1 //===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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).
17 //
18 //===----------------------------------------------------------------------===//
19
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/ConstantFolding.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/ConstantRange.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/GetElementPtrTypeIterator.h"
30 #include "llvm/IR/GlobalAlias.h"
31 #include "llvm/IR/Operator.h"
32 #include "llvm/IR/PatternMatch.h"
33 #include "llvm/IR/ValueHandle.h"
34 using namespace llvm;
35 using namespace llvm::PatternMatch;
36
37 #define DEBUG_TYPE "instsimplify"
38
39 enum { RecursionLimit = 3 };
40
41 STATISTIC(NumExpand, "Number of expansions");
42 STATISTIC(NumReassoc, "Number of reassociations");
43
44 struct Query {
45 const DataLayout *DL;
46 const TargetLibraryInfo *TLI;
47 const DominatorTree *DT;
48
QueryQuery49 Query(const DataLayout *DL, const TargetLibraryInfo *tli,
50 const DominatorTree *dt) : DL(DL), TLI(tli), DT(dt) {}
51 };
52
53 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
54 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
55 unsigned);
56 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
57 unsigned);
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);
61
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.
getFalse(Type * Ty)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);
68 }
69
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.
getTrue(Type * Ty)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);
76 }
77
78 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
isSameCompare(Value * V,CmpInst::Predicate Pred,Value * LHS,Value * RHS)79 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
80 Value *RHS) {
81 CmpInst *Cmp = dyn_cast<CmpInst>(V);
82 if (!Cmp)
83 return false;
84 CmpInst::Predicate CPred = Cmp->getPredicate();
85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
86 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
87 return true;
88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
89 CRHS == LHS;
90 }
91
92 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
ValueDominatesPHI(Value * V,PHINode * P,const DominatorTree * DT)93 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
94 Instruction *I = dyn_cast<Instruction>(V);
95 if (!I)
96 // Arguments and constants dominate all instructions.
97 return true;
98
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())
103 return false;
104
105 // If we have a DominatorTree then do a precise test.
106 if (DT) {
107 if (!DT->isReachableFromEntry(P->getParent()))
108 return true;
109 if (!DT->isReachableFromEntry(I->getParent()))
110 return false;
111 return DT->dominates(I, P);
112 }
113
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() &&
117 !isa<InvokeInst>(I))
118 return true;
119
120 return false;
121 }
122
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.
ExpandBinOp(unsigned Opcode,Value * LHS,Value * RHS,unsigned OpcToExpand,const Query & Q,unsigned MaxRecurse)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.
133 if (!MaxRecurse--)
134 return nullptr;
135
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)) {
148 ++NumExpand;
149 return LHS;
150 }
151 // Otherwise return "L op' R" if it simplifies.
152 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
153 ++NumExpand;
154 return V;
155 }
156 }
157 }
158
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)) {
171 ++NumExpand;
172 return RHS;
173 }
174 // Otherwise return "L op' R" if it simplifies.
175 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
176 ++NumExpand;
177 return V;
178 }
179 }
180 }
181
182 return nullptr;
183 }
184
185 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
186 /// operations. Returns the simpler value, or null if none was found.
SimplifyAssociativeBinOp(unsigned Opc,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)187 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
188 const Query &Q, unsigned MaxRecurse) {
189 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
190 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
191
192 // Recursion is always used, so bail out at once if we already hit the limit.
193 if (!MaxRecurse--)
194 return nullptr;
195
196 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
197 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
198
199 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
200 if (Op0 && Op0->getOpcode() == Opcode) {
201 Value *A = Op0->getOperand(0);
202 Value *B = Op0->getOperand(1);
203 Value *C = RHS;
204
205 // Does "B op C" simplify?
206 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
207 // It does! Return "A op V" if it simplifies or is already available.
208 // If V equals B then "A op V" is just the LHS.
209 if (V == B) return LHS;
210 // Otherwise return "A op V" if it simplifies.
211 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
212 ++NumReassoc;
213 return W;
214 }
215 }
216 }
217
218 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
219 if (Op1 && Op1->getOpcode() == Opcode) {
220 Value *A = LHS;
221 Value *B = Op1->getOperand(0);
222 Value *C = Op1->getOperand(1);
223
224 // Does "A op B" simplify?
225 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
226 // It does! Return "V op C" if it simplifies or is already available.
227 // If V equals B then "V op C" is just the RHS.
228 if (V == B) return RHS;
229 // Otherwise return "V op C" if it simplifies.
230 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
231 ++NumReassoc;
232 return W;
233 }
234 }
235 }
236
237 // The remaining transforms require commutativity as well as associativity.
238 if (!Instruction::isCommutative(Opcode))
239 return nullptr;
240
241 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
242 if (Op0 && Op0->getOpcode() == Opcode) {
243 Value *A = Op0->getOperand(0);
244 Value *B = Op0->getOperand(1);
245 Value *C = RHS;
246
247 // Does "C op A" simplify?
248 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
249 // It does! Return "V op B" if it simplifies or is already available.
250 // If V equals A then "V op B" is just the LHS.
251 if (V == A) return LHS;
252 // Otherwise return "V op B" if it simplifies.
253 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
254 ++NumReassoc;
255 return W;
256 }
257 }
258 }
259
260 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
261 if (Op1 && Op1->getOpcode() == Opcode) {
262 Value *A = LHS;
263 Value *B = Op1->getOperand(0);
264 Value *C = Op1->getOperand(1);
265
266 // Does "C op A" simplify?
267 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
268 // It does! Return "B op V" if it simplifies or is already available.
269 // If V equals C then "B op V" is just the RHS.
270 if (V == C) return RHS;
271 // Otherwise return "B op V" if it simplifies.
272 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
273 ++NumReassoc;
274 return W;
275 }
276 }
277 }
278
279 return nullptr;
280 }
281
282 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
283 /// instruction as an operand, try to simplify the binop by seeing whether
284 /// evaluating it on both branches of the select results in the same value.
285 /// Returns the common value if so, otherwise returns null.
ThreadBinOpOverSelect(unsigned Opcode,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)286 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
287 const Query &Q, unsigned MaxRecurse) {
288 // Recursion is always used, so bail out at once if we already hit the limit.
289 if (!MaxRecurse--)
290 return nullptr;
291
292 SelectInst *SI;
293 if (isa<SelectInst>(LHS)) {
294 SI = cast<SelectInst>(LHS);
295 } else {
296 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
297 SI = cast<SelectInst>(RHS);
298 }
299
300 // Evaluate the BinOp on the true and false branches of the select.
301 Value *TV;
302 Value *FV;
303 if (SI == LHS) {
304 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
305 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
306 } else {
307 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
308 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
309 }
310
311 // If they simplified to the same value, then return the common value.
312 // If they both failed to simplify then return null.
313 if (TV == FV)
314 return TV;
315
316 // If one branch simplified to undef, return the other one.
317 if (TV && isa<UndefValue>(TV))
318 return FV;
319 if (FV && isa<UndefValue>(FV))
320 return TV;
321
322 // If applying the operation did not change the true and false select values,
323 // then the result of the binop is the select itself.
324 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
325 return SI;
326
327 // If one branch simplified and the other did not, and the simplified
328 // value is equal to the unsimplified one, return the simplified value.
329 // For example, select (cond, X, X & Z) & Z -> X & Z.
330 if ((FV && !TV) || (TV && !FV)) {
331 // Check that the simplified value has the form "X op Y" where "op" is the
332 // same as the original operation.
333 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
334 if (Simplified && Simplified->getOpcode() == Opcode) {
335 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
336 // We already know that "op" is the same as for the simplified value. See
337 // if the operands match too. If so, return the simplified value.
338 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
339 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
340 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
341 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
342 Simplified->getOperand(1) == UnsimplifiedRHS)
343 return Simplified;
344 if (Simplified->isCommutative() &&
345 Simplified->getOperand(1) == UnsimplifiedLHS &&
346 Simplified->getOperand(0) == UnsimplifiedRHS)
347 return Simplified;
348 }
349 }
350
351 return nullptr;
352 }
353
354 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
355 /// try to simplify the comparison by seeing whether both branches of the select
356 /// result in the same value. Returns the common value if so, otherwise returns
357 /// null.
ThreadCmpOverSelect(CmpInst::Predicate Pred,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)358 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
359 Value *RHS, const Query &Q,
360 unsigned MaxRecurse) {
361 // Recursion is always used, so bail out at once if we already hit the limit.
362 if (!MaxRecurse--)
363 return nullptr;
364
365 // Make sure the select is on the LHS.
366 if (!isa<SelectInst>(LHS)) {
367 std::swap(LHS, RHS);
368 Pred = CmpInst::getSwappedPredicate(Pred);
369 }
370 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
371 SelectInst *SI = cast<SelectInst>(LHS);
372 Value *Cond = SI->getCondition();
373 Value *TV = SI->getTrueValue();
374 Value *FV = SI->getFalseValue();
375
376 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
377 // Does "cmp TV, RHS" simplify?
378 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
379 if (TCmp == Cond) {
380 // It not only simplified, it simplified to the select condition. Replace
381 // it with 'true'.
382 TCmp = getTrue(Cond->getType());
383 } else if (!TCmp) {
384 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
385 // condition then we can replace it with 'true'. Otherwise give up.
386 if (!isSameCompare(Cond, Pred, TV, RHS))
387 return nullptr;
388 TCmp = getTrue(Cond->getType());
389 }
390
391 // Does "cmp FV, RHS" simplify?
392 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
393 if (FCmp == Cond) {
394 // It not only simplified, it simplified to the select condition. Replace
395 // it with 'false'.
396 FCmp = getFalse(Cond->getType());
397 } else if (!FCmp) {
398 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
399 // condition then we can replace it with 'false'. Otherwise give up.
400 if (!isSameCompare(Cond, Pred, FV, RHS))
401 return nullptr;
402 FCmp = getFalse(Cond->getType());
403 }
404
405 // If both sides simplified to the same value, then use it as the result of
406 // the original comparison.
407 if (TCmp == FCmp)
408 return TCmp;
409
410 // The remaining cases only make sense if the select condition has the same
411 // type as the result of the comparison, so bail out if this is not so.
412 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
413 return nullptr;
414 // If the false value simplified to false, then the result of the compare
415 // is equal to "Cond && TCmp". This also catches the case when the false
416 // value simplified to false and the true value to true, returning "Cond".
417 if (match(FCmp, m_Zero()))
418 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
419 return V;
420 // If the true value simplified to true, then the result of the compare
421 // is equal to "Cond || FCmp".
422 if (match(TCmp, m_One()))
423 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
424 return V;
425 // Finally, if the false value simplified to true and the true value to
426 // false, then the result of the compare is equal to "!Cond".
427 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
428 if (Value *V =
429 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
430 Q, MaxRecurse))
431 return V;
432
433 return nullptr;
434 }
435
436 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
437 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
438 /// it on the incoming phi values yields the same result for every value. If so
439 /// returns the common value, otherwise returns null.
ThreadBinOpOverPHI(unsigned Opcode,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)440 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
441 const Query &Q, unsigned MaxRecurse) {
442 // Recursion is always used, so bail out at once if we already hit the limit.
443 if (!MaxRecurse--)
444 return nullptr;
445
446 PHINode *PI;
447 if (isa<PHINode>(LHS)) {
448 PI = cast<PHINode>(LHS);
449 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
450 if (!ValueDominatesPHI(RHS, PI, Q.DT))
451 return nullptr;
452 } else {
453 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
454 PI = cast<PHINode>(RHS);
455 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
456 if (!ValueDominatesPHI(LHS, PI, Q.DT))
457 return nullptr;
458 }
459
460 // Evaluate the BinOp on the incoming phi values.
461 Value *CommonValue = nullptr;
462 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
463 Value *Incoming = PI->getIncomingValue(i);
464 // If the incoming value is the phi node itself, it can safely be skipped.
465 if (Incoming == PI) continue;
466 Value *V = PI == LHS ?
467 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
468 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
469 // If the operation failed to simplify, or simplified to a different value
470 // to previously, then give up.
471 if (!V || (CommonValue && V != CommonValue))
472 return nullptr;
473 CommonValue = V;
474 }
475
476 return CommonValue;
477 }
478
479 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
480 /// try to simplify the comparison by seeing whether comparing with all of the
481 /// incoming phi values yields the same result every time. If so returns the
482 /// common result, otherwise returns null.
ThreadCmpOverPHI(CmpInst::Predicate Pred,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)483 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
484 const Query &Q, unsigned MaxRecurse) {
485 // Recursion is always used, so bail out at once if we already hit the limit.
486 if (!MaxRecurse--)
487 return nullptr;
488
489 // Make sure the phi is on the LHS.
490 if (!isa<PHINode>(LHS)) {
491 std::swap(LHS, RHS);
492 Pred = CmpInst::getSwappedPredicate(Pred);
493 }
494 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
495 PHINode *PI = cast<PHINode>(LHS);
496
497 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
498 if (!ValueDominatesPHI(RHS, PI, Q.DT))
499 return nullptr;
500
501 // Evaluate the BinOp on the incoming phi values.
502 Value *CommonValue = nullptr;
503 for (unsigned i = 0, e = PI->getNumIncomingValues(); i != e; ++i) {
504 Value *Incoming = PI->getIncomingValue(i);
505 // If the incoming value is the phi node itself, it can safely be skipped.
506 if (Incoming == PI) continue;
507 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
508 // If the operation failed to simplify, or simplified to a different value
509 // to previously, then give up.
510 if (!V || (CommonValue && V != CommonValue))
511 return nullptr;
512 CommonValue = V;
513 }
514
515 return CommonValue;
516 }
517
518 /// SimplifyAddInst - Given operands for an Add, see if we can
519 /// fold the result. If not, this returns null.
SimplifyAddInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const Query & Q,unsigned MaxRecurse)520 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
521 const Query &Q, unsigned MaxRecurse) {
522 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
523 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
524 Constant *Ops[] = { CLHS, CRHS };
525 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
526 Q.DL, Q.TLI);
527 }
528
529 // Canonicalize the constant to the RHS.
530 std::swap(Op0, Op1);
531 }
532
533 // X + undef -> undef
534 if (match(Op1, m_Undef()))
535 return Op1;
536
537 // X + 0 -> X
538 if (match(Op1, m_Zero()))
539 return Op0;
540
541 // X + (Y - X) -> Y
542 // (Y - X) + X -> Y
543 // Eg: X + -X -> 0
544 Value *Y = nullptr;
545 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
546 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
547 return Y;
548
549 // X + ~X -> -1 since ~X = -X-1
550 if (match(Op0, m_Not(m_Specific(Op1))) ||
551 match(Op1, m_Not(m_Specific(Op0))))
552 return Constant::getAllOnesValue(Op0->getType());
553
554 /// i1 add -> xor.
555 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
556 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
557 return V;
558
559 // Try some generic simplifications for associative operations.
560 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
561 MaxRecurse))
562 return V;
563
564 // Threading Add over selects and phi nodes is pointless, so don't bother.
565 // Threading over the select in "A + select(cond, B, C)" means evaluating
566 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
567 // only if B and C are equal. If B and C are equal then (since we assume
568 // that operands have already been simplified) "select(cond, B, C)" should
569 // have been simplified to the common value of B and C already. Analysing
570 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
571 // for threading over phi nodes.
572
573 return nullptr;
574 }
575
SimplifyAddInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)576 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
577 const DataLayout *DL, const TargetLibraryInfo *TLI,
578 const DominatorTree *DT) {
579 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
580 RecursionLimit);
581 }
582
583 /// \brief Compute the base pointer and cumulative constant offsets for V.
584 ///
585 /// This strips all constant offsets off of V, leaving it the base pointer, and
586 /// accumulates the total constant offset applied in the returned constant. It
587 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
588 /// no constant offsets applied.
589 ///
590 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
591 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
592 /// folding.
stripAndComputeConstantOffsets(const DataLayout * DL,Value * & V,bool AllowNonInbounds=false)593 static Constant *stripAndComputeConstantOffsets(const DataLayout *DL,
594 Value *&V,
595 bool AllowNonInbounds = false) {
596 assert(V->getType()->getScalarType()->isPointerTy());
597
598 // Without DataLayout, just be conservative for now. Theoretically, more could
599 // be done in this case.
600 if (!DL)
601 return ConstantInt::get(IntegerType::get(V->getContext(), 64), 0);
602
603 Type *IntPtrTy = DL->getIntPtrType(V->getType())->getScalarType();
604 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
605
606 // Even though we don't look through PHI nodes, we could be called on an
607 // instruction in an unreachable block, which may be on a cycle.
608 SmallPtrSet<Value *, 4> Visited;
609 Visited.insert(V);
610 do {
611 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
612 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
613 !GEP->accumulateConstantOffset(*DL, Offset))
614 break;
615 V = GEP->getPointerOperand();
616 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
617 V = cast<Operator>(V)->getOperand(0);
618 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
619 if (GA->mayBeOverridden())
620 break;
621 V = GA->getAliasee();
622 } else {
623 break;
624 }
625 assert(V->getType()->getScalarType()->isPointerTy() &&
626 "Unexpected operand type!");
627 } while (Visited.insert(V));
628
629 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
630 if (V->getType()->isVectorTy())
631 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
632 OffsetIntPtr);
633 return OffsetIntPtr;
634 }
635
636 /// \brief Compute the constant difference between two pointer values.
637 /// If the difference is not a constant, returns zero.
computePointerDifference(const DataLayout * DL,Value * LHS,Value * RHS)638 static Constant *computePointerDifference(const DataLayout *DL,
639 Value *LHS, Value *RHS) {
640 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
641 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
642
643 // If LHS and RHS are not related via constant offsets to the same base
644 // value, there is nothing we can do here.
645 if (LHS != RHS)
646 return nullptr;
647
648 // Otherwise, the difference of LHS - RHS can be computed as:
649 // LHS - RHS
650 // = (LHSOffset + Base) - (RHSOffset + Base)
651 // = LHSOffset - RHSOffset
652 return ConstantExpr::getSub(LHSOffset, RHSOffset);
653 }
654
655 /// SimplifySubInst - Given operands for a Sub, see if we can
656 /// fold the result. If not, this returns null.
SimplifySubInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const Query & Q,unsigned MaxRecurse)657 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
658 const Query &Q, unsigned MaxRecurse) {
659 if (Constant *CLHS = dyn_cast<Constant>(Op0))
660 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
661 Constant *Ops[] = { CLHS, CRHS };
662 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
663 Ops, Q.DL, Q.TLI);
664 }
665
666 // X - undef -> undef
667 // undef - X -> undef
668 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
669 return UndefValue::get(Op0->getType());
670
671 // X - 0 -> X
672 if (match(Op1, m_Zero()))
673 return Op0;
674
675 // X - X -> 0
676 if (Op0 == Op1)
677 return Constant::getNullValue(Op0->getType());
678
679 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
680 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
681 Value *X = nullptr, *Y = nullptr, *Z = Op1;
682 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
683 // See if "V === Y - Z" simplifies.
684 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
685 // It does! Now see if "X + V" simplifies.
686 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
687 // It does, we successfully reassociated!
688 ++NumReassoc;
689 return W;
690 }
691 // See if "V === X - Z" simplifies.
692 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
693 // It does! Now see if "Y + V" simplifies.
694 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
695 // It does, we successfully reassociated!
696 ++NumReassoc;
697 return W;
698 }
699 }
700
701 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
702 // For example, X - (X + 1) -> -1
703 X = Op0;
704 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
705 // See if "V === X - Y" simplifies.
706 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
707 // It does! Now see if "V - Z" simplifies.
708 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
709 // It does, we successfully reassociated!
710 ++NumReassoc;
711 return W;
712 }
713 // See if "V === X - Z" simplifies.
714 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
715 // It does! Now see if "V - Y" simplifies.
716 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
717 // It does, we successfully reassociated!
718 ++NumReassoc;
719 return W;
720 }
721 }
722
723 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
724 // For example, X - (X - Y) -> Y.
725 Z = Op0;
726 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
727 // See if "V === Z - X" simplifies.
728 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
729 // It does! Now see if "V + Y" simplifies.
730 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
731 // It does, we successfully reassociated!
732 ++NumReassoc;
733 return W;
734 }
735
736 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
737 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
738 match(Op1, m_Trunc(m_Value(Y))))
739 if (X->getType() == Y->getType())
740 // See if "V === X - Y" simplifies.
741 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
742 // It does! Now see if "trunc V" simplifies.
743 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
744 // It does, return the simplified "trunc V".
745 return W;
746
747 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
748 if (match(Op0, m_PtrToInt(m_Value(X))) &&
749 match(Op1, m_PtrToInt(m_Value(Y))))
750 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
751 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
752
753 // i1 sub -> xor.
754 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
755 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
756 return V;
757
758 // Threading Sub over selects and phi nodes is pointless, so don't bother.
759 // Threading over the select in "A - select(cond, B, C)" means evaluating
760 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
761 // only if B and C are equal. If B and C are equal then (since we assume
762 // that operands have already been simplified) "select(cond, B, C)" should
763 // have been simplified to the common value of B and C already. Analysing
764 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
765 // for threading over phi nodes.
766
767 return nullptr;
768 }
769
SimplifySubInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)770 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
771 const DataLayout *DL, const TargetLibraryInfo *TLI,
772 const DominatorTree *DT) {
773 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
774 RecursionLimit);
775 }
776
777 /// Given operands for an FAdd, see if we can fold the result. If not, this
778 /// returns null.
SimplifyFAddInst(Value * Op0,Value * Op1,FastMathFlags FMF,const Query & Q,unsigned MaxRecurse)779 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
780 const Query &Q, unsigned MaxRecurse) {
781 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
782 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
783 Constant *Ops[] = { CLHS, CRHS };
784 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(),
785 Ops, Q.DL, Q.TLI);
786 }
787
788 // Canonicalize the constant to the RHS.
789 std::swap(Op0, Op1);
790 }
791
792 // fadd X, -0 ==> X
793 if (match(Op1, m_NegZero()))
794 return Op0;
795
796 // fadd X, 0 ==> X, when we know X is not -0
797 if (match(Op1, m_Zero()) &&
798 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
799 return Op0;
800
801 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
802 // where nnan and ninf have to occur at least once somewhere in this
803 // expression
804 Value *SubOp = nullptr;
805 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
806 SubOp = Op1;
807 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
808 SubOp = Op0;
809 if (SubOp) {
810 Instruction *FSub = cast<Instruction>(SubOp);
811 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
812 (FMF.noInfs() || FSub->hasNoInfs()))
813 return Constant::getNullValue(Op0->getType());
814 }
815
816 return nullptr;
817 }
818
819 /// Given operands for an FSub, see if we can fold the result. If not, this
820 /// returns null.
SimplifyFSubInst(Value * Op0,Value * Op1,FastMathFlags FMF,const Query & Q,unsigned MaxRecurse)821 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
822 const Query &Q, unsigned MaxRecurse) {
823 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
824 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
825 Constant *Ops[] = { CLHS, CRHS };
826 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(),
827 Ops, Q.DL, Q.TLI);
828 }
829 }
830
831 // fsub X, 0 ==> X
832 if (match(Op1, m_Zero()))
833 return Op0;
834
835 // fsub X, -0 ==> X, when we know X is not -0
836 if (match(Op1, m_NegZero()) &&
837 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
838 return Op0;
839
840 // fsub 0, (fsub -0.0, X) ==> X
841 Value *X;
842 if (match(Op0, m_AnyZero())) {
843 if (match(Op1, m_FSub(m_NegZero(), m_Value(X))))
844 return X;
845 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
846 return X;
847 }
848
849 // fsub nnan ninf x, x ==> 0.0
850 if (FMF.noNaNs() && FMF.noInfs() && Op0 == Op1)
851 return Constant::getNullValue(Op0->getType());
852
853 return nullptr;
854 }
855
856 /// Given the operands for an FMul, see if we can fold the result
SimplifyFMulInst(Value * Op0,Value * Op1,FastMathFlags FMF,const Query & Q,unsigned MaxRecurse)857 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
858 FastMathFlags FMF,
859 const Query &Q,
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(),
865 Ops, Q.DL, Q.TLI);
866 }
867
868 // Canonicalize the constant to the RHS.
869 std::swap(Op0, Op1);
870 }
871
872 // fmul X, 1.0 ==> X
873 if (match(Op1, m_FPOne()))
874 return Op0;
875
876 // fmul nnan nsz X, 0 ==> 0
877 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
878 return Op1;
879
880 return nullptr;
881 }
882
883 /// SimplifyMulInst - Given operands for a Mul, see if we can
884 /// fold the result. If not, this returns null.
SimplifyMulInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)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(),
891 Ops, Q.DL, Q.TLI);
892 }
893
894 // Canonicalize the constant to the RHS.
895 std::swap(Op0, Op1);
896 }
897
898 // X * undef -> 0
899 if (match(Op1, m_Undef()))
900 return Constant::getNullValue(Op0->getType());
901
902 // X * 0 -> 0
903 if (match(Op1, m_Zero()))
904 return Op1;
905
906 // X * 1 -> X
907 if (match(Op1, m_One()))
908 return Op0;
909
910 // (X / Y) * Y -> X if the division is exact.
911 Value *X = nullptr;
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)
914 return X;
915
916 // i1 mul -> and.
917 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
918 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
919 return V;
920
921 // Try some generic simplifications for associative operations.
922 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
923 MaxRecurse))
924 return V;
925
926 // Mul distributes over Add. Try some generic simplifications based on this.
927 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
928 Q, MaxRecurse))
929 return V;
930
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,
935 MaxRecurse))
936 return V;
937
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,
942 MaxRecurse))
943 return V;
944
945 return nullptr;
946 }
947
SimplifyFAddInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)948 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
949 const DataLayout *DL, const TargetLibraryInfo *TLI,
950 const DominatorTree *DT) {
951 return ::SimplifyFAddInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
952 }
953
SimplifyFSubInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)954 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
955 const DataLayout *DL, const TargetLibraryInfo *TLI,
956 const DominatorTree *DT) {
957 return ::SimplifyFSubInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
958 }
959
SimplifyFMulInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)960 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1,
961 FastMathFlags FMF,
962 const DataLayout *DL,
963 const TargetLibraryInfo *TLI,
964 const DominatorTree *DT) {
965 return ::SimplifyFMulInst(Op0, Op1, FMF, Query (DL, TLI, DT), RecursionLimit);
966 }
967
SimplifyMulInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)968 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout *DL,
969 const TargetLibraryInfo *TLI,
970 const DominatorTree *DT) {
971 return ::SimplifyMulInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
972 }
973
974 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
975 /// fold the result. If not, this returns null.
SimplifyDiv(Instruction::BinaryOps Opcode,Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)976 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
977 const Query &Q, unsigned MaxRecurse) {
978 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
979 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
980 Constant *Ops[] = { C0, C1 };
981 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
982 }
983 }
984
985 bool isSigned = Opcode == Instruction::SDiv;
986
987 // X / undef -> undef
988 if (match(Op1, m_Undef()))
989 return Op1;
990
991 // undef / X -> 0
992 if (match(Op0, m_Undef()))
993 return Constant::getNullValue(Op0->getType());
994
995 // 0 / X -> 0, we don't need to preserve faults!
996 if (match(Op0, m_Zero()))
997 return Op0;
998
999 // X / 1 -> X
1000 if (match(Op1, m_One()))
1001 return Op0;
1002
1003 if (Op0->getType()->isIntegerTy(1))
1004 // It can't be division by zero, hence it must be division by one.
1005 return Op0;
1006
1007 // X / X -> 1
1008 if (Op0 == Op1)
1009 return ConstantInt::get(Op0->getType(), 1);
1010
1011 // (X * Y) / Y -> X if the multiplication does not overflow.
1012 Value *X = nullptr, *Y = nullptr;
1013 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1014 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1015 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1016 // If the Mul knows it does not overflow, then we are good to go.
1017 if ((isSigned && Mul->hasNoSignedWrap()) ||
1018 (!isSigned && Mul->hasNoUnsignedWrap()))
1019 return X;
1020 // If X has the form X = A / Y then X * Y cannot overflow.
1021 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1022 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1023 return X;
1024 }
1025
1026 // (X rem Y) / Y -> 0
1027 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1028 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1029 return Constant::getNullValue(Op0->getType());
1030
1031 // If the operation is with the result of a select instruction, check whether
1032 // operating on either branch of the select always yields the same value.
1033 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1034 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1035 return V;
1036
1037 // If the operation is with the result of a phi instruction, check whether
1038 // operating on all incoming values of the phi always yields the same value.
1039 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1040 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1041 return V;
1042
1043 return nullptr;
1044 }
1045
1046 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1047 /// fold the result. If not, this returns null.
SimplifySDivInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1048 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1049 unsigned MaxRecurse) {
1050 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1051 return V;
1052
1053 return nullptr;
1054 }
1055
SimplifySDivInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1056 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1057 const TargetLibraryInfo *TLI,
1058 const DominatorTree *DT) {
1059 return ::SimplifySDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1060 }
1061
1062 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1063 /// fold the result. If not, this returns null.
SimplifyUDivInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1064 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1065 unsigned MaxRecurse) {
1066 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1067 return V;
1068
1069 return nullptr;
1070 }
1071
SimplifyUDivInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1072 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1073 const TargetLibraryInfo *TLI,
1074 const DominatorTree *DT) {
1075 return ::SimplifyUDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1076 }
1077
SimplifyFDivInst(Value * Op0,Value * Op1,const Query & Q,unsigned)1078 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, const Query &Q,
1079 unsigned) {
1080 // undef / X -> undef (the undef could be a snan).
1081 if (match(Op0, m_Undef()))
1082 return Op0;
1083
1084 // X / undef -> undef
1085 if (match(Op1, m_Undef()))
1086 return Op1;
1087
1088 return nullptr;
1089 }
1090
SimplifyFDivInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1091 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, const DataLayout *DL,
1092 const TargetLibraryInfo *TLI,
1093 const DominatorTree *DT) {
1094 return ::SimplifyFDivInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1095 }
1096
1097 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1098 /// fold the result. If not, this returns null.
SimplifyRem(Instruction::BinaryOps Opcode,Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1099 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1100 const Query &Q, unsigned MaxRecurse) {
1101 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1102 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1103 Constant *Ops[] = { C0, C1 };
1104 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1105 }
1106 }
1107
1108 // X % undef -> undef
1109 if (match(Op1, m_Undef()))
1110 return Op1;
1111
1112 // undef % X -> 0
1113 if (match(Op0, m_Undef()))
1114 return Constant::getNullValue(Op0->getType());
1115
1116 // 0 % X -> 0, we don't need to preserve faults!
1117 if (match(Op0, m_Zero()))
1118 return Op0;
1119
1120 // X % 0 -> undef, we don't need to preserve faults!
1121 if (match(Op1, m_Zero()))
1122 return UndefValue::get(Op0->getType());
1123
1124 // X % 1 -> 0
1125 if (match(Op1, m_One()))
1126 return Constant::getNullValue(Op0->getType());
1127
1128 if (Op0->getType()->isIntegerTy(1))
1129 // It can't be remainder by zero, hence it must be remainder by one.
1130 return Constant::getNullValue(Op0->getType());
1131
1132 // X % X -> 0
1133 if (Op0 == Op1)
1134 return Constant::getNullValue(Op0->getType());
1135
1136 // If the operation is with the result of a select instruction, check whether
1137 // operating on either branch of the select always yields the same value.
1138 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1139 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1140 return V;
1141
1142 // If the operation is with the result of a phi instruction, check whether
1143 // operating on all incoming values of the phi always yields the same value.
1144 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1145 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1146 return V;
1147
1148 return nullptr;
1149 }
1150
1151 /// SimplifySRemInst - Given operands for an SRem, see if we can
1152 /// fold the result. If not, this returns null.
SimplifySRemInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1153 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1154 unsigned MaxRecurse) {
1155 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1156 return V;
1157
1158 return nullptr;
1159 }
1160
SimplifySRemInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1161 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1162 const TargetLibraryInfo *TLI,
1163 const DominatorTree *DT) {
1164 return ::SimplifySRemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1165 }
1166
1167 /// SimplifyURemInst - Given operands for a URem, see if we can
1168 /// fold the result. If not, this returns null.
SimplifyURemInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1169 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1170 unsigned MaxRecurse) {
1171 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1172 return V;
1173
1174 return nullptr;
1175 }
1176
SimplifyURemInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1177 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1178 const TargetLibraryInfo *TLI,
1179 const DominatorTree *DT) {
1180 return ::SimplifyURemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1181 }
1182
SimplifyFRemInst(Value * Op0,Value * Op1,const Query &,unsigned)1183 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, const Query &,
1184 unsigned) {
1185 // undef % X -> undef (the undef could be a snan).
1186 if (match(Op0, m_Undef()))
1187 return Op0;
1188
1189 // X % undef -> undef
1190 if (match(Op1, m_Undef()))
1191 return Op1;
1192
1193 return nullptr;
1194 }
1195
SimplifyFRemInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1196 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, const DataLayout *DL,
1197 const TargetLibraryInfo *TLI,
1198 const DominatorTree *DT) {
1199 return ::SimplifyFRemInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1200 }
1201
1202 /// isUndefShift - Returns true if a shift by \c Amount always yields undef.
isUndefShift(Value * Amount)1203 static bool isUndefShift(Value *Amount) {
1204 Constant *C = dyn_cast<Constant>(Amount);
1205 if (!C)
1206 return false;
1207
1208 // X shift by undef -> undef because it may shift by the bitwidth.
1209 if (isa<UndefValue>(C))
1210 return true;
1211
1212 // Shifting by the bitwidth or more is undefined.
1213 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1214 if (CI->getValue().getLimitedValue() >=
1215 CI->getType()->getScalarSizeInBits())
1216 return true;
1217
1218 // If all lanes of a vector shift are undefined the whole shift is.
1219 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1220 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1221 if (!isUndefShift(C->getAggregateElement(I)))
1222 return false;
1223 return true;
1224 }
1225
1226 return false;
1227 }
1228
1229 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1230 /// fold the result. If not, this returns null.
SimplifyShift(unsigned Opcode,Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1231 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1232 const Query &Q, unsigned MaxRecurse) {
1233 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1234 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1235 Constant *Ops[] = { C0, C1 };
1236 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1237 }
1238 }
1239
1240 // 0 shift by X -> 0
1241 if (match(Op0, m_Zero()))
1242 return Op0;
1243
1244 // X shift by 0 -> X
1245 if (match(Op1, m_Zero()))
1246 return Op0;
1247
1248 // Fold undefined shifts.
1249 if (isUndefShift(Op1))
1250 return UndefValue::get(Op0->getType());
1251
1252 // If the operation is with the result of a select instruction, check whether
1253 // operating on either branch of the select always yields the same value.
1254 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1255 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1256 return V;
1257
1258 // If the operation is with the result of a phi instruction, check whether
1259 // operating on all incoming values of the phi always yields the same value.
1260 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1261 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1262 return V;
1263
1264 return nullptr;
1265 }
1266
1267 /// SimplifyShlInst - Given operands for an Shl, see if we can
1268 /// fold the result. If not, this returns null.
SimplifyShlInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const Query & Q,unsigned MaxRecurse)1269 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1270 const Query &Q, unsigned MaxRecurse) {
1271 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1272 return V;
1273
1274 // undef << X -> 0
1275 if (match(Op0, m_Undef()))
1276 return Constant::getNullValue(Op0->getType());
1277
1278 // (X >> A) << A -> X
1279 Value *X;
1280 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1281 return X;
1282 return nullptr;
1283 }
1284
SimplifyShlInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1285 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1286 const DataLayout *DL, const TargetLibraryInfo *TLI,
1287 const DominatorTree *DT) {
1288 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query (DL, TLI, DT),
1289 RecursionLimit);
1290 }
1291
1292 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1293 /// fold the result. If not, this returns null.
SimplifyLShrInst(Value * Op0,Value * Op1,bool isExact,const Query & Q,unsigned MaxRecurse)1294 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1295 const Query &Q, unsigned MaxRecurse) {
1296 if (Value *V = SimplifyShift(Instruction::LShr, Op0, Op1, Q, MaxRecurse))
1297 return V;
1298
1299 // X >> X -> 0
1300 if (Op0 == Op1)
1301 return Constant::getNullValue(Op0->getType());
1302
1303 // undef >>l X -> 0
1304 if (match(Op0, m_Undef()))
1305 return Constant::getNullValue(Op0->getType());
1306
1307 // (X << A) >> A -> X
1308 Value *X;
1309 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1310 cast<OverflowingBinaryOperator>(Op0)->hasNoUnsignedWrap())
1311 return X;
1312
1313 return nullptr;
1314 }
1315
SimplifyLShrInst(Value * Op0,Value * Op1,bool isExact,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1316 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1317 const DataLayout *DL,
1318 const TargetLibraryInfo *TLI,
1319 const DominatorTree *DT) {
1320 return ::SimplifyLShrInst(Op0, Op1, isExact, Query (DL, TLI, DT),
1321 RecursionLimit);
1322 }
1323
1324 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1325 /// fold the result. If not, this returns null.
SimplifyAShrInst(Value * Op0,Value * Op1,bool isExact,const Query & Q,unsigned MaxRecurse)1326 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1327 const Query &Q, unsigned MaxRecurse) {
1328 if (Value *V = SimplifyShift(Instruction::AShr, Op0, Op1, Q, MaxRecurse))
1329 return V;
1330
1331 // X >> X -> 0
1332 if (Op0 == Op1)
1333 return Constant::getNullValue(Op0->getType());
1334
1335 // all ones >>a X -> all ones
1336 if (match(Op0, m_AllOnes()))
1337 return Op0;
1338
1339 // undef >>a X -> all ones
1340 if (match(Op0, m_Undef()))
1341 return Constant::getAllOnesValue(Op0->getType());
1342
1343 // (X << A) >> A -> X
1344 Value *X;
1345 if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
1346 cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap())
1347 return X;
1348
1349 return nullptr;
1350 }
1351
SimplifyAShrInst(Value * Op0,Value * Op1,bool isExact,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1352 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1353 const DataLayout *DL,
1354 const TargetLibraryInfo *TLI,
1355 const DominatorTree *DT) {
1356 return ::SimplifyAShrInst(Op0, Op1, isExact, Query (DL, TLI, DT),
1357 RecursionLimit);
1358 }
1359
1360 /// SimplifyAndInst - Given operands for an And, see if we can
1361 /// fold the result. If not, this returns null.
SimplifyAndInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1362 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1363 unsigned MaxRecurse) {
1364 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1365 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1366 Constant *Ops[] = { CLHS, CRHS };
1367 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1368 Ops, Q.DL, Q.TLI);
1369 }
1370
1371 // Canonicalize the constant to the RHS.
1372 std::swap(Op0, Op1);
1373 }
1374
1375 // X & undef -> 0
1376 if (match(Op1, m_Undef()))
1377 return Constant::getNullValue(Op0->getType());
1378
1379 // X & X = X
1380 if (Op0 == Op1)
1381 return Op0;
1382
1383 // X & 0 = 0
1384 if (match(Op1, m_Zero()))
1385 return Op1;
1386
1387 // X & -1 = X
1388 if (match(Op1, m_AllOnes()))
1389 return Op0;
1390
1391 // A & ~A = ~A & A = 0
1392 if (match(Op0, m_Not(m_Specific(Op1))) ||
1393 match(Op1, m_Not(m_Specific(Op0))))
1394 return Constant::getNullValue(Op0->getType());
1395
1396 // (A | ?) & A = A
1397 Value *A = nullptr, *B = nullptr;
1398 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1399 (A == Op1 || B == Op1))
1400 return Op1;
1401
1402 // A & (A | ?) = A
1403 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1404 (A == Op0 || B == Op0))
1405 return Op0;
1406
1407 // A & (-A) = A if A is a power of two or zero.
1408 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1409 match(Op1, m_Neg(m_Specific(Op0)))) {
1410 if (isKnownToBeAPowerOfTwo(Op0, /*OrZero*/true))
1411 return Op0;
1412 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true))
1413 return Op1;
1414 }
1415
1416 // Try some generic simplifications for associative operations.
1417 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1418 MaxRecurse))
1419 return V;
1420
1421 // And distributes over Or. Try some generic simplifications based on this.
1422 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1423 Q, MaxRecurse))
1424 return V;
1425
1426 // And distributes over Xor. Try some generic simplifications based on this.
1427 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1428 Q, MaxRecurse))
1429 return V;
1430
1431 // If the operation is with the result of a select instruction, check whether
1432 // operating on either branch of the select always yields the same value.
1433 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1434 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1435 MaxRecurse))
1436 return V;
1437
1438 // If the operation is with the result of a phi instruction, check whether
1439 // operating on all incoming values of the phi always yields the same value.
1440 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1441 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1442 MaxRecurse))
1443 return V;
1444
1445 return nullptr;
1446 }
1447
SimplifyAndInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1448 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout *DL,
1449 const TargetLibraryInfo *TLI,
1450 const DominatorTree *DT) {
1451 return ::SimplifyAndInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1452 }
1453
1454 /// SimplifyOrInst - Given operands for an Or, see if we can
1455 /// fold the result. If not, this returns null.
SimplifyOrInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1456 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1457 unsigned MaxRecurse) {
1458 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1459 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1460 Constant *Ops[] = { CLHS, CRHS };
1461 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1462 Ops, Q.DL, Q.TLI);
1463 }
1464
1465 // Canonicalize the constant to the RHS.
1466 std::swap(Op0, Op1);
1467 }
1468
1469 // X | undef -> -1
1470 if (match(Op1, m_Undef()))
1471 return Constant::getAllOnesValue(Op0->getType());
1472
1473 // X | X = X
1474 if (Op0 == Op1)
1475 return Op0;
1476
1477 // X | 0 = X
1478 if (match(Op1, m_Zero()))
1479 return Op0;
1480
1481 // X | -1 = -1
1482 if (match(Op1, m_AllOnes()))
1483 return Op1;
1484
1485 // A | ~A = ~A | A = -1
1486 if (match(Op0, m_Not(m_Specific(Op1))) ||
1487 match(Op1, m_Not(m_Specific(Op0))))
1488 return Constant::getAllOnesValue(Op0->getType());
1489
1490 // (A & ?) | A = A
1491 Value *A = nullptr, *B = nullptr;
1492 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1493 (A == Op1 || B == Op1))
1494 return Op1;
1495
1496 // A | (A & ?) = A
1497 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1498 (A == Op0 || B == Op0))
1499 return Op0;
1500
1501 // ~(A & ?) | A = -1
1502 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1503 (A == Op1 || B == Op1))
1504 return Constant::getAllOnesValue(Op1->getType());
1505
1506 // A | ~(A & ?) = -1
1507 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1508 (A == Op0 || B == Op0))
1509 return Constant::getAllOnesValue(Op0->getType());
1510
1511 // Try some generic simplifications for associative operations.
1512 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1513 MaxRecurse))
1514 return V;
1515
1516 // Or distributes over And. Try some generic simplifications based on this.
1517 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1518 MaxRecurse))
1519 return V;
1520
1521 // If the operation is with the result of a select instruction, check whether
1522 // operating on either branch of the select always yields the same value.
1523 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1524 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1525 MaxRecurse))
1526 return V;
1527
1528 // (A & C)|(B & D)
1529 Value *C = nullptr, *D = nullptr;
1530 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1531 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1532 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1533 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1534 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1535 // (A & C1)|(B & C2)
1536 // If we have: ((V + N) & C1) | (V & C2)
1537 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1538 // replace with V+N.
1539 Value *V1, *V2;
1540 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1541 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1542 // Add commutes, try both ways.
1543 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1544 return A;
1545 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1546 return A;
1547 }
1548 // Or commutes, try both ways.
1549 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1550 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1551 // Add commutes, try both ways.
1552 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1553 return B;
1554 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1555 return B;
1556 }
1557 }
1558 }
1559
1560 // If the operation is with the result of a phi instruction, check whether
1561 // operating on all incoming values of the phi always yields the same value.
1562 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1563 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1564 return V;
1565
1566 return nullptr;
1567 }
1568
SimplifyOrInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1569 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout *DL,
1570 const TargetLibraryInfo *TLI,
1571 const DominatorTree *DT) {
1572 return ::SimplifyOrInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1573 }
1574
1575 /// SimplifyXorInst - Given operands for a Xor, see if we can
1576 /// fold the result. If not, this returns null.
SimplifyXorInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1577 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1578 unsigned MaxRecurse) {
1579 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1580 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1581 Constant *Ops[] = { CLHS, CRHS };
1582 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1583 Ops, Q.DL, Q.TLI);
1584 }
1585
1586 // Canonicalize the constant to the RHS.
1587 std::swap(Op0, Op1);
1588 }
1589
1590 // A ^ undef -> undef
1591 if (match(Op1, m_Undef()))
1592 return Op1;
1593
1594 // A ^ 0 = A
1595 if (match(Op1, m_Zero()))
1596 return Op0;
1597
1598 // A ^ A = 0
1599 if (Op0 == Op1)
1600 return Constant::getNullValue(Op0->getType());
1601
1602 // A ^ ~A = ~A ^ A = -1
1603 if (match(Op0, m_Not(m_Specific(Op1))) ||
1604 match(Op1, m_Not(m_Specific(Op0))))
1605 return Constant::getAllOnesValue(Op0->getType());
1606
1607 // Try some generic simplifications for associative operations.
1608 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1609 MaxRecurse))
1610 return V;
1611
1612 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1613 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1614 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1615 // only if B and C are equal. If B and C are equal then (since we assume
1616 // that operands have already been simplified) "select(cond, B, C)" should
1617 // have been simplified to the common value of B and C already. Analysing
1618 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1619 // for threading over phi nodes.
1620
1621 return nullptr;
1622 }
1623
SimplifyXorInst(Value * Op0,Value * Op1,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)1624 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout *DL,
1625 const TargetLibraryInfo *TLI,
1626 const DominatorTree *DT) {
1627 return ::SimplifyXorInst(Op0, Op1, Query (DL, TLI, DT), RecursionLimit);
1628 }
1629
GetCompareTy(Value * Op)1630 static Type *GetCompareTy(Value *Op) {
1631 return CmpInst::makeCmpResultType(Op->getType());
1632 }
1633
1634 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1635 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1636 /// otherwise return null. Helper function for analyzing max/min idioms.
ExtractEquivalentCondition(Value * V,CmpInst::Predicate Pred,Value * LHS,Value * RHS)1637 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1638 Value *LHS, Value *RHS) {
1639 SelectInst *SI = dyn_cast<SelectInst>(V);
1640 if (!SI)
1641 return nullptr;
1642 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1643 if (!Cmp)
1644 return nullptr;
1645 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1646 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1647 return Cmp;
1648 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1649 LHS == CmpRHS && RHS == CmpLHS)
1650 return Cmp;
1651 return nullptr;
1652 }
1653
1654 // A significant optimization not implemented here is assuming that alloca
1655 // addresses are not equal to incoming argument values. They don't *alias*,
1656 // as we say, but that doesn't mean they aren't equal, so we take a
1657 // conservative approach.
1658 //
1659 // This is inspired in part by C++11 5.10p1:
1660 // "Two pointers of the same type compare equal if and only if they are both
1661 // null, both point to the same function, or both represent the same
1662 // address."
1663 //
1664 // This is pretty permissive.
1665 //
1666 // It's also partly due to C11 6.5.9p6:
1667 // "Two pointers compare equal if and only if both are null pointers, both are
1668 // pointers to the same object (including a pointer to an object and a
1669 // subobject at its beginning) or function, both are pointers to one past the
1670 // last element of the same array object, or one is a pointer to one past the
1671 // end of one array object and the other is a pointer to the start of a
1672 // different array object that happens to immediately follow the first array
1673 // object in the address space.)
1674 //
1675 // C11's version is more restrictive, however there's no reason why an argument
1676 // couldn't be a one-past-the-end value for a stack object in the caller and be
1677 // equal to the beginning of a stack object in the callee.
1678 //
1679 // If the C and C++ standards are ever made sufficiently restrictive in this
1680 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1681 // this optimization.
computePointerICmp(const DataLayout * DL,const TargetLibraryInfo * TLI,CmpInst::Predicate Pred,Value * LHS,Value * RHS)1682 static Constant *computePointerICmp(const DataLayout *DL,
1683 const TargetLibraryInfo *TLI,
1684 CmpInst::Predicate Pred,
1685 Value *LHS, Value *RHS) {
1686 // First, skip past any trivial no-ops.
1687 LHS = LHS->stripPointerCasts();
1688 RHS = RHS->stripPointerCasts();
1689
1690 // A non-null pointer is not equal to a null pointer.
1691 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) &&
1692 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1693 return ConstantInt::get(GetCompareTy(LHS),
1694 !CmpInst::isTrueWhenEqual(Pred));
1695
1696 // We can only fold certain predicates on pointer comparisons.
1697 switch (Pred) {
1698 default:
1699 return nullptr;
1700
1701 // Equality comaprisons are easy to fold.
1702 case CmpInst::ICMP_EQ:
1703 case CmpInst::ICMP_NE:
1704 break;
1705
1706 // We can only handle unsigned relational comparisons because 'inbounds' on
1707 // a GEP only protects against unsigned wrapping.
1708 case CmpInst::ICMP_UGT:
1709 case CmpInst::ICMP_UGE:
1710 case CmpInst::ICMP_ULT:
1711 case CmpInst::ICMP_ULE:
1712 // However, we have to switch them to their signed variants to handle
1713 // negative indices from the base pointer.
1714 Pred = ICmpInst::getSignedPredicate(Pred);
1715 break;
1716 }
1717
1718 // Strip off any constant offsets so that we can reason about them.
1719 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
1720 // here and compare base addresses like AliasAnalysis does, however there are
1721 // numerous hazards. AliasAnalysis and its utilities rely on special rules
1722 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
1723 // doesn't need to guarantee pointer inequality when it says NoAlias.
1724 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
1725 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
1726
1727 // If LHS and RHS are related via constant offsets to the same base
1728 // value, we can replace it with an icmp which just compares the offsets.
1729 if (LHS == RHS)
1730 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
1731
1732 // Various optimizations for (in)equality comparisons.
1733 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
1734 // Different non-empty allocations that exist at the same time have
1735 // different addresses (if the program can tell). Global variables always
1736 // exist, so they always exist during the lifetime of each other and all
1737 // allocas. Two different allocas usually have different addresses...
1738 //
1739 // However, if there's an @llvm.stackrestore dynamically in between two
1740 // allocas, they may have the same address. It's tempting to reduce the
1741 // scope of the problem by only looking at *static* allocas here. That would
1742 // cover the majority of allocas while significantly reducing the likelihood
1743 // of having an @llvm.stackrestore pop up in the middle. However, it's not
1744 // actually impossible for an @llvm.stackrestore to pop up in the middle of
1745 // an entry block. Also, if we have a block that's not attached to a
1746 // function, we can't tell if it's "static" under the current definition.
1747 // Theoretically, this problem could be fixed by creating a new kind of
1748 // instruction kind specifically for static allocas. Such a new instruction
1749 // could be required to be at the top of the entry block, thus preventing it
1750 // from being subject to a @llvm.stackrestore. Instcombine could even
1751 // convert regular allocas into these special allocas. It'd be nifty.
1752 // However, until then, this problem remains open.
1753 //
1754 // So, we'll assume that two non-empty allocas have different addresses
1755 // for now.
1756 //
1757 // With all that, if the offsets are within the bounds of their allocations
1758 // (and not one-past-the-end! so we can't use inbounds!), and their
1759 // allocations aren't the same, the pointers are not equal.
1760 //
1761 // Note that it's not necessary to check for LHS being a global variable
1762 // address, due to canonicalization and constant folding.
1763 if (isa<AllocaInst>(LHS) &&
1764 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
1765 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
1766 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
1767 uint64_t LHSSize, RHSSize;
1768 if (LHSOffsetCI && RHSOffsetCI &&
1769 getObjectSize(LHS, LHSSize, DL, TLI) &&
1770 getObjectSize(RHS, RHSSize, DL, TLI)) {
1771 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
1772 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
1773 if (!LHSOffsetValue.isNegative() &&
1774 !RHSOffsetValue.isNegative() &&
1775 LHSOffsetValue.ult(LHSSize) &&
1776 RHSOffsetValue.ult(RHSSize)) {
1777 return ConstantInt::get(GetCompareTy(LHS),
1778 !CmpInst::isTrueWhenEqual(Pred));
1779 }
1780 }
1781
1782 // Repeat the above check but this time without depending on DataLayout
1783 // or being able to compute a precise size.
1784 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
1785 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
1786 LHSOffset->isNullValue() &&
1787 RHSOffset->isNullValue())
1788 return ConstantInt::get(GetCompareTy(LHS),
1789 !CmpInst::isTrueWhenEqual(Pred));
1790 }
1791
1792 // Even if an non-inbounds GEP occurs along the path we can still optimize
1793 // equality comparisons concerning the result. We avoid walking the whole
1794 // chain again by starting where the last calls to
1795 // stripAndComputeConstantOffsets left off and accumulate the offsets.
1796 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
1797 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
1798 if (LHS == RHS)
1799 return ConstantExpr::getICmp(Pred,
1800 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
1801 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
1802 }
1803
1804 // Otherwise, fail.
1805 return nullptr;
1806 }
1807
1808 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
1809 /// fold the result. If not, this returns null.
SimplifyICmpInst(unsigned Predicate,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)1810 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
1811 const Query &Q, unsigned MaxRecurse) {
1812 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
1813 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
1814
1815 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
1816 if (Constant *CRHS = dyn_cast<Constant>(RHS))
1817 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
1818
1819 // If we have a constant, make sure it is on the RHS.
1820 std::swap(LHS, RHS);
1821 Pred = CmpInst::getSwappedPredicate(Pred);
1822 }
1823
1824 Type *ITy = GetCompareTy(LHS); // The return type.
1825 Type *OpTy = LHS->getType(); // The operand type.
1826
1827 // icmp X, X -> true/false
1828 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
1829 // because X could be 0.
1830 if (LHS == RHS || isa<UndefValue>(RHS))
1831 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
1832
1833 // Special case logic when the operands have i1 type.
1834 if (OpTy->getScalarType()->isIntegerTy(1)) {
1835 switch (Pred) {
1836 default: break;
1837 case ICmpInst::ICMP_EQ:
1838 // X == 1 -> X
1839 if (match(RHS, m_One()))
1840 return LHS;
1841 break;
1842 case ICmpInst::ICMP_NE:
1843 // X != 0 -> X
1844 if (match(RHS, m_Zero()))
1845 return LHS;
1846 break;
1847 case ICmpInst::ICMP_UGT:
1848 // X >u 0 -> X
1849 if (match(RHS, m_Zero()))
1850 return LHS;
1851 break;
1852 case ICmpInst::ICMP_UGE:
1853 // X >=u 1 -> X
1854 if (match(RHS, m_One()))
1855 return LHS;
1856 break;
1857 case ICmpInst::ICMP_SLT:
1858 // X <s 0 -> X
1859 if (match(RHS, m_Zero()))
1860 return LHS;
1861 break;
1862 case ICmpInst::ICMP_SLE:
1863 // X <=s -1 -> X
1864 if (match(RHS, m_One()))
1865 return LHS;
1866 break;
1867 }
1868 }
1869
1870 // If we are comparing with zero then try hard since this is a common case.
1871 if (match(RHS, m_Zero())) {
1872 bool LHSKnownNonNegative, LHSKnownNegative;
1873 switch (Pred) {
1874 default: llvm_unreachable("Unknown ICmp predicate!");
1875 case ICmpInst::ICMP_ULT:
1876 return getFalse(ITy);
1877 case ICmpInst::ICMP_UGE:
1878 return getTrue(ITy);
1879 case ICmpInst::ICMP_EQ:
1880 case ICmpInst::ICMP_ULE:
1881 if (isKnownNonZero(LHS, Q.DL))
1882 return getFalse(ITy);
1883 break;
1884 case ICmpInst::ICMP_NE:
1885 case ICmpInst::ICMP_UGT:
1886 if (isKnownNonZero(LHS, Q.DL))
1887 return getTrue(ITy);
1888 break;
1889 case ICmpInst::ICMP_SLT:
1890 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1891 if (LHSKnownNegative)
1892 return getTrue(ITy);
1893 if (LHSKnownNonNegative)
1894 return getFalse(ITy);
1895 break;
1896 case ICmpInst::ICMP_SLE:
1897 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1898 if (LHSKnownNegative)
1899 return getTrue(ITy);
1900 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL))
1901 return getFalse(ITy);
1902 break;
1903 case ICmpInst::ICMP_SGE:
1904 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1905 if (LHSKnownNegative)
1906 return getFalse(ITy);
1907 if (LHSKnownNonNegative)
1908 return getTrue(ITy);
1909 break;
1910 case ICmpInst::ICMP_SGT:
1911 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL);
1912 if (LHSKnownNegative)
1913 return getFalse(ITy);
1914 if (LHSKnownNonNegative && isKnownNonZero(LHS, Q.DL))
1915 return getTrue(ITy);
1916 break;
1917 }
1918 }
1919
1920 // See if we are doing a comparison with a constant integer.
1921 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
1922 // Rule out tautological comparisons (eg., ult 0 or uge 0).
1923 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
1924 if (RHS_CR.isEmptySet())
1925 return ConstantInt::getFalse(CI->getContext());
1926 if (RHS_CR.isFullSet())
1927 return ConstantInt::getTrue(CI->getContext());
1928
1929 // Many binary operators with constant RHS have easy to compute constant
1930 // range. Use them to check whether the comparison is a tautology.
1931 unsigned Width = CI->getBitWidth();
1932 APInt Lower = APInt(Width, 0);
1933 APInt Upper = APInt(Width, 0);
1934 ConstantInt *CI2;
1935 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
1936 // 'urem x, CI2' produces [0, CI2).
1937 Upper = CI2->getValue();
1938 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
1939 // 'srem x, CI2' produces (-|CI2|, |CI2|).
1940 Upper = CI2->getValue().abs();
1941 Lower = (-Upper) + 1;
1942 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
1943 // 'udiv CI2, x' produces [0, CI2].
1944 Upper = CI2->getValue() + 1;
1945 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
1946 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
1947 APInt NegOne = APInt::getAllOnesValue(Width);
1948 if (!CI2->isZero())
1949 Upper = NegOne.udiv(CI2->getValue()) + 1;
1950 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) {
1951 if (CI2->isMinSignedValue()) {
1952 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
1953 Lower = CI2->getValue();
1954 Upper = Lower.lshr(1) + 1;
1955 } else {
1956 // 'sdiv CI2, x' produces [-|CI2|, |CI2|].
1957 Upper = CI2->getValue().abs() + 1;
1958 Lower = (-Upper) + 1;
1959 }
1960 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
1961 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2].
1962 APInt IntMin = APInt::getSignedMinValue(Width);
1963 APInt IntMax = APInt::getSignedMaxValue(Width);
1964 APInt Val = CI2->getValue().abs();
1965 if (!Val.isMinValue()) {
1966 Lower = IntMin.sdiv(Val);
1967 Upper = IntMax.sdiv(Val) + 1;
1968 }
1969 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
1970 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
1971 APInt NegOne = APInt::getAllOnesValue(Width);
1972 if (CI2->getValue().ult(Width))
1973 Upper = NegOne.lshr(CI2->getValue()) + 1;
1974 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) {
1975 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2].
1976 unsigned ShiftAmount = Width - 1;
1977 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
1978 ShiftAmount = CI2->getValue().countTrailingZeros();
1979 Lower = CI2->getValue().lshr(ShiftAmount);
1980 Upper = CI2->getValue() + 1;
1981 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
1982 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
1983 APInt IntMin = APInt::getSignedMinValue(Width);
1984 APInt IntMax = APInt::getSignedMaxValue(Width);
1985 if (CI2->getValue().ult(Width)) {
1986 Lower = IntMin.ashr(CI2->getValue());
1987 Upper = IntMax.ashr(CI2->getValue()) + 1;
1988 }
1989 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) {
1990 unsigned ShiftAmount = Width - 1;
1991 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
1992 ShiftAmount = CI2->getValue().countTrailingZeros();
1993 if (CI2->isNegative()) {
1994 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)]
1995 Lower = CI2->getValue();
1996 Upper = CI2->getValue().ashr(ShiftAmount) + 1;
1997 } else {
1998 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2]
1999 Lower = CI2->getValue().ashr(ShiftAmount);
2000 Upper = CI2->getValue() + 1;
2001 }
2002 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
2003 // 'or x, CI2' produces [CI2, UINT_MAX].
2004 Lower = CI2->getValue();
2005 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
2006 // 'and x, CI2' produces [0, CI2].
2007 Upper = CI2->getValue() + 1;
2008 }
2009 if (Lower != Upper) {
2010 ConstantRange LHS_CR = ConstantRange(Lower, Upper);
2011 if (RHS_CR.contains(LHS_CR))
2012 return ConstantInt::getTrue(RHS->getContext());
2013 if (RHS_CR.inverse().contains(LHS_CR))
2014 return ConstantInt::getFalse(RHS->getContext());
2015 }
2016 }
2017
2018 // Compare of cast, for example (zext X) != 0 -> X != 0
2019 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
2020 Instruction *LI = cast<CastInst>(LHS);
2021 Value *SrcOp = LI->getOperand(0);
2022 Type *SrcTy = SrcOp->getType();
2023 Type *DstTy = LI->getType();
2024
2025 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
2026 // if the integer type is the same size as the pointer type.
2027 if (MaxRecurse && Q.DL && isa<PtrToIntInst>(LI) &&
2028 Q.DL->getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
2029 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2030 // Transfer the cast to the constant.
2031 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
2032 ConstantExpr::getIntToPtr(RHSC, SrcTy),
2033 Q, MaxRecurse-1))
2034 return V;
2035 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
2036 if (RI->getOperand(0)->getType() == SrcTy)
2037 // Compare without the cast.
2038 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2039 Q, MaxRecurse-1))
2040 return V;
2041 }
2042 }
2043
2044 if (isa<ZExtInst>(LHS)) {
2045 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
2046 // same type.
2047 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
2048 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2049 // Compare X and Y. Note that signed predicates become unsigned.
2050 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2051 SrcOp, RI->getOperand(0), Q,
2052 MaxRecurse-1))
2053 return V;
2054 }
2055 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
2056 // too. If not, then try to deduce the result of the comparison.
2057 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2058 // Compute the constant that would happen if we truncated to SrcTy then
2059 // reextended to DstTy.
2060 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2061 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
2062
2063 // If the re-extended constant didn't change then this is effectively
2064 // also a case of comparing two zero-extended values.
2065 if (RExt == CI && MaxRecurse)
2066 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2067 SrcOp, Trunc, Q, MaxRecurse-1))
2068 return V;
2069
2070 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
2071 // there. Use this to work out the result of the comparison.
2072 if (RExt != CI) {
2073 switch (Pred) {
2074 default: llvm_unreachable("Unknown ICmp predicate!");
2075 // LHS <u RHS.
2076 case ICmpInst::ICMP_EQ:
2077 case ICmpInst::ICMP_UGT:
2078 case ICmpInst::ICMP_UGE:
2079 return ConstantInt::getFalse(CI->getContext());
2080
2081 case ICmpInst::ICMP_NE:
2082 case ICmpInst::ICMP_ULT:
2083 case ICmpInst::ICMP_ULE:
2084 return ConstantInt::getTrue(CI->getContext());
2085
2086 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
2087 // is non-negative then LHS <s RHS.
2088 case ICmpInst::ICMP_SGT:
2089 case ICmpInst::ICMP_SGE:
2090 return CI->getValue().isNegative() ?
2091 ConstantInt::getTrue(CI->getContext()) :
2092 ConstantInt::getFalse(CI->getContext());
2093
2094 case ICmpInst::ICMP_SLT:
2095 case ICmpInst::ICMP_SLE:
2096 return CI->getValue().isNegative() ?
2097 ConstantInt::getFalse(CI->getContext()) :
2098 ConstantInt::getTrue(CI->getContext());
2099 }
2100 }
2101 }
2102 }
2103
2104 if (isa<SExtInst>(LHS)) {
2105 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
2106 // same type.
2107 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
2108 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2109 // Compare X and Y. Note that the predicate does not change.
2110 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2111 Q, MaxRecurse-1))
2112 return V;
2113 }
2114 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
2115 // too. If not, then try to deduce the result of the comparison.
2116 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2117 // Compute the constant that would happen if we truncated to SrcTy then
2118 // reextended to DstTy.
2119 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2120 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
2121
2122 // If the re-extended constant didn't change then this is effectively
2123 // also a case of comparing two sign-extended values.
2124 if (RExt == CI && MaxRecurse)
2125 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
2126 return V;
2127
2128 // Otherwise the upper bits of LHS are all equal, while RHS has varying
2129 // bits there. Use this to work out the result of the comparison.
2130 if (RExt != CI) {
2131 switch (Pred) {
2132 default: llvm_unreachable("Unknown ICmp predicate!");
2133 case ICmpInst::ICMP_EQ:
2134 return ConstantInt::getFalse(CI->getContext());
2135 case ICmpInst::ICMP_NE:
2136 return ConstantInt::getTrue(CI->getContext());
2137
2138 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
2139 // LHS >s RHS.
2140 case ICmpInst::ICMP_SGT:
2141 case ICmpInst::ICMP_SGE:
2142 return CI->getValue().isNegative() ?
2143 ConstantInt::getTrue(CI->getContext()) :
2144 ConstantInt::getFalse(CI->getContext());
2145 case ICmpInst::ICMP_SLT:
2146 case ICmpInst::ICMP_SLE:
2147 return CI->getValue().isNegative() ?
2148 ConstantInt::getFalse(CI->getContext()) :
2149 ConstantInt::getTrue(CI->getContext());
2150
2151 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2152 // LHS >u RHS.
2153 case ICmpInst::ICMP_UGT:
2154 case ICmpInst::ICMP_UGE:
2155 // Comparison is true iff the LHS <s 0.
2156 if (MaxRecurse)
2157 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2158 Constant::getNullValue(SrcTy),
2159 Q, MaxRecurse-1))
2160 return V;
2161 break;
2162 case ICmpInst::ICMP_ULT:
2163 case ICmpInst::ICMP_ULE:
2164 // Comparison is true iff the LHS >=s 0.
2165 if (MaxRecurse)
2166 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2167 Constant::getNullValue(SrcTy),
2168 Q, MaxRecurse-1))
2169 return V;
2170 break;
2171 }
2172 }
2173 }
2174 }
2175 }
2176
2177 // If a bit is known to be zero for A and known to be one for B,
2178 // then A and B cannot be equal.
2179 if (ICmpInst::isEquality(Pred)) {
2180 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2181 uint32_t BitWidth = CI->getBitWidth();
2182 APInt LHSKnownZero(BitWidth, 0);
2183 APInt LHSKnownOne(BitWidth, 0);
2184 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne);
2185 APInt RHSKnownZero(BitWidth, 0);
2186 APInt RHSKnownOne(BitWidth, 0);
2187 computeKnownBits(RHS, RHSKnownZero, RHSKnownOne);
2188 if (((LHSKnownOne & RHSKnownZero) != 0) ||
2189 ((LHSKnownZero & RHSKnownOne) != 0))
2190 return (Pred == ICmpInst::ICMP_EQ)
2191 ? ConstantInt::getFalse(CI->getContext())
2192 : ConstantInt::getTrue(CI->getContext());
2193 }
2194 }
2195
2196 // Special logic for binary operators.
2197 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2198 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2199 if (MaxRecurse && (LBO || RBO)) {
2200 // Analyze the case when either LHS or RHS is an add instruction.
2201 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2202 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2203 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2204 if (LBO && LBO->getOpcode() == Instruction::Add) {
2205 A = LBO->getOperand(0); B = LBO->getOperand(1);
2206 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2207 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2208 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2209 }
2210 if (RBO && RBO->getOpcode() == Instruction::Add) {
2211 C = RBO->getOperand(0); D = RBO->getOperand(1);
2212 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2213 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2214 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2215 }
2216
2217 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2218 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2219 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2220 Constant::getNullValue(RHS->getType()),
2221 Q, MaxRecurse-1))
2222 return V;
2223
2224 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2225 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2226 if (Value *V = SimplifyICmpInst(Pred,
2227 Constant::getNullValue(LHS->getType()),
2228 C == LHS ? D : C, Q, MaxRecurse-1))
2229 return V;
2230
2231 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2232 if (A && C && (A == C || A == D || B == C || B == D) &&
2233 NoLHSWrapProblem && NoRHSWrapProblem) {
2234 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2235 Value *Y, *Z;
2236 if (A == C) {
2237 // C + B == C + D -> B == D
2238 Y = B;
2239 Z = D;
2240 } else if (A == D) {
2241 // D + B == C + D -> B == C
2242 Y = B;
2243 Z = C;
2244 } else if (B == C) {
2245 // A + C == C + D -> A == D
2246 Y = A;
2247 Z = D;
2248 } else {
2249 assert(B == D);
2250 // A + D == C + D -> A == C
2251 Y = A;
2252 Z = C;
2253 }
2254 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2255 return V;
2256 }
2257 }
2258
2259 // 0 - (zext X) pred C
2260 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2261 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2262 if (RHSC->getValue().isStrictlyPositive()) {
2263 if (Pred == ICmpInst::ICMP_SLT)
2264 return ConstantInt::getTrue(RHSC->getContext());
2265 if (Pred == ICmpInst::ICMP_SGE)
2266 return ConstantInt::getFalse(RHSC->getContext());
2267 if (Pred == ICmpInst::ICMP_EQ)
2268 return ConstantInt::getFalse(RHSC->getContext());
2269 if (Pred == ICmpInst::ICMP_NE)
2270 return ConstantInt::getTrue(RHSC->getContext());
2271 }
2272 if (RHSC->getValue().isNonNegative()) {
2273 if (Pred == ICmpInst::ICMP_SLE)
2274 return ConstantInt::getTrue(RHSC->getContext());
2275 if (Pred == ICmpInst::ICMP_SGT)
2276 return ConstantInt::getFalse(RHSC->getContext());
2277 }
2278 }
2279 }
2280
2281 // icmp pred (urem X, Y), Y
2282 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2283 bool KnownNonNegative, KnownNegative;
2284 switch (Pred) {
2285 default:
2286 break;
2287 case ICmpInst::ICMP_SGT:
2288 case ICmpInst::ICMP_SGE:
2289 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL);
2290 if (!KnownNonNegative)
2291 break;
2292 // fall-through
2293 case ICmpInst::ICMP_EQ:
2294 case ICmpInst::ICMP_UGT:
2295 case ICmpInst::ICMP_UGE:
2296 return getFalse(ITy);
2297 case ICmpInst::ICMP_SLT:
2298 case ICmpInst::ICMP_SLE:
2299 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL);
2300 if (!KnownNonNegative)
2301 break;
2302 // fall-through
2303 case ICmpInst::ICMP_NE:
2304 case ICmpInst::ICMP_ULT:
2305 case ICmpInst::ICMP_ULE:
2306 return getTrue(ITy);
2307 }
2308 }
2309
2310 // icmp pred X, (urem Y, X)
2311 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2312 bool KnownNonNegative, KnownNegative;
2313 switch (Pred) {
2314 default:
2315 break;
2316 case ICmpInst::ICMP_SGT:
2317 case ICmpInst::ICMP_SGE:
2318 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL);
2319 if (!KnownNonNegative)
2320 break;
2321 // fall-through
2322 case ICmpInst::ICMP_NE:
2323 case ICmpInst::ICMP_UGT:
2324 case ICmpInst::ICMP_UGE:
2325 return getTrue(ITy);
2326 case ICmpInst::ICMP_SLT:
2327 case ICmpInst::ICMP_SLE:
2328 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL);
2329 if (!KnownNonNegative)
2330 break;
2331 // fall-through
2332 case ICmpInst::ICMP_EQ:
2333 case ICmpInst::ICMP_ULT:
2334 case ICmpInst::ICMP_ULE:
2335 return getFalse(ITy);
2336 }
2337 }
2338
2339 // x udiv y <=u x.
2340 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2341 // icmp pred (X /u Y), X
2342 if (Pred == ICmpInst::ICMP_UGT)
2343 return getFalse(ITy);
2344 if (Pred == ICmpInst::ICMP_ULE)
2345 return getTrue(ITy);
2346 }
2347
2348 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2349 LBO->getOperand(1) == RBO->getOperand(1)) {
2350 switch (LBO->getOpcode()) {
2351 default: break;
2352 case Instruction::UDiv:
2353 case Instruction::LShr:
2354 if (ICmpInst::isSigned(Pred))
2355 break;
2356 // fall-through
2357 case Instruction::SDiv:
2358 case Instruction::AShr:
2359 if (!LBO->isExact() || !RBO->isExact())
2360 break;
2361 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2362 RBO->getOperand(0), Q, MaxRecurse-1))
2363 return V;
2364 break;
2365 case Instruction::Shl: {
2366 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2367 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2368 if (!NUW && !NSW)
2369 break;
2370 if (!NSW && ICmpInst::isSigned(Pred))
2371 break;
2372 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2373 RBO->getOperand(0), Q, MaxRecurse-1))
2374 return V;
2375 break;
2376 }
2377 }
2378 }
2379
2380 // Simplify comparisons involving max/min.
2381 Value *A, *B;
2382 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2383 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2384
2385 // Signed variants on "max(a,b)>=a -> true".
2386 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2387 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2388 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2389 // We analyze this as smax(A, B) pred A.
2390 P = Pred;
2391 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2392 (A == LHS || B == LHS)) {
2393 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2394 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2395 // We analyze this as smax(A, B) swapped-pred A.
2396 P = CmpInst::getSwappedPredicate(Pred);
2397 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2398 (A == RHS || B == RHS)) {
2399 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2400 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2401 // We analyze this as smax(-A, -B) swapped-pred -A.
2402 // Note that we do not need to actually form -A or -B thanks to EqP.
2403 P = CmpInst::getSwappedPredicate(Pred);
2404 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2405 (A == LHS || B == LHS)) {
2406 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2407 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2408 // We analyze this as smax(-A, -B) pred -A.
2409 // Note that we do not need to actually form -A or -B thanks to EqP.
2410 P = Pred;
2411 }
2412 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2413 // Cases correspond to "max(A, B) p A".
2414 switch (P) {
2415 default:
2416 break;
2417 case CmpInst::ICMP_EQ:
2418 case CmpInst::ICMP_SLE:
2419 // Equivalent to "A EqP B". This may be the same as the condition tested
2420 // in the max/min; if so, we can just return that.
2421 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2422 return V;
2423 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2424 return V;
2425 // Otherwise, see if "A EqP B" simplifies.
2426 if (MaxRecurse)
2427 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2428 return V;
2429 break;
2430 case CmpInst::ICMP_NE:
2431 case CmpInst::ICMP_SGT: {
2432 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2433 // Equivalent to "A InvEqP B". This may be the same as the condition
2434 // tested in the max/min; if so, we can just return that.
2435 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2436 return V;
2437 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2438 return V;
2439 // Otherwise, see if "A InvEqP B" simplifies.
2440 if (MaxRecurse)
2441 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2442 return V;
2443 break;
2444 }
2445 case CmpInst::ICMP_SGE:
2446 // Always true.
2447 return getTrue(ITy);
2448 case CmpInst::ICMP_SLT:
2449 // Always false.
2450 return getFalse(ITy);
2451 }
2452 }
2453
2454 // Unsigned variants on "max(a,b)>=a -> true".
2455 P = CmpInst::BAD_ICMP_PREDICATE;
2456 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2457 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2458 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2459 // We analyze this as umax(A, B) pred A.
2460 P = Pred;
2461 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2462 (A == LHS || B == LHS)) {
2463 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2464 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2465 // We analyze this as umax(A, B) swapped-pred A.
2466 P = CmpInst::getSwappedPredicate(Pred);
2467 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2468 (A == RHS || B == RHS)) {
2469 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2470 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2471 // We analyze this as umax(-A, -B) swapped-pred -A.
2472 // Note that we do not need to actually form -A or -B thanks to EqP.
2473 P = CmpInst::getSwappedPredicate(Pred);
2474 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2475 (A == LHS || B == LHS)) {
2476 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2477 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2478 // We analyze this as umax(-A, -B) pred -A.
2479 // Note that we do not need to actually form -A or -B thanks to EqP.
2480 P = Pred;
2481 }
2482 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2483 // Cases correspond to "max(A, B) p A".
2484 switch (P) {
2485 default:
2486 break;
2487 case CmpInst::ICMP_EQ:
2488 case CmpInst::ICMP_ULE:
2489 // Equivalent to "A EqP B". This may be the same as the condition tested
2490 // in the max/min; if so, we can just return that.
2491 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2492 return V;
2493 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2494 return V;
2495 // Otherwise, see if "A EqP B" simplifies.
2496 if (MaxRecurse)
2497 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2498 return V;
2499 break;
2500 case CmpInst::ICMP_NE:
2501 case CmpInst::ICMP_UGT: {
2502 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2503 // Equivalent to "A InvEqP B". This may be the same as the condition
2504 // tested in the max/min; if so, we can just return that.
2505 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2506 return V;
2507 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2508 return V;
2509 // Otherwise, see if "A InvEqP B" simplifies.
2510 if (MaxRecurse)
2511 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2512 return V;
2513 break;
2514 }
2515 case CmpInst::ICMP_UGE:
2516 // Always true.
2517 return getTrue(ITy);
2518 case CmpInst::ICMP_ULT:
2519 // Always false.
2520 return getFalse(ITy);
2521 }
2522 }
2523
2524 // Variants on "max(x,y) >= min(x,z)".
2525 Value *C, *D;
2526 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2527 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2528 (A == C || A == D || B == C || B == D)) {
2529 // max(x, ?) pred min(x, ?).
2530 if (Pred == CmpInst::ICMP_SGE)
2531 // Always true.
2532 return getTrue(ITy);
2533 if (Pred == CmpInst::ICMP_SLT)
2534 // Always false.
2535 return getFalse(ITy);
2536 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2537 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
2538 (A == C || A == D || B == C || B == D)) {
2539 // min(x, ?) pred max(x, ?).
2540 if (Pred == CmpInst::ICMP_SLE)
2541 // Always true.
2542 return getTrue(ITy);
2543 if (Pred == CmpInst::ICMP_SGT)
2544 // Always false.
2545 return getFalse(ITy);
2546 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
2547 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
2548 (A == C || A == D || B == C || B == D)) {
2549 // max(x, ?) pred min(x, ?).
2550 if (Pred == CmpInst::ICMP_UGE)
2551 // Always true.
2552 return getTrue(ITy);
2553 if (Pred == CmpInst::ICMP_ULT)
2554 // Always false.
2555 return getFalse(ITy);
2556 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2557 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
2558 (A == C || A == D || B == C || B == D)) {
2559 // min(x, ?) pred max(x, ?).
2560 if (Pred == CmpInst::ICMP_ULE)
2561 // Always true.
2562 return getTrue(ITy);
2563 if (Pred == CmpInst::ICMP_UGT)
2564 // Always false.
2565 return getFalse(ITy);
2566 }
2567
2568 // Simplify comparisons of related pointers using a powerful, recursive
2569 // GEP-walk when we have target data available..
2570 if (LHS->getType()->isPointerTy())
2571 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS))
2572 return C;
2573
2574 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
2575 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
2576 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
2577 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
2578 (ICmpInst::isEquality(Pred) ||
2579 (GLHS->isInBounds() && GRHS->isInBounds() &&
2580 Pred == ICmpInst::getSignedPredicate(Pred)))) {
2581 // The bases are equal and the indices are constant. Build a constant
2582 // expression GEP with the same indices and a null base pointer to see
2583 // what constant folding can make out of it.
2584 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
2585 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
2586 Constant *NewLHS = ConstantExpr::getGetElementPtr(Null, IndicesLHS);
2587
2588 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
2589 Constant *NewRHS = ConstantExpr::getGetElementPtr(Null, IndicesRHS);
2590 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
2591 }
2592 }
2593 }
2594
2595 // If the comparison is with the result of a select instruction, check whether
2596 // comparing with either branch of the select always yields the same value.
2597 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2598 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2599 return V;
2600
2601 // If the comparison is with the result of a phi instruction, check whether
2602 // doing the compare with each incoming phi value yields a common result.
2603 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2604 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2605 return V;
2606
2607 return nullptr;
2608 }
2609
SimplifyICmpInst(unsigned Predicate,Value * LHS,Value * RHS,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2610 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2611 const DataLayout *DL,
2612 const TargetLibraryInfo *TLI,
2613 const DominatorTree *DT) {
2614 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
2615 RecursionLimit);
2616 }
2617
2618 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
2619 /// fold the result. If not, this returns null.
SimplifyFCmpInst(unsigned Predicate,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)2620 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2621 const Query &Q, unsigned MaxRecurse) {
2622 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2623 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
2624
2625 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2626 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2627 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
2628
2629 // If we have a constant, make sure it is on the RHS.
2630 std::swap(LHS, RHS);
2631 Pred = CmpInst::getSwappedPredicate(Pred);
2632 }
2633
2634 // Fold trivial predicates.
2635 if (Pred == FCmpInst::FCMP_FALSE)
2636 return ConstantInt::get(GetCompareTy(LHS), 0);
2637 if (Pred == FCmpInst::FCMP_TRUE)
2638 return ConstantInt::get(GetCompareTy(LHS), 1);
2639
2640 if (isa<UndefValue>(RHS)) // fcmp pred X, undef -> undef
2641 return UndefValue::get(GetCompareTy(LHS));
2642
2643 // fcmp x,x -> true/false. Not all compares are foldable.
2644 if (LHS == RHS) {
2645 if (CmpInst::isTrueWhenEqual(Pred))
2646 return ConstantInt::get(GetCompareTy(LHS), 1);
2647 if (CmpInst::isFalseWhenEqual(Pred))
2648 return ConstantInt::get(GetCompareTy(LHS), 0);
2649 }
2650
2651 // Handle fcmp with constant RHS
2652 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2653 // If the constant is a nan, see if we can fold the comparison based on it.
2654 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2655 if (CFP->getValueAPF().isNaN()) {
2656 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
2657 return ConstantInt::getFalse(CFP->getContext());
2658 assert(FCmpInst::isUnordered(Pred) &&
2659 "Comparison must be either ordered or unordered!");
2660 // True if unordered.
2661 return ConstantInt::getTrue(CFP->getContext());
2662 }
2663 // Check whether the constant is an infinity.
2664 if (CFP->getValueAPF().isInfinity()) {
2665 if (CFP->getValueAPF().isNegative()) {
2666 switch (Pred) {
2667 case FCmpInst::FCMP_OLT:
2668 // No value is ordered and less than negative infinity.
2669 return ConstantInt::getFalse(CFP->getContext());
2670 case FCmpInst::FCMP_UGE:
2671 // All values are unordered with or at least negative infinity.
2672 return ConstantInt::getTrue(CFP->getContext());
2673 default:
2674 break;
2675 }
2676 } else {
2677 switch (Pred) {
2678 case FCmpInst::FCMP_OGT:
2679 // No value is ordered and greater than infinity.
2680 return ConstantInt::getFalse(CFP->getContext());
2681 case FCmpInst::FCMP_ULE:
2682 // All values are unordered with and at most infinity.
2683 return ConstantInt::getTrue(CFP->getContext());
2684 default:
2685 break;
2686 }
2687 }
2688 }
2689 }
2690 }
2691
2692 // If the comparison is with the result of a select instruction, check whether
2693 // comparing with either branch of the select always yields the same value.
2694 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2695 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
2696 return V;
2697
2698 // If the comparison is with the result of a phi instruction, check whether
2699 // doing the compare with each incoming phi value yields a common result.
2700 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2701 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
2702 return V;
2703
2704 return nullptr;
2705 }
2706
SimplifyFCmpInst(unsigned Predicate,Value * LHS,Value * RHS,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2707 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2708 const DataLayout *DL,
2709 const TargetLibraryInfo *TLI,
2710 const DominatorTree *DT) {
2711 return ::SimplifyFCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
2712 RecursionLimit);
2713 }
2714
2715 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
2716 /// the result. If not, this returns null.
SimplifySelectInst(Value * CondVal,Value * TrueVal,Value * FalseVal,const Query & Q,unsigned MaxRecurse)2717 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
2718 Value *FalseVal, const Query &Q,
2719 unsigned MaxRecurse) {
2720 // select true, X, Y -> X
2721 // select false, X, Y -> Y
2722 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
2723 if (CB->isAllOnesValue())
2724 return TrueVal;
2725 if (CB->isNullValue())
2726 return FalseVal;
2727 }
2728
2729 // select C, X, X -> X
2730 if (TrueVal == FalseVal)
2731 return TrueVal;
2732
2733 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
2734 if (isa<Constant>(TrueVal))
2735 return TrueVal;
2736 return FalseVal;
2737 }
2738 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
2739 return FalseVal;
2740 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
2741 return TrueVal;
2742
2743 return nullptr;
2744 }
2745
SimplifySelectInst(Value * Cond,Value * TrueVal,Value * FalseVal,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2746 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
2747 const DataLayout *DL,
2748 const TargetLibraryInfo *TLI,
2749 const DominatorTree *DT) {
2750 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Query (DL, TLI, DT),
2751 RecursionLimit);
2752 }
2753
2754 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
2755 /// fold the result. If not, this returns null.
SimplifyGEPInst(ArrayRef<Value * > Ops,const Query & Q,unsigned)2756 static Value *SimplifyGEPInst(ArrayRef<Value *> Ops, const Query &Q, unsigned) {
2757 // The type of the GEP pointer operand.
2758 PointerType *PtrTy = cast<PointerType>(Ops[0]->getType()->getScalarType());
2759
2760 // getelementptr P -> P.
2761 if (Ops.size() == 1)
2762 return Ops[0];
2763
2764 if (isa<UndefValue>(Ops[0])) {
2765 // Compute the (pointer) type returned by the GEP instruction.
2766 Type *LastType = GetElementPtrInst::getIndexedType(PtrTy, Ops.slice(1));
2767 Type *GEPTy = PointerType::get(LastType, PtrTy->getAddressSpace());
2768 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
2769 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
2770 return UndefValue::get(GEPTy);
2771 }
2772
2773 if (Ops.size() == 2) {
2774 // getelementptr P, 0 -> P.
2775 if (match(Ops[1], m_Zero()))
2776 return Ops[0];
2777 // getelementptr P, N -> P if P points to a type of zero size.
2778 if (Q.DL) {
2779 Type *Ty = PtrTy->getElementType();
2780 if (Ty->isSized() && Q.DL->getTypeAllocSize(Ty) == 0)
2781 return Ops[0];
2782 }
2783 }
2784
2785 // Check to see if this is constant foldable.
2786 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
2787 if (!isa<Constant>(Ops[i]))
2788 return nullptr;
2789
2790 return ConstantExpr::getGetElementPtr(cast<Constant>(Ops[0]), Ops.slice(1));
2791 }
2792
SimplifyGEPInst(ArrayRef<Value * > Ops,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2793 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout *DL,
2794 const TargetLibraryInfo *TLI,
2795 const DominatorTree *DT) {
2796 return ::SimplifyGEPInst(Ops, Query (DL, TLI, DT), RecursionLimit);
2797 }
2798
2799 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
2800 /// can fold the result. If not, this returns null.
SimplifyInsertValueInst(Value * Agg,Value * Val,ArrayRef<unsigned> Idxs,const Query & Q,unsigned)2801 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
2802 ArrayRef<unsigned> Idxs, const Query &Q,
2803 unsigned) {
2804 if (Constant *CAgg = dyn_cast<Constant>(Agg))
2805 if (Constant *CVal = dyn_cast<Constant>(Val))
2806 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
2807
2808 // insertvalue x, undef, n -> x
2809 if (match(Val, m_Undef()))
2810 return Agg;
2811
2812 // insertvalue x, (extractvalue y, n), n
2813 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
2814 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
2815 EV->getIndices() == Idxs) {
2816 // insertvalue undef, (extractvalue y, n), n -> y
2817 if (match(Agg, m_Undef()))
2818 return EV->getAggregateOperand();
2819
2820 // insertvalue y, (extractvalue y, n), n -> y
2821 if (Agg == EV->getAggregateOperand())
2822 return Agg;
2823 }
2824
2825 return nullptr;
2826 }
2827
SimplifyInsertValueInst(Value * Agg,Value * Val,ArrayRef<unsigned> Idxs,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2828 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
2829 ArrayRef<unsigned> Idxs,
2830 const DataLayout *DL,
2831 const TargetLibraryInfo *TLI,
2832 const DominatorTree *DT) {
2833 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query (DL, TLI, DT),
2834 RecursionLimit);
2835 }
2836
2837 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
SimplifyPHINode(PHINode * PN,const Query & Q)2838 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
2839 // If all of the PHI's incoming values are the same then replace the PHI node
2840 // with the common value.
2841 Value *CommonValue = nullptr;
2842 bool HasUndefInput = false;
2843 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2844 Value *Incoming = PN->getIncomingValue(i);
2845 // If the incoming value is the phi node itself, it can safely be skipped.
2846 if (Incoming == PN) continue;
2847 if (isa<UndefValue>(Incoming)) {
2848 // Remember that we saw an undef value, but otherwise ignore them.
2849 HasUndefInput = true;
2850 continue;
2851 }
2852 if (CommonValue && Incoming != CommonValue)
2853 return nullptr; // Not the same, bail out.
2854 CommonValue = Incoming;
2855 }
2856
2857 // If CommonValue is null then all of the incoming values were either undef or
2858 // equal to the phi node itself.
2859 if (!CommonValue)
2860 return UndefValue::get(PN->getType());
2861
2862 // If we have a PHI node like phi(X, undef, X), where X is defined by some
2863 // instruction, we cannot return X as the result of the PHI node unless it
2864 // dominates the PHI block.
2865 if (HasUndefInput)
2866 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
2867
2868 return CommonValue;
2869 }
2870
SimplifyTruncInst(Value * Op,Type * Ty,const Query & Q,unsigned)2871 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
2872 if (Constant *C = dyn_cast<Constant>(Op))
2873 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI);
2874
2875 return nullptr;
2876 }
2877
SimplifyTruncInst(Value * Op,Type * Ty,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2878 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout *DL,
2879 const TargetLibraryInfo *TLI,
2880 const DominatorTree *DT) {
2881 return ::SimplifyTruncInst(Op, Ty, Query (DL, TLI, DT), RecursionLimit);
2882 }
2883
2884 //=== Helper functions for higher up the class hierarchy.
2885
2886 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
2887 /// fold the result. If not, this returns null.
SimplifyBinOp(unsigned Opcode,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)2888 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
2889 const Query &Q, unsigned MaxRecurse) {
2890 switch (Opcode) {
2891 case Instruction::Add:
2892 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2893 Q, MaxRecurse);
2894 case Instruction::FAdd:
2895 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
2896
2897 case Instruction::Sub:
2898 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2899 Q, MaxRecurse);
2900 case Instruction::FSub:
2901 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
2902
2903 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
2904 case Instruction::FMul:
2905 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
2906 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
2907 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
2908 case Instruction::FDiv: return SimplifyFDivInst(LHS, RHS, Q, MaxRecurse);
2909 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
2910 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
2911 case Instruction::FRem: return SimplifyFRemInst(LHS, RHS, Q, MaxRecurse);
2912 case Instruction::Shl:
2913 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
2914 Q, MaxRecurse);
2915 case Instruction::LShr:
2916 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
2917 case Instruction::AShr:
2918 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
2919 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
2920 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
2921 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
2922 default:
2923 if (Constant *CLHS = dyn_cast<Constant>(LHS))
2924 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
2925 Constant *COps[] = {CLHS, CRHS};
2926 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL,
2927 Q.TLI);
2928 }
2929
2930 // If the operation is associative, try some generic simplifications.
2931 if (Instruction::isAssociative(Opcode))
2932 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
2933 return V;
2934
2935 // If the operation is with the result of a select instruction check whether
2936 // operating on either branch of the select always yields the same value.
2937 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
2938 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
2939 return V;
2940
2941 // If the operation is with the result of a phi instruction, check whether
2942 // operating on all incoming values of the phi always yields the same value.
2943 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
2944 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
2945 return V;
2946
2947 return nullptr;
2948 }
2949 }
2950
SimplifyBinOp(unsigned Opcode,Value * LHS,Value * RHS,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2951 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
2952 const DataLayout *DL, const TargetLibraryInfo *TLI,
2953 const DominatorTree *DT) {
2954 return ::SimplifyBinOp(Opcode, LHS, RHS, Query (DL, TLI, DT), RecursionLimit);
2955 }
2956
2957 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
2958 /// fold the result.
SimplifyCmpInst(unsigned Predicate,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)2959 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2960 const Query &Q, unsigned MaxRecurse) {
2961 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
2962 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
2963 return SimplifyFCmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
2964 }
2965
SimplifyCmpInst(unsigned Predicate,Value * LHS,Value * RHS,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)2966 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2967 const DataLayout *DL, const TargetLibraryInfo *TLI,
2968 const DominatorTree *DT) {
2969 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query (DL, TLI, DT),
2970 RecursionLimit);
2971 }
2972
IsIdempotent(Intrinsic::ID ID)2973 static bool IsIdempotent(Intrinsic::ID ID) {
2974 switch (ID) {
2975 default: return false;
2976
2977 // Unary idempotent: f(f(x)) = f(x)
2978 case Intrinsic::fabs:
2979 case Intrinsic::floor:
2980 case Intrinsic::ceil:
2981 case Intrinsic::trunc:
2982 case Intrinsic::rint:
2983 case Intrinsic::nearbyint:
2984 case Intrinsic::round:
2985 return true;
2986 }
2987 }
2988
2989 template <typename IterTy>
SimplifyIntrinsic(Intrinsic::ID IID,IterTy ArgBegin,IterTy ArgEnd,const Query & Q,unsigned MaxRecurse)2990 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd,
2991 const Query &Q, unsigned MaxRecurse) {
2992 // Perform idempotent optimizations
2993 if (!IsIdempotent(IID))
2994 return nullptr;
2995
2996 // Unary Ops
2997 if (std::distance(ArgBegin, ArgEnd) == 1)
2998 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
2999 if (II->getIntrinsicID() == IID)
3000 return II;
3001
3002 return nullptr;
3003 }
3004
3005 template <typename IterTy>
SimplifyCall(Value * V,IterTy ArgBegin,IterTy ArgEnd,const Query & Q,unsigned MaxRecurse)3006 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
3007 const Query &Q, unsigned MaxRecurse) {
3008 Type *Ty = V->getType();
3009 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
3010 Ty = PTy->getElementType();
3011 FunctionType *FTy = cast<FunctionType>(Ty);
3012
3013 // call undef -> undef
3014 if (isa<UndefValue>(V))
3015 return UndefValue::get(FTy->getReturnType());
3016
3017 Function *F = dyn_cast<Function>(V);
3018 if (!F)
3019 return nullptr;
3020
3021 if (unsigned IID = F->getIntrinsicID())
3022 if (Value *Ret =
3023 SimplifyIntrinsic((Intrinsic::ID) IID, ArgBegin, ArgEnd, Q, MaxRecurse))
3024 return Ret;
3025
3026 if (!canConstantFoldCallTo(F))
3027 return nullptr;
3028
3029 SmallVector<Constant *, 4> ConstantArgs;
3030 ConstantArgs.reserve(ArgEnd - ArgBegin);
3031 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
3032 Constant *C = dyn_cast<Constant>(*I);
3033 if (!C)
3034 return nullptr;
3035 ConstantArgs.push_back(C);
3036 }
3037
3038 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
3039 }
3040
SimplifyCall(Value * V,User::op_iterator ArgBegin,User::op_iterator ArgEnd,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)3041 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
3042 User::op_iterator ArgEnd, const DataLayout *DL,
3043 const TargetLibraryInfo *TLI,
3044 const DominatorTree *DT) {
3045 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT),
3046 RecursionLimit);
3047 }
3048
SimplifyCall(Value * V,ArrayRef<Value * > Args,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)3049 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
3050 const DataLayout *DL, const TargetLibraryInfo *TLI,
3051 const DominatorTree *DT) {
3052 return ::SimplifyCall(V, Args.begin(), Args.end(), Query(DL, TLI, DT),
3053 RecursionLimit);
3054 }
3055
3056 /// SimplifyInstruction - See if we can compute a simplified version of this
3057 /// instruction. If not, this returns null.
SimplifyInstruction(Instruction * I,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)3058 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout *DL,
3059 const TargetLibraryInfo *TLI,
3060 const DominatorTree *DT) {
3061 Value *Result;
3062
3063 switch (I->getOpcode()) {
3064 default:
3065 Result = ConstantFoldInstruction(I, DL, TLI);
3066 break;
3067 case Instruction::FAdd:
3068 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
3069 I->getFastMathFlags(), DL, TLI, DT);
3070 break;
3071 case Instruction::Add:
3072 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
3073 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3074 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3075 DL, TLI, DT);
3076 break;
3077 case Instruction::FSub:
3078 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
3079 I->getFastMathFlags(), DL, TLI, DT);
3080 break;
3081 case Instruction::Sub:
3082 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
3083 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3084 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3085 DL, TLI, DT);
3086 break;
3087 case Instruction::FMul:
3088 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
3089 I->getFastMathFlags(), DL, TLI, DT);
3090 break;
3091 case Instruction::Mul:
3092 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3093 break;
3094 case Instruction::SDiv:
3095 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3096 break;
3097 case Instruction::UDiv:
3098 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3099 break;
3100 case Instruction::FDiv:
3101 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3102 break;
3103 case Instruction::SRem:
3104 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3105 break;
3106 case Instruction::URem:
3107 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3108 break;
3109 case Instruction::FRem:
3110 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3111 break;
3112 case Instruction::Shl:
3113 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
3114 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3115 cast<BinaryOperator>(I)->hasNoUnsignedWrap(),
3116 DL, TLI, DT);
3117 break;
3118 case Instruction::LShr:
3119 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
3120 cast<BinaryOperator>(I)->isExact(),
3121 DL, TLI, DT);
3122 break;
3123 case Instruction::AShr:
3124 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
3125 cast<BinaryOperator>(I)->isExact(),
3126 DL, TLI, DT);
3127 break;
3128 case Instruction::And:
3129 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3130 break;
3131 case Instruction::Or:
3132 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3133 break;
3134 case Instruction::Xor:
3135 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3136 break;
3137 case Instruction::ICmp:
3138 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
3139 I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3140 break;
3141 case Instruction::FCmp:
3142 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
3143 I->getOperand(0), I->getOperand(1), DL, TLI, DT);
3144 break;
3145 case Instruction::Select:
3146 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
3147 I->getOperand(2), DL, TLI, DT);
3148 break;
3149 case Instruction::GetElementPtr: {
3150 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
3151 Result = SimplifyGEPInst(Ops, DL, TLI, DT);
3152 break;
3153 }
3154 case Instruction::InsertValue: {
3155 InsertValueInst *IV = cast<InsertValueInst>(I);
3156 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
3157 IV->getInsertedValueOperand(),
3158 IV->getIndices(), DL, TLI, DT);
3159 break;
3160 }
3161 case Instruction::PHI:
3162 Result = SimplifyPHINode(cast<PHINode>(I), Query (DL, TLI, DT));
3163 break;
3164 case Instruction::Call: {
3165 CallSite CS(cast<CallInst>(I));
3166 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
3167 DL, TLI, DT);
3168 break;
3169 }
3170 case Instruction::Trunc:
3171 Result = SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT);
3172 break;
3173 }
3174
3175 /// If called on unreachable code, the above logic may report that the
3176 /// instruction simplified to itself. Make life easier for users by
3177 /// detecting that case here, returning a safe value instead.
3178 return Result == I ? UndefValue::get(I->getType()) : Result;
3179 }
3180
3181 /// \brief Implementation of recursive simplification through an instructions
3182 /// uses.
3183 ///
3184 /// This is the common implementation of the recursive simplification routines.
3185 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
3186 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
3187 /// instructions to process and attempt to simplify it using
3188 /// InstructionSimplify.
3189 ///
3190 /// This routine returns 'true' only when *it* simplifies something. The passed
3191 /// in simplified value does not count toward this.
replaceAndRecursivelySimplifyImpl(Instruction * I,Value * SimpleV,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)3192 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
3193 const DataLayout *DL,
3194 const TargetLibraryInfo *TLI,
3195 const DominatorTree *DT) {
3196 bool Simplified = false;
3197 SmallSetVector<Instruction *, 8> Worklist;
3198
3199 // If we have an explicit value to collapse to, do that round of the
3200 // simplification loop by hand initially.
3201 if (SimpleV) {
3202 for (User *U : I->users())
3203 if (U != I)
3204 Worklist.insert(cast<Instruction>(U));
3205
3206 // Replace the instruction with its simplified value.
3207 I->replaceAllUsesWith(SimpleV);
3208
3209 // Gracefully handle edge cases where the instruction is not wired into any
3210 // parent block.
3211 if (I->getParent())
3212 I->eraseFromParent();
3213 } else {
3214 Worklist.insert(I);
3215 }
3216
3217 // Note that we must test the size on each iteration, the worklist can grow.
3218 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
3219 I = Worklist[Idx];
3220
3221 // See if this instruction simplifies.
3222 SimpleV = SimplifyInstruction(I, DL, TLI, DT);
3223 if (!SimpleV)
3224 continue;
3225
3226 Simplified = true;
3227
3228 // Stash away all the uses of the old instruction so we can check them for
3229 // recursive simplifications after a RAUW. This is cheaper than checking all
3230 // uses of To on the recursive step in most cases.
3231 for (User *U : I->users())
3232 Worklist.insert(cast<Instruction>(U));
3233
3234 // Replace the instruction with its simplified value.
3235 I->replaceAllUsesWith(SimpleV);
3236
3237 // Gracefully handle edge cases where the instruction is not wired into any
3238 // parent block.
3239 if (I->getParent())
3240 I->eraseFromParent();
3241 }
3242 return Simplified;
3243 }
3244
recursivelySimplifyInstruction(Instruction * I,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)3245 bool llvm::recursivelySimplifyInstruction(Instruction *I,
3246 const DataLayout *DL,
3247 const TargetLibraryInfo *TLI,
3248 const DominatorTree *DT) {
3249 return replaceAndRecursivelySimplifyImpl(I, nullptr, DL, TLI, DT);
3250 }
3251
replaceAndRecursivelySimplify(Instruction * I,Value * SimpleV,const DataLayout * DL,const TargetLibraryInfo * TLI,const DominatorTree * DT)3252 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
3253 const DataLayout *DL,
3254 const TargetLibraryInfo *TLI,
3255 const DominatorTree *DT) {
3256 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
3257 assert(SimpleV && "Must provide a simplified value.");
3258 return replaceAndRecursivelySimplifyImpl(I, SimpleV, DL, TLI, DT);
3259 }
3260