1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements routines for folding instructions into simpler forms
11 // that do not require creating new instructions. This does constant folding
12 // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13 // returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14 // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15 // simplified: This is usually true and assuming it simplifies the logic (if
16 // they have not been simplified then results are correct but maybe suboptimal).
18 //===----------------------------------------------------------------------===//
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/CaptureTracking.h"
26 #include "llvm/Analysis/CmpInstAnalysis.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/LoopAnalysisManager.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/Analysis/VectorUtils.h"
32 #include "llvm/IR/ConstantRange.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/GetElementPtrTypeIterator.h"
36 #include "llvm/IR/GlobalAlias.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/IR/ValueHandle.h"
40 #include "llvm/Support/KnownBits.h"
43 using namespace llvm::PatternMatch;
45 #define DEBUG_TYPE "instsimplify"
47 enum { RecursionLimit = 3 };
49 STATISTIC(NumExpand, "Number of expansions");
50 STATISTIC(NumReassoc, "Number of reassociations");
52 static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
53 static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
55 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
56 const SimplifyQuery &, unsigned);
57 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
59 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
60 const SimplifyQuery &Q, unsigned MaxRecurse);
61 static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
62 static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 static Value *SimplifyCastInst(unsigned, Value *, Type *,
64 const SimplifyQuery &, unsigned);
65 static Value *SimplifyGEPInst(Type *, ArrayRef<Value *>, const SimplifyQuery &,
68 static Value *foldSelectWithBinaryOp(Value *Cond, Value *TrueVal,
70 BinaryOperator::BinaryOps BinOpCode;
71 if (auto *BO = dyn_cast<BinaryOperator>(Cond))
72 BinOpCode = BO->getOpcode();
76 CmpInst::Predicate ExpectedPred, Pred1, Pred2;
77 if (BinOpCode == BinaryOperator::Or) {
78 ExpectedPred = ICmpInst::ICMP_NE;
79 } else if (BinOpCode == BinaryOperator::And) {
80 ExpectedPred = ICmpInst::ICMP_EQ;
84 // %A = icmp eq %TV, %FV
85 // %B = icmp eq %X, %Y (and one of these is a select operand)
87 // %D = select %C, %TV, %FV
91 // %A = icmp ne %TV, %FV
92 // %B = icmp ne %X, %Y (and one of these is a select operand)
94 // %D = select %C, %TV, %FV
98 if (!match(Cond, m_c_BinOp(m_c_ICmp(Pred1, m_Specific(TrueVal),
99 m_Specific(FalseVal)),
100 m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
101 Pred1 != Pred2 || Pred1 != ExpectedPred)
104 if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
105 return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
110 /// For a boolean type or a vector of boolean type, return false or a vector
111 /// with every element false.
112 static Constant *getFalse(Type *Ty) {
113 return ConstantInt::getFalse(Ty);
116 /// For a boolean type or a vector of boolean type, return true or a vector
117 /// with every element true.
118 static Constant *getTrue(Type *Ty) {
119 return ConstantInt::getTrue(Ty);
122 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
123 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
125 CmpInst *Cmp = dyn_cast<CmpInst>(V);
128 CmpInst::Predicate CPred = Cmp->getPredicate();
129 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
130 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
132 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
136 /// Does the given value dominate the specified phi node?
137 static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
138 Instruction *I = dyn_cast<Instruction>(V);
140 // Arguments and constants dominate all instructions.
143 // If we are processing instructions (and/or basic blocks) that have not been
144 // fully added to a function, the parent nodes may still be null. Simply
145 // return the conservative answer in these cases.
146 if (!I->getParent() || !P->getParent() || !I->getFunction())
149 // If we have a DominatorTree then do a precise test.
151 return DT->dominates(I, P);
153 // Otherwise, if the instruction is in the entry block and is not an invoke,
154 // then it obviously dominates all phi nodes.
155 if (I->getParent() == &I->getFunction()->getEntryBlock() &&
162 /// Simplify "A op (B op' C)" by distributing op over op', turning it into
163 /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
164 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
165 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
166 /// Returns the simplified value, or null if no simplification was performed.
167 static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
168 Instruction::BinaryOps OpcodeToExpand,
169 const SimplifyQuery &Q, unsigned MaxRecurse) {
170 // Recursion is always used, so bail out at once if we already hit the limit.
174 // Check whether the expression has the form "(A op' B) op C".
175 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
176 if (Op0->getOpcode() == OpcodeToExpand) {
177 // It does! Try turning it into "(A op C) op' (B op C)".
178 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
179 // Do "A op C" and "B op C" both simplify?
180 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
181 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
182 // They do! Return "L op' R" if it simplifies or is already available.
183 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
184 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
185 && L == B && R == A)) {
189 // Otherwise return "L op' R" if it simplifies.
190 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
197 // Check whether the expression has the form "A op (B op' C)".
198 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
199 if (Op1->getOpcode() == OpcodeToExpand) {
200 // It does! Try turning it into "(A op B) op' (A op C)".
201 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
202 // Do "A op B" and "A op C" both simplify?
203 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
204 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
205 // They do! Return "L op' R" if it simplifies or is already available.
206 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
207 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
208 && L == C && R == B)) {
212 // Otherwise return "L op' R" if it simplifies.
213 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
223 /// Generic simplifications for associative binary operations.
224 /// Returns the simpler value, or null if none was found.
225 static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
226 Value *LHS, Value *RHS,
227 const SimplifyQuery &Q,
228 unsigned MaxRecurse) {
229 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
231 // Recursion is always used, so bail out at once if we already hit the limit.
235 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
236 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
238 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
239 if (Op0 && Op0->getOpcode() == Opcode) {
240 Value *A = Op0->getOperand(0);
241 Value *B = Op0->getOperand(1);
244 // Does "B op C" simplify?
245 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
246 // It does! Return "A op V" if it simplifies or is already available.
247 // If V equals B then "A op V" is just the LHS.
248 if (V == B) return LHS;
249 // Otherwise return "A op V" if it simplifies.
250 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
257 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
258 if (Op1 && Op1->getOpcode() == Opcode) {
260 Value *B = Op1->getOperand(0);
261 Value *C = Op1->getOperand(1);
263 // Does "A op B" simplify?
264 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
265 // It does! Return "V op C" if it simplifies or is already available.
266 // If V equals B then "V op C" is just the RHS.
267 if (V == B) return RHS;
268 // Otherwise return "V op C" if it simplifies.
269 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
276 // The remaining transforms require commutativity as well as associativity.
277 if (!Instruction::isCommutative(Opcode))
280 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
281 if (Op0 && Op0->getOpcode() == Opcode) {
282 Value *A = Op0->getOperand(0);
283 Value *B = Op0->getOperand(1);
286 // Does "C op A" simplify?
287 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
288 // It does! Return "V op B" if it simplifies or is already available.
289 // If V equals A then "V op B" is just the LHS.
290 if (V == A) return LHS;
291 // Otherwise return "V op B" if it simplifies.
292 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
299 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
300 if (Op1 && Op1->getOpcode() == Opcode) {
302 Value *B = Op1->getOperand(0);
303 Value *C = Op1->getOperand(1);
305 // Does "C op A" simplify?
306 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
307 // It does! Return "B op V" if it simplifies or is already available.
308 // If V equals C then "B op V" is just the RHS.
309 if (V == C) return RHS;
310 // Otherwise return "B op V" if it simplifies.
311 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
321 /// In the case of a binary operation with a select instruction as an operand,
322 /// try to simplify the binop by seeing whether evaluating it on both branches
323 /// of the select results in the same value. Returns the common value if so,
324 /// otherwise returns null.
325 static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
326 Value *RHS, const SimplifyQuery &Q,
327 unsigned MaxRecurse) {
328 // Recursion is always used, so bail out at once if we already hit the limit.
333 if (isa<SelectInst>(LHS)) {
334 SI = cast<SelectInst>(LHS);
336 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
337 SI = cast<SelectInst>(RHS);
340 // Evaluate the BinOp on the true and false branches of the select.
344 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
345 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
347 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
348 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
351 // If they simplified to the same value, then return the common value.
352 // If they both failed to simplify then return null.
356 // If one branch simplified to undef, return the other one.
357 if (TV && isa<UndefValue>(TV))
359 if (FV && isa<UndefValue>(FV))
362 // If applying the operation did not change the true and false select values,
363 // then the result of the binop is the select itself.
364 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
367 // If one branch simplified and the other did not, and the simplified
368 // value is equal to the unsimplified one, return the simplified value.
369 // For example, select (cond, X, X & Z) & Z -> X & Z.
370 if ((FV && !TV) || (TV && !FV)) {
371 // Check that the simplified value has the form "X op Y" where "op" is the
372 // same as the original operation.
373 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
374 if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
375 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
376 // We already know that "op" is the same as for the simplified value. See
377 // if the operands match too. If so, return the simplified value.
378 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
379 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
380 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
381 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
382 Simplified->getOperand(1) == UnsimplifiedRHS)
384 if (Simplified->isCommutative() &&
385 Simplified->getOperand(1) == UnsimplifiedLHS &&
386 Simplified->getOperand(0) == UnsimplifiedRHS)
394 /// In the case of a comparison with a select instruction, try to simplify the
395 /// comparison by seeing whether both branches of the select result in the same
396 /// value. Returns the common value if so, otherwise returns null.
397 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
398 Value *RHS, const SimplifyQuery &Q,
399 unsigned MaxRecurse) {
400 // Recursion is always used, so bail out at once if we already hit the limit.
404 // Make sure the select is on the LHS.
405 if (!isa<SelectInst>(LHS)) {
407 Pred = CmpInst::getSwappedPredicate(Pred);
409 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
410 SelectInst *SI = cast<SelectInst>(LHS);
411 Value *Cond = SI->getCondition();
412 Value *TV = SI->getTrueValue();
413 Value *FV = SI->getFalseValue();
415 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
416 // Does "cmp TV, RHS" simplify?
417 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
419 // It not only simplified, it simplified to the select condition. Replace
421 TCmp = getTrue(Cond->getType());
423 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
424 // condition then we can replace it with 'true'. Otherwise give up.
425 if (!isSameCompare(Cond, Pred, TV, RHS))
427 TCmp = getTrue(Cond->getType());
430 // Does "cmp FV, RHS" simplify?
431 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
433 // It not only simplified, it simplified to the select condition. Replace
435 FCmp = getFalse(Cond->getType());
437 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
438 // condition then we can replace it with 'false'. Otherwise give up.
439 if (!isSameCompare(Cond, Pred, FV, RHS))
441 FCmp = getFalse(Cond->getType());
444 // If both sides simplified to the same value, then use it as the result of
445 // the original comparison.
449 // The remaining cases only make sense if the select condition has the same
450 // type as the result of the comparison, so bail out if this is not so.
451 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
453 // If the false value simplified to false, then the result of the compare
454 // is equal to "Cond && TCmp". This also catches the case when the false
455 // value simplified to false and the true value to true, returning "Cond".
456 if (match(FCmp, m_Zero()))
457 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
459 // If the true value simplified to true, then the result of the compare
460 // is equal to "Cond || FCmp".
461 if (match(TCmp, m_One()))
462 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
464 // Finally, if the false value simplified to true and the true value to
465 // false, then the result of the compare is equal to "!Cond".
466 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
468 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
475 /// In the case of a binary operation with an operand that is a PHI instruction,
476 /// try to simplify the binop by seeing whether evaluating it on the incoming
477 /// phi values yields the same result for every value. If so returns the common
478 /// value, otherwise returns null.
479 static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
480 Value *RHS, const SimplifyQuery &Q,
481 unsigned MaxRecurse) {
482 // Recursion is always used, so bail out at once if we already hit the limit.
487 if (isa<PHINode>(LHS)) {
488 PI = cast<PHINode>(LHS);
489 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
490 if (!valueDominatesPHI(RHS, PI, Q.DT))
493 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
494 PI = cast<PHINode>(RHS);
495 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
496 if (!valueDominatesPHI(LHS, PI, Q.DT))
500 // Evaluate the BinOp on the incoming phi values.
501 Value *CommonValue = nullptr;
502 for (Value *Incoming : PI->incoming_values()) {
503 // If the incoming value is the phi node itself, it can safely be skipped.
504 if (Incoming == PI) continue;
505 Value *V = PI == LHS ?
506 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
507 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
508 // If the operation failed to simplify, or simplified to a different value
509 // to previously, then give up.
510 if (!V || (CommonValue && V != CommonValue))
518 /// In the case of a comparison with a PHI instruction, try to simplify the
519 /// comparison by seeing whether comparing with all of the incoming phi values
520 /// yields the same result every time. If so returns the common result,
521 /// otherwise returns null.
522 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
523 const SimplifyQuery &Q, unsigned MaxRecurse) {
524 // Recursion is always used, so bail out at once if we already hit the limit.
528 // Make sure the phi is on the LHS.
529 if (!isa<PHINode>(LHS)) {
531 Pred = CmpInst::getSwappedPredicate(Pred);
533 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
534 PHINode *PI = cast<PHINode>(LHS);
536 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
537 if (!valueDominatesPHI(RHS, PI, Q.DT))
540 // Evaluate the BinOp on the incoming phi values.
541 Value *CommonValue = nullptr;
542 for (Value *Incoming : PI->incoming_values()) {
543 // If the incoming value is the phi node itself, it can safely be skipped.
544 if (Incoming == PI) continue;
545 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
546 // If the operation failed to simplify, or simplified to a different value
547 // to previously, then give up.
548 if (!V || (CommonValue && V != CommonValue))
556 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
557 Value *&Op0, Value *&Op1,
558 const SimplifyQuery &Q) {
559 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
560 if (auto *CRHS = dyn_cast<Constant>(Op1))
561 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
563 // Canonicalize the constant to the RHS if this is a commutative operation.
564 if (Instruction::isCommutative(Opcode))
570 /// Given operands for an Add, see if we can fold the result.
571 /// If not, this returns null.
572 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
573 const SimplifyQuery &Q, unsigned MaxRecurse) {
574 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
577 // X + undef -> undef
578 if (match(Op1, m_Undef()))
582 if (match(Op1, m_Zero()))
585 // If two operands are negative, return 0.
586 if (isKnownNegation(Op0, Op1))
587 return Constant::getNullValue(Op0->getType());
593 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
594 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
597 // X + ~X -> -1 since ~X = -X-1
598 Type *Ty = Op0->getType();
599 if (match(Op0, m_Not(m_Specific(Op1))) ||
600 match(Op1, m_Not(m_Specific(Op0))))
601 return Constant::getAllOnesValue(Ty);
603 // add nsw/nuw (xor Y, signmask), signmask --> Y
604 // The no-wrapping add guarantees that the top bit will be set by the add.
605 // Therefore, the xor must be clearing the already set sign bit of Y.
606 if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
607 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
610 // add nuw %x, -1 -> -1, because %x can only be 0.
611 if (IsNUW && match(Op1, m_AllOnes()))
612 return Op1; // Which is -1.
615 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
616 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
619 // Try some generic simplifications for associative operations.
620 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
624 // Threading Add over selects and phi nodes is pointless, so don't bother.
625 // Threading over the select in "A + select(cond, B, C)" means evaluating
626 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
627 // only if B and C are equal. If B and C are equal then (since we assume
628 // that operands have already been simplified) "select(cond, B, C)" should
629 // have been simplified to the common value of B and C already. Analysing
630 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
631 // for threading over phi nodes.
636 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
637 const SimplifyQuery &Query) {
638 return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
641 /// Compute the base pointer and cumulative constant offsets for V.
643 /// This strips all constant offsets off of V, leaving it the base pointer, and
644 /// accumulates the total constant offset applied in the returned constant. It
645 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
646 /// no constant offsets applied.
648 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
649 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
651 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
652 bool AllowNonInbounds = false) {
653 assert(V->getType()->isPtrOrPtrVectorTy());
655 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
656 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
658 // Even though we don't look through PHI nodes, we could be called on an
659 // instruction in an unreachable block, which may be on a cycle.
660 SmallPtrSet<Value *, 4> Visited;
663 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
664 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
665 !GEP->accumulateConstantOffset(DL, Offset))
667 V = GEP->getPointerOperand();
668 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
669 V = cast<Operator>(V)->getOperand(0);
670 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
671 if (GA->isInterposable())
673 V = GA->getAliasee();
675 if (auto CS = CallSite(V))
676 if (Value *RV = CS.getReturnedArgOperand()) {
682 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
683 } while (Visited.insert(V).second);
685 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
686 if (V->getType()->isVectorTy())
687 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
692 /// Compute the constant difference between two pointer values.
693 /// If the difference is not a constant, returns zero.
694 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
696 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
697 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
699 // If LHS and RHS are not related via constant offsets to the same base
700 // value, there is nothing we can do here.
704 // Otherwise, the difference of LHS - RHS can be computed as:
706 // = (LHSOffset + Base) - (RHSOffset + Base)
707 // = LHSOffset - RHSOffset
708 return ConstantExpr::getSub(LHSOffset, RHSOffset);
711 /// Given operands for a Sub, see if we can fold the result.
712 /// If not, this returns null.
713 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
714 const SimplifyQuery &Q, unsigned MaxRecurse) {
715 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
718 // X - undef -> undef
719 // undef - X -> undef
720 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
721 return UndefValue::get(Op0->getType());
724 if (match(Op1, m_Zero()))
729 return Constant::getNullValue(Op0->getType());
731 // Is this a negation?
732 if (match(Op0, m_Zero())) {
733 // 0 - X -> 0 if the sub is NUW.
735 return Constant::getNullValue(Op0->getType());
737 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
738 if (Known.Zero.isMaxSignedValue()) {
739 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
740 // Op1 must be 0 because negating the minimum signed value is undefined.
742 return Constant::getNullValue(Op0->getType());
744 // 0 - X -> X if X is 0 or the minimum signed value.
749 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
750 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
751 Value *X = nullptr, *Y = nullptr, *Z = Op1;
752 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
753 // See if "V === Y - Z" simplifies.
754 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
755 // It does! Now see if "X + V" simplifies.
756 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
757 // It does, we successfully reassociated!
761 // See if "V === X - Z" simplifies.
762 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
763 // It does! Now see if "Y + V" simplifies.
764 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
765 // It does, we successfully reassociated!
771 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
772 // For example, X - (X + 1) -> -1
774 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
775 // See if "V === X - Y" simplifies.
776 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
777 // It does! Now see if "V - Z" simplifies.
778 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
779 // It does, we successfully reassociated!
783 // See if "V === X - Z" simplifies.
784 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
785 // It does! Now see if "V - Y" simplifies.
786 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
787 // It does, we successfully reassociated!
793 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
794 // For example, X - (X - Y) -> Y.
796 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
797 // See if "V === Z - X" simplifies.
798 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
799 // It does! Now see if "V + Y" simplifies.
800 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
801 // It does, we successfully reassociated!
806 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
807 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
808 match(Op1, m_Trunc(m_Value(Y))))
809 if (X->getType() == Y->getType())
810 // See if "V === X - Y" simplifies.
811 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
812 // It does! Now see if "trunc V" simplifies.
813 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
815 // It does, return the simplified "trunc V".
818 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
819 if (match(Op0, m_PtrToInt(m_Value(X))) &&
820 match(Op1, m_PtrToInt(m_Value(Y))))
821 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
822 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
825 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
826 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
829 // Threading Sub over selects and phi nodes is pointless, so don't bother.
830 // Threading over the select in "A - select(cond, B, C)" means evaluating
831 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
832 // only if B and C are equal. If B and C are equal then (since we assume
833 // that operands have already been simplified) "select(cond, B, C)" should
834 // have been simplified to the common value of B and C already. Analysing
835 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
836 // for threading over phi nodes.
841 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
842 const SimplifyQuery &Q) {
843 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
846 /// Given operands for a Mul, see if we can fold the result.
847 /// If not, this returns null.
848 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
849 unsigned MaxRecurse) {
850 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
855 if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
856 return Constant::getNullValue(Op0->getType());
859 if (match(Op1, m_One()))
862 // (X / Y) * Y -> X if the division is exact.
864 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
865 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
869 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
870 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
873 // Try some generic simplifications for associative operations.
874 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
878 // Mul distributes over Add. Try some generic simplifications based on this.
879 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
883 // If the operation is with the result of a select instruction, check whether
884 // operating on either branch of the select always yields the same value.
885 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
886 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
890 // If the operation is with the result of a phi instruction, check whether
891 // operating on all incoming values of the phi always yields the same value.
892 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
893 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
900 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
901 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
904 /// Check for common or similar folds of integer division or integer remainder.
905 /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
906 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
907 Type *Ty = Op0->getType();
909 // X / undef -> undef
910 // X % undef -> undef
911 if (match(Op1, m_Undef()))
916 // We don't need to preserve faults!
917 if (match(Op1, m_Zero()))
918 return UndefValue::get(Ty);
920 // If any element of a constant divisor vector is zero or undef, the whole op
922 auto *Op1C = dyn_cast<Constant>(Op1);
923 if (Op1C && Ty->isVectorTy()) {
924 unsigned NumElts = Ty->getVectorNumElements();
925 for (unsigned i = 0; i != NumElts; ++i) {
926 Constant *Elt = Op1C->getAggregateElement(i);
927 if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
928 return UndefValue::get(Ty);
934 if (match(Op0, m_Undef()))
935 return Constant::getNullValue(Ty);
939 if (match(Op0, m_Zero()))
940 return Constant::getNullValue(Op0->getType());
945 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
949 // If this is a boolean op (single-bit element type), we can't have
950 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
951 // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
953 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
954 (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
955 return IsDiv ? Op0 : Constant::getNullValue(Ty);
960 /// Given a predicate and two operands, return true if the comparison is true.
961 /// This is a helper for div/rem simplification where we return some other value
962 /// when we can prove a relationship between the operands.
963 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
964 const SimplifyQuery &Q, unsigned MaxRecurse) {
965 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
966 Constant *C = dyn_cast_or_null<Constant>(V);
967 return (C && C->isAllOnesValue());
970 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
971 /// to simplify X % Y to X.
972 static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
973 unsigned MaxRecurse, bool IsSigned) {
974 // Recursion is always used, so bail out at once if we already hit the limit.
981 // We require that 1 operand is a simple constant. That could be extended to
982 // 2 variables if we computed the sign bit for each.
984 // Make sure that a constant is not the minimum signed value because taking
985 // the abs() of that is undefined.
986 Type *Ty = X->getType();
988 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
989 // Is the variable divisor magnitude always greater than the constant
990 // dividend magnitude?
991 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
992 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
993 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
994 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
995 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
998 if (match(Y, m_APInt(C))) {
999 // Special-case: we can't take the abs() of a minimum signed value. If
1000 // that's the divisor, then all we have to do is prove that the dividend
1001 // is also not the minimum signed value.
1002 if (C->isMinSignedValue())
1003 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1005 // Is the variable dividend magnitude always less than the constant
1006 // divisor magnitude?
1007 // |X| < |C| --> X > -abs(C) and X < abs(C)
1008 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1009 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1010 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1011 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1017 // IsSigned == false.
1018 // Is the dividend unsigned less than the divisor?
1019 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1022 /// These are simplifications common to SDiv and UDiv.
1023 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1024 const SimplifyQuery &Q, unsigned MaxRecurse) {
1025 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1028 if (Value *V = simplifyDivRem(Op0, Op1, true))
1031 bool IsSigned = Opcode == Instruction::SDiv;
1033 // (X * Y) / Y -> X if the multiplication does not overflow.
1035 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1036 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1037 // If the Mul does not overflow, then we are good to go.
1038 if ((IsSigned && Mul->hasNoSignedWrap()) ||
1039 (!IsSigned && Mul->hasNoUnsignedWrap()))
1041 // If X has the form X = A / Y, then X * Y cannot overflow.
1042 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1043 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1047 // (X rem Y) / Y -> 0
1048 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1049 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1050 return Constant::getNullValue(Op0->getType());
1052 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1053 ConstantInt *C1, *C2;
1054 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1055 match(Op1, m_ConstantInt(C2))) {
1057 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1059 return Constant::getNullValue(Op0->getType());
1062 // If the operation is with the result of a select instruction, check whether
1063 // operating on either branch of the select always yields the same value.
1064 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1065 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1068 // If the operation is with the result of a phi instruction, check whether
1069 // operating on all incoming values of the phi always yields the same value.
1070 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1071 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1074 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1075 return Constant::getNullValue(Op0->getType());
1080 /// These are simplifications common to SRem and URem.
1081 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1082 const SimplifyQuery &Q, unsigned MaxRecurse) {
1083 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1086 if (Value *V = simplifyDivRem(Op0, Op1, false))
1089 // (X % Y) % Y -> X % Y
1090 if ((Opcode == Instruction::SRem &&
1091 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1092 (Opcode == Instruction::URem &&
1093 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1096 // (X << Y) % X -> 0
1097 if ((Opcode == Instruction::SRem &&
1098 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1099 (Opcode == Instruction::URem &&
1100 match(Op0, m_NUWShl(m_Specific(Op1), m_Value()))))
1101 return Constant::getNullValue(Op0->getType());
1103 // If the operation is with the result of a select instruction, check whether
1104 // operating on either branch of the select always yields the same value.
1105 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1106 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1109 // If the operation is with the result of a phi instruction, check whether
1110 // operating on all incoming values of the phi always yields the same value.
1111 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1112 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1115 // If X / Y == 0, then X % Y == X.
1116 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1122 /// Given operands for an SDiv, see if we can fold the result.
1123 /// If not, this returns null.
1124 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1125 unsigned MaxRecurse) {
1126 // If two operands are negated and no signed overflow, return -1.
1127 if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1128 return Constant::getAllOnesValue(Op0->getType());
1130 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1133 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1134 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1137 /// Given operands for a UDiv, see if we can fold the result.
1138 /// If not, this returns null.
1139 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1140 unsigned MaxRecurse) {
1141 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1144 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1145 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1148 /// Given operands for an SRem, see if we can fold the result.
1149 /// If not, this returns null.
1150 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1151 unsigned MaxRecurse) {
1152 // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1153 // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1155 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1156 return ConstantInt::getNullValue(Op0->getType());
1158 // If the two operands are negated, return 0.
1159 if (isKnownNegation(Op0, Op1))
1160 return ConstantInt::getNullValue(Op0->getType());
1162 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1165 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1166 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1169 /// Given operands for a URem, see if we can fold the result.
1170 /// If not, this returns null.
1171 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1172 unsigned MaxRecurse) {
1173 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1176 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1177 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1180 /// Returns true if a shift by \c Amount always yields undef.
1181 static bool isUndefShift(Value *Amount) {
1182 Constant *C = dyn_cast<Constant>(Amount);
1186 // X shift by undef -> undef because it may shift by the bitwidth.
1187 if (isa<UndefValue>(C))
1190 // Shifting by the bitwidth or more is undefined.
1191 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1192 if (CI->getValue().getLimitedValue() >=
1193 CI->getType()->getScalarSizeInBits())
1196 // If all lanes of a vector shift are undefined the whole shift is.
1197 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1198 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1199 if (!isUndefShift(C->getAggregateElement(I)))
1207 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1208 /// If not, this returns null.
1209 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1210 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1211 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1214 // 0 shift by X -> 0
1215 if (match(Op0, m_Zero()))
1216 return Constant::getNullValue(Op0->getType());
1218 // X shift by 0 -> X
1219 // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1222 if (match(Op1, m_Zero()) ||
1223 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1226 // Fold undefined shifts.
1227 if (isUndefShift(Op1))
1228 return UndefValue::get(Op0->getType());
1230 // If the operation is with the result of a select instruction, check whether
1231 // operating on either branch of the select always yields the same value.
1232 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1233 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1236 // If the operation is with the result of a phi instruction, check whether
1237 // operating on all incoming values of the phi always yields the same value.
1238 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1239 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1242 // If any bits in the shift amount make that value greater than or equal to
1243 // the number of bits in the type, the shift is undefined.
1244 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1245 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1246 return UndefValue::get(Op0->getType());
1248 // If all valid bits in the shift amount are known zero, the first operand is
1250 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1251 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1257 /// Given operands for an Shl, LShr or AShr, see if we can
1258 /// fold the result. If not, this returns null.
1259 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1260 Value *Op1, bool isExact, const SimplifyQuery &Q,
1261 unsigned MaxRecurse) {
1262 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1267 return Constant::getNullValue(Op0->getType());
1270 // undef >> X -> undef (if it's exact)
1271 if (match(Op0, m_Undef()))
1272 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1274 // The low bit cannot be shifted out of an exact shift if it is set.
1276 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1277 if (Op0Known.One[0])
1284 /// Given operands for an Shl, see if we can fold the result.
1285 /// If not, this returns null.
1286 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1287 const SimplifyQuery &Q, unsigned MaxRecurse) {
1288 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1292 // undef << X -> undef if (if it's NSW/NUW)
1293 if (match(Op0, m_Undef()))
1294 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1296 // (X >> A) << A -> X
1298 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1301 // shl nuw i8 C, %x -> C iff C has sign bit set.
1302 if (isNUW && match(Op0, m_Negative()))
1304 // NOTE: could use computeKnownBits() / LazyValueInfo,
1305 // but the cost-benefit analysis suggests it isn't worth it.
1310 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1311 const SimplifyQuery &Q) {
1312 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1315 /// Given operands for an LShr, see if we can fold the result.
1316 /// If not, this returns null.
1317 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1318 const SimplifyQuery &Q, unsigned MaxRecurse) {
1319 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1323 // (X << A) >> A -> X
1325 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1328 // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1329 // We can return X as we do in the above case since OR alters no bits in X.
1330 // SimplifyDemandedBits in InstCombine can do more general optimization for
1331 // bit manipulation. This pattern aims to provide opportunities for other
1332 // optimizers by supporting a simple but common case in InstSimplify.
1334 const APInt *ShRAmt, *ShLAmt;
1335 if (match(Op1, m_APInt(ShRAmt)) &&
1336 match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1337 *ShRAmt == *ShLAmt) {
1338 const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1339 const unsigned Width = Op0->getType()->getScalarSizeInBits();
1340 const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1341 if (EffWidthY <= ShRAmt->getZExtValue())
1348 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1349 const SimplifyQuery &Q) {
1350 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1353 /// Given operands for an AShr, see if we can fold the result.
1354 /// If not, this returns null.
1355 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1356 const SimplifyQuery &Q, unsigned MaxRecurse) {
1357 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1361 // all ones >>a X -> -1
1362 // Do not return Op0 because it may contain undef elements if it's a vector.
1363 if (match(Op0, m_AllOnes()))
1364 return Constant::getAllOnesValue(Op0->getType());
1366 // (X << A) >> A -> X
1368 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1371 // Arithmetic shifting an all-sign-bit value is a no-op.
1372 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1373 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1379 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1380 const SimplifyQuery &Q) {
1381 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1384 /// Commuted variants are assumed to be handled by calling this function again
1385 /// with the parameters swapped.
1386 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1387 ICmpInst *UnsignedICmp, bool IsAnd) {
1390 ICmpInst::Predicate EqPred;
1391 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1392 !ICmpInst::isEquality(EqPred))
1395 ICmpInst::Predicate UnsignedPred;
1396 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1397 ICmpInst::isUnsigned(UnsignedPred))
1399 else if (match(UnsignedICmp,
1400 m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1401 ICmpInst::isUnsigned(UnsignedPred))
1402 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1406 // X < Y && Y != 0 --> X < Y
1407 // X < Y || Y != 0 --> Y != 0
1408 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1409 return IsAnd ? UnsignedICmp : ZeroICmp;
1411 // X >= Y || Y != 0 --> true
1412 // X >= Y || Y == 0 --> X >= Y
1413 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1414 if (EqPred == ICmpInst::ICMP_NE)
1415 return getTrue(UnsignedICmp->getType());
1416 return UnsignedICmp;
1419 // X < Y && Y == 0 --> false
1420 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1422 return getFalse(UnsignedICmp->getType());
1427 /// Commuted variants are assumed to be handled by calling this function again
1428 /// with the parameters swapped.
1429 static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1430 ICmpInst::Predicate Pred0, Pred1;
1432 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1433 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1436 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1437 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1438 // can eliminate Op1 from this 'and'.
1439 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1442 // Check for any combination of predicates that are guaranteed to be disjoint.
1443 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1444 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1445 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1446 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1447 return getFalse(Op0->getType());
1452 /// Commuted variants are assumed to be handled by calling this function again
1453 /// with the parameters swapped.
1454 static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1455 ICmpInst::Predicate Pred0, Pred1;
1457 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1458 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1461 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1462 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1463 // can eliminate Op0 from this 'or'.
1464 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1467 // Check for any combination of predicates that cover the entire range of
1469 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1470 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1471 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1472 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1473 return getTrue(Op0->getType());
1478 /// Test if a pair of compares with a shared operand and 2 constants has an
1479 /// empty set intersection, full set union, or if one compare is a superset of
1481 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1483 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1484 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1487 const APInt *C0, *C1;
1488 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1489 !match(Cmp1->getOperand(1), m_APInt(C1)))
1492 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1493 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1495 // For and-of-compares, check if the intersection is empty:
1496 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1497 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1498 return getFalse(Cmp0->getType());
1500 // For or-of-compares, check if the union is full:
1501 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1502 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1503 return getTrue(Cmp0->getType());
1505 // Is one range a superset of the other?
1506 // If this is and-of-compares, take the smaller set:
1507 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1508 // If this is or-of-compares, take the larger set:
1509 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1510 if (Range0.contains(Range1))
1511 return IsAnd ? Cmp1 : Cmp0;
1512 if (Range1.contains(Range0))
1513 return IsAnd ? Cmp0 : Cmp1;
1518 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1520 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1521 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1522 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1525 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1528 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1529 Value *X = Cmp0->getOperand(0);
1530 Value *Y = Cmp1->getOperand(0);
1532 // If one of the compares is a masked version of a (not) null check, then
1533 // that compare implies the other, so we eliminate the other. Optionally, look
1534 // through a pointer-to-int cast to match a null check of a pointer type.
1536 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1537 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1538 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1539 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1540 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1541 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1544 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1545 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1546 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1547 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1548 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1549 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1555 static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1556 // (icmp (add V, C0), C1) & (icmp V, C0)
1557 ICmpInst::Predicate Pred0, Pred1;
1558 const APInt *C0, *C1;
1560 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1563 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1566 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1567 if (AddInst->getOperand(1) != Op1->getOperand(1))
1570 Type *ITy = Op0->getType();
1571 bool isNSW = AddInst->hasNoSignedWrap();
1572 bool isNUW = AddInst->hasNoUnsignedWrap();
1574 const APInt Delta = *C1 - *C0;
1575 if (C0->isStrictlyPositive()) {
1577 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1578 return getFalse(ITy);
1579 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1580 return getFalse(ITy);
1583 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1584 return getFalse(ITy);
1585 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1586 return getFalse(ITy);
1589 if (C0->getBoolValue() && isNUW) {
1591 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1592 return getFalse(ITy);
1594 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1595 return getFalse(ITy);
1601 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1602 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1604 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1607 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1609 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1612 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1615 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1618 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1620 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1626 static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1627 // (icmp (add V, C0), C1) | (icmp V, C0)
1628 ICmpInst::Predicate Pred0, Pred1;
1629 const APInt *C0, *C1;
1631 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1634 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1637 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1638 if (AddInst->getOperand(1) != Op1->getOperand(1))
1641 Type *ITy = Op0->getType();
1642 bool isNSW = AddInst->hasNoSignedWrap();
1643 bool isNUW = AddInst->hasNoUnsignedWrap();
1645 const APInt Delta = *C1 - *C0;
1646 if (C0->isStrictlyPositive()) {
1648 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1649 return getTrue(ITy);
1650 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1651 return getTrue(ITy);
1654 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1655 return getTrue(ITy);
1656 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1657 return getTrue(ITy);
1660 if (C0->getBoolValue() && isNUW) {
1662 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1663 return getTrue(ITy);
1665 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1666 return getTrue(ITy);
1672 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1673 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1675 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1678 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1680 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1683 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1686 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1689 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1691 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1697 static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1698 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1699 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1700 if (LHS0->getType() != RHS0->getType())
1703 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1704 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1705 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1706 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1707 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1708 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1709 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1710 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1711 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1712 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1713 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1714 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1715 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1718 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1719 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1720 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1721 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1722 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1723 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1724 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1725 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1726 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1727 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1734 static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1735 // Look through casts of the 'and' operands to find compares.
1736 auto *Cast0 = dyn_cast<CastInst>(Op0);
1737 auto *Cast1 = dyn_cast<CastInst>(Op1);
1738 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1739 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1740 Op0 = Cast0->getOperand(0);
1741 Op1 = Cast1->getOperand(0);
1745 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1746 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1748 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1749 simplifyOrOfICmps(ICmp0, ICmp1);
1751 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1752 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1754 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1761 // If we looked through casts, we can only handle a constant simplification
1762 // because we are not allowed to create a cast instruction here.
1763 if (auto *C = dyn_cast<Constant>(V))
1764 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1769 /// Given operands for an And, see if we can fold the result.
1770 /// If not, this returns null.
1771 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1772 unsigned MaxRecurse) {
1773 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1777 if (match(Op1, m_Undef()))
1778 return Constant::getNullValue(Op0->getType());
1785 if (match(Op1, m_Zero()))
1786 return Constant::getNullValue(Op0->getType());
1789 if (match(Op1, m_AllOnes()))
1792 // A & ~A = ~A & A = 0
1793 if (match(Op0, m_Not(m_Specific(Op1))) ||
1794 match(Op1, m_Not(m_Specific(Op0))))
1795 return Constant::getNullValue(Op0->getType());
1798 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1802 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1805 // A mask that only clears known zeros of a shifted value is a no-op.
1809 if (match(Op1, m_APInt(Mask))) {
1810 // If all bits in the inverted and shifted mask are clear:
1811 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1812 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1813 (~(*Mask)).lshr(*ShAmt).isNullValue())
1816 // If all bits in the inverted and shifted mask are clear:
1817 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1818 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1819 (~(*Mask)).shl(*ShAmt).isNullValue())
1823 // A & (-A) = A if A is a power of two or zero.
1824 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1825 match(Op1, m_Neg(m_Specific(Op0)))) {
1826 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1829 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1834 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1837 // Try some generic simplifications for associative operations.
1838 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1842 // And distributes over Or. Try some generic simplifications based on this.
1843 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1847 // And distributes over Xor. Try some generic simplifications based on this.
1848 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1852 // If the operation is with the result of a select instruction, check whether
1853 // operating on either branch of the select always yields the same value.
1854 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1855 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1859 // If the operation is with the result of a phi instruction, check whether
1860 // operating on all incoming values of the phi always yields the same value.
1861 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1862 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1869 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1870 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1873 /// Given operands for an Or, see if we can fold the result.
1874 /// If not, this returns null.
1875 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1876 unsigned MaxRecurse) {
1877 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1882 // Do not return Op1 because it may contain undef elements if it's a vector.
1883 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1884 return Constant::getAllOnesValue(Op0->getType());
1888 if (Op0 == Op1 || match(Op1, m_Zero()))
1891 // A | ~A = ~A | A = -1
1892 if (match(Op0, m_Not(m_Specific(Op1))) ||
1893 match(Op1, m_Not(m_Specific(Op0))))
1894 return Constant::getAllOnesValue(Op0->getType());
1897 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1901 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1904 // ~(A & ?) | A = -1
1905 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1906 return Constant::getAllOnesValue(Op1->getType());
1908 // A | ~(A & ?) = -1
1909 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1910 return Constant::getAllOnesValue(Op0->getType());
1913 // (A & ~B) | (A ^ B) -> (A ^ B)
1914 // (~B & A) | (A ^ B) -> (A ^ B)
1915 // (A & ~B) | (B ^ A) -> (B ^ A)
1916 // (~B & A) | (B ^ A) -> (B ^ A)
1917 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1918 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1919 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1922 // Commute the 'or' operands.
1923 // (A ^ B) | (A & ~B) -> (A ^ B)
1924 // (A ^ B) | (~B & A) -> (A ^ B)
1925 // (B ^ A) | (A & ~B) -> (B ^ A)
1926 // (B ^ A) | (~B & A) -> (B ^ A)
1927 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1928 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1929 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1932 // (A & B) | (~A ^ B) -> (~A ^ B)
1933 // (B & A) | (~A ^ B) -> (~A ^ B)
1934 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1935 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1936 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1937 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1938 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1941 // (~A ^ B) | (A & B) -> (~A ^ B)
1942 // (~A ^ B) | (B & A) -> (~A ^ B)
1943 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1944 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1945 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1946 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1947 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1950 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1953 // Try some generic simplifications for associative operations.
1954 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1958 // Or distributes over And. Try some generic simplifications based on this.
1959 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1963 // If the operation is with the result of a select instruction, check whether
1964 // operating on either branch of the select always yields the same value.
1965 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1966 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1970 // (A & C1)|(B & C2)
1971 const APInt *C1, *C2;
1972 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1973 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1975 // (A & C1)|(B & C2)
1976 // If we have: ((V + N) & C1) | (V & C2)
1977 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1978 // replace with V+N.
1980 if (C2->isMask() && // C2 == 0+1+
1981 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1982 // Add commutes, try both ways.
1983 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1986 // Or commutes, try both ways.
1988 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1989 // Add commutes, try both ways.
1990 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1996 // If the operation is with the result of a phi instruction, check whether
1997 // operating on all incoming values of the phi always yields the same value.
1998 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1999 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2005 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2006 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
2009 /// Given operands for a Xor, see if we can fold the result.
2010 /// If not, this returns null.
2011 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2012 unsigned MaxRecurse) {
2013 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2016 // A ^ undef -> undef
2017 if (match(Op1, m_Undef()))
2021 if (match(Op1, m_Zero()))
2026 return Constant::getNullValue(Op0->getType());
2028 // A ^ ~A = ~A ^ A = -1
2029 if (match(Op0, m_Not(m_Specific(Op1))) ||
2030 match(Op1, m_Not(m_Specific(Op0))))
2031 return Constant::getAllOnesValue(Op0->getType());
2033 // Try some generic simplifications for associative operations.
2034 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2038 // Threading Xor over selects and phi nodes is pointless, so don't bother.
2039 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2040 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2041 // only if B and C are equal. If B and C are equal then (since we assume
2042 // that operands have already been simplified) "select(cond, B, C)" should
2043 // have been simplified to the common value of B and C already. Analysing
2044 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2045 // for threading over phi nodes.
2050 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2051 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2055 static Type *GetCompareTy(Value *Op) {
2056 return CmpInst::makeCmpResultType(Op->getType());
2059 /// Rummage around inside V looking for something equivalent to the comparison
2060 /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2061 /// Helper function for analyzing max/min idioms.
2062 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2063 Value *LHS, Value *RHS) {
2064 SelectInst *SI = dyn_cast<SelectInst>(V);
2067 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2070 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2071 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2073 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2074 LHS == CmpRHS && RHS == CmpLHS)
2079 // A significant optimization not implemented here is assuming that alloca
2080 // addresses are not equal to incoming argument values. They don't *alias*,
2081 // as we say, but that doesn't mean they aren't equal, so we take a
2082 // conservative approach.
2084 // This is inspired in part by C++11 5.10p1:
2085 // "Two pointers of the same type compare equal if and only if they are both
2086 // null, both point to the same function, or both represent the same
2089 // This is pretty permissive.
2091 // It's also partly due to C11 6.5.9p6:
2092 // "Two pointers compare equal if and only if both are null pointers, both are
2093 // pointers to the same object (including a pointer to an object and a
2094 // subobject at its beginning) or function, both are pointers to one past the
2095 // last element of the same array object, or one is a pointer to one past the
2096 // end of one array object and the other is a pointer to the start of a
2097 // different array object that happens to immediately follow the first array
2098 // object in the address space.)
2100 // C11's version is more restrictive, however there's no reason why an argument
2101 // couldn't be a one-past-the-end value for a stack object in the caller and be
2102 // equal to the beginning of a stack object in the callee.
2104 // If the C and C++ standards are ever made sufficiently restrictive in this
2105 // area, it may be possible to update LLVM's semantics accordingly and reinstate
2106 // this optimization.
2108 computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2109 const DominatorTree *DT, CmpInst::Predicate Pred,
2110 AssumptionCache *AC, const Instruction *CxtI,
2111 Value *LHS, Value *RHS) {
2112 // First, skip past any trivial no-ops.
2113 LHS = LHS->stripPointerCasts();
2114 RHS = RHS->stripPointerCasts();
2116 // A non-null pointer is not equal to a null pointer.
2117 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2118 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2119 return ConstantInt::get(GetCompareTy(LHS),
2120 !CmpInst::isTrueWhenEqual(Pred));
2122 // We can only fold certain predicates on pointer comparisons.
2127 // Equality comaprisons are easy to fold.
2128 case CmpInst::ICMP_EQ:
2129 case CmpInst::ICMP_NE:
2132 // We can only handle unsigned relational comparisons because 'inbounds' on
2133 // a GEP only protects against unsigned wrapping.
2134 case CmpInst::ICMP_UGT:
2135 case CmpInst::ICMP_UGE:
2136 case CmpInst::ICMP_ULT:
2137 case CmpInst::ICMP_ULE:
2138 // However, we have to switch them to their signed variants to handle
2139 // negative indices from the base pointer.
2140 Pred = ICmpInst::getSignedPredicate(Pred);
2144 // Strip off any constant offsets so that we can reason about them.
2145 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2146 // here and compare base addresses like AliasAnalysis does, however there are
2147 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2148 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2149 // doesn't need to guarantee pointer inequality when it says NoAlias.
2150 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2151 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2153 // If LHS and RHS are related via constant offsets to the same base
2154 // value, we can replace it with an icmp which just compares the offsets.
2156 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2158 // Various optimizations for (in)equality comparisons.
2159 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2160 // Different non-empty allocations that exist at the same time have
2161 // different addresses (if the program can tell). Global variables always
2162 // exist, so they always exist during the lifetime of each other and all
2163 // allocas. Two different allocas usually have different addresses...
2165 // However, if there's an @llvm.stackrestore dynamically in between two
2166 // allocas, they may have the same address. It's tempting to reduce the
2167 // scope of the problem by only looking at *static* allocas here. That would
2168 // cover the majority of allocas while significantly reducing the likelihood
2169 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2170 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2171 // an entry block. Also, if we have a block that's not attached to a
2172 // function, we can't tell if it's "static" under the current definition.
2173 // Theoretically, this problem could be fixed by creating a new kind of
2174 // instruction kind specifically for static allocas. Such a new instruction
2175 // could be required to be at the top of the entry block, thus preventing it
2176 // from being subject to a @llvm.stackrestore. Instcombine could even
2177 // convert regular allocas into these special allocas. It'd be nifty.
2178 // However, until then, this problem remains open.
2180 // So, we'll assume that two non-empty allocas have different addresses
2183 // With all that, if the offsets are within the bounds of their allocations
2184 // (and not one-past-the-end! so we can't use inbounds!), and their
2185 // allocations aren't the same, the pointers are not equal.
2187 // Note that it's not necessary to check for LHS being a global variable
2188 // address, due to canonicalization and constant folding.
2189 if (isa<AllocaInst>(LHS) &&
2190 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2191 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2192 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2193 uint64_t LHSSize, RHSSize;
2194 ObjectSizeOpts Opts;
2195 Opts.NullIsUnknownSize =
2196 NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2197 if (LHSOffsetCI && RHSOffsetCI &&
2198 getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2199 getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2200 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2201 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2202 if (!LHSOffsetValue.isNegative() &&
2203 !RHSOffsetValue.isNegative() &&
2204 LHSOffsetValue.ult(LHSSize) &&
2205 RHSOffsetValue.ult(RHSSize)) {
2206 return ConstantInt::get(GetCompareTy(LHS),
2207 !CmpInst::isTrueWhenEqual(Pred));
2211 // Repeat the above check but this time without depending on DataLayout
2212 // or being able to compute a precise size.
2213 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2214 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2215 LHSOffset->isNullValue() &&
2216 RHSOffset->isNullValue())
2217 return ConstantInt::get(GetCompareTy(LHS),
2218 !CmpInst::isTrueWhenEqual(Pred));
2221 // Even if an non-inbounds GEP occurs along the path we can still optimize
2222 // equality comparisons concerning the result. We avoid walking the whole
2223 // chain again by starting where the last calls to
2224 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2225 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2226 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2228 return ConstantExpr::getICmp(Pred,
2229 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2230 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2232 // If one side of the equality comparison must come from a noalias call
2233 // (meaning a system memory allocation function), and the other side must
2234 // come from a pointer that cannot overlap with dynamically-allocated
2235 // memory within the lifetime of the current function (allocas, byval
2236 // arguments, globals), then determine the comparison result here.
2237 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2238 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2239 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2241 // Is the set of underlying objects all noalias calls?
2242 auto IsNAC = [](ArrayRef<Value *> Objects) {
2243 return all_of(Objects, isNoAliasCall);
2246 // Is the set of underlying objects all things which must be disjoint from
2247 // noalias calls. For allocas, we consider only static ones (dynamic
2248 // allocas might be transformed into calls to malloc not simultaneously
2249 // live with the compared-to allocation). For globals, we exclude symbols
2250 // that might be resolve lazily to symbols in another dynamically-loaded
2251 // library (and, thus, could be malloc'ed by the implementation).
2252 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2253 return all_of(Objects, [](Value *V) {
2254 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2255 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2256 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2257 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2258 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2259 !GV->isThreadLocal();
2260 if (const Argument *A = dyn_cast<Argument>(V))
2261 return A->hasByValAttr();
2266 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2267 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2268 return ConstantInt::get(GetCompareTy(LHS),
2269 !CmpInst::isTrueWhenEqual(Pred));
2271 // Fold comparisons for non-escaping pointer even if the allocation call
2272 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2273 // dynamic allocation call could be either of the operands.
2274 Value *MI = nullptr;
2275 if (isAllocLikeFn(LHS, TLI) &&
2276 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2278 else if (isAllocLikeFn(RHS, TLI) &&
2279 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2281 // FIXME: We should also fold the compare when the pointer escapes, but the
2282 // compare dominates the pointer escape
2283 if (MI && !PointerMayBeCaptured(MI, true, true))
2284 return ConstantInt::get(GetCompareTy(LHS),
2285 CmpInst::isFalseWhenEqual(Pred));
2292 /// Fold an icmp when its operands have i1 scalar type.
2293 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2294 Value *RHS, const SimplifyQuery &Q) {
2295 Type *ITy = GetCompareTy(LHS); // The return type.
2296 Type *OpTy = LHS->getType(); // The operand type.
2297 if (!OpTy->isIntOrIntVectorTy(1))
2300 // A boolean compared to true/false can be simplified in 14 out of the 20
2301 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2302 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2303 if (match(RHS, m_Zero())) {
2305 case CmpInst::ICMP_NE: // X != 0 -> X
2306 case CmpInst::ICMP_UGT: // X >u 0 -> X
2307 case CmpInst::ICMP_SLT: // X <s 0 -> X
2310 case CmpInst::ICMP_ULT: // X <u 0 -> false
2311 case CmpInst::ICMP_SGT: // X >s 0 -> false
2312 return getFalse(ITy);
2314 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2315 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2316 return getTrue(ITy);
2320 } else if (match(RHS, m_One())) {
2322 case CmpInst::ICMP_EQ: // X == 1 -> X
2323 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2324 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2327 case CmpInst::ICMP_UGT: // X >u 1 -> false
2328 case CmpInst::ICMP_SLT: // X <s -1 -> false
2329 return getFalse(ITy);
2331 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2332 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2333 return getTrue(ITy);
2342 case ICmpInst::ICMP_UGE:
2343 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2344 return getTrue(ITy);
2346 case ICmpInst::ICMP_SGE:
2347 /// For signed comparison, the values for an i1 are 0 and -1
2348 /// respectively. This maps into a truth table of:
2349 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2350 /// 0 | 0 | 1 (0 >= 0) | 1
2351 /// 0 | 1 | 1 (0 >= -1) | 1
2352 /// 1 | 0 | 0 (-1 >= 0) | 0
2353 /// 1 | 1 | 1 (-1 >= -1) | 1
2354 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2355 return getTrue(ITy);
2357 case ICmpInst::ICMP_ULE:
2358 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2359 return getTrue(ITy);
2366 /// Try hard to fold icmp with zero RHS because this is a common case.
2367 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2368 Value *RHS, const SimplifyQuery &Q) {
2369 if (!match(RHS, m_Zero()))
2372 Type *ITy = GetCompareTy(LHS); // The return type.
2375 llvm_unreachable("Unknown ICmp predicate!");
2376 case ICmpInst::ICMP_ULT:
2377 return getFalse(ITy);
2378 case ICmpInst::ICMP_UGE:
2379 return getTrue(ITy);
2380 case ICmpInst::ICMP_EQ:
2381 case ICmpInst::ICMP_ULE:
2382 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2383 return getFalse(ITy);
2385 case ICmpInst::ICMP_NE:
2386 case ICmpInst::ICMP_UGT:
2387 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2388 return getTrue(ITy);
2390 case ICmpInst::ICMP_SLT: {
2391 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2392 if (LHSKnown.isNegative())
2393 return getTrue(ITy);
2394 if (LHSKnown.isNonNegative())
2395 return getFalse(ITy);
2398 case ICmpInst::ICMP_SLE: {
2399 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2400 if (LHSKnown.isNegative())
2401 return getTrue(ITy);
2402 if (LHSKnown.isNonNegative() &&
2403 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2404 return getFalse(ITy);
2407 case ICmpInst::ICMP_SGE: {
2408 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2409 if (LHSKnown.isNegative())
2410 return getFalse(ITy);
2411 if (LHSKnown.isNonNegative())
2412 return getTrue(ITy);
2415 case ICmpInst::ICMP_SGT: {
2416 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2417 if (LHSKnown.isNegative())
2418 return getFalse(ITy);
2419 if (LHSKnown.isNonNegative() &&
2420 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2421 return getTrue(ITy);
2429 /// Many binary operators with a constant operand have an easy-to-compute
2430 /// range of outputs. This can be used to fold a comparison to always true or
2432 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2433 unsigned Width = Lower.getBitWidth();
2435 switch (BO.getOpcode()) {
2436 case Instruction::Add:
2437 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2438 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2439 if (BO.hasNoUnsignedWrap()) {
2440 // 'add nuw x, C' produces [C, UINT_MAX].
2442 } else if (BO.hasNoSignedWrap()) {
2443 if (C->isNegative()) {
2444 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2445 Lower = APInt::getSignedMinValue(Width);
2446 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2448 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2449 Lower = APInt::getSignedMinValue(Width) + *C;
2450 Upper = APInt::getSignedMaxValue(Width) + 1;
2456 case Instruction::And:
2457 if (match(BO.getOperand(1), m_APInt(C)))
2458 // 'and x, C' produces [0, C].
2462 case Instruction::Or:
2463 if (match(BO.getOperand(1), m_APInt(C)))
2464 // 'or x, C' produces [C, UINT_MAX].
2468 case Instruction::AShr:
2469 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2470 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2471 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2472 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2473 } else if (match(BO.getOperand(0), m_APInt(C))) {
2474 unsigned ShiftAmount = Width - 1;
2475 if (!C->isNullValue() && BO.isExact())
2476 ShiftAmount = C->countTrailingZeros();
2477 if (C->isNegative()) {
2478 // 'ashr C, x' produces [C, C >> (Width-1)]
2480 Upper = C->ashr(ShiftAmount) + 1;
2482 // 'ashr C, x' produces [C >> (Width-1), C]
2483 Lower = C->ashr(ShiftAmount);
2489 case Instruction::LShr:
2490 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2491 // 'lshr x, C' produces [0, UINT_MAX >> C].
2492 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2493 } else if (match(BO.getOperand(0), m_APInt(C))) {
2494 // 'lshr C, x' produces [C >> (Width-1), C].
2495 unsigned ShiftAmount = Width - 1;
2496 if (!C->isNullValue() && BO.isExact())
2497 ShiftAmount = C->countTrailingZeros();
2498 Lower = C->lshr(ShiftAmount);
2503 case Instruction::Shl:
2504 if (match(BO.getOperand(0), m_APInt(C))) {
2505 if (BO.hasNoUnsignedWrap()) {
2506 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2508 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2509 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2510 if (C->isNegative()) {
2511 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2512 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2513 Lower = C->shl(ShiftAmount);
2516 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2517 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2519 Upper = C->shl(ShiftAmount) + 1;
2525 case Instruction::SDiv:
2526 if (match(BO.getOperand(1), m_APInt(C))) {
2527 APInt IntMin = APInt::getSignedMinValue(Width);
2528 APInt IntMax = APInt::getSignedMaxValue(Width);
2529 if (C->isAllOnesValue()) {
2530 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2531 // where C != -1 and C != 0 and C != 1
2534 } else if (C->countLeadingZeros() < Width - 1) {
2535 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2536 // where C != -1 and C != 0 and C != 1
2537 Lower = IntMin.sdiv(*C);
2538 Upper = IntMax.sdiv(*C);
2539 if (Lower.sgt(Upper))
2540 std::swap(Lower, Upper);
2542 assert(Upper != Lower && "Upper part of range has wrapped!");
2544 } else if (match(BO.getOperand(0), m_APInt(C))) {
2545 if (C->isMinSignedValue()) {
2546 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2548 Upper = Lower.lshr(1) + 1;
2550 // 'sdiv C, x' produces [-|C|, |C|].
2551 Upper = C->abs() + 1;
2552 Lower = (-Upper) + 1;
2557 case Instruction::UDiv:
2558 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2559 // 'udiv x, C' produces [0, UINT_MAX / C].
2560 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2561 } else if (match(BO.getOperand(0), m_APInt(C))) {
2562 // 'udiv C, x' produces [0, C].
2567 case Instruction::SRem:
2568 if (match(BO.getOperand(1), m_APInt(C))) {
2569 // 'srem x, C' produces (-|C|, |C|).
2571 Lower = (-Upper) + 1;
2575 case Instruction::URem:
2576 if (match(BO.getOperand(1), m_APInt(C)))
2577 // 'urem x, C' produces [0, C).
2586 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2588 Type *ITy = GetCompareTy(RHS); // The return type.
2591 // Sign-bit checks can be optimized to true/false after unsigned
2592 // floating-point casts:
2593 // icmp slt (bitcast (uitofp X)), 0 --> false
2594 // icmp sgt (bitcast (uitofp X)), -1 --> true
2595 if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2596 if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2597 return ConstantInt::getFalse(ITy);
2598 if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2599 return ConstantInt::getTrue(ITy);
2603 if (!match(RHS, m_APInt(C)))
2606 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2607 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2608 if (RHS_CR.isEmptySet())
2609 return ConstantInt::getFalse(ITy);
2610 if (RHS_CR.isFullSet())
2611 return ConstantInt::getTrue(ITy);
2613 // Find the range of possible values for binary operators.
2614 unsigned Width = C->getBitWidth();
2615 APInt Lower = APInt(Width, 0);
2616 APInt Upper = APInt(Width, 0);
2617 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2618 setLimitsForBinOp(*BO, Lower, Upper);
2620 ConstantRange LHS_CR =
2621 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2623 if (auto *I = dyn_cast<Instruction>(LHS))
2624 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2625 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2627 if (!LHS_CR.isFullSet()) {
2628 if (RHS_CR.contains(LHS_CR))
2629 return ConstantInt::getTrue(ITy);
2630 if (RHS_CR.inverse().contains(LHS_CR))
2631 return ConstantInt::getFalse(ITy);
2637 /// TODO: A large part of this logic is duplicated in InstCombine's
2638 /// foldICmpBinOp(). We should be able to share that and avoid the code
2640 static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2641 Value *RHS, const SimplifyQuery &Q,
2642 unsigned MaxRecurse) {
2643 Type *ITy = GetCompareTy(LHS); // The return type.
2645 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2646 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2647 if (MaxRecurse && (LBO || RBO)) {
2648 // Analyze the case when either LHS or RHS is an add instruction.
2649 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2650 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2651 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2652 if (LBO && LBO->getOpcode() == Instruction::Add) {
2653 A = LBO->getOperand(0);
2654 B = LBO->getOperand(1);
2656 ICmpInst::isEquality(Pred) ||
2657 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2658 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2660 if (RBO && RBO->getOpcode() == Instruction::Add) {
2661 C = RBO->getOperand(0);
2662 D = RBO->getOperand(1);
2664 ICmpInst::isEquality(Pred) ||
2665 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2666 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2669 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2670 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2671 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2672 Constant::getNullValue(RHS->getType()), Q,
2676 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2677 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2679 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2680 C == LHS ? D : C, Q, MaxRecurse - 1))
2683 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2684 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2686 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2689 // C + B == C + D -> B == D
2692 } else if (A == D) {
2693 // D + B == C + D -> B == C
2696 } else if (B == C) {
2697 // A + C == C + D -> A == D
2702 // A + D == C + D -> A == C
2706 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2713 // icmp pred (or X, Y), X
2714 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2715 if (Pred == ICmpInst::ICMP_ULT)
2716 return getFalse(ITy);
2717 if (Pred == ICmpInst::ICMP_UGE)
2718 return getTrue(ITy);
2720 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2721 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2722 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2723 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2724 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2725 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2726 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2729 // icmp pred X, (or X, Y)
2730 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2731 if (Pred == ICmpInst::ICMP_ULE)
2732 return getTrue(ITy);
2733 if (Pred == ICmpInst::ICMP_UGT)
2734 return getFalse(ITy);
2736 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2737 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2738 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2739 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2740 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2741 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2742 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2747 // icmp pred (and X, Y), X
2748 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2749 if (Pred == ICmpInst::ICMP_UGT)
2750 return getFalse(ITy);
2751 if (Pred == ICmpInst::ICMP_ULE)
2752 return getTrue(ITy);
2754 // icmp pred X, (and X, Y)
2755 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2756 if (Pred == ICmpInst::ICMP_UGE)
2757 return getTrue(ITy);
2758 if (Pred == ICmpInst::ICMP_ULT)
2759 return getFalse(ITy);
2762 // 0 - (zext X) pred C
2763 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2764 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2765 if (RHSC->getValue().isStrictlyPositive()) {
2766 if (Pred == ICmpInst::ICMP_SLT)
2767 return ConstantInt::getTrue(RHSC->getContext());
2768 if (Pred == ICmpInst::ICMP_SGE)
2769 return ConstantInt::getFalse(RHSC->getContext());
2770 if (Pred == ICmpInst::ICMP_EQ)
2771 return ConstantInt::getFalse(RHSC->getContext());
2772 if (Pred == ICmpInst::ICMP_NE)
2773 return ConstantInt::getTrue(RHSC->getContext());
2775 if (RHSC->getValue().isNonNegative()) {
2776 if (Pred == ICmpInst::ICMP_SLE)
2777 return ConstantInt::getTrue(RHSC->getContext());
2778 if (Pred == ICmpInst::ICMP_SGT)
2779 return ConstantInt::getFalse(RHSC->getContext());
2784 // icmp pred (urem X, Y), Y
2785 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2789 case ICmpInst::ICMP_SGT:
2790 case ICmpInst::ICMP_SGE: {
2791 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2792 if (!Known.isNonNegative())
2796 case ICmpInst::ICMP_EQ:
2797 case ICmpInst::ICMP_UGT:
2798 case ICmpInst::ICMP_UGE:
2799 return getFalse(ITy);
2800 case ICmpInst::ICMP_SLT:
2801 case ICmpInst::ICMP_SLE: {
2802 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2803 if (!Known.isNonNegative())
2807 case ICmpInst::ICMP_NE:
2808 case ICmpInst::ICMP_ULT:
2809 case ICmpInst::ICMP_ULE:
2810 return getTrue(ITy);
2814 // icmp pred X, (urem Y, X)
2815 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2819 case ICmpInst::ICMP_SGT:
2820 case ICmpInst::ICMP_SGE: {
2821 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2822 if (!Known.isNonNegative())
2826 case ICmpInst::ICMP_NE:
2827 case ICmpInst::ICMP_UGT:
2828 case ICmpInst::ICMP_UGE:
2829 return getTrue(ITy);
2830 case ICmpInst::ICMP_SLT:
2831 case ICmpInst::ICMP_SLE: {
2832 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2833 if (!Known.isNonNegative())
2837 case ICmpInst::ICMP_EQ:
2838 case ICmpInst::ICMP_ULT:
2839 case ICmpInst::ICMP_ULE:
2840 return getFalse(ITy);
2846 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2847 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2848 // icmp pred (X op Y), X
2849 if (Pred == ICmpInst::ICMP_UGT)
2850 return getFalse(ITy);
2851 if (Pred == ICmpInst::ICMP_ULE)
2852 return getTrue(ITy);
2857 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2858 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2859 // icmp pred X, (X op Y)
2860 if (Pred == ICmpInst::ICMP_ULT)
2861 return getFalse(ITy);
2862 if (Pred == ICmpInst::ICMP_UGE)
2863 return getTrue(ITy);
2870 // where CI2 is a power of 2 and CI isn't
2871 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2872 const APInt *CI2Val, *CIVal = &CI->getValue();
2873 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2874 CI2Val->isPowerOf2()) {
2875 if (!CIVal->isPowerOf2()) {
2876 // CI2 << X can equal zero in some circumstances,
2877 // this simplification is unsafe if CI is zero.
2879 // We know it is safe if:
2880 // - The shift is nsw, we can't shift out the one bit.
2881 // - The shift is nuw, we can't shift out the one bit.
2884 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2885 CI2Val->isOneValue() || !CI->isZero()) {
2886 if (Pred == ICmpInst::ICMP_EQ)
2887 return ConstantInt::getFalse(RHS->getContext());
2888 if (Pred == ICmpInst::ICMP_NE)
2889 return ConstantInt::getTrue(RHS->getContext());
2892 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2893 if (Pred == ICmpInst::ICMP_UGT)
2894 return ConstantInt::getFalse(RHS->getContext());
2895 if (Pred == ICmpInst::ICMP_ULE)
2896 return ConstantInt::getTrue(RHS->getContext());
2901 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2902 LBO->getOperand(1) == RBO->getOperand(1)) {
2903 switch (LBO->getOpcode()) {
2906 case Instruction::UDiv:
2907 case Instruction::LShr:
2908 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2910 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2911 RBO->getOperand(0), Q, MaxRecurse - 1))
2914 case Instruction::SDiv:
2915 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2917 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2918 RBO->getOperand(0), Q, MaxRecurse - 1))
2921 case Instruction::AShr:
2922 if (!LBO->isExact() || !RBO->isExact())
2924 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2925 RBO->getOperand(0), Q, MaxRecurse - 1))
2928 case Instruction::Shl: {
2929 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2930 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2933 if (!NSW && ICmpInst::isSigned(Pred))
2935 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2936 RBO->getOperand(0), Q, MaxRecurse - 1))
2945 /// Simplify integer comparisons where at least one operand of the compare
2946 /// matches an integer min/max idiom.
2947 static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2948 Value *RHS, const SimplifyQuery &Q,
2949 unsigned MaxRecurse) {
2950 Type *ITy = GetCompareTy(LHS); // The return type.
2952 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2953 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2955 // Signed variants on "max(a,b)>=a -> true".
2956 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2958 std::swap(A, B); // smax(A, B) pred A.
2959 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2960 // We analyze this as smax(A, B) pred A.
2962 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2963 (A == LHS || B == LHS)) {
2965 std::swap(A, B); // A pred smax(A, B).
2966 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2967 // We analyze this as smax(A, B) swapped-pred A.
2968 P = CmpInst::getSwappedPredicate(Pred);
2969 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2970 (A == RHS || B == RHS)) {
2972 std::swap(A, B); // smin(A, B) pred A.
2973 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2974 // We analyze this as smax(-A, -B) swapped-pred -A.
2975 // Note that we do not need to actually form -A or -B thanks to EqP.
2976 P = CmpInst::getSwappedPredicate(Pred);
2977 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2978 (A == LHS || B == LHS)) {
2980 std::swap(A, B); // A pred smin(A, B).
2981 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2982 // We analyze this as smax(-A, -B) pred -A.
2983 // Note that we do not need to actually form -A or -B thanks to EqP.
2986 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2987 // Cases correspond to "max(A, B) p A".
2991 case CmpInst::ICMP_EQ:
2992 case CmpInst::ICMP_SLE:
2993 // Equivalent to "A EqP B". This may be the same as the condition tested
2994 // in the max/min; if so, we can just return that.
2995 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2997 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2999 // Otherwise, see if "A EqP B" simplifies.
3001 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3004 case CmpInst::ICMP_NE:
3005 case CmpInst::ICMP_SGT: {
3006 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3007 // Equivalent to "A InvEqP B". This may be the same as the condition
3008 // tested in the max/min; if so, we can just return that.
3009 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3011 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3013 // Otherwise, see if "A InvEqP B" simplifies.
3015 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3019 case CmpInst::ICMP_SGE:
3021 return getTrue(ITy);
3022 case CmpInst::ICMP_SLT:
3024 return getFalse(ITy);
3028 // Unsigned variants on "max(a,b)>=a -> true".
3029 P = CmpInst::BAD_ICMP_PREDICATE;
3030 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3032 std::swap(A, B); // umax(A, B) pred A.
3033 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3034 // We analyze this as umax(A, B) pred A.
3036 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
3037 (A == LHS || B == LHS)) {
3039 std::swap(A, B); // A pred umax(A, B).
3040 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3041 // We analyze this as umax(A, B) swapped-pred A.
3042 P = CmpInst::getSwappedPredicate(Pred);
3043 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3044 (A == RHS || B == RHS)) {
3046 std::swap(A, B); // umin(A, B) pred A.
3047 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3048 // We analyze this as umax(-A, -B) swapped-pred -A.
3049 // Note that we do not need to actually form -A or -B thanks to EqP.
3050 P = CmpInst::getSwappedPredicate(Pred);
3051 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3052 (A == LHS || B == LHS)) {
3054 std::swap(A, B); // A pred umin(A, B).
3055 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3056 // We analyze this as umax(-A, -B) pred -A.
3057 // Note that we do not need to actually form -A or -B thanks to EqP.
3060 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3061 // Cases correspond to "max(A, B) p A".
3065 case CmpInst::ICMP_EQ:
3066 case CmpInst::ICMP_ULE:
3067 // Equivalent to "A EqP B". This may be the same as the condition tested
3068 // in the max/min; if so, we can just return that.
3069 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3071 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3073 // Otherwise, see if "A EqP B" simplifies.
3075 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3078 case CmpInst::ICMP_NE:
3079 case CmpInst::ICMP_UGT: {
3080 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3081 // Equivalent to "A InvEqP B". This may be the same as the condition
3082 // tested in the max/min; if so, we can just return that.
3083 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3085 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3087 // Otherwise, see if "A InvEqP B" simplifies.
3089 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3093 case CmpInst::ICMP_UGE:
3095 return getTrue(ITy);
3096 case CmpInst::ICMP_ULT:
3098 return getFalse(ITy);
3102 // Variants on "max(x,y) >= min(x,z)".
3104 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3105 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3106 (A == C || A == D || B == C || B == D)) {
3107 // max(x, ?) pred min(x, ?).
3108 if (Pred == CmpInst::ICMP_SGE)
3110 return getTrue(ITy);
3111 if (Pred == CmpInst::ICMP_SLT)
3113 return getFalse(ITy);
3114 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3115 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3116 (A == C || A == D || B == C || B == D)) {
3117 // min(x, ?) pred max(x, ?).
3118 if (Pred == CmpInst::ICMP_SLE)
3120 return getTrue(ITy);
3121 if (Pred == CmpInst::ICMP_SGT)
3123 return getFalse(ITy);
3124 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3125 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3126 (A == C || A == D || B == C || B == D)) {
3127 // max(x, ?) pred min(x, ?).
3128 if (Pred == CmpInst::ICMP_UGE)
3130 return getTrue(ITy);
3131 if (Pred == CmpInst::ICMP_ULT)
3133 return getFalse(ITy);
3134 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3135 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3136 (A == C || A == D || B == C || B == D)) {
3137 // min(x, ?) pred max(x, ?).
3138 if (Pred == CmpInst::ICMP_ULE)
3140 return getTrue(ITy);
3141 if (Pred == CmpInst::ICMP_UGT)
3143 return getFalse(ITy);
3149 /// Given operands for an ICmpInst, see if we can fold the result.
3150 /// If not, this returns null.
3151 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3152 const SimplifyQuery &Q, unsigned MaxRecurse) {
3153 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3154 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3156 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3157 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3158 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3160 // If we have a constant, make sure it is on the RHS.
3161 std::swap(LHS, RHS);
3162 Pred = CmpInst::getSwappedPredicate(Pred);
3165 Type *ITy = GetCompareTy(LHS); // The return type.
3167 // icmp X, X -> true/false
3168 // icmp X, undef -> true/false because undef could be X.
3169 if (LHS == RHS || isa<UndefValue>(RHS))
3170 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3172 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3175 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3178 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3181 // If both operands have range metadata, use the metadata
3182 // to simplify the comparison.
3183 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3184 auto RHS_Instr = cast<Instruction>(RHS);
3185 auto LHS_Instr = cast<Instruction>(LHS);
3187 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3188 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3189 auto RHS_CR = getConstantRangeFromMetadata(
3190 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3191 auto LHS_CR = getConstantRangeFromMetadata(
3192 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3194 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3195 if (Satisfied_CR.contains(LHS_CR))
3196 return ConstantInt::getTrue(RHS->getContext());
3198 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3199 CmpInst::getInversePredicate(Pred), RHS_CR);
3200 if (InversedSatisfied_CR.contains(LHS_CR))
3201 return ConstantInt::getFalse(RHS->getContext());
3205 // Compare of cast, for example (zext X) != 0 -> X != 0
3206 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3207 Instruction *LI = cast<CastInst>(LHS);
3208 Value *SrcOp = LI->getOperand(0);
3209 Type *SrcTy = SrcOp->getType();
3210 Type *DstTy = LI->getType();
3212 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3213 // if the integer type is the same size as the pointer type.
3214 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3215 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3216 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3217 // Transfer the cast to the constant.
3218 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3219 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3222 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3223 if (RI->getOperand(0)->getType() == SrcTy)
3224 // Compare without the cast.
3225 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3231 if (isa<ZExtInst>(LHS)) {
3232 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3234 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3235 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3236 // Compare X and Y. Note that signed predicates become unsigned.
3237 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3238 SrcOp, RI->getOperand(0), Q,
3242 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3243 // too. If not, then try to deduce the result of the comparison.
3244 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3245 // Compute the constant that would happen if we truncated to SrcTy then
3246 // reextended to DstTy.
3247 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3248 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3250 // If the re-extended constant didn't change then this is effectively
3251 // also a case of comparing two zero-extended values.
3252 if (RExt == CI && MaxRecurse)
3253 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3254 SrcOp, Trunc, Q, MaxRecurse-1))
3257 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3258 // there. Use this to work out the result of the comparison.
3261 default: llvm_unreachable("Unknown ICmp predicate!");
3263 case ICmpInst::ICMP_EQ:
3264 case ICmpInst::ICMP_UGT:
3265 case ICmpInst::ICMP_UGE:
3266 return ConstantInt::getFalse(CI->getContext());
3268 case ICmpInst::ICMP_NE:
3269 case ICmpInst::ICMP_ULT:
3270 case ICmpInst::ICMP_ULE:
3271 return ConstantInt::getTrue(CI->getContext());
3273 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3274 // is non-negative then LHS <s RHS.
3275 case ICmpInst::ICMP_SGT:
3276 case ICmpInst::ICMP_SGE:
3277 return CI->getValue().isNegative() ?
3278 ConstantInt::getTrue(CI->getContext()) :
3279 ConstantInt::getFalse(CI->getContext());
3281 case ICmpInst::ICMP_SLT:
3282 case ICmpInst::ICMP_SLE:
3283 return CI->getValue().isNegative() ?
3284 ConstantInt::getFalse(CI->getContext()) :
3285 ConstantInt::getTrue(CI->getContext());
3291 if (isa<SExtInst>(LHS)) {
3292 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3294 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3295 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3296 // Compare X and Y. Note that the predicate does not change.
3297 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3301 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3302 // too. If not, then try to deduce the result of the comparison.
3303 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3304 // Compute the constant that would happen if we truncated to SrcTy then
3305 // reextended to DstTy.
3306 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3307 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3309 // If the re-extended constant didn't change then this is effectively
3310 // also a case of comparing two sign-extended values.
3311 if (RExt == CI && MaxRecurse)
3312 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3315 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3316 // bits there. Use this to work out the result of the comparison.
3319 default: llvm_unreachable("Unknown ICmp predicate!");
3320 case ICmpInst::ICMP_EQ:
3321 return ConstantInt::getFalse(CI->getContext());
3322 case ICmpInst::ICMP_NE:
3323 return ConstantInt::getTrue(CI->getContext());
3325 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3327 case ICmpInst::ICMP_SGT:
3328 case ICmpInst::ICMP_SGE:
3329 return CI->getValue().isNegative() ?
3330 ConstantInt::getTrue(CI->getContext()) :
3331 ConstantInt::getFalse(CI->getContext());
3332 case ICmpInst::ICMP_SLT:
3333 case ICmpInst::ICMP_SLE:
3334 return CI->getValue().isNegative() ?
3335 ConstantInt::getFalse(CI->getContext()) :
3336 ConstantInt::getTrue(CI->getContext());
3338 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3340 case ICmpInst::ICMP_UGT:
3341 case ICmpInst::ICMP_UGE:
3342 // Comparison is true iff the LHS <s 0.
3344 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3345 Constant::getNullValue(SrcTy),
3349 case ICmpInst::ICMP_ULT:
3350 case ICmpInst::ICMP_ULE:
3351 // Comparison is true iff the LHS >=s 0.
3353 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3354 Constant::getNullValue(SrcTy),
3364 // icmp eq|ne X, Y -> false|true if X != Y
3365 if (ICmpInst::isEquality(Pred) &&
3366 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3367 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3370 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3373 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3376 // Simplify comparisons of related pointers using a powerful, recursive
3377 // GEP-walk when we have target data available..
3378 if (LHS->getType()->isPointerTy())
3379 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3382 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3383 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3384 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3385 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3386 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3387 Q.DL.getTypeSizeInBits(CRHS->getType()))
3388 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3389 CLHS->getPointerOperand(),
3390 CRHS->getPointerOperand()))
3393 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3394 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3395 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3396 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3397 (ICmpInst::isEquality(Pred) ||
3398 (GLHS->isInBounds() && GRHS->isInBounds() &&
3399 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3400 // The bases are equal and the indices are constant. Build a constant
3401 // expression GEP with the same indices and a null base pointer to see
3402 // what constant folding can make out of it.
3403 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3404 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3405 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3406 GLHS->getSourceElementType(), Null, IndicesLHS);
3408 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3409 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3410 GLHS->getSourceElementType(), Null, IndicesRHS);
3411 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3416 // If the comparison is with the result of a select instruction, check whether
3417 // comparing with either branch of the select always yields the same value.
3418 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3419 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3422 // If the comparison is with the result of a phi instruction, check whether
3423 // doing the compare with each incoming phi value yields a common result.
3424 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3425 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3431 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3432 const SimplifyQuery &Q) {
3433 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3436 /// Given operands for an FCmpInst, see if we can fold the result.
3437 /// If not, this returns null.
3438 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3439 FastMathFlags FMF, const SimplifyQuery &Q,
3440 unsigned MaxRecurse) {
3441 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3442 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3444 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3445 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3446 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3448 // If we have a constant, make sure it is on the RHS.
3449 std::swap(LHS, RHS);
3450 Pred = CmpInst::getSwappedPredicate(Pred);
3453 // Fold trivial predicates.
3454 Type *RetTy = GetCompareTy(LHS);
3455 if (Pred == FCmpInst::FCMP_FALSE)
3456 return getFalse(RetTy);
3457 if (Pred == FCmpInst::FCMP_TRUE)
3458 return getTrue(RetTy);
3460 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3462 if (Pred == FCmpInst::FCMP_UNO)
3463 return getFalse(RetTy);
3464 if (Pred == FCmpInst::FCMP_ORD)
3465 return getTrue(RetTy);
3468 // NaN is unordered; NaN is not ordered.
3469 assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&
3470 "Comparison must be either ordered or unordered");
3471 if (match(RHS, m_NaN()))
3472 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3474 // fcmp pred x, undef and fcmp pred undef, x
3475 // fold to true if unordered, false if ordered
3476 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3477 // Choosing NaN for the undef will always make unordered comparison succeed
3478 // and ordered comparison fail.
3479 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3482 // fcmp x,x -> true/false. Not all compares are foldable.
3484 if (CmpInst::isTrueWhenEqual(Pred))
3485 return getTrue(RetTy);
3486 if (CmpInst::isFalseWhenEqual(Pred))
3487 return getFalse(RetTy);
3490 // Handle fcmp with constant RHS.
3492 if (match(RHS, m_APFloat(C))) {
3493 // Check whether the constant is an infinity.
3494 if (C->isInfinity()) {
3495 if (C->isNegative()) {
3497 case FCmpInst::FCMP_OLT:
3498 // No value is ordered and less than negative infinity.
3499 return getFalse(RetTy);
3500 case FCmpInst::FCMP_UGE:
3501 // All values are unordered with or at least negative infinity.
3502 return getTrue(RetTy);
3508 case FCmpInst::FCMP_OGT:
3509 // No value is ordered and greater than infinity.
3510 return getFalse(RetTy);
3511 case FCmpInst::FCMP_ULE:
3512 // All values are unordered with and at most infinity.
3513 return getTrue(RetTy);
3521 case FCmpInst::FCMP_UGE:
3522 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3523 return getTrue(RetTy);
3525 case FCmpInst::FCMP_OLT:
3527 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3528 return getFalse(RetTy);
3533 } else if (C->isNegative()) {
3534 assert(!C->isNaN() && "Unexpected NaN constant!");
3535 // TODO: We can catch more cases by using a range check rather than
3536 // relying on CannotBeOrderedLessThanZero.
3538 case FCmpInst::FCMP_UGE:
3539 case FCmpInst::FCMP_UGT:
3540 case FCmpInst::FCMP_UNE:
3541 // (X >= 0) implies (X > C) when (C < 0)
3542 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3543 return getTrue(RetTy);
3545 case FCmpInst::FCMP_OEQ:
3546 case FCmpInst::FCMP_OLE:
3547 case FCmpInst::FCMP_OLT:
3548 // (X >= 0) implies !(X < C) when (C < 0)
3549 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3550 return getFalse(RetTy);
3558 // If the comparison is with the result of a select instruction, check whether
3559 // comparing with either branch of the select always yields the same value.
3560 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3561 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3564 // If the comparison is with the result of a phi instruction, check whether
3565 // doing the compare with each incoming phi value yields a common result.
3566 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3567 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3573 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3574 FastMathFlags FMF, const SimplifyQuery &Q) {
3575 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3578 /// See if V simplifies when its operand Op is replaced with RepOp.
3579 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3580 const SimplifyQuery &Q,
3581 unsigned MaxRecurse) {
3582 // Trivial replacement.
3586 // We cannot replace a constant, and shouldn't even try.
3587 if (isa<Constant>(Op))
3590 auto *I = dyn_cast<Instruction>(V);
3594 // If this is a binary operator, try to simplify it with the replaced op.
3595 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3597 // %cmp = icmp eq i32 %x, 2147483647
3598 // %add = add nsw i32 %x, 1
3599 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3601 // We can't replace %sel with %add unless we strip away the flags.
3602 if (isa<OverflowingBinaryOperator>(B))
3603 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3605 if (isa<PossiblyExactOperator>(B))
3610 if (B->getOperand(0) == Op)
3611 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3613 if (B->getOperand(1) == Op)
3614 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3619 // Same for CmpInsts.
3620 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3622 if (C->getOperand(0) == Op)
3623 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3625 if (C->getOperand(1) == Op)
3626 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3632 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3634 SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3635 transform(GEP->operands(), NewOps.begin(),
3636 [&](Value *V) { return V == Op ? RepOp : V; });
3637 return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3642 // TODO: We could hand off more cases to instsimplify here.
3644 // If all operands are constant after substituting Op for RepOp then we can
3645 // constant fold the instruction.
3646 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3647 // Build a list of all constant operands.
3648 SmallVector<Constant *, 8> ConstOps;
3649 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3650 if (I->getOperand(i) == Op)
3651 ConstOps.push_back(CRepOp);
3652 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3653 ConstOps.push_back(COp);
3658 // All operands were constants, fold it.
3659 if (ConstOps.size() == I->getNumOperands()) {
3660 if (CmpInst *C = dyn_cast<CmpInst>(I))
3661 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3662 ConstOps[1], Q.DL, Q.TLI);
3664 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3665 if (!LI->isVolatile())
3666 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3668 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3675 /// Try to simplify a select instruction when its condition operand is an
3676 /// integer comparison where one operand of the compare is a constant.
3677 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3678 const APInt *Y, bool TrueWhenUnset) {
3681 // (X & Y) == 0 ? X & ~Y : X --> X
3682 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3683 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3685 return TrueWhenUnset ? FalseVal : TrueVal;
3687 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3688 // (X & Y) != 0 ? X : X & ~Y --> X
3689 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3691 return TrueWhenUnset ? FalseVal : TrueVal;
3693 if (Y->isPowerOf2()) {
3694 // (X & Y) == 0 ? X | Y : X --> X | Y
3695 // (X & Y) != 0 ? X | Y : X --> X
3696 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3698 return TrueWhenUnset ? TrueVal : FalseVal;
3700 // (X & Y) == 0 ? X : X | Y --> X
3701 // (X & Y) != 0 ? X : X | Y --> X | Y
3702 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3704 return TrueWhenUnset ? TrueVal : FalseVal;
3710 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3712 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3713 ICmpInst::Predicate Pred,
3714 Value *TrueVal, Value *FalseVal) {
3717 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3720 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3721 Pred == ICmpInst::ICMP_EQ);
3724 /// Try to simplify a select instruction when its condition operand is an
3725 /// integer comparison.
3726 static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3727 Value *FalseVal, const SimplifyQuery &Q,
3728 unsigned MaxRecurse) {
3729 ICmpInst::Predicate Pred;
3730 Value *CmpLHS, *CmpRHS;
3731 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3734 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3737 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3738 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3739 Pred == ICmpInst::ICMP_EQ))
3743 // Check for other compares that behave like bit test.
3744 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3748 // If we have an equality comparison, then we know the value in one of the
3749 // arms of the select. See if substituting this value into the arm and
3750 // simplifying the result yields the same value as the other arm.
3751 if (Pred == ICmpInst::ICMP_EQ) {
3752 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3754 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3757 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3759 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3762 } else if (Pred == ICmpInst::ICMP_NE) {
3763 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3765 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3768 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3770 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3778 /// Given operands for a SelectInst, see if we can fold the result.
3779 /// If not, this returns null.
3780 static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3781 const SimplifyQuery &Q, unsigned MaxRecurse) {
3782 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3783 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3784 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3785 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3787 // select undef, X, Y -> X or Y
3788 if (isa<UndefValue>(CondC))
3789 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3791 // TODO: Vector constants with undef elements don't simplify.
3793 // select true, X, Y -> X
3794 if (CondC->isAllOnesValue())
3796 // select false, X, Y -> Y
3797 if (CondC->isNullValue())
3801 // select ?, X, X -> X
3802 if (TrueVal == FalseVal)
3805 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3807 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3811 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3814 if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
3820 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3821 const SimplifyQuery &Q) {
3822 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3825 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3826 /// If not, this returns null.
3827 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3828 const SimplifyQuery &Q, unsigned) {
3829 // The type of the GEP pointer operand.
3831 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3833 // getelementptr P -> P.
3834 if (Ops.size() == 1)
3837 // Compute the (pointer) type returned by the GEP instruction.
3838 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3839 Type *GEPTy = PointerType::get(LastType, AS);
3840 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3841 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3842 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3843 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3845 if (isa<UndefValue>(Ops[0]))
3846 return UndefValue::get(GEPTy);
3848 if (Ops.size() == 2) {
3849 // getelementptr P, 0 -> P.
3850 if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3854 if (Ty->isSized()) {
3857 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3858 // getelementptr P, N -> P if P points to a type of zero size.
3859 if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3862 // The following transforms are only safe if the ptrtoint cast
3863 // doesn't truncate the pointers.
3864 if (Ops[1]->getType()->getScalarSizeInBits() ==
3865 Q.DL.getIndexSizeInBits(AS)) {
3866 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3867 if (match(P, m_Zero()))
3868 return Constant::getNullValue(GEPTy);
3870 if (match(P, m_PtrToInt(m_Value(Temp))))
3871 if (Temp->getType() == GEPTy)
3876 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3877 if (TyAllocSize == 1 &&
3878 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3879 if (Value *R = PtrToIntOrZero(P))
3882 // getelementptr V, (ashr (sub P, V), C) -> Q
3883 // if P points to a type of size 1 << C.
3885 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3886 m_ConstantInt(C))) &&
3887 TyAllocSize == 1ULL << C)
3888 if (Value *R = PtrToIntOrZero(P))
3891 // getelementptr V, (sdiv (sub P, V), C) -> Q
3892 // if P points to a type of size C.
3894 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3895 m_SpecificInt(TyAllocSize))))
3896 if (Value *R = PtrToIntOrZero(P))
3902 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3903 all_of(Ops.slice(1).drop_back(1),
3904 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3906 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3907 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3908 APInt BasePtrOffset(IdxWidth, 0);
3909 Value *StrippedBasePtr =
3910 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3913 // gep (gep V, C), (sub 0, V) -> C
3914 if (match(Ops.back(),
3915 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3916 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3917 return ConstantExpr::getIntToPtr(CI, GEPTy);
3919 // gep (gep V, C), (xor V, -1) -> C-1
3920 if (match(Ops.back(),
3921 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3922 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3923 return ConstantExpr::getIntToPtr(CI, GEPTy);
3928 // Check to see if this is constant foldable.
3929 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3932 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3934 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3939 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3940 const SimplifyQuery &Q) {
3941 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3944 /// Given operands for an InsertValueInst, see if we can fold the result.
3945 /// If not, this returns null.
3946 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3947 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3949 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3950 if (Constant *CVal = dyn_cast<Constant>(Val))
3951 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3953 // insertvalue x, undef, n -> x
3954 if (match(Val, m_Undef()))
3957 // insertvalue x, (extractvalue y, n), n
3958 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3959 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3960 EV->getIndices() == Idxs) {
3961 // insertvalue undef, (extractvalue y, n), n -> y
3962 if (match(Agg, m_Undef()))
3963 return EV->getAggregateOperand();
3965 // insertvalue y, (extractvalue y, n), n -> y
3966 if (Agg == EV->getAggregateOperand())
3973 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3974 ArrayRef<unsigned> Idxs,
3975 const SimplifyQuery &Q) {
3976 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3979 Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3980 const SimplifyQuery &Q) {
3981 // Try to constant fold.
3982 auto *VecC = dyn_cast<Constant>(Vec);
3983 auto *ValC = dyn_cast<Constant>(Val);
3984 auto *IdxC = dyn_cast<Constant>(Idx);
3985 if (VecC && ValC && IdxC)
3986 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3988 // Fold into undef if index is out of bounds.
3989 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3990 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3991 if (CI->uge(NumElements))
3992 return UndefValue::get(Vec->getType());
3995 // If index is undef, it might be out of bounds (see above case)
3996 if (isa<UndefValue>(Idx))
3997 return UndefValue::get(Vec->getType());
4002 /// Given operands for an ExtractValueInst, see if we can fold the result.
4003 /// If not, this returns null.
4004 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4005 const SimplifyQuery &, unsigned) {
4006 if (auto *CAgg = dyn_cast<Constant>(Agg))
4007 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4009 // extractvalue x, (insertvalue y, elt, n), n -> elt
4010 unsigned NumIdxs = Idxs.size();
4011 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4012 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4013 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4014 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4015 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4016 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4017 Idxs.slice(0, NumCommonIdxs)) {
4018 if (NumIdxs == NumInsertValueIdxs)
4019 return IVI->getInsertedValueOperand();
4027 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4028 const SimplifyQuery &Q) {
4029 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
4032 /// Given operands for an ExtractElementInst, see if we can fold the result.
4033 /// If not, this returns null.
4034 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
4036 if (auto *CVec = dyn_cast<Constant>(Vec)) {
4037 if (auto *CIdx = dyn_cast<Constant>(Idx))
4038 return ConstantFoldExtractElementInstruction(CVec, CIdx);
4040 // The index is not relevant if our vector is a splat.
4041 if (auto *Splat = CVec->getSplatValue())
4044 if (isa<UndefValue>(Vec))
4045 return UndefValue::get(Vec->getType()->getVectorElementType());
4048 // If extracting a specified index from the vector, see if we can recursively
4049 // find a previously computed scalar that was inserted into the vector.
4050 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
4051 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4052 // definitely out of bounds, thus undefined result
4053 return UndefValue::get(Vec->getType()->getVectorElementType());
4054 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4058 // An undef extract index can be arbitrarily chosen to be an out-of-range
4059 // index value, which would result in the instruction being undef.
4060 if (isa<UndefValue>(Idx))
4061 return UndefValue::get(Vec->getType()->getVectorElementType());
4066 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
4067 const SimplifyQuery &Q) {
4068 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
4071 /// See if we can fold the given phi. If not, returns null.
4072 static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4073 // If all of the PHI's incoming values are the same then replace the PHI node
4074 // with the common value.
4075 Value *CommonValue = nullptr;
4076 bool HasUndefInput = false;
4077 for (Value *Incoming : PN->incoming_values()) {
4078 // If the incoming value is the phi node itself, it can safely be skipped.
4079 if (Incoming == PN) continue;
4080 if (isa<UndefValue>(Incoming)) {
4081 // Remember that we saw an undef value, but otherwise ignore them.
4082 HasUndefInput = true;
4085 if (CommonValue && Incoming != CommonValue)
4086 return nullptr; // Not the same, bail out.
4087 CommonValue = Incoming;
4090 // If CommonValue is null then all of the incoming values were either undef or
4091 // equal to the phi node itself.
4093 return UndefValue::get(PN->getType());
4095 // If we have a PHI node like phi(X, undef, X), where X is defined by some
4096 // instruction, we cannot return X as the result of the PHI node unless it
4097 // dominates the PHI block.
4099 return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4104 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4105 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4106 if (auto *C = dyn_cast<Constant>(Op))
4107 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4109 if (auto *CI = dyn_cast<CastInst>(Op)) {
4110 auto *Src = CI->getOperand(0);
4111 Type *SrcTy = Src->getType();
4112 Type *MidTy = CI->getType();
4114 if (Src->getType() == Ty) {
4115 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4116 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4118 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4120 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4122 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4123 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4124 SrcIntPtrTy, MidIntPtrTy,
4125 DstIntPtrTy) == Instruction::BitCast)
4131 if (CastOpc == Instruction::BitCast)
4132 if (Op->getType() == Ty)
4138 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4139 const SimplifyQuery &Q) {
4140 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4143 /// For the given destination element of a shuffle, peek through shuffles to
4144 /// match a root vector source operand that contains that element in the same
4145 /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4146 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4147 int MaskVal, Value *RootVec,
4148 unsigned MaxRecurse) {
4152 // Bail out if any mask value is undefined. That kind of shuffle may be
4153 // simplified further based on demanded bits or other folds.
4157 // The mask value chooses which source operand we need to look at next.
4158 int InVecNumElts = Op0->getType()->getVectorNumElements();
4159 int RootElt = MaskVal;
4160 Value *SourceOp = Op0;
4161 if (MaskVal >= InVecNumElts) {
4162 RootElt = MaskVal - InVecNumElts;
4166 // If the source operand is a shuffle itself, look through it to find the
4167 // matching root vector.
4168 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4169 return foldIdentityShuffles(
4170 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4171 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4174 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4177 // The source operand is not a shuffle. Initialize the root vector value for
4178 // this shuffle if that has not been done yet.
4182 // Give up as soon as a source operand does not match the existing root value.
4183 if (RootVec != SourceOp)
4186 // The element must be coming from the same lane in the source vector
4187 // (although it may have crossed lanes in intermediate shuffles).
4188 if (RootElt != DestElt)
4194 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4195 Type *RetTy, const SimplifyQuery &Q,
4196 unsigned MaxRecurse) {
4197 if (isa<UndefValue>(Mask))
4198 return UndefValue::get(RetTy);
4200 Type *InVecTy = Op0->getType();
4201 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4202 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4204 SmallVector<int, 32> Indices;
4205 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4206 assert(MaskNumElts == Indices.size() &&
4207 "Size of Indices not same as number of mask elements?");
4209 // Canonicalization: If mask does not select elements from an input vector,
4210 // replace that input vector with undef.
4211 bool MaskSelects0 = false, MaskSelects1 = false;
4212 for (unsigned i = 0; i != MaskNumElts; ++i) {
4213 if (Indices[i] == -1)
4215 if ((unsigned)Indices[i] < InVecNumElts)
4216 MaskSelects0 = true;
4218 MaskSelects1 = true;
4221 Op0 = UndefValue::get(InVecTy);
4223 Op1 = UndefValue::get(InVecTy);
4225 auto *Op0Const = dyn_cast<Constant>(Op0);
4226 auto *Op1Const = dyn_cast<Constant>(Op1);
4228 // If all operands are constant, constant fold the shuffle.
4229 if (Op0Const && Op1Const)
4230 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4232 // Canonicalization: if only one input vector is constant, it shall be the
4234 if (Op0Const && !Op1Const) {
4235 std::swap(Op0, Op1);
4236 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4239 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4240 // value type is same as the input vectors' type.
4241 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4242 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4243 OpShuf->getMask()->getSplatValue())
4246 // Don't fold a shuffle with undef mask elements. This may get folded in a
4247 // better way using demanded bits or other analysis.
4248 // TODO: Should we allow this?
4249 if (find(Indices, -1) != Indices.end())
4252 // Check if every element of this shuffle can be mapped back to the
4253 // corresponding element of a single root vector. If so, we don't need this
4254 // shuffle. This handles simple identity shuffles as well as chains of
4255 // shuffles that may widen/narrow and/or move elements across lanes and back.
4256 Value *RootVec = nullptr;
4257 for (unsigned i = 0; i != MaskNumElts; ++i) {
4258 // Note that recursion is limited for each vector element, so if any element
4259 // exceeds the limit, this will fail to simplify.
4261 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4263 // We can't replace a widening/narrowing shuffle with one of its operands.
4264 if (!RootVec || RootVec->getType() != RetTy)
4270 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4271 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4272 Type *RetTy, const SimplifyQuery &Q) {
4273 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4276 static Constant *propagateNaN(Constant *In) {
4277 // If the input is a vector with undef elements, just return a default NaN.
4279 return ConstantFP::getNaN(In->getType());
4281 // Propagate the existing NaN constant when possible.
4282 // TODO: Should we quiet a signaling NaN?
4286 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4287 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4288 return ConstantFP::getNaN(Op0->getType());
4290 if (match(Op0, m_NaN()))
4291 return propagateNaN(cast<Constant>(Op0));
4292 if (match(Op1, m_NaN()))
4293 return propagateNaN(cast<Constant>(Op1));
4298 /// Given operands for an FAdd, see if we can fold the result. If not, this
4300 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4301 const SimplifyQuery &Q, unsigned MaxRecurse) {
4302 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4305 if (Constant *C = simplifyFPBinop(Op0, Op1))
4309 if (match(Op1, m_NegZeroFP()))
4312 // fadd X, 0 ==> X, when we know X is not -0
4313 if (match(Op1, m_PosZeroFP()) &&
4314 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4317 // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4318 // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4319 // Negative zeros are allowed because we always end up with positive zero:
4320 // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4321 // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4322 // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4323 // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4324 if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4325 match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4326 return ConstantFP::getNullValue(Op0->getType());
4331 /// Given operands for an FSub, see if we can fold the result. If not, this
4333 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4334 const SimplifyQuery &Q, unsigned MaxRecurse) {
4335 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4338 if (Constant *C = simplifyFPBinop(Op0, Op1))
4342 if (match(Op1, m_PosZeroFP()))
4345 // fsub X, -0 ==> X, when we know X is not -0
4346 if (match(Op1, m_NegZeroFP()) &&
4347 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4350 // fsub -0.0, (fsub -0.0, X) ==> X
4352 if (match(Op0, m_NegZeroFP()) &&
4353 match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4356 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4357 if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4358 match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4361 // fsub nnan x, x ==> 0.0
4362 if (FMF.noNaNs() && Op0 == Op1)
4363 return Constant::getNullValue(Op0->getType());
4368 /// Given the operands for an FMul, see if we can fold the result
4369 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4370 const SimplifyQuery &Q, unsigned MaxRecurse) {
4371 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4374 if (Constant *C = simplifyFPBinop(Op0, Op1))
4377 // fmul X, 1.0 ==> X
4378 if (match(Op1, m_FPOne()))
4381 // fmul nnan nsz X, 0 ==> 0
4382 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4383 return ConstantFP::getNullValue(Op0->getType());
4385 // sqrt(X) * sqrt(X) --> X, if we can:
4386 // 1. Remove the intermediate rounding (reassociate).
4387 // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4388 // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4390 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4391 FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4397 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4398 const SimplifyQuery &Q) {
4399 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4403 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4404 const SimplifyQuery &Q) {
4405 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4408 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4409 const SimplifyQuery &Q) {
4410 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4413 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4414 const SimplifyQuery &Q, unsigned) {
4415 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4418 if (Constant *C = simplifyFPBinop(Op0, Op1))
4422 if (match(Op1, m_FPOne()))
4426 // Requires that NaNs are off (X could be zero) and signed zeroes are
4427 // ignored (X could be positive or negative, so the output sign is unknown).
4428 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4429 return ConstantFP::getNullValue(Op0->getType());
4432 // X / X -> 1.0 is legal when NaNs are ignored.
4433 // We can ignore infinities because INF/INF is NaN.
4435 return ConstantFP::get(Op0->getType(), 1.0);
4437 // (X * Y) / Y --> X if we can reassociate to the above form.
4439 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4442 // -X / X -> -1.0 and
4443 // X / -X -> -1.0 are legal when NaNs are ignored.
4444 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4445 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4446 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4447 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4448 BinaryOperator::getFNegArgument(Op1) == Op0))
4449 return ConstantFP::get(Op0->getType(), -1.0);
4455 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4456 const SimplifyQuery &Q) {
4457 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4460 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4461 const SimplifyQuery &Q, unsigned) {
4462 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4465 if (Constant *C = simplifyFPBinop(Op0, Op1))
4468 // Unlike fdiv, the result of frem always matches the sign of the dividend.
4469 // The constant match may include undef elements in a vector, so return a full
4470 // zero constant as the result.
4473 if (match(Op0, m_PosZeroFP()))
4474 return ConstantFP::getNullValue(Op0->getType());
4476 if (match(Op0, m_NegZeroFP()))
4477 return ConstantFP::getNegativeZero(Op0->getType());
4483 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4484 const SimplifyQuery &Q) {
4485 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4488 //=== Helper functions for higher up the class hierarchy.
4490 /// Given operands for a BinaryOperator, see if we can fold the result.
4491 /// If not, this returns null.
4492 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4493 const SimplifyQuery &Q, unsigned MaxRecurse) {
4495 case Instruction::Add:
4496 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4497 case Instruction::Sub:
4498 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4499 case Instruction::Mul:
4500 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4501 case Instruction::SDiv:
4502 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4503 case Instruction::UDiv:
4504 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4505 case Instruction::SRem:
4506 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4507 case Instruction::URem:
4508 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4509 case Instruction::Shl:
4510 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4511 case Instruction::LShr:
4512 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4513 case Instruction::AShr:
4514 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4515 case Instruction::And:
4516 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4517 case Instruction::Or:
4518 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4519 case Instruction::Xor:
4520 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4521 case Instruction::FAdd:
4522 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4523 case Instruction::FSub:
4524 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4525 case Instruction::FMul:
4526 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4527 case Instruction::FDiv:
4528 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4529 case Instruction::FRem:
4530 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4532 llvm_unreachable("Unexpected opcode");
4536 /// Given operands for a BinaryOperator, see if we can fold the result.
4537 /// If not, this returns null.
4538 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4539 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4540 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4541 const FastMathFlags &FMF, const SimplifyQuery &Q,
4542 unsigned MaxRecurse) {
4544 case Instruction::FAdd:
4545 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4546 case Instruction::FSub:
4547 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4548 case Instruction::FMul:
4549 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4550 case Instruction::FDiv:
4551 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4553 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4557 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4558 const SimplifyQuery &Q) {
4559 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4562 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4563 FastMathFlags FMF, const SimplifyQuery &Q) {
4564 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4567 /// Given operands for a CmpInst, see if we can fold the result.
4568 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4569 const SimplifyQuery &Q, unsigned MaxRecurse) {
4570 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4571 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4572 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4575 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4576 const SimplifyQuery &Q) {
4577 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4580 static bool IsIdempotent(Intrinsic::ID ID) {
4582 default: return false;
4584 // Unary idempotent: f(f(x)) = f(x)
4585 case Intrinsic::fabs:
4586 case Intrinsic::floor:
4587 case Intrinsic::ceil:
4588 case Intrinsic::trunc:
4589 case Intrinsic::rint:
4590 case Intrinsic::nearbyint:
4591 case Intrinsic::round:
4592 case Intrinsic::canonicalize:
4597 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4598 const DataLayout &DL) {
4599 GlobalValue *PtrSym;
4601 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4604 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4605 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4606 Type *Int32PtrTy = Int32Ty->getPointerTo();
4607 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4609 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4610 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4613 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4614 if (OffsetInt % 4 != 0)
4617 Constant *C = ConstantExpr::getGetElementPtr(
4618 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4619 ConstantInt::get(Int64Ty, OffsetInt / 4));
4620 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4624 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4628 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4629 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4634 if (LoadedCE->getOpcode() != Instruction::Sub)
4637 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4638 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4640 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4642 Constant *LoadedRHS = LoadedCE->getOperand(1);
4643 GlobalValue *LoadedRHSSym;
4644 APInt LoadedRHSOffset;
4645 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4647 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4650 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4653 static bool maskIsAllZeroOrUndef(Value *Mask) {
4654 auto *ConstMask = dyn_cast<Constant>(Mask);
4657 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4659 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4661 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4662 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4669 static Value *simplifyUnaryIntrinsic(Function *F, Value *Op0,
4670 const SimplifyQuery &Q) {
4671 // Idempotent functions return the same result when called repeatedly.
4672 Intrinsic::ID IID = F->getIntrinsicID();
4673 if (IsIdempotent(IID))
4674 if (auto *II = dyn_cast<IntrinsicInst>(Op0))
4675 if (II->getIntrinsicID() == IID)
4680 case Intrinsic::fabs:
4681 if (SignBitMustBeZero(Op0, Q.TLI)) return Op0;
4683 case Intrinsic::bswap:
4684 // bswap(bswap(x)) -> x
4685 if (match(Op0, m_BSwap(m_Value(X)))) return X;
4687 case Intrinsic::bitreverse:
4688 // bitreverse(bitreverse(x)) -> x
4689 if (match(Op0, m_BitReverse(m_Value(X)))) return X;
4691 case Intrinsic::exp:
4693 if (Q.CxtI->hasAllowReassoc() &&
4694 match(Op0, m_Intrinsic<Intrinsic::log>(m_Value(X)))) return X;
4696 case Intrinsic::exp2:
4697 // exp2(log2(x)) -> x
4698 if (Q.CxtI->hasAllowReassoc() &&
4699 match(Op0, m_Intrinsic<Intrinsic::log2>(m_Value(X)))) return X;
4701 case Intrinsic::log:
4703 if (Q.CxtI->hasAllowReassoc() &&
4704 match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X)))) return X;
4706 case Intrinsic::log2:
4707 // log2(exp2(x)) -> x
4708 if (Q.CxtI->hasAllowReassoc() &&
4709 match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) return X;
4718 static Value *simplifyBinaryIntrinsic(Function *F, Value *Op0, Value *Op1,
4719 const SimplifyQuery &Q) {
4720 Intrinsic::ID IID = F->getIntrinsicID();
4721 Type *ReturnType = F->getReturnType();
4723 case Intrinsic::usub_with_overflow:
4724 case Intrinsic::ssub_with_overflow:
4725 // X - X -> { 0, false }
4727 return Constant::getNullValue(ReturnType);
4728 // X - undef -> undef
4729 // undef - X -> undef
4730 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4731 return UndefValue::get(ReturnType);
4733 case Intrinsic::uadd_with_overflow:
4734 case Intrinsic::sadd_with_overflow:
4735 // X + undef -> undef
4736 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4737 return UndefValue::get(ReturnType);
4739 case Intrinsic::umul_with_overflow:
4740 case Intrinsic::smul_with_overflow:
4741 // 0 * X -> { 0, false }
4742 // X * 0 -> { 0, false }
4743 if (match(Op0, m_Zero()) || match(Op1, m_Zero()))
4744 return Constant::getNullValue(ReturnType);
4745 // undef * X -> { 0, false }
4746 // X * undef -> { 0, false }
4747 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
4748 return Constant::getNullValue(ReturnType);
4750 case Intrinsic::load_relative:
4751 if (auto *C0 = dyn_cast<Constant>(Op0))
4752 if (auto *C1 = dyn_cast<Constant>(Op1))
4753 return SimplifyRelativeLoad(C0, C1, Q.DL);
4755 case Intrinsic::powi:
4756 if (auto *Power = dyn_cast<ConstantInt>(Op1)) {
4757 // powi(x, 0) -> 1.0
4758 if (Power->isZero())
4759 return ConstantFP::get(Op0->getType(), 1.0);
4765 case Intrinsic::maxnum:
4766 case Intrinsic::minnum:
4767 // If one argument is NaN, return the other argument.
4768 if (match(Op0, m_NaN())) return Op1;
4769 if (match(Op1, m_NaN())) return Op0;
4778 template <typename IterTy>
4779 static Value *simplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4780 const SimplifyQuery &Q) {
4781 // Intrinsics with no operands have some kind of side effect. Don't simplify.
4782 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4783 if (NumOperands == 0)
4786 Intrinsic::ID IID = F->getIntrinsicID();
4787 if (NumOperands == 1)
4788 return simplifyUnaryIntrinsic(F, ArgBegin[0], Q);
4790 if (NumOperands == 2)
4791 return simplifyBinaryIntrinsic(F, ArgBegin[0], ArgBegin[1], Q);
4793 // Handle intrinsics with 3 or more arguments.
4795 case Intrinsic::masked_load: {
4796 Value *MaskArg = ArgBegin[2];
4797 Value *PassthruArg = ArgBegin[3];
4798 // If the mask is all zeros or undef, the "passthru" argument is the result.
4799 if (maskIsAllZeroOrUndef(MaskArg))
4803 case Intrinsic::fshl:
4804 case Intrinsic::fshr: {
4805 Value *ShAmtArg = ArgBegin[2];
4806 const APInt *ShAmtC;
4807 if (match(ShAmtArg, m_APInt(ShAmtC))) {
4808 // If there's effectively no shift, return the 1st arg or 2nd arg.
4809 // TODO: For vectors, we could check each element of a non-splat constant.
4810 APInt BitWidth = APInt(ShAmtC->getBitWidth(), ShAmtC->getBitWidth());
4811 if (ShAmtC->urem(BitWidth).isNullValue())
4812 return ArgBegin[IID == Intrinsic::fshl ? 0 : 1];
4821 template <typename IterTy>
4822 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4823 IterTy ArgEnd, const SimplifyQuery &Q,
4824 unsigned MaxRecurse) {
4825 Type *Ty = V->getType();
4826 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4827 Ty = PTy->getElementType();
4828 FunctionType *FTy = cast<FunctionType>(Ty);
4830 // call undef -> undef
4831 // call null -> undef
4832 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4833 return UndefValue::get(FTy->getReturnType());
4835 Function *F = dyn_cast<Function>(V);
4839 if (F->isIntrinsic())
4840 if (Value *Ret = simplifyIntrinsic(F, ArgBegin, ArgEnd, Q))
4843 if (!canConstantFoldCallTo(CS, F))
4846 SmallVector<Constant *, 4> ConstantArgs;
4847 ConstantArgs.reserve(ArgEnd - ArgBegin);
4848 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4849 Constant *C = dyn_cast<Constant>(*I);
4852 ConstantArgs.push_back(C);
4855 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4858 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4859 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4860 const SimplifyQuery &Q) {
4861 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4864 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4865 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4866 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4869 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4870 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4871 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4875 /// See if we can compute a simplified version of this instruction.
4876 /// If not, this returns null.
4878 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4879 OptimizationRemarkEmitter *ORE) {
4880 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4883 switch (I->getOpcode()) {
4885 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4887 case Instruction::FAdd:
4888 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4889 I->getFastMathFlags(), Q);
4891 case Instruction::Add:
4892 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4893 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4894 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4896 case Instruction::FSub:
4897 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4898 I->getFastMathFlags(), Q);
4900 case Instruction::Sub:
4901 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4902 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4903 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4905 case Instruction::FMul:
4906 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4907 I->getFastMathFlags(), Q);
4909 case Instruction::Mul:
4910 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4912 case Instruction::SDiv:
4913 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4915 case Instruction::UDiv:
4916 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4918 case Instruction::FDiv:
4919 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4920 I->getFastMathFlags(), Q);
4922 case Instruction::SRem:
4923 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4925 case Instruction::URem:
4926 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4928 case Instruction::FRem:
4929 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4930 I->getFastMathFlags(), Q);
4932 case Instruction::Shl:
4933 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4934 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4935 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4937 case Instruction::LShr:
4938 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4939 cast<BinaryOperator>(I)->isExact(), Q);
4941 case Instruction::AShr:
4942 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4943 cast<BinaryOperator>(I)->isExact(), Q);
4945 case Instruction::And:
4946 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4948 case Instruction::Or:
4949 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4951 case Instruction::Xor:
4952 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4954 case Instruction::ICmp:
4955 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4956 I->getOperand(0), I->getOperand(1), Q);
4958 case Instruction::FCmp:
4960 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4961 I->getOperand(1), I->getFastMathFlags(), Q);
4963 case Instruction::Select:
4964 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4965 I->getOperand(2), Q);
4967 case Instruction::GetElementPtr: {
4968 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4969 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4973 case Instruction::InsertValue: {
4974 InsertValueInst *IV = cast<InsertValueInst>(I);
4975 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4976 IV->getInsertedValueOperand(),
4977 IV->getIndices(), Q);
4980 case Instruction::InsertElement: {
4981 auto *IE = cast<InsertElementInst>(I);
4982 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4983 IE->getOperand(2), Q);
4986 case Instruction::ExtractValue: {
4987 auto *EVI = cast<ExtractValueInst>(I);
4988 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4989 EVI->getIndices(), Q);
4992 case Instruction::ExtractElement: {
4993 auto *EEI = cast<ExtractElementInst>(I);
4994 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4995 EEI->getIndexOperand(), Q);
4998 case Instruction::ShuffleVector: {
4999 auto *SVI = cast<ShuffleVectorInst>(I);
5000 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5001 SVI->getMask(), SVI->getType(), Q);
5004 case Instruction::PHI:
5005 Result = SimplifyPHINode(cast<PHINode>(I), Q);
5007 case Instruction::Call: {
5008 CallSite CS(cast<CallInst>(I));
5009 Result = SimplifyCall(CS, Q);
5012 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
5013 #include "llvm/IR/Instruction.def"
5014 #undef HANDLE_CAST_INST
5016 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
5018 case Instruction::Alloca:
5019 // No simplifications for Alloca and it can't be constant folded.
5024 // In general, it is possible for computeKnownBits to determine all bits in a
5025 // value even when the operands are not all constants.
5026 if (!Result && I->getType()->isIntOrIntVectorTy()) {
5027 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
5028 if (Known.isConstant())
5029 Result = ConstantInt::get(I->getType(), Known.getConstant());
5032 /// If called on unreachable code, the above logic may report that the
5033 /// instruction simplified to itself. Make life easier for users by
5034 /// detecting that case here, returning a safe value instead.
5035 return Result == I ? UndefValue::get(I->getType()) : Result;
5038 /// Implementation of recursive simplification through an instruction's
5041 /// This is the common implementation of the recursive simplification routines.
5042 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
5043 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
5044 /// instructions to process and attempt to simplify it using
5045 /// InstructionSimplify.
5047 /// This routine returns 'true' only when *it* simplifies something. The passed
5048 /// in simplified value does not count toward this.
5049 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
5050 const TargetLibraryInfo *TLI,
5051 const DominatorTree *DT,
5052 AssumptionCache *AC) {
5053 bool Simplified = false;
5054 SmallSetVector<Instruction *, 8> Worklist;
5055 const DataLayout &DL = I->getModule()->getDataLayout();
5057 // If we have an explicit value to collapse to, do that round of the
5058 // simplification loop by hand initially.
5060 for (User *U : I->users())
5062 Worklist.insert(cast<Instruction>(U));
5064 // Replace the instruction with its simplified value.
5065 I->replaceAllUsesWith(SimpleV);
5067 // Gracefully handle edge cases where the instruction is not wired into any
5069 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
5070 !I->mayHaveSideEffects())
5071 I->eraseFromParent();
5076 // Note that we must test the size on each iteration, the worklist can grow.
5077 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
5080 // See if this instruction simplifies.
5081 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
5087 // Stash away all the uses of the old instruction so we can check them for
5088 // recursive simplifications after a RAUW. This is cheaper than checking all
5089 // uses of To on the recursive step in most cases.
5090 for (User *U : I->users())
5091 Worklist.insert(cast<Instruction>(U));
5093 // Replace the instruction with its simplified value.
5094 I->replaceAllUsesWith(SimpleV);
5096 // Gracefully handle edge cases where the instruction is not wired into any
5098 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
5099 !I->mayHaveSideEffects())
5100 I->eraseFromParent();
5105 bool llvm::recursivelySimplifyInstruction(Instruction *I,
5106 const TargetLibraryInfo *TLI,
5107 const DominatorTree *DT,
5108 AssumptionCache *AC) {
5109 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
5112 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
5113 const TargetLibraryInfo *TLI,
5114 const DominatorTree *DT,
5115 AssumptionCache *AC) {
5116 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5117 assert(SimpleV && "Must provide a simplified value.");
5118 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
5122 const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
5123 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
5124 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
5125 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
5126 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
5127 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5128 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5129 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5132 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5133 const DataLayout &DL) {
5134 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5137 template <class T, class... TArgs>
5138 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5140 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5141 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5142 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5143 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5145 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,