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/AliasAnalysis.h"
24 #include "llvm/Analysis/ConstantFolding.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/Analysis/VectorUtils.h"
28 #include "llvm/IR/ConstantRange.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/GlobalAlias.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/ValueHandle.h"
36 #include <algorithm>
37 using namespace llvm;
38 using namespace llvm::PatternMatch;
39
40 #define DEBUG_TYPE "instsimplify"
41
42 enum { RecursionLimit = 3 };
43
44 STATISTIC(NumExpand, "Number of expansions");
45 STATISTIC(NumReassoc, "Number of reassociations");
46
47 namespace {
48 struct Query {
49 const DataLayout &DL;
50 const TargetLibraryInfo *TLI;
51 const DominatorTree *DT;
52 AssumptionCache *AC;
53 const Instruction *CxtI;
54
Query__anona24799d40211::Query55 Query(const DataLayout &DL, const TargetLibraryInfo *tli,
56 const DominatorTree *dt, AssumptionCache *ac = nullptr,
57 const Instruction *cxti = nullptr)
58 : DL(DL), TLI(tli), DT(dt), AC(ac), CxtI(cxti) {}
59 };
60 } // end anonymous namespace
61
62 static Value *SimplifyAndInst(Value *, Value *, const Query &, unsigned);
63 static Value *SimplifyBinOp(unsigned, Value *, Value *, const Query &,
64 unsigned);
65 static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
66 const Query &, unsigned);
67 static Value *SimplifyCmpInst(unsigned, Value *, Value *, const Query &,
68 unsigned);
69 static Value *SimplifyOrInst(Value *, Value *, const Query &, unsigned);
70 static Value *SimplifyXorInst(Value *, Value *, const Query &, unsigned);
71 static Value *SimplifyTruncInst(Value *, Type *, const Query &, unsigned);
72
73 /// getFalse - For a boolean type, or a vector of boolean type, return false, or
74 /// a vector with every element false, as appropriate for the type.
getFalse(Type * Ty)75 static Constant *getFalse(Type *Ty) {
76 assert(Ty->getScalarType()->isIntegerTy(1) &&
77 "Expected i1 type or a vector of i1!");
78 return Constant::getNullValue(Ty);
79 }
80
81 /// getTrue - For a boolean type, or a vector of boolean type, return true, or
82 /// a vector with every element true, as appropriate for the type.
getTrue(Type * Ty)83 static Constant *getTrue(Type *Ty) {
84 assert(Ty->getScalarType()->isIntegerTy(1) &&
85 "Expected i1 type or a vector of i1!");
86 return Constant::getAllOnesValue(Ty);
87 }
88
89 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
isSameCompare(Value * V,CmpInst::Predicate Pred,Value * LHS,Value * RHS)90 static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
91 Value *RHS) {
92 CmpInst *Cmp = dyn_cast<CmpInst>(V);
93 if (!Cmp)
94 return false;
95 CmpInst::Predicate CPred = Cmp->getPredicate();
96 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
97 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
98 return true;
99 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
100 CRHS == LHS;
101 }
102
103 /// ValueDominatesPHI - Does the given value dominate the specified phi node?
ValueDominatesPHI(Value * V,PHINode * P,const DominatorTree * DT)104 static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
105 Instruction *I = dyn_cast<Instruction>(V);
106 if (!I)
107 // Arguments and constants dominate all instructions.
108 return true;
109
110 // If we are processing instructions (and/or basic blocks) that have not been
111 // fully added to a function, the parent nodes may still be null. Simply
112 // return the conservative answer in these cases.
113 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
114 return false;
115
116 // If we have a DominatorTree then do a precise test.
117 if (DT) {
118 if (!DT->isReachableFromEntry(P->getParent()))
119 return true;
120 if (!DT->isReachableFromEntry(I->getParent()))
121 return false;
122 return DT->dominates(I, P);
123 }
124
125 // Otherwise, if the instruction is in the entry block and is not an invoke,
126 // then it obviously dominates all phi nodes.
127 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
128 !isa<InvokeInst>(I))
129 return true;
130
131 return false;
132 }
133
134 /// ExpandBinOp - Simplify "A op (B op' C)" by distributing op over op', turning
135 /// it into "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
136 /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
137 /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
138 /// 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)139 static Value *ExpandBinOp(unsigned Opcode, Value *LHS, Value *RHS,
140 unsigned OpcToExpand, const Query &Q,
141 unsigned MaxRecurse) {
142 Instruction::BinaryOps OpcodeToExpand = (Instruction::BinaryOps)OpcToExpand;
143 // Recursion is always used, so bail out at once if we already hit the limit.
144 if (!MaxRecurse--)
145 return nullptr;
146
147 // Check whether the expression has the form "(A op' B) op C".
148 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
149 if (Op0->getOpcode() == OpcodeToExpand) {
150 // It does! Try turning it into "(A op C) op' (B op C)".
151 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
152 // Do "A op C" and "B op C" both simplify?
153 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
154 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
155 // They do! Return "L op' R" if it simplifies or is already available.
156 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
157 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
158 && L == B && R == A)) {
159 ++NumExpand;
160 return LHS;
161 }
162 // Otherwise return "L op' R" if it simplifies.
163 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
164 ++NumExpand;
165 return V;
166 }
167 }
168 }
169
170 // Check whether the expression has the form "A op (B op' C)".
171 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
172 if (Op1->getOpcode() == OpcodeToExpand) {
173 // It does! Try turning it into "(A op B) op' (A op C)".
174 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
175 // Do "A op B" and "A op C" both simplify?
176 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
177 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
178 // They do! Return "L op' R" if it simplifies or is already available.
179 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
180 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
181 && L == C && R == B)) {
182 ++NumExpand;
183 return RHS;
184 }
185 // Otherwise return "L op' R" if it simplifies.
186 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
187 ++NumExpand;
188 return V;
189 }
190 }
191 }
192
193 return nullptr;
194 }
195
196 /// SimplifyAssociativeBinOp - Generic simplifications for associative binary
197 /// operations. Returns the simpler value, or null if none was found.
SimplifyAssociativeBinOp(unsigned Opc,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)198 static Value *SimplifyAssociativeBinOp(unsigned Opc, Value *LHS, Value *RHS,
199 const Query &Q, unsigned MaxRecurse) {
200 Instruction::BinaryOps Opcode = (Instruction::BinaryOps)Opc;
201 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
202
203 // Recursion is always used, so bail out at once if we already hit the limit.
204 if (!MaxRecurse--)
205 return nullptr;
206
207 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
208 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
209
210 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
211 if (Op0 && Op0->getOpcode() == Opcode) {
212 Value *A = Op0->getOperand(0);
213 Value *B = Op0->getOperand(1);
214 Value *C = RHS;
215
216 // Does "B op C" simplify?
217 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
218 // It does! Return "A op V" if it simplifies or is already available.
219 // If V equals B then "A op V" is just the LHS.
220 if (V == B) return LHS;
221 // Otherwise return "A op V" if it simplifies.
222 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
223 ++NumReassoc;
224 return W;
225 }
226 }
227 }
228
229 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
230 if (Op1 && Op1->getOpcode() == Opcode) {
231 Value *A = LHS;
232 Value *B = Op1->getOperand(0);
233 Value *C = Op1->getOperand(1);
234
235 // Does "A op B" simplify?
236 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
237 // It does! Return "V op C" if it simplifies or is already available.
238 // If V equals B then "V op C" is just the RHS.
239 if (V == B) return RHS;
240 // Otherwise return "V op C" if it simplifies.
241 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
242 ++NumReassoc;
243 return W;
244 }
245 }
246 }
247
248 // The remaining transforms require commutativity as well as associativity.
249 if (!Instruction::isCommutative(Opcode))
250 return nullptr;
251
252 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
253 if (Op0 && Op0->getOpcode() == Opcode) {
254 Value *A = Op0->getOperand(0);
255 Value *B = Op0->getOperand(1);
256 Value *C = RHS;
257
258 // Does "C op A" simplify?
259 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
260 // It does! Return "V op B" if it simplifies or is already available.
261 // If V equals A then "V op B" is just the LHS.
262 if (V == A) return LHS;
263 // Otherwise return "V op B" if it simplifies.
264 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
265 ++NumReassoc;
266 return W;
267 }
268 }
269 }
270
271 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
272 if (Op1 && Op1->getOpcode() == Opcode) {
273 Value *A = LHS;
274 Value *B = Op1->getOperand(0);
275 Value *C = Op1->getOperand(1);
276
277 // Does "C op A" simplify?
278 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
279 // It does! Return "B op V" if it simplifies or is already available.
280 // If V equals C then "B op V" is just the RHS.
281 if (V == C) return RHS;
282 // Otherwise return "B op V" if it simplifies.
283 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
284 ++NumReassoc;
285 return W;
286 }
287 }
288 }
289
290 return nullptr;
291 }
292
293 /// ThreadBinOpOverSelect - In the case of a binary operation with a select
294 /// instruction as an operand, try to simplify the binop by seeing whether
295 /// evaluating it on both branches of the select results in the same value.
296 /// Returns the common value if so, otherwise returns null.
ThreadBinOpOverSelect(unsigned Opcode,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)297 static Value *ThreadBinOpOverSelect(unsigned Opcode, Value *LHS, Value *RHS,
298 const Query &Q, unsigned MaxRecurse) {
299 // Recursion is always used, so bail out at once if we already hit the limit.
300 if (!MaxRecurse--)
301 return nullptr;
302
303 SelectInst *SI;
304 if (isa<SelectInst>(LHS)) {
305 SI = cast<SelectInst>(LHS);
306 } else {
307 assert(isa<SelectInst>(RHS) && "No select instruction operand!");
308 SI = cast<SelectInst>(RHS);
309 }
310
311 // Evaluate the BinOp on the true and false branches of the select.
312 Value *TV;
313 Value *FV;
314 if (SI == LHS) {
315 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
316 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
317 } else {
318 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
319 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
320 }
321
322 // If they simplified to the same value, then return the common value.
323 // If they both failed to simplify then return null.
324 if (TV == FV)
325 return TV;
326
327 // If one branch simplified to undef, return the other one.
328 if (TV && isa<UndefValue>(TV))
329 return FV;
330 if (FV && isa<UndefValue>(FV))
331 return TV;
332
333 // If applying the operation did not change the true and false select values,
334 // then the result of the binop is the select itself.
335 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
336 return SI;
337
338 // If one branch simplified and the other did not, and the simplified
339 // value is equal to the unsimplified one, return the simplified value.
340 // For example, select (cond, X, X & Z) & Z -> X & Z.
341 if ((FV && !TV) || (TV && !FV)) {
342 // Check that the simplified value has the form "X op Y" where "op" is the
343 // same as the original operation.
344 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
345 if (Simplified && Simplified->getOpcode() == Opcode) {
346 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
347 // We already know that "op" is the same as for the simplified value. See
348 // if the operands match too. If so, return the simplified value.
349 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
350 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
351 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
352 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
353 Simplified->getOperand(1) == UnsimplifiedRHS)
354 return Simplified;
355 if (Simplified->isCommutative() &&
356 Simplified->getOperand(1) == UnsimplifiedLHS &&
357 Simplified->getOperand(0) == UnsimplifiedRHS)
358 return Simplified;
359 }
360 }
361
362 return nullptr;
363 }
364
365 /// ThreadCmpOverSelect - In the case of a comparison with a select instruction,
366 /// try to simplify the comparison by seeing whether both branches of the select
367 /// result in the same value. Returns the common value if so, otherwise returns
368 /// null.
ThreadCmpOverSelect(CmpInst::Predicate Pred,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)369 static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
370 Value *RHS, const Query &Q,
371 unsigned MaxRecurse) {
372 // Recursion is always used, so bail out at once if we already hit the limit.
373 if (!MaxRecurse--)
374 return nullptr;
375
376 // Make sure the select is on the LHS.
377 if (!isa<SelectInst>(LHS)) {
378 std::swap(LHS, RHS);
379 Pred = CmpInst::getSwappedPredicate(Pred);
380 }
381 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
382 SelectInst *SI = cast<SelectInst>(LHS);
383 Value *Cond = SI->getCondition();
384 Value *TV = SI->getTrueValue();
385 Value *FV = SI->getFalseValue();
386
387 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
388 // Does "cmp TV, RHS" simplify?
389 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
390 if (TCmp == Cond) {
391 // It not only simplified, it simplified to the select condition. Replace
392 // it with 'true'.
393 TCmp = getTrue(Cond->getType());
394 } else if (!TCmp) {
395 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
396 // condition then we can replace it with 'true'. Otherwise give up.
397 if (!isSameCompare(Cond, Pred, TV, RHS))
398 return nullptr;
399 TCmp = getTrue(Cond->getType());
400 }
401
402 // Does "cmp FV, RHS" simplify?
403 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
404 if (FCmp == Cond) {
405 // It not only simplified, it simplified to the select condition. Replace
406 // it with 'false'.
407 FCmp = getFalse(Cond->getType());
408 } else if (!FCmp) {
409 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
410 // condition then we can replace it with 'false'. Otherwise give up.
411 if (!isSameCompare(Cond, Pred, FV, RHS))
412 return nullptr;
413 FCmp = getFalse(Cond->getType());
414 }
415
416 // If both sides simplified to the same value, then use it as the result of
417 // the original comparison.
418 if (TCmp == FCmp)
419 return TCmp;
420
421 // The remaining cases only make sense if the select condition has the same
422 // type as the result of the comparison, so bail out if this is not so.
423 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
424 return nullptr;
425 // If the false value simplified to false, then the result of the compare
426 // is equal to "Cond && TCmp". This also catches the case when the false
427 // value simplified to false and the true value to true, returning "Cond".
428 if (match(FCmp, m_Zero()))
429 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
430 return V;
431 // If the true value simplified to true, then the result of the compare
432 // is equal to "Cond || FCmp".
433 if (match(TCmp, m_One()))
434 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
435 return V;
436 // Finally, if the false value simplified to true and the true value to
437 // false, then the result of the compare is equal to "!Cond".
438 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
439 if (Value *V =
440 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
441 Q, MaxRecurse))
442 return V;
443
444 return nullptr;
445 }
446
447 /// ThreadBinOpOverPHI - In the case of a binary operation with an operand that
448 /// is a PHI instruction, try to simplify the binop by seeing whether evaluating
449 /// it on the incoming phi values yields the same result for every value. If so
450 /// returns the common value, otherwise returns null.
ThreadBinOpOverPHI(unsigned Opcode,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)451 static Value *ThreadBinOpOverPHI(unsigned Opcode, Value *LHS, Value *RHS,
452 const Query &Q, unsigned MaxRecurse) {
453 // Recursion is always used, so bail out at once if we already hit the limit.
454 if (!MaxRecurse--)
455 return nullptr;
456
457 PHINode *PI;
458 if (isa<PHINode>(LHS)) {
459 PI = cast<PHINode>(LHS);
460 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
461 if (!ValueDominatesPHI(RHS, PI, Q.DT))
462 return nullptr;
463 } else {
464 assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
465 PI = cast<PHINode>(RHS);
466 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
467 if (!ValueDominatesPHI(LHS, PI, Q.DT))
468 return nullptr;
469 }
470
471 // Evaluate the BinOp on the incoming phi values.
472 Value *CommonValue = nullptr;
473 for (Value *Incoming : PI->incoming_values()) {
474 // If the incoming value is the phi node itself, it can safely be skipped.
475 if (Incoming == PI) continue;
476 Value *V = PI == LHS ?
477 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
478 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
479 // If the operation failed to simplify, or simplified to a different value
480 // to previously, then give up.
481 if (!V || (CommonValue && V != CommonValue))
482 return nullptr;
483 CommonValue = V;
484 }
485
486 return CommonValue;
487 }
488
489 /// ThreadCmpOverPHI - In the case of a comparison with a PHI instruction, try
490 /// try to simplify the comparison by seeing whether comparing with all of the
491 /// incoming phi values yields the same result every time. If so returns the
492 /// common result, otherwise returns null.
ThreadCmpOverPHI(CmpInst::Predicate Pred,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)493 static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
494 const Query &Q, unsigned MaxRecurse) {
495 // Recursion is always used, so bail out at once if we already hit the limit.
496 if (!MaxRecurse--)
497 return nullptr;
498
499 // Make sure the phi is on the LHS.
500 if (!isa<PHINode>(LHS)) {
501 std::swap(LHS, RHS);
502 Pred = CmpInst::getSwappedPredicate(Pred);
503 }
504 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
505 PHINode *PI = cast<PHINode>(LHS);
506
507 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
508 if (!ValueDominatesPHI(RHS, PI, Q.DT))
509 return nullptr;
510
511 // Evaluate the BinOp on the incoming phi values.
512 Value *CommonValue = nullptr;
513 for (Value *Incoming : PI->incoming_values()) {
514 // If the incoming value is the phi node itself, it can safely be skipped.
515 if (Incoming == PI) continue;
516 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
517 // If the operation failed to simplify, or simplified to a different value
518 // to previously, then give up.
519 if (!V || (CommonValue && V != CommonValue))
520 return nullptr;
521 CommonValue = V;
522 }
523
524 return CommonValue;
525 }
526
527 /// SimplifyAddInst - Given operands for an Add, see if we can
528 /// fold the result. If not, this returns null.
SimplifyAddInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const Query & Q,unsigned MaxRecurse)529 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
530 const Query &Q, unsigned MaxRecurse) {
531 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
532 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
533 Constant *Ops[] = { CLHS, CRHS };
534 return ConstantFoldInstOperands(Instruction::Add, CLHS->getType(), Ops,
535 Q.DL, Q.TLI);
536 }
537
538 // Canonicalize the constant to the RHS.
539 std::swap(Op0, Op1);
540 }
541
542 // X + undef -> undef
543 if (match(Op1, m_Undef()))
544 return Op1;
545
546 // X + 0 -> X
547 if (match(Op1, m_Zero()))
548 return Op0;
549
550 // X + (Y - X) -> Y
551 // (Y - X) + X -> Y
552 // Eg: X + -X -> 0
553 Value *Y = nullptr;
554 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
555 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
556 return Y;
557
558 // X + ~X -> -1 since ~X = -X-1
559 if (match(Op0, m_Not(m_Specific(Op1))) ||
560 match(Op1, m_Not(m_Specific(Op0))))
561 return Constant::getAllOnesValue(Op0->getType());
562
563 /// i1 add -> xor.
564 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
565 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
566 return V;
567
568 // Try some generic simplifications for associative operations.
569 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
570 MaxRecurse))
571 return V;
572
573 // Threading Add over selects and phi nodes is pointless, so don't bother.
574 // Threading over the select in "A + select(cond, B, C)" means evaluating
575 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
576 // only if B and C are equal. If B and C are equal then (since we assume
577 // that operands have already been simplified) "select(cond, B, C)" should
578 // have been simplified to the common value of B and C already. Analysing
579 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
580 // for threading over phi nodes.
581
582 return nullptr;
583 }
584
SimplifyAddInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)585 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
586 const DataLayout &DL, const TargetLibraryInfo *TLI,
587 const DominatorTree *DT, AssumptionCache *AC,
588 const Instruction *CxtI) {
589 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
590 RecursionLimit);
591 }
592
593 /// \brief Compute the base pointer and cumulative constant offsets for V.
594 ///
595 /// This strips all constant offsets off of V, leaving it the base pointer, and
596 /// accumulates the total constant offset applied in the returned constant. It
597 /// returns 0 if V is not a pointer, and returns the constant '0' if there are
598 /// no constant offsets applied.
599 ///
600 /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
601 /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
602 /// folding.
stripAndComputeConstantOffsets(const DataLayout & DL,Value * & V,bool AllowNonInbounds=false)603 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
604 bool AllowNonInbounds = false) {
605 assert(V->getType()->getScalarType()->isPointerTy());
606
607 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
608 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
609
610 // Even though we don't look through PHI nodes, we could be called on an
611 // instruction in an unreachable block, which may be on a cycle.
612 SmallPtrSet<Value *, 4> Visited;
613 Visited.insert(V);
614 do {
615 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
616 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
617 !GEP->accumulateConstantOffset(DL, Offset))
618 break;
619 V = GEP->getPointerOperand();
620 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
621 V = cast<Operator>(V)->getOperand(0);
622 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
623 if (GA->mayBeOverridden())
624 break;
625 V = GA->getAliasee();
626 } else {
627 break;
628 }
629 assert(V->getType()->getScalarType()->isPointerTy() &&
630 "Unexpected operand type!");
631 } while (Visited.insert(V).second);
632
633 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
634 if (V->getType()->isVectorTy())
635 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
636 OffsetIntPtr);
637 return OffsetIntPtr;
638 }
639
640 /// \brief Compute the constant difference between two pointer values.
641 /// If the difference is not a constant, returns zero.
computePointerDifference(const DataLayout & DL,Value * LHS,Value * RHS)642 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
643 Value *RHS) {
644 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
645 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
646
647 // If LHS and RHS are not related via constant offsets to the same base
648 // value, there is nothing we can do here.
649 if (LHS != RHS)
650 return nullptr;
651
652 // Otherwise, the difference of LHS - RHS can be computed as:
653 // LHS - RHS
654 // = (LHSOffset + Base) - (RHSOffset + Base)
655 // = LHSOffset - RHSOffset
656 return ConstantExpr::getSub(LHSOffset, RHSOffset);
657 }
658
659 /// SimplifySubInst - Given operands for a Sub, see if we can
660 /// fold the result. If not, this returns null.
SimplifySubInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const Query & Q,unsigned MaxRecurse)661 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
662 const Query &Q, unsigned MaxRecurse) {
663 if (Constant *CLHS = dyn_cast<Constant>(Op0))
664 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
665 Constant *Ops[] = { CLHS, CRHS };
666 return ConstantFoldInstOperands(Instruction::Sub, CLHS->getType(),
667 Ops, Q.DL, Q.TLI);
668 }
669
670 // X - undef -> undef
671 // undef - X -> undef
672 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
673 return UndefValue::get(Op0->getType());
674
675 // X - 0 -> X
676 if (match(Op1, m_Zero()))
677 return Op0;
678
679 // X - X -> 0
680 if (Op0 == Op1)
681 return Constant::getNullValue(Op0->getType());
682
683 // 0 - X -> 0 if the sub is NUW.
684 if (isNUW && match(Op0, m_Zero()))
685 return Op0;
686
687 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
688 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
689 Value *X = nullptr, *Y = nullptr, *Z = Op1;
690 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
691 // See if "V === Y - Z" simplifies.
692 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
693 // It does! Now see if "X + V" simplifies.
694 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
695 // It does, we successfully reassociated!
696 ++NumReassoc;
697 return W;
698 }
699 // See if "V === X - Z" simplifies.
700 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
701 // It does! Now see if "Y + V" simplifies.
702 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
703 // It does, we successfully reassociated!
704 ++NumReassoc;
705 return W;
706 }
707 }
708
709 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
710 // For example, X - (X + 1) -> -1
711 X = Op0;
712 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
713 // See if "V === X - Y" simplifies.
714 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
715 // It does! Now see if "V - Z" simplifies.
716 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
717 // It does, we successfully reassociated!
718 ++NumReassoc;
719 return W;
720 }
721 // See if "V === X - Z" simplifies.
722 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
723 // It does! Now see if "V - Y" simplifies.
724 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
725 // It does, we successfully reassociated!
726 ++NumReassoc;
727 return W;
728 }
729 }
730
731 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
732 // For example, X - (X - Y) -> Y.
733 Z = Op0;
734 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
735 // See if "V === Z - X" simplifies.
736 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
737 // It does! Now see if "V + Y" simplifies.
738 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
739 // It does, we successfully reassociated!
740 ++NumReassoc;
741 return W;
742 }
743
744 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
745 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
746 match(Op1, m_Trunc(m_Value(Y))))
747 if (X->getType() == Y->getType())
748 // See if "V === X - Y" simplifies.
749 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
750 // It does! Now see if "trunc V" simplifies.
751 if (Value *W = SimplifyTruncInst(V, Op0->getType(), Q, MaxRecurse-1))
752 // It does, return the simplified "trunc V".
753 return W;
754
755 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
756 if (match(Op0, m_PtrToInt(m_Value(X))) &&
757 match(Op1, m_PtrToInt(m_Value(Y))))
758 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
759 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
760
761 // i1 sub -> xor.
762 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
763 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
764 return V;
765
766 // Threading Sub over selects and phi nodes is pointless, so don't bother.
767 // Threading over the select in "A - select(cond, B, C)" means evaluating
768 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
769 // only if B and C are equal. If B and C are equal then (since we assume
770 // that operands have already been simplified) "select(cond, B, C)" should
771 // have been simplified to the common value of B and C already. Analysing
772 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
773 // for threading over phi nodes.
774
775 return nullptr;
776 }
777
SimplifySubInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)778 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
779 const DataLayout &DL, const TargetLibraryInfo *TLI,
780 const DominatorTree *DT, AssumptionCache *AC,
781 const Instruction *CxtI) {
782 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
783 RecursionLimit);
784 }
785
786 /// Given operands for an FAdd, see if we can fold the result. If not, this
787 /// returns null.
SimplifyFAddInst(Value * Op0,Value * Op1,FastMathFlags FMF,const Query & Q,unsigned MaxRecurse)788 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
789 const Query &Q, unsigned MaxRecurse) {
790 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
791 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
792 Constant *Ops[] = { CLHS, CRHS };
793 return ConstantFoldInstOperands(Instruction::FAdd, CLHS->getType(),
794 Ops, Q.DL, Q.TLI);
795 }
796
797 // Canonicalize the constant to the RHS.
798 std::swap(Op0, Op1);
799 }
800
801 // fadd X, -0 ==> X
802 if (match(Op1, m_NegZero()))
803 return Op0;
804
805 // fadd X, 0 ==> X, when we know X is not -0
806 if (match(Op1, m_Zero()) &&
807 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
808 return Op0;
809
810 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
811 // where nnan and ninf have to occur at least once somewhere in this
812 // expression
813 Value *SubOp = nullptr;
814 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
815 SubOp = Op1;
816 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
817 SubOp = Op0;
818 if (SubOp) {
819 Instruction *FSub = cast<Instruction>(SubOp);
820 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
821 (FMF.noInfs() || FSub->hasNoInfs()))
822 return Constant::getNullValue(Op0->getType());
823 }
824
825 return nullptr;
826 }
827
828 /// Given operands for an FSub, see if we can fold the result. If not, this
829 /// returns null.
SimplifyFSubInst(Value * Op0,Value * Op1,FastMathFlags FMF,const Query & Q,unsigned MaxRecurse)830 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
831 const Query &Q, unsigned MaxRecurse) {
832 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
833 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
834 Constant *Ops[] = { CLHS, CRHS };
835 return ConstantFoldInstOperands(Instruction::FSub, CLHS->getType(),
836 Ops, Q.DL, Q.TLI);
837 }
838 }
839
840 // fsub X, 0 ==> X
841 if (match(Op1, m_Zero()))
842 return Op0;
843
844 // fsub X, -0 ==> X, when we know X is not -0
845 if (match(Op1, m_NegZero()) &&
846 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0)))
847 return Op0;
848
849 // fsub 0, (fsub -0.0, X) ==> X
850 Value *X;
851 if (match(Op0, m_AnyZero())) {
852 if (match(Op1, m_FSub(m_NegZero(), m_Value(X))))
853 return X;
854 if (FMF.noSignedZeros() && match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
855 return X;
856 }
857
858 // fsub nnan x, x ==> 0.0
859 if (FMF.noNaNs() && Op0 == Op1)
860 return Constant::getNullValue(Op0->getType());
861
862 return nullptr;
863 }
864
865 /// 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)866 static Value *SimplifyFMulInst(Value *Op0, Value *Op1,
867 FastMathFlags FMF,
868 const Query &Q,
869 unsigned MaxRecurse) {
870 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
871 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
872 Constant *Ops[] = { CLHS, CRHS };
873 return ConstantFoldInstOperands(Instruction::FMul, CLHS->getType(),
874 Ops, Q.DL, Q.TLI);
875 }
876
877 // Canonicalize the constant to the RHS.
878 std::swap(Op0, Op1);
879 }
880
881 // fmul X, 1.0 ==> X
882 if (match(Op1, m_FPOne()))
883 return Op0;
884
885 // fmul nnan nsz X, 0 ==> 0
886 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
887 return Op1;
888
889 return nullptr;
890 }
891
892 /// SimplifyMulInst - Given operands for a Mul, see if we can
893 /// fold the result. If not, this returns null.
SimplifyMulInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)894 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const Query &Q,
895 unsigned MaxRecurse) {
896 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
897 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
898 Constant *Ops[] = { CLHS, CRHS };
899 return ConstantFoldInstOperands(Instruction::Mul, CLHS->getType(),
900 Ops, Q.DL, Q.TLI);
901 }
902
903 // Canonicalize the constant to the RHS.
904 std::swap(Op0, Op1);
905 }
906
907 // X * undef -> 0
908 if (match(Op1, m_Undef()))
909 return Constant::getNullValue(Op0->getType());
910
911 // X * 0 -> 0
912 if (match(Op1, m_Zero()))
913 return Op1;
914
915 // X * 1 -> X
916 if (match(Op1, m_One()))
917 return Op0;
918
919 // (X / Y) * Y -> X if the division is exact.
920 Value *X = nullptr;
921 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
922 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
923 return X;
924
925 // i1 mul -> and.
926 if (MaxRecurse && Op0->getType()->isIntegerTy(1))
927 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
928 return V;
929
930 // Try some generic simplifications for associative operations.
931 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
932 MaxRecurse))
933 return V;
934
935 // Mul distributes over Add. Try some generic simplifications based on this.
936 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
937 Q, MaxRecurse))
938 return V;
939
940 // If the operation is with the result of a select instruction, check whether
941 // operating on either branch of the select always yields the same value.
942 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
943 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
944 MaxRecurse))
945 return V;
946
947 // If the operation is with the result of a phi instruction, check whether
948 // operating on all incoming values of the phi always yields the same value.
949 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
950 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
951 MaxRecurse))
952 return V;
953
954 return nullptr;
955 }
956
SimplifyFAddInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)957 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
958 const DataLayout &DL,
959 const TargetLibraryInfo *TLI,
960 const DominatorTree *DT, AssumptionCache *AC,
961 const Instruction *CxtI) {
962 return ::SimplifyFAddInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
963 RecursionLimit);
964 }
965
SimplifyFSubInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)966 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
967 const DataLayout &DL,
968 const TargetLibraryInfo *TLI,
969 const DominatorTree *DT, AssumptionCache *AC,
970 const Instruction *CxtI) {
971 return ::SimplifyFSubInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
972 RecursionLimit);
973 }
974
SimplifyFMulInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)975 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
976 const DataLayout &DL,
977 const TargetLibraryInfo *TLI,
978 const DominatorTree *DT, AssumptionCache *AC,
979 const Instruction *CxtI) {
980 return ::SimplifyFMulInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
981 RecursionLimit);
982 }
983
SimplifyMulInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)984 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const DataLayout &DL,
985 const TargetLibraryInfo *TLI,
986 const DominatorTree *DT, AssumptionCache *AC,
987 const Instruction *CxtI) {
988 return ::SimplifyMulInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
989 RecursionLimit);
990 }
991
992 /// SimplifyDiv - Given operands for an SDiv or UDiv, see if we can
993 /// fold the result. If not, this returns null.
SimplifyDiv(Instruction::BinaryOps Opcode,Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)994 static Value *SimplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
995 const Query &Q, unsigned MaxRecurse) {
996 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
997 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
998 Constant *Ops[] = { C0, C1 };
999 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1000 }
1001 }
1002
1003 bool isSigned = Opcode == Instruction::SDiv;
1004
1005 // X / undef -> undef
1006 if (match(Op1, m_Undef()))
1007 return Op1;
1008
1009 // X / 0 -> undef, we don't need to preserve faults!
1010 if (match(Op1, m_Zero()))
1011 return UndefValue::get(Op1->getType());
1012
1013 // undef / X -> 0
1014 if (match(Op0, m_Undef()))
1015 return Constant::getNullValue(Op0->getType());
1016
1017 // 0 / X -> 0, we don't need to preserve faults!
1018 if (match(Op0, m_Zero()))
1019 return Op0;
1020
1021 // X / 1 -> X
1022 if (match(Op1, m_One()))
1023 return Op0;
1024
1025 if (Op0->getType()->isIntegerTy(1))
1026 // It can't be division by zero, hence it must be division by one.
1027 return Op0;
1028
1029 // X / X -> 1
1030 if (Op0 == Op1)
1031 return ConstantInt::get(Op0->getType(), 1);
1032
1033 // (X * Y) / Y -> X if the multiplication does not overflow.
1034 Value *X = nullptr, *Y = nullptr;
1035 if (match(Op0, m_Mul(m_Value(X), m_Value(Y))) && (X == Op1 || Y == Op1)) {
1036 if (Y != Op1) std::swap(X, Y); // Ensure expression is (X * Y) / Y, Y = Op1
1037 OverflowingBinaryOperator *Mul = cast<OverflowingBinaryOperator>(Op0);
1038 // If the Mul knows it does not overflow, then we are good to go.
1039 if ((isSigned && Mul->hasNoSignedWrap()) ||
1040 (!isSigned && Mul->hasNoUnsignedWrap()))
1041 return X;
1042 // If X has the form X = A / Y then X * Y cannot overflow.
1043 if (BinaryOperator *Div = dyn_cast<BinaryOperator>(X))
1044 if (Div->getOpcode() == Opcode && Div->getOperand(1) == Y)
1045 return X;
1046 }
1047
1048 // (X rem Y) / Y -> 0
1049 if ((isSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1050 (!isSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1051 return Constant::getNullValue(Op0->getType());
1052
1053 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1054 ConstantInt *C1, *C2;
1055 if (!isSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1056 match(Op1, m_ConstantInt(C2))) {
1057 bool Overflow;
1058 C1->getValue().umul_ov(C2->getValue(), Overflow);
1059 if (Overflow)
1060 return Constant::getNullValue(Op0->getType());
1061 }
1062
1063 // If the operation is with the result of a select instruction, check whether
1064 // operating on either branch of the select always yields the same value.
1065 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1066 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1067 return V;
1068
1069 // If the operation is with the result of a phi instruction, check whether
1070 // operating on all incoming values of the phi always yields the same value.
1071 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1072 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1073 return V;
1074
1075 return nullptr;
1076 }
1077
1078 /// SimplifySDivInst - Given operands for an SDiv, see if we can
1079 /// fold the result. If not, this returns null.
SimplifySDivInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1080 static Value *SimplifySDivInst(Value *Op0, Value *Op1, const Query &Q,
1081 unsigned MaxRecurse) {
1082 if (Value *V = SimplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse))
1083 return V;
1084
1085 return nullptr;
1086 }
1087
SimplifySDivInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1088 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1089 const TargetLibraryInfo *TLI,
1090 const DominatorTree *DT, AssumptionCache *AC,
1091 const Instruction *CxtI) {
1092 return ::SimplifySDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1093 RecursionLimit);
1094 }
1095
1096 /// SimplifyUDivInst - Given operands for a UDiv, see if we can
1097 /// fold the result. If not, this returns null.
SimplifyUDivInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1098 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const Query &Q,
1099 unsigned MaxRecurse) {
1100 if (Value *V = SimplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse))
1101 return V;
1102
1103 return nullptr;
1104 }
1105
SimplifyUDivInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1106 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const DataLayout &DL,
1107 const TargetLibraryInfo *TLI,
1108 const DominatorTree *DT, AssumptionCache *AC,
1109 const Instruction *CxtI) {
1110 return ::SimplifyUDivInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1111 RecursionLimit);
1112 }
1113
SimplifyFDivInst(Value * Op0,Value * Op1,FastMathFlags FMF,const Query & Q,unsigned)1114 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1115 const Query &Q, unsigned) {
1116 // undef / X -> undef (the undef could be a snan).
1117 if (match(Op0, m_Undef()))
1118 return Op0;
1119
1120 // X / undef -> undef
1121 if (match(Op1, m_Undef()))
1122 return Op1;
1123
1124 // 0 / X -> 0
1125 // Requires that NaNs are off (X could be zero) and signed zeroes are
1126 // ignored (X could be positive or negative, so the output sign is unknown).
1127 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1128 return Op0;
1129
1130 if (FMF.noNaNs()) {
1131 // X / X -> 1.0 is legal when NaNs are ignored.
1132 if (Op0 == Op1)
1133 return ConstantFP::get(Op0->getType(), 1.0);
1134
1135 // -X / X -> -1.0 and
1136 // X / -X -> -1.0 are legal when NaNs are ignored.
1137 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
1138 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
1139 BinaryOperator::getFNegArgument(Op0) == Op1) ||
1140 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
1141 BinaryOperator::getFNegArgument(Op1) == Op0))
1142 return ConstantFP::get(Op0->getType(), -1.0);
1143 }
1144
1145 return nullptr;
1146 }
1147
SimplifyFDivInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1148 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1149 const DataLayout &DL,
1150 const TargetLibraryInfo *TLI,
1151 const DominatorTree *DT, AssumptionCache *AC,
1152 const Instruction *CxtI) {
1153 return ::SimplifyFDivInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1154 RecursionLimit);
1155 }
1156
1157 /// SimplifyRem - Given operands for an SRem or URem, see if we can
1158 /// fold the result. If not, this returns null.
SimplifyRem(Instruction::BinaryOps Opcode,Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1159 static Value *SimplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1160 const Query &Q, unsigned MaxRecurse) {
1161 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1162 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1163 Constant *Ops[] = { C0, C1 };
1164 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1165 }
1166 }
1167
1168 // X % undef -> undef
1169 if (match(Op1, m_Undef()))
1170 return Op1;
1171
1172 // undef % X -> 0
1173 if (match(Op0, m_Undef()))
1174 return Constant::getNullValue(Op0->getType());
1175
1176 // 0 % X -> 0, we don't need to preserve faults!
1177 if (match(Op0, m_Zero()))
1178 return Op0;
1179
1180 // X % 0 -> undef, we don't need to preserve faults!
1181 if (match(Op1, m_Zero()))
1182 return UndefValue::get(Op0->getType());
1183
1184 // X % 1 -> 0
1185 if (match(Op1, m_One()))
1186 return Constant::getNullValue(Op0->getType());
1187
1188 if (Op0->getType()->isIntegerTy(1))
1189 // It can't be remainder by zero, hence it must be remainder by one.
1190 return Constant::getNullValue(Op0->getType());
1191
1192 // X % X -> 0
1193 if (Op0 == Op1)
1194 return Constant::getNullValue(Op0->getType());
1195
1196 // (X % Y) % Y -> X % Y
1197 if ((Opcode == Instruction::SRem &&
1198 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1199 (Opcode == Instruction::URem &&
1200 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1201 return Op0;
1202
1203 // If the operation is with the result of a select instruction, check whether
1204 // operating on either branch of the select always yields the same value.
1205 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1206 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1207 return V;
1208
1209 // If the operation is with the result of a phi instruction, check whether
1210 // operating on all incoming values of the phi always yields the same value.
1211 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1212 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1213 return V;
1214
1215 return nullptr;
1216 }
1217
1218 /// SimplifySRemInst - Given operands for an SRem, see if we can
1219 /// fold the result. If not, this returns null.
SimplifySRemInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1220 static Value *SimplifySRemInst(Value *Op0, Value *Op1, const Query &Q,
1221 unsigned MaxRecurse) {
1222 if (Value *V = SimplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse))
1223 return V;
1224
1225 return nullptr;
1226 }
1227
SimplifySRemInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1228 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1229 const TargetLibraryInfo *TLI,
1230 const DominatorTree *DT, AssumptionCache *AC,
1231 const Instruction *CxtI) {
1232 return ::SimplifySRemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1233 RecursionLimit);
1234 }
1235
1236 /// SimplifyURemInst - Given operands for a URem, see if we can
1237 /// fold the result. If not, this returns null.
SimplifyURemInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1238 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const Query &Q,
1239 unsigned MaxRecurse) {
1240 if (Value *V = SimplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse))
1241 return V;
1242
1243 return nullptr;
1244 }
1245
SimplifyURemInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1246 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const DataLayout &DL,
1247 const TargetLibraryInfo *TLI,
1248 const DominatorTree *DT, AssumptionCache *AC,
1249 const Instruction *CxtI) {
1250 return ::SimplifyURemInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1251 RecursionLimit);
1252 }
1253
SimplifyFRemInst(Value * Op0,Value * Op1,FastMathFlags FMF,const Query &,unsigned)1254 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1255 const Query &, unsigned) {
1256 // undef % X -> undef (the undef could be a snan).
1257 if (match(Op0, m_Undef()))
1258 return Op0;
1259
1260 // X % undef -> undef
1261 if (match(Op1, m_Undef()))
1262 return Op1;
1263
1264 // 0 % X -> 0
1265 // Requires that NaNs are off (X could be zero) and signed zeroes are
1266 // ignored (X could be positive or negative, so the output sign is unknown).
1267 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
1268 return Op0;
1269
1270 return nullptr;
1271 }
1272
SimplifyFRemInst(Value * Op0,Value * Op1,FastMathFlags FMF,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1273 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
1274 const DataLayout &DL,
1275 const TargetLibraryInfo *TLI,
1276 const DominatorTree *DT, AssumptionCache *AC,
1277 const Instruction *CxtI) {
1278 return ::SimplifyFRemInst(Op0, Op1, FMF, Query(DL, TLI, DT, AC, CxtI),
1279 RecursionLimit);
1280 }
1281
1282 /// isUndefShift - Returns true if a shift by \c Amount always yields undef.
isUndefShift(Value * Amount)1283 static bool isUndefShift(Value *Amount) {
1284 Constant *C = dyn_cast<Constant>(Amount);
1285 if (!C)
1286 return false;
1287
1288 // X shift by undef -> undef because it may shift by the bitwidth.
1289 if (isa<UndefValue>(C))
1290 return true;
1291
1292 // Shifting by the bitwidth or more is undefined.
1293 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1294 if (CI->getValue().getLimitedValue() >=
1295 CI->getType()->getScalarSizeInBits())
1296 return true;
1297
1298 // If all lanes of a vector shift are undefined the whole shift is.
1299 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1300 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1301 if (!isUndefShift(C->getAggregateElement(I)))
1302 return false;
1303 return true;
1304 }
1305
1306 return false;
1307 }
1308
1309 /// SimplifyShift - Given operands for an Shl, LShr or AShr, see if we can
1310 /// fold the result. If not, this returns null.
SimplifyShift(unsigned Opcode,Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1311 static Value *SimplifyShift(unsigned Opcode, Value *Op0, Value *Op1,
1312 const Query &Q, unsigned MaxRecurse) {
1313 if (Constant *C0 = dyn_cast<Constant>(Op0)) {
1314 if (Constant *C1 = dyn_cast<Constant>(Op1)) {
1315 Constant *Ops[] = { C0, C1 };
1316 return ConstantFoldInstOperands(Opcode, C0->getType(), Ops, Q.DL, Q.TLI);
1317 }
1318 }
1319
1320 // 0 shift by X -> 0
1321 if (match(Op0, m_Zero()))
1322 return Op0;
1323
1324 // X shift by 0 -> X
1325 if (match(Op1, m_Zero()))
1326 return Op0;
1327
1328 // Fold undefined shifts.
1329 if (isUndefShift(Op1))
1330 return UndefValue::get(Op0->getType());
1331
1332 // If the operation is with the result of a select instruction, check whether
1333 // operating on either branch of the select always yields the same value.
1334 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1335 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1336 return V;
1337
1338 // If the operation is with the result of a phi instruction, check whether
1339 // operating on all incoming values of the phi always yields the same value.
1340 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1341 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1342 return V;
1343
1344 return nullptr;
1345 }
1346
1347 /// \brief Given operands for an Shl, LShr or AShr, see if we can
1348 /// fold the result. If not, this returns null.
SimplifyRightShift(unsigned Opcode,Value * Op0,Value * Op1,bool isExact,const Query & Q,unsigned MaxRecurse)1349 static Value *SimplifyRightShift(unsigned Opcode, Value *Op0, Value *Op1,
1350 bool isExact, const Query &Q,
1351 unsigned MaxRecurse) {
1352 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1353 return V;
1354
1355 // X >> X -> 0
1356 if (Op0 == Op1)
1357 return Constant::getNullValue(Op0->getType());
1358
1359 // undef >> X -> 0
1360 // undef >> X -> undef (if it's exact)
1361 if (match(Op0, m_Undef()))
1362 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1363
1364 // The low bit cannot be shifted out of an exact shift if it is set.
1365 if (isExact) {
1366 unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
1367 APInt Op0KnownZero(BitWidth, 0);
1368 APInt Op0KnownOne(BitWidth, 0);
1369 computeKnownBits(Op0, Op0KnownZero, Op0KnownOne, Q.DL, /*Depth=*/0, Q.AC,
1370 Q.CxtI, Q.DT);
1371 if (Op0KnownOne[0])
1372 return Op0;
1373 }
1374
1375 return nullptr;
1376 }
1377
1378 /// SimplifyShlInst - Given operands for an Shl, see if we can
1379 /// fold the result. If not, this returns null.
SimplifyShlInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const Query & Q,unsigned MaxRecurse)1380 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1381 const Query &Q, unsigned MaxRecurse) {
1382 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1383 return V;
1384
1385 // undef << X -> 0
1386 // undef << X -> undef if (if it's NSW/NUW)
1387 if (match(Op0, m_Undef()))
1388 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1389
1390 // (X >> A) << A -> X
1391 Value *X;
1392 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1393 return X;
1394 return nullptr;
1395 }
1396
SimplifyShlInst(Value * Op0,Value * Op1,bool isNSW,bool isNUW,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1397 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1398 const DataLayout &DL, const TargetLibraryInfo *TLI,
1399 const DominatorTree *DT, AssumptionCache *AC,
1400 const Instruction *CxtI) {
1401 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Query(DL, TLI, DT, AC, CxtI),
1402 RecursionLimit);
1403 }
1404
1405 /// SimplifyLShrInst - Given operands for an LShr, see if we can
1406 /// fold the result. If not, this returns null.
SimplifyLShrInst(Value * Op0,Value * Op1,bool isExact,const Query & Q,unsigned MaxRecurse)1407 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1408 const Query &Q, unsigned MaxRecurse) {
1409 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1410 MaxRecurse))
1411 return V;
1412
1413 // (X << A) >> A -> X
1414 Value *X;
1415 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1416 return X;
1417
1418 return nullptr;
1419 }
1420
SimplifyLShrInst(Value * Op0,Value * Op1,bool isExact,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1421 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1422 const DataLayout &DL,
1423 const TargetLibraryInfo *TLI,
1424 const DominatorTree *DT, AssumptionCache *AC,
1425 const Instruction *CxtI) {
1426 return ::SimplifyLShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1427 RecursionLimit);
1428 }
1429
1430 /// SimplifyAShrInst - Given operands for an AShr, see if we can
1431 /// fold the result. If not, this returns null.
SimplifyAShrInst(Value * Op0,Value * Op1,bool isExact,const Query & Q,unsigned MaxRecurse)1432 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1433 const Query &Q, unsigned MaxRecurse) {
1434 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1435 MaxRecurse))
1436 return V;
1437
1438 // all ones >>a X -> all ones
1439 if (match(Op0, m_AllOnes()))
1440 return Op0;
1441
1442 // (X << A) >> A -> X
1443 Value *X;
1444 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1445 return X;
1446
1447 // Arithmetic shifting an all-sign-bit value is a no-op.
1448 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1449 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1450 return Op0;
1451
1452 return nullptr;
1453 }
1454
SimplifyAShrInst(Value * Op0,Value * Op1,bool isExact,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1455 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1456 const DataLayout &DL,
1457 const TargetLibraryInfo *TLI,
1458 const DominatorTree *DT, AssumptionCache *AC,
1459 const Instruction *CxtI) {
1460 return ::SimplifyAShrInst(Op0, Op1, isExact, Query(DL, TLI, DT, AC, CxtI),
1461 RecursionLimit);
1462 }
1463
simplifyUnsignedRangeCheck(ICmpInst * ZeroICmp,ICmpInst * UnsignedICmp,bool IsAnd)1464 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1465 ICmpInst *UnsignedICmp, bool IsAnd) {
1466 Value *X, *Y;
1467
1468 ICmpInst::Predicate EqPred;
1469 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1470 !ICmpInst::isEquality(EqPred))
1471 return nullptr;
1472
1473 ICmpInst::Predicate UnsignedPred;
1474 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1475 ICmpInst::isUnsigned(UnsignedPred))
1476 ;
1477 else if (match(UnsignedICmp,
1478 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
1479 ICmpInst::isUnsigned(UnsignedPred))
1480 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1481 else
1482 return nullptr;
1483
1484 // X < Y && Y != 0 --> X < Y
1485 // X < Y || Y != 0 --> Y != 0
1486 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1487 return IsAnd ? UnsignedICmp : ZeroICmp;
1488
1489 // X >= Y || Y != 0 --> true
1490 // X >= Y || Y == 0 --> X >= Y
1491 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1492 if (EqPred == ICmpInst::ICMP_NE)
1493 return getTrue(UnsignedICmp->getType());
1494 return UnsignedICmp;
1495 }
1496
1497 // X < Y && Y == 0 --> false
1498 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1499 IsAnd)
1500 return getFalse(UnsignedICmp->getType());
1501
1502 return nullptr;
1503 }
1504
1505 // Simplify (and (icmp ...) (icmp ...)) to true when we can tell that the range
1506 // of possible values cannot be satisfied.
SimplifyAndOfICmps(ICmpInst * Op0,ICmpInst * Op1)1507 static Value *SimplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1508 ICmpInst::Predicate Pred0, Pred1;
1509 ConstantInt *CI1, *CI2;
1510 Value *V;
1511
1512 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1513 return X;
1514
1515 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)),
1516 m_ConstantInt(CI2))))
1517 return nullptr;
1518
1519 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1))))
1520 return nullptr;
1521
1522 Type *ITy = Op0->getType();
1523
1524 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1525 bool isNSW = AddInst->hasNoSignedWrap();
1526 bool isNUW = AddInst->hasNoUnsignedWrap();
1527
1528 const APInt &CI1V = CI1->getValue();
1529 const APInt &CI2V = CI2->getValue();
1530 const APInt Delta = CI2V - CI1V;
1531 if (CI1V.isStrictlyPositive()) {
1532 if (Delta == 2) {
1533 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1534 return getFalse(ITy);
1535 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1536 return getFalse(ITy);
1537 }
1538 if (Delta == 1) {
1539 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1540 return getFalse(ITy);
1541 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1542 return getFalse(ITy);
1543 }
1544 }
1545 if (CI1V.getBoolValue() && isNUW) {
1546 if (Delta == 2)
1547 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1548 return getFalse(ITy);
1549 if (Delta == 1)
1550 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1551 return getFalse(ITy);
1552 }
1553
1554 return nullptr;
1555 }
1556
1557 /// SimplifyAndInst - Given operands for an And, see if we can
1558 /// fold the result. If not, this returns null.
SimplifyAndInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1559 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const Query &Q,
1560 unsigned MaxRecurse) {
1561 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1562 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1563 Constant *Ops[] = { CLHS, CRHS };
1564 return ConstantFoldInstOperands(Instruction::And, CLHS->getType(),
1565 Ops, Q.DL, Q.TLI);
1566 }
1567
1568 // Canonicalize the constant to the RHS.
1569 std::swap(Op0, Op1);
1570 }
1571
1572 // X & undef -> 0
1573 if (match(Op1, m_Undef()))
1574 return Constant::getNullValue(Op0->getType());
1575
1576 // X & X = X
1577 if (Op0 == Op1)
1578 return Op0;
1579
1580 // X & 0 = 0
1581 if (match(Op1, m_Zero()))
1582 return Op1;
1583
1584 // X & -1 = X
1585 if (match(Op1, m_AllOnes()))
1586 return Op0;
1587
1588 // A & ~A = ~A & A = 0
1589 if (match(Op0, m_Not(m_Specific(Op1))) ||
1590 match(Op1, m_Not(m_Specific(Op0))))
1591 return Constant::getNullValue(Op0->getType());
1592
1593 // (A | ?) & A = A
1594 Value *A = nullptr, *B = nullptr;
1595 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1596 (A == Op1 || B == Op1))
1597 return Op1;
1598
1599 // A & (A | ?) = A
1600 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1601 (A == Op0 || B == Op0))
1602 return Op0;
1603
1604 // A & (-A) = A if A is a power of two or zero.
1605 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1606 match(Op1, m_Neg(m_Specific(Op0)))) {
1607 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1608 Q.DT))
1609 return Op0;
1610 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1611 Q.DT))
1612 return Op1;
1613 }
1614
1615 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1616 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1617 if (Value *V = SimplifyAndOfICmps(ICILHS, ICIRHS))
1618 return V;
1619 if (Value *V = SimplifyAndOfICmps(ICIRHS, ICILHS))
1620 return V;
1621 }
1622 }
1623
1624 // Try some generic simplifications for associative operations.
1625 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1626 MaxRecurse))
1627 return V;
1628
1629 // And distributes over Or. Try some generic simplifications based on this.
1630 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1631 Q, MaxRecurse))
1632 return V;
1633
1634 // And distributes over Xor. Try some generic simplifications based on this.
1635 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1636 Q, MaxRecurse))
1637 return V;
1638
1639 // If the operation is with the result of a select instruction, check whether
1640 // operating on either branch of the select always yields the same value.
1641 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1642 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1643 MaxRecurse))
1644 return V;
1645
1646 // If the operation is with the result of a phi instruction, check whether
1647 // operating on all incoming values of the phi always yields the same value.
1648 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1649 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1650 MaxRecurse))
1651 return V;
1652
1653 return nullptr;
1654 }
1655
SimplifyAndInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1656 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const DataLayout &DL,
1657 const TargetLibraryInfo *TLI,
1658 const DominatorTree *DT, AssumptionCache *AC,
1659 const Instruction *CxtI) {
1660 return ::SimplifyAndInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1661 RecursionLimit);
1662 }
1663
1664 // Simplify (or (icmp ...) (icmp ...)) to true when we can tell that the union
1665 // contains all possible values.
SimplifyOrOfICmps(ICmpInst * Op0,ICmpInst * Op1)1666 static Value *SimplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1667 ICmpInst::Predicate Pred0, Pred1;
1668 ConstantInt *CI1, *CI2;
1669 Value *V;
1670
1671 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1672 return X;
1673
1674 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)),
1675 m_ConstantInt(CI2))))
1676 return nullptr;
1677
1678 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1))))
1679 return nullptr;
1680
1681 Type *ITy = Op0->getType();
1682
1683 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1684 bool isNSW = AddInst->hasNoSignedWrap();
1685 bool isNUW = AddInst->hasNoUnsignedWrap();
1686
1687 const APInt &CI1V = CI1->getValue();
1688 const APInt &CI2V = CI2->getValue();
1689 const APInt Delta = CI2V - CI1V;
1690 if (CI1V.isStrictlyPositive()) {
1691 if (Delta == 2) {
1692 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1693 return getTrue(ITy);
1694 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1695 return getTrue(ITy);
1696 }
1697 if (Delta == 1) {
1698 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1699 return getTrue(ITy);
1700 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1701 return getTrue(ITy);
1702 }
1703 }
1704 if (CI1V.getBoolValue() && isNUW) {
1705 if (Delta == 2)
1706 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1707 return getTrue(ITy);
1708 if (Delta == 1)
1709 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1710 return getTrue(ITy);
1711 }
1712
1713 return nullptr;
1714 }
1715
1716 /// SimplifyOrInst - Given operands for an Or, see if we can
1717 /// fold the result. If not, this returns null.
SimplifyOrInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1718 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const Query &Q,
1719 unsigned MaxRecurse) {
1720 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1721 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1722 Constant *Ops[] = { CLHS, CRHS };
1723 return ConstantFoldInstOperands(Instruction::Or, CLHS->getType(),
1724 Ops, Q.DL, Q.TLI);
1725 }
1726
1727 // Canonicalize the constant to the RHS.
1728 std::swap(Op0, Op1);
1729 }
1730
1731 // X | undef -> -1
1732 if (match(Op1, m_Undef()))
1733 return Constant::getAllOnesValue(Op0->getType());
1734
1735 // X | X = X
1736 if (Op0 == Op1)
1737 return Op0;
1738
1739 // X | 0 = X
1740 if (match(Op1, m_Zero()))
1741 return Op0;
1742
1743 // X | -1 = -1
1744 if (match(Op1, m_AllOnes()))
1745 return Op1;
1746
1747 // A | ~A = ~A | A = -1
1748 if (match(Op0, m_Not(m_Specific(Op1))) ||
1749 match(Op1, m_Not(m_Specific(Op0))))
1750 return Constant::getAllOnesValue(Op0->getType());
1751
1752 // (A & ?) | A = A
1753 Value *A = nullptr, *B = nullptr;
1754 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1755 (A == Op1 || B == Op1))
1756 return Op1;
1757
1758 // A | (A & ?) = A
1759 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1760 (A == Op0 || B == Op0))
1761 return Op0;
1762
1763 // ~(A & ?) | A = -1
1764 if (match(Op0, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1765 (A == Op1 || B == Op1))
1766 return Constant::getAllOnesValue(Op1->getType());
1767
1768 // A | ~(A & ?) = -1
1769 if (match(Op1, m_Not(m_And(m_Value(A), m_Value(B)))) &&
1770 (A == Op0 || B == Op0))
1771 return Constant::getAllOnesValue(Op0->getType());
1772
1773 if (auto *ICILHS = dyn_cast<ICmpInst>(Op0)) {
1774 if (auto *ICIRHS = dyn_cast<ICmpInst>(Op1)) {
1775 if (Value *V = SimplifyOrOfICmps(ICILHS, ICIRHS))
1776 return V;
1777 if (Value *V = SimplifyOrOfICmps(ICIRHS, ICILHS))
1778 return V;
1779 }
1780 }
1781
1782 // Try some generic simplifications for associative operations.
1783 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1784 MaxRecurse))
1785 return V;
1786
1787 // Or distributes over And. Try some generic simplifications based on this.
1788 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1789 MaxRecurse))
1790 return V;
1791
1792 // If the operation is with the result of a select instruction, check whether
1793 // operating on either branch of the select always yields the same value.
1794 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1795 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1796 MaxRecurse))
1797 return V;
1798
1799 // (A & C)|(B & D)
1800 Value *C = nullptr, *D = nullptr;
1801 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1802 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1803 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
1804 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
1805 if (C1 && C2 && (C1->getValue() == ~C2->getValue())) {
1806 // (A & C1)|(B & C2)
1807 // If we have: ((V + N) & C1) | (V & C2)
1808 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1809 // replace with V+N.
1810 Value *V1, *V2;
1811 if ((C2->getValue() & (C2->getValue() + 1)) == 0 && // C2 == 0+1+
1812 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1813 // Add commutes, try both ways.
1814 if (V1 == B &&
1815 MaskedValueIsZero(V2, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1816 return A;
1817 if (V2 == B &&
1818 MaskedValueIsZero(V1, C2->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1819 return A;
1820 }
1821 // Or commutes, try both ways.
1822 if ((C1->getValue() & (C1->getValue() + 1)) == 0 &&
1823 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1824 // Add commutes, try both ways.
1825 if (V1 == A &&
1826 MaskedValueIsZero(V2, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1827 return B;
1828 if (V2 == A &&
1829 MaskedValueIsZero(V1, C1->getValue(), Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1830 return B;
1831 }
1832 }
1833 }
1834
1835 // If the operation is with the result of a phi instruction, check whether
1836 // operating on all incoming values of the phi always yields the same value.
1837 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1838 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1839 return V;
1840
1841 return nullptr;
1842 }
1843
SimplifyOrInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1844 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const DataLayout &DL,
1845 const TargetLibraryInfo *TLI,
1846 const DominatorTree *DT, AssumptionCache *AC,
1847 const Instruction *CxtI) {
1848 return ::SimplifyOrInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1849 RecursionLimit);
1850 }
1851
1852 /// SimplifyXorInst - Given operands for a Xor, see if we can
1853 /// fold the result. If not, this returns null.
SimplifyXorInst(Value * Op0,Value * Op1,const Query & Q,unsigned MaxRecurse)1854 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const Query &Q,
1855 unsigned MaxRecurse) {
1856 if (Constant *CLHS = dyn_cast<Constant>(Op0)) {
1857 if (Constant *CRHS = dyn_cast<Constant>(Op1)) {
1858 Constant *Ops[] = { CLHS, CRHS };
1859 return ConstantFoldInstOperands(Instruction::Xor, CLHS->getType(),
1860 Ops, Q.DL, Q.TLI);
1861 }
1862
1863 // Canonicalize the constant to the RHS.
1864 std::swap(Op0, Op1);
1865 }
1866
1867 // A ^ undef -> undef
1868 if (match(Op1, m_Undef()))
1869 return Op1;
1870
1871 // A ^ 0 = A
1872 if (match(Op1, m_Zero()))
1873 return Op0;
1874
1875 // A ^ A = 0
1876 if (Op0 == Op1)
1877 return Constant::getNullValue(Op0->getType());
1878
1879 // A ^ ~A = ~A ^ A = -1
1880 if (match(Op0, m_Not(m_Specific(Op1))) ||
1881 match(Op1, m_Not(m_Specific(Op0))))
1882 return Constant::getAllOnesValue(Op0->getType());
1883
1884 // Try some generic simplifications for associative operations.
1885 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1886 MaxRecurse))
1887 return V;
1888
1889 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1890 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1891 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1892 // only if B and C are equal. If B and C are equal then (since we assume
1893 // that operands have already been simplified) "select(cond, B, C)" should
1894 // have been simplified to the common value of B and C already. Analysing
1895 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1896 // for threading over phi nodes.
1897
1898 return nullptr;
1899 }
1900
SimplifyXorInst(Value * Op0,Value * Op1,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)1901 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const DataLayout &DL,
1902 const TargetLibraryInfo *TLI,
1903 const DominatorTree *DT, AssumptionCache *AC,
1904 const Instruction *CxtI) {
1905 return ::SimplifyXorInst(Op0, Op1, Query(DL, TLI, DT, AC, CxtI),
1906 RecursionLimit);
1907 }
1908
GetCompareTy(Value * Op)1909 static Type *GetCompareTy(Value *Op) {
1910 return CmpInst::makeCmpResultType(Op->getType());
1911 }
1912
1913 /// ExtractEquivalentCondition - Rummage around inside V looking for something
1914 /// equivalent to the comparison "LHS Pred RHS". Return such a value if found,
1915 /// otherwise return null. Helper function for analyzing max/min idioms.
ExtractEquivalentCondition(Value * V,CmpInst::Predicate Pred,Value * LHS,Value * RHS)1916 static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1917 Value *LHS, Value *RHS) {
1918 SelectInst *SI = dyn_cast<SelectInst>(V);
1919 if (!SI)
1920 return nullptr;
1921 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1922 if (!Cmp)
1923 return nullptr;
1924 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1925 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1926 return Cmp;
1927 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1928 LHS == CmpRHS && RHS == CmpLHS)
1929 return Cmp;
1930 return nullptr;
1931 }
1932
1933 // A significant optimization not implemented here is assuming that alloca
1934 // addresses are not equal to incoming argument values. They don't *alias*,
1935 // as we say, but that doesn't mean they aren't equal, so we take a
1936 // conservative approach.
1937 //
1938 // This is inspired in part by C++11 5.10p1:
1939 // "Two pointers of the same type compare equal if and only if they are both
1940 // null, both point to the same function, or both represent the same
1941 // address."
1942 //
1943 // This is pretty permissive.
1944 //
1945 // It's also partly due to C11 6.5.9p6:
1946 // "Two pointers compare equal if and only if both are null pointers, both are
1947 // pointers to the same object (including a pointer to an object and a
1948 // subobject at its beginning) or function, both are pointers to one past the
1949 // last element of the same array object, or one is a pointer to one past the
1950 // end of one array object and the other is a pointer to the start of a
1951 // different array object that happens to immediately follow the first array
1952 // object in the address space.)
1953 //
1954 // C11's version is more restrictive, however there's no reason why an argument
1955 // couldn't be a one-past-the-end value for a stack object in the caller and be
1956 // equal to the beginning of a stack object in the callee.
1957 //
1958 // If the C and C++ standards are ever made sufficiently restrictive in this
1959 // area, it may be possible to update LLVM's semantics accordingly and reinstate
1960 // this optimization.
computePointerICmp(const DataLayout & DL,const TargetLibraryInfo * TLI,CmpInst::Predicate Pred,Value * LHS,Value * RHS)1961 static Constant *computePointerICmp(const DataLayout &DL,
1962 const TargetLibraryInfo *TLI,
1963 CmpInst::Predicate Pred, Value *LHS,
1964 Value *RHS) {
1965 // First, skip past any trivial no-ops.
1966 LHS = LHS->stripPointerCasts();
1967 RHS = RHS->stripPointerCasts();
1968
1969 // A non-null pointer is not equal to a null pointer.
1970 if (llvm::isKnownNonNull(LHS, TLI) && isa<ConstantPointerNull>(RHS) &&
1971 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
1972 return ConstantInt::get(GetCompareTy(LHS),
1973 !CmpInst::isTrueWhenEqual(Pred));
1974
1975 // We can only fold certain predicates on pointer comparisons.
1976 switch (Pred) {
1977 default:
1978 return nullptr;
1979
1980 // Equality comaprisons are easy to fold.
1981 case CmpInst::ICMP_EQ:
1982 case CmpInst::ICMP_NE:
1983 break;
1984
1985 // We can only handle unsigned relational comparisons because 'inbounds' on
1986 // a GEP only protects against unsigned wrapping.
1987 case CmpInst::ICMP_UGT:
1988 case CmpInst::ICMP_UGE:
1989 case CmpInst::ICMP_ULT:
1990 case CmpInst::ICMP_ULE:
1991 // However, we have to switch them to their signed variants to handle
1992 // negative indices from the base pointer.
1993 Pred = ICmpInst::getSignedPredicate(Pred);
1994 break;
1995 }
1996
1997 // Strip off any constant offsets so that we can reason about them.
1998 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
1999 // here and compare base addresses like AliasAnalysis does, however there are
2000 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2001 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2002 // doesn't need to guarantee pointer inequality when it says NoAlias.
2003 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2004 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2005
2006 // If LHS and RHS are related via constant offsets to the same base
2007 // value, we can replace it with an icmp which just compares the offsets.
2008 if (LHS == RHS)
2009 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2010
2011 // Various optimizations for (in)equality comparisons.
2012 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2013 // Different non-empty allocations that exist at the same time have
2014 // different addresses (if the program can tell). Global variables always
2015 // exist, so they always exist during the lifetime of each other and all
2016 // allocas. Two different allocas usually have different addresses...
2017 //
2018 // However, if there's an @llvm.stackrestore dynamically in between two
2019 // allocas, they may have the same address. It's tempting to reduce the
2020 // scope of the problem by only looking at *static* allocas here. That would
2021 // cover the majority of allocas while significantly reducing the likelihood
2022 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2023 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2024 // an entry block. Also, if we have a block that's not attached to a
2025 // function, we can't tell if it's "static" under the current definition.
2026 // Theoretically, this problem could be fixed by creating a new kind of
2027 // instruction kind specifically for static allocas. Such a new instruction
2028 // could be required to be at the top of the entry block, thus preventing it
2029 // from being subject to a @llvm.stackrestore. Instcombine could even
2030 // convert regular allocas into these special allocas. It'd be nifty.
2031 // However, until then, this problem remains open.
2032 //
2033 // So, we'll assume that two non-empty allocas have different addresses
2034 // for now.
2035 //
2036 // With all that, if the offsets are within the bounds of their allocations
2037 // (and not one-past-the-end! so we can't use inbounds!), and their
2038 // allocations aren't the same, the pointers are not equal.
2039 //
2040 // Note that it's not necessary to check for LHS being a global variable
2041 // address, due to canonicalization and constant folding.
2042 if (isa<AllocaInst>(LHS) &&
2043 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2044 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2045 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2046 uint64_t LHSSize, RHSSize;
2047 if (LHSOffsetCI && RHSOffsetCI &&
2048 getObjectSize(LHS, LHSSize, DL, TLI) &&
2049 getObjectSize(RHS, RHSSize, DL, TLI)) {
2050 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2051 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2052 if (!LHSOffsetValue.isNegative() &&
2053 !RHSOffsetValue.isNegative() &&
2054 LHSOffsetValue.ult(LHSSize) &&
2055 RHSOffsetValue.ult(RHSSize)) {
2056 return ConstantInt::get(GetCompareTy(LHS),
2057 !CmpInst::isTrueWhenEqual(Pred));
2058 }
2059 }
2060
2061 // Repeat the above check but this time without depending on DataLayout
2062 // or being able to compute a precise size.
2063 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2064 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2065 LHSOffset->isNullValue() &&
2066 RHSOffset->isNullValue())
2067 return ConstantInt::get(GetCompareTy(LHS),
2068 !CmpInst::isTrueWhenEqual(Pred));
2069 }
2070
2071 // Even if an non-inbounds GEP occurs along the path we can still optimize
2072 // equality comparisons concerning the result. We avoid walking the whole
2073 // chain again by starting where the last calls to
2074 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2075 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2076 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2077 if (LHS == RHS)
2078 return ConstantExpr::getICmp(Pred,
2079 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2080 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2081
2082 // If one side of the equality comparison must come from a noalias call
2083 // (meaning a system memory allocation function), and the other side must
2084 // come from a pointer that cannot overlap with dynamically-allocated
2085 // memory within the lifetime of the current function (allocas, byval
2086 // arguments, globals), then determine the comparison result here.
2087 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2088 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2089 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2090
2091 // Is the set of underlying objects all noalias calls?
2092 auto IsNAC = [](SmallVectorImpl<Value *> &Objects) {
2093 return std::all_of(Objects.begin(), Objects.end(), isNoAliasCall);
2094 };
2095
2096 // Is the set of underlying objects all things which must be disjoint from
2097 // noalias calls. For allocas, we consider only static ones (dynamic
2098 // allocas might be transformed into calls to malloc not simultaneously
2099 // live with the compared-to allocation). For globals, we exclude symbols
2100 // that might be resolve lazily to symbols in another dynamically-loaded
2101 // library (and, thus, could be malloc'ed by the implementation).
2102 auto IsAllocDisjoint = [](SmallVectorImpl<Value *> &Objects) {
2103 return std::all_of(Objects.begin(), Objects.end(),
2104 [](Value *V){
2105 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2106 return AI->getParent() && AI->getParent()->getParent() &&
2107 AI->isStaticAlloca();
2108 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2109 return (GV->hasLocalLinkage() ||
2110 GV->hasHiddenVisibility() ||
2111 GV->hasProtectedVisibility() ||
2112 GV->hasUnnamedAddr()) &&
2113 !GV->isThreadLocal();
2114 if (const Argument *A = dyn_cast<Argument>(V))
2115 return A->hasByValAttr();
2116 return false;
2117 });
2118 };
2119
2120 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2121 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2122 return ConstantInt::get(GetCompareTy(LHS),
2123 !CmpInst::isTrueWhenEqual(Pred));
2124 }
2125
2126 // Otherwise, fail.
2127 return nullptr;
2128 }
2129
2130 /// SimplifyICmpInst - Given operands for an ICmpInst, see if we can
2131 /// fold the result. If not, this returns null.
SimplifyICmpInst(unsigned Predicate,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)2132 static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
2133 const Query &Q, unsigned MaxRecurse) {
2134 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
2135 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
2136
2137 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
2138 if (Constant *CRHS = dyn_cast<Constant>(RHS))
2139 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
2140
2141 // If we have a constant, make sure it is on the RHS.
2142 std::swap(LHS, RHS);
2143 Pred = CmpInst::getSwappedPredicate(Pred);
2144 }
2145
2146 Type *ITy = GetCompareTy(LHS); // The return type.
2147 Type *OpTy = LHS->getType(); // The operand type.
2148
2149 // icmp X, X -> true/false
2150 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
2151 // because X could be 0.
2152 if (LHS == RHS || isa<UndefValue>(RHS))
2153 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
2154
2155 // Special case logic when the operands have i1 type.
2156 if (OpTy->getScalarType()->isIntegerTy(1)) {
2157 switch (Pred) {
2158 default: break;
2159 case ICmpInst::ICMP_EQ:
2160 // X == 1 -> X
2161 if (match(RHS, m_One()))
2162 return LHS;
2163 break;
2164 case ICmpInst::ICMP_NE:
2165 // X != 0 -> X
2166 if (match(RHS, m_Zero()))
2167 return LHS;
2168 break;
2169 case ICmpInst::ICMP_UGT:
2170 // X >u 0 -> X
2171 if (match(RHS, m_Zero()))
2172 return LHS;
2173 break;
2174 case ICmpInst::ICMP_UGE:
2175 // X >=u 1 -> X
2176 if (match(RHS, m_One()))
2177 return LHS;
2178 if (isImpliedCondition(RHS, LHS, Q.DL))
2179 return getTrue(ITy);
2180 break;
2181 case ICmpInst::ICMP_SGE:
2182 /// For signed comparison, the values for an i1 are 0 and -1
2183 /// respectively. This maps into a truth table of:
2184 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2185 /// 0 | 0 | 1 (0 >= 0) | 1
2186 /// 0 | 1 | 1 (0 >= -1) | 1
2187 /// 1 | 0 | 0 (-1 >= 0) | 0
2188 /// 1 | 1 | 1 (-1 >= -1) | 1
2189 if (isImpliedCondition(LHS, RHS, Q.DL))
2190 return getTrue(ITy);
2191 break;
2192 case ICmpInst::ICMP_SLT:
2193 // X <s 0 -> X
2194 if (match(RHS, m_Zero()))
2195 return LHS;
2196 break;
2197 case ICmpInst::ICMP_SLE:
2198 // X <=s -1 -> X
2199 if (match(RHS, m_One()))
2200 return LHS;
2201 break;
2202 case ICmpInst::ICMP_ULE:
2203 if (isImpliedCondition(LHS, RHS, Q.DL))
2204 return getTrue(ITy);
2205 break;
2206 }
2207 }
2208
2209 // If we are comparing with zero then try hard since this is a common case.
2210 if (match(RHS, m_Zero())) {
2211 bool LHSKnownNonNegative, LHSKnownNegative;
2212 switch (Pred) {
2213 default: llvm_unreachable("Unknown ICmp predicate!");
2214 case ICmpInst::ICMP_ULT:
2215 return getFalse(ITy);
2216 case ICmpInst::ICMP_UGE:
2217 return getTrue(ITy);
2218 case ICmpInst::ICMP_EQ:
2219 case ICmpInst::ICMP_ULE:
2220 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2221 return getFalse(ITy);
2222 break;
2223 case ICmpInst::ICMP_NE:
2224 case ICmpInst::ICMP_UGT:
2225 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2226 return getTrue(ITy);
2227 break;
2228 case ICmpInst::ICMP_SLT:
2229 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2230 Q.CxtI, Q.DT);
2231 if (LHSKnownNegative)
2232 return getTrue(ITy);
2233 if (LHSKnownNonNegative)
2234 return getFalse(ITy);
2235 break;
2236 case ICmpInst::ICMP_SLE:
2237 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2238 Q.CxtI, Q.DT);
2239 if (LHSKnownNegative)
2240 return getTrue(ITy);
2241 if (LHSKnownNonNegative &&
2242 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2243 return getFalse(ITy);
2244 break;
2245 case ICmpInst::ICMP_SGE:
2246 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2247 Q.CxtI, Q.DT);
2248 if (LHSKnownNegative)
2249 return getFalse(ITy);
2250 if (LHSKnownNonNegative)
2251 return getTrue(ITy);
2252 break;
2253 case ICmpInst::ICMP_SGT:
2254 ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, Q.DL, 0, Q.AC,
2255 Q.CxtI, Q.DT);
2256 if (LHSKnownNegative)
2257 return getFalse(ITy);
2258 if (LHSKnownNonNegative &&
2259 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2260 return getTrue(ITy);
2261 break;
2262 }
2263 }
2264
2265 // See if we are doing a comparison with a constant integer.
2266 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2267 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2268 ConstantRange RHS_CR = ICmpInst::makeConstantRange(Pred, CI->getValue());
2269 if (RHS_CR.isEmptySet())
2270 return ConstantInt::getFalse(CI->getContext());
2271 if (RHS_CR.isFullSet())
2272 return ConstantInt::getTrue(CI->getContext());
2273
2274 // Many binary operators with constant RHS have easy to compute constant
2275 // range. Use them to check whether the comparison is a tautology.
2276 unsigned Width = CI->getBitWidth();
2277 APInt Lower = APInt(Width, 0);
2278 APInt Upper = APInt(Width, 0);
2279 ConstantInt *CI2;
2280 if (match(LHS, m_URem(m_Value(), m_ConstantInt(CI2)))) {
2281 // 'urem x, CI2' produces [0, CI2).
2282 Upper = CI2->getValue();
2283 } else if (match(LHS, m_SRem(m_Value(), m_ConstantInt(CI2)))) {
2284 // 'srem x, CI2' produces (-|CI2|, |CI2|).
2285 Upper = CI2->getValue().abs();
2286 Lower = (-Upper) + 1;
2287 } else if (match(LHS, m_UDiv(m_ConstantInt(CI2), m_Value()))) {
2288 // 'udiv CI2, x' produces [0, CI2].
2289 Upper = CI2->getValue() + 1;
2290 } else if (match(LHS, m_UDiv(m_Value(), m_ConstantInt(CI2)))) {
2291 // 'udiv x, CI2' produces [0, UINT_MAX / CI2].
2292 APInt NegOne = APInt::getAllOnesValue(Width);
2293 if (!CI2->isZero())
2294 Upper = NegOne.udiv(CI2->getValue()) + 1;
2295 } else if (match(LHS, m_SDiv(m_ConstantInt(CI2), m_Value()))) {
2296 if (CI2->isMinSignedValue()) {
2297 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2298 Lower = CI2->getValue();
2299 Upper = Lower.lshr(1) + 1;
2300 } else {
2301 // 'sdiv CI2, x' produces [-|CI2|, |CI2|].
2302 Upper = CI2->getValue().abs() + 1;
2303 Lower = (-Upper) + 1;
2304 }
2305 } else if (match(LHS, m_SDiv(m_Value(), m_ConstantInt(CI2)))) {
2306 APInt IntMin = APInt::getSignedMinValue(Width);
2307 APInt IntMax = APInt::getSignedMaxValue(Width);
2308 APInt Val = CI2->getValue();
2309 if (Val.isAllOnesValue()) {
2310 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2311 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2312 Lower = IntMin + 1;
2313 Upper = IntMax + 1;
2314 } else if (Val.countLeadingZeros() < Width - 1) {
2315 // 'sdiv x, CI2' produces [INT_MIN / CI2, INT_MAX / CI2]
2316 // where CI2 != -1 and CI2 != 0 and CI2 != 1
2317 Lower = IntMin.sdiv(Val);
2318 Upper = IntMax.sdiv(Val);
2319 if (Lower.sgt(Upper))
2320 std::swap(Lower, Upper);
2321 Upper = Upper + 1;
2322 assert(Upper != Lower && "Upper part of range has wrapped!");
2323 }
2324 } else if (match(LHS, m_NUWShl(m_ConstantInt(CI2), m_Value()))) {
2325 // 'shl nuw CI2, x' produces [CI2, CI2 << CLZ(CI2)]
2326 Lower = CI2->getValue();
2327 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2328 } else if (match(LHS, m_NSWShl(m_ConstantInt(CI2), m_Value()))) {
2329 if (CI2->isNegative()) {
2330 // 'shl nsw CI2, x' produces [CI2 << CLO(CI2)-1, CI2]
2331 unsigned ShiftAmount = CI2->getValue().countLeadingOnes() - 1;
2332 Lower = CI2->getValue().shl(ShiftAmount);
2333 Upper = CI2->getValue() + 1;
2334 } else {
2335 // 'shl nsw CI2, x' produces [CI2, CI2 << CLZ(CI2)-1]
2336 unsigned ShiftAmount = CI2->getValue().countLeadingZeros() - 1;
2337 Lower = CI2->getValue();
2338 Upper = CI2->getValue().shl(ShiftAmount) + 1;
2339 }
2340 } else if (match(LHS, m_LShr(m_Value(), m_ConstantInt(CI2)))) {
2341 // 'lshr x, CI2' produces [0, UINT_MAX >> CI2].
2342 APInt NegOne = APInt::getAllOnesValue(Width);
2343 if (CI2->getValue().ult(Width))
2344 Upper = NegOne.lshr(CI2->getValue()) + 1;
2345 } else if (match(LHS, m_LShr(m_ConstantInt(CI2), m_Value()))) {
2346 // 'lshr CI2, x' produces [CI2 >> (Width-1), CI2].
2347 unsigned ShiftAmount = Width - 1;
2348 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2349 ShiftAmount = CI2->getValue().countTrailingZeros();
2350 Lower = CI2->getValue().lshr(ShiftAmount);
2351 Upper = CI2->getValue() + 1;
2352 } else if (match(LHS, m_AShr(m_Value(), m_ConstantInt(CI2)))) {
2353 // 'ashr x, CI2' produces [INT_MIN >> CI2, INT_MAX >> CI2].
2354 APInt IntMin = APInt::getSignedMinValue(Width);
2355 APInt IntMax = APInt::getSignedMaxValue(Width);
2356 if (CI2->getValue().ult(Width)) {
2357 Lower = IntMin.ashr(CI2->getValue());
2358 Upper = IntMax.ashr(CI2->getValue()) + 1;
2359 }
2360 } else if (match(LHS, m_AShr(m_ConstantInt(CI2), m_Value()))) {
2361 unsigned ShiftAmount = Width - 1;
2362 if (!CI2->isZero() && cast<BinaryOperator>(LHS)->isExact())
2363 ShiftAmount = CI2->getValue().countTrailingZeros();
2364 if (CI2->isNegative()) {
2365 // 'ashr CI2, x' produces [CI2, CI2 >> (Width-1)]
2366 Lower = CI2->getValue();
2367 Upper = CI2->getValue().ashr(ShiftAmount) + 1;
2368 } else {
2369 // 'ashr CI2, x' produces [CI2 >> (Width-1), CI2]
2370 Lower = CI2->getValue().ashr(ShiftAmount);
2371 Upper = CI2->getValue() + 1;
2372 }
2373 } else if (match(LHS, m_Or(m_Value(), m_ConstantInt(CI2)))) {
2374 // 'or x, CI2' produces [CI2, UINT_MAX].
2375 Lower = CI2->getValue();
2376 } else if (match(LHS, m_And(m_Value(), m_ConstantInt(CI2)))) {
2377 // 'and x, CI2' produces [0, CI2].
2378 Upper = CI2->getValue() + 1;
2379 } else if (match(LHS, m_NUWAdd(m_Value(), m_ConstantInt(CI2)))) {
2380 // 'add nuw x, CI2' produces [CI2, UINT_MAX].
2381 Lower = CI2->getValue();
2382 }
2383
2384 ConstantRange LHS_CR = Lower != Upper ? ConstantRange(Lower, Upper)
2385 : ConstantRange(Width, true);
2386
2387 if (auto *I = dyn_cast<Instruction>(LHS))
2388 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2389 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2390
2391 if (!LHS_CR.isFullSet()) {
2392 if (RHS_CR.contains(LHS_CR))
2393 return ConstantInt::getTrue(RHS->getContext());
2394 if (RHS_CR.inverse().contains(LHS_CR))
2395 return ConstantInt::getFalse(RHS->getContext());
2396 }
2397 }
2398
2399 // If both operands have range metadata, use the metadata
2400 // to simplify the comparison.
2401 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
2402 auto RHS_Instr = dyn_cast<Instruction>(RHS);
2403 auto LHS_Instr = dyn_cast<Instruction>(LHS);
2404
2405 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
2406 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
2407 auto RHS_CR = getConstantRangeFromMetadata(
2408 *RHS_Instr->getMetadata(LLVMContext::MD_range));
2409 auto LHS_CR = getConstantRangeFromMetadata(
2410 *LHS_Instr->getMetadata(LLVMContext::MD_range));
2411
2412 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
2413 if (Satisfied_CR.contains(LHS_CR))
2414 return ConstantInt::getTrue(RHS->getContext());
2415
2416 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
2417 CmpInst::getInversePredicate(Pred), RHS_CR);
2418 if (InversedSatisfied_CR.contains(LHS_CR))
2419 return ConstantInt::getFalse(RHS->getContext());
2420 }
2421 }
2422
2423 // Compare of cast, for example (zext X) != 0 -> X != 0
2424 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
2425 Instruction *LI = cast<CastInst>(LHS);
2426 Value *SrcOp = LI->getOperand(0);
2427 Type *SrcTy = SrcOp->getType();
2428 Type *DstTy = LI->getType();
2429
2430 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
2431 // if the integer type is the same size as the pointer type.
2432 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
2433 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
2434 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2435 // Transfer the cast to the constant.
2436 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
2437 ConstantExpr::getIntToPtr(RHSC, SrcTy),
2438 Q, MaxRecurse-1))
2439 return V;
2440 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
2441 if (RI->getOperand(0)->getType() == SrcTy)
2442 // Compare without the cast.
2443 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2444 Q, MaxRecurse-1))
2445 return V;
2446 }
2447 }
2448
2449 if (isa<ZExtInst>(LHS)) {
2450 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
2451 // same type.
2452 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
2453 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2454 // Compare X and Y. Note that signed predicates become unsigned.
2455 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2456 SrcOp, RI->getOperand(0), Q,
2457 MaxRecurse-1))
2458 return V;
2459 }
2460 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
2461 // too. If not, then try to deduce the result of the comparison.
2462 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2463 // Compute the constant that would happen if we truncated to SrcTy then
2464 // reextended to DstTy.
2465 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2466 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
2467
2468 // If the re-extended constant didn't change then this is effectively
2469 // also a case of comparing two zero-extended values.
2470 if (RExt == CI && MaxRecurse)
2471 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
2472 SrcOp, Trunc, Q, MaxRecurse-1))
2473 return V;
2474
2475 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
2476 // there. Use this to work out the result of the comparison.
2477 if (RExt != CI) {
2478 switch (Pred) {
2479 default: llvm_unreachable("Unknown ICmp predicate!");
2480 // LHS <u RHS.
2481 case ICmpInst::ICMP_EQ:
2482 case ICmpInst::ICMP_UGT:
2483 case ICmpInst::ICMP_UGE:
2484 return ConstantInt::getFalse(CI->getContext());
2485
2486 case ICmpInst::ICMP_NE:
2487 case ICmpInst::ICMP_ULT:
2488 case ICmpInst::ICMP_ULE:
2489 return ConstantInt::getTrue(CI->getContext());
2490
2491 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
2492 // is non-negative then LHS <s RHS.
2493 case ICmpInst::ICMP_SGT:
2494 case ICmpInst::ICMP_SGE:
2495 return CI->getValue().isNegative() ?
2496 ConstantInt::getTrue(CI->getContext()) :
2497 ConstantInt::getFalse(CI->getContext());
2498
2499 case ICmpInst::ICMP_SLT:
2500 case ICmpInst::ICMP_SLE:
2501 return CI->getValue().isNegative() ?
2502 ConstantInt::getFalse(CI->getContext()) :
2503 ConstantInt::getTrue(CI->getContext());
2504 }
2505 }
2506 }
2507 }
2508
2509 if (isa<SExtInst>(LHS)) {
2510 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
2511 // same type.
2512 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
2513 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
2514 // Compare X and Y. Note that the predicate does not change.
2515 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
2516 Q, MaxRecurse-1))
2517 return V;
2518 }
2519 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
2520 // too. If not, then try to deduce the result of the comparison.
2521 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
2522 // Compute the constant that would happen if we truncated to SrcTy then
2523 // reextended to DstTy.
2524 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
2525 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
2526
2527 // If the re-extended constant didn't change then this is effectively
2528 // also a case of comparing two sign-extended values.
2529 if (RExt == CI && MaxRecurse)
2530 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
2531 return V;
2532
2533 // Otherwise the upper bits of LHS are all equal, while RHS has varying
2534 // bits there. Use this to work out the result of the comparison.
2535 if (RExt != CI) {
2536 switch (Pred) {
2537 default: llvm_unreachable("Unknown ICmp predicate!");
2538 case ICmpInst::ICMP_EQ:
2539 return ConstantInt::getFalse(CI->getContext());
2540 case ICmpInst::ICMP_NE:
2541 return ConstantInt::getTrue(CI->getContext());
2542
2543 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
2544 // LHS >s RHS.
2545 case ICmpInst::ICMP_SGT:
2546 case ICmpInst::ICMP_SGE:
2547 return CI->getValue().isNegative() ?
2548 ConstantInt::getTrue(CI->getContext()) :
2549 ConstantInt::getFalse(CI->getContext());
2550 case ICmpInst::ICMP_SLT:
2551 case ICmpInst::ICMP_SLE:
2552 return CI->getValue().isNegative() ?
2553 ConstantInt::getFalse(CI->getContext()) :
2554 ConstantInt::getTrue(CI->getContext());
2555
2556 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
2557 // LHS >u RHS.
2558 case ICmpInst::ICMP_UGT:
2559 case ICmpInst::ICMP_UGE:
2560 // Comparison is true iff the LHS <s 0.
2561 if (MaxRecurse)
2562 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
2563 Constant::getNullValue(SrcTy),
2564 Q, MaxRecurse-1))
2565 return V;
2566 break;
2567 case ICmpInst::ICMP_ULT:
2568 case ICmpInst::ICMP_ULE:
2569 // Comparison is true iff the LHS >=s 0.
2570 if (MaxRecurse)
2571 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
2572 Constant::getNullValue(SrcTy),
2573 Q, MaxRecurse-1))
2574 return V;
2575 break;
2576 }
2577 }
2578 }
2579 }
2580 }
2581
2582 // icmp eq|ne X, Y -> false|true if X != Y
2583 if ((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) &&
2584 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
2585 LLVMContext &Ctx = LHS->getType()->getContext();
2586 return Pred == ICmpInst::ICMP_NE ?
2587 ConstantInt::getTrue(Ctx) : ConstantInt::getFalse(Ctx);
2588 }
2589
2590 // Special logic for binary operators.
2591 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2592 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2593 if (MaxRecurse && (LBO || RBO)) {
2594 // Analyze the case when either LHS or RHS is an add instruction.
2595 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2596 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2597 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2598 if (LBO && LBO->getOpcode() == Instruction::Add) {
2599 A = LBO->getOperand(0); B = LBO->getOperand(1);
2600 NoLHSWrapProblem = ICmpInst::isEquality(Pred) ||
2601 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2602 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2603 }
2604 if (RBO && RBO->getOpcode() == Instruction::Add) {
2605 C = RBO->getOperand(0); D = RBO->getOperand(1);
2606 NoRHSWrapProblem = ICmpInst::isEquality(Pred) ||
2607 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2608 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2609 }
2610
2611 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2612 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2613 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2614 Constant::getNullValue(RHS->getType()),
2615 Q, MaxRecurse-1))
2616 return V;
2617
2618 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2619 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2620 if (Value *V = SimplifyICmpInst(Pred,
2621 Constant::getNullValue(LHS->getType()),
2622 C == LHS ? D : C, Q, MaxRecurse-1))
2623 return V;
2624
2625 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2626 if (A && C && (A == C || A == D || B == C || B == D) &&
2627 NoLHSWrapProblem && NoRHSWrapProblem) {
2628 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2629 Value *Y, *Z;
2630 if (A == C) {
2631 // C + B == C + D -> B == D
2632 Y = B;
2633 Z = D;
2634 } else if (A == D) {
2635 // D + B == C + D -> B == C
2636 Y = B;
2637 Z = C;
2638 } else if (B == C) {
2639 // A + C == C + D -> A == D
2640 Y = A;
2641 Z = D;
2642 } else {
2643 assert(B == D);
2644 // A + D == C + D -> A == C
2645 Y = A;
2646 Z = C;
2647 }
2648 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse-1))
2649 return V;
2650 }
2651 }
2652
2653 // icmp pred (or X, Y), X
2654 if (LBO && match(LBO, m_CombineOr(m_Or(m_Value(), m_Specific(RHS)),
2655 m_Or(m_Specific(RHS), m_Value())))) {
2656 if (Pred == ICmpInst::ICMP_ULT)
2657 return getFalse(ITy);
2658 if (Pred == ICmpInst::ICMP_UGE)
2659 return getTrue(ITy);
2660 }
2661 // icmp pred X, (or X, Y)
2662 if (RBO && match(RBO, m_CombineOr(m_Or(m_Value(), m_Specific(LHS)),
2663 m_Or(m_Specific(LHS), m_Value())))) {
2664 if (Pred == ICmpInst::ICMP_ULE)
2665 return getTrue(ITy);
2666 if (Pred == ICmpInst::ICMP_UGT)
2667 return getFalse(ITy);
2668 }
2669
2670 // icmp pred (and X, Y), X
2671 if (LBO && match(LBO, m_CombineOr(m_And(m_Value(), m_Specific(RHS)),
2672 m_And(m_Specific(RHS), m_Value())))) {
2673 if (Pred == ICmpInst::ICMP_UGT)
2674 return getFalse(ITy);
2675 if (Pred == ICmpInst::ICMP_ULE)
2676 return getTrue(ITy);
2677 }
2678 // icmp pred X, (and X, Y)
2679 if (RBO && match(RBO, m_CombineOr(m_And(m_Value(), m_Specific(LHS)),
2680 m_And(m_Specific(LHS), m_Value())))) {
2681 if (Pred == ICmpInst::ICMP_UGE)
2682 return getTrue(ITy);
2683 if (Pred == ICmpInst::ICMP_ULT)
2684 return getFalse(ITy);
2685 }
2686
2687 // 0 - (zext X) pred C
2688 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2689 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2690 if (RHSC->getValue().isStrictlyPositive()) {
2691 if (Pred == ICmpInst::ICMP_SLT)
2692 return ConstantInt::getTrue(RHSC->getContext());
2693 if (Pred == ICmpInst::ICMP_SGE)
2694 return ConstantInt::getFalse(RHSC->getContext());
2695 if (Pred == ICmpInst::ICMP_EQ)
2696 return ConstantInt::getFalse(RHSC->getContext());
2697 if (Pred == ICmpInst::ICMP_NE)
2698 return ConstantInt::getTrue(RHSC->getContext());
2699 }
2700 if (RHSC->getValue().isNonNegative()) {
2701 if (Pred == ICmpInst::ICMP_SLE)
2702 return ConstantInt::getTrue(RHSC->getContext());
2703 if (Pred == ICmpInst::ICMP_SGT)
2704 return ConstantInt::getFalse(RHSC->getContext());
2705 }
2706 }
2707 }
2708
2709 // icmp pred (urem X, Y), Y
2710 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2711 bool KnownNonNegative, KnownNegative;
2712 switch (Pred) {
2713 default:
2714 break;
2715 case ICmpInst::ICMP_SGT:
2716 case ICmpInst::ICMP_SGE:
2717 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2718 Q.CxtI, Q.DT);
2719 if (!KnownNonNegative)
2720 break;
2721 // fall-through
2722 case ICmpInst::ICMP_EQ:
2723 case ICmpInst::ICMP_UGT:
2724 case ICmpInst::ICMP_UGE:
2725 return getFalse(ITy);
2726 case ICmpInst::ICMP_SLT:
2727 case ICmpInst::ICMP_SLE:
2728 ComputeSignBit(RHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2729 Q.CxtI, Q.DT);
2730 if (!KnownNonNegative)
2731 break;
2732 // fall-through
2733 case ICmpInst::ICMP_NE:
2734 case ICmpInst::ICMP_ULT:
2735 case ICmpInst::ICMP_ULE:
2736 return getTrue(ITy);
2737 }
2738 }
2739
2740 // icmp pred X, (urem Y, X)
2741 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2742 bool KnownNonNegative, KnownNegative;
2743 switch (Pred) {
2744 default:
2745 break;
2746 case ICmpInst::ICMP_SGT:
2747 case ICmpInst::ICMP_SGE:
2748 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2749 Q.CxtI, Q.DT);
2750 if (!KnownNonNegative)
2751 break;
2752 // fall-through
2753 case ICmpInst::ICMP_NE:
2754 case ICmpInst::ICMP_UGT:
2755 case ICmpInst::ICMP_UGE:
2756 return getTrue(ITy);
2757 case ICmpInst::ICMP_SLT:
2758 case ICmpInst::ICMP_SLE:
2759 ComputeSignBit(LHS, KnownNonNegative, KnownNegative, Q.DL, 0, Q.AC,
2760 Q.CxtI, Q.DT);
2761 if (!KnownNonNegative)
2762 break;
2763 // fall-through
2764 case ICmpInst::ICMP_EQ:
2765 case ICmpInst::ICMP_ULT:
2766 case ICmpInst::ICMP_ULE:
2767 return getFalse(ITy);
2768 }
2769 }
2770
2771 // x udiv y <=u x.
2772 if (LBO && match(LBO, m_UDiv(m_Specific(RHS), m_Value()))) {
2773 // icmp pred (X /u Y), X
2774 if (Pred == ICmpInst::ICMP_UGT)
2775 return getFalse(ITy);
2776 if (Pred == ICmpInst::ICMP_ULE)
2777 return getTrue(ITy);
2778 }
2779
2780 // handle:
2781 // CI2 << X == CI
2782 // CI2 << X != CI
2783 //
2784 // where CI2 is a power of 2 and CI isn't
2785 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2786 const APInt *CI2Val, *CIVal = &CI->getValue();
2787 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2788 CI2Val->isPowerOf2()) {
2789 if (!CIVal->isPowerOf2()) {
2790 // CI2 << X can equal zero in some circumstances,
2791 // this simplification is unsafe if CI is zero.
2792 //
2793 // We know it is safe if:
2794 // - The shift is nsw, we can't shift out the one bit.
2795 // - The shift is nuw, we can't shift out the one bit.
2796 // - CI2 is one
2797 // - CI isn't zero
2798 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2799 *CI2Val == 1 || !CI->isZero()) {
2800 if (Pred == ICmpInst::ICMP_EQ)
2801 return ConstantInt::getFalse(RHS->getContext());
2802 if (Pred == ICmpInst::ICMP_NE)
2803 return ConstantInt::getTrue(RHS->getContext());
2804 }
2805 }
2806 if (CIVal->isSignBit() && *CI2Val == 1) {
2807 if (Pred == ICmpInst::ICMP_UGT)
2808 return ConstantInt::getFalse(RHS->getContext());
2809 if (Pred == ICmpInst::ICMP_ULE)
2810 return ConstantInt::getTrue(RHS->getContext());
2811 }
2812 }
2813 }
2814
2815 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2816 LBO->getOperand(1) == RBO->getOperand(1)) {
2817 switch (LBO->getOpcode()) {
2818 default: break;
2819 case Instruction::UDiv:
2820 case Instruction::LShr:
2821 if (ICmpInst::isSigned(Pred))
2822 break;
2823 // fall-through
2824 case Instruction::SDiv:
2825 case Instruction::AShr:
2826 if (!LBO->isExact() || !RBO->isExact())
2827 break;
2828 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2829 RBO->getOperand(0), Q, MaxRecurse-1))
2830 return V;
2831 break;
2832 case Instruction::Shl: {
2833 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2834 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2835 if (!NUW && !NSW)
2836 break;
2837 if (!NSW && ICmpInst::isSigned(Pred))
2838 break;
2839 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2840 RBO->getOperand(0), Q, MaxRecurse-1))
2841 return V;
2842 break;
2843 }
2844 }
2845 }
2846
2847 // Simplify comparisons involving max/min.
2848 Value *A, *B;
2849 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2850 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2851
2852 // Signed variants on "max(a,b)>=a -> true".
2853 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2854 if (A != RHS) std::swap(A, B); // smax(A, B) pred A.
2855 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2856 // We analyze this as smax(A, B) pred A.
2857 P = Pred;
2858 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2859 (A == LHS || B == LHS)) {
2860 if (A != LHS) std::swap(A, B); // A pred smax(A, B).
2861 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2862 // We analyze this as smax(A, B) swapped-pred A.
2863 P = CmpInst::getSwappedPredicate(Pred);
2864 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2865 (A == RHS || B == RHS)) {
2866 if (A != RHS) std::swap(A, B); // smin(A, B) pred A.
2867 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2868 // We analyze this as smax(-A, -B) swapped-pred -A.
2869 // Note that we do not need to actually form -A or -B thanks to EqP.
2870 P = CmpInst::getSwappedPredicate(Pred);
2871 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2872 (A == LHS || B == LHS)) {
2873 if (A != LHS) std::swap(A, B); // A pred smin(A, B).
2874 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2875 // We analyze this as smax(-A, -B) pred -A.
2876 // Note that we do not need to actually form -A or -B thanks to EqP.
2877 P = Pred;
2878 }
2879 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2880 // Cases correspond to "max(A, B) p A".
2881 switch (P) {
2882 default:
2883 break;
2884 case CmpInst::ICMP_EQ:
2885 case CmpInst::ICMP_SLE:
2886 // Equivalent to "A EqP B". This may be the same as the condition tested
2887 // in the max/min; if so, we can just return that.
2888 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2889 return V;
2890 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2891 return V;
2892 // Otherwise, see if "A EqP B" simplifies.
2893 if (MaxRecurse)
2894 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2895 return V;
2896 break;
2897 case CmpInst::ICMP_NE:
2898 case CmpInst::ICMP_SGT: {
2899 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2900 // Equivalent to "A InvEqP B". This may be the same as the condition
2901 // tested in the max/min; if so, we can just return that.
2902 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2903 return V;
2904 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2905 return V;
2906 // Otherwise, see if "A InvEqP B" simplifies.
2907 if (MaxRecurse)
2908 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2909 return V;
2910 break;
2911 }
2912 case CmpInst::ICMP_SGE:
2913 // Always true.
2914 return getTrue(ITy);
2915 case CmpInst::ICMP_SLT:
2916 // Always false.
2917 return getFalse(ITy);
2918 }
2919 }
2920
2921 // Unsigned variants on "max(a,b)>=a -> true".
2922 P = CmpInst::BAD_ICMP_PREDICATE;
2923 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2924 if (A != RHS) std::swap(A, B); // umax(A, B) pred A.
2925 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2926 // We analyze this as umax(A, B) pred A.
2927 P = Pred;
2928 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2929 (A == LHS || B == LHS)) {
2930 if (A != LHS) std::swap(A, B); // A pred umax(A, B).
2931 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2932 // We analyze this as umax(A, B) swapped-pred A.
2933 P = CmpInst::getSwappedPredicate(Pred);
2934 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2935 (A == RHS || B == RHS)) {
2936 if (A != RHS) std::swap(A, B); // umin(A, B) pred A.
2937 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2938 // We analyze this as umax(-A, -B) swapped-pred -A.
2939 // Note that we do not need to actually form -A or -B thanks to EqP.
2940 P = CmpInst::getSwappedPredicate(Pred);
2941 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2942 (A == LHS || B == LHS)) {
2943 if (A != LHS) std::swap(A, B); // A pred umin(A, B).
2944 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2945 // We analyze this as umax(-A, -B) pred -A.
2946 // Note that we do not need to actually form -A or -B thanks to EqP.
2947 P = Pred;
2948 }
2949 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2950 // Cases correspond to "max(A, B) p A".
2951 switch (P) {
2952 default:
2953 break;
2954 case CmpInst::ICMP_EQ:
2955 case CmpInst::ICMP_ULE:
2956 // Equivalent to "A EqP B". This may be the same as the condition tested
2957 // in the max/min; if so, we can just return that.
2958 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2959 return V;
2960 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2961 return V;
2962 // Otherwise, see if "A EqP B" simplifies.
2963 if (MaxRecurse)
2964 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse-1))
2965 return V;
2966 break;
2967 case CmpInst::ICMP_NE:
2968 case CmpInst::ICMP_UGT: {
2969 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2970 // Equivalent to "A InvEqP B". This may be the same as the condition
2971 // tested in the max/min; if so, we can just return that.
2972 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2973 return V;
2974 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2975 return V;
2976 // Otherwise, see if "A InvEqP B" simplifies.
2977 if (MaxRecurse)
2978 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse-1))
2979 return V;
2980 break;
2981 }
2982 case CmpInst::ICMP_UGE:
2983 // Always true.
2984 return getTrue(ITy);
2985 case CmpInst::ICMP_ULT:
2986 // Always false.
2987 return getFalse(ITy);
2988 }
2989 }
2990
2991 // Variants on "max(x,y) >= min(x,z)".
2992 Value *C, *D;
2993 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2994 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2995 (A == C || A == D || B == C || B == D)) {
2996 // max(x, ?) pred min(x, ?).
2997 if (Pred == CmpInst::ICMP_SGE)
2998 // Always true.
2999 return getTrue(ITy);
3000 if (Pred == CmpInst::ICMP_SLT)
3001 // Always false.
3002 return getFalse(ITy);
3003 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3004 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3005 (A == C || A == D || B == C || B == D)) {
3006 // min(x, ?) pred max(x, ?).
3007 if (Pred == CmpInst::ICMP_SLE)
3008 // Always true.
3009 return getTrue(ITy);
3010 if (Pred == CmpInst::ICMP_SGT)
3011 // Always false.
3012 return getFalse(ITy);
3013 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3014 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3015 (A == C || A == D || B == C || B == D)) {
3016 // max(x, ?) pred min(x, ?).
3017 if (Pred == CmpInst::ICMP_UGE)
3018 // Always true.
3019 return getTrue(ITy);
3020 if (Pred == CmpInst::ICMP_ULT)
3021 // Always false.
3022 return getFalse(ITy);
3023 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3024 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3025 (A == C || A == D || B == C || B == D)) {
3026 // min(x, ?) pred max(x, ?).
3027 if (Pred == CmpInst::ICMP_ULE)
3028 // Always true.
3029 return getTrue(ITy);
3030 if (Pred == CmpInst::ICMP_UGT)
3031 // Always false.
3032 return getFalse(ITy);
3033 }
3034
3035 // Simplify comparisons of related pointers using a powerful, recursive
3036 // GEP-walk when we have target data available..
3037 if (LHS->getType()->isPointerTy())
3038 if (Constant *C = computePointerICmp(Q.DL, Q.TLI, Pred, LHS, RHS))
3039 return C;
3040
3041 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3042 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3043 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3044 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3045 (ICmpInst::isEquality(Pred) ||
3046 (GLHS->isInBounds() && GRHS->isInBounds() &&
3047 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3048 // The bases are equal and the indices are constant. Build a constant
3049 // expression GEP with the same indices and a null base pointer to see
3050 // what constant folding can make out of it.
3051 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3052 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3053 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3054 GLHS->getSourceElementType(), Null, IndicesLHS);
3055
3056 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3057 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3058 GLHS->getSourceElementType(), Null, IndicesRHS);
3059 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3060 }
3061 }
3062 }
3063
3064 // If a bit is known to be zero for A and known to be one for B,
3065 // then A and B cannot be equal.
3066 if (ICmpInst::isEquality(Pred)) {
3067 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3068 uint32_t BitWidth = CI->getBitWidth();
3069 APInt LHSKnownZero(BitWidth, 0);
3070 APInt LHSKnownOne(BitWidth, 0);
3071 computeKnownBits(LHS, LHSKnownZero, LHSKnownOne, Q.DL, /*Depth=*/0, Q.AC,
3072 Q.CxtI, Q.DT);
3073 const APInt &RHSVal = CI->getValue();
3074 if (((LHSKnownZero & RHSVal) != 0) || ((LHSKnownOne & ~RHSVal) != 0))
3075 return Pred == ICmpInst::ICMP_EQ
3076 ? ConstantInt::getFalse(CI->getContext())
3077 : ConstantInt::getTrue(CI->getContext());
3078 }
3079 }
3080
3081 // If the comparison is with the result of a select instruction, check whether
3082 // comparing with either branch of the select always yields the same value.
3083 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3084 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3085 return V;
3086
3087 // If the comparison is with the result of a phi instruction, check whether
3088 // doing the compare with each incoming phi value yields a common result.
3089 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3090 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3091 return V;
3092
3093 return nullptr;
3094 }
3095
SimplifyICmpInst(unsigned Predicate,Value * LHS,Value * RHS,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,Instruction * CxtI)3096 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3097 const DataLayout &DL,
3098 const TargetLibraryInfo *TLI,
3099 const DominatorTree *DT, AssumptionCache *AC,
3100 Instruction *CxtI) {
3101 return ::SimplifyICmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3102 RecursionLimit);
3103 }
3104
3105 /// SimplifyFCmpInst - Given operands for an FCmpInst, see if we can
3106 /// fold the result. If not, this returns null.
SimplifyFCmpInst(unsigned Predicate,Value * LHS,Value * RHS,FastMathFlags FMF,const Query & Q,unsigned MaxRecurse)3107 static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3108 FastMathFlags FMF, const Query &Q,
3109 unsigned MaxRecurse) {
3110 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3111 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3112
3113 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3114 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3115 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3116
3117 // If we have a constant, make sure it is on the RHS.
3118 std::swap(LHS, RHS);
3119 Pred = CmpInst::getSwappedPredicate(Pred);
3120 }
3121
3122 // Fold trivial predicates.
3123 if (Pred == FCmpInst::FCMP_FALSE)
3124 return ConstantInt::get(GetCompareTy(LHS), 0);
3125 if (Pred == FCmpInst::FCMP_TRUE)
3126 return ConstantInt::get(GetCompareTy(LHS), 1);
3127
3128 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3129 if (FMF.noNaNs()) {
3130 if (Pred == FCmpInst::FCMP_UNO)
3131 return ConstantInt::get(GetCompareTy(LHS), 0);
3132 if (Pred == FCmpInst::FCMP_ORD)
3133 return ConstantInt::get(GetCompareTy(LHS), 1);
3134 }
3135
3136 // fcmp pred x, undef and fcmp pred undef, x
3137 // fold to true if unordered, false if ordered
3138 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3139 // Choosing NaN for the undef will always make unordered comparison succeed
3140 // and ordered comparison fail.
3141 return ConstantInt::get(GetCompareTy(LHS), CmpInst::isUnordered(Pred));
3142 }
3143
3144 // fcmp x,x -> true/false. Not all compares are foldable.
3145 if (LHS == RHS) {
3146 if (CmpInst::isTrueWhenEqual(Pred))
3147 return ConstantInt::get(GetCompareTy(LHS), 1);
3148 if (CmpInst::isFalseWhenEqual(Pred))
3149 return ConstantInt::get(GetCompareTy(LHS), 0);
3150 }
3151
3152 // Handle fcmp with constant RHS
3153 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
3154 // If the constant is a nan, see if we can fold the comparison based on it.
3155 if (CFP->getValueAPF().isNaN()) {
3156 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3157 return ConstantInt::getFalse(CFP->getContext());
3158 assert(FCmpInst::isUnordered(Pred) &&
3159 "Comparison must be either ordered or unordered!");
3160 // True if unordered.
3161 return ConstantInt::getTrue(CFP->getContext());
3162 }
3163 // Check whether the constant is an infinity.
3164 if (CFP->getValueAPF().isInfinity()) {
3165 if (CFP->getValueAPF().isNegative()) {
3166 switch (Pred) {
3167 case FCmpInst::FCMP_OLT:
3168 // No value is ordered and less than negative infinity.
3169 return ConstantInt::getFalse(CFP->getContext());
3170 case FCmpInst::FCMP_UGE:
3171 // All values are unordered with or at least negative infinity.
3172 return ConstantInt::getTrue(CFP->getContext());
3173 default:
3174 break;
3175 }
3176 } else {
3177 switch (Pred) {
3178 case FCmpInst::FCMP_OGT:
3179 // No value is ordered and greater than infinity.
3180 return ConstantInt::getFalse(CFP->getContext());
3181 case FCmpInst::FCMP_ULE:
3182 // All values are unordered with and at most infinity.
3183 return ConstantInt::getTrue(CFP->getContext());
3184 default:
3185 break;
3186 }
3187 }
3188 }
3189 if (CFP->getValueAPF().isZero()) {
3190 switch (Pred) {
3191 case FCmpInst::FCMP_UGE:
3192 if (CannotBeOrderedLessThanZero(LHS))
3193 return ConstantInt::getTrue(CFP->getContext());
3194 break;
3195 case FCmpInst::FCMP_OLT:
3196 // X < 0
3197 if (CannotBeOrderedLessThanZero(LHS))
3198 return ConstantInt::getFalse(CFP->getContext());
3199 break;
3200 default:
3201 break;
3202 }
3203 }
3204 }
3205
3206 // If the comparison is with the result of a select instruction, check whether
3207 // comparing with either branch of the select always yields the same value.
3208 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3209 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3210 return V;
3211
3212 // If the comparison is with the result of a phi instruction, check whether
3213 // doing the compare with each incoming phi value yields a common result.
3214 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3215 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3216 return V;
3217
3218 return nullptr;
3219 }
3220
SimplifyFCmpInst(unsigned Predicate,Value * LHS,Value * RHS,FastMathFlags FMF,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3221 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3222 FastMathFlags FMF, const DataLayout &DL,
3223 const TargetLibraryInfo *TLI,
3224 const DominatorTree *DT, AssumptionCache *AC,
3225 const Instruction *CxtI) {
3226 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF,
3227 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3228 }
3229
3230 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is
3231 /// replaced with RepOp.
SimplifyWithOpReplaced(Value * V,Value * Op,Value * RepOp,const Query & Q,unsigned MaxRecurse)3232 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3233 const Query &Q,
3234 unsigned MaxRecurse) {
3235 // Trivial replacement.
3236 if (V == Op)
3237 return RepOp;
3238
3239 auto *I = dyn_cast<Instruction>(V);
3240 if (!I)
3241 return nullptr;
3242
3243 // If this is a binary operator, try to simplify it with the replaced op.
3244 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3245 // Consider:
3246 // %cmp = icmp eq i32 %x, 2147483647
3247 // %add = add nsw i32 %x, 1
3248 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3249 //
3250 // We can't replace %sel with %add unless we strip away the flags.
3251 if (isa<OverflowingBinaryOperator>(B))
3252 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3253 return nullptr;
3254 if (isa<PossiblyExactOperator>(B))
3255 if (B->isExact())
3256 return nullptr;
3257
3258 if (MaxRecurse) {
3259 if (B->getOperand(0) == Op)
3260 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3261 MaxRecurse - 1);
3262 if (B->getOperand(1) == Op)
3263 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3264 MaxRecurse - 1);
3265 }
3266 }
3267
3268 // Same for CmpInsts.
3269 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3270 if (MaxRecurse) {
3271 if (C->getOperand(0) == Op)
3272 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3273 MaxRecurse - 1);
3274 if (C->getOperand(1) == Op)
3275 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3276 MaxRecurse - 1);
3277 }
3278 }
3279
3280 // TODO: We could hand off more cases to instsimplify here.
3281
3282 // If all operands are constant after substituting Op for RepOp then we can
3283 // constant fold the instruction.
3284 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3285 // Build a list of all constant operands.
3286 SmallVector<Constant *, 8> ConstOps;
3287 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3288 if (I->getOperand(i) == Op)
3289 ConstOps.push_back(CRepOp);
3290 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3291 ConstOps.push_back(COp);
3292 else
3293 break;
3294 }
3295
3296 // All operands were constants, fold it.
3297 if (ConstOps.size() == I->getNumOperands()) {
3298 if (CmpInst *C = dyn_cast<CmpInst>(I))
3299 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3300 ConstOps[1], Q.DL, Q.TLI);
3301
3302 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3303 if (!LI->isVolatile())
3304 return ConstantFoldLoadFromConstPtr(ConstOps[0], Q.DL);
3305
3306 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), ConstOps,
3307 Q.DL, Q.TLI);
3308 }
3309 }
3310
3311 return nullptr;
3312 }
3313
3314 /// SimplifySelectInst - Given operands for a SelectInst, see if we can fold
3315 /// the result. If not, this returns null.
SimplifySelectInst(Value * CondVal,Value * TrueVal,Value * FalseVal,const Query & Q,unsigned MaxRecurse)3316 static Value *SimplifySelectInst(Value *CondVal, Value *TrueVal,
3317 Value *FalseVal, const Query &Q,
3318 unsigned MaxRecurse) {
3319 // select true, X, Y -> X
3320 // select false, X, Y -> Y
3321 if (Constant *CB = dyn_cast<Constant>(CondVal)) {
3322 if (CB->isAllOnesValue())
3323 return TrueVal;
3324 if (CB->isNullValue())
3325 return FalseVal;
3326 }
3327
3328 // select C, X, X -> X
3329 if (TrueVal == FalseVal)
3330 return TrueVal;
3331
3332 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
3333 if (isa<Constant>(TrueVal))
3334 return TrueVal;
3335 return FalseVal;
3336 }
3337 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
3338 return FalseVal;
3339 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
3340 return TrueVal;
3341
3342 if (const auto *ICI = dyn_cast<ICmpInst>(CondVal)) {
3343 unsigned BitWidth = Q.DL.getTypeSizeInBits(TrueVal->getType());
3344 ICmpInst::Predicate Pred = ICI->getPredicate();
3345 Value *CmpLHS = ICI->getOperand(0);
3346 Value *CmpRHS = ICI->getOperand(1);
3347 APInt MinSignedValue = APInt::getSignBit(BitWidth);
3348 Value *X;
3349 const APInt *Y;
3350 bool TrueWhenUnset;
3351 bool IsBitTest = false;
3352 if (ICmpInst::isEquality(Pred) &&
3353 match(CmpLHS, m_And(m_Value(X), m_APInt(Y))) &&
3354 match(CmpRHS, m_Zero())) {
3355 IsBitTest = true;
3356 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
3357 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
3358 X = CmpLHS;
3359 Y = &MinSignedValue;
3360 IsBitTest = true;
3361 TrueWhenUnset = false;
3362 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
3363 X = CmpLHS;
3364 Y = &MinSignedValue;
3365 IsBitTest = true;
3366 TrueWhenUnset = true;
3367 }
3368 if (IsBitTest) {
3369 const APInt *C;
3370 // (X & Y) == 0 ? X & ~Y : X --> X
3371 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3372 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3373 *Y == ~*C)
3374 return TrueWhenUnset ? FalseVal : TrueVal;
3375 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3376 // (X & Y) != 0 ? X : X & ~Y --> X
3377 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3378 *Y == ~*C)
3379 return TrueWhenUnset ? FalseVal : TrueVal;
3380
3381 if (Y->isPowerOf2()) {
3382 // (X & Y) == 0 ? X | Y : X --> X | Y
3383 // (X & Y) != 0 ? X | Y : X --> X
3384 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3385 *Y == *C)
3386 return TrueWhenUnset ? TrueVal : FalseVal;
3387 // (X & Y) == 0 ? X : X | Y --> X
3388 // (X & Y) != 0 ? X : X | Y --> X | Y
3389 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3390 *Y == *C)
3391 return TrueWhenUnset ? TrueVal : FalseVal;
3392 }
3393 }
3394 if (ICI->hasOneUse()) {
3395 const APInt *C;
3396 if (match(CmpRHS, m_APInt(C))) {
3397 // X < MIN ? T : F --> F
3398 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3399 return FalseVal;
3400 // X < MIN ? T : F --> F
3401 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3402 return FalseVal;
3403 // X > MAX ? T : F --> F
3404 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3405 return FalseVal;
3406 // X > MAX ? T : F --> F
3407 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3408 return FalseVal;
3409 }
3410 }
3411
3412 // If we have an equality comparison then we know the value in one of the
3413 // arms of the select. See if substituting this value into the arm and
3414 // simplifying the result yields the same value as the other arm.
3415 if (Pred == ICmpInst::ICMP_EQ) {
3416 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3417 TrueVal ||
3418 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3419 TrueVal)
3420 return FalseVal;
3421 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3422 FalseVal ||
3423 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3424 FalseVal)
3425 return FalseVal;
3426 } else if (Pred == ICmpInst::ICMP_NE) {
3427 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3428 FalseVal ||
3429 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3430 FalseVal)
3431 return TrueVal;
3432 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3433 TrueVal ||
3434 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3435 TrueVal)
3436 return TrueVal;
3437 }
3438 }
3439
3440 return nullptr;
3441 }
3442
SimplifySelectInst(Value * Cond,Value * TrueVal,Value * FalseVal,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3443 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3444 const DataLayout &DL,
3445 const TargetLibraryInfo *TLI,
3446 const DominatorTree *DT, AssumptionCache *AC,
3447 const Instruction *CxtI) {
3448 return ::SimplifySelectInst(Cond, TrueVal, FalseVal,
3449 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3450 }
3451
3452 /// SimplifyGEPInst - Given operands for an GetElementPtrInst, see if we can
3453 /// fold the result. If not, this returns null.
SimplifyGEPInst(Type * SrcTy,ArrayRef<Value * > Ops,const Query & Q,unsigned)3454 static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3455 const Query &Q, unsigned) {
3456 // The type of the GEP pointer operand.
3457 unsigned AS =
3458 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3459
3460 // getelementptr P -> P.
3461 if (Ops.size() == 1)
3462 return Ops[0];
3463
3464 // Compute the (pointer) type returned by the GEP instruction.
3465 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3466 Type *GEPTy = PointerType::get(LastType, AS);
3467 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3468 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3469
3470 if (isa<UndefValue>(Ops[0]))
3471 return UndefValue::get(GEPTy);
3472
3473 if (Ops.size() == 2) {
3474 // getelementptr P, 0 -> P.
3475 if (match(Ops[1], m_Zero()))
3476 return Ops[0];
3477
3478 Type *Ty = SrcTy;
3479 if (Ty->isSized()) {
3480 Value *P;
3481 uint64_t C;
3482 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3483 // getelementptr P, N -> P if P points to a type of zero size.
3484 if (TyAllocSize == 0)
3485 return Ops[0];
3486
3487 // The following transforms are only safe if the ptrtoint cast
3488 // doesn't truncate the pointers.
3489 if (Ops[1]->getType()->getScalarSizeInBits() ==
3490 Q.DL.getPointerSizeInBits(AS)) {
3491 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3492 if (match(P, m_Zero()))
3493 return Constant::getNullValue(GEPTy);
3494 Value *Temp;
3495 if (match(P, m_PtrToInt(m_Value(Temp))))
3496 if (Temp->getType() == GEPTy)
3497 return Temp;
3498 return nullptr;
3499 };
3500
3501 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3502 if (TyAllocSize == 1 &&
3503 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3504 if (Value *R = PtrToIntOrZero(P))
3505 return R;
3506
3507 // getelementptr V, (ashr (sub P, V), C) -> Q
3508 // if P points to a type of size 1 << C.
3509 if (match(Ops[1],
3510 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3511 m_ConstantInt(C))) &&
3512 TyAllocSize == 1ULL << C)
3513 if (Value *R = PtrToIntOrZero(P))
3514 return R;
3515
3516 // getelementptr V, (sdiv (sub P, V), C) -> Q
3517 // if P points to a type of size C.
3518 if (match(Ops[1],
3519 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3520 m_SpecificInt(TyAllocSize))))
3521 if (Value *R = PtrToIntOrZero(P))
3522 return R;
3523 }
3524 }
3525 }
3526
3527 // Check to see if this is constant foldable.
3528 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
3529 if (!isa<Constant>(Ops[i]))
3530 return nullptr;
3531
3532 return ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3533 Ops.slice(1));
3534 }
3535
SimplifyGEPInst(ArrayRef<Value * > Ops,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3536 Value *llvm::SimplifyGEPInst(ArrayRef<Value *> Ops, const DataLayout &DL,
3537 const TargetLibraryInfo *TLI,
3538 const DominatorTree *DT, AssumptionCache *AC,
3539 const Instruction *CxtI) {
3540 return ::SimplifyGEPInst(
3541 cast<PointerType>(Ops[0]->getType()->getScalarType())->getElementType(),
3542 Ops, Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3543 }
3544
3545 /// SimplifyInsertValueInst - Given operands for an InsertValueInst, see if we
3546 /// can fold the result. If not, this returns null.
SimplifyInsertValueInst(Value * Agg,Value * Val,ArrayRef<unsigned> Idxs,const Query & Q,unsigned)3547 static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3548 ArrayRef<unsigned> Idxs, const Query &Q,
3549 unsigned) {
3550 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3551 if (Constant *CVal = dyn_cast<Constant>(Val))
3552 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3553
3554 // insertvalue x, undef, n -> x
3555 if (match(Val, m_Undef()))
3556 return Agg;
3557
3558 // insertvalue x, (extractvalue y, n), n
3559 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3560 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3561 EV->getIndices() == Idxs) {
3562 // insertvalue undef, (extractvalue y, n), n -> y
3563 if (match(Agg, m_Undef()))
3564 return EV->getAggregateOperand();
3565
3566 // insertvalue y, (extractvalue y, n), n -> y
3567 if (Agg == EV->getAggregateOperand())
3568 return Agg;
3569 }
3570
3571 return nullptr;
3572 }
3573
SimplifyInsertValueInst(Value * Agg,Value * Val,ArrayRef<unsigned> Idxs,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3574 Value *llvm::SimplifyInsertValueInst(
3575 Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, const DataLayout &DL,
3576 const TargetLibraryInfo *TLI, const DominatorTree *DT, AssumptionCache *AC,
3577 const Instruction *CxtI) {
3578 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Query(DL, TLI, DT, AC, CxtI),
3579 RecursionLimit);
3580 }
3581
3582 /// SimplifyExtractValueInst - Given operands for an ExtractValueInst, see if we
3583 /// can fold the result. If not, this returns null.
SimplifyExtractValueInst(Value * Agg,ArrayRef<unsigned> Idxs,const Query &,unsigned)3584 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3585 const Query &, unsigned) {
3586 if (auto *CAgg = dyn_cast<Constant>(Agg))
3587 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3588
3589 // extractvalue x, (insertvalue y, elt, n), n -> elt
3590 unsigned NumIdxs = Idxs.size();
3591 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3592 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3593 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3594 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3595 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3596 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3597 Idxs.slice(0, NumCommonIdxs)) {
3598 if (NumIdxs == NumInsertValueIdxs)
3599 return IVI->getInsertedValueOperand();
3600 break;
3601 }
3602 }
3603
3604 return nullptr;
3605 }
3606
SimplifyExtractValueInst(Value * Agg,ArrayRef<unsigned> Idxs,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3607 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3608 const DataLayout &DL,
3609 const TargetLibraryInfo *TLI,
3610 const DominatorTree *DT,
3611 AssumptionCache *AC,
3612 const Instruction *CxtI) {
3613 return ::SimplifyExtractValueInst(Agg, Idxs, Query(DL, TLI, DT, AC, CxtI),
3614 RecursionLimit);
3615 }
3616
3617 /// SimplifyExtractElementInst - Given operands for an ExtractElementInst, see if we
3618 /// can fold the result. If not, this returns null.
SimplifyExtractElementInst(Value * Vec,Value * Idx,const Query &,unsigned)3619 static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const Query &,
3620 unsigned) {
3621 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3622 if (auto *CIdx = dyn_cast<Constant>(Idx))
3623 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3624
3625 // The index is not relevant if our vector is a splat.
3626 if (auto *Splat = CVec->getSplatValue())
3627 return Splat;
3628
3629 if (isa<UndefValue>(Vec))
3630 return UndefValue::get(Vec->getType()->getVectorElementType());
3631 }
3632
3633 // If extracting a specified index from the vector, see if we can recursively
3634 // find a previously computed scalar that was inserted into the vector.
3635 if (auto *IdxC = dyn_cast<ConstantInt>(Idx))
3636 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3637 return Elt;
3638
3639 return nullptr;
3640 }
3641
SimplifyExtractElementInst(Value * Vec,Value * Idx,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3642 Value *llvm::SimplifyExtractElementInst(
3643 Value *Vec, Value *Idx, const DataLayout &DL, const TargetLibraryInfo *TLI,
3644 const DominatorTree *DT, AssumptionCache *AC, const Instruction *CxtI) {
3645 return ::SimplifyExtractElementInst(Vec, Idx, Query(DL, TLI, DT, AC, CxtI),
3646 RecursionLimit);
3647 }
3648
3649 /// SimplifyPHINode - See if we can fold the given phi. If not, returns null.
SimplifyPHINode(PHINode * PN,const Query & Q)3650 static Value *SimplifyPHINode(PHINode *PN, const Query &Q) {
3651 // If all of the PHI's incoming values are the same then replace the PHI node
3652 // with the common value.
3653 Value *CommonValue = nullptr;
3654 bool HasUndefInput = false;
3655 for (Value *Incoming : PN->incoming_values()) {
3656 // If the incoming value is the phi node itself, it can safely be skipped.
3657 if (Incoming == PN) continue;
3658 if (isa<UndefValue>(Incoming)) {
3659 // Remember that we saw an undef value, but otherwise ignore them.
3660 HasUndefInput = true;
3661 continue;
3662 }
3663 if (CommonValue && Incoming != CommonValue)
3664 return nullptr; // Not the same, bail out.
3665 CommonValue = Incoming;
3666 }
3667
3668 // If CommonValue is null then all of the incoming values were either undef or
3669 // equal to the phi node itself.
3670 if (!CommonValue)
3671 return UndefValue::get(PN->getType());
3672
3673 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3674 // instruction, we cannot return X as the result of the PHI node unless it
3675 // dominates the PHI block.
3676 if (HasUndefInput)
3677 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3678
3679 return CommonValue;
3680 }
3681
SimplifyTruncInst(Value * Op,Type * Ty,const Query & Q,unsigned)3682 static Value *SimplifyTruncInst(Value *Op, Type *Ty, const Query &Q, unsigned) {
3683 if (Constant *C = dyn_cast<Constant>(Op))
3684 return ConstantFoldInstOperands(Instruction::Trunc, Ty, C, Q.DL, Q.TLI);
3685
3686 return nullptr;
3687 }
3688
SimplifyTruncInst(Value * Op,Type * Ty,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3689 Value *llvm::SimplifyTruncInst(Value *Op, Type *Ty, const DataLayout &DL,
3690 const TargetLibraryInfo *TLI,
3691 const DominatorTree *DT, AssumptionCache *AC,
3692 const Instruction *CxtI) {
3693 return ::SimplifyTruncInst(Op, Ty, Query(DL, TLI, DT, AC, CxtI),
3694 RecursionLimit);
3695 }
3696
3697 //=== Helper functions for higher up the class hierarchy.
3698
3699 /// SimplifyBinOp - Given operands for a BinaryOperator, see if we can
3700 /// fold the result. If not, this returns null.
SimplifyBinOp(unsigned Opcode,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)3701 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3702 const Query &Q, unsigned MaxRecurse) {
3703 switch (Opcode) {
3704 case Instruction::Add:
3705 return SimplifyAddInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3706 Q, MaxRecurse);
3707 case Instruction::FAdd:
3708 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3709
3710 case Instruction::Sub:
3711 return SimplifySubInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3712 Q, MaxRecurse);
3713 case Instruction::FSub:
3714 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3715
3716 case Instruction::Mul: return SimplifyMulInst (LHS, RHS, Q, MaxRecurse);
3717 case Instruction::FMul:
3718 return SimplifyFMulInst (LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3719 case Instruction::SDiv: return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
3720 case Instruction::UDiv: return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
3721 case Instruction::FDiv:
3722 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3723 case Instruction::SRem: return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
3724 case Instruction::URem: return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
3725 case Instruction::FRem:
3726 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3727 case Instruction::Shl:
3728 return SimplifyShlInst(LHS, RHS, /*isNSW*/false, /*isNUW*/false,
3729 Q, MaxRecurse);
3730 case Instruction::LShr:
3731 return SimplifyLShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3732 case Instruction::AShr:
3733 return SimplifyAShrInst(LHS, RHS, /*isExact*/false, Q, MaxRecurse);
3734 case Instruction::And: return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
3735 case Instruction::Or: return SimplifyOrInst (LHS, RHS, Q, MaxRecurse);
3736 case Instruction::Xor: return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
3737 default:
3738 if (Constant *CLHS = dyn_cast<Constant>(LHS))
3739 if (Constant *CRHS = dyn_cast<Constant>(RHS)) {
3740 Constant *COps[] = {CLHS, CRHS};
3741 return ConstantFoldInstOperands(Opcode, LHS->getType(), COps, Q.DL,
3742 Q.TLI);
3743 }
3744
3745 // If the operation is associative, try some generic simplifications.
3746 if (Instruction::isAssociative(Opcode))
3747 if (Value *V = SimplifyAssociativeBinOp(Opcode, LHS, RHS, Q, MaxRecurse))
3748 return V;
3749
3750 // If the operation is with the result of a select instruction check whether
3751 // operating on either branch of the select always yields the same value.
3752 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3753 if (Value *V = ThreadBinOpOverSelect(Opcode, LHS, RHS, Q, MaxRecurse))
3754 return V;
3755
3756 // If the operation is with the result of a phi instruction, check whether
3757 // operating on all incoming values of the phi always yields the same value.
3758 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3759 if (Value *V = ThreadBinOpOverPHI(Opcode, LHS, RHS, Q, MaxRecurse))
3760 return V;
3761
3762 return nullptr;
3763 }
3764 }
3765
3766 /// SimplifyFPBinOp - Given operands for a BinaryOperator, see if we can
3767 /// fold the result. If not, this returns null.
3768 /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
3769 /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
SimplifyFPBinOp(unsigned Opcode,Value * LHS,Value * RHS,const FastMathFlags & FMF,const Query & Q,unsigned MaxRecurse)3770 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3771 const FastMathFlags &FMF, const Query &Q,
3772 unsigned MaxRecurse) {
3773 switch (Opcode) {
3774 case Instruction::FAdd:
3775 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
3776 case Instruction::FSub:
3777 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
3778 case Instruction::FMul:
3779 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
3780 default:
3781 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
3782 }
3783 }
3784
SimplifyBinOp(unsigned Opcode,Value * LHS,Value * RHS,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3785 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3786 const DataLayout &DL, const TargetLibraryInfo *TLI,
3787 const DominatorTree *DT, AssumptionCache *AC,
3788 const Instruction *CxtI) {
3789 return ::SimplifyBinOp(Opcode, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3790 RecursionLimit);
3791 }
3792
SimplifyFPBinOp(unsigned Opcode,Value * LHS,Value * RHS,const FastMathFlags & FMF,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3793 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
3794 const FastMathFlags &FMF, const DataLayout &DL,
3795 const TargetLibraryInfo *TLI,
3796 const DominatorTree *DT, AssumptionCache *AC,
3797 const Instruction *CxtI) {
3798 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Query(DL, TLI, DT, AC, CxtI),
3799 RecursionLimit);
3800 }
3801
3802 /// SimplifyCmpInst - Given operands for a CmpInst, see if we can
3803 /// fold the result.
SimplifyCmpInst(unsigned Predicate,Value * LHS,Value * RHS,const Query & Q,unsigned MaxRecurse)3804 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3805 const Query &Q, unsigned MaxRecurse) {
3806 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
3807 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
3808 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
3809 }
3810
SimplifyCmpInst(unsigned Predicate,Value * LHS,Value * RHS,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3811 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3812 const DataLayout &DL, const TargetLibraryInfo *TLI,
3813 const DominatorTree *DT, AssumptionCache *AC,
3814 const Instruction *CxtI) {
3815 return ::SimplifyCmpInst(Predicate, LHS, RHS, Query(DL, TLI, DT, AC, CxtI),
3816 RecursionLimit);
3817 }
3818
IsIdempotent(Intrinsic::ID ID)3819 static bool IsIdempotent(Intrinsic::ID ID) {
3820 switch (ID) {
3821 default: return false;
3822
3823 // Unary idempotent: f(f(x)) = f(x)
3824 case Intrinsic::fabs:
3825 case Intrinsic::floor:
3826 case Intrinsic::ceil:
3827 case Intrinsic::trunc:
3828 case Intrinsic::rint:
3829 case Intrinsic::nearbyint:
3830 case Intrinsic::round:
3831 return true;
3832 }
3833 }
3834
3835 template <typename IterTy>
SimplifyIntrinsic(Function * F,IterTy ArgBegin,IterTy ArgEnd,const Query & Q,unsigned MaxRecurse)3836 static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
3837 const Query &Q, unsigned MaxRecurse) {
3838 Intrinsic::ID IID = F->getIntrinsicID();
3839 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
3840 Type *ReturnType = F->getReturnType();
3841
3842 // Binary Ops
3843 if (NumOperands == 2) {
3844 Value *LHS = *ArgBegin;
3845 Value *RHS = *(ArgBegin + 1);
3846 if (IID == Intrinsic::usub_with_overflow ||
3847 IID == Intrinsic::ssub_with_overflow) {
3848 // X - X -> { 0, false }
3849 if (LHS == RHS)
3850 return Constant::getNullValue(ReturnType);
3851
3852 // X - undef -> undef
3853 // undef - X -> undef
3854 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
3855 return UndefValue::get(ReturnType);
3856 }
3857
3858 if (IID == Intrinsic::uadd_with_overflow ||
3859 IID == Intrinsic::sadd_with_overflow) {
3860 // X + undef -> undef
3861 if (isa<UndefValue>(RHS))
3862 return UndefValue::get(ReturnType);
3863 }
3864
3865 if (IID == Intrinsic::umul_with_overflow ||
3866 IID == Intrinsic::smul_with_overflow) {
3867 // X * 0 -> { 0, false }
3868 if (match(RHS, m_Zero()))
3869 return Constant::getNullValue(ReturnType);
3870
3871 // X * undef -> { 0, false }
3872 if (match(RHS, m_Undef()))
3873 return Constant::getNullValue(ReturnType);
3874 }
3875 }
3876
3877 // Perform idempotent optimizations
3878 if (!IsIdempotent(IID))
3879 return nullptr;
3880
3881 // Unary Ops
3882 if (NumOperands == 1)
3883 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin))
3884 if (II->getIntrinsicID() == IID)
3885 return II;
3886
3887 return nullptr;
3888 }
3889
3890 template <typename IterTy>
SimplifyCall(Value * V,IterTy ArgBegin,IterTy ArgEnd,const Query & Q,unsigned MaxRecurse)3891 static Value *SimplifyCall(Value *V, IterTy ArgBegin, IterTy ArgEnd,
3892 const Query &Q, unsigned MaxRecurse) {
3893 Type *Ty = V->getType();
3894 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
3895 Ty = PTy->getElementType();
3896 FunctionType *FTy = cast<FunctionType>(Ty);
3897
3898 // call undef -> undef
3899 if (isa<UndefValue>(V))
3900 return UndefValue::get(FTy->getReturnType());
3901
3902 Function *F = dyn_cast<Function>(V);
3903 if (!F)
3904 return nullptr;
3905
3906 if (F->isIntrinsic())
3907 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
3908 return Ret;
3909
3910 if (!canConstantFoldCallTo(F))
3911 return nullptr;
3912
3913 SmallVector<Constant *, 4> ConstantArgs;
3914 ConstantArgs.reserve(ArgEnd - ArgBegin);
3915 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
3916 Constant *C = dyn_cast<Constant>(*I);
3917 if (!C)
3918 return nullptr;
3919 ConstantArgs.push_back(C);
3920 }
3921
3922 return ConstantFoldCall(F, ConstantArgs, Q.TLI);
3923 }
3924
SimplifyCall(Value * V,User::op_iterator ArgBegin,User::op_iterator ArgEnd,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3925 Value *llvm::SimplifyCall(Value *V, User::op_iterator ArgBegin,
3926 User::op_iterator ArgEnd, const DataLayout &DL,
3927 const TargetLibraryInfo *TLI, const DominatorTree *DT,
3928 AssumptionCache *AC, const Instruction *CxtI) {
3929 return ::SimplifyCall(V, ArgBegin, ArgEnd, Query(DL, TLI, DT, AC, CxtI),
3930 RecursionLimit);
3931 }
3932
SimplifyCall(Value * V,ArrayRef<Value * > Args,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC,const Instruction * CxtI)3933 Value *llvm::SimplifyCall(Value *V, ArrayRef<Value *> Args,
3934 const DataLayout &DL, const TargetLibraryInfo *TLI,
3935 const DominatorTree *DT, AssumptionCache *AC,
3936 const Instruction *CxtI) {
3937 return ::SimplifyCall(V, Args.begin(), Args.end(),
3938 Query(DL, TLI, DT, AC, CxtI), RecursionLimit);
3939 }
3940
3941 /// SimplifyInstruction - See if we can compute a simplified version of this
3942 /// instruction. If not, this returns null.
SimplifyInstruction(Instruction * I,const DataLayout & DL,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC)3943 Value *llvm::SimplifyInstruction(Instruction *I, const DataLayout &DL,
3944 const TargetLibraryInfo *TLI,
3945 const DominatorTree *DT, AssumptionCache *AC) {
3946 Value *Result;
3947
3948 switch (I->getOpcode()) {
3949 default:
3950 Result = ConstantFoldInstruction(I, DL, TLI);
3951 break;
3952 case Instruction::FAdd:
3953 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
3954 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3955 break;
3956 case Instruction::Add:
3957 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
3958 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3959 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3960 TLI, DT, AC, I);
3961 break;
3962 case Instruction::FSub:
3963 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
3964 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3965 break;
3966 case Instruction::Sub:
3967 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
3968 cast<BinaryOperator>(I)->hasNoSignedWrap(),
3969 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
3970 TLI, DT, AC, I);
3971 break;
3972 case Instruction::FMul:
3973 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
3974 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3975 break;
3976 case Instruction::Mul:
3977 Result =
3978 SimplifyMulInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
3979 break;
3980 case Instruction::SDiv:
3981 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3982 AC, I);
3983 break;
3984 case Instruction::UDiv:
3985 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3986 AC, I);
3987 break;
3988 case Instruction::FDiv:
3989 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
3990 I->getFastMathFlags(), DL, TLI, DT, AC, I);
3991 break;
3992 case Instruction::SRem:
3993 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3994 AC, I);
3995 break;
3996 case Instruction::URem:
3997 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT,
3998 AC, I);
3999 break;
4000 case Instruction::FRem:
4001 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4002 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4003 break;
4004 case Instruction::Shl:
4005 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4006 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4007 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), DL,
4008 TLI, DT, AC, I);
4009 break;
4010 case Instruction::LShr:
4011 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4012 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
4013 AC, I);
4014 break;
4015 case Instruction::AShr:
4016 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4017 cast<BinaryOperator>(I)->isExact(), DL, TLI, DT,
4018 AC, I);
4019 break;
4020 case Instruction::And:
4021 Result =
4022 SimplifyAndInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
4023 break;
4024 case Instruction::Or:
4025 Result =
4026 SimplifyOrInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
4027 break;
4028 case Instruction::Xor:
4029 Result =
4030 SimplifyXorInst(I->getOperand(0), I->getOperand(1), DL, TLI, DT, AC, I);
4031 break;
4032 case Instruction::ICmp:
4033 Result =
4034 SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), I->getOperand(0),
4035 I->getOperand(1), DL, TLI, DT, AC, I);
4036 break;
4037 case Instruction::FCmp:
4038 Result = SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(),
4039 I->getOperand(0), I->getOperand(1),
4040 I->getFastMathFlags(), DL, TLI, DT, AC, I);
4041 break;
4042 case Instruction::Select:
4043 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4044 I->getOperand(2), DL, TLI, DT, AC, I);
4045 break;
4046 case Instruction::GetElementPtr: {
4047 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end());
4048 Result = SimplifyGEPInst(Ops, DL, TLI, DT, AC, I);
4049 break;
4050 }
4051 case Instruction::InsertValue: {
4052 InsertValueInst *IV = cast<InsertValueInst>(I);
4053 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4054 IV->getInsertedValueOperand(),
4055 IV->getIndices(), DL, TLI, DT, AC, I);
4056 break;
4057 }
4058 case Instruction::ExtractValue: {
4059 auto *EVI = cast<ExtractValueInst>(I);
4060 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4061 EVI->getIndices(), DL, TLI, DT, AC, I);
4062 break;
4063 }
4064 case Instruction::ExtractElement: {
4065 auto *EEI = cast<ExtractElementInst>(I);
4066 Result = SimplifyExtractElementInst(
4067 EEI->getVectorOperand(), EEI->getIndexOperand(), DL, TLI, DT, AC, I);
4068 break;
4069 }
4070 case Instruction::PHI:
4071 Result = SimplifyPHINode(cast<PHINode>(I), Query(DL, TLI, DT, AC, I));
4072 break;
4073 case Instruction::Call: {
4074 CallSite CS(cast<CallInst>(I));
4075 Result = SimplifyCall(CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), DL,
4076 TLI, DT, AC, I);
4077 break;
4078 }
4079 case Instruction::Trunc:
4080 Result =
4081 SimplifyTruncInst(I->getOperand(0), I->getType(), DL, TLI, DT, AC, I);
4082 break;
4083 }
4084
4085 // In general, it is possible for computeKnownBits to determine all bits in a
4086 // value even when the operands are not all constants.
4087 if (!Result && I->getType()->isIntegerTy()) {
4088 unsigned BitWidth = I->getType()->getScalarSizeInBits();
4089 APInt KnownZero(BitWidth, 0);
4090 APInt KnownOne(BitWidth, 0);
4091 computeKnownBits(I, KnownZero, KnownOne, DL, /*Depth*/0, AC, I, DT);
4092 if ((KnownZero | KnownOne).isAllOnesValue())
4093 Result = ConstantInt::get(I->getContext(), KnownOne);
4094 }
4095
4096 /// If called on unreachable code, the above logic may report that the
4097 /// instruction simplified to itself. Make life easier for users by
4098 /// detecting that case here, returning a safe value instead.
4099 return Result == I ? UndefValue::get(I->getType()) : Result;
4100 }
4101
4102 /// \brief Implementation of recursive simplification through an instructions
4103 /// uses.
4104 ///
4105 /// This is the common implementation of the recursive simplification routines.
4106 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4107 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4108 /// instructions to process and attempt to simplify it using
4109 /// InstructionSimplify.
4110 ///
4111 /// This routine returns 'true' only when *it* simplifies something. The passed
4112 /// in simplified value does not count toward this.
replaceAndRecursivelySimplifyImpl(Instruction * I,Value * SimpleV,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC)4113 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4114 const TargetLibraryInfo *TLI,
4115 const DominatorTree *DT,
4116 AssumptionCache *AC) {
4117 bool Simplified = false;
4118 SmallSetVector<Instruction *, 8> Worklist;
4119 const DataLayout &DL = I->getModule()->getDataLayout();
4120
4121 // If we have an explicit value to collapse to, do that round of the
4122 // simplification loop by hand initially.
4123 if (SimpleV) {
4124 for (User *U : I->users())
4125 if (U != I)
4126 Worklist.insert(cast<Instruction>(U));
4127
4128 // Replace the instruction with its simplified value.
4129 I->replaceAllUsesWith(SimpleV);
4130
4131 // Gracefully handle edge cases where the instruction is not wired into any
4132 // parent block.
4133 if (I->getParent())
4134 I->eraseFromParent();
4135 } else {
4136 Worklist.insert(I);
4137 }
4138
4139 // Note that we must test the size on each iteration, the worklist can grow.
4140 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4141 I = Worklist[Idx];
4142
4143 // See if this instruction simplifies.
4144 SimpleV = SimplifyInstruction(I, DL, TLI, DT, AC);
4145 if (!SimpleV)
4146 continue;
4147
4148 Simplified = true;
4149
4150 // Stash away all the uses of the old instruction so we can check them for
4151 // recursive simplifications after a RAUW. This is cheaper than checking all
4152 // uses of To on the recursive step in most cases.
4153 for (User *U : I->users())
4154 Worklist.insert(cast<Instruction>(U));
4155
4156 // Replace the instruction with its simplified value.
4157 I->replaceAllUsesWith(SimpleV);
4158
4159 // Gracefully handle edge cases where the instruction is not wired into any
4160 // parent block.
4161 if (I->getParent())
4162 I->eraseFromParent();
4163 }
4164 return Simplified;
4165 }
4166
recursivelySimplifyInstruction(Instruction * I,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC)4167 bool llvm::recursivelySimplifyInstruction(Instruction *I,
4168 const TargetLibraryInfo *TLI,
4169 const DominatorTree *DT,
4170 AssumptionCache *AC) {
4171 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4172 }
4173
replaceAndRecursivelySimplify(Instruction * I,Value * SimpleV,const TargetLibraryInfo * TLI,const DominatorTree * DT,AssumptionCache * AC)4174 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4175 const TargetLibraryInfo *TLI,
4176 const DominatorTree *DT,
4177 AssumptionCache *AC) {
4178 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
4179 assert(SimpleV && "Must provide a simplified value.");
4180 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4181 }
4182