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/ConstantFolding.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/GlobalAlias.h"
31 #include "llvm/IR/Operator.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/IR/ValueHandle.h"
35 using namespace llvm::PatternMatch;
37 #define DEBUG_TYPE "instsimplify"
39 enum { RecursionLimit = 3 };
41 STATISTIC(NumExpand, "Number of expansions");
42 STATISTIC(NumFactor , "Number of factorizations");
43 STATISTIC(NumReassoc, "Number of reassociations");
47 const TargetLibraryInfo *TLI;
48 const DominatorTree *DT;
50 Query(const DataLayout *DL, const TargetLibraryInfo *tli,
51 const DominatorTree *dt) : DL(DL), TLI(tli), DT(dt) {}
54 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
55 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
57 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
59 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
60 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
61 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned);
63 /// getFalse - For a boolean type, or a vector of boolean type, return false, or
64 /// a vector with every element false, as appropriate for the type.
65 static Constant *getFalse(Type *Ty) {
66 assert(Ty->getScalarType()->isIntegerTy(1) &&
67 "Expected i1 type or a vector of i1!");
68 return Constant::getNullValue(Ty);
71 /// getTrue - For a boolean type, or a vector of boolean type, return true, or
72 /// a vector with every element true, as appropriate for the type.
73 static Constant *getTrue(Type *Ty) {
74 assert(Ty->getScalarType()->isIntegerTy(1) &&
75 "Expected i1 type or a vector of i1!");
76 return Constant::getAllOnesValue(Ty);
79 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
80 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
82 CmpInst *Cmp = dyn_cast<CmpInst>(V);
85 CmpInst::Predicate CPred = Cmp->getPredicate();
86 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
87 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
89 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
93 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
94 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
95 Instruction *I = dyn_cast<Instruction>(V);
97 // Arguments and constants dominate all instructions.
100 // If we are processing instructions (and/or basic blocks) that have not been
101 // fully added to a function, the parent nodes may still be null. Simply
102 // return the conservative answer in these cases.
103 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
106 // If we have a DominatorTree then do a precise test.
108 if (!DT->isReachableFromEntry(P->getParent()))
110 if (!DT->isReachableFromEntry(I->getParent()))
112 return DT->dominates(I, P);
115 // Otherwise, if the instruction is in the entry block, and is not an invoke,
116 // then it obviously dominates all phi nodes.
117 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
124 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning
125 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
126 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
127 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
128 /// Returns the simplified value, or null if no simplification was performed.
129 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
130 unsigned OpcToExpand, const Query &Q,
131 unsigned MaxRecurse) {
132 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
133 // Recursion is always used, so bail out at once if we already hit the limit.
137 // Check whether the expression has the form "(A op' B) op C".
138 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
139 if (Op0->getOpcode() == OpcodeToExpand) {
140 // It does! Try turning it into "(A op C) op' (B op C)".
141 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
142 // Do "A op C" and "B op C" both simplify?
143 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
144 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
145 // They do! Return "L op' R" if it simplifies or is already available.
146 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
147 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
148 && L == B && R == A)) {
152 // Otherwise return "L op' R" if it simplifies.
153 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
160 // Check whether the expression has the form "A op (B op' C)".
161 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
162 if (Op1->getOpcode() == OpcodeToExpand) {
163 // It does! Try turning it into "(A op B) op' (A op C)".
164 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
165 // Do "A op B" and "A op C" both simplify?
166 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
167 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
168 // They do! Return "L op' R" if it simplifies or is already available.
169 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
170 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
171 && L == C && R == B)) {
175 // Otherwise return "L op' R" if it simplifies.
176 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
186 /// FactorizeBinOp - Simplify "LHS Opcode RHS" by factorizing out a common term
187 /// using the operation OpCodeToExtract. For example, when Opcode is Add and
188 /// OpCodeToExtract is Mul then this tries to turn "(A*B)+(A*C)" into "A*(B+C)".
189 /// Returns the simplified value, or null if no simplification was performed.
190 static Value *FactorizeBinOp(unsigned Opcode, Value *LHS, Value *RHS,
191 unsigned OpcToExtract, const Query &Q,
192 unsigned MaxRecurse) {
193 Instruction::BinaryOps OpcodeToExtract = (Instruction::BinaryOps)OpcToExtract;
194 // Recursion is always used, so bail out at once if we already hit the limit.
198 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
199 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
201 if (!Op0 || Op0->getOpcode() != OpcodeToExtract ||
202 !Op1 || Op1->getOpcode() != OpcodeToExtract)
205 // The expression has the form "(A op' B) op (C op' D)".
206 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1);
207 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1);
209 // Use left distributivity, i.e. "X op' (Y op Z) = (X op' Y) op (X op' Z)".
210 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
211 // commutative case, "(A op' B) op (C op' A)"?
212 if (A == C || (Instruction::isCommutative(OpcodeToExtract) && A == D)) {
213 Value *DD = A == C ? D : C;
214 // Form "A op' (B op DD)" if it simplifies completely.
215 // Does "B op DD" simplify?
216 if (Value *V = SimplifyBinOp(Opcode, B, DD, Q, MaxRecurse)) {
217 // It does! Return "A op' V" if it simplifies or is already available.
218 // If V equals B then "A op' V" is just the LHS. If V equals DD then
219 // "A op' V" is just the RHS.
220 if (V == B || V == DD) {
222 return V == B ? LHS : RHS;
224 // Otherwise return "A op' V" if it simplifies.
225 if (Value *W = SimplifyBinOp(OpcodeToExtract, A, V, Q, MaxRecurse)) {
232 // Use right distributivity, i.e. "(X op Y) op' Z = (X op' Z) op (Y op' Z)".
233 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
234 // commutative case, "(A op' B) op (B op' D)"?
235 if (B == D || (Instruction::isCommutative(OpcodeToExtract) && B == C)) {
236 Value *CC = B == D ? C : D;
237 // Form "(A op CC) op' B" if it simplifies completely..
238 // Does "A op CC" simplify?
239 if (Value *V = SimplifyBinOp(Opcode, A, CC, Q, MaxRecurse)) {
240 // It does! Return "V op' B" if it simplifies or is already available.
241 // If V equals A then "V op' B" is just the LHS. If V equals CC then
242 // "V op' B" is just the RHS.
243 if (V == A || V == CC) {
245 return V == A ? LHS : RHS;
247 // Otherwise return "V op' B" if it simplifies.
248 if (Value *W = SimplifyBinOp(OpcodeToExtract, V, B, Q, MaxRecurse)) {
258 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
259 /// operations. Returns the simpler value, or null if none was found.
260 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
261 const Query &Q, unsigned MaxRecurse) {
262 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
263 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
265 // Recursion is always used, so bail out at once if we already hit the limit.
269 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
270 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
272 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
273 if (Op0 && Op0->getOpcode() == Opcode) {
274 Value *A = Op0->getOperand(0);
275 Value *B = Op0->getOperand(1);
278 // Does "B op C" simplify?
279 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
280 // It does! Return "A op V" if it simplifies or is already available.
281 // If V equals B then "A op V" is just the LHS.
282 if (V == B) return LHS;
283 // Otherwise return "A op V" if it simplifies.
284 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
291 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
292 if (Op1 && Op1->getOpcode() == Opcode) {
294 Value *B = Op1->getOperand(0);
295 Value *C = Op1->getOperand(1);
297 // Does "A op B" simplify?
298 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
299 // It does! Return "V op C" if it simplifies or is already available.
300 // If V equals B then "V op C" is just the RHS.
301 if (V == B) return RHS;
302 // Otherwise return "V op C" if it simplifies.
303 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
310 // The remaining transforms require commutativity as well as associativity.
311 if (!Instruction::isCommutative(Opcode))
314 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
315 if (Op0 && Op0->getOpcode() == Opcode) {
316 Value *A = Op0->getOperand(0);
317 Value *B = Op0->getOperand(1);
320 // Does "C op A" simplify?
321 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
322 // It does! Return "V op B" if it simplifies or is already available.
323 // If V equals A then "V op B" is just the LHS.
324 if (V == A) return LHS;
325 // Otherwise return "V op B" if it simplifies.
326 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
333 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
334 if (Op1 && Op1->getOpcode() == Opcode) {
336 Value *B = Op1->getOperand(0);
337 Value *C = Op1->getOperand(1);
339 // Does "C op A" simplify?
340 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
341 // It does! Return "B op V" if it simplifies or is already available.
342 // If V equals C then "B op V" is just the RHS.
343 if (V == C) return RHS;
344 // Otherwise return "B op V" if it simplifies.
345 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
355 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
356 /// instruction as an operand, try to simplify the binop by seeing whether
357 /// evaluating it on both branches of the select results in the same value.
358 /// Returns the common value if so, otherwise returns null.
359 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
360 const Query &Q, unsigned MaxRecurse) {
361 // Recursion is always used, so bail out at once if we already hit the limit.
366 if (isa<SelectInst>(LHS)) {
367 SI = cast<SelectInst>(LHS);
369 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
370 SI = cast<SelectInst>(RHS);
373 // Evaluate the BinOp on the true and false branches of the select.
377 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
378 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
380 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
381 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
384 // If they simplified to the same value, then return the common value.
385 // If they both failed to simplify then return null.
389 // If one branch simplified to undef, return the other one.
390 if (TV && isa<UndefValue>(TV))
392 if (FV && isa<UndefValue>(FV))
395 // If applying the operation did not change the true and false select values,
396 // then the result of the binop is the select itself.
397 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
400 // If one branch simplified and the other did not, and the simplified
401 // value is equal to the unsimplified one, return the simplified value.
402 // For example, select (cond, X, X & Z) & Z -> X & Z.
403 if ((FV && !TV) || (TV && !FV)) {
404 // Check that the simplified value has the form "X op Y" where "op" is the
405 // same as the original operation.
406 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
407 if (Simplified && Simplified->getOpcode() == Opcode) {
408 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
409 // We already know that "op" is the same as for the simplified value. See
410 // if the operands match too. If so, return the simplified value.
411 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
412 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
413 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
414 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
415 Simplified->getOperand(1) == UnsimplifiedRHS)
417 if (Simplified->isCommutative() &&
418 Simplified->getOperand(1) == UnsimplifiedLHS &&
419 Simplified->getOperand(0) == UnsimplifiedRHS)
427 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
428 /// try to simplify the comparison by seeing whether both branches of the select
429 /// result in the same value. Returns the common value if so, otherwise returns
431 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
432 Value *RHS, const Query &Q,
433 unsigned MaxRecurse) {
434 // Recursion is always used, so bail out at once if we already hit the limit.
438 // Make sure the select is on the LHS.
439 if (!isa<SelectInst>(LHS)) {
441 Pred = CmpInst::getSwappedPredicate(Pred);
443 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
444 SelectInst *SI = cast<SelectInst>(LHS);
445 Value *Cond = SI->getCondition();
446 Value *TV = SI->getTrueValue();
447 Value *FV = SI->getFalseValue();
449 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
450 // Does "cmp TV, RHS" simplify?
451 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
453 // It not only simplified, it simplified to the select condition. Replace
455 TCmp = getTrue(Cond->getType());
457 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
458 // condition then we can replace it with 'true'. Otherwise give up.
459 if (!isSameCompare(Cond, Pred, TV, RHS))
461 TCmp = getTrue(Cond->getType());
464 // Does "cmp FV, RHS" simplify?
465 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
467 // It not only simplified, it simplified to the select condition. Replace
469 FCmp = getFalse(Cond->getType());
471 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
472 // condition then we can replace it with 'false'. Otherwise give up.
473 if (!isSameCompare(Cond, Pred, FV, RHS))
475 FCmp = getFalse(Cond->getType());
478 // If both sides simplified to the same value, then use it as the result of
479 // the original comparison.
483 // The remaining cases only make sense if the select condition has the same
484 // type as the result of the comparison, so bail out if this is not so.
485 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
487 // If the false value simplified to false, then the result of the compare
488 // is equal to "Cond && TCmp". This also catches the case when the false
489 // value simplified to false and the true value to true, returning "Cond".
490 if (match(FCmp, m_Zero()))
491 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
493 // If the true value simplified to true, then the result of the compare
494 // is equal to "Cond || FCmp".
495 if (match(TCmp, m_One()))
496 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
498 // Finally, if the false value simplified to true and the true value to
499 // false, then the result of the compare is equal to "!Cond".
500 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
502 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
509 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
510 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
511 /// it on the incoming phi values yields the same result for every value. If so
512 /// returns the common value, otherwise returns null.
513 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
514 const Query &Q, unsigned MaxRecurse) {
515 // Recursion is always used, so bail out at once if we already hit the limit.
520 if (isa<PHINode>(LHS)) {
521 PI = cast<PHINode>(LHS);
522 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
523 if (!ValueDominatesPHI(RHS, PI, Q.DT))
526 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
527 PI = cast<PHINode>(RHS);
528 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
529 if (!ValueDominatesPHI(LHS, PI, Q.DT))
533 // Evaluate the BinOp on the incoming phi values.
534 Value *CommonValue = nullptr;
535 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
536 Value *Incoming = PI->getIncomingValue(i);
537 // If the incoming value is the phi node itself, it can safely be skipped.
538 if (Incoming == PI) continue;
539 Value *V = PI == LHS ?
540 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
541 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
542 // If the operation failed to simplify, or simplified to a different value
543 // to previously, then give up.
544 if (!V || (CommonValue && V != CommonValue))
552 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
553 /// try to simplify the comparison by seeing whether comparing with all of the
554 /// incoming phi values yields the same result every time. If so returns the
555 /// common result, otherwise returns null.
556 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
557 const Query &Q, unsigned MaxRecurse) {
558 // Recursion is always used, so bail out at once if we already hit the limit.
562 // Make sure the phi is on the LHS.
563 if (!isa<PHINode>(LHS)) {
565 Pred = CmpInst::getSwappedPredicate(Pred);
567 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
568 PHINode *PI = cast<PHINode>(LHS);
570 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
571 if (!ValueDominatesPHI(RHS, PI, Q.DT))
574 // Evaluate the BinOp on the incoming phi values.
575 Value *CommonValue = nullptr;
576 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
577 Value *Incoming = PI->getIncomingValue(i);
578 // If the incoming value is the phi node itself, it can safely be skipped.
579 if (Incoming == PI) continue;
580 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
581 // If the operation failed to simplify, or simplified to a different value
582 // to previously, then give up.
583 if (!V || (CommonValue && V != CommonValue))
591 /// SimplifyAddInst - Given operands for an Add, see if we can
592 /// fold the result. If not, this returns null.
593 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
594 const Query &Q, unsigned MaxRecurse) {
595 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
596 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
597 Constant *Ops[] = { CLHS, CRHS };
598 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
602 // Canonicalize the constant to the RHS.
606 // X + undef -> undef
607 if (match(Op1, m_Undef()))
611 if (match(Op1, m_Zero()))
618 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
619 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
622 // X + ~X -> -1 since ~X = -X-1
623 if (match(Op0, m_Not(m_Specific(Op1))) ||
624 match(Op1, m_Not(m_Specific(Op0))))
625 return Constant::getAllOnesValue(Op0->getType());
628 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
629 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
632 // Try some generic simplifications for associative operations.
633 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
637 // Mul distributes over Add. Try some generic simplifications based on this.
638 if (Value *V = FactorizeBinOp(Instruction::Add, Op0, Op1, Instruction::Mul,
642 // Threading Add over selects and phi nodes is pointless, so don't bother.
643 // Threading over the select in "A + select(cond, B, C)" means evaluating
644 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
645 // only if B and C are equal. If B and C are equal then (since we assume
646 // that operands have already been simplified) "select(cond, B, C)" should
647 // have been simplified to the common value of B and C already. Analysing
648 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
649 // for threading over phi nodes.
654 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
655 const DataLayout *DL, const TargetLibraryInfo *TLI,
656 const DominatorTree *DT) {
657 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
661 /// \brief Compute the base pointer and cumulative constant offsets for V.
663 /// This strips all constant offsets off of V, leaving it the base pointer, and
664 /// accumulates the total constant offset applied in the returned constant. It
665 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
666 /// no constant offsets applied.
668 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
669 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
671 static Constant *stripAndComputeConstantOffsets(const DataLayout *DL,
673 bool AllowNonInbounds = false) {
674 assert(V->getType()->getScalarType()->isPointerTy());
676 // Without DataLayout, just be conservative for now. Theoretically, more could
677 // be done in this case.
679 return ConstantInt::get(IntegerType::get(V->getContext(), 64), 0);
681 Type *IntPtrTy = DL->getIntPtrType(V->getType())->getScalarType();
682 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
684 // Even though we don't look through PHI nodes, we could be called on an
685 // instruction in an unreachable block, which may be on a cycle.
686 SmallPtrSet<Value *, 4> Visited;
689 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
690 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
691 !GEP->accumulateConstantOffset(*DL, Offset))
693 V = GEP->getPointerOperand();
694 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
695 V = cast<Operator>(V)->getOperand(0);
696 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
697 if (GA->mayBeOverridden())
699 V = GA->getAliasee();
703 assert(V->getType()->getScalarType()->isPointerTy() &&
704 "Unexpected operand type!");
705 } while (Visited.insert(V));
707 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
708 if (V->getType()->isVectorTy())
709 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
714 /// \brief Compute the constant difference between two pointer values.
715 /// If the difference is not a constant, returns zero.
716 static Constant *computePointerDifference(const DataLayout *DL,
717 Value *LHS, Value *RHS) {
718 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
719 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
721 // If LHS and RHS are not related via constant offsets to the same base
722 // value, there is nothing we can do here.
726 // Otherwise, the difference of LHS - RHS can be computed as:
728 // = (LHSOffset + Base) - (RHSOffset + Base)
729 // = LHSOffset - RHSOffset
730 return ConstantExpr::getSub(LHSOffset, RHSOffset);
733 /// SimplifySubInst - Given operands for a Sub, see if we can
734 /// fold the result. If not, this returns null.
735 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
736 const Query &Q, unsigned MaxRecurse) {
737 if (Constant *CLHS = dyn_cast<Constant>(Op0))
738 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
739 Constant *Ops[] = { CLHS, CRHS };
740 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
744 // X - undef -> undef
745 // undef - X -> undef
746 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
747 return UndefValue::get(Op0->getType());
750 if (match(Op1, m_Zero()))
755 return Constant::getNullValue(Op0->getType());
760 if (match(Op0, m_Mul(m_Specific(Op1), m_ConstantInt<2>())) ||
761 match(Op0, m_Shl(m_Specific(Op1), m_One())))
764 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
765 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
766 Value *Y = nullptr, *Z = Op1;
767 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
768 // See if "V === Y - Z" simplifies.
769 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
770 // It does! Now see if "X + V" simplifies.
771 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
772 // It does, we successfully reassociated!
776 // See if "V === X - Z" simplifies.
777 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
778 // It does! Now see if "Y + V" simplifies.
779 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
780 // It does, we successfully reassociated!
786 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
787 // For example, X - (X + 1) -> -1
789 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
790 // See if "V === X - Y" simplifies.
791 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
792 // It does! Now see if "V - Z" simplifies.
793 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
794 // It does, we successfully reassociated!
798 // See if "V === X - Z" simplifies.
799 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
800 // It does! Now see if "V - Y" simplifies.
801 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
802 // It does, we successfully reassociated!
808 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
809 // For example, X - (X - Y) -> Y.
811 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
812 // See if "V === Z - X" simplifies.
813 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
814 // It does! Now see if "V + Y" simplifies.
815 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
816 // It does, we successfully reassociated!
821 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
822 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
823 match(Op1, m_Trunc(m_Value(Y))))
824 if (X->getType() == Y->getType())
825 // See if "V === X - Y" simplifies.
826 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
827 // It does! Now see if "trunc V" simplifies.
828 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
829 // It does, return the simplified "trunc V".
832 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
833 if (match(Op0, m_PtrToInt(m_Value(X))) &&
834 match(Op1, m_PtrToInt(m_Value(Y))))
835 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
836 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
838 // Mul distributes over Sub. Try some generic simplifications based on this.
839 if (Value *V = FactorizeBinOp(Instruction::Sub, Op0, Op1, Instruction::Mul,
844 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
845 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
848 // Threading Sub over selects and phi nodes is pointless, so don't bother.
849 // Threading over the select in "A - select(cond, B, C)" means evaluating
850 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
851 // only if B and C are equal. If B and C are equal then (since we assume
852 // that operands have already been simplified) "select(cond, B, C)" should
853 // have been simplified to the common value of B and C already. Analysing
854 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
855 // for threading over phi nodes.
860 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
861 const DataLayout *DL, const TargetLibraryInfo *TLI,
862 const DominatorTree *DT) {
863 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
867 /// Given operands for an FAdd, see if we can fold the result. If not, this
869 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
870 const Query &Q, unsigned MaxRecurse) {
871 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
872 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
873 Constant *Ops[] = { CLHS, CRHS };
874 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(),
878 // Canonicalize the constant to the RHS.
883 if (match(Op1, m_NegZero()))
886 // fadd X, 0 ==> X, when we know X is not -0
887 if (match(Op1, m_Zero()) &&
888 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
891 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
892 // where nnan and ninf have to occur at least once somewhere in this
894 Value *SubOp = nullptr;
895 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
897 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
900 Instruction *FSub = cast<Instruction>(SubOp);
901 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
902 (FMF.noInfs() || FSub->hasNoInfs()))
903 return Constant::getNullValue(Op0->getType());
909 /// Given operands for an FSub, see if we can fold the result. If not, this
911 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
912 const Query &Q, unsigned MaxRecurse) {
913 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
914 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
915 Constant *Ops[] = { CLHS, CRHS };
916 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(),
922 if (match(Op1, m_Zero()))
925 // fsub X, -0 ==> X, when we know X is not -0
926 if (match(Op1, m_NegZero()) &&
927 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
930 // fsub 0, (fsub -0.0, X) ==> X
932 if (match(Op0, m_AnyZero())) {
933 if (match(Op1, m_FSub(m_NegZero(), m_Value(X))))
935 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
939 // fsub nnan ninf x, x ==> 0.0
940 if (FMF.noNaNs() && FMF.noInfs() && Op0 == Op1)
941 return Constant::getNullValue(Op0->getType());
946 /// Given the operands for an FMul, see if we can fold the result
947 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
950 unsigned MaxRecurse) {
951 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
952 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
953 Constant *Ops[] = { CLHS, CRHS };
954 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(),
958 // Canonicalize the constant to the RHS.
963 if (match(Op1, m_FPOne()))
966 // fmul nnan nsz X, 0 ==> 0
967 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
973 /// SimplifyMulInst - Given operands for a Mul, see if we can
974 /// fold the result. If not, this returns null.
975 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
976 unsigned MaxRecurse) {
977 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
978 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
979 Constant *Ops[] = { CLHS, CRHS };
980 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(),
984 // Canonicalize the constant to the RHS.
989 if (match(Op1, m_Undef()))
990 return Constant::getNullValue(Op0->getType());
993 if (match(Op1, m_Zero()))
997 if (match(Op1, m_One()))
1000 // (X / Y) * Y -> X if the division is exact.
1002 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
1003 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
1007 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
1008 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
1011 // Try some generic simplifications for associative operations.
1012 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
1016 // Mul distributes over Add. Try some generic simplifications based on this.
1017 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
1021 // If the operation is with the result of a select instruction, check whether
1022 // operating on either branch of the select always yields the same value.
1023 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1024 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
1028 // If the operation is with the result of a phi instruction, check whether
1029 // operating on all incoming values of the phi always yields the same value.
1030 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1031 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
1038 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1039 const DataLayout *DL, const TargetLibraryInfo *TLI,
1040 const DominatorTree *DT) {
1041 return ::SimplifyFAddInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
1044 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1045 const DataLayout *DL, const TargetLibraryInfo *TLI,
1046 const DominatorTree *DT) {
1047 return ::SimplifyFSubInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
1050 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1,
1052 const DataLayout *DL,
1053 const TargetLibraryInfo *TLI,
1054 const DominatorTree *DT) {
1055 return ::SimplifyFMulInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
1058 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout *DL,
1059 const TargetLibraryInfo *TLI,
1060 const DominatorTree *DT) {
1061 return ::SimplifyMulInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1064 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
1065 /// fold the result. If not, this returns null.
1066 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1067 const Query &Q, unsigned MaxRecurse) {
1068 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1069 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1070 Constant *Ops[] = { C0, C1 };
1071 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1075 bool isSigned = Opcode == Instruction::SDiv;
1077 // X / undef -> undef
1078 if (match(Op1, m_Undef()))
1082 if (match(Op0, m_Undef()))
1083 return Constant::getNullValue(Op0->getType());
1085 // 0 / X -> 0, we don't need to preserve faults!
1086 if (match(Op0, m_Zero()))
1090 if (match(Op1, m_One()))
1093 if (Op0->getType()->isIntegerTy(1))
1094 // It can't be division by zero, hence it must be division by one.
1099 return ConstantInt::get(Op0->getType(), 1);
1101 // (X * Y) / Y -> X if the multiplication does not overflow.
1102 Value *X = nullptr, *Y = nullptr;
1103 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1104 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1105 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1106 // If the Mul knows it does not overflow, then we are good to go.
1107 if ((isSigned && Mul->hasNoSignedWrap()) ||
1108 (!isSigned && Mul->hasNoUnsignedWrap()))
1110 // If X has the form X = A / Y then X * Y cannot overflow.
1111 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1112 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1116 // (X rem Y) / Y -> 0
1117 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1118 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1119 return Constant::getNullValue(Op0->getType());
1121 // If the operation is with the result of a select instruction, check whether
1122 // operating on either branch of the select always yields the same value.
1123 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1124 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1127 // If the operation is with the result of a phi instruction, check whether
1128 // operating on all incoming values of the phi always yields the same value.
1129 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1130 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1136 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1137 /// fold the result. If not, this returns null.
1138 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1139 unsigned MaxRecurse) {
1140 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1146 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1147 const TargetLibraryInfo *TLI,
1148 const DominatorTree *DT) {
1149 return ::SimplifySDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1152 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1153 /// fold the result. If not, this returns null.
1154 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1155 unsigned MaxRecurse) {
1156 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1162 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1163 const TargetLibraryInfo *TLI,
1164 const DominatorTree *DT) {
1165 return ::SimplifyUDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1168 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, const Query &Q,
1170 // undef / X -> undef (the undef could be a snan).
1171 if (match(Op0, m_Undef()))
1174 // X / undef -> undef
1175 if (match(Op1, m_Undef()))
1181 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1182 const TargetLibraryInfo *TLI,
1183 const DominatorTree *DT) {
1184 return ::SimplifyFDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1187 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1188 /// fold the result. If not, this returns null.
1189 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1190 const Query &Q, unsigned MaxRecurse) {
1191 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1192 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1193 Constant *Ops[] = { C0, C1 };
1194 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1198 // X % undef -> undef
1199 if (match(Op1, m_Undef()))
1203 if (match(Op0, m_Undef()))
1204 return Constant::getNullValue(Op0->getType());
1206 // 0 % X -> 0, we don't need to preserve faults!
1207 if (match(Op0, m_Zero()))
1210 // X % 0 -> undef, we don't need to preserve faults!
1211 if (match(Op1, m_Zero()))
1212 return UndefValue::get(Op0->getType());
1215 if (match(Op1, m_One()))
1216 return Constant::getNullValue(Op0->getType());
1218 if (Op0->getType()->isIntegerTy(1))
1219 // It can't be remainder by zero, hence it must be remainder by one.
1220 return Constant::getNullValue(Op0->getType());
1224 return Constant::getNullValue(Op0->getType());
1226 // If the operation is with the result of a select instruction, check whether
1227 // operating on either branch of the select always yields the same value.
1228 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1229 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1232 // If the operation is with the result of a phi instruction, check whether
1233 // operating on all incoming values of the phi always yields the same value.
1234 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1235 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1241 /// SimplifySRemInst - Given operands for an SRem, see if we can
1242 /// fold the result. If not, this returns null.
1243 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1244 unsigned MaxRecurse) {
1245 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1251 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1252 const TargetLibraryInfo *TLI,
1253 const DominatorTree *DT) {
1254 return ::SimplifySRemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1257 /// SimplifyURemInst - Given operands for a URem, see if we can
1258 /// fold the result. If not, this returns null.
1259 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1260 unsigned MaxRecurse) {
1261 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1267 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1268 const TargetLibraryInfo *TLI,
1269 const DominatorTree *DT) {
1270 return ::SimplifyURemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1273 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, const Query &,
1275 // undef % X -> undef (the undef could be a snan).
1276 if (match(Op0, m_Undef()))
1279 // X % undef -> undef
1280 if (match(Op1, m_Undef()))
1286 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1287 const TargetLibraryInfo *TLI,
1288 const DominatorTree *DT) {
1289 return ::SimplifyFRemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1292 /// isUndefShift - Returns true if a shift by \c Amount always yields undef.
1293 static bool isUndefShift(Value *Amount) {
1294 Constant *C = dyn_cast<Constant>(Amount);
1298 // X shift by undef -> undef because it may shift by the bitwidth.
1299 if (isa<UndefValue>(C))
1302 // Shifting by the bitwidth or more is undefined.
1303 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1304 if (CI->getValue().getLimitedValue() >=
1305 CI->getType()->getScalarSizeInBits())
1308 // If all lanes of a vector shift are undefined the whole shift is.
1309 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1310 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1311 if (!isUndefShift(C->getAggregateElement(I)))
1319 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1320 /// fold the result. If not, this returns null.
1321 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1322 const Query &Q, unsigned MaxRecurse) {
1323 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1324 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1325 Constant *Ops[] = { C0, C1 };
1326 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1330 // 0 shift by X -> 0
1331 if (match(Op0, m_Zero()))
1334 // X shift by 0 -> X
1335 if (match(Op1, m_Zero()))
1338 // Fold undefined shifts.
1339 if (isUndefShift(Op1))
1340 return UndefValue::get(Op0->getType());
1342 // If the operation is with the result of a select instruction, check whether
1343 // operating on either branch of the select always yields the same value.
1344 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1345 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1348 // If the operation is with the result of a phi instruction, check whether
1349 // operating on all incoming values of the phi always yields the same value.
1350 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1351 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1357 /// SimplifyShlInst - Given operands for an Shl, see if we can
1358 /// fold the result. If not, this returns null.
1359 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1360 const Query &Q, unsigned MaxRecurse) {
1361 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1365 if (match(Op0, m_Undef()))
1366 return Constant::getNullValue(Op0->getType());
1368 // (X >> A) << A -> X
1370 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1375 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1376 const DataLayout *DL, const TargetLibraryInfo *TLI,
1377 const DominatorTree *DT) {
1378 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
1382 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1383 /// fold the result. If not, this returns null.
1384 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1385 const Query &Q, unsigned MaxRecurse) {
1386 if (Value *V = SimplifyShift(Instruction::LShr, Op0, Op1, Q, MaxRecurse))
1391 return Constant::getNullValue(Op0->getType());
1394 if (match(Op0, m_Undef()))
1395 return Constant::getNullValue(Op0->getType());
1397 // (X << A) >> A -> X
1399 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1400 cast<OverflowingBinaryOperator>(Op0)->hasNoUnsignedWrap())
1406 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1407 const DataLayout *DL,
1408 const TargetLibraryInfo *TLI,
1409 const DominatorTree *DT) {
1410 return ::SimplifyLShrInst(Op0, Op1, isExact, Query (DL, TLI, DT),
1414 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1415 /// fold the result. If not, this returns null.
1416 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1417 const Query &Q, unsigned MaxRecurse) {
1418 if (Value *V = SimplifyShift(Instruction::AShr, Op0, Op1, Q, MaxRecurse))
1423 return Constant::getNullValue(Op0->getType());
1425 // all ones >>a X -> all ones
1426 if (match(Op0, m_AllOnes()))
1429 // undef >>a X -> all ones
1430 if (match(Op0, m_Undef()))
1431 return Constant::getAllOnesValue(Op0->getType());
1433 // (X << A) >> A -> X
1435 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1436 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap())
1442 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1443 const DataLayout *DL,
1444 const TargetLibraryInfo *TLI,
1445 const DominatorTree *DT) {
1446 return ::SimplifyAShrInst(Op0, Op1, isExact, Query (DL, TLI, DT),
1450 /// SimplifyAndInst - Given operands for an And, see if we can
1451 /// fold the result. If not, this returns null.
1452 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1453 unsigned MaxRecurse) {
1454 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1455 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1456 Constant *Ops[] = { CLHS, CRHS };
1457 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1461 // Canonicalize the constant to the RHS.
1462 std::swap(Op0, Op1);
1466 if (match(Op1, m_Undef()))
1467 return Constant::getNullValue(Op0->getType());
1474 if (match(Op1, m_Zero()))
1478 if (match(Op1, m_AllOnes()))
1481 // A & ~A = ~A & A = 0
1482 if (match(Op0, m_Not(m_Specific(Op1))) ||
1483 match(Op1, m_Not(m_Specific(Op0))))
1484 return Constant::getNullValue(Op0->getType());
1487 Value *A = nullptr, *B = nullptr;
1488 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1489 (A == Op1 || B == Op1))
1493 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1494 (A == Op0 || B == Op0))
1497 // A & (-A) = A if A is a power of two or zero.
1498 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1499 match(Op1, m_Neg(m_Specific(Op0)))) {
1500 if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/true))
1502 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true))
1506 // Try some generic simplifications for associative operations.
1507 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1511 // And distributes over Or. Try some generic simplifications based on this.
1512 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1516 // And distributes over Xor. Try some generic simplifications based on this.
1517 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1521 // Or distributes over And. Try some generic simplifications based on this.
1522 if (Value *V = FactorizeBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1526 // If the operation is with the result of a select instruction, check whether
1527 // operating on either branch of the select always yields the same value.
1528 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1529 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1533 // If the operation is with the result of a phi instruction, check whether
1534 // operating on all incoming values of the phi always yields the same value.
1535 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1536 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1543 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *DL,
1544 const TargetLibraryInfo *TLI,
1545 const DominatorTree *DT) {
1546 return ::SimplifyAndInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1549 /// SimplifyOrInst - Given operands for an Or, see if we can
1550 /// fold the result. If not, this returns null.
1551 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1552 unsigned MaxRecurse) {
1553 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1554 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1555 Constant *Ops[] = { CLHS, CRHS };
1556 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1560 // Canonicalize the constant to the RHS.
1561 std::swap(Op0, Op1);
1565 if (match(Op1, m_Undef()))
1566 return Constant::getAllOnesValue(Op0->getType());
1573 if (match(Op1, m_Zero()))
1577 if (match(Op1, m_AllOnes()))
1580 // A | ~A = ~A | A = -1
1581 if (match(Op0, m_Not(m_Specific(Op1))) ||
1582 match(Op1, m_Not(m_Specific(Op0))))
1583 return Constant::getAllOnesValue(Op0->getType());
1586 Value *A = nullptr, *B = nullptr;
1587 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1588 (A == Op1 || B == Op1))
1592 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1593 (A == Op0 || B == Op0))
1596 // ~(A & ?) | A = -1
1597 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1598 (A == Op1 || B == Op1))
1599 return Constant::getAllOnesValue(Op1->getType());
1601 // A | ~(A & ?) = -1
1602 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1603 (A == Op0 || B == Op0))
1604 return Constant::getAllOnesValue(Op0->getType());
1606 // Try some generic simplifications for associative operations.
1607 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1611 // Or distributes over And. Try some generic simplifications based on this.
1612 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1616 // And distributes over Or. Try some generic simplifications based on this.
1617 if (Value *V = FactorizeBinOp(Instruction::Or, Op0, Op1, Instruction::And,
1621 // If the operation is with the result of a select instruction, check whether
1622 // operating on either branch of the select always yields the same value.
1623 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1624 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1629 Value *C = nullptr, *D = nullptr;
1630 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1631 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1632 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1633 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1634 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1635 // (A & C1)|(B & C2)
1636 // If we have: ((V + N) & C1) | (V & C2)
1637 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1638 // replace with V+N.
1640 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1641 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1642 // Add commutes, try both ways.
1643 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1645 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1648 // Or commutes, try both ways.
1649 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1650 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1651 // Add commutes, try both ways.
1652 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1654 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1660 // If the operation is with the result of a phi instruction, check whether
1661 // operating on all incoming values of the phi always yields the same value.
1662 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1663 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1669 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *DL,
1670 const TargetLibraryInfo *TLI,
1671 const DominatorTree *DT) {
1672 return ::SimplifyOrInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1675 /// SimplifyXorInst - Given operands for a Xor, see if we can
1676 /// fold the result. If not, this returns null.
1677 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1678 unsigned MaxRecurse) {
1679 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1680 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1681 Constant *Ops[] = { CLHS, CRHS };
1682 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1686 // Canonicalize the constant to the RHS.
1687 std::swap(Op0, Op1);
1690 // A ^ undef -> undef
1691 if (match(Op1, m_Undef()))
1695 if (match(Op1, m_Zero()))
1700 return Constant::getNullValue(Op0->getType());
1702 // A ^ ~A = ~A ^ A = -1
1703 if (match(Op0, m_Not(m_Specific(Op1))) ||
1704 match(Op1, m_Not(m_Specific(Op0))))
1705 return Constant::getAllOnesValue(Op0->getType());
1707 // Try some generic simplifications for associative operations.
1708 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1712 // And distributes over Xor. Try some generic simplifications based on this.
1713 if (Value *V = FactorizeBinOp(Instruction::Xor, Op0, Op1, Instruction::And,
1717 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1718 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1719 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1720 // only if B and C are equal. If B and C are equal then (since we assume
1721 // that operands have already been simplified) "select(cond, B, C)" should
1722 // have been simplified to the common value of B and C already. Analysing
1723 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1724 // for threading over phi nodes.
1729 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *DL,
1730 const TargetLibraryInfo *TLI,
1731 const DominatorTree *DT) {
1732 return ::SimplifyXorInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1735 static Type *GetCompareTy(Value *Op) {
1736 return CmpInst::makeCmpResultType(Op->getType());
1739 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1740 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1741 /// otherwise return null. Helper function for analyzing max/min idioms.
1742 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1743 Value *LHS, Value *RHS) {
1744 SelectInst *SI = dyn_cast<SelectInst>(V);
1747 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1750 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1751 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1753 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1754 LHS == CmpRHS && RHS == CmpLHS)
1759 // A significant optimization not implemented here is assuming that alloca
1760 // addresses are not equal to incoming argument values. They don't *alias*,
1761 // as we say, but that doesn't mean they aren't equal, so we take a
1762 // conservative approach.
1764 // This is inspired in part by C++11 5.10p1:
1765 // "Two pointers of the same type compare equal if and only if they are both
1766 // null, both point to the same function, or both represent the same
1769 // This is pretty permissive.
1771 // It's also partly due to C11 6.5.9p6:
1772 // "Two pointers compare equal if and only if both are null pointers, both are
1773 // pointers to the same object (including a pointer to an object and a
1774 // subobject at its beginning) or function, both are pointers to one past the
1775 // last element of the same array object, or one is a pointer to one past the
1776 // end of one array object and the other is a pointer to the start of a
1777 // different array object that happens to immediately follow the first array
1778 // object in the address space.)
1780 // C11's version is more restrictive, however there's no reason why an argument
1781 // couldn't be a one-past-the-end value for a stack object in the caller and be
1782 // equal to the beginning of a stack object in the callee.
1784 // If the C and C++ standards are ever made sufficiently restrictive in this
1785 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1786 // this optimization.
1787 static Constant *computePointerICmp(const DataLayout *DL,
1788 const TargetLibraryInfo *TLI,
1789 CmpInst::Predicate Pred,
1790 Value *LHS, Value *RHS) {
1791 // First, skip past any trivial no-ops.
1792 LHS = LHS->stripPointerCasts();
1793 RHS = RHS->stripPointerCasts();
1795 // A non-null pointer is not equal to a null pointer.
1796 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) &&
1797 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1798 return ConstantInt::get(GetCompareTy(LHS),
1799 !CmpInst::isTrueWhenEqual(Pred));
1801 // We can only fold certain predicates on pointer comparisons.
1806 // Equality comaprisons are easy to fold.
1807 case CmpInst::ICMP_EQ:
1808 case CmpInst::ICMP_NE:
1811 // We can only handle unsigned relational comparisons because 'inbounds' on
1812 // a GEP only protects against unsigned wrapping.
1813 case CmpInst::ICMP_UGT:
1814 case CmpInst::ICMP_UGE:
1815 case CmpInst::ICMP_ULT:
1816 case CmpInst::ICMP_ULE:
1817 // However, we have to switch them to their signed variants to handle
1818 // negative indices from the base pointer.
1819 Pred = ICmpInst::getSignedPredicate(Pred);
1823 // Strip off any constant offsets so that we can reason about them.
1824 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
1825 // here and compare base addresses like AliasAnalysis does, however there are
1826 // numerous hazards. AliasAnalysis and its utilities rely on special rules
1827 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
1828 // doesn't need to guarantee pointer inequality when it says NoAlias.
1829 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
1830 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
1832 // If LHS and RHS are related via constant offsets to the same base
1833 // value, we can replace it with an icmp which just compares the offsets.
1835 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
1837 // Various optimizations for (in)equality comparisons.
1838 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
1839 // Different non-empty allocations that exist at the same time have
1840 // different addresses (if the program can tell). Global variables always
1841 // exist, so they always exist during the lifetime of each other and all
1842 // allocas. Two different allocas usually have different addresses...
1844 // However, if there's an @llvm.stackrestore dynamically in between two
1845 // allocas, they may have the same address. It's tempting to reduce the
1846 // scope of the problem by only looking at *static* allocas here. That would
1847 // cover the majority of allocas while significantly reducing the likelihood
1848 // of having an @llvm.stackrestore pop up in the middle. However, it's not
1849 // actually impossible for an @llvm.stackrestore to pop up in the middle of
1850 // an entry block. Also, if we have a block that's not attached to a
1851 // function, we can't tell if it's "static" under the current definition.
1852 // Theoretically, this problem could be fixed by creating a new kind of
1853 // instruction kind specifically for static allocas. Such a new instruction
1854 // could be required to be at the top of the entry block, thus preventing it
1855 // from being subject to a @llvm.stackrestore. Instcombine could even
1856 // convert regular allocas into these special allocas. It'd be nifty.
1857 // However, until then, this problem remains open.
1859 // So, we'll assume that two non-empty allocas have different addresses
1862 // With all that, if the offsets are within the bounds of their allocations
1863 // (and not one-past-the-end! so we can't use inbounds!), and their
1864 // allocations aren't the same, the pointers are not equal.
1866 // Note that it's not necessary to check for LHS being a global variable
1867 // address, due to canonicalization and constant folding.
1868 if (isa<AllocaInst>(LHS) &&
1869 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
1870 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
1871 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
1872 uint64_t LHSSize, RHSSize;
1873 if (LHSOffsetCI && RHSOffsetCI &&
1874 getObjectSize(LHS, LHSSize, DL, TLI) &&
1875 getObjectSize(RHS, RHSSize, DL, TLI)) {
1876 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
1877 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
1878 if (!LHSOffsetValue.isNegative() &&
1879 !RHSOffsetValue.isNegative() &&
1880 LHSOffsetValue.ult(LHSSize) &&
1881 RHSOffsetValue.ult(RHSSize)) {
1882 return ConstantInt::get(GetCompareTy(LHS),
1883 !CmpInst::isTrueWhenEqual(Pred));
1887 // Repeat the above check but this time without depending on DataLayout
1888 // or being able to compute a precise size.
1889 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
1890 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
1891 LHSOffset->isNullValue() &&
1892 RHSOffset->isNullValue())
1893 return ConstantInt::get(GetCompareTy(LHS),
1894 !CmpInst::isTrueWhenEqual(Pred));
1897 // Even if an non-inbounds GEP occurs along the path we can still optimize
1898 // equality comparisons concerning the result. We avoid walking the whole
1899 // chain again by starting where the last calls to
1900 // stripAndComputeConstantOffsets left off and accumulate the offsets.
1901 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
1902 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
1904 return ConstantExpr::getICmp(Pred,
1905 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
1906 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
1913 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
1914 /// fold the result. If not, this returns null.
1915 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
1916 const Query &Q, unsigned MaxRecurse) {
1917 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
1918 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
1920 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
1921 if (Constant *CRHS = dyn_cast<Constant>(RHS))
1922 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
1924 // If we have a constant, make sure it is on the RHS.
1925 std::swap(LHS, RHS);
1926 Pred = CmpInst::getSwappedPredicate(Pred);
1929 Type *ITy = GetCompareTy(LHS); // The return type.
1930 Type *OpTy = LHS->getType(); // The operand type.
1932 // icmp X, X -> true/false
1933 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
1934 // because X could be 0.
1935 if (LHS == RHS || isa<UndefValue>(RHS))
1936 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
1938 // Special case logic when the operands have i1 type.
1939 if (OpTy->getScalarType()->isIntegerTy(1)) {
1942 case ICmpInst::ICMP_EQ:
1944 if (match(RHS, m_One()))
1947 case ICmpInst::ICMP_NE:
1949 if (match(RHS, m_Zero()))
1952 case ICmpInst::ICMP_UGT:
1954 if (match(RHS, m_Zero()))
1957 case ICmpInst::ICMP_UGE:
1959 if (match(RHS, m_One()))
1962 case ICmpInst::ICMP_SLT:
1964 if (match(RHS, m_Zero()))
1967 case ICmpInst::ICMP_SLE:
1969 if (match(RHS, m_One()))
1975 // If we are comparing with zero then try hard since this is a common case.
1976 if (match(RHS, m_Zero())) {
1977 bool LHSKnownNonNegative, LHSKnownNegative;
1979 default: llvm_unreachable("Unknown ICmp predicate!");
1980 case ICmpInst::ICMP_ULT:
1981 return getFalse(ITy);
1982 case ICmpInst::ICMP_UGE:
1983 return getTrue(ITy);
1984 case ICmpInst::ICMP_EQ:
1985 case ICmpInst::ICMP_ULE:
1986 if (isKnownNonZero(LHS, Q.DL))
1987 return getFalse(ITy);
1989 case ICmpInst::ICMP_NE:
1990 case ICmpInst::ICMP_UGT:
1991 if (isKnownNonZero(LHS, Q.DL))
1992 return getTrue(ITy);
1994 case ICmpInst::ICMP_SLT:
1995 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1996 if (LHSKnownNegative)
1997 return getTrue(ITy);
1998 if (LHSKnownNonNegative)
1999 return getFalse(ITy);
2001 case ICmpInst::ICMP_SLE:
2002 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
2003 if (LHSKnownNegative)
2004 return getTrue(ITy);
2005 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL))
2006 return getFalse(ITy);
2008 case ICmpInst::ICMP_SGE:
2009 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
2010 if (LHSKnownNegative)
2011 return getFalse(ITy);
2012 if (LHSKnownNonNegative)
2013 return getTrue(ITy);
2015 case ICmpInst::ICMP_SGT:
2016 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
2017 if (LHSKnownNegative)
2018 return getFalse(ITy);
2019 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL))
2020 return getTrue(ITy);
2025 // See if we are doing a comparison with a constant integer.
2026 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2027 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2028 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
2029 if (RHS_CR.isEmptySet())
2030 return ConstantInt::getFalse(CI->getContext());
2031 if (RHS_CR.isFullSet())
2032 return ConstantInt::getTrue(CI->getContext());
2034 // Many binary operators with constant RHS have easy to compute constant
2035 // range. Use them to check whether the comparison is a tautology.
2036 unsigned Width = CI->getBitWidth();
2037 APInt Lower = APInt(Width, 0);
2038 APInt Upper = APInt(Width, 0);
2040 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
2041 // 'urem x, CI2' produces [0, CI2).
2042 Upper = CI2->getValue();
2043 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
2044 // 'srem x, CI2' produces (-|CI2|, |CI2|).
2045 Upper = CI2->getValue().abs();
2046 Lower = (-Upper) + 1;
2047 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
2048 // 'udiv CI2, x' produces [0, CI2].
2049 Upper = CI2->getValue() + 1;
2050 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
2051 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
2052 APInt NegOne = APInt::getAllOnesValue(Width);
2054 Upper = NegOne.udiv(CI2->getValue()) + 1;
2055 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) {
2056 // 'sdiv CI2, x' produces [-|CI2|, |CI2|].
2057 Upper = CI2->getValue().abs() + 1;
2058 Lower = (-Upper) + 1;
2059 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
2060 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2].
2061 APInt IntMin = APInt::getSignedMinValue(Width);
2062 APInt IntMax = APInt::getSignedMaxValue(Width);
2063 APInt Val = CI2->getValue().abs();
2064 if (!Val.isMinValue()) {
2065 Lower = IntMin.sdiv(Val);
2066 Upper = IntMax.sdiv(Val) + 1;
2068 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
2069 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
2070 APInt NegOne = APInt::getAllOnesValue(Width);
2071 if (CI2->getValue().ult(Width))
2072 Upper = NegOne.lshr(CI2->getValue()) + 1;
2073 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) {
2074 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2].
2075 unsigned ShiftAmount = Width - 1;
2076 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2077 ShiftAmount = CI2->getValue().countTrailingZeros();
2078 Lower = CI2->getValue().lshr(ShiftAmount);
2079 Upper = CI2->getValue() + 1;
2080 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
2081 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
2082 APInt IntMin = APInt::getSignedMinValue(Width);
2083 APInt IntMax = APInt::getSignedMaxValue(Width);
2084 if (CI2->getValue().ult(Width)) {
2085 Lower = IntMin.ashr(CI2->getValue());
2086 Upper = IntMax.ashr(CI2->getValue()) + 1;
2088 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) {
2089 unsigned ShiftAmount = Width - 1;
2090 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2091 ShiftAmount = CI2->getValue().countTrailingZeros();
2092 if (CI2->isNegative()) {
2093 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)]
2094 Lower = CI2->getValue();
2095 Upper = CI2->getValue().ashr(ShiftAmount) + 1;
2097 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2]
2098 Lower = CI2->getValue().ashr(ShiftAmount);
2099 Upper = CI2->getValue() + 1;
2101 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
2102 // 'or x, CI2' produces [CI2, UINT_MAX].
2103 Lower = CI2->getValue();
2104 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
2105 // 'and x, CI2' produces [0, CI2].
2106 Upper = CI2->getValue() + 1;
2108 if (Lower != Upper) {
2109 ConstantRange LHS_CR = ConstantRange(Lower, Upper);
2110 if (RHS_CR.contains(LHS_CR))
2111 return ConstantInt::getTrue(RHS->getContext());
2112 if (RHS_CR.inverse().contains(LHS_CR))
2113 return ConstantInt::getFalse(RHS->getContext());
2117 // Compare of cast, for example (zext X) != 0 -> X != 0
2118 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
2119 Instruction *LI = cast<CastInst>(LHS);
2120 Value *SrcOp = LI->getOperand(0);
2121 Type *SrcTy = SrcOp->getType();
2122 Type *DstTy = LI->getType();
2124 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
2125 // if the integer type is the same size as the pointer type.
2126 if (MaxRecurse && Q.DL && isa<PtrToIntInst>(LI) &&
2127 Q.DL->getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
2128 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2129 // Transfer the cast to the constant.
2130 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
2131 ConstantExpr::getIntToPtr(RHSC, SrcTy),
2134 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
2135 if (RI->getOperand(0)->getType() == SrcTy)
2136 // Compare without the cast.
2137 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2143 if (isa<ZExtInst>(LHS)) {
2144 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
2146 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
2147 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2148 // Compare X and Y. Note that signed predicates become unsigned.
2149 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2150 SrcOp, RI->getOperand(0), Q,
2154 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
2155 // too. If not, then try to deduce the result of the comparison.
2156 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2157 // Compute the constant that would happen if we truncated to SrcTy then
2158 // reextended to DstTy.
2159 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2160 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
2162 // If the re-extended constant didn't change then this is effectively
2163 // also a case of comparing two zero-extended values.
2164 if (RExt == CI && MaxRecurse)
2165 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2166 SrcOp, Trunc, Q, MaxRecurse-1))
2169 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
2170 // there. Use this to work out the result of the comparison.
2173 default: llvm_unreachable("Unknown ICmp predicate!");
2175 case ICmpInst::ICMP_EQ:
2176 case ICmpInst::ICMP_UGT:
2177 case ICmpInst::ICMP_UGE:
2178 return ConstantInt::getFalse(CI->getContext());
2180 case ICmpInst::ICMP_NE:
2181 case ICmpInst::ICMP_ULT:
2182 case ICmpInst::ICMP_ULE:
2183 return ConstantInt::getTrue(CI->getContext());
2185 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
2186 // is non-negative then LHS <s RHS.
2187 case ICmpInst::ICMP_SGT:
2188 case ICmpInst::ICMP_SGE:
2189 return CI->getValue().isNegative() ?
2190 ConstantInt::getTrue(CI->getContext()) :
2191 ConstantInt::getFalse(CI->getContext());
2193 case ICmpInst::ICMP_SLT:
2194 case ICmpInst::ICMP_SLE:
2195 return CI->getValue().isNegative() ?
2196 ConstantInt::getFalse(CI->getContext()) :
2197 ConstantInt::getTrue(CI->getContext());
2203 if (isa<SExtInst>(LHS)) {
2204 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
2206 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
2207 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2208 // Compare X and Y. Note that the predicate does not change.
2209 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2213 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
2214 // too. If not, then try to deduce the result of the comparison.
2215 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2216 // Compute the constant that would happen if we truncated to SrcTy then
2217 // reextended to DstTy.
2218 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2219 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
2221 // If the re-extended constant didn't change then this is effectively
2222 // also a case of comparing two sign-extended values.
2223 if (RExt == CI && MaxRecurse)
2224 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
2227 // Otherwise the upper bits of LHS are all equal, while RHS has varying
2228 // bits there. Use this to work out the result of the comparison.
2231 default: llvm_unreachable("Unknown ICmp predicate!");
2232 case ICmpInst::ICMP_EQ:
2233 return ConstantInt::getFalse(CI->getContext());
2234 case ICmpInst::ICMP_NE:
2235 return ConstantInt::getTrue(CI->getContext());
2237 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
2239 case ICmpInst::ICMP_SGT:
2240 case ICmpInst::ICMP_SGE:
2241 return CI->getValue().isNegative() ?
2242 ConstantInt::getTrue(CI->getContext()) :
2243 ConstantInt::getFalse(CI->getContext());
2244 case ICmpInst::ICMP_SLT:
2245 case ICmpInst::ICMP_SLE:
2246 return CI->getValue().isNegative() ?
2247 ConstantInt::getFalse(CI->getContext()) :
2248 ConstantInt::getTrue(CI->getContext());
2250 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2252 case ICmpInst::ICMP_UGT:
2253 case ICmpInst::ICMP_UGE:
2254 // Comparison is true iff the LHS <s 0.
2256 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2257 Constant::getNullValue(SrcTy),
2261 case ICmpInst::ICMP_ULT:
2262 case ICmpInst::ICMP_ULE:
2263 // Comparison is true iff the LHS >=s 0.
2265 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2266 Constant::getNullValue(SrcTy),
2276 // If a bit is known to be zero for A and known to be one for B,
2277 // then A and B cannot be equal.
2278 if (ICmpInst::isEquality(Pred)) {
2279 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2280 uint32_t BitWidth = CI->getBitWidth();
2281 APInt LHSKnownZero(BitWidth, 0);
2282 APInt LHSKnownOne(BitWidth, 0);
2283 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
2284 APInt RHSKnownZero(BitWidth, 0);
2285 APInt RHSKnownOne(BitWidth, 0);
2286 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
2287 if (((LHSKnownOne & RHSKnownZero) != 0) ||
2288 ((LHSKnownZero & RHSKnownOne) != 0))
2289 return (Pred == ICmpInst::ICMP_EQ)
2290 ? ConstantInt::getFalse(CI->getContext())
2291 : ConstantInt::getTrue(CI->getContext());
2295 // Special logic for binary operators.
2296 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2297 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2298 if (MaxRecurse && (LBO || RBO)) {
2299 // Analyze the case when either LHS or RHS is an add instruction.
2300 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2301 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2302 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2303 if (LBO && LBO->getOpcode() == Instruction::Add) {
2304 A = LBO->getOperand(0); B = LBO->getOperand(1);
2305 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2306 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2307 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2309 if (RBO && RBO->getOpcode() == Instruction::Add) {
2310 C = RBO->getOperand(0); D = RBO->getOperand(1);
2311 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2312 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2313 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2316 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2317 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2318 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2319 Constant::getNullValue(RHS->getType()),
2323 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2324 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2325 if (Value *V = SimplifyICmpInst(Pred,
2326 Constant::getNullValue(LHS->getType()),
2327 C == LHS ? D : C, Q, MaxRecurse-1))
2330 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2331 if (A && C && (A == C || A == D || B == C || B == D) &&
2332 NoLHSWrapProblem && NoRHSWrapProblem) {
2333 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2336 // C + B == C + D -> B == D
2339 } else if (A == D) {
2340 // D + B == C + D -> B == C
2343 } else if (B == C) {
2344 // A + C == C + D -> A == D
2349 // A + D == C + D -> A == C
2353 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2358 // 0 - (zext X) pred C
2359 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2360 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2361 if (RHSC->getValue().isStrictlyPositive()) {
2362 if (Pred == ICmpInst::ICMP_SLT)
2363 return ConstantInt::getTrue(RHSC->getContext());
2364 if (Pred == ICmpInst::ICMP_SGE)
2365 return ConstantInt::getFalse(RHSC->getContext());
2366 if (Pred == ICmpInst::ICMP_EQ)
2367 return ConstantInt::getFalse(RHSC->getContext());
2368 if (Pred == ICmpInst::ICMP_NE)
2369 return ConstantInt::getTrue(RHSC->getContext());
2371 if (RHSC->getValue().isNonNegative()) {
2372 if (Pred == ICmpInst::ICMP_SLE)
2373 return ConstantInt::getTrue(RHSC->getContext());
2374 if (Pred == ICmpInst::ICMP_SGT)
2375 return ConstantInt::getFalse(RHSC->getContext());
2380 // icmp pred (urem X, Y), Y
2381 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2382 bool KnownNonNegative, KnownNegative;
2386 case ICmpInst::ICMP_SGT:
2387 case ICmpInst::ICMP_SGE:
2388 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL);
2389 if (!KnownNonNegative)
2392 case ICmpInst::ICMP_EQ:
2393 case ICmpInst::ICMP_UGT:
2394 case ICmpInst::ICMP_UGE:
2395 return getFalse(ITy);
2396 case ICmpInst::ICMP_SLT:
2397 case ICmpInst::ICMP_SLE:
2398 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL);
2399 if (!KnownNonNegative)
2402 case ICmpInst::ICMP_NE:
2403 case ICmpInst::ICMP_ULT:
2404 case ICmpInst::ICMP_ULE:
2405 return getTrue(ITy);
2409 // icmp pred X, (urem Y, X)
2410 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2411 bool KnownNonNegative, KnownNegative;
2415 case ICmpInst::ICMP_SGT:
2416 case ICmpInst::ICMP_SGE:
2417 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL);
2418 if (!KnownNonNegative)
2421 case ICmpInst::ICMP_NE:
2422 case ICmpInst::ICMP_UGT:
2423 case ICmpInst::ICMP_UGE:
2424 return getTrue(ITy);
2425 case ICmpInst::ICMP_SLT:
2426 case ICmpInst::ICMP_SLE:
2427 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL);
2428 if (!KnownNonNegative)
2431 case ICmpInst::ICMP_EQ:
2432 case ICmpInst::ICMP_ULT:
2433 case ICmpInst::ICMP_ULE:
2434 return getFalse(ITy);
2439 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2440 // icmp pred (X /u Y), X
2441 if (Pred == ICmpInst::ICMP_UGT)
2442 return getFalse(ITy);
2443 if (Pred == ICmpInst::ICMP_ULE)
2444 return getTrue(ITy);
2447 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2448 LBO->getOperand(1) == RBO->getOperand(1)) {
2449 switch (LBO->getOpcode()) {
2451 case Instruction::UDiv:
2452 case Instruction::LShr:
2453 if (ICmpInst::isSigned(Pred))
2456 case Instruction::SDiv:
2457 case Instruction::AShr:
2458 if (!LBO->isExact() || !RBO->isExact())
2460 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2461 RBO->getOperand(0), Q, MaxRecurse-1))
2464 case Instruction::Shl: {
2465 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2466 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2469 if (!NSW && ICmpInst::isSigned(Pred))
2471 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2472 RBO->getOperand(0), Q, MaxRecurse-1))
2479 // Simplify comparisons involving max/min.
2481 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2482 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2484 // Signed variants on "max(a,b)>=a -> true".
2485 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2486 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2487 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2488 // We analyze this as smax(A, B) pred A.
2490 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2491 (A == LHS || B == LHS)) {
2492 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2493 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2494 // We analyze this as smax(A, B) swapped-pred A.
2495 P = CmpInst::getSwappedPredicate(Pred);
2496 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2497 (A == RHS || B == RHS)) {
2498 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2499 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2500 // We analyze this as smax(-A, -B) swapped-pred -A.
2501 // Note that we do not need to actually form -A or -B thanks to EqP.
2502 P = CmpInst::getSwappedPredicate(Pred);
2503 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2504 (A == LHS || B == LHS)) {
2505 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2506 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2507 // We analyze this as smax(-A, -B) pred -A.
2508 // Note that we do not need to actually form -A or -B thanks to EqP.
2511 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2512 // Cases correspond to "max(A, B) p A".
2516 case CmpInst::ICMP_EQ:
2517 case CmpInst::ICMP_SLE:
2518 // Equivalent to "A EqP B". This may be the same as the condition tested
2519 // in the max/min; if so, we can just return that.
2520 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2522 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2524 // Otherwise, see if "A EqP B" simplifies.
2526 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2529 case CmpInst::ICMP_NE:
2530 case CmpInst::ICMP_SGT: {
2531 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2532 // Equivalent to "A InvEqP B". This may be the same as the condition
2533 // tested in the max/min; if so, we can just return that.
2534 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2536 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2538 // Otherwise, see if "A InvEqP B" simplifies.
2540 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2544 case CmpInst::ICMP_SGE:
2546 return getTrue(ITy);
2547 case CmpInst::ICMP_SLT:
2549 return getFalse(ITy);
2553 // Unsigned variants on "max(a,b)>=a -> true".
2554 P = CmpInst::BAD_ICMP_PREDICATE;
2555 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2556 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2557 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2558 // We analyze this as umax(A, B) pred A.
2560 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2561 (A == LHS || B == LHS)) {
2562 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2563 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2564 // We analyze this as umax(A, B) swapped-pred A.
2565 P = CmpInst::getSwappedPredicate(Pred);
2566 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2567 (A == RHS || B == RHS)) {
2568 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2569 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2570 // We analyze this as umax(-A, -B) swapped-pred -A.
2571 // Note that we do not need to actually form -A or -B thanks to EqP.
2572 P = CmpInst::getSwappedPredicate(Pred);
2573 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2574 (A == LHS || B == LHS)) {
2575 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2576 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2577 // We analyze this as umax(-A, -B) pred -A.
2578 // Note that we do not need to actually form -A or -B thanks to EqP.
2581 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2582 // Cases correspond to "max(A, B) p A".
2586 case CmpInst::ICMP_EQ:
2587 case CmpInst::ICMP_ULE:
2588 // Equivalent to "A EqP B". This may be the same as the condition tested
2589 // in the max/min; if so, we can just return that.
2590 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2592 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2594 // Otherwise, see if "A EqP B" simplifies.
2596 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2599 case CmpInst::ICMP_NE:
2600 case CmpInst::ICMP_UGT: {
2601 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2602 // Equivalent to "A InvEqP B". This may be the same as the condition
2603 // tested in the max/min; if so, we can just return that.
2604 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2606 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2608 // Otherwise, see if "A InvEqP B" simplifies.
2610 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2614 case CmpInst::ICMP_UGE:
2616 return getTrue(ITy);
2617 case CmpInst::ICMP_ULT:
2619 return getFalse(ITy);
2623 // Variants on "max(x,y) >= min(x,z)".
2625 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2626 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2627 (A == C || A == D || B == C || B == D)) {
2628 // max(x, ?) pred min(x, ?).
2629 if (Pred == CmpInst::ICMP_SGE)
2631 return getTrue(ITy);
2632 if (Pred == CmpInst::ICMP_SLT)
2634 return getFalse(ITy);
2635 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2636 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2637 (A == C || A == D || B == C || B == D)) {
2638 // min(x, ?) pred max(x, ?).
2639 if (Pred == CmpInst::ICMP_SLE)
2641 return getTrue(ITy);
2642 if (Pred == CmpInst::ICMP_SGT)
2644 return getFalse(ITy);
2645 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2646 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2647 (A == C || A == D || B == C || B == D)) {
2648 // max(x, ?) pred min(x, ?).
2649 if (Pred == CmpInst::ICMP_UGE)
2651 return getTrue(ITy);
2652 if (Pred == CmpInst::ICMP_ULT)
2654 return getFalse(ITy);
2655 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2656 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2657 (A == C || A == D || B == C || B == D)) {
2658 // min(x, ?) pred max(x, ?).
2659 if (Pred == CmpInst::ICMP_ULE)
2661 return getTrue(ITy);
2662 if (Pred == CmpInst::ICMP_UGT)
2664 return getFalse(ITy);
2667 // Simplify comparisons of related pointers using a powerful, recursive
2668 // GEP-walk when we have target data available..
2669 if (LHS->getType()->isPointerTy())
2670 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS))
2673 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
2674 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
2675 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
2676 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
2677 (ICmpInst::isEquality(Pred) ||
2678 (GLHS->isInBounds() && GRHS->isInBounds() &&
2679 Pred == ICmpInst::getSignedPredicate(Pred)))) {
2680 // The bases are equal and the indices are constant. Build a constant
2681 // expression GEP with the same indices and a null base pointer to see
2682 // what constant folding can make out of it.
2683 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
2684 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
2685 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS);
2687 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
2688 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS);
2689 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
2694 // If the comparison is with the result of a select instruction, check whether
2695 // comparing with either branch of the select always yields the same value.
2696 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2697 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2700 // If the comparison is with the result of a phi instruction, check whether
2701 // doing the compare with each incoming phi value yields a common result.
2702 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2703 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2709 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2710 const DataLayout *DL,
2711 const TargetLibraryInfo *TLI,
2712 const DominatorTree *DT) {
2713 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
2717 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
2718 /// fold the result. If not, this returns null.
2719 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2720 const Query &Q, unsigned MaxRecurse) {
2721 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2722 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
2724 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2725 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2726 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
2728 // If we have a constant, make sure it is on the RHS.
2729 std::swap(LHS, RHS);
2730 Pred = CmpInst::getSwappedPredicate(Pred);
2733 // Fold trivial predicates.
2734 if (Pred == FCmpInst::FCMP_FALSE)
2735 return ConstantInt::get(GetCompareTy(LHS), 0);
2736 if (Pred == FCmpInst::FCMP_TRUE)
2737 return ConstantInt::get(GetCompareTy(LHS), 1);
2739 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef
2740 return UndefValue::get(GetCompareTy(LHS));
2742 // fcmp x,x -> true/false. Not all compares are foldable.
2744 if (CmpInst::isTrueWhenEqual(Pred))
2745 return ConstantInt::get(GetCompareTy(LHS), 1);
2746 if (CmpInst::isFalseWhenEqual(Pred))
2747 return ConstantInt::get(GetCompareTy(LHS), 0);
2750 // Handle fcmp with constant RHS
2751 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2752 // If the constant is a nan, see if we can fold the comparison based on it.
2753 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2754 if (CFP->getValueAPF().isNaN()) {
2755 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
2756 return ConstantInt::getFalse(CFP->getContext());
2757 assert(FCmpInst::isUnordered(Pred) &&
2758 "Comparison must be either ordered or unordered!");
2759 // True if unordered.
2760 return ConstantInt::getTrue(CFP->getContext());
2762 // Check whether the constant is an infinity.
2763 if (CFP->getValueAPF().isInfinity()) {
2764 if (CFP->getValueAPF().isNegative()) {
2766 case FCmpInst::FCMP_OLT:
2767 // No value is ordered and less than negative infinity.
2768 return ConstantInt::getFalse(CFP->getContext());
2769 case FCmpInst::FCMP_UGE:
2770 // All values are unordered with or at least negative infinity.
2771 return ConstantInt::getTrue(CFP->getContext());
2777 case FCmpInst::FCMP_OGT:
2778 // No value is ordered and greater than infinity.
2779 return ConstantInt::getFalse(CFP->getContext());
2780 case FCmpInst::FCMP_ULE:
2781 // All values are unordered with and at most infinity.
2782 return ConstantInt::getTrue(CFP->getContext());
2791 // If the comparison is with the result of a select instruction, check whether
2792 // comparing with either branch of the select always yields the same value.
2793 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2794 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2797 // If the comparison is with the result of a phi instruction, check whether
2798 // doing the compare with each incoming phi value yields a common result.
2799 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2800 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2806 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2807 const DataLayout *DL,
2808 const TargetLibraryInfo *TLI,
2809 const DominatorTree *DT) {
2810 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
2814 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
2815 /// the result. If not, this returns null.
2816 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
2817 Value *FalseVal, const Query &Q,
2818 unsigned MaxRecurse) {
2819 // select true, X, Y -> X
2820 // select false, X, Y -> Y
2821 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
2822 if (CB->isAllOnesValue())
2824 if (CB->isNullValue())
2828 // select C, X, X -> X
2829 if (TrueVal == FalseVal)
2832 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
2833 if (isa<Constant>(TrueVal))
2837 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
2839 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
2845 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
2846 const DataLayout *DL,
2847 const TargetLibraryInfo *TLI,
2848 const DominatorTree *DT) {
2849 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Query (DL, TLI, DT),
2853 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
2854 /// fold the result. If not, this returns null.
2855 static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) {
2856 // The type of the GEP pointer operand.
2857 PointerType *PtrTy = cast<PointerType>(Ops[0]->getType()->getScalarType());
2859 // getelementptr P -> P.
2860 if (Ops.size() == 1)
2863 if (isa<UndefValue>(Ops[0])) {
2864 // Compute the (pointer) type returned by the GEP instruction.
2865 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1));
2866 Type *GEPTy = PointerType::get(LastType, PtrTy->getAddressSpace());
2867 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
2868 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
2869 return UndefValue::get(GEPTy);
2872 if (Ops.size() == 2) {
2873 // getelementptr P, 0 -> P.
2874 if (match(Ops[1], m_Zero()))
2876 // getelementptr P, N -> P if P points to a type of zero size.
2878 Type *Ty = PtrTy->getElementType();
2879 if (Ty->isSized() && Q.DL->getTypeAllocSize(Ty) == 0)
2884 // Check to see if this is constant foldable.
2885 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2886 if (!isa<Constant>(Ops[i]))
2889 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1));
2892 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *DL,
2893 const TargetLibraryInfo *TLI,
2894 const DominatorTree *DT) {
2895 return ::SimplifyGEPInst(Ops, Query (DL, TLI, DT), RecursionLimit);
2898 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
2899 /// can fold the result. If not, this returns null.
2900 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
2901 ArrayRef<unsigned> Idxs, const Query &Q,
2903 if (Constant *CAgg = dyn_cast<Constant>(Agg))
2904 if (Constant *CVal = dyn_cast<Constant>(Val))
2905 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
2907 // insertvalue x, undef, n -> x
2908 if (match(Val, m_Undef()))
2911 // insertvalue x, (extractvalue y, n), n
2912 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
2913 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
2914 EV->getIndices() == Idxs) {
2915 // insertvalue undef, (extractvalue y, n), n -> y
2916 if (match(Agg, m_Undef()))
2917 return EV->getAggregateOperand();
2919 // insertvalue y, (extractvalue y, n), n -> y
2920 if (Agg == EV->getAggregateOperand())
2927 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
2928 ArrayRef<unsigned> Idxs,
2929 const DataLayout *DL,
2930 const TargetLibraryInfo *TLI,
2931 const DominatorTree *DT) {
2932 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query (DL, TLI, DT),
2936 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
2937 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
2938 // If all of the PHI's incoming values are the same then replace the PHI node
2939 // with the common value.
2940 Value *CommonValue = nullptr;
2941 bool HasUndefInput = false;
2942 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2943 Value *Incoming = PN->getIncomingValue(i);
2944 // If the incoming value is the phi node itself, it can safely be skipped.
2945 if (Incoming == PN) continue;
2946 if (isa<UndefValue>(Incoming)) {
2947 // Remember that we saw an undef value, but otherwise ignore them.
2948 HasUndefInput = true;
2951 if (CommonValue && Incoming != CommonValue)
2952 return nullptr; // Not the same, bail out.
2953 CommonValue = Incoming;
2956 // If CommonValue is null then all of the incoming values were either undef or
2957 // equal to the phi node itself.
2959 return UndefValue::get(PN->getType());
2961 // If we have a PHI node like phi(X, undef, X), where X is defined by some
2962 // instruction, we cannot return X as the result of the PHI node unless it
2963 // dominates the PHI block.
2965 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
2970 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
2971 if (Constant *C = dyn_cast<Constant>(Op))
2972 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI);
2977 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *DL,
2978 const TargetLibraryInfo *TLI,
2979 const DominatorTree *DT) {
2980 return ::SimplifyTruncInst(Op, Ty, Query (DL, TLI, DT), RecursionLimit);
2983 //=== Helper functions for higher up the class hierarchy.
2985 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
2986 /// fold the result. If not, this returns null.
2987 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
2988 const Query &Q, unsigned MaxRecurse) {
2990 case Instruction::Add:
2991 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2993 case Instruction::FAdd:
2994 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
2996 case Instruction::Sub:
2997 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2999 case Instruction::FSub:
3000 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3002 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
3003 case Instruction::FMul:
3004 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3005 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
3006 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
3007 case Instruction::FDiv: return SimplifyFDivInst(LHS, RHS, Q, MaxRecurse);
3008 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
3009 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
3010 case Instruction::FRem: return SimplifyFRemInst(LHS, RHS, Q, MaxRecurse);
3011 case Instruction::Shl:
3012 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3014 case Instruction::LShr:
3015 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3016 case Instruction::AShr:
3017 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3018 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
3019 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
3020 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
3022 if (Constant *CLHS = dyn_cast<Constant>(LHS))
3023 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
3024 Constant *COps[] = {CLHS, CRHS};
3025 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL,
3029 // If the operation is associative, try some generic simplifications.
3030 if (Instruction::isAssociative(Opcode))
3031 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
3034 // If the operation is with the result of a select instruction check whether
3035 // operating on either branch of the select always yields the same value.
3036 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3037 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
3040 // If the operation is with the result of a phi instruction, check whether
3041 // operating on all incoming values of the phi always yields the same value.
3042 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3043 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
3050 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3051 const DataLayout *DL, const TargetLibraryInfo *TLI,
3052 const DominatorTree *DT) {
3053 return ::SimplifyBinOp(Opcode, LHS, RHS, Query (DL, TLI, DT), RecursionLimit);
3056 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
3057 /// fold the result.
3058 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3059 const Query &Q, unsigned MaxRecurse) {
3060 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
3061 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3062 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3065 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3066 const DataLayout *DL, const TargetLibraryInfo *TLI,
3067 const DominatorTree *DT) {
3068 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
3072 static bool IsIdempotent(Intrinsic::ID ID) {
3074 default: return false;
3076 // Unary idempotent: f(f(x)) = f(x)
3077 case Intrinsic::fabs:
3078 case Intrinsic::floor:
3079 case Intrinsic::ceil:
3080 case Intrinsic::trunc:
3081 case Intrinsic::rint:
3082 case Intrinsic::nearbyint:
3083 case Intrinsic::round:
3088 template <typename IterTy>
3089 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd,
3090 const Query &Q, unsigned MaxRecurse) {
3091 // Perform idempotent optimizations
3092 if (!IsIdempotent(IID))
3096 if (std::distance(ArgBegin, ArgEnd) == 1)
3097 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
3098 if (II->getIntrinsicID() == IID)
3104 template <typename IterTy>
3105 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
3106 const Query &Q, unsigned MaxRecurse) {
3107 Type *Ty = V->getType();
3108 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
3109 Ty = PTy->getElementType();
3110 FunctionType *FTy = cast<FunctionType>(Ty);
3112 // call undef -> undef
3113 if (isa<UndefValue>(V))
3114 return UndefValue::get(FTy->getReturnType());
3116 Function *F = dyn_cast<Function>(V);
3120 if (unsigned IID = F->getIntrinsicID())
3122 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse))
3125 if (!canConstantFoldCallTo(F))
3128 SmallVector<Constant *, 4> ConstantArgs;
3129 ConstantArgs.reserve(ArgEnd - ArgBegin);
3130 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
3131 Constant *C = dyn_cast<Constant>(*I);
3134 ConstantArgs.push_back(C);
3137 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
3140 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
3141 User::op_iterator ArgEnd, const DataLayout *DL,
3142 const TargetLibraryInfo *TLI,
3143 const DominatorTree *DT) {
3144 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT),
3148 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
3149 const DataLayout *DL, const TargetLibraryInfo *TLI,
3150 const DominatorTree *DT) {
3151 return ::SimplifyCall(V, Args.begin(), Args.end(), Query(DL, TLI, DT),
3155 /// SimplifyInstruction - See if we can compute a simplified version of this
3156 /// instruction. If not, this returns null.
3157 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *DL,
3158 const TargetLibraryInfo *TLI,
3159 const DominatorTree *DT) {
3162 switch (I->getOpcode()) {
3164 Result = ConstantFoldInstruction(I, DL, TLI);
3166 case Instruction::FAdd:
3167 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
3168 I->getFastMathFlags(), DL, TLI, DT);
3170 case Instruction::Add:
3171 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
3172 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3173 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3176 case Instruction::FSub:
3177 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
3178 I->getFastMathFlags(), DL, TLI, DT);
3180 case Instruction::Sub:
3181 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
3182 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3183 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3186 case Instruction::FMul:
3187 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
3188 I->getFastMathFlags(), DL, TLI, DT);
3190 case Instruction::Mul:
3191 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3193 case Instruction::SDiv:
3194 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3196 case Instruction::UDiv:
3197 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3199 case Instruction::FDiv:
3200 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3202 case Instruction::SRem:
3203 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3205 case Instruction::URem:
3206 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3208 case Instruction::FRem:
3209 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3211 case Instruction::Shl:
3212 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
3213 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3214 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3217 case Instruction::LShr:
3218 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
3219 cast<BinaryOperator>(I)->isExact(),
3222 case Instruction::AShr:
3223 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
3224 cast<BinaryOperator>(I)->isExact(),
3227 case Instruction::And:
3228 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3230 case Instruction::Or:
3231 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3233 case Instruction::Xor:
3234 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3236 case Instruction::ICmp:
3237 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
3238 I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3240 case Instruction::FCmp:
3241 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
3242 I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3244 case Instruction::Select:
3245 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
3246 I->getOperand(2), DL, TLI, DT);
3248 case Instruction::GetElementPtr: {
3249 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
3250 Result = SimplifyGEPInst(Ops, DL, TLI, DT);
3253 case Instruction::InsertValue: {
3254 InsertValueInst *IV = cast<InsertValueInst>(I);
3255 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
3256 IV->getInsertedValueOperand(),
3257 IV->getIndices(), DL, TLI, DT);
3260 case Instruction::PHI:
3261 Result = SimplifyPHINode(cast<PHINode>(I), Query (DL, TLI, DT));
3263 case Instruction::Call: {
3264 CallSite CS(cast<CallInst>(I));
3265 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
3269 case Instruction::Trunc:
3270 Result = SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT);
3274 /// If called on unreachable code, the above logic may report that the
3275 /// instruction simplified to itself. Make life easier for users by
3276 /// detecting that case here, returning a safe value instead.
3277 return Result == I ? UndefValue::get(I->getType()) : Result;
3280 /// \brief Implementation of recursive simplification through an instructions
3283 /// This is the common implementation of the recursive simplification routines.
3284 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
3285 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
3286 /// instructions to process and attempt to simplify it using
3287 /// InstructionSimplify.
3289 /// This routine returns 'true' only when *it* simplifies something. The passed
3290 /// in simplified value does not count toward this.
3291 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
3292 const DataLayout *DL,
3293 const TargetLibraryInfo *TLI,
3294 const DominatorTree *DT) {
3295 bool Simplified = false;
3296 SmallSetVector<Instruction *, 8> Worklist;
3298 // If we have an explicit value to collapse to, do that round of the
3299 // simplification loop by hand initially.
3301 for (User *U : I->users())
3303 Worklist.insert(cast<Instruction>(U));
3305 // Replace the instruction with its simplified value.
3306 I->replaceAllUsesWith(SimpleV);
3308 // Gracefully handle edge cases where the instruction is not wired into any
3311 I->eraseFromParent();
3316 // Note that we must test the size on each iteration, the worklist can grow.
3317 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
3320 // See if this instruction simplifies.
3321 SimpleV = SimplifyInstruction(I, DL, TLI, DT);
3327 // Stash away all the uses of the old instruction so we can check them for
3328 // recursive simplifications after a RAUW. This is cheaper than checking all
3329 // uses of To on the recursive step in most cases.
3330 for (User *U : I->users())
3331 Worklist.insert(cast<Instruction>(U));
3333 // Replace the instruction with its simplified value.
3334 I->replaceAllUsesWith(SimpleV);
3336 // Gracefully handle edge cases where the instruction is not wired into any
3339 I->eraseFromParent();
3344 bool llvm::recursivelySimplifyInstruction(Instruction *I,
3345 const DataLayout *DL,
3346 const TargetLibraryInfo *TLI,
3347 const DominatorTree *DT) {
3348 return replaceAndRecursivelySimplifyImpl(I, nullptr, DL, TLI, DT);
3351 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
3352 const DataLayout *DL,
3353 const TargetLibraryInfo *TLI,
3354 const DominatorTree *DT) {
3355 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
3356 assert(SimpleV && "Must provide a simplified value.");
3357 return replaceAndRecursivelySimplifyImpl(I, SimpleV, DL, TLI, DT);