match(Ops[1], m_Zero()))
3806 if (Ty->isSized()) {
3809 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3810 // getelementptr P, N -> P if P points to a type of zero size.
3811 if (TyAllocSize == 0)
3814 // The following transforms are only safe if the ptrtoint cast
3815 // doesn't truncate the pointers.
3816 if (Ops[1]->getType()->getScalarSizeInBits() ==
3817 Q.DL.getPointerSizeInBits(AS)) {
3818 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3819 if (match(P, m_Zero()))
3820 return Constant::getNullValue(GEPTy);
3822 if (match(P, m_PtrToInt(m_Value(Temp))))
3823 if (Temp->getType() == GEPTy)
3828 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3829 if (TyAllocSize == 1 &&
3830 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3831 if (Value *R = PtrToIntOrZero(P))
3834 // getelementptr V, (ashr (sub P, V), C) -> Q
3835 // if P points to a type of size 1 << C.
3837 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3838 m_ConstantInt(C))) &&
3839 TyAllocSize == 1ULL << C)
3840 if (Value *R = PtrToIntOrZero(P))
3843 // getelementptr V, (sdiv (sub P, V), C) -> Q
3844 // if P points to a type of size C.
3846 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3847 m_SpecificInt(TyAllocSize))))
3848 if (Value *R = PtrToIntOrZero(P))
3854 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3855 all_of(Ops.slice(1).drop_back(1),
3856 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3858 Q.DL.getPointerSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3859 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == PtrWidth) {
3860 APInt BasePtrOffset(PtrWidth, 0);
3861 Value *StrippedBasePtr =
3862 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3865 // gep (gep V, C), (sub 0, V) -> C
3866 if (match(Ops.back(),
3867 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3868 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3869 return ConstantExpr::getIntToPtr(CI, GEPTy);
3871 // gep (gep V, C), (xor V, -1) -> C-1
3872 if (match(Ops.back(),
3873 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3874 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3875 return ConstantExpr::getIntToPtr(CI, GEPTy);
3880 // Check to see if this is constant foldable.
3881 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3884 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3886 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3891 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3892 const SimplifyQuery &Q) {
3893 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3896 /// Given operands for an InsertValueInst, see if we can fold the result.
3897 /// If not, this returns null.
3898 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3899 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3901 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3902 if (Constant *CVal = dyn_cast<Constant>(Val))
3903 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3905 // insertvalue x, undef, n -> x
3906 if (match(Val, m_Undef()))
3909 // insertvalue x, (extractvalue y, n), n
3910 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3911 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3912 EV->getIndices() == Idxs) {
3913 // insertvalue undef, (extractvalue y, n), n -> y
3914 if (match(Agg, m_Undef()))
3915 return EV->getAggregateOperand();
3917 // insertvalue y, (extractvalue y, n), n -> y
3918 if (Agg == EV->getAggregateOperand())
3925 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3926 ArrayRef<unsigned> Idxs,
3927 const SimplifyQuery &Q) {
3928 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3931 /// Given operands for an ExtractValueInst, see if we can fold the result.
3932 /// If not, this returns null.
3933 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3934 const SimplifyQuery &, unsigned) {
3935 if (auto *CAgg = dyn_cast<Constant>(Agg))
3936 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3938 // extractvalue x, (insertvalue y, elt, n), n -> elt
3939 unsigned NumIdxs = Idxs.size();
3940 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3941 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3942 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3943 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3944 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3945 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3946 Idxs.slice(0, NumCommonIdxs)) {
3947 if (NumIdxs == NumInsertValueIdxs)
3948 return IVI->getInsertedValueOperand();
3956 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3957 const SimplifyQuery &Q) {
3958 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3961 /// Given operands for an ExtractElementInst, see if we can fold the result.
3962 /// If not, this returns null.
3963 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3965 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3966 if (auto *CIdx = dyn_cast<Constant>(Idx))
3967 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3969 // The index is not relevant if our vector is a splat.
3970 if (auto *Splat = CVec->getSplatValue())
3973 if (isa<UndefValue>(Vec))
3974 return UndefValue::get(Vec->getType()->getVectorElementType());
3977 // If extracting a specified index from the vector, see if we can recursively
3978 // find a previously computed scalar that was inserted into the vector.
3979 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3980 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3986 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3987 const SimplifyQuery &Q) {
3988 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3991 /// See if we can fold the given phi. If not, returns null.
3992 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3993 // If all of the PHI's incoming values are the same then replace the PHI node
3994 // with the common value.
3995 Value *CommonValue = nullptr;
3996 bool HasUndefInput = false;
3997 for (Value *Incoming : PN->incoming_values()) {
3998 // If the incoming value is the phi node itself, it can safely be skipped.
3999 if (Incoming == PN) continue;
4000 if (isa<UndefValue>(Incoming)) {
4001 // Remember that we saw an undef value, but otherwise ignore them.
4002 HasUndefInput = true;
4005 if (CommonValue && Incoming != CommonValue)
4006 return nullptr; // Not the same, bail out.
4007 CommonValue = Incoming;
4010 // If CommonValue is null then all of the incoming values were either undef or
4011 // equal to the phi node itself.
4013 return UndefValue::get(PN->getType());
4015 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4016 // instruction, we cannot return X as the result of the PHI node unless it
4017 // dominates the PHI block.
4019 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4024 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4025 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4026 if (auto *C = dyn_cast<Constant>(Op))
4027 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4029 if (auto *CI = dyn_cast<CastInst>(Op)) {
4030 auto *Src = CI->getOperand(0);
4031 Type *SrcTy = Src->getType();
4032 Type *MidTy = CI->getType();
4034 if (Src->getType() == Ty) {
4035 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4036 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4038 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4040 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4042 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4043 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4044 SrcIntPtrTy, MidIntPtrTy,
4045 DstIntPtrTy) == Instruction::BitCast)
4051 if (CastOpc == Instruction::BitCast)
4052 if (Op->getType() == Ty)
4058 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4059 const SimplifyQuery &Q) {
4060 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4063 /// For the given destination element of a shuffle, peek through shuffles to
4064 /// match a root vector source operand that contains that element in the same
4065 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4066 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4067 int MaskVal, Value *RootVec,
4068 unsigned MaxRecurse) {
4072 // Bail out if any mask value is undefined. That kind of shuffle may be
4073 // simplified further based on demanded bits or other folds.
4077 // The mask value chooses which source operand we need to look at next.
4078 int InVecNumElts = Op0->getType()->getVectorNumElements();
4079 int RootElt = MaskVal;
4080 Value *SourceOp = Op0;
4081 if (MaskVal >= InVecNumElts) {
4082 RootElt = MaskVal - InVecNumElts;
4086 // If the source operand is a shuffle itself, look through it to find the
4087 // matching root vector.
4088 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4089 return foldIdentityShuffles(
4090 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4091 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4094 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4097 // The source operand is not a shuffle. Initialize the root vector value for
4098 // this shuffle if that has not been done yet.
4102 // Give up as soon as a source operand does not match the existing root value.
4103 if (RootVec != SourceOp)
4106 // The element must be coming from the same lane in the source vector
4107 // (although it may have crossed lanes in intermediate shuffles).
4108 if (RootElt != DestElt)
4114 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4115 Type *RetTy, const SimplifyQuery &Q,
4116 unsigned MaxRecurse) {
4117 if (isa<UndefValue>(Mask))
4118 return UndefValue::get(RetTy);
4120 Type *InVecTy = Op0->getType();
4121 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4122 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4124 SmallVector<int, 32> Indices;
4125 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4126 assert(MaskNumElts == Indices.size() &&
4127 "Size of Indices not same as number of mask elements?");
4129 // Canonicalization: If mask does not select elements from an input vector,
4130 // replace that input vector with undef.
4131 bool MaskSelects0 = false, MaskSelects1 = false;
4132 for (unsigned i = 0; i != MaskNumElts; ++i) {
4133 if (Indices[i] == -1)
4135 if ((unsigned)Indices[i] < InVecNumElts)
4136 MaskSelects0 = true;
4138 MaskSelects1 = true;
4141 Op0 = UndefValue::get(InVecTy);
4143 Op1 = UndefValue::get(InVecTy);
4145 auto *Op0Const = dyn_cast<Constant>(Op0);
4146 auto *Op1Const = dyn_cast<Constant>(Op1);
4148 // If all operands are constant, constant fold the shuffle.
4149 if (Op0Const && Op1Const)
4150 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4152 // Canonicalization: if only one input vector is constant, it shall be the
4154 if (Op0Const && !Op1Const) {
4155 std::swap(Op0, Op1);
4156 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4159 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4160 // value type is same as the input vectors' type.
4161 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4162 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4163 OpShuf->getMask()->getSplatValue())
4166 // Don't fold a shuffle with undef mask elements. This may get folded in a
4167 // better way using demanded bits or other analysis.
4168 // TODO: Should we allow this?
4169 if (find(Indices, -1) != Indices.end())
4172 // Check if every element of this shuffle can be mapped back to the
4173 // corresponding element of a single root vector. If so, we don't need this
4174 // shuffle. This handles simple identity shuffles as well as chains of
4175 // shuffles that may widen/narrow and/or move elements across lanes and back.
4176 Value *RootVec = nullptr;
4177 for (unsigned i = 0; i != MaskNumElts; ++i) {
4178 // Note that recursion is limited for each vector element, so if any element
4179 // exceeds the limit, this will fail to simplify.
4181 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4183 // We can't replace a widening/narrowing shuffle with one of its operands.
4184 if (!RootVec || RootVec->getType() != RetTy)
4190 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4191 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4192 Type *RetTy, const SimplifyQuery &Q) {
4193 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4196 //=== Helper functions for higher up the class hierarchy.
4198 /// Given operands for a BinaryOperator, see if we can fold the result.
4199 /// If not, this returns null.
4200 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4201 const SimplifyQuery &Q, unsigned MaxRecurse) {
4203 case Instruction::Add:
4204 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4205 case Instruction::FAdd:
4206 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4207 case Instruction::Sub:
4208 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4209 case Instruction::FSub:
4210 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4211 case Instruction::Mul:
4212 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4213 case Instruction::FMul:
4214 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4215 case Instruction::SDiv:
4216 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4217 case Instruction::UDiv:
4218 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4219 case Instruction::FDiv:
4220 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4221 case Instruction::SRem:
4222 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4223 case Instruction::URem:
4224 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4225 case Instruction::FRem:
4226 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4227 case Instruction::Shl:
4228 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4229 case Instruction::LShr:
4230 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4231 case Instruction::AShr:
4232 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4233 case Instruction::And:
4234 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4235 case Instruction::Or:
4236 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4237 case Instruction::Xor:
4238 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4240 llvm_unreachable("Unexpected opcode");
4244 /// Given operands for a BinaryOperator, see if we can fold the result.
4245 /// If not, this returns null.
4246 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4247 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4248 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4249 const FastMathFlags &FMF, const SimplifyQuery &Q,
4250 unsigned MaxRecurse) {
4252 case Instruction::FAdd:
4253 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4254 case Instruction::FSub:
4255 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4256 case Instruction::FMul:
4257 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4258 case Instruction::FDiv:
4259 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4261 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4265 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4266 const SimplifyQuery &Q) {
4267 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4270 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4271 FastMathFlags FMF, const SimplifyQuery &Q) {
4272 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4275 /// Given operands for a CmpInst, see if we can fold the result.
4276 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4277 const SimplifyQuery &Q, unsigned MaxRecurse) {
4278 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4279 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4280 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4283 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4284 const SimplifyQuery &Q) {
4285 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4288 static bool IsIdempotent(Intrinsic::ID ID) {
4290 default: return false;
4292 // Unary idempotent: f(f(x)) = f(x)
4293 case Intrinsic::fabs:
4294 case Intrinsic::floor:
4295 case Intrinsic::ceil:
4296 case Intrinsic::trunc:
4297 case Intrinsic::rint:
4298 case Intrinsic::nearbyint:
4299 case Intrinsic::round:
4304 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4305 const DataLayout &DL) {
4306 GlobalValue *PtrSym;
4308 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4311 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4312 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4313 Type *Int32PtrTy = Int32Ty->getPointerTo();
4314 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4316 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4317 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4320 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4321 if (OffsetInt % 4 != 0)
4324 Constant *C = ConstantExpr::getGetElementPtr(
4325 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4326 ConstantInt::get(Int64Ty, OffsetInt / 4));
4327 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4331 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4335 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4336 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4341 if (LoadedCE->getOpcode() != Instruction::Sub)
4344 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4345 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4347 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4349 Constant *LoadedRHS = LoadedCE->getOperand(1);
4350 GlobalValue *LoadedRHSSym;
4351 APInt LoadedRHSOffset;
4352 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4354 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4357 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4360 static bool maskIsAllZeroOrUndef(Value *Mask) {
4361 auto *ConstMask = dyn_cast<Constant>(Mask);
4364 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4366 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4368 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4369 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4376 template <typename IterTy>
4377 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4378 const SimplifyQuery &Q, unsigned MaxRecurse) {
4379 Intrinsic::ID IID = F->getIntrinsicID();
4380 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4383 if (NumOperands == 1) {
4384 // Perform idempotent optimizations
4385 if (IsIdempotent(IID)) {
4386 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4387 if (II->getIntrinsicID() == IID)
4393 case Intrinsic::fabs: {
4394 if (SignBitMustBeZero(*ArgBegin, Q.TLI))
4404 if (NumOperands == 2) {
4405 Value *LHS = *ArgBegin;
4406 Value *RHS = *(ArgBegin + 1);
4407 Type *ReturnType = F->getReturnType();
4410 case Intrinsic::usub_with_overflow:
4411 case Intrinsic::ssub_with_overflow: {
4412 // X - X -> { 0, false }
4414 return Constant::getNullValue(ReturnType);
4416 // X - undef -> undef
4417 // undef - X -> undef
4418 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4419 return UndefValue::get(ReturnType);
4423 case Intrinsic::uadd_with_overflow:
4424 case Intrinsic::sadd_with_overflow: {
4425 // X + undef -> undef
4426 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4427 return UndefValue::get(ReturnType);
4431 case Intrinsic::umul_with_overflow:
4432 case Intrinsic::smul_with_overflow: {
4433 // 0 * X -> { 0, false }
4434 // X * 0 -> { 0, false }
4435 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4436 return Constant::getNullValue(ReturnType);
4438 // undef * X -> { 0, false }
4439 // X * undef -> { 0, false }
4440 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4441 return Constant::getNullValue(ReturnType);
4445 case Intrinsic::load_relative: {
4446 Constant *C0 = dyn_cast<Constant>(LHS);
4447 Constant *C1 = dyn_cast<Constant>(RHS);
4449 return SimplifyRelativeLoad(C0, C1, Q.DL);
4457 // Simplify calls to llvm.masked.load.*
4459 case Intrinsic::masked_load: {
4460 Value *MaskArg = ArgBegin[2];
4461 Value *PassthruArg = ArgBegin[3];
4462 // If the mask is all zeros or undef, the "passthru" argument is the result.
4463 if (maskIsAllZeroOrUndef(MaskArg))
4472 template <typename IterTy>
4473 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4474 IterTy ArgEnd, const SimplifyQuery &Q,
4475 unsigned MaxRecurse) {
4476 Type *Ty = V->getType();
4477 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4478 Ty = PTy->getElementType();
4479 FunctionType *FTy = cast<FunctionType>(Ty);
4481 // call undef -> undef
4482 // call null -> undef
4483 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4484 return UndefValue::get(FTy->getReturnType());
4486 Function *F = dyn_cast<Function>(V);
4490 if (F->isIntrinsic())
4491 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4494 if (!canConstantFoldCallTo(CS, F))
4497 SmallVector<Constant *, 4> ConstantArgs;
4498 ConstantArgs.reserve(ArgEnd - ArgBegin);
4499 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4500 Constant *C = dyn_cast<Constant>(*I);
4503 ConstantArgs.push_back(C);
4506 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4509 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4510 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4511 const SimplifyQuery &Q) {
4512 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4515 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4516 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4517 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4520 /// See if we can compute a simplified version of this instruction.
4521 /// If not, this returns null.
4523 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4524 OptimizationRemarkEmitter *ORE) {
4525 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4528 switch (I->getOpcode()) {
4530 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4532 case Instruction::FAdd:
4533 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4534 I->getFastMathFlags(), Q);
4536 case Instruction::Add:
4537 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4538 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4539 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4541 case Instruction::FSub:
4542 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4543 I->getFastMathFlags(), Q);
4545 case Instruction::Sub:
4546 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4547 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4548 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4550 case Instruction::FMul:
4551 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4552 I->getFastMathFlags(), Q);
4554 case Instruction::Mul:
4555 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4557 case Instruction::SDiv:
4558 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4560 case Instruction::UDiv:
4561 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4563 case Instruction::FDiv:
4564 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4565 I->getFastMathFlags(), Q);
4567 case Instruction::SRem:
4568 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4570 case Instruction::URem:
4571 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4573 case Instruction::FRem:
4574 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4575 I->getFastMathFlags(), Q);
4577 case Instruction::Shl:
4578 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4579 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4580 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4582 case Instruction::LShr:
4583 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4584 cast<BinaryOperator>(I)->isExact(), Q);
4586 case Instruction::AShr:
4587 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4588 cast<BinaryOperator>(I)->isExact(), Q);
4590 case Instruction::And:
4591 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4593 case Instruction::Or:
4594 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4596 case Instruction::Xor:
4597 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4599 case Instruction::ICmp:
4600 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4601 I->getOperand(0), I->getOperand(1), Q);
4603 case Instruction::FCmp:
4605 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4606 I->getOperand(1), I->getFastMathFlags(), Q);
4608 case Instruction::Select:
4609 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4610 I->getOperand(2), Q);
4612 case Instruction::GetElementPtr: {
4613 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4614 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4618 case Instruction::InsertValue: {
4619 InsertValueInst *IV = cast<InsertValueInst>(I);
4620 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4621 IV->getInsertedValueOperand(),
4622 IV->getIndices(), Q);
4625 case Instruction::ExtractValue: {
4626 auto *EVI = cast<ExtractValueInst>(I);
4627 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4628 EVI->getIndices(), Q);
4631 case Instruction::ExtractElement: {
4632 auto *EEI = cast<ExtractElementInst>(I);
4633 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4634 EEI->getIndexOperand(), Q);
4637 case Instruction::ShuffleVector: {
4638 auto *SVI = cast<ShuffleVectorInst>(I);
4639 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4640 SVI->getMask(), SVI->getType(), Q);
4643 case Instruction::PHI:
4644 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4646 case Instruction::Call: {
4647 CallSite CS(cast<CallInst>(I));
4648 Result = SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4652 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4653 #include "llvm/IR/Instruction.def"
4654 #undef HANDLE_CAST_INST
4656 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4658 case Instruction::Alloca:
4659 // No simplifications for Alloca and it can't be constant folded.
4664 // In general, it is possible for computeKnownBits to determine all bits in a
4665 // value even when the operands are not all constants.
4666 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4667 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4668 if (Known.isConstant())
4669 Result = ConstantInt::get(I->getType(), Known.getConstant());
4672 /// If called on unreachable code, the above logic may report that the
4673 /// instruction simplified to itself. Make life easier for users by
4674 /// detecting that case here, returning a safe value instead.
4675 return Result == I ? UndefValue::get(I->getType()) : Result;
4678 /// \brief Implementation of recursive simplification through an instruction's
4681 /// This is the common implementation of the recursive simplification routines.
4682 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4683 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4684 /// instructions to process and attempt to simplify it using
4685 /// InstructionSimplify.
4687 /// This routine returns 'true' only when *it* simplifies something. The passed
4688 /// in simplified value does not count toward this.
4689 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4690 const TargetLibraryInfo *TLI,
4691 const DominatorTree *DT,
4692 AssumptionCache *AC) {
4693 bool Simplified = false;
4694 SmallSetVector<Instruction *, 8> Worklist;
4695 const DataLayout &DL = I->getModule()->getDataLayout();
4697 // If we have an explicit value to collapse to, do that round of the
4698 // simplification loop by hand initially.
4700 for (User *U : I->users())
4702 Worklist.insert(cast<Instruction>(U));
4704 // Replace the instruction with its simplified value.
4705 I->replaceAllUsesWith(SimpleV);
4707 // Gracefully handle edge cases where the instruction is not wired into any
4709 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4710 !I->mayHaveSideEffects())
4711 I->eraseFromParent();
4716 // Note that we must test the size on each iteration, the worklist can grow.
4717 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4720 // See if this instruction simplifies.
4721 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4727 // Stash away all the uses of the old instruction so we can check them for
4728 // recursive simplifications after a RAUW. This is cheaper than checking all
4729 // uses of To on the recursive step in most cases.
4730 for (User *U : I->users())
4731 Worklist.insert(cast<Instruction>(U));
4733 // Replace the instruction with its simplified value.
4734 I->replaceAllUsesWith(SimpleV);
4736 // Gracefully handle edge cases where the instruction is not wired into any
4738 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4739 !I->mayHaveSideEffects())
4740 I->eraseFromParent();
4745 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4746 const TargetLibraryInfo *TLI,
4747 const DominatorTree *DT,
4748 AssumptionCache *AC) {
4749 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4752 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4753 const TargetLibraryInfo *TLI,
4754 const DominatorTree *DT,
4755 AssumptionCache *AC) {
4756 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4757 assert(SimpleV && "Must provide a simplified value.");
4758 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4762 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4763 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4764 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4765 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4766 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4767 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4768 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4769 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4772 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
4773 const DataLayout &DL) {
4774 return {DL, &AR.TLI, &AR.DT, &AR.AC};
4777 template <class T, class... TArgs>
4778 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
4780 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
4781 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
4782 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
4783 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4785 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,