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 #define DEBUG_TYPE "instsimplify"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/Dominators.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/DataLayout.h"
29 #include "llvm/GlobalAlias.h"
30 #include "llvm/Operator.h"
31 #include "llvm/Support/ConstantRange.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Support/PatternMatch.h"
34 #include "llvm/Support/ValueHandle.h"
36 using namespace llvm::PatternMatch;
38 enum { RecursionLimit = 3 };
40 STATISTIC(NumExpand, "Number of expansions");
41 STATISTIC(NumFactor , "Number of factorizations");
42 STATISTIC(NumReassoc, "Number of reassociations");
46 const TargetLibraryInfo *TLI;
47 const DominatorTree *DT;
49 Query(const DataLayout *td, const TargetLibraryInfo *tli,
50 const DominatorTree *dt) : TD(td), TLI(tli), DT(dt) {}
53 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
54 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
56 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
58 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
59 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
60 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned);
62 /// getFalse - For a boolean type, or a vector of boolean type, return false, or
63 /// a vector with every element false, as appropriate for the type.
64 static Constant *getFalse(Type *Ty) {
65 assert(Ty->getScalarType()->isIntegerTy(1) &&
66 "Expected i1 type or a vector of i1!");
67 return Constant::getNullValue(Ty);
70 /// getTrue - For a boolean type, or a vector of boolean type, return true, or
71 /// a vector with every element true, as appropriate for the type.
72 static Constant *getTrue(Type *Ty) {
73 assert(Ty->getScalarType()->isIntegerTy(1) &&
74 "Expected i1 type or a vector of i1!");
75 return Constant::getAllOnesValue(Ty);
78 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
79 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
81 CmpInst *Cmp = dyn_cast<CmpInst>(V);
84 CmpInst::Predicate CPred = Cmp->getPredicate();
85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
86 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
92 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
93 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
94 Instruction *I = dyn_cast<Instruction>(V);
96 // Arguments and constants dominate all instructions.
99 // If we are processing instructions (and/or basic blocks) that have not been
100 // fully added to a function, the parent nodes may still be null. Simply
101 // return the conservative answer in these cases.
102 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
105 // If we have a DominatorTree then do a precise test.
107 if (!DT->isReachableFromEntry(P->getParent()))
109 if (!DT->isReachableFromEntry(I->getParent()))
111 return DT->dominates(I, P);
114 // Otherwise, if the instruction is in the entry block, and is not an invoke,
115 // then it obviously dominates all phi nodes.
116 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
123 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning
124 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
125 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
126 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
127 /// Returns the simplified value, or null if no simplification was performed.
128 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
129 unsigned OpcToExpand, const Query &Q,
130 unsigned MaxRecurse) {
131 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
132 // Recursion is always used, so bail out at once if we already hit the limit.
136 // Check whether the expression has the form "(A op' B) op C".
137 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
138 if (Op0->getOpcode() == OpcodeToExpand) {
139 // It does! Try turning it into "(A op C) op' (B op C)".
140 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
141 // Do "A op C" and "B op C" both simplify?
142 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
143 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
144 // They do! Return "L op' R" if it simplifies or is already available.
145 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
146 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
147 && L == B && R == A)) {
151 // Otherwise return "L op' R" if it simplifies.
152 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
159 // Check whether the expression has the form "A op (B op' C)".
160 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
161 if (Op1->getOpcode() == OpcodeToExpand) {
162 // It does! Try turning it into "(A op B) op' (A op C)".
163 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
164 // Do "A op B" and "A op C" both simplify?
165 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
166 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
167 // They do! Return "L op' R" if it simplifies or is already available.
168 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
169 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
170 && L == C && R == B)) {
174 // Otherwise return "L op' R" if it simplifies.
175 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
185 /// FactorizeBinOp - Simplify "LHS Opcode RHS" by factorizing out a common term
186 /// using the operation OpCodeToExtract. For example, when Opcode is Add and
187 /// OpCodeToExtract is Mul then this tries to turn "(A*B)+(A*C)" into "A*(B+C)".
188 /// Returns the simplified value, or null if no simplification was performed.
189 static Value *FactorizeBinOp(unsigned Opcode, Value *LHS, Value *RHS,
190 unsigned OpcToExtract, const Query &Q,
191 unsigned MaxRecurse) {
192 Instruction::BinaryOps OpcodeToExtract = (Instruction::BinaryOps)OpcToExtract;
193 // Recursion is always used, so bail out at once if we already hit the limit.
197 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
198 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
200 if (!Op0 || Op0->getOpcode() != OpcodeToExtract ||
201 !Op1 || Op1->getOpcode() != OpcodeToExtract)
204 // The expression has the form "(A op' B) op (C op' D)".
205 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1);
206 Value *C = Op1->getOperand(0), *D = Op1->getOperand(1);
208 // Use left distributivity, i.e. "X op' (Y op Z) = (X op' Y) op (X op' Z)".
209 // Does the instruction have the form "(A op' B) op (A op' D)" or, in the
210 // commutative case, "(A op' B) op (C op' A)"?
211 if (A == C || (Instruction::isCommutative(OpcodeToExtract) && A == D)) {
212 Value *DD = A == C ? D : C;
213 // Form "A op' (B op DD)" if it simplifies completely.
214 // Does "B op DD" simplify?
215 if (Value *V = SimplifyBinOp(Opcode, B, DD, Q, MaxRecurse)) {
216 // It does! Return "A op' V" if it simplifies or is already available.
217 // If V equals B then "A op' V" is just the LHS. If V equals DD then
218 // "A op' V" is just the RHS.
219 if (V == B || V == DD) {
221 return V == B ? LHS : RHS;
223 // Otherwise return "A op' V" if it simplifies.
224 if (Value *W = SimplifyBinOp(OpcodeToExtract, A, V, Q, MaxRecurse)) {
231 // Use right distributivity, i.e. "(X op Y) op' Z = (X op' Z) op (Y op' Z)".
232 // Does the instruction have the form "(A op' B) op (C op' B)" or, in the
233 // commutative case, "(A op' B) op (B op' D)"?
234 if (B == D || (Instruction::isCommutative(OpcodeToExtract) && B == C)) {
235 Value *CC = B == D ? C : D;
236 // Form "(A op CC) op' B" if it simplifies completely..
237 // Does "A op CC" simplify?
238 if (Value *V = SimplifyBinOp(Opcode, A, CC, Q, MaxRecurse)) {
239 // It does! Return "V op' B" if it simplifies or is already available.
240 // If V equals A then "V op' B" is just the LHS. If V equals CC then
241 // "V op' B" is just the RHS.
242 if (V == A || V == CC) {
244 return V == A ? LHS : RHS;
246 // Otherwise return "V op' B" if it simplifies.
247 if (Value *W = SimplifyBinOp(OpcodeToExtract, V, B, Q, MaxRecurse)) {
257 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
258 /// operations. Returns the simpler value, or null if none was found.
259 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
260 const Query &Q, unsigned MaxRecurse) {
261 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
262 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
264 // Recursion is always used, so bail out at once if we already hit the limit.
268 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
269 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
271 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
272 if (Op0 && Op0->getOpcode() == Opcode) {
273 Value *A = Op0->getOperand(0);
274 Value *B = Op0->getOperand(1);
277 // Does "B op C" simplify?
278 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
279 // It does! Return "A op V" if it simplifies or is already available.
280 // If V equals B then "A op V" is just the LHS.
281 if (V == B) return LHS;
282 // Otherwise return "A op V" if it simplifies.
283 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
290 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
291 if (Op1 && Op1->getOpcode() == Opcode) {
293 Value *B = Op1->getOperand(0);
294 Value *C = Op1->getOperand(1);
296 // Does "A op B" simplify?
297 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
298 // It does! Return "V op C" if it simplifies or is already available.
299 // If V equals B then "V op C" is just the RHS.
300 if (V == B) return RHS;
301 // Otherwise return "V op C" if it simplifies.
302 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
309 // The remaining transforms require commutativity as well as associativity.
310 if (!Instruction::isCommutative(Opcode))
313 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
314 if (Op0 && Op0->getOpcode() == Opcode) {
315 Value *A = Op0->getOperand(0);
316 Value *B = Op0->getOperand(1);
319 // Does "C op A" simplify?
320 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
321 // It does! Return "V op B" if it simplifies or is already available.
322 // If V equals A then "V op B" is just the LHS.
323 if (V == A) return LHS;
324 // Otherwise return "V op B" if it simplifies.
325 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
332 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
333 if (Op1 && Op1->getOpcode() == Opcode) {
335 Value *B = Op1->getOperand(0);
336 Value *C = Op1->getOperand(1);
338 // Does "C op A" simplify?
339 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
340 // It does! Return "B op V" if it simplifies or is already available.
341 // If V equals C then "B op V" is just the RHS.
342 if (V == C) return RHS;
343 // Otherwise return "B op V" if it simplifies.
344 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
354 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
355 /// instruction as an operand, try to simplify the binop by seeing whether
356 /// evaluating it on both branches of the select results in the same value.
357 /// Returns the common value if so, otherwise returns null.
358 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
359 const Query &Q, unsigned MaxRecurse) {
360 // Recursion is always used, so bail out at once if we already hit the limit.
365 if (isa<SelectInst>(LHS)) {
366 SI = cast<SelectInst>(LHS);
368 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
369 SI = cast<SelectInst>(RHS);
372 // Evaluate the BinOp on the true and false branches of the select.
376 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
377 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
379 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
380 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
383 // If they simplified to the same value, then return the common value.
384 // If they both failed to simplify then return null.
388 // If one branch simplified to undef, return the other one.
389 if (TV && isa<UndefValue>(TV))
391 if (FV && isa<UndefValue>(FV))
394 // If applying the operation did not change the true and false select values,
395 // then the result of the binop is the select itself.
396 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
399 // If one branch simplified and the other did not, and the simplified
400 // value is equal to the unsimplified one, return the simplified value.
401 // For example, select (cond, X, X & Z) & Z -> X & Z.
402 if ((FV && !TV) || (TV && !FV)) {
403 // Check that the simplified value has the form "X op Y" where "op" is the
404 // same as the original operation.
405 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
406 if (Simplified && Simplified->getOpcode() == Opcode) {
407 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
408 // We already know that "op" is the same as for the simplified value. See
409 // if the operands match too. If so, return the simplified value.
410 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
411 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
412 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
413 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
414 Simplified->getOperand(1) == UnsimplifiedRHS)
416 if (Simplified->isCommutative() &&
417 Simplified->getOperand(1) == UnsimplifiedLHS &&
418 Simplified->getOperand(0) == UnsimplifiedRHS)
426 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
427 /// try to simplify the comparison by seeing whether both branches of the select
428 /// result in the same value. Returns the common value if so, otherwise returns
430 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
431 Value *RHS, const Query &Q,
432 unsigned MaxRecurse) {
433 // Recursion is always used, so bail out at once if we already hit the limit.
437 // Make sure the select is on the LHS.
438 if (!isa<SelectInst>(LHS)) {
440 Pred = CmpInst::getSwappedPredicate(Pred);
442 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
443 SelectInst *SI = cast<SelectInst>(LHS);
444 Value *Cond = SI->getCondition();
445 Value *TV = SI->getTrueValue();
446 Value *FV = SI->getFalseValue();
448 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
449 // Does "cmp TV, RHS" simplify?
450 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
452 // It not only simplified, it simplified to the select condition. Replace
454 TCmp = getTrue(Cond->getType());
456 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
457 // condition then we can replace it with 'true'. Otherwise give up.
458 if (!isSameCompare(Cond, Pred, TV, RHS))
460 TCmp = getTrue(Cond->getType());
463 // Does "cmp FV, RHS" simplify?
464 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
466 // It not only simplified, it simplified to the select condition. Replace
468 FCmp = getFalse(Cond->getType());
470 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
471 // condition then we can replace it with 'false'. Otherwise give up.
472 if (!isSameCompare(Cond, Pred, FV, RHS))
474 FCmp = getFalse(Cond->getType());
477 // If both sides simplified to the same value, then use it as the result of
478 // the original comparison.
482 // The remaining cases only make sense if the select condition has the same
483 // type as the result of the comparison, so bail out if this is not so.
484 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
486 // If the false value simplified to false, then the result of the compare
487 // is equal to "Cond && TCmp". This also catches the case when the false
488 // value simplified to false and the true value to true, returning "Cond".
489 if (match(FCmp, m_Zero()))
490 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
492 // If the true value simplified to true, then the result of the compare
493 // is equal to "Cond || FCmp".
494 if (match(TCmp, m_One()))
495 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
497 // Finally, if the false value simplified to true and the true value to
498 // false, then the result of the compare is equal to "!Cond".
499 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
501 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
508 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
509 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
510 /// it on the incoming phi values yields the same result for every value. If so
511 /// returns the common value, otherwise returns null.
512 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
513 const Query &Q, unsigned MaxRecurse) {
514 // Recursion is always used, so bail out at once if we already hit the limit.
519 if (isa<PHINode>(LHS)) {
520 PI = cast<PHINode>(LHS);
521 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
522 if (!ValueDominatesPHI(RHS, PI, Q.DT))
525 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
526 PI = cast<PHINode>(RHS);
527 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
528 if (!ValueDominatesPHI(LHS, PI, Q.DT))
532 // Evaluate the BinOp on the incoming phi values.
533 Value *CommonValue = 0;
534 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
535 Value *Incoming = PI->getIncomingValue(i);
536 // If the incoming value is the phi node itself, it can safely be skipped.
537 if (Incoming == PI) continue;
538 Value *V = PI == LHS ?
539 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
540 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
541 // If the operation failed to simplify, or simplified to a different value
542 // to previously, then give up.
543 if (!V || (CommonValue && V != CommonValue))
551 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
552 /// try to simplify the comparison by seeing whether comparing with all of the
553 /// incoming phi values yields the same result every time. If so returns the
554 /// common result, otherwise returns null.
555 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
556 const Query &Q, unsigned MaxRecurse) {
557 // Recursion is always used, so bail out at once if we already hit the limit.
561 // Make sure the phi is on the LHS.
562 if (!isa<PHINode>(LHS)) {
564 Pred = CmpInst::getSwappedPredicate(Pred);
566 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
567 PHINode *PI = cast<PHINode>(LHS);
569 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
570 if (!ValueDominatesPHI(RHS, PI, Q.DT))
573 // Evaluate the BinOp on the incoming phi values.
574 Value *CommonValue = 0;
575 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
576 Value *Incoming = PI->getIncomingValue(i);
577 // If the incoming value is the phi node itself, it can safely be skipped.
578 if (Incoming == PI) continue;
579 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
580 // If the operation failed to simplify, or simplified to a different value
581 // to previously, then give up.
582 if (!V || (CommonValue && V != CommonValue))
590 /// SimplifyAddInst - Given operands for an Add, see if we can
591 /// fold the result. If not, this returns null.
592 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
593 const Query &Q, unsigned MaxRecurse) {
594 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
595 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
596 Constant *Ops[] = { CLHS, CRHS };
597 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
601 // Canonicalize the constant to the RHS.
605 // X + undef -> undef
606 if (match(Op1, m_Undef()))
610 if (match(Op1, m_Zero()))
617 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
618 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
621 // X + ~X -> -1 since ~X = -X-1
622 if (match(Op0, m_Not(m_Specific(Op1))) ||
623 match(Op1, m_Not(m_Specific(Op0))))
624 return Constant::getAllOnesValue(Op0->getType());
627 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
628 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
631 // Try some generic simplifications for associative operations.
632 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
636 // Mul distributes over Add. Try some generic simplifications based on this.
637 if (Value *V = FactorizeBinOp(Instruction::Add, Op0, Op1, Instruction::Mul,
641 // Threading Add over selects and phi nodes is pointless, so don't bother.
642 // Threading over the select in "A + select(cond, B, C)" means evaluating
643 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
644 // only if B and C are equal. If B and C are equal then (since we assume
645 // that operands have already been simplified) "select(cond, B, C)" should
646 // have been simplified to the common value of B and C already. Analysing
647 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
648 // for threading over phi nodes.
653 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
654 const DataLayout *TD, const TargetLibraryInfo *TLI,
655 const DominatorTree *DT) {
656 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query (TD, TLI, DT),
660 /// \brief Compute the base pointer and cumulative constant offsets for V.
662 /// This strips all constant offsets off of V, leaving it the base pointer, and
663 /// accumulates the total constant offset applied in the returned constant. It
664 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
665 /// no constant offsets applied.
666 static Constant *stripAndComputeConstantOffsets(const DataLayout &TD,
668 if (!V->getType()->isPointerTy())
671 unsigned IntPtrWidth = TD.getPointerSizeInBits();
672 APInt Offset = APInt::getNullValue(IntPtrWidth);
674 // Even though we don't look through PHI nodes, we could be called on an
675 // instruction in an unreachable block, which may be on a cycle.
676 SmallPtrSet<Value *, 4> Visited;
679 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
680 if (!GEP->isInBounds() || !GEP->accumulateConstantOffset(TD, Offset))
682 V = GEP->getPointerOperand();
683 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
684 V = cast<Operator>(V)->getOperand(0);
685 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
686 if (GA->mayBeOverridden())
688 V = GA->getAliasee();
692 assert(V->getType()->isPointerTy() && "Unexpected operand type!");
693 } while (Visited.insert(V));
695 Type *IntPtrTy = TD.getIntPtrType(V->getContext());
696 return ConstantInt::get(IntPtrTy, Offset);
699 /// \brief Compute the constant difference between two pointer values.
700 /// If the difference is not a constant, returns zero.
701 static Constant *computePointerDifference(const DataLayout &TD,
702 Value *LHS, Value *RHS) {
703 Constant *LHSOffset = stripAndComputeConstantOffsets(TD, LHS);
706 Constant *RHSOffset = stripAndComputeConstantOffsets(TD, RHS);
710 // If LHS and RHS are not related via constant offsets to the same base
711 // value, there is nothing we can do here.
715 // Otherwise, the difference of LHS - RHS can be computed as:
717 // = (LHSOffset + Base) - (RHSOffset + Base)
718 // = LHSOffset - RHSOffset
719 return ConstantExpr::getSub(LHSOffset, RHSOffset);
722 /// SimplifySubInst - Given operands for a Sub, see if we can
723 /// fold the result. If not, this returns null.
724 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
725 const Query &Q, unsigned MaxRecurse) {
726 if (Constant *CLHS = dyn_cast<Constant>(Op0))
727 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
728 Constant *Ops[] = { CLHS, CRHS };
729 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
733 // X - undef -> undef
734 // undef - X -> undef
735 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
736 return UndefValue::get(Op0->getType());
739 if (match(Op1, m_Zero()))
744 return Constant::getNullValue(Op0->getType());
749 if (match(Op0, m_Mul(m_Specific(Op1), m_ConstantInt<2>())) ||
750 match(Op0, m_Shl(m_Specific(Op1), m_One())))
753 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
754 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
755 Value *Y = 0, *Z = Op1;
756 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
757 // See if "V === Y - Z" simplifies.
758 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
759 // It does! Now see if "X + V" simplifies.
760 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
761 // It does, we successfully reassociated!
765 // See if "V === X - Z" simplifies.
766 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
767 // It does! Now see if "Y + V" simplifies.
768 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
769 // It does, we successfully reassociated!
775 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
776 // For example, X - (X + 1) -> -1
778 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
779 // See if "V === X - Y" simplifies.
780 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
781 // It does! Now see if "V - Z" simplifies.
782 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
783 // It does, we successfully reassociated!
787 // See if "V === X - Z" simplifies.
788 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
789 // It does! Now see if "V - Y" simplifies.
790 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
791 // It does, we successfully reassociated!
797 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
798 // For example, X - (X - Y) -> Y.
800 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
801 // See if "V === Z - X" simplifies.
802 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
803 // It does! Now see if "V + Y" simplifies.
804 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
805 // It does, we successfully reassociated!
810 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
811 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
812 match(Op1, m_Trunc(m_Value(Y))))
813 if (X->getType() == Y->getType())
814 // See if "V === X - Y" simplifies.
815 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
816 // It does! Now see if "trunc V" simplifies.
817 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
818 // It does, return the simplified "trunc V".
821 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
822 if (Q.TD && match(Op0, m_PtrToInt(m_Value(X))) &&
823 match(Op1, m_PtrToInt(m_Value(Y))))
824 if (Constant *Result = computePointerDifference(*Q.TD, X, Y))
825 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
827 // Mul distributes over Sub. Try some generic simplifications based on this.
828 if (Value *V = FactorizeBinOp(Instruction::Sub, Op0, Op1, Instruction::Mul,
833 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
834 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
837 // Threading Sub over selects and phi nodes is pointless, so don't bother.
838 // Threading over the select in "A - select(cond, B, C)" means evaluating
839 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
840 // only if B and C are equal. If B and C are equal then (since we assume
841 // that operands have already been simplified) "select(cond, B, C)" should
842 // have been simplified to the common value of B and C already. Analysing
843 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
844 // for threading over phi nodes.
849 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
850 const DataLayout *TD, const TargetLibraryInfo *TLI,
851 const DominatorTree *DT) {
852 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query (TD, TLI, DT),
856 /// Given the operands for an FMul, see if we can fold the result
857 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
860 unsigned MaxRecurse) {
861 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
862 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
863 Constant *Ops[] = { CLHS, CRHS };
864 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(),
869 // Check for some fast-math optimizations
871 if (FMF.noSignedZeros()) {
872 // fmul N S 0, x ==> 0
873 if (match(Op0, m_Zero()))
875 if (match(Op1, m_Zero()))
883 /// SimplifyMulInst - Given operands for a Mul, see if we can
884 /// fold the result. If not, this returns null.
885 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
886 unsigned MaxRecurse) {
887 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
888 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
889 Constant *Ops[] = { CLHS, CRHS };
890 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(),
894 // Canonicalize the constant to the RHS.
899 if (match(Op1, m_Undef()))
900 return Constant::getNullValue(Op0->getType());
903 if (match(Op1, m_Zero()))
907 if (match(Op1, m_One()))
910 // (X / Y) * Y -> X if the division is exact.
912 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
913 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
917 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
918 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
921 // Try some generic simplifications for associative operations.
922 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
926 // Mul distributes over Add. Try some generic simplifications based on this.
927 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
931 // If the operation is with the result of a select instruction, check whether
932 // operating on either branch of the select always yields the same value.
933 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
934 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
938 // If the operation is with the result of a phi instruction, check whether
939 // operating on all incoming values of the phi always yields the same value.
940 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
941 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
948 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1,
950 const DataLayout *TD,
951 const TargetLibraryInfo *TLI,
952 const DominatorTree *DT) {
953 return ::SimplifyFMulInst(Op0, Op1, FMF, Query (TD, TLI, DT), RecursionLimit);
956 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout *TD,
957 const TargetLibraryInfo *TLI,
958 const DominatorTree *DT) {
959 return ::SimplifyMulInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
962 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
963 /// fold the result. If not, this returns null.
964 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
965 const Query &Q, unsigned MaxRecurse) {
966 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
967 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
968 Constant *Ops[] = { C0, C1 };
969 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.TD, Q.TLI);
973 bool isSigned = Opcode == Instruction::SDiv;
975 // X / undef -> undef
976 if (match(Op1, m_Undef()))
980 if (match(Op0, m_Undef()))
981 return Constant::getNullValue(Op0->getType());
983 // 0 / X -> 0, we don't need to preserve faults!
984 if (match(Op0, m_Zero()))
988 if (match(Op1, m_One()))
991 if (Op0->getType()->isIntegerTy(1))
992 // It can't be division by zero, hence it must be division by one.
997 return ConstantInt::get(Op0->getType(), 1);
999 // (X * Y) / Y -> X if the multiplication does not overflow.
1000 Value *X = 0, *Y = 0;
1001 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1002 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1003 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1004 // If the Mul knows it does not overflow, then we are good to go.
1005 if ((isSigned && Mul->hasNoSignedWrap()) ||
1006 (!isSigned && Mul->hasNoUnsignedWrap()))
1008 // If X has the form X = A / Y then X * Y cannot overflow.
1009 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1010 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1014 // (X rem Y) / Y -> 0
1015 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1016 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1017 return Constant::getNullValue(Op0->getType());
1019 // If the operation is with the result of a select instruction, check whether
1020 // operating on either branch of the select always yields the same value.
1021 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1022 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1025 // If the operation is with the result of a phi instruction, check whether
1026 // operating on all incoming values of the phi always yields the same value.
1027 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1028 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1034 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1035 /// fold the result. If not, this returns null.
1036 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1037 unsigned MaxRecurse) {
1038 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1044 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout *TD,
1045 const TargetLibraryInfo *TLI,
1046 const DominatorTree *DT) {
1047 return ::SimplifySDivInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1050 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1051 /// fold the result. If not, this returns null.
1052 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1053 unsigned MaxRecurse) {
1054 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1060 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout *TD,
1061 const TargetLibraryInfo *TLI,
1062 const DominatorTree *DT) {
1063 return ::SimplifyUDivInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1066 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, const Query &Q,
1068 // undef / X -> undef (the undef could be a snan).
1069 if (match(Op0, m_Undef()))
1072 // X / undef -> undef
1073 if (match(Op1, m_Undef()))
1079 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, const DataLayout *TD,
1080 const TargetLibraryInfo *TLI,
1081 const DominatorTree *DT) {
1082 return ::SimplifyFDivInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1085 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1086 /// fold the result. If not, this returns null.
1087 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1088 const Query &Q, unsigned MaxRecurse) {
1089 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1090 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1091 Constant *Ops[] = { C0, C1 };
1092 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.TD, Q.TLI);
1096 // X % undef -> undef
1097 if (match(Op1, m_Undef()))
1101 if (match(Op0, m_Undef()))
1102 return Constant::getNullValue(Op0->getType());
1104 // 0 % X -> 0, we don't need to preserve faults!
1105 if (match(Op0, m_Zero()))
1108 // X % 0 -> undef, we don't need to preserve faults!
1109 if (match(Op1, m_Zero()))
1110 return UndefValue::get(Op0->getType());
1113 if (match(Op1, m_One()))
1114 return Constant::getNullValue(Op0->getType());
1116 if (Op0->getType()->isIntegerTy(1))
1117 // It can't be remainder by zero, hence it must be remainder by one.
1118 return Constant::getNullValue(Op0->getType());
1122 return Constant::getNullValue(Op0->getType());
1124 // If the operation is with the result of a select instruction, check whether
1125 // operating on either branch of the select always yields the same value.
1126 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1127 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1130 // If the operation is with the result of a phi instruction, check whether
1131 // operating on all incoming values of the phi always yields the same value.
1132 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1133 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1139 /// SimplifySRemInst - Given operands for an SRem, see if we can
1140 /// fold the result. If not, this returns null.
1141 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1142 unsigned MaxRecurse) {
1143 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1149 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *TD,
1150 const TargetLibraryInfo *TLI,
1151 const DominatorTree *DT) {
1152 return ::SimplifySRemInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1155 /// SimplifyURemInst - Given operands for a URem, see if we can
1156 /// fold the result. If not, this returns null.
1157 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1158 unsigned MaxRecurse) {
1159 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1165 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *TD,
1166 const TargetLibraryInfo *TLI,
1167 const DominatorTree *DT) {
1168 return ::SimplifyURemInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1171 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, const Query &,
1173 // undef % X -> undef (the undef could be a snan).
1174 if (match(Op0, m_Undef()))
1177 // X % undef -> undef
1178 if (match(Op1, m_Undef()))
1184 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, const DataLayout *TD,
1185 const TargetLibraryInfo *TLI,
1186 const DominatorTree *DT) {
1187 return ::SimplifyFRemInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1190 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1191 /// fold the result. If not, this returns null.
1192 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1193 const Query &Q, unsigned MaxRecurse) {
1194 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1195 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1196 Constant *Ops[] = { C0, C1 };
1197 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.TD, Q.TLI);
1201 // 0 shift by X -> 0
1202 if (match(Op0, m_Zero()))
1205 // X shift by 0 -> X
1206 if (match(Op1, m_Zero()))
1209 // X shift by undef -> undef because it may shift by the bitwidth.
1210 if (match(Op1, m_Undef()))
1213 // Shifting by the bitwidth or more is undefined.
1214 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1))
1215 if (CI->getValue().getLimitedValue() >=
1216 Op0->getType()->getScalarSizeInBits())
1217 return UndefValue::get(Op0->getType());
1219 // If the operation is with the result of a select instruction, check whether
1220 // operating on either branch of the select always yields the same value.
1221 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1222 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1225 // If the operation is with the result of a phi instruction, check whether
1226 // operating on all incoming values of the phi always yields the same value.
1227 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1228 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1234 /// SimplifyShlInst - Given operands for an Shl, see if we can
1235 /// fold the result. If not, this returns null.
1236 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1237 const Query &Q, unsigned MaxRecurse) {
1238 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1242 if (match(Op0, m_Undef()))
1243 return Constant::getNullValue(Op0->getType());
1245 // (X >> A) << A -> X
1247 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1252 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1253 const DataLayout *TD, const TargetLibraryInfo *TLI,
1254 const DominatorTree *DT) {
1255 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query (TD, TLI, DT),
1259 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1260 /// fold the result. If not, this returns null.
1261 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1262 const Query &Q, unsigned MaxRecurse) {
1263 if (Value *V = SimplifyShift(Instruction::LShr, Op0, Op1, Q, MaxRecurse))
1267 if (match(Op0, m_Undef()))
1268 return Constant::getNullValue(Op0->getType());
1270 // (X << A) >> A -> X
1272 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1273 cast<OverflowingBinaryOperator>(Op0)->hasNoUnsignedWrap())
1279 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1280 const DataLayout *TD,
1281 const TargetLibraryInfo *TLI,
1282 const DominatorTree *DT) {
1283 return ::SimplifyLShrInst(Op0, Op1, isExact, Query (TD, TLI, DT),
1287 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1288 /// fold the result. If not, this returns null.
1289 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1290 const Query &Q, unsigned MaxRecurse) {
1291 if (Value *V = SimplifyShift(Instruction::AShr, Op0, Op1, Q, MaxRecurse))
1294 // all ones >>a X -> all ones
1295 if (match(Op0, m_AllOnes()))
1298 // undef >>a X -> all ones
1299 if (match(Op0, m_Undef()))
1300 return Constant::getAllOnesValue(Op0->getType());
1302 // (X << A) >> A -> X
1304 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1305 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap())
1311 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1312 const DataLayout *TD,
1313 const TargetLibraryInfo *TLI,
1314 const DominatorTree *DT) {
1315 return ::SimplifyAShrInst(Op0, Op1, isExact, Query (TD, TLI, DT),
1319 /// SimplifyAndInst - Given operands for an And, see if we can
1320 /// fold the result. If not, this returns null.
1321 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1322 unsigned MaxRecurse) {
1323 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1324 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1325 Constant *Ops[] = { CLHS, CRHS };
1326 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1330 // Canonicalize the constant to the RHS.
1331 std::swap(Op0, Op1);
1335 if (match(Op1, m_Undef()))
1336 return Constant::getNullValue(Op0->getType());
1343 if (match(Op1, m_Zero()))
1347 if (match(Op1, m_AllOnes()))
1350 // A & ~A = ~A & A = 0
1351 if (match(Op0, m_Not(m_Specific(Op1))) ||
1352 match(Op1, m_Not(m_Specific(Op0))))
1353 return Constant::getNullValue(Op0->getType());
1356 Value *A = 0, *B = 0;
1357 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1358 (A == Op1 || B == Op1))
1362 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1363 (A == Op0 || B == Op0))
1366 // A & (-A) = A if A is a power of two or zero.
1367 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1368 match(Op1, m_Neg(m_Specific(Op0)))) {
1369 if (isPowerOfTwo(Op0, Q.TD, /*OrZero*/true))
1371 if (isPowerOfTwo(Op1, Q.TD, /*OrZero*/true))
1375 // Try some generic simplifications for associative operations.
1376 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1380 // And distributes over Or. Try some generic simplifications based on this.
1381 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1385 // And distributes over Xor. Try some generic simplifications based on this.
1386 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1390 // Or distributes over And. Try some generic simplifications based on this.
1391 if (Value *V = FactorizeBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1395 // If the operation is with the result of a select instruction, check whether
1396 // operating on either branch of the select always yields the same value.
1397 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1398 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1402 // If the operation is with the result of a phi instruction, check whether
1403 // operating on all incoming values of the phi always yields the same value.
1404 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1405 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1412 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *TD,
1413 const TargetLibraryInfo *TLI,
1414 const DominatorTree *DT) {
1415 return ::SimplifyAndInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1418 /// SimplifyOrInst - Given operands for an Or, see if we can
1419 /// fold the result. If not, this returns null.
1420 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1421 unsigned MaxRecurse) {
1422 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1423 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1424 Constant *Ops[] = { CLHS, CRHS };
1425 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1429 // Canonicalize the constant to the RHS.
1430 std::swap(Op0, Op1);
1434 if (match(Op1, m_Undef()))
1435 return Constant::getAllOnesValue(Op0->getType());
1442 if (match(Op1, m_Zero()))
1446 if (match(Op1, m_AllOnes()))
1449 // A | ~A = ~A | A = -1
1450 if (match(Op0, m_Not(m_Specific(Op1))) ||
1451 match(Op1, m_Not(m_Specific(Op0))))
1452 return Constant::getAllOnesValue(Op0->getType());
1455 Value *A = 0, *B = 0;
1456 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1457 (A == Op1 || B == Op1))
1461 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1462 (A == Op0 || B == Op0))
1465 // ~(A & ?) | A = -1
1466 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1467 (A == Op1 || B == Op1))
1468 return Constant::getAllOnesValue(Op1->getType());
1470 // A | ~(A & ?) = -1
1471 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1472 (A == Op0 || B == Op0))
1473 return Constant::getAllOnesValue(Op0->getType());
1475 // Try some generic simplifications for associative operations.
1476 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1480 // Or distributes over And. Try some generic simplifications based on this.
1481 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1485 // And distributes over Or. Try some generic simplifications based on this.
1486 if (Value *V = FactorizeBinOp(Instruction::Or, Op0, Op1, Instruction::And,
1490 // If the operation is with the result of a select instruction, check whether
1491 // operating on either branch of the select always yields the same value.
1492 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1493 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1497 // If the operation is with the result of a phi instruction, check whether
1498 // operating on all incoming values of the phi always yields the same value.
1499 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1500 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1506 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *TD,
1507 const TargetLibraryInfo *TLI,
1508 const DominatorTree *DT) {
1509 return ::SimplifyOrInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1512 /// SimplifyXorInst - Given operands for a Xor, see if we can
1513 /// fold the result. If not, this returns null.
1514 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1515 unsigned MaxRecurse) {
1516 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1517 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1518 Constant *Ops[] = { CLHS, CRHS };
1519 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1523 // Canonicalize the constant to the RHS.
1524 std::swap(Op0, Op1);
1527 // A ^ undef -> undef
1528 if (match(Op1, m_Undef()))
1532 if (match(Op1, m_Zero()))
1537 return Constant::getNullValue(Op0->getType());
1539 // A ^ ~A = ~A ^ A = -1
1540 if (match(Op0, m_Not(m_Specific(Op1))) ||
1541 match(Op1, m_Not(m_Specific(Op0))))
1542 return Constant::getAllOnesValue(Op0->getType());
1544 // Try some generic simplifications for associative operations.
1545 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1549 // And distributes over Xor. Try some generic simplifications based on this.
1550 if (Value *V = FactorizeBinOp(Instruction::Xor, Op0, Op1, Instruction::And,
1554 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1555 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1556 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1557 // only if B and C are equal. If B and C are equal then (since we assume
1558 // that operands have already been simplified) "select(cond, B, C)" should
1559 // have been simplified to the common value of B and C already. Analysing
1560 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1561 // for threading over phi nodes.
1566 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *TD,
1567 const TargetLibraryInfo *TLI,
1568 const DominatorTree *DT) {
1569 return ::SimplifyXorInst(Op0, Op1, Query (TD, TLI, DT), RecursionLimit);
1572 static Type *GetCompareTy(Value *Op) {
1573 return CmpInst::makeCmpResultType(Op->getType());
1576 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1577 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1578 /// otherwise return null. Helper function for analyzing max/min idioms.
1579 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1580 Value *LHS, Value *RHS) {
1581 SelectInst *SI = dyn_cast<SelectInst>(V);
1584 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1587 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1588 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1590 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1591 LHS == CmpRHS && RHS == CmpLHS)
1596 static Constant *computePointerICmp(const DataLayout &TD,
1597 CmpInst::Predicate Pred,
1598 Value *LHS, Value *RHS) {
1599 // We can only fold certain predicates on pointer comparisons.
1604 // Equality comaprisons are easy to fold.
1605 case CmpInst::ICMP_EQ:
1606 case CmpInst::ICMP_NE:
1609 // We can only handle unsigned relational comparisons because 'inbounds' on
1610 // a GEP only protects against unsigned wrapping.
1611 case CmpInst::ICMP_UGT:
1612 case CmpInst::ICMP_UGE:
1613 case CmpInst::ICMP_ULT:
1614 case CmpInst::ICMP_ULE:
1615 // However, we have to switch them to their signed variants to handle
1616 // negative indices from the base pointer.
1617 Pred = ICmpInst::getSignedPredicate(Pred);
1621 Constant *LHSOffset = stripAndComputeConstantOffsets(TD, LHS);
1624 Constant *RHSOffset = stripAndComputeConstantOffsets(TD, RHS);
1628 // If LHS and RHS are not related via constant offsets to the same base
1629 // value, there is nothing we can do here.
1633 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
1636 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
1637 /// fold the result. If not, this returns null.
1638 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
1639 const Query &Q, unsigned MaxRecurse) {
1640 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
1641 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
1643 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
1644 if (Constant *CRHS = dyn_cast<Constant>(RHS))
1645 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.TD, Q.TLI);
1647 // If we have a constant, make sure it is on the RHS.
1648 std::swap(LHS, RHS);
1649 Pred = CmpInst::getSwappedPredicate(Pred);
1652 Type *ITy = GetCompareTy(LHS); // The return type.
1653 Type *OpTy = LHS->getType(); // The operand type.
1655 // icmp X, X -> true/false
1656 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
1657 // because X could be 0.
1658 if (LHS == RHS || isa<UndefValue>(RHS))
1659 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
1661 // Special case logic when the operands have i1 type.
1662 if (OpTy->getScalarType()->isIntegerTy(1)) {
1665 case ICmpInst::ICMP_EQ:
1667 if (match(RHS, m_One()))
1670 case ICmpInst::ICMP_NE:
1672 if (match(RHS, m_Zero()))
1675 case ICmpInst::ICMP_UGT:
1677 if (match(RHS, m_Zero()))
1680 case ICmpInst::ICMP_UGE:
1682 if (match(RHS, m_One()))
1685 case ICmpInst::ICMP_SLT:
1687 if (match(RHS, m_Zero()))
1690 case ICmpInst::ICMP_SLE:
1692 if (match(RHS, m_One()))
1698 // icmp <object*>, <object*/null> - Different identified objects have
1699 // different addresses (unless null), and what's more the address of an
1700 // identified local is never equal to another argument (again, barring null).
1701 // Note that generalizing to the case where LHS is a global variable address
1702 // or null is pointless, since if both LHS and RHS are constants then we
1703 // already constant folded the compare, and if only one of them is then we
1704 // moved it to RHS already.
1705 Value *LHSPtr = LHS->stripPointerCasts();
1706 Value *RHSPtr = RHS->stripPointerCasts();
1707 if (LHSPtr == RHSPtr)
1708 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
1710 // Be more aggressive about stripping pointer adjustments when checking a
1711 // comparison of an alloca address to another object. We can rip off all
1712 // inbounds GEP operations, even if they are variable.
1713 LHSPtr = LHSPtr->stripInBoundsOffsets();
1714 if (llvm::isIdentifiedObject(LHSPtr)) {
1715 RHSPtr = RHSPtr->stripInBoundsOffsets();
1716 if (llvm::isKnownNonNull(LHSPtr) || llvm::isKnownNonNull(RHSPtr)) {
1717 // If both sides are different identified objects, they aren't equal
1718 // unless they're null.
1719 if (LHSPtr != RHSPtr && llvm::isIdentifiedObject(RHSPtr) &&
1720 Pred == CmpInst::ICMP_EQ)
1721 return ConstantInt::get(ITy, false);
1723 // A local identified object (alloca or noalias call) can't equal any
1724 // incoming argument, unless they're both null or they belong to
1725 // different functions. The latter happens during inlining.
1726 if (Instruction *LHSInst = dyn_cast<Instruction>(LHSPtr))
1727 if (Argument *RHSArg = dyn_cast<Argument>(RHSPtr))
1728 if (LHSInst->getParent()->getParent() == RHSArg->getParent() &&
1729 Pred == CmpInst::ICMP_EQ)
1730 return ConstantInt::get(ITy, false);
1733 // Assume that the constant null is on the right.
1734 if (llvm::isKnownNonNull(LHSPtr) && isa<ConstantPointerNull>(RHSPtr)) {
1735 if (Pred == CmpInst::ICMP_EQ)
1736 return ConstantInt::get(ITy, false);
1737 else if (Pred == CmpInst::ICMP_NE)
1738 return ConstantInt::get(ITy, true);
1740 } else if (Argument *LHSArg = dyn_cast<Argument>(LHSPtr)) {
1741 RHSPtr = RHSPtr->stripInBoundsOffsets();
1742 // An alloca can't be equal to an argument unless they come from separate
1743 // functions via inlining.
1744 if (AllocaInst *RHSInst = dyn_cast<AllocaInst>(RHSPtr)) {
1745 if (LHSArg->getParent() == RHSInst->getParent()->getParent()) {
1746 if (Pred == CmpInst::ICMP_EQ)
1747 return ConstantInt::get(ITy, false);
1748 else if (Pred == CmpInst::ICMP_NE)
1749 return ConstantInt::get(ITy, true);
1754 // If we are comparing with zero then try hard since this is a common case.
1755 if (match(RHS, m_Zero())) {
1756 bool LHSKnownNonNegative, LHSKnownNegative;
1758 default: llvm_unreachable("Unknown ICmp predicate!");
1759 case ICmpInst::ICMP_ULT:
1760 return getFalse(ITy);
1761 case ICmpInst::ICMP_UGE:
1762 return getTrue(ITy);
1763 case ICmpInst::ICMP_EQ:
1764 case ICmpInst::ICMP_ULE:
1765 if (isKnownNonZero(LHS, Q.TD))
1766 return getFalse(ITy);
1768 case ICmpInst::ICMP_NE:
1769 case ICmpInst::ICMP_UGT:
1770 if (isKnownNonZero(LHS, Q.TD))
1771 return getTrue(ITy);
1773 case ICmpInst::ICMP_SLT:
1774 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD);
1775 if (LHSKnownNegative)
1776 return getTrue(ITy);
1777 if (LHSKnownNonNegative)
1778 return getFalse(ITy);
1780 case ICmpInst::ICMP_SLE:
1781 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD);
1782 if (LHSKnownNegative)
1783 return getTrue(ITy);
1784 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.TD))
1785 return getFalse(ITy);
1787 case ICmpInst::ICMP_SGE:
1788 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD);
1789 if (LHSKnownNegative)
1790 return getFalse(ITy);
1791 if (LHSKnownNonNegative)
1792 return getTrue(ITy);
1794 case ICmpInst::ICMP_SGT:
1795 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.TD);
1796 if (LHSKnownNegative)
1797 return getFalse(ITy);
1798 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.TD))
1799 return getTrue(ITy);
1804 // See if we are doing a comparison with a constant integer.
1805 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1806 // Rule out tautological comparisons (eg., ult 0 or uge 0).
1807 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
1808 if (RHS_CR.isEmptySet())
1809 return ConstantInt::getFalse(CI->getContext());
1810 if (RHS_CR.isFullSet())
1811 return ConstantInt::getTrue(CI->getContext());
1813 // Many binary operators with constant RHS have easy to compute constant
1814 // range. Use them to check whether the comparison is a tautology.
1815 uint32_t Width = CI->getBitWidth();
1816 APInt Lower = APInt(Width, 0);
1817 APInt Upper = APInt(Width, 0);
1819 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
1820 // 'urem x, CI2' produces [0, CI2).
1821 Upper = CI2->getValue();
1822 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
1823 // 'srem x, CI2' produces (-|CI2|, |CI2|).
1824 Upper = CI2->getValue().abs();
1825 Lower = (-Upper) + 1;
1826 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
1827 // 'udiv CI2, x' produces [0, CI2].
1828 Upper = CI2->getValue() + 1;
1829 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
1830 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
1831 APInt NegOne = APInt::getAllOnesValue(Width);
1833 Upper = NegOne.udiv(CI2->getValue()) + 1;
1834 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
1835 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2].
1836 APInt IntMin = APInt::getSignedMinValue(Width);
1837 APInt IntMax = APInt::getSignedMaxValue(Width);
1838 APInt Val = CI2->getValue().abs();
1839 if (!Val.isMinValue()) {
1840 Lower = IntMin.sdiv(Val);
1841 Upper = IntMax.sdiv(Val) + 1;
1843 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
1844 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
1845 APInt NegOne = APInt::getAllOnesValue(Width);
1846 if (CI2->getValue().ult(Width))
1847 Upper = NegOne.lshr(CI2->getValue()) + 1;
1848 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
1849 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
1850 APInt IntMin = APInt::getSignedMinValue(Width);
1851 APInt IntMax = APInt::getSignedMaxValue(Width);
1852 if (CI2->getValue().ult(Width)) {
1853 Lower = IntMin.ashr(CI2->getValue());
1854 Upper = IntMax.ashr(CI2->getValue()) + 1;
1856 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
1857 // 'or x, CI2' produces [CI2, UINT_MAX].
1858 Lower = CI2->getValue();
1859 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
1860 // 'and x, CI2' produces [0, CI2].
1861 Upper = CI2->getValue() + 1;
1863 if (Lower != Upper) {
1864 ConstantRange LHS_CR = ConstantRange(Lower, Upper);
1865 if (RHS_CR.contains(LHS_CR))
1866 return ConstantInt::getTrue(RHS->getContext());
1867 if (RHS_CR.inverse().contains(LHS_CR))
1868 return ConstantInt::getFalse(RHS->getContext());
1872 // Compare of cast, for example (zext X) != 0 -> X != 0
1873 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
1874 Instruction *LI = cast<CastInst>(LHS);
1875 Value *SrcOp = LI->getOperand(0);
1876 Type *SrcTy = SrcOp->getType();
1877 Type *DstTy = LI->getType();
1879 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
1880 // if the integer type is the same size as the pointer type.
1881 if (MaxRecurse && Q.TD && isa<PtrToIntInst>(LI) &&
1882 Q.TD->getPointerSizeInBits() == DstTy->getPrimitiveSizeInBits()) {
1883 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1884 // Transfer the cast to the constant.
1885 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
1886 ConstantExpr::getIntToPtr(RHSC, SrcTy),
1889 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
1890 if (RI->getOperand(0)->getType() == SrcTy)
1891 // Compare without the cast.
1892 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
1898 if (isa<ZExtInst>(LHS)) {
1899 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
1901 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
1902 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
1903 // Compare X and Y. Note that signed predicates become unsigned.
1904 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
1905 SrcOp, RI->getOperand(0), Q,
1909 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
1910 // too. If not, then try to deduce the result of the comparison.
1911 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1912 // Compute the constant that would happen if we truncated to SrcTy then
1913 // reextended to DstTy.
1914 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
1915 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
1917 // If the re-extended constant didn't change then this is effectively
1918 // also a case of comparing two zero-extended values.
1919 if (RExt == CI && MaxRecurse)
1920 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
1921 SrcOp, Trunc, Q, MaxRecurse-1))
1924 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
1925 // there. Use this to work out the result of the comparison.
1928 default: llvm_unreachable("Unknown ICmp predicate!");
1930 case ICmpInst::ICMP_EQ:
1931 case ICmpInst::ICMP_UGT:
1932 case ICmpInst::ICMP_UGE:
1933 return ConstantInt::getFalse(CI->getContext());
1935 case ICmpInst::ICMP_NE:
1936 case ICmpInst::ICMP_ULT:
1937 case ICmpInst::ICMP_ULE:
1938 return ConstantInt::getTrue(CI->getContext());
1940 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
1941 // is non-negative then LHS <s RHS.
1942 case ICmpInst::ICMP_SGT:
1943 case ICmpInst::ICMP_SGE:
1944 return CI->getValue().isNegative() ?
1945 ConstantInt::getTrue(CI->getContext()) :
1946 ConstantInt::getFalse(CI->getContext());
1948 case ICmpInst::ICMP_SLT:
1949 case ICmpInst::ICMP_SLE:
1950 return CI->getValue().isNegative() ?
1951 ConstantInt::getFalse(CI->getContext()) :
1952 ConstantInt::getTrue(CI->getContext());
1958 if (isa<SExtInst>(LHS)) {
1959 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
1961 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
1962 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
1963 // Compare X and Y. Note that the predicate does not change.
1964 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
1968 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
1969 // too. If not, then try to deduce the result of the comparison.
1970 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1971 // Compute the constant that would happen if we truncated to SrcTy then
1972 // reextended to DstTy.
1973 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
1974 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
1976 // If the re-extended constant didn't change then this is effectively
1977 // also a case of comparing two sign-extended values.
1978 if (RExt == CI && MaxRecurse)
1979 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
1982 // Otherwise the upper bits of LHS are all equal, while RHS has varying
1983 // bits there. Use this to work out the result of the comparison.
1986 default: llvm_unreachable("Unknown ICmp predicate!");
1987 case ICmpInst::ICMP_EQ:
1988 return ConstantInt::getFalse(CI->getContext());
1989 case ICmpInst::ICMP_NE:
1990 return ConstantInt::getTrue(CI->getContext());
1992 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
1994 case ICmpInst::ICMP_SGT:
1995 case ICmpInst::ICMP_SGE:
1996 return CI->getValue().isNegative() ?
1997 ConstantInt::getTrue(CI->getContext()) :
1998 ConstantInt::getFalse(CI->getContext());
1999 case ICmpInst::ICMP_SLT:
2000 case ICmpInst::ICMP_SLE:
2001 return CI->getValue().isNegative() ?
2002 ConstantInt::getFalse(CI->getContext()) :
2003 ConstantInt::getTrue(CI->getContext());
2005 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2007 case ICmpInst::ICMP_UGT:
2008 case ICmpInst::ICMP_UGE:
2009 // Comparison is true iff the LHS <s 0.
2011 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2012 Constant::getNullValue(SrcTy),
2016 case ICmpInst::ICMP_ULT:
2017 case ICmpInst::ICMP_ULE:
2018 // Comparison is true iff the LHS >=s 0.
2020 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2021 Constant::getNullValue(SrcTy),
2031 // Special logic for binary operators.
2032 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2033 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2034 if (MaxRecurse && (LBO || RBO)) {
2035 // Analyze the case when either LHS or RHS is an add instruction.
2036 Value *A = 0, *B = 0, *C = 0, *D = 0;
2037 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2038 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2039 if (LBO && LBO->getOpcode() == Instruction::Add) {
2040 A = LBO->getOperand(0); B = LBO->getOperand(1);
2041 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2042 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2043 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2045 if (RBO && RBO->getOpcode() == Instruction::Add) {
2046 C = RBO->getOperand(0); D = RBO->getOperand(1);
2047 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2048 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2049 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2052 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2053 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2054 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2055 Constant::getNullValue(RHS->getType()),
2059 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2060 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2061 if (Value *V = SimplifyICmpInst(Pred,
2062 Constant::getNullValue(LHS->getType()),
2063 C == LHS ? D : C, Q, MaxRecurse-1))
2066 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2067 if (A && C && (A == C || A == D || B == C || B == D) &&
2068 NoLHSWrapProblem && NoRHSWrapProblem) {
2069 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2072 // C + B == C + D -> B == D
2075 } else if (A == D) {
2076 // D + B == C + D -> B == C
2079 } else if (B == C) {
2080 // A + C == C + D -> A == D
2085 // A + D == C + D -> A == C
2089 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2094 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2095 bool KnownNonNegative, KnownNegative;
2099 case ICmpInst::ICMP_SGT:
2100 case ICmpInst::ICMP_SGE:
2101 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD);
2102 if (!KnownNonNegative)
2105 case ICmpInst::ICMP_EQ:
2106 case ICmpInst::ICMP_UGT:
2107 case ICmpInst::ICMP_UGE:
2108 return getFalse(ITy);
2109 case ICmpInst::ICMP_SLT:
2110 case ICmpInst::ICMP_SLE:
2111 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.TD);
2112 if (!KnownNonNegative)
2115 case ICmpInst::ICMP_NE:
2116 case ICmpInst::ICMP_ULT:
2117 case ICmpInst::ICMP_ULE:
2118 return getTrue(ITy);
2121 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2122 bool KnownNonNegative, KnownNegative;
2126 case ICmpInst::ICMP_SGT:
2127 case ICmpInst::ICMP_SGE:
2128 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD);
2129 if (!KnownNonNegative)
2132 case ICmpInst::ICMP_NE:
2133 case ICmpInst::ICMP_UGT:
2134 case ICmpInst::ICMP_UGE:
2135 return getTrue(ITy);
2136 case ICmpInst::ICMP_SLT:
2137 case ICmpInst::ICMP_SLE:
2138 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.TD);
2139 if (!KnownNonNegative)
2142 case ICmpInst::ICMP_EQ:
2143 case ICmpInst::ICMP_ULT:
2144 case ICmpInst::ICMP_ULE:
2145 return getFalse(ITy);
2150 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2151 // icmp pred (X /u Y), X
2152 if (Pred == ICmpInst::ICMP_UGT)
2153 return getFalse(ITy);
2154 if (Pred == ICmpInst::ICMP_ULE)
2155 return getTrue(ITy);
2158 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2159 LBO->getOperand(1) == RBO->getOperand(1)) {
2160 switch (LBO->getOpcode()) {
2162 case Instruction::UDiv:
2163 case Instruction::LShr:
2164 if (ICmpInst::isSigned(Pred))
2167 case Instruction::SDiv:
2168 case Instruction::AShr:
2169 if (!LBO->isExact() || !RBO->isExact())
2171 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2172 RBO->getOperand(0), Q, MaxRecurse-1))
2175 case Instruction::Shl: {
2176 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2177 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2180 if (!NSW && ICmpInst::isSigned(Pred))
2182 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2183 RBO->getOperand(0), Q, MaxRecurse-1))
2190 // Simplify comparisons involving max/min.
2192 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2193 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2195 // Signed variants on "max(a,b)>=a -> true".
2196 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2197 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2198 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2199 // We analyze this as smax(A, B) pred A.
2201 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2202 (A == LHS || B == LHS)) {
2203 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2204 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2205 // We analyze this as smax(A, B) swapped-pred A.
2206 P = CmpInst::getSwappedPredicate(Pred);
2207 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2208 (A == RHS || B == RHS)) {
2209 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2210 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2211 // We analyze this as smax(-A, -B) swapped-pred -A.
2212 // Note that we do not need to actually form -A or -B thanks to EqP.
2213 P = CmpInst::getSwappedPredicate(Pred);
2214 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2215 (A == LHS || B == LHS)) {
2216 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2217 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2218 // We analyze this as smax(-A, -B) pred -A.
2219 // Note that we do not need to actually form -A or -B thanks to EqP.
2222 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2223 // Cases correspond to "max(A, B) p A".
2227 case CmpInst::ICMP_EQ:
2228 case CmpInst::ICMP_SLE:
2229 // Equivalent to "A EqP B". This may be the same as the condition tested
2230 // in the max/min; if so, we can just return that.
2231 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2233 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2235 // Otherwise, see if "A EqP B" simplifies.
2237 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2240 case CmpInst::ICMP_NE:
2241 case CmpInst::ICMP_SGT: {
2242 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2243 // Equivalent to "A InvEqP B". This may be the same as the condition
2244 // tested in the max/min; if so, we can just return that.
2245 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2247 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2249 // Otherwise, see if "A InvEqP B" simplifies.
2251 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2255 case CmpInst::ICMP_SGE:
2257 return getTrue(ITy);
2258 case CmpInst::ICMP_SLT:
2260 return getFalse(ITy);
2264 // Unsigned variants on "max(a,b)>=a -> true".
2265 P = CmpInst::BAD_ICMP_PREDICATE;
2266 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2267 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2268 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2269 // We analyze this as umax(A, B) pred A.
2271 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2272 (A == LHS || B == LHS)) {
2273 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2274 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2275 // We analyze this as umax(A, B) swapped-pred A.
2276 P = CmpInst::getSwappedPredicate(Pred);
2277 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2278 (A == RHS || B == RHS)) {
2279 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2280 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2281 // We analyze this as umax(-A, -B) swapped-pred -A.
2282 // Note that we do not need to actually form -A or -B thanks to EqP.
2283 P = CmpInst::getSwappedPredicate(Pred);
2284 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2285 (A == LHS || B == LHS)) {
2286 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2287 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2288 // We analyze this as umax(-A, -B) pred -A.
2289 // Note that we do not need to actually form -A or -B thanks to EqP.
2292 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2293 // Cases correspond to "max(A, B) p A".
2297 case CmpInst::ICMP_EQ:
2298 case CmpInst::ICMP_ULE:
2299 // Equivalent to "A EqP B". This may be the same as the condition tested
2300 // in the max/min; if so, we can just return that.
2301 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2303 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2305 // Otherwise, see if "A EqP B" simplifies.
2307 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2310 case CmpInst::ICMP_NE:
2311 case CmpInst::ICMP_UGT: {
2312 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2313 // Equivalent to "A InvEqP B". This may be the same as the condition
2314 // tested in the max/min; if so, we can just return that.
2315 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2317 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2319 // Otherwise, see if "A InvEqP B" simplifies.
2321 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2325 case CmpInst::ICMP_UGE:
2327 return getTrue(ITy);
2328 case CmpInst::ICMP_ULT:
2330 return getFalse(ITy);
2334 // Variants on "max(x,y) >= min(x,z)".
2336 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2337 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2338 (A == C || A == D || B == C || B == D)) {
2339 // max(x, ?) pred min(x, ?).
2340 if (Pred == CmpInst::ICMP_SGE)
2342 return getTrue(ITy);
2343 if (Pred == CmpInst::ICMP_SLT)
2345 return getFalse(ITy);
2346 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2347 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2348 (A == C || A == D || B == C || B == D)) {
2349 // min(x, ?) pred max(x, ?).
2350 if (Pred == CmpInst::ICMP_SLE)
2352 return getTrue(ITy);
2353 if (Pred == CmpInst::ICMP_SGT)
2355 return getFalse(ITy);
2356 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2357 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2358 (A == C || A == D || B == C || B == D)) {
2359 // max(x, ?) pred min(x, ?).
2360 if (Pred == CmpInst::ICMP_UGE)
2362 return getTrue(ITy);
2363 if (Pred == CmpInst::ICMP_ULT)
2365 return getFalse(ITy);
2366 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2367 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2368 (A == C || A == D || B == C || B == D)) {
2369 // min(x, ?) pred max(x, ?).
2370 if (Pred == CmpInst::ICMP_ULE)
2372 return getTrue(ITy);
2373 if (Pred == CmpInst::ICMP_UGT)
2375 return getFalse(ITy);
2378 // Simplify comparisons of related pointers using a powerful, recursive
2379 // GEP-walk when we have target data available..
2380 if (Q.TD && LHS->getType()->isPointerTy() && RHS->getType()->isPointerTy())
2381 if (Constant *C = computePointerICmp(*Q.TD, Pred, LHS, RHS))
2384 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
2385 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
2386 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
2387 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
2388 (ICmpInst::isEquality(Pred) ||
2389 (GLHS->isInBounds() && GRHS->isInBounds() &&
2390 Pred == ICmpInst::getSignedPredicate(Pred)))) {
2391 // The bases are equal and the indices are constant. Build a constant
2392 // expression GEP with the same indices and a null base pointer to see
2393 // what constant folding can make out of it.
2394 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
2395 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
2396 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS);
2398 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
2399 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS);
2400 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
2405 // If the comparison is with the result of a select instruction, check whether
2406 // comparing with either branch of the select always yields the same value.
2407 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2408 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2411 // If the comparison is with the result of a phi instruction, check whether
2412 // doing the compare with each incoming phi value yields a common result.
2413 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2414 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2420 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2421 const DataLayout *TD,
2422 const TargetLibraryInfo *TLI,
2423 const DominatorTree *DT) {
2424 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query (TD, TLI, DT),
2428 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
2429 /// fold the result. If not, this returns null.
2430 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2431 const Query &Q, unsigned MaxRecurse) {
2432 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2433 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
2435 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2436 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2437 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.TD, Q.TLI);
2439 // If we have a constant, make sure it is on the RHS.
2440 std::swap(LHS, RHS);
2441 Pred = CmpInst::getSwappedPredicate(Pred);
2444 // Fold trivial predicates.
2445 if (Pred == FCmpInst::FCMP_FALSE)
2446 return ConstantInt::get(GetCompareTy(LHS), 0);
2447 if (Pred == FCmpInst::FCMP_TRUE)
2448 return ConstantInt::get(GetCompareTy(LHS), 1);
2450 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef
2451 return UndefValue::get(GetCompareTy(LHS));
2453 // fcmp x,x -> true/false. Not all compares are foldable.
2455 if (CmpInst::isTrueWhenEqual(Pred))
2456 return ConstantInt::get(GetCompareTy(LHS), 1);
2457 if (CmpInst::isFalseWhenEqual(Pred))
2458 return ConstantInt::get(GetCompareTy(LHS), 0);
2461 // Handle fcmp with constant RHS
2462 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2463 // If the constant is a nan, see if we can fold the comparison based on it.
2464 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2465 if (CFP->getValueAPF().isNaN()) {
2466 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
2467 return ConstantInt::getFalse(CFP->getContext());
2468 assert(FCmpInst::isUnordered(Pred) &&
2469 "Comparison must be either ordered or unordered!");
2470 // True if unordered.
2471 return ConstantInt::getTrue(CFP->getContext());
2473 // Check whether the constant is an infinity.
2474 if (CFP->getValueAPF().isInfinity()) {
2475 if (CFP->getValueAPF().isNegative()) {
2477 case FCmpInst::FCMP_OLT:
2478 // No value is ordered and less than negative infinity.
2479 return ConstantInt::getFalse(CFP->getContext());
2480 case FCmpInst::FCMP_UGE:
2481 // All values are unordered with or at least negative infinity.
2482 return ConstantInt::getTrue(CFP->getContext());
2488 case FCmpInst::FCMP_OGT:
2489 // No value is ordered and greater than infinity.
2490 return ConstantInt::getFalse(CFP->getContext());
2491 case FCmpInst::FCMP_ULE:
2492 // All values are unordered with and at most infinity.
2493 return ConstantInt::getTrue(CFP->getContext());
2502 // If the comparison is with the result of a select instruction, check whether
2503 // comparing with either branch of the select always yields the same value.
2504 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2505 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2508 // If the comparison is with the result of a phi instruction, check whether
2509 // doing the compare with each incoming phi value yields a common result.
2510 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2511 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2517 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2518 const DataLayout *TD,
2519 const TargetLibraryInfo *TLI,
2520 const DominatorTree *DT) {
2521 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query (TD, TLI, DT),
2525 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
2526 /// the result. If not, this returns null.
2527 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
2528 Value *FalseVal, const Query &Q,
2529 unsigned MaxRecurse) {
2530 // select true, X, Y -> X
2531 // select false, X, Y -> Y
2532 if (ConstantInt *CB = dyn_cast<ConstantInt>(CondVal))
2533 return CB->getZExtValue() ? TrueVal : FalseVal;
2535 // select C, X, X -> X
2536 if (TrueVal == FalseVal)
2539 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
2540 if (isa<Constant>(TrueVal))
2544 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
2546 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
2552 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
2553 const DataLayout *TD,
2554 const TargetLibraryInfo *TLI,
2555 const DominatorTree *DT) {
2556 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Query (TD, TLI, DT),
2560 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
2561 /// fold the result. If not, this returns null.
2562 static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) {
2563 // The type of the GEP pointer operand.
2564 PointerType *PtrTy = dyn_cast<PointerType>(Ops[0]->getType());
2565 // The GEP pointer operand is not a pointer, it's a vector of pointers.
2569 // getelementptr P -> P.
2570 if (Ops.size() == 1)
2573 if (isa<UndefValue>(Ops[0])) {
2574 // Compute the (pointer) type returned by the GEP instruction.
2575 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1));
2576 Type *GEPTy = PointerType::get(LastType, PtrTy->getAddressSpace());
2577 return UndefValue::get(GEPTy);
2580 if (Ops.size() == 2) {
2581 // getelementptr P, 0 -> P.
2582 if (ConstantInt *C = dyn_cast<ConstantInt>(Ops[1]))
2585 // getelementptr P, N -> P if P points to a type of zero size.
2587 Type *Ty = PtrTy->getElementType();
2588 if (Ty->isSized() && Q.TD->getTypeAllocSize(Ty) == 0)
2593 // Check to see if this is constant foldable.
2594 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2595 if (!isa<Constant>(Ops[i]))
2598 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1));
2601 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *TD,
2602 const TargetLibraryInfo *TLI,
2603 const DominatorTree *DT) {
2604 return ::SimplifyGEPInst(Ops, Query (TD, TLI, DT), RecursionLimit);
2607 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
2608 /// can fold the result. If not, this returns null.
2609 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
2610 ArrayRef<unsigned> Idxs, const Query &Q,
2612 if (Constant *CAgg = dyn_cast<Constant>(Agg))
2613 if (Constant *CVal = dyn_cast<Constant>(Val))
2614 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
2616 // insertvalue x, undef, n -> x
2617 if (match(Val, m_Undef()))
2620 // insertvalue x, (extractvalue y, n), n
2621 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
2622 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
2623 EV->getIndices() == Idxs) {
2624 // insertvalue undef, (extractvalue y, n), n -> y
2625 if (match(Agg, m_Undef()))
2626 return EV->getAggregateOperand();
2628 // insertvalue y, (extractvalue y, n), n -> y
2629 if (Agg == EV->getAggregateOperand())
2636 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
2637 ArrayRef<unsigned> Idxs,
2638 const DataLayout *TD,
2639 const TargetLibraryInfo *TLI,
2640 const DominatorTree *DT) {
2641 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query (TD, TLI, DT),
2645 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
2646 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
2647 // If all of the PHI's incoming values are the same then replace the PHI node
2648 // with the common value.
2649 Value *CommonValue = 0;
2650 bool HasUndefInput = false;
2651 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2652 Value *Incoming = PN->getIncomingValue(i);
2653 // If the incoming value is the phi node itself, it can safely be skipped.
2654 if (Incoming == PN) continue;
2655 if (isa<UndefValue>(Incoming)) {
2656 // Remember that we saw an undef value, but otherwise ignore them.
2657 HasUndefInput = true;
2660 if (CommonValue && Incoming != CommonValue)
2661 return 0; // Not the same, bail out.
2662 CommonValue = Incoming;
2665 // If CommonValue is null then all of the incoming values were either undef or
2666 // equal to the phi node itself.
2668 return UndefValue::get(PN->getType());
2670 // If we have a PHI node like phi(X, undef, X), where X is defined by some
2671 // instruction, we cannot return X as the result of the PHI node unless it
2672 // dominates the PHI block.
2674 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : 0;
2679 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
2680 if (Constant *C = dyn_cast<Constant>(Op))
2681 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.TD, Q.TLI);
2686 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *TD,
2687 const TargetLibraryInfo *TLI,
2688 const DominatorTree *DT) {
2689 return ::SimplifyTruncInst(Op, Ty, Query (TD, TLI, DT), RecursionLimit);
2692 //=== Helper functions for higher up the class hierarchy.
2694 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
2695 /// fold the result. If not, this returns null.
2696 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
2697 const Query &Q, unsigned MaxRecurse) {
2699 case Instruction::Add:
2700 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2702 case Instruction::Sub:
2703 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2705 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
2706 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
2707 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
2708 case Instruction::FDiv: return SimplifyFDivInst(LHS, RHS, Q, MaxRecurse);
2709 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
2710 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
2711 case Instruction::FRem: return SimplifyFRemInst(LHS, RHS, Q, MaxRecurse);
2712 case Instruction::Shl:
2713 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2715 case Instruction::LShr:
2716 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
2717 case Instruction::AShr:
2718 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
2719 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
2720 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
2721 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
2723 if (Constant *CLHS = dyn_cast<Constant>(LHS))
2724 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
2725 Constant *COps[] = {CLHS, CRHS};
2726 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.TD,
2730 // If the operation is associative, try some generic simplifications.
2731 if (Instruction::isAssociative(Opcode))
2732 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
2735 // If the operation is with the result of a select instruction check whether
2736 // operating on either branch of the select always yields the same value.
2737 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2738 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
2741 // If the operation is with the result of a phi instruction, check whether
2742 // operating on all incoming values of the phi always yields the same value.
2743 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2744 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
2751 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
2752 const DataLayout *TD, const TargetLibraryInfo *TLI,
2753 const DominatorTree *DT) {
2754 return ::SimplifyBinOp(Opcode, LHS, RHS, Query (TD, TLI, DT), RecursionLimit);
2757 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
2758 /// fold the result.
2759 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2760 const Query &Q, unsigned MaxRecurse) {
2761 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
2762 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
2763 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
2766 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2767 const DataLayout *TD, const TargetLibraryInfo *TLI,
2768 const DominatorTree *DT) {
2769 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query (TD, TLI, DT),
2773 static Value *SimplifyCallInst(CallInst *CI, const Query &) {
2774 // call undef -> undef
2775 if (isa<UndefValue>(CI->getCalledValue()))
2776 return UndefValue::get(CI->getType());
2781 /// SimplifyInstruction - See if we can compute a simplified version of this
2782 /// instruction. If not, this returns null.
2783 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *TD,
2784 const TargetLibraryInfo *TLI,
2785 const DominatorTree *DT) {
2788 switch (I->getOpcode()) {
2790 Result = ConstantFoldInstruction(I, TD, TLI);
2792 case Instruction::Add:
2793 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
2794 cast<BinaryOperator>(I)->hasNoSignedWrap(),
2795 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
2798 case Instruction::Sub:
2799 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
2800 cast<BinaryOperator>(I)->hasNoSignedWrap(),
2801 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
2804 case Instruction::FMul:
2805 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
2806 I->getFastMathFlags(), TD, TLI, DT);
2808 case Instruction::Mul:
2809 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2811 case Instruction::SDiv:
2812 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2814 case Instruction::UDiv:
2815 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2817 case Instruction::FDiv:
2818 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2820 case Instruction::SRem:
2821 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2823 case Instruction::URem:
2824 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2826 case Instruction::FRem:
2827 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2829 case Instruction::Shl:
2830 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
2831 cast<BinaryOperator>(I)->hasNoSignedWrap(),
2832 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
2835 case Instruction::LShr:
2836 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
2837 cast<BinaryOperator>(I)->isExact(),
2840 case Instruction::AShr:
2841 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
2842 cast<BinaryOperator>(I)->isExact(),
2845 case Instruction::And:
2846 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2848 case Instruction::Or:
2849 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2851 case Instruction::Xor:
2852 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2854 case Instruction::ICmp:
2855 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
2856 I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2858 case Instruction::FCmp:
2859 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
2860 I->getOperand(0), I->getOperand(1), TD, TLI, DT);
2862 case Instruction::Select:
2863 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
2864 I->getOperand(2), TD, TLI, DT);
2866 case Instruction::GetElementPtr: {
2867 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
2868 Result = SimplifyGEPInst(Ops, TD, TLI, DT);
2871 case Instruction::InsertValue: {
2872 InsertValueInst *IV = cast<InsertValueInst>(I);
2873 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
2874 IV->getInsertedValueOperand(),
2875 IV->getIndices(), TD, TLI, DT);
2878 case Instruction::PHI:
2879 Result = SimplifyPHINode(cast<PHINode>(I), Query (TD, TLI, DT));
2881 case Instruction::Call:
2882 Result = SimplifyCallInst(cast<CallInst>(I), Query (TD, TLI, DT));
2884 case Instruction::Trunc:
2885 Result = SimplifyTruncInst(I->getOperand(0), I->getType(), TD, TLI, DT);
2889 /// If called on unreachable code, the above logic may report that the
2890 /// instruction simplified to itself. Make life easier for users by
2891 /// detecting that case here, returning a safe value instead.
2892 return Result == I ? UndefValue::get(I->getType()) : Result;
2895 /// \brief Implementation of recursive simplification through an instructions
2898 /// This is the common implementation of the recursive simplification routines.
2899 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
2900 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
2901 /// instructions to process and attempt to simplify it using
2902 /// InstructionSimplify.
2904 /// This routine returns 'true' only when *it* simplifies something. The passed
2905 /// in simplified value does not count toward this.
2906 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
2907 const DataLayout *TD,
2908 const TargetLibraryInfo *TLI,
2909 const DominatorTree *DT) {
2910 bool Simplified = false;
2911 SmallSetVector<Instruction *, 8> Worklist;
2913 // If we have an explicit value to collapse to, do that round of the
2914 // simplification loop by hand initially.
2916 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
2919 Worklist.insert(cast<Instruction>(*UI));
2921 // Replace the instruction with its simplified value.
2922 I->replaceAllUsesWith(SimpleV);
2924 // Gracefully handle edge cases where the instruction is not wired into any
2927 I->eraseFromParent();
2932 // Note that we must test the size on each iteration, the worklist can grow.
2933 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
2936 // See if this instruction simplifies.
2937 SimpleV = SimplifyInstruction(I, TD, TLI, DT);
2943 // Stash away all the uses of the old instruction so we can check them for
2944 // recursive simplifications after a RAUW. This is cheaper than checking all
2945 // uses of To on the recursive step in most cases.
2946 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
2948 Worklist.insert(cast<Instruction>(*UI));
2950 // Replace the instruction with its simplified value.
2951 I->replaceAllUsesWith(SimpleV);
2953 // Gracefully handle edge cases where the instruction is not wired into any
2956 I->eraseFromParent();
2961 bool llvm::recursivelySimplifyInstruction(Instruction *I,
2962 const DataLayout *TD,
2963 const TargetLibraryInfo *TLI,
2964 const DominatorTree *DT) {
2965 return replaceAndRecursivelySimplifyImpl(I, 0, TD, TLI, DT);
2968 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
2969 const DataLayout *TD,
2970 const TargetLibraryInfo *TLI,
2971 const DominatorTree *DT) {
2972 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
2973 assert(SimpleV && "Must provide a simplified value.");
2974 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TD, TLI, DT);