Line data Source code
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/AssumptionCache.h"
25 : #include "llvm/Analysis/CaptureTracking.h"
26 : #include "llvm/Analysis/CmpInstAnalysis.h"
27 : #include "llvm/Analysis/ConstantFolding.h"
28 : #include "llvm/Analysis/LoopAnalysisManager.h"
29 : #include "llvm/Analysis/MemoryBuiltins.h"
30 : #include "llvm/Analysis/ValueTracking.h"
31 : #include "llvm/Analysis/VectorUtils.h"
32 : #include "llvm/IR/ConstantRange.h"
33 : #include "llvm/IR/DataLayout.h"
34 : #include "llvm/IR/Dominators.h"
35 : #include "llvm/IR/GetElementPtrTypeIterator.h"
36 : #include "llvm/IR/GlobalAlias.h"
37 : #include "llvm/IR/Operator.h"
38 : #include "llvm/IR/PatternMatch.h"
39 : #include "llvm/IR/ValueHandle.h"
40 : #include "llvm/Support/KnownBits.h"
41 : #include <algorithm>
42 : using namespace llvm;
43 : using namespace llvm::PatternMatch;
44 :
45 : #define DEBUG_TYPE "instsimplify"
46 :
47 : enum { RecursionLimit = 3 };
48 :
49 : STATISTIC(NumExpand, "Number of expansions");
50 : STATISTIC(NumReassoc, "Number of reassociations");
51 :
52 : static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
53 : static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
54 : unsigned);
55 : static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
56 : const SimplifyQuery &, unsigned);
57 : static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
58 : unsigned);
59 : static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
60 : const SimplifyQuery &Q, unsigned MaxRecurse);
61 : static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
62 : static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
63 : static Value *SimplifyCastInst(unsigned, Value *, Type *,
64 : const SimplifyQuery &, unsigned);
65 : static Value *SimplifyGEPInst(Type *, ArrayRef<Value *>, const SimplifyQuery &,
66 : unsigned);
67 :
68 118531 : static Value *foldSelectWithBinaryOp(Value *Cond, Value *TrueVal,
69 : Value *FalseVal) {
70 : BinaryOperator::BinaryOps BinOpCode;
71 : if (auto *BO = dyn_cast<BinaryOperator>(Cond))
72 : BinOpCode = BO->getOpcode();
73 : else
74 : return nullptr;
75 :
76 : CmpInst::Predicate ExpectedPred, Pred1, Pred2;
77 999 : if (BinOpCode == BinaryOperator::Or) {
78 : ExpectedPred = ICmpInst::ICMP_NE;
79 348 : } else if (BinOpCode == BinaryOperator::And) {
80 : ExpectedPred = ICmpInst::ICMP_EQ;
81 : } else
82 : return nullptr;
83 :
84 : // %A = icmp eq %TV, %FV
85 : // %B = icmp eq %X, %Y (and one of these is a select operand)
86 : // %C = and %A, %B
87 : // %D = select %C, %TV, %FV
88 : // -->
89 : // %FV
90 :
91 : // %A = icmp ne %TV, %FV
92 : // %B = icmp ne %X, %Y (and one of these is a select operand)
93 : // %C = or %A, %B
94 : // %D = select %C, %TV, %FV
95 : // -->
96 : // %TV
97 : Value *X, *Y;
98 933 : if (!match(Cond, m_c_BinOp(m_c_ICmp(Pred1, m_Specific(TrueVal),
99 : m_Specific(FalseVal)),
100 578 : m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
101 953 : Pred1 != Pred2 || Pred1 != ExpectedPred)
102 933 : return nullptr;
103 :
104 20 : if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
105 30 : return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
106 :
107 : return nullptr;
108 : }
109 :
110 : /// For a boolean type or a vector of boolean type, return false or a vector
111 : /// with every element false.
112 : static Constant *getFalse(Type *Ty) {
113 4705 : return ConstantInt::getFalse(Ty);
114 : }
115 :
116 : /// For a boolean type or a vector of boolean type, return true or a vector
117 : /// with every element true.
118 : static Constant *getTrue(Type *Ty) {
119 455 : return ConstantInt::getTrue(Ty);
120 : }
121 :
122 : /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
123 13812 : static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
124 : Value *RHS) {
125 : CmpInst *Cmp = dyn_cast<CmpInst>(V);
126 : if (!Cmp)
127 : return false;
128 : CmpInst::Predicate CPred = Cmp->getPredicate();
129 : Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
130 12451 : if (CPred == Pred && CLHS == LHS && CRHS == RHS)
131 : return true;
132 12428 : return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
133 4848 : CRHS == LHS;
134 : }
135 :
136 : /// Does the given value dominate the specified phi node?
137 198374 : static bool valueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
138 : Instruction *I = dyn_cast<Instruction>(V);
139 : if (!I)
140 : // Arguments and constants dominate all instructions.
141 : return true;
142 :
143 : // If we are processing instructions (and/or basic blocks) that have not been
144 : // fully added to a function, the parent nodes may still be null. Simply
145 : // return the conservative answer in these cases.
146 167482 : if (!I->getParent() || !P->getParent() || !I->getFunction())
147 11387 : return false;
148 :
149 : // If we have a DominatorTree then do a precise test.
150 72354 : if (DT)
151 57988 : return DT->dominates(I, P);
152 :
153 : // Otherwise, if the instruction is in the entry block and is not an invoke,
154 : // then it obviously dominates all phi nodes.
155 28732 : if (I->getParent() == &I->getFunction()->getEntryBlock() &&
156 : !isa<InvokeInst>(I))
157 295 : return true;
158 :
159 : return false;
160 : }
161 :
162 : /// Simplify "A op (B op' C)" by distributing op over op', turning it into
163 : /// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
164 : /// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
165 : /// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
166 : /// Returns the simplified value, or null if no simplification was performed.
167 865295 : static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
168 : Instruction::BinaryOps OpcodeToExpand,
169 : const SimplifyQuery &Q, unsigned MaxRecurse) {
170 : // Recursion is always used, so bail out at once if we already hit the limit.
171 865295 : if (!MaxRecurse--)
172 : return nullptr;
173 :
174 : // Check whether the expression has the form "(A op' B) op C".
175 : if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
176 188998 : if (Op0->getOpcode() == OpcodeToExpand) {
177 : // It does! Try turning it into "(A op C) op' (B op C)".
178 : Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
179 : // Do "A op C" and "B op C" both simplify?
180 39270 : if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
181 2321 : if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
182 : // They do! Return "L op' R" if it simplifies or is already available.
183 : // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
184 750 : if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
185 692 : && L == B && R == A)) {
186 : ++NumExpand;
187 : return LHS;
188 : }
189 : // Otherwise return "L op' R" if it simplifies.
190 692 : if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
191 : ++NumExpand;
192 : return V;
193 : }
194 : }
195 : }
196 :
197 : // Check whether the expression has the form "A op (B op' C)".
198 : if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
199 155658 : if (Op1->getOpcode() == OpcodeToExpand) {
200 : // It does! Try turning it into "(A op B) op' (A op C)".
201 : Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
202 : // Do "A op B" and "A op C" both simplify?
203 30915 : if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
204 1090 : if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
205 : // They do! Return "L op' R" if it simplifies or is already available.
206 : // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
207 41 : if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
208 41 : && L == C && R == B)) {
209 : ++NumExpand;
210 : return RHS;
211 : }
212 : // Otherwise return "L op' R" if it simplifies.
213 41 : if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
214 : ++NumExpand;
215 2 : return V;
216 : }
217 : }
218 : }
219 :
220 : return nullptr;
221 : }
222 :
223 : /// Generic simplifications for associative binary operations.
224 : /// Returns the simpler value, or null if none was found.
225 11889938 : static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
226 : Value *LHS, Value *RHS,
227 : const SimplifyQuery &Q,
228 : unsigned MaxRecurse) {
229 : assert(Instruction::isAssociative(Opcode) && "Not an associative operation!");
230 :
231 : // Recursion is always used, so bail out at once if we already hit the limit.
232 11889938 : if (!MaxRecurse--)
233 : return nullptr;
234 :
235 : BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
236 : BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
237 :
238 : // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
239 11775167 : if (Op0 && Op0->getOpcode() == Opcode) {
240 : Value *A = Op0->getOperand(0);
241 : Value *B = Op0->getOperand(1);
242 : Value *C = RHS;
243 :
244 : // Does "B op C" simplify?
245 137367 : if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
246 : // It does! Return "A op V" if it simplifies or is already available.
247 : // If V equals B then "A op V" is just the LHS.
248 89798 : if (V == B) return LHS;
249 : // Otherwise return "A op V" if it simplifies.
250 89256 : if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
251 : ++NumReassoc;
252 : return W;
253 : }
254 : }
255 : }
256 :
257 : // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
258 11771500 : if (Op1 && Op1->getOpcode() == Opcode) {
259 : Value *A = LHS;
260 : Value *B = Op1->getOperand(0);
261 : Value *C = Op1->getOperand(1);
262 :
263 : // Does "A op B" simplify?
264 43944 : if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
265 : // It does! Return "V op C" if it simplifies or is already available.
266 : // If V equals B then "V op C" is just the RHS.
267 26 : if (V == B) return RHS;
268 : // Otherwise return "V op C" if it simplifies.
269 2 : if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
270 : ++NumReassoc;
271 : return W;
272 : }
273 : }
274 : }
275 :
276 : // The remaining transforms require commutativity as well as associativity.
277 : if (!Instruction::isCommutative(Opcode))
278 : return nullptr;
279 :
280 : // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
281 11771475 : if (Op0 && Op0->getOpcode() == Opcode) {
282 : Value *A = Op0->getOperand(0);
283 : Value *B = Op0->getOperand(1);
284 : Value *C = RHS;
285 :
286 : // Does "C op A" simplify?
287 133700 : if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
288 : // It does! Return "V op B" if it simplifies or is already available.
289 : // If V equals A then "V op B" is just the LHS.
290 165 : if (V == A) return LHS;
291 : // Otherwise return "V op B" if it simplifies.
292 54 : if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
293 : ++NumReassoc;
294 : return W;
295 : }
296 : }
297 : }
298 :
299 : // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
300 11771359 : if (Op1 && Op1->getOpcode() == Opcode) {
301 : Value *A = LHS;
302 : Value *B = Op1->getOperand(0);
303 : Value *C = Op1->getOperand(1);
304 :
305 : // Does "C op A" simplify?
306 43919 : if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
307 : // It does! Return "B op V" if it simplifies or is already available.
308 : // If V equals C then "B op V" is just the RHS.
309 107 : if (V == C) return RHS;
310 : // Otherwise return "B op V" if it simplifies.
311 99 : if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
312 : ++NumReassoc;
313 3 : return W;
314 : }
315 : }
316 : }
317 :
318 : return nullptr;
319 : }
320 :
321 : /// In the case of a binary operation with a select instruction as an operand,
322 : /// try to simplify the binop by seeing whether evaluating it on both branches
323 : /// of the select results in the same value. Returns the common value if so,
324 : /// otherwise returns null.
325 6745 : static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
326 : Value *RHS, const SimplifyQuery &Q,
327 : unsigned MaxRecurse) {
328 : // Recursion is always used, so bail out at once if we already hit the limit.
329 6745 : if (!MaxRecurse--)
330 : return nullptr;
331 :
332 : SelectInst *SI;
333 : if (isa<SelectInst>(LHS)) {
334 : SI = cast<SelectInst>(LHS);
335 : } else {
336 : assert(isa<SelectInst>(RHS) && "No select instruction operand!");
337 : SI = cast<SelectInst>(RHS);
338 : }
339 :
340 : // Evaluate the BinOp on the true and false branches of the select.
341 : Value *TV;
342 : Value *FV;
343 6438 : if (SI == LHS) {
344 5409 : TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
345 5409 : FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
346 : } else {
347 1029 : TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
348 1029 : FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
349 : }
350 :
351 : // If they simplified to the same value, then return the common value.
352 : // If they both failed to simplify then return null.
353 6438 : if (TV == FV)
354 : return TV;
355 :
356 : // If one branch simplified to undef, return the other one.
357 4197 : if (TV && isa<UndefValue>(TV))
358 : return FV;
359 4194 : if (FV && isa<UndefValue>(FV))
360 : return TV;
361 :
362 : // If applying the operation did not change the true and false select values,
363 : // then the result of the binop is the select itself.
364 4187 : if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
365 : return SI;
366 :
367 : // If one branch simplified and the other did not, and the simplified
368 : // value is equal to the unsimplified one, return the simplified value.
369 : // For example, select (cond, X, X & Z) & Z -> X & Z.
370 4185 : if ((FV && !TV) || (TV && !FV)) {
371 : // Check that the simplified value has the form "X op Y" where "op" is the
372 : // same as the original operation.
373 3341 : Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
374 957 : if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
375 : // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
376 : // We already know that "op" is the same as for the simplified value. See
377 : // if the operands match too. If so, return the simplified value.
378 143 : Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
379 143 : Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
380 143 : Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
381 286 : if (Simplified->getOperand(0) == UnsimplifiedLHS &&
382 : Simplified->getOperand(1) == UnsimplifiedRHS)
383 : return Simplified;
384 143 : if (Simplified->isCommutative() &&
385 0 : Simplified->getOperand(1) == UnsimplifiedLHS &&
386 : Simplified->getOperand(0) == UnsimplifiedRHS)
387 0 : return Simplified;
388 : }
389 : }
390 :
391 : return nullptr;
392 : }
393 :
394 : /// In the case of a comparison with a select instruction, try to simplify the
395 : /// comparison by seeing whether both branches of the select result in the same
396 : /// value. Returns the common value if so, otherwise returns null.
397 14235 : static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
398 : Value *RHS, const SimplifyQuery &Q,
399 : unsigned MaxRecurse) {
400 : // Recursion is always used, so bail out at once if we already hit the limit.
401 14235 : if (!MaxRecurse--)
402 : return nullptr;
403 :
404 : // Make sure the select is on the LHS.
405 : if (!isa<SelectInst>(LHS)) {
406 : std::swap(LHS, RHS);
407 3178 : Pred = CmpInst::getSwappedPredicate(Pred);
408 : }
409 : assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!");
410 : SelectInst *SI = cast<SelectInst>(LHS);
411 : Value *Cond = SI->getCondition();
412 : Value *TV = SI->getTrueValue();
413 : Value *FV = SI->getFalseValue();
414 :
415 : // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
416 : // Does "cmp TV, RHS" simplify?
417 14071 : Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
418 14071 : if (TCmp == Cond) {
419 : // It not only simplified, it simplified to the select condition. Replace
420 : // it with 'true'.
421 37 : TCmp = getTrue(Cond->getType());
422 14034 : } else if (!TCmp) {
423 : // It didn't simplify. However if "cmp TV, RHS" is equal to the select
424 : // condition then we can replace it with 'true'. Otherwise give up.
425 10993 : if (!isSameCompare(Cond, Pred, TV, RHS))
426 : return nullptr;
427 8 : TCmp = getTrue(Cond->getType());
428 : }
429 :
430 : // Does "cmp FV, RHS" simplify?
431 3086 : Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
432 3086 : if (FCmp == Cond) {
433 : // It not only simplified, it simplified to the select condition. Replace
434 : // it with 'false'.
435 1 : FCmp = getFalse(Cond->getType());
436 3085 : } else if (!FCmp) {
437 : // It didn't simplify. However if "cmp FV, RHS" is equal to the select
438 : // condition then we can replace it with 'false'. Otherwise give up.
439 2819 : if (!isSameCompare(Cond, Pred, FV, RHS))
440 : return nullptr;
441 15 : FCmp = getFalse(Cond->getType());
442 : }
443 :
444 : // If both sides simplified to the same value, then use it as the result of
445 : // the original comparison.
446 282 : if (TCmp == FCmp)
447 : return TCmp;
448 :
449 : // The remaining cases only make sense if the select condition has the same
450 : // type as the result of the comparison, so bail out if this is not so.
451 588 : if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
452 : return nullptr;
453 : // If the false value simplified to false, then the result of the compare
454 : // is equal to "Cond && TCmp". This also catches the case when the false
455 : // value simplified to false and the true value to true, returning "Cond".
456 195 : if (match(FCmp, m_Zero()))
457 59 : if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
458 : return V;
459 : // If the true value simplified to true, then the result of the compare
460 : // is equal to "Cond || FCmp".
461 136 : if (match(TCmp, m_One()))
462 4 : if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
463 : return V;
464 : // Finally, if the false value simplified to true and the true value to
465 : // false, then the result of the compare is equal to "!Cond".
466 262 : if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
467 126 : if (Value *V =
468 126 : SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
469 : Q, MaxRecurse))
470 7 : return V;
471 :
472 : return nullptr;
473 : }
474 :
475 : /// In the case of a binary operation with an operand that is a PHI instruction,
476 : /// try to simplify the binop by seeing whether evaluating it on the incoming
477 : /// phi values yields the same result for every value. If so returns the common
478 : /// value, otherwise returns null.
479 73856 : static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
480 : Value *RHS, const SimplifyQuery &Q,
481 : unsigned MaxRecurse) {
482 : // Recursion is always used, so bail out at once if we already hit the limit.
483 73856 : if (!MaxRecurse--)
484 : return nullptr;
485 :
486 : PHINode *PI;
487 : if (isa<PHINode>(LHS)) {
488 : PI = cast<PHINode>(LHS);
489 : // Bail out if RHS and the phi may be mutually interdependent due to a loop.
490 50162 : if (!valueDominatesPHI(RHS, PI, Q.DT))
491 : return nullptr;
492 : } else {
493 : assert(isa<PHINode>(RHS) && "No PHI instruction operand!");
494 : PI = cast<PHINode>(RHS);
495 : // Bail out if LHS and the phi may be mutually interdependent due to a loop.
496 15154 : if (!valueDominatesPHI(LHS, PI, Q.DT))
497 : return nullptr;
498 : }
499 :
500 : // Evaluate the BinOp on the incoming phi values.
501 : Value *CommonValue = nullptr;
502 56531 : for (Value *Incoming : PI->incoming_values()) {
503 : // If the incoming value is the phi node itself, it can safely be skipped.
504 56524 : if (Incoming == PI) continue;
505 56524 : Value *V = PI == LHS ?
506 54102 : SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
507 2422 : SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
508 : // If the operation failed to simplify, or simplified to a different value
509 : // to previously, then give up.
510 56524 : if (!V || (CommonValue && V != CommonValue))
511 : return nullptr;
512 : CommonValue = V;
513 : }
514 :
515 : return CommonValue;
516 : }
517 :
518 : /// In the case of a comparison with a PHI instruction, try to simplify the
519 : /// comparison by seeing whether comparing with all of the incoming phi values
520 : /// yields the same result every time. If so returns the common result,
521 : /// otherwise returns null.
522 124377 : static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
523 : const SimplifyQuery &Q, unsigned MaxRecurse) {
524 : // Recursion is always used, so bail out at once if we already hit the limit.
525 124377 : if (!MaxRecurse--)
526 : return nullptr;
527 :
528 : // Make sure the phi is on the LHS.
529 : if (!isa<PHINode>(LHS)) {
530 : std::swap(LHS, RHS);
531 21771 : Pred = CmpInst::getSwappedPredicate(Pred);
532 : }
533 : assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
534 : PHINode *PI = cast<PHINode>(LHS);
535 :
536 : // Bail out if RHS and the phi may be mutually interdependent due to a loop.
537 121315 : if (!valueDominatesPHI(RHS, PI, Q.DT))
538 : return nullptr;
539 :
540 : // Evaluate the BinOp on the incoming phi values.
541 : Value *CommonValue = nullptr;
542 111594 : for (Value *Incoming : PI->incoming_values()) {
543 : // If the incoming value is the phi node itself, it can safely be skipped.
544 111238 : if (Incoming == PI) continue;
545 111229 : Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
546 : // If the operation failed to simplify, or simplified to a different value
547 : // to previously, then give up.
548 111229 : if (!V || (CommonValue && V != CommonValue))
549 : return nullptr;
550 : CommonValue = V;
551 : }
552 :
553 : return CommonValue;
554 : }
555 :
556 0 : static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
557 : Value *&Op0, Value *&Op1,
558 : const SimplifyQuery &Q) {
559 0 : if (auto *CLHS = dyn_cast<Constant>(Op0)) {
560 0 : if (auto *CRHS = dyn_cast<Constant>(Op1))
561 0 : return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
562 :
563 : // Canonicalize the constant to the RHS if this is a commutative operation.
564 : if (Instruction::isCommutative(Opcode))
565 : std::swap(Op0, Op1);
566 : }
567 : return nullptr;
568 : }
569 :
570 : /// Given operands for an Add, see if we can fold the result.
571 : /// If not, this returns null.
572 11344480 : static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
573 : const SimplifyQuery &Q, unsigned MaxRecurse) {
574 11344480 : if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
575 : return C;
576 :
577 : // X + undef -> undef
578 22437082 : if (match(Op1, m_Undef()))
579 : return Op1;
580 :
581 : // X + 0 -> X
582 11218537 : if (match(Op1, m_Zero()))
583 6840 : return Op0;
584 :
585 : // If two operands are negative, return 0.
586 11211697 : if (isKnownNegation(Op0, Op1))
587 7 : return Constant::getNullValue(Op0->getType());
588 :
589 : // X + (Y - X) -> Y
590 : // (Y - X) + X -> Y
591 : // Eg: X + -X -> 0
592 11211690 : Value *Y = nullptr;
593 22423379 : if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
594 11211689 : match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
595 278 : return Y;
596 :
597 : // X + ~X -> -1 since ~X = -X-1
598 11211412 : Type *Ty = Op0->getType();
599 22422822 : if (match(Op0, m_Not(m_Specific(Op1))) ||
600 11211410 : match(Op1, m_Not(m_Specific(Op0))))
601 2 : return Constant::getAllOnesValue(Ty);
602 :
603 : // add nsw/nuw (xor Y, signmask), signmask --> Y
604 : // The no-wrapping add guarantees that the top bit will be set by the add.
605 : // Therefore, the xor must be clearing the already set sign bit of Y.
606 11211419 : if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
607 9 : match(Op0, m_Xor(m_Value(Y), m_SignMask())))
608 5 : return Y;
609 :
610 : // add nuw %x, -1 -> -1, because %x can only be 0.
611 11211405 : if (IsNUW && match(Op1, m_AllOnes()))
612 29 : return Op1; // Which is -1.
613 :
614 : /// i1 add -> xor.
615 22341258 : if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
616 107 : if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
617 : return V;
618 :
619 : // Try some generic simplifications for associative operations.
620 11211371 : if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
621 : MaxRecurse))
622 163 : return V;
623 :
624 : // Threading Add over selects and phi nodes is pointless, so don't bother.
625 : // Threading over the select in "A + select(cond, B, C)" means evaluating
626 : // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
627 : // only if B and C are equal. If B and C are equal then (since we assume
628 : // that operands have already been simplified) "select(cond, B, C)" should
629 : // have been simplified to the common value of B and C already. Analysing
630 : // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
631 : // for threading over phi nodes.
632 :
633 : return nullptr;
634 : }
635 :
636 7626350 : Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
637 : const SimplifyQuery &Query) {
638 7626350 : return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
639 : }
640 :
641 : /// Compute the base pointer and cumulative constant offsets for V.
642 : ///
643 : /// This strips all constant offsets off of V, leaving it the base pointer, and
644 : /// accumulates the total constant offset applied in the returned constant. It
645 : /// returns 0 if V is not a pointer, and returns the constant '0' if there are
646 : /// no constant offsets applied.
647 : ///
648 : /// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
649 : /// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
650 : /// folding.
651 2314670 : static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
652 : bool AllowNonInbounds = false) {
653 : assert(V->getType()->isPtrOrPtrVectorTy());
654 :
655 2314670 : Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
656 2314670 : APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
657 :
658 : // Even though we don't look through PHI nodes, we could be called on an
659 : // instruction in an unreachable block, which may be on a cycle.
660 : SmallPtrSet<Value *, 4> Visited;
661 2314670 : Visited.insert(V);
662 : do {
663 2378485 : if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
664 170860 : if ((!AllowNonInbounds && !GEP->isInBounds()) ||
665 82784 : !GEP->accumulateConstantOffset(DL, Offset))
666 : break;
667 62884 : V = GEP->getPointerOperand();
668 1376392 : } else if (Operator::getOpcode(V) == Instruction::BitCast) {
669 1856 : V = cast<Operator>(V)->getOperand(0);
670 : } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
671 : if (GA->isInterposable())
672 : break;
673 0 : V = GA->getAliasee();
674 : } else {
675 2289481 : if (auto CS = CallSite(V))
676 72405 : if (Value *RV = CS.getReturnedArgOperand()) {
677 3 : V = RV;
678 3 : continue;
679 : }
680 2289478 : break;
681 : }
682 : assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
683 63815 : } while (Visited.insert(V).second);
684 :
685 2314670 : Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
686 4629340 : if (V->getType()->isVectorTy())
687 2 : return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
688 2 : OffsetIntPtr);
689 : return OffsetIntPtr;
690 : }
691 :
692 : /// Compute the constant difference between two pointer values.
693 : /// If the difference is not a constant, returns zero.
694 107886 : static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
695 : Value *RHS) {
696 107886 : Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
697 107886 : Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
698 :
699 : // If LHS and RHS are not related via constant offsets to the same base
700 : // value, there is nothing we can do here.
701 107886 : if (LHS != RHS)
702 : return nullptr;
703 :
704 : // Otherwise, the difference of LHS - RHS can be computed as:
705 : // LHS - RHS
706 : // = (LHSOffset + Base) - (RHSOffset + Base)
707 : // = LHSOffset - RHSOffset
708 15 : return ConstantExpr::getSub(LHSOffset, RHSOffset);
709 : }
710 :
711 : /// Given operands for a Sub, see if we can fold the result.
712 : /// If not, this returns null.
713 280552 : static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
714 : const SimplifyQuery &Q, unsigned MaxRecurse) {
715 280552 : if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
716 : return C;
717 :
718 : // X - undef -> undef
719 : // undef - X -> undef
720 556422 : if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
721 2 : return UndefValue::get(Op0->getType());
722 :
723 : // X - 0 -> X
724 278209 : if (match(Op1, m_Zero()))
725 641 : return Op0;
726 :
727 : // X - X -> 0
728 277568 : if (Op0 == Op1)
729 214 : return Constant::getNullValue(Op0->getType());
730 :
731 : // Is this a negation?
732 277354 : if (match(Op0, m_Zero())) {
733 : // 0 - X -> 0 if the sub is NUW.
734 12085 : if (isNUW)
735 13 : return Constant::getNullValue(Op0->getType());
736 :
737 24150 : KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
738 12078 : if (Known.Zero.isMaxSignedValue()) {
739 : // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
740 : // Op1 must be 0 because negating the minimum signed value is undefined.
741 6 : if (isNSW)
742 6 : return Constant::getNullValue(Op0->getType());
743 :
744 : // 0 - X -> X if X is 0 or the minimum signed value.
745 3 : return Op1;
746 : }
747 : }
748 :
749 : // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
750 : // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
751 277341 : Value *X = nullptr, *Y = nullptr, *Z = Op1;
752 277341 : if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
753 : // See if "V === Y - Z" simplifies.
754 9132 : if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
755 : // It does! Now see if "X + V" simplifies.
756 651 : if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
757 : // It does, we successfully reassociated!
758 : ++NumReassoc;
759 : return W;
760 : }
761 : // See if "V === X - Z" simplifies.
762 9113 : if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
763 : // It does! Now see if "Y + V" simplifies.
764 300 : if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
765 : // It does, we successfully reassociated!
766 : ++NumReassoc;
767 : return W;
768 : }
769 : }
770 :
771 : // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
772 : // For example, X - (X + 1) -> -1
773 277294 : X = Op0;
774 277294 : if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
775 : // See if "V === X - Y" simplifies.
776 2171 : if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
777 : // It does! Now see if "V - Z" simplifies.
778 51 : if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
779 : // It does, we successfully reassociated!
780 : ++NumReassoc;
781 : return W;
782 : }
783 : // See if "V === X - Z" simplifies.
784 2127 : if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
785 : // It does! Now see if "V - Y" simplifies.
786 266 : if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
787 : // It does, we successfully reassociated!
788 : ++NumReassoc;
789 : return W;
790 : }
791 : }
792 :
793 : // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
794 : // For example, X - (X - Y) -> Y.
795 277250 : Z = Op0;
796 277250 : if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
797 : // See if "V === Z - X" simplifies.
798 941 : if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
799 : // It does! Now see if "V + Y" simplifies.
800 55 : if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
801 : // It does, we successfully reassociated!
802 : ++NumReassoc;
803 : return W;
804 : }
805 :
806 : // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
807 277611 : if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
808 277607 : match(Op1, m_Trunc(m_Value(Y))))
809 4 : if (X->getType() == Y->getType())
810 : // See if "V === X - Y" simplifies.
811 4 : if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
812 : // It does! Now see if "trunc V" simplifies.
813 1 : if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
814 : Q, MaxRecurse - 1))
815 : // It does, return the simplified "trunc V".
816 : return W;
817 :
818 : // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
819 391834 : if (match(Op0, m_PtrToInt(m_Value(X))) &&
820 283948 : match(Op1, m_PtrToInt(m_Value(Y))))
821 107886 : if (Constant *Result = computePointerDifference(Q.DL, X, Y))
822 15 : return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
823 :
824 : // i1 sub -> xor.
825 553753 : if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
826 6 : if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
827 3 : return V;
828 :
829 : // Threading Sub over selects and phi nodes is pointless, so don't bother.
830 : // Threading over the select in "A - select(cond, B, C)" means evaluating
831 : // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
832 : // only if B and C are equal. If B and C are equal then (since we assume
833 : // that operands have already been simplified) "select(cond, B, C)" should
834 : // have been simplified to the common value of B and C already. Analysing
835 : // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
836 : // for threading over phi nodes.
837 :
838 : return nullptr;
839 : }
840 :
841 206209 : Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
842 : const SimplifyQuery &Q) {
843 206209 : return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
844 : }
845 :
846 : /// Given operands for a Mul, see if we can fold the result.
847 : /// If not, this returns null.
848 115655 : static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
849 : unsigned MaxRecurse) {
850 115655 : if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
851 : return C;
852 :
853 : // X * undef -> 0
854 : // X * 0 -> 0
855 110403 : if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
856 875 : return Constant::getNullValue(Op0->getType());
857 :
858 : // X * 1 -> X
859 109528 : if (match(Op1, m_One()))
860 1792 : return Op0;
861 :
862 : // (X / Y) * Y -> X if the division is exact.
863 107736 : Value *X = nullptr;
864 215441 : if (Q.IIQ.UseInstrInfo &&
865 107705 : (match(Op0,
866 214997 : m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
867 107292 : match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
868 415 : return X;
869 :
870 : // i1 mul -> and.
871 191869 : if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
872 4 : if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
873 : return V;
874 :
875 : // Try some generic simplifications for associative operations.
876 107319 : if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
877 : MaxRecurse))
878 : return V;
879 :
880 : // Mul distributes over Add. Try some generic simplifications based on this.
881 107317 : if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
882 : Q, MaxRecurse))
883 : return V;
884 :
885 : // If the operation is with the result of a select instruction, check whether
886 : // operating on either branch of the select always yields the same value.
887 201998 : if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
888 1371 : if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
889 : MaxRecurse))
890 : return V;
891 :
892 : // If the operation is with the result of a phi instruction, check whether
893 : // operating on all incoming values of the phi always yields the same value.
894 190685 : if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
895 18749 : if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
896 : MaxRecurse))
897 0 : return V;
898 :
899 : return nullptr;
900 : }
901 :
902 48940 : Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
903 48940 : return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
904 : }
905 :
906 : /// Check for common or similar folds of integer division or integer remainder.
907 : /// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
908 123466 : static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
909 123466 : Type *Ty = Op0->getType();
910 :
911 : // X / undef -> undef
912 : // X % undef -> undef
913 123466 : if (match(Op1, m_Undef()))
914 : return Op1;
915 :
916 : // X / 0 -> undef
917 : // X % 0 -> undef
918 : // We don't need to preserve faults!
919 123451 : if (match(Op1, m_Zero()))
920 66 : return UndefValue::get(Ty);
921 :
922 : // If any element of a constant divisor vector is zero or undef, the whole op
923 : // is undef.
924 : auto *Op1C = dyn_cast<Constant>(Op1);
925 96660 : if (Op1C && Ty->isVectorTy()) {
926 : unsigned NumElts = Ty->getVectorNumElements();
927 1582 : for (unsigned i = 0; i != NumElts; ++i) {
928 1275 : Constant *Elt = Op1C->getAggregateElement(i);
929 1275 : if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
930 20 : return UndefValue::get(Ty);
931 : }
932 : }
933 :
934 : // undef / X -> 0
935 : // undef % X -> 0
936 123365 : if (match(Op0, m_Undef()))
937 0 : return Constant::getNullValue(Ty);
938 :
939 : // 0 / X -> 0
940 : // 0 % X -> 0
941 123365 : if (match(Op0, m_Zero()))
942 19 : return Constant::getNullValue(Op0->getType());
943 :
944 : // X / X -> 1
945 : // X % X -> 0
946 123346 : if (Op0 == Op1)
947 22 : return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
948 :
949 : // X / 1 -> X
950 : // X % 1 -> 0
951 : // If this is a boolean op (single-bit element type), we can't have
952 : // division-by-zero or remainder-by-zero, so assume the divisor is 1.
953 : // Similarly, if we're zero-extending a boolean divisor, then assume it's a 1.
954 : Value *X;
955 369644 : if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
956 246464 : (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
957 173 : return IsDiv ? Op0 : Constant::getNullValue(Ty);
958 :
959 : return nullptr;
960 : }
961 :
962 : /// Given a predicate and two operands, return true if the comparison is true.
963 : /// This is a helper for div/rem simplification where we return some other value
964 : /// when we can prove a relationship between the operands.
965 122126 : static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
966 : const SimplifyQuery &Q, unsigned MaxRecurse) {
967 122126 : Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
968 : Constant *C = dyn_cast_or_null<Constant>(V);
969 985 : return (C && C->isAllOnesValue());
970 : }
971 :
972 : /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
973 : /// to simplify X % Y to X.
974 123132 : static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
975 : unsigned MaxRecurse, bool IsSigned) {
976 : // Recursion is always used, so bail out at once if we already hit the limit.
977 123132 : if (!MaxRecurse--)
978 : return false;
979 :
980 123130 : if (IsSigned) {
981 : // |X| / |Y| --> 0
982 : //
983 : // We require that 1 operand is a simple constant. That could be extended to
984 : // 2 variables if we computed the sign bit for each.
985 : //
986 : // Make sure that a constant is not the minimum signed value because taking
987 : // the abs() of that is undefined.
988 79066 : Type *Ty = X->getType();
989 : const APInt *C;
990 79066 : if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
991 : // Is the variable divisor magnitude always greater than the constant
992 : // dividend magnitude?
993 : // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
994 61 : Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
995 183 : Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
996 122 : if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
997 61 : isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
998 0 : return true;
999 : }
1000 79066 : if (match(Y, m_APInt(C))) {
1001 : // Special-case: we can't take the abs() of a minimum signed value. If
1002 : // that's the divisor, then all we have to do is prove that the dividend
1003 : // is also not the minimum signed value.
1004 77703 : if (C->isMinSignedValue())
1005 9 : return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
1006 :
1007 : // Is the variable dividend magnitude always less than the constant
1008 : // divisor magnitude?
1009 : // |X| < |C| --> X > -abs(C) and X < abs(C)
1010 77694 : Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1011 233082 : Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1012 77931 : if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1013 237 : isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1014 : return true;
1015 : }
1016 79038 : return false;
1017 : }
1018 :
1019 : // IsSigned == false.
1020 : // Is the dividend unsigned less than the divisor?
1021 44064 : return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1022 : }
1023 :
1024 : /// These are simplifications common to SDiv and UDiv.
1025 94402 : static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1026 : const SimplifyQuery &Q, unsigned MaxRecurse) {
1027 94402 : if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1028 : return C;
1029 :
1030 93479 : if (Value *V = simplifyDivRem(Op0, Op1, true))
1031 : return V;
1032 :
1033 93240 : bool IsSigned = Opcode == Instruction::SDiv;
1034 :
1035 : // (X * Y) / Y -> X if the multiplication does not overflow.
1036 : Value *X;
1037 93240 : if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1038 81 : auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1039 : // If the Mul does not overflow, then we are good to go.
1040 81 : if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1041 57 : (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)))
1042 4 : return X;
1043 : // If X has the form X = A / Y, then X * Y cannot overflow.
1044 77 : if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1045 130 : (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1046 4 : return X;
1047 : }
1048 :
1049 : // (X rem Y) / Y -> 0
1050 93232 : if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1051 34636 : (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1052 2 : return Constant::getNullValue(Op0->getType());
1053 :
1054 : // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1055 : ConstantInt *C1, *C2;
1056 93236 : if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1057 32 : match(Op1, m_ConstantInt(C2))) {
1058 : bool Overflow;
1059 52 : (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1060 26 : if (Overflow)
1061 1 : return Constant::getNullValue(Op0->getType());
1062 : }
1063 :
1064 : // If the operation is with the result of a select instruction, check whether
1065 : // operating on either branch of the select always yields the same value.
1066 182352 : if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1067 166 : if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1068 : return V;
1069 :
1070 : // If the operation is with the result of a phi instruction, check whether
1071 : // operating on all incoming values of the phi always yields the same value.
1072 180548 : if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1073 3012 : if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1074 : return V;
1075 :
1076 93229 : if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1077 16 : return Constant::getNullValue(Op0->getType());
1078 :
1079 : return nullptr;
1080 : }
1081 :
1082 : /// These are simplifications common to SRem and URem.
1083 30212 : static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1084 : const SimplifyQuery &Q, unsigned MaxRecurse) {
1085 30212 : if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1086 : return C;
1087 :
1088 29987 : if (Value *V = simplifyDivRem(Op0, Op1, false))
1089 : return V;
1090 :
1091 : // (X % Y) % Y -> X % Y
1092 3157 : if ((Opcode == Instruction::SRem &&
1093 29911 : match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1094 26754 : (Opcode == Instruction::URem &&
1095 26754 : match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1096 2 : return Op0;
1097 :
1098 : // (X << Y) % X -> 0
1099 29909 : if (Q.IIQ.UseInstrInfo &&
1100 3156 : ((Opcode == Instruction::SRem &&
1101 29909 : match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1102 26753 : (Opcode == Instruction::URem &&
1103 26753 : match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1104 4 : return Constant::getNullValue(Op0->getType());
1105 :
1106 : // If the operation is with the result of a select instruction, check whether
1107 : // operating on either branch of the select always yields the same value.
1108 56935 : if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1109 114 : if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1110 : return V;
1111 :
1112 : // If the operation is with the result of a phi instruction, check whether
1113 : // operating on all incoming values of the phi always yields the same value.
1114 55725 : if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1115 1769 : if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1116 : return V;
1117 :
1118 : // If X / Y == 0, then X % Y == X.
1119 29903 : if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1120 14 : return Op0;
1121 :
1122 : return nullptr;
1123 : }
1124 :
1125 : /// Given operands for an SDiv, see if we can fold the result.
1126 : /// If not, this returns null.
1127 76231 : static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1128 : unsigned MaxRecurse) {
1129 : // If two operands are negated and no signed overflow, return -1.
1130 76231 : if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1131 7 : return Constant::getAllOnesValue(Op0->getType());
1132 :
1133 76224 : return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1134 : }
1135 :
1136 71977 : Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1137 71977 : return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1138 : }
1139 :
1140 : /// Given operands for a UDiv, see if we can fold the result.
1141 : /// If not, this returns null.
1142 4527 : static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1143 : unsigned MaxRecurse) {
1144 4527 : return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1145 : }
1146 :
1147 13651 : Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1148 13651 : return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1149 : }
1150 :
1151 : /// Given operands for an SRem, see if we can fold the result.
1152 : /// If not, this returns null.
1153 3212 : static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1154 : unsigned MaxRecurse) {
1155 : // If the divisor is 0, the result is undefined, so assume the divisor is -1.
1156 : // srem Op0, (sext i1 X) --> srem Op0, -1 --> 0
1157 : Value *X;
1158 3219 : if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1159 2 : return ConstantInt::getNullValue(Op0->getType());
1160 :
1161 : // If the two operands are negated, return 0.
1162 3210 : if (isKnownNegation(Op0, Op1))
1163 6 : return ConstantInt::getNullValue(Op0->getType());
1164 :
1165 3204 : return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1166 : }
1167 :
1168 3077 : Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1169 3077 : return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1170 : }
1171 :
1172 : /// Given operands for a URem, see if we can fold the result.
1173 : /// If not, this returns null.
1174 3708 : static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1175 : unsigned MaxRecurse) {
1176 3708 : return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1177 : }
1178 :
1179 23300 : Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1180 23300 : return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1181 : }
1182 :
1183 : /// Returns true if a shift by \c Amount always yields undef.
1184 232799 : static bool isUndefShift(Value *Amount) {
1185 : Constant *C = dyn_cast<Constant>(Amount);
1186 : if (!C)
1187 : return false;
1188 :
1189 : // X shift by undef -> undef because it may shift by the bitwidth.
1190 204046 : if (isa<UndefValue>(C))
1191 : return true;
1192 :
1193 : // Shifting by the bitwidth or more is undefined.
1194 : if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1195 201632 : if (CI->getValue().getLimitedValue() >=
1196 201632 : CI->getType()->getScalarSizeInBits())
1197 : return true;
1198 :
1199 : // If all lanes of a vector shift are undefined the whole shift is.
1200 203968 : if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1201 2327 : for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1202 2323 : if (!isUndefShift(C->getAggregateElement(I)))
1203 : return false;
1204 : return true;
1205 : }
1206 :
1207 : return false;
1208 : }
1209 :
1210 : /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1211 : /// If not, this returns null.
1212 239702 : static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1213 : Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1214 239702 : if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1215 : return C;
1216 :
1217 : // 0 shift by X -> 0
1218 231180 : if (match(Op0, m_Zero()))
1219 569 : return Constant::getNullValue(Op0->getType());
1220 :
1221 : // X shift by 0 -> X
1222 : // Shift-by-sign-extended bool must be shift-by-0 because shift-by-all-ones
1223 : // would be poison.
1224 : Value *X;
1225 461093 : if (match(Op1, m_Zero()) ||
1226 230516 : (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1227 135 : return Op0;
1228 :
1229 : // Fold undefined shifts.
1230 230476 : if (isUndefShift(Op1))
1231 20 : return UndefValue::get(Op0->getType());
1232 :
1233 : // If the operation is with the result of a select instruction, check whether
1234 : // operating on either branch of the select always yields the same value.
1235 420118 : if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1236 3407 : if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1237 : return V;
1238 :
1239 : // If the operation is with the result of a phi instruction, check whether
1240 : // operating on all incoming values of the phi always yields the same value.
1241 408492 : if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1242 15610 : if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1243 : return V;
1244 :
1245 : // If any bits in the shift amount make that value greater than or equal to
1246 : // the number of bits in the type, the shift is undefined.
1247 460910 : KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1248 230455 : if (Known.One.getLimitedValue() >= Known.getBitWidth())
1249 4 : return UndefValue::get(Op0->getType());
1250 :
1251 : // If all valid bits in the shift amount are known zero, the first operand is
1252 : // unchanged.
1253 : unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1254 230451 : if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1255 6 : return Op0;
1256 :
1257 : return nullptr;
1258 : }
1259 :
1260 : /// Given operands for an Shl, LShr or AShr, see if we can
1261 : /// fold the result. If not, this returns null.
1262 134861 : static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1263 : Value *Op1, bool isExact, const SimplifyQuery &Q,
1264 : unsigned MaxRecurse) {
1265 134861 : if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1266 : return V;
1267 :
1268 : // X >> X -> 0
1269 132687 : if (Op0 == Op1)
1270 3 : return Constant::getNullValue(Op0->getType());
1271 :
1272 : // undef >> X -> 0
1273 : // undef >> X -> undef (if it's exact)
1274 132684 : if (match(Op0, m_Undef()))
1275 3 : return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1276 :
1277 : // The low bit cannot be shifted out of an exact shift if it is set.
1278 132681 : if (isExact) {
1279 59086 : KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1280 29545 : if (Op0Known.One[0])
1281 4 : return Op0;
1282 : }
1283 :
1284 : return nullptr;
1285 : }
1286 :
1287 : /// Given operands for an Shl, see if we can fold the result.
1288 : /// If not, this returns null.
1289 104841 : static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1290 : const SimplifyQuery &Q, unsigned MaxRecurse) {
1291 104841 : if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1292 : return V;
1293 :
1294 : // undef << X -> 0
1295 : // undef << X -> undef if (if it's NSW/NUW)
1296 97758 : if (match(Op0, m_Undef()))
1297 4 : return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1298 :
1299 : // (X >> A) << A -> X
1300 : Value *X;
1301 195487 : if (Q.IIQ.UseInstrInfo &&
1302 97733 : match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1303 57 : return X;
1304 :
1305 : // shl nuw i8 C, %x -> C iff C has sign bit set.
1306 105069 : if (isNUW && match(Op0, m_Negative()))
1307 6 : return Op0;
1308 : // NOTE: could use computeKnownBits() / LazyValueInfo,
1309 : // but the cost-benefit analysis suggests it isn't worth it.
1310 :
1311 : return nullptr;
1312 : }
1313 :
1314 66882 : Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1315 : const SimplifyQuery &Q) {
1316 66882 : return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1317 : }
1318 :
1319 : /// Given operands for an LShr, see if we can fold the result.
1320 : /// If not, this returns null.
1321 66618 : static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1322 : const SimplifyQuery &Q, unsigned MaxRecurse) {
1323 66618 : if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1324 : MaxRecurse))
1325 : return V;
1326 :
1327 : // (X << A) >> A -> X
1328 : Value *X;
1329 64653 : if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1330 12 : return X;
1331 :
1332 : // ((X << A) | Y) >> A -> X if effective width of Y is not larger than A.
1333 : // We can return X as we do in the above case since OR alters no bits in X.
1334 : // SimplifyDemandedBits in InstCombine can do more general optimization for
1335 : // bit manipulation. This pattern aims to provide opportunities for other
1336 : // optimizers by supporting a simple but common case in InstSimplify.
1337 : Value *Y;
1338 : const APInt *ShRAmt, *ShLAmt;
1339 123400 : if (match(Op1, m_APInt(ShRAmt)) &&
1340 123400 : match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1341 147 : *ShRAmt == *ShLAmt) {
1342 66 : const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1343 55 : const unsigned Width = Op0->getType()->getScalarSizeInBits();
1344 55 : const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1345 110 : if (ShRAmt->uge(EffWidthY))
1346 44 : return X;
1347 : }
1348 :
1349 : return nullptr;
1350 : }
1351 :
1352 50098 : Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1353 : const SimplifyQuery &Q) {
1354 50098 : return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1355 : }
1356 :
1357 : /// Given operands for an AShr, see if we can fold the result.
1358 : /// If not, this returns null.
1359 68243 : static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1360 : const SimplifyQuery &Q, unsigned MaxRecurse) {
1361 68243 : if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1362 : MaxRecurse))
1363 : return V;
1364 :
1365 : // all ones >>a X -> -1
1366 : // Do not return Op0 because it may contain undef elements if it's a vector.
1367 68024 : if (match(Op0, m_AllOnes()))
1368 8 : return Constant::getAllOnesValue(Op0->getType());
1369 :
1370 : // (X << A) >> A -> X
1371 : Value *X;
1372 68016 : if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1373 3 : return X;
1374 :
1375 : // Arithmetic shifting an all-sign-bit value is a no-op.
1376 68013 : unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1377 68013 : if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1378 7 : return Op0;
1379 :
1380 : return nullptr;
1381 : }
1382 :
1383 54666 : Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1384 : const SimplifyQuery &Q) {
1385 54666 : return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1386 : }
1387 :
1388 : /// Commuted variants are assumed to be handled by calling this function again
1389 : /// with the parameters swapped.
1390 35550 : static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1391 : ICmpInst *UnsignedICmp, bool IsAnd) {
1392 : Value *X, *Y;
1393 :
1394 : ICmpInst::Predicate EqPred;
1395 35550 : if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1396 12878 : !ICmpInst::isEquality(EqPred))
1397 23222 : return nullptr;
1398 :
1399 : ICmpInst::Predicate UnsignedPred;
1400 12395 : if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1401 67 : ICmpInst::isUnsigned(UnsignedPred))
1402 : ;
1403 : else if (match(UnsignedICmp,
1404 13230 : m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1405 911 : ICmpInst::isUnsigned(UnsignedPred))
1406 137 : UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1407 : else
1408 12182 : return nullptr;
1409 :
1410 : // X < Y && Y != 0 --> X < Y
1411 : // X < Y || Y != 0 --> Y != 0
1412 146 : if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1413 11 : return IsAnd ? UnsignedICmp : ZeroICmp;
1414 :
1415 : // X >= Y || Y != 0 --> true
1416 : // X >= Y || Y == 0 --> X >= Y
1417 137 : if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1418 5 : if (EqPred == ICmpInst::ICMP_NE)
1419 3 : return getTrue(UnsignedICmp->getType());
1420 : return UnsignedICmp;
1421 : }
1422 :
1423 : // X < Y && Y == 0 --> false
1424 132 : if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1425 : IsAnd)
1426 4 : return getFalse(UnsignedICmp->getType());
1427 :
1428 : return nullptr;
1429 : }
1430 :
1431 : /// Commuted variants are assumed to be handled by calling this function again
1432 : /// with the parameters swapped.
1433 16154 : static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1434 : ICmpInst::Predicate Pred0, Pred1;
1435 : Value *A ,*B;
1436 : if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1437 : !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1438 15932 : return nullptr;
1439 :
1440 : // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1441 : // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1442 : // can eliminate Op1 from this 'and'.
1443 222 : if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1444 : return Op0;
1445 :
1446 : // Check for any combination of predicates that are guaranteed to be disjoint.
1447 349 : if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1448 163 : (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1449 340 : (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1450 13 : (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1451 39 : return getFalse(Op0->getType());
1452 :
1453 : return nullptr;
1454 : }
1455 :
1456 : /// Commuted variants are assumed to be handled by calling this function again
1457 : /// with the parameters swapped.
1458 19193 : static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1459 : ICmpInst::Predicate Pred0, Pred1;
1460 : Value *A ,*B;
1461 : if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1462 : !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1463 18876 : return nullptr;
1464 :
1465 : // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1466 : // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1467 : // can eliminate Op0 from this 'or'.
1468 317 : if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1469 : return Op1;
1470 :
1471 : // Check for any combination of predicates that cover the entire range of
1472 : // possibilities.
1473 409 : if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1474 192 : (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1475 401 : (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1476 13 : (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1477 37 : return getTrue(Op0->getType());
1478 :
1479 : return nullptr;
1480 : }
1481 :
1482 : /// Test if a pair of compares with a shared operand and 2 constants has an
1483 : /// empty set intersection, full set union, or if one compare is a superset of
1484 : /// the other.
1485 17549 : static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1486 : bool IsAnd) {
1487 : // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1488 17549 : if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1489 : return nullptr;
1490 :
1491 : const APInt *C0, *C1;
1492 7764 : if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1493 1064 : !match(Cmp1->getOperand(1), m_APInt(C1)))
1494 2405 : return nullptr;
1495 :
1496 2835 : auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1497 2835 : auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1498 :
1499 : // For and-of-compares, check if the intersection is empty:
1500 : // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1501 945 : if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1502 53 : return getFalse(Cmp0->getType());
1503 :
1504 : // For or-of-compares, check if the union is full:
1505 : // (icmp X, C0) || (icmp X, C1) --> full set --> true
1506 892 : if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1507 43 : return getTrue(Cmp0->getType());
1508 :
1509 : // Is one range a superset of the other?
1510 : // If this is and-of-compares, take the smaller set:
1511 : // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1512 : // If this is or-of-compares, take the larger set:
1513 : // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1514 849 : if (Range0.contains(Range1))
1515 150 : return IsAnd ? Cmp1 : Cmp0;
1516 744 : if (Range1.contains(Range0))
1517 172 : return IsAnd ? Cmp0 : Cmp1;
1518 :
1519 : return nullptr;
1520 : }
1521 :
1522 17232 : static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1523 : bool IsAnd) {
1524 : ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1525 6839 : if (!match(Cmp0->getOperand(1), m_Zero()) ||
1526 21035 : !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1527 15812 : return nullptr;
1528 :
1529 1420 : if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1530 : return nullptr;
1531 :
1532 : // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1533 : Value *X = Cmp0->getOperand(0);
1534 : Value *Y = Cmp1->getOperand(0);
1535 :
1536 : // If one of the compares is a masked version of a (not) null check, then
1537 : // that compare implies the other, so we eliminate the other. Optionally, look
1538 : // through a pointer-to-int cast to match a null check of a pointer type.
1539 :
1540 : // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1541 : // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1542 : // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1543 : // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1544 2534 : if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1545 1262 : match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1546 9 : return Cmp1;
1547 :
1548 : // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1549 : // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1550 : // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1551 : // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1552 2516 : if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1553 1253 : match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1554 9 : return Cmp0;
1555 :
1556 : return nullptr;
1557 : }
1558 :
1559 15686 : static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1,
1560 : const InstrInfoQuery &IIQ) {
1561 : // (icmp (add V, C0), C1) & (icmp V, C0)
1562 : ICmpInst::Predicate Pred0, Pred1;
1563 : const APInt *C0, *C1;
1564 : Value *V;
1565 15686 : if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1566 : return nullptr;
1567 :
1568 595 : if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1569 565 : return nullptr;
1570 :
1571 : auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1572 60 : if (AddInst->getOperand(1) != Op1->getOperand(1))
1573 : return nullptr;
1574 :
1575 17 : Type *ITy = Op0->getType();
1576 17 : bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1577 : bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1578 :
1579 17 : const APInt Delta = *C1 - *C0;
1580 17 : if (C0->isStrictlyPositive()) {
1581 17 : if (Delta == 2) {
1582 6 : if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1583 2 : return getFalse(ITy);
1584 4 : if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1585 2 : return getFalse(ITy);
1586 : }
1587 13 : if (Delta == 1) {
1588 6 : if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1589 2 : return getFalse(ITy);
1590 4 : if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1591 2 : return getFalse(ITy);
1592 : }
1593 : }
1594 9 : if (C0->getBoolValue() && isNUW) {
1595 4 : if (Delta == 2)
1596 2 : if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1597 2 : return getFalse(ITy);
1598 2 : if (Delta == 1)
1599 2 : if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1600 2 : return getFalse(ITy);
1601 : }
1602 :
1603 : return nullptr;
1604 : }
1605 :
1606 8116 : static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1607 : const InstrInfoQuery &IIQ) {
1608 8116 : if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1609 : return X;
1610 8109 : if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1611 : return X;
1612 :
1613 8105 : if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1614 : return X;
1615 8049 : if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1616 : return X;
1617 :
1618 8030 : if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1619 : return X;
1620 :
1621 7857 : if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1622 : return X;
1623 :
1624 7849 : if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, IIQ))
1625 : return X;
1626 7837 : if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, IIQ))
1627 0 : return X;
1628 :
1629 : return nullptr;
1630 : }
1631 :
1632 18718 : static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1,
1633 : const InstrInfoQuery &IIQ) {
1634 : // (icmp (add V, C0), C1) | (icmp V, C0)
1635 : ICmpInst::Predicate Pred0, Pred1;
1636 : const APInt *C0, *C1;
1637 : Value *V;
1638 18718 : if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1639 : return nullptr;
1640 :
1641 120 : if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1642 23 : return nullptr;
1643 :
1644 : auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1645 97 : if (AddInst->getOperand(1) != Op1->getOperand(1))
1646 : return nullptr;
1647 :
1648 12 : Type *ITy = Op0->getType();
1649 12 : bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1650 12 : bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1651 :
1652 12 : const APInt Delta = *C1 - *C0;
1653 12 : if (C0->isStrictlyPositive()) {
1654 12 : if (Delta == 2) {
1655 6 : if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1656 2 : return getTrue(ITy);
1657 4 : if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1658 2 : return getTrue(ITy);
1659 : }
1660 8 : if (Delta == 1) {
1661 6 : if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1662 2 : return getTrue(ITy);
1663 4 : if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1664 2 : return getTrue(ITy);
1665 : }
1666 : }
1667 4 : if (C0->getBoolValue() && isNUW) {
1668 4 : if (Delta == 2)
1669 2 : if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1670 2 : return getTrue(ITy);
1671 2 : if (Delta == 1)
1672 2 : if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1673 2 : return getTrue(ITy);
1674 : }
1675 :
1676 : return nullptr;
1677 : }
1678 :
1679 9663 : static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1680 : const InstrInfoQuery &IIQ) {
1681 9663 : if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1682 : return X;
1683 9662 : if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1684 : return X;
1685 :
1686 9656 : if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1687 : return X;
1688 9537 : if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1689 : return X;
1690 :
1691 9519 : if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1692 : return X;
1693 :
1694 9375 : if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1695 : return X;
1696 :
1697 9365 : if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, IIQ))
1698 : return X;
1699 9353 : if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, IIQ))
1700 0 : return X;
1701 :
1702 : return nullptr;
1703 : }
1704 :
1705 1275 : static Value *simplifyAndOrOfFCmps(const TargetLibraryInfo *TLI,
1706 : FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1707 : Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1708 : Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1709 1275 : if (LHS0->getType() != RHS0->getType())
1710 : return nullptr;
1711 :
1712 : FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1713 1272 : if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1714 1257 : (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1715 : // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1716 : // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1717 : // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1718 : // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1719 : // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1720 : // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1721 : // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1722 : // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1723 56 : if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1724 42 : (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1725 8 : return RHS;
1726 :
1727 : // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1728 : // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1729 : // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1730 : // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1731 : // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1732 : // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1733 : // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1734 : // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1735 40 : if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1736 34 : (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1737 8 : return LHS;
1738 : }
1739 :
1740 : return nullptr;
1741 : }
1742 :
1743 465082 : static Value *simplifyAndOrOfCmps(const SimplifyQuery &Q,
1744 : Value *Op0, Value *Op1, bool IsAnd) {
1745 : // Look through casts of the 'and' operands to find compares.
1746 : auto *Cast0 = dyn_cast<CastInst>(Op0);
1747 : auto *Cast1 = dyn_cast<CastInst>(Op1);
1748 467677 : if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1749 : Cast0->getSrcTy() == Cast1->getSrcTy()) {
1750 : Op0 = Cast0->getOperand(0);
1751 : Op1 = Cast1->getOperand(0);
1752 : }
1753 :
1754 : Value *V = nullptr;
1755 : auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1756 : auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1757 465082 : if (ICmp0 && ICmp1)
1758 17779 : V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q.IIQ)
1759 9663 : : simplifyOrOfICmps(ICmp0, ICmp1, Q.IIQ);
1760 :
1761 : auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1762 : auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1763 465082 : if (FCmp0 && FCmp1)
1764 1275 : V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1765 :
1766 465082 : if (!V)
1767 : return nullptr;
1768 605 : if (!Cast0)
1769 : return V;
1770 :
1771 : // If we looked through casts, we can only handle a constant simplification
1772 : // because we are not allowed to create a cast instruction here.
1773 : if (auto *C = dyn_cast<Constant>(V))
1774 16 : return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1775 :
1776 : return nullptr;
1777 : }
1778 :
1779 : /// Given operands for an And, see if we can fold the result.
1780 : /// If not, this returns null.
1781 314124 : static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1782 : unsigned MaxRecurse) {
1783 314124 : if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1784 : return C;
1785 :
1786 : // X & undef -> 0
1787 607672 : if (match(Op1, m_Undef()))
1788 2 : return Constant::getNullValue(Op0->getType());
1789 :
1790 : // X & X = X
1791 303834 : if (Op0 == Op1)
1792 : return Op0;
1793 :
1794 : // X & 0 = 0
1795 303758 : if (match(Op1, m_Zero()))
1796 1326 : return Constant::getNullValue(Op0->getType());
1797 :
1798 : // X & -1 = X
1799 302432 : if (match(Op1, m_AllOnes()))
1800 4776 : return Op0;
1801 :
1802 : // A & ~A = ~A & A = 0
1803 595275 : if (match(Op0, m_Not(m_Specific(Op1))) ||
1804 297619 : match(Op1, m_Not(m_Specific(Op0))))
1805 60 : return Constant::getNullValue(Op0->getType());
1806 :
1807 : // (A | ?) & A = A
1808 595192 : if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1809 : return Op1;
1810 :
1811 : // A & (A | ?) = A
1812 595102 : if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1813 : return Op0;
1814 :
1815 : // A mask that only clears known zeros of a shifted value is a no-op.
1816 : Value *X;
1817 : const APInt *Mask;
1818 : const APInt *ShAmt;
1819 297498 : if (match(Op1, m_APInt(Mask))) {
1820 : // If all bits in the inverted and shifted mask are clear:
1821 : // and (shl X, ShAmt), Mask --> shl X, ShAmt
1822 208472 : if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1823 226920 : (~(*Mask)).lshr(*ShAmt).isNullValue())
1824 232 : return Op0;
1825 :
1826 : // If all bits in the inverted and shifted mask are clear:
1827 : // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1828 210429 : if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1829 237633 : (~(*Mask)).shl(*ShAmt).isNullValue())
1830 59 : return Op0;
1831 : }
1832 :
1833 : // A & (-A) = A if A is a power of two or zero.
1834 594414 : if (match(Op0, m_Neg(m_Specific(Op1))) ||
1835 594404 : match(Op1, m_Neg(m_Specific(Op0)))) {
1836 10 : if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1837 10 : Q.DT))
1838 2 : return Op0;
1839 8 : if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1840 8 : Q.DT))
1841 0 : return Op1;
1842 : }
1843 :
1844 297205 : if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
1845 : return V;
1846 :
1847 : // Try some generic simplifications for associative operations.
1848 296918 : if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1849 : MaxRecurse))
1850 : return V;
1851 :
1852 : // And distributes over Or. Try some generic simplifications based on this.
1853 295683 : if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1854 : Q, MaxRecurse))
1855 : return V;
1856 :
1857 : // And distributes over Xor. Try some generic simplifications based on this.
1858 295018 : if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1859 : Q, MaxRecurse))
1860 : return V;
1861 :
1862 : // If the operation is with the result of a select instruction, check whether
1863 : // operating on either branch of the select always yields the same value.
1864 571840 : if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1865 713 : if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1866 : MaxRecurse))
1867 : return V;
1868 :
1869 : // If the operation is with the result of a phi instruction, check whether
1870 : // operating on all incoming values of the phi always yields the same value.
1871 554137 : if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1872 20591 : if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1873 : MaxRecurse))
1874 : return V;
1875 :
1876 : // Assuming the effective width of Y is not larger than A, i.e. all bits
1877 : // from X and Y are disjoint in (X << A) | Y,
1878 : // if the mask of this AND op covers all bits of X or Y, while it covers
1879 : // no bits from the other, we can bypass this AND op. E.g.,
1880 : // ((X << A) | Y) & Mask -> Y,
1881 : // if Mask = ((1 << effective_width_of(Y)) - 1)
1882 : // ((X << A) | Y) & Mask -> X << A,
1883 : // if Mask = ((1 << effective_width_of(X)) - 1) << A
1884 : // SimplifyDemandedBits in InstCombine can optimize the general case.
1885 : // This pattern aims to help other passes for a common case.
1886 : Value *Y, *XShifted;
1887 496692 : if (match(Op1, m_APInt(Mask)) &&
1888 496545 : match(Op0, m_c_Or(m_CombineAnd(m_NUWShl(m_Value(X), m_APInt(ShAmt)),
1889 : m_Value(XShifted)),
1890 : m_Value(Y)))) {
1891 147 : const unsigned Width = Op0->getType()->getScalarSizeInBits();
1892 147 : const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
1893 276 : const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1894 147 : const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1895 147 : if (EffWidthY <= ShftCnt) {
1896 73 : const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
1897 128 : Q.DT);
1898 73 : const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
1899 73 : const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
1900 73 : const APInt EffBitsX = APInt::getLowBitsSet(Width, EffWidthX) << ShftCnt;
1901 : // If the mask is extracting all bits from X or Y as is, we can skip
1902 : // this AND op.
1903 193 : if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
1904 11 : return Y;
1905 71 : if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
1906 7 : return XShifted;
1907 : }
1908 : }
1909 :
1910 : return nullptr;
1911 : }
1912 :
1913 229278 : Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1914 229278 : return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1915 : }
1916 :
1917 : /// Given operands for an Or, see if we can fold the result.
1918 : /// If not, this returns null.
1919 182141 : static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1920 : unsigned MaxRecurse) {
1921 182141 : if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1922 : return C;
1923 :
1924 : // X | undef -> -1
1925 : // X | -1 = -1
1926 : // Do not return Op1 because it may contain undef elements if it's a vector.
1927 514674 : if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1928 62 : return Constant::getAllOnesValue(Op0->getType());
1929 :
1930 : // X | X = X
1931 : // X | 0 = X
1932 342871 : if (Op0 == Op1 || match(Op1, m_Zero()))
1933 1295 : return Op0;
1934 :
1935 : // A | ~A = ~A | A = -1
1936 340305 : if (match(Op0, m_Not(m_Specific(Op1))) ||
1937 170102 : match(Op1, m_Not(m_Specific(Op0))))
1938 2060 : return Constant::getAllOnesValue(Op0->getType());
1939 :
1940 : // (A & ?) | A = A
1941 336286 : if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1942 : return Op1;
1943 :
1944 : // A | (A & ?) = A
1945 336118 : if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1946 : return Op0;
1947 :
1948 : // ~(A & ?) | A = -1
1949 167895 : if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1950 2 : return Constant::getAllOnesValue(Op1->getType());
1951 :
1952 : // A | ~(A & ?) = -1
1953 167893 : if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1954 0 : return Constant::getAllOnesValue(Op0->getType());
1955 :
1956 : Value *A, *B;
1957 : // (A & ~B) | (A ^ B) -> (A ^ B)
1958 : // (~B & A) | (A ^ B) -> (A ^ B)
1959 : // (A & ~B) | (B ^ A) -> (B ^ A)
1960 : // (~B & A) | (B ^ A) -> (B ^ A)
1961 171462 : if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1962 10705 : (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1963 7134 : match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1964 4 : return Op1;
1965 :
1966 : // Commute the 'or' operands.
1967 : // (A ^ B) | (A & ~B) -> (A ^ B)
1968 : // (A ^ B) | (~B & A) -> (A ^ B)
1969 : // (B ^ A) | (A & ~B) -> (B ^ A)
1970 : // (B ^ A) | (~B & A) -> (B ^ A)
1971 168701 : if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1972 2434 : (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1973 1620 : match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1974 4 : return Op0;
1975 :
1976 : // (A & B) | (~A ^ B) -> (~A ^ B)
1977 : // (B & A) | (~A ^ B) -> (~A ^ B)
1978 : // (A & B) | (B ^ ~A) -> (B ^ ~A)
1979 : // (B & A) | (B ^ ~A) -> (B ^ ~A)
1980 196870 : if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1981 86953 : (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1982 57966 : match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1983 4 : return Op1;
1984 :
1985 : // (~A ^ B) | (A & B) -> (~A ^ B)
1986 : // (~A ^ B) | (B & A) -> (~A ^ B)
1987 : // (B ^ ~A) | (A & B) -> (B ^ ~A)
1988 : // (B ^ ~A) | (B & A) -> (B ^ ~A)
1989 186156 : if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1990 54823 : (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1991 36546 : match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1992 4 : return Op0;
1993 :
1994 167877 : if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
1995 : return V;
1996 :
1997 : // Try some generic simplifications for associative operations.
1998 167559 : if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1999 : MaxRecurse))
2000 : return V;
2001 :
2002 : // Or distributes over And. Try some generic simplifications based on this.
2003 167277 : if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
2004 : MaxRecurse))
2005 : return V;
2006 :
2007 : // If the operation is with the result of a select instruction, check whether
2008 : // operating on either branch of the select always yields the same value.
2009 326898 : if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2010 974 : if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2011 : MaxRecurse))
2012 : return V;
2013 :
2014 : // (A & C1)|(B & C2)
2015 : const APInt *C1, *C2;
2016 190640 : if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2017 187638 : match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2018 9006 : if (*C1 == ~*C2) {
2019 : // (A & C1)|(B & C2)
2020 : // If we have: ((V + N) & C1) | (V & C2)
2021 : // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
2022 : // replace with V+N.
2023 : Value *N;
2024 399 : if (C2->isMask() && // C2 == 0+1+
2025 395 : match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2026 : // Add commutes, try both ways.
2027 4 : if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2028 7 : return A;
2029 : }
2030 : // Or commutes, try both ways.
2031 323 : if (C1->isMask() &&
2032 320 : match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2033 : // Add commutes, try both ways.
2034 3 : if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2035 3 : return B;
2036 : }
2037 : }
2038 : }
2039 :
2040 : // If the operation is with the result of a phi instruction, check whether
2041 : // operating on all incoming values of the phi always yields the same value.
2042 318985 : if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2043 14125 : if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2044 0 : return V;
2045 :
2046 : return nullptr;
2047 : }
2048 :
2049 64925 : Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2050 64925 : return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
2051 : }
2052 :
2053 : /// Given operands for a Xor, see if we can fold the result.
2054 : /// If not, this returns null.
2055 112412 : static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2056 : unsigned MaxRecurse) {
2057 112412 : if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2058 : return C;
2059 :
2060 : // A ^ undef -> undef
2061 218152 : if (match(Op1, m_Undef()))
2062 : return Op1;
2063 :
2064 : // A ^ 0 = A
2065 109076 : if (match(Op1, m_Zero()))
2066 2277 : return Op0;
2067 :
2068 : // A ^ A = 0
2069 106799 : if (Op0 == Op1)
2070 23 : return Constant::getNullValue(Op0->getType());
2071 :
2072 : // A ^ ~A = ~A ^ A = -1
2073 213552 : if (match(Op0, m_Not(m_Specific(Op1))) ||
2074 106776 : match(Op1, m_Not(m_Specific(Op0))))
2075 5 : return Constant::getAllOnesValue(Op0->getType());
2076 :
2077 : // Try some generic simplifications for associative operations.
2078 106771 : if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2079 : MaxRecurse))
2080 2137 : return V;
2081 :
2082 : // Threading Xor over selects and phi nodes is pointless, so don't bother.
2083 : // Threading over the select in "A ^ select(cond, B, C)" means evaluating
2084 : // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
2085 : // only if B and C are equal. If B and C are equal then (since we assume
2086 : // that operands have already been simplified) "select(cond, B, C)" should
2087 : // have been simplified to the common value of B and C already. Analysing
2088 : // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
2089 : // for threading over phi nodes.
2090 :
2091 : return nullptr;
2092 : }
2093 :
2094 77974 : Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2095 77974 : return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2096 : }
2097 :
2098 :
2099 : static Type *GetCompareTy(Value *Op) {
2100 2352205 : return CmpInst::makeCmpResultType(Op->getType());
2101 : }
2102 :
2103 : /// Rummage around inside V looking for something equivalent to the comparison
2104 : /// "LHS Pred RHS". Return such a value if found, otherwise return null.
2105 : /// Helper function for analyzing max/min idioms.
2106 352 : static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
2107 : Value *LHS, Value *RHS) {
2108 : SelectInst *SI = dyn_cast<SelectInst>(V);
2109 : if (!SI)
2110 : return nullptr;
2111 : CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
2112 : if (!Cmp)
2113 : return nullptr;
2114 : Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
2115 181 : if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2116 : return Cmp;
2117 169 : if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2118 169 : LHS == CmpRHS && RHS == CmpLHS)
2119 44 : return Cmp;
2120 : return nullptr;
2121 : }
2122 :
2123 : // A significant optimization not implemented here is assuming that alloca
2124 : // addresses are not equal to incoming argument values. They don't *alias*,
2125 : // as we say, but that doesn't mean they aren't equal, so we take a
2126 : // conservative approach.
2127 : //
2128 : // This is inspired in part by C++11 5.10p1:
2129 : // "Two pointers of the same type compare equal if and only if they are both
2130 : // null, both point to the same function, or both represent the same
2131 : // address."
2132 : //
2133 : // This is pretty permissive.
2134 : //
2135 : // It's also partly due to C11 6.5.9p6:
2136 : // "Two pointers compare equal if and only if both are null pointers, both are
2137 : // pointers to the same object (including a pointer to an object and a
2138 : // subobject at its beginning) or function, both are pointers to one past the
2139 : // last element of the same array object, or one is a pointer to one past the
2140 : // end of one array object and the other is a pointer to the start of a
2141 : // different array object that happens to immediately follow the first array
2142 : // object in the address space.)
2143 : //
2144 : // C11's version is more restrictive, however there's no reason why an argument
2145 : // couldn't be a one-past-the-end value for a stack object in the caller and be
2146 : // equal to the beginning of a stack object in the callee.
2147 : //
2148 : // If the C and C++ standards are ever made sufficiently restrictive in this
2149 : // area, it may be possible to update LLVM's semantics accordingly and reinstate
2150 : // this optimization.
2151 : static Constant *
2152 0 : computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2153 : const DominatorTree *DT, CmpInst::Predicate Pred,
2154 : AssumptionCache *AC, const Instruction *CxtI,
2155 : const InstrInfoQuery &IIQ, Value *LHS, Value *RHS) {
2156 : // First, skip past any trivial no-ops.
2157 0 : LHS = LHS->stripPointerCasts();
2158 0 : RHS = RHS->stripPointerCasts();
2159 :
2160 : // A non-null pointer is not equal to a null pointer.
2161 0 : if (llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2162 0 : IIQ.UseInstrInfo) &&
2163 0 : isa<ConstantPointerNull>(RHS) &&
2164 0 : (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2165 0 : return ConstantInt::get(GetCompareTy(LHS),
2166 0 : !CmpInst::isTrueWhenEqual(Pred));
2167 :
2168 : // We can only fold certain predicates on pointer comparisons.
2169 0 : switch (Pred) {
2170 : default:
2171 : return nullptr;
2172 :
2173 : // Equality comaprisons are easy to fold.
2174 : case CmpInst::ICMP_EQ:
2175 : case CmpInst::ICMP_NE:
2176 : break;
2177 :
2178 : // We can only handle unsigned relational comparisons because 'inbounds' on
2179 : // a GEP only protects against unsigned wrapping.
2180 0 : case CmpInst::ICMP_UGT:
2181 : case CmpInst::ICMP_UGE:
2182 : case CmpInst::ICMP_ULT:
2183 : case CmpInst::ICMP_ULE:
2184 : // However, we have to switch them to their signed variants to handle
2185 : // negative indices from the base pointer.
2186 0 : Pred = ICmpInst::getSignedPredicate(Pred);
2187 0 : break;
2188 : }
2189 :
2190 : // Strip off any constant offsets so that we can reason about them.
2191 : // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2192 : // here and compare base addresses like AliasAnalysis does, however there are
2193 : // numerous hazards. AliasAnalysis and its utilities rely on special rules
2194 : // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2195 : // doesn't need to guarantee pointer inequality when it says NoAlias.
2196 0 : Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2197 0 : Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2198 :
2199 : // If LHS and RHS are related via constant offsets to the same base
2200 : // value, we can replace it with an icmp which just compares the offsets.
2201 0 : if (LHS == RHS)
2202 0 : return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2203 :
2204 : // Various optimizations for (in)equality comparisons.
2205 0 : if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2206 : // Different non-empty allocations that exist at the same time have
2207 : // different addresses (if the program can tell). Global variables always
2208 : // exist, so they always exist during the lifetime of each other and all
2209 : // allocas. Two different allocas usually have different addresses...
2210 : //
2211 : // However, if there's an @llvm.stackrestore dynamically in between two
2212 : // allocas, they may have the same address. It's tempting to reduce the
2213 : // scope of the problem by only looking at *static* allocas here. That would
2214 : // cover the majority of allocas while significantly reducing the likelihood
2215 : // of having an @llvm.stackrestore pop up in the middle. However, it's not
2216 : // actually impossible for an @llvm.stackrestore to pop up in the middle of
2217 : // an entry block. Also, if we have a block that's not attached to a
2218 : // function, we can't tell if it's "static" under the current definition.
2219 : // Theoretically, this problem could be fixed by creating a new kind of
2220 : // instruction kind specifically for static allocas. Such a new instruction
2221 : // could be required to be at the top of the entry block, thus preventing it
2222 : // from being subject to a @llvm.stackrestore. Instcombine could even
2223 : // convert regular allocas into these special allocas. It'd be nifty.
2224 : // However, until then, this problem remains open.
2225 : //
2226 : // So, we'll assume that two non-empty allocas have different addresses
2227 : // for now.
2228 : //
2229 : // With all that, if the offsets are within the bounds of their allocations
2230 : // (and not one-past-the-end! so we can't use inbounds!), and their
2231 : // allocations aren't the same, the pointers are not equal.
2232 : //
2233 : // Note that it's not necessary to check for LHS being a global variable
2234 : // address, due to canonicalization and constant folding.
2235 : if (isa<AllocaInst>(LHS) &&
2236 0 : (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2237 : ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2238 : ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2239 : uint64_t LHSSize, RHSSize;
2240 0 : ObjectSizeOpts Opts;
2241 0 : Opts.NullIsUnknownSize =
2242 0 : NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2243 0 : if (LHSOffsetCI && RHSOffsetCI &&
2244 0 : getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2245 0 : getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2246 : const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2247 : const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2248 0 : if (!LHSOffsetValue.isNegative() &&
2249 0 : !RHSOffsetValue.isNegative() &&
2250 0 : LHSOffsetValue.ult(LHSSize) &&
2251 0 : RHSOffsetValue.ult(RHSSize)) {
2252 0 : return ConstantInt::get(GetCompareTy(LHS),
2253 0 : !CmpInst::isTrueWhenEqual(Pred));
2254 : }
2255 : }
2256 :
2257 : // Repeat the above check but this time without depending on DataLayout
2258 : // or being able to compute a precise size.
2259 0 : if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2260 0 : !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2261 0 : LHSOffset->isNullValue() &&
2262 0 : RHSOffset->isNullValue())
2263 0 : return ConstantInt::get(GetCompareTy(LHS),
2264 0 : !CmpInst::isTrueWhenEqual(Pred));
2265 : }
2266 :
2267 : // Even if an non-inbounds GEP occurs along the path we can still optimize
2268 : // equality comparisons concerning the result. We avoid walking the whole
2269 : // chain again by starting where the last calls to
2270 : // stripAndComputeConstantOffsets left off and accumulate the offsets.
2271 0 : Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2272 0 : Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2273 0 : if (LHS == RHS)
2274 0 : return ConstantExpr::getICmp(Pred,
2275 : ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2276 0 : ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2277 :
2278 : // If one side of the equality comparison must come from a noalias call
2279 : // (meaning a system memory allocation function), and the other side must
2280 : // come from a pointer that cannot overlap with dynamically-allocated
2281 : // memory within the lifetime of the current function (allocas, byval
2282 : // arguments, globals), then determine the comparison result here.
2283 : SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2284 0 : GetUnderlyingObjects(LHS, LHSUObjs, DL);
2285 0 : GetUnderlyingObjects(RHS, RHSUObjs, DL);
2286 :
2287 : // Is the set of underlying objects all noalias calls?
2288 : auto IsNAC = [](ArrayRef<Value *> Objects) {
2289 : return all_of(Objects, isNoAliasCall);
2290 : };
2291 :
2292 : // Is the set of underlying objects all things which must be disjoint from
2293 : // noalias calls. For allocas, we consider only static ones (dynamic
2294 : // allocas might be transformed into calls to malloc not simultaneously
2295 : // live with the compared-to allocation). For globals, we exclude symbols
2296 : // that might be resolve lazily to symbols in another dynamically-loaded
2297 : // library (and, thus, could be malloc'ed by the implementation).
2298 : auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2299 : return all_of(Objects, [](Value *V) {
2300 : if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2301 : return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2302 : if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2303 : return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2304 : GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2305 : !GV->isThreadLocal();
2306 : if (const Argument *A = dyn_cast<Argument>(V))
2307 : return A->hasByValAttr();
2308 : return false;
2309 : });
2310 : };
2311 :
2312 0 : if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2313 0 : (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2314 0 : return ConstantInt::get(GetCompareTy(LHS),
2315 0 : !CmpInst::isTrueWhenEqual(Pred));
2316 :
2317 : // Fold comparisons for non-escaping pointer even if the allocation call
2318 : // cannot be elided. We cannot fold malloc comparison to null. Also, the
2319 : // dynamic allocation call could be either of the operands.
2320 : Value *MI = nullptr;
2321 0 : if (isAllocLikeFn(LHS, TLI) &&
2322 0 : llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2323 0 : MI = LHS;
2324 0 : else if (isAllocLikeFn(RHS, TLI) &&
2325 0 : llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2326 0 : MI = RHS;
2327 : // FIXME: We should also fold the compare when the pointer escapes, but the
2328 : // compare dominates the pointer escape
2329 0 : if (MI && !PointerMayBeCaptured(MI, true, true))
2330 0 : return ConstantInt::get(GetCompareTy(LHS),
2331 0 : CmpInst::isFalseWhenEqual(Pred));
2332 : }
2333 :
2334 : // Otherwise, fail.
2335 : return nullptr;
2336 : }
2337 :
2338 : /// Fold an icmp when its operands have i1 scalar type.
2339 0 : static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2340 : Value *RHS, const SimplifyQuery &Q) {
2341 : Type *ITy = GetCompareTy(LHS); // The return type.
2342 0 : Type *OpTy = LHS->getType(); // The operand type.
2343 0 : if (!OpTy->isIntOrIntVectorTy(1))
2344 0 : return nullptr;
2345 :
2346 : // A boolean compared to true/false can be simplified in 14 out of the 20
2347 : // (10 predicates * 2 constants) possible combinations. Cases not handled here
2348 : // require a 'not' of the LHS, so those must be transformed in InstCombine.
2349 0 : if (match(RHS, m_Zero())) {
2350 0 : switch (Pred) {
2351 0 : case CmpInst::ICMP_NE: // X != 0 -> X
2352 : case CmpInst::ICMP_UGT: // X >u 0 -> X
2353 : case CmpInst::ICMP_SLT: // X <s 0 -> X
2354 0 : return LHS;
2355 :
2356 : case CmpInst::ICMP_ULT: // X <u 0 -> false
2357 : case CmpInst::ICMP_SGT: // X >s 0 -> false
2358 0 : return getFalse(ITy);
2359 :
2360 : case CmpInst::ICMP_UGE: // X >=u 0 -> true
2361 : case CmpInst::ICMP_SLE: // X <=s 0 -> true
2362 0 : return getTrue(ITy);
2363 :
2364 : default: break;
2365 : }
2366 0 : } else if (match(RHS, m_One())) {
2367 0 : switch (Pred) {
2368 0 : case CmpInst::ICMP_EQ: // X == 1 -> X
2369 : case CmpInst::ICMP_UGE: // X >=u 1 -> X
2370 : case CmpInst::ICMP_SLE: // X <=s -1 -> X
2371 0 : return LHS;
2372 :
2373 : case CmpInst::ICMP_UGT: // X >u 1 -> false
2374 : case CmpInst::ICMP_SLT: // X <s -1 -> false
2375 0 : return getFalse(ITy);
2376 :
2377 : case CmpInst::ICMP_ULE: // X <=u 1 -> true
2378 : case CmpInst::ICMP_SGE: // X >=s -1 -> true
2379 0 : return getTrue(ITy);
2380 :
2381 : default: break;
2382 : }
2383 : }
2384 :
2385 0 : switch (Pred) {
2386 : default:
2387 : break;
2388 0 : case ICmpInst::ICMP_UGE:
2389 0 : if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2390 0 : return getTrue(ITy);
2391 : break;
2392 0 : case ICmpInst::ICMP_SGE:
2393 : /// For signed comparison, the values for an i1 are 0 and -1
2394 : /// respectively. This maps into a truth table of:
2395 : /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2396 : /// 0 | 0 | 1 (0 >= 0) | 1
2397 : /// 0 | 1 | 1 (0 >= -1) | 1
2398 : /// 1 | 0 | 0 (-1 >= 0) | 0
2399 : /// 1 | 1 | 1 (-1 >= -1) | 1
2400 0 : if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2401 0 : return getTrue(ITy);
2402 : break;
2403 0 : case ICmpInst::ICMP_ULE:
2404 0 : if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2405 0 : return getTrue(ITy);
2406 : break;
2407 : }
2408 :
2409 : return nullptr;
2410 : }
2411 :
2412 : /// Try hard to fold icmp with zero RHS because this is a common case.
2413 1567158 : static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2414 : Value *RHS, const SimplifyQuery &Q) {
2415 1567158 : if (!match(RHS, m_Zero()))
2416 : return nullptr;
2417 :
2418 : Type *ITy = GetCompareTy(LHS); // The return type.
2419 761695 : switch (Pred) {
2420 0 : default:
2421 0 : llvm_unreachable("Unknown ICmp predicate!");
2422 : case ICmpInst::ICMP_ULT:
2423 112 : return getFalse(ITy);
2424 : case ICmpInst::ICMP_UGE:
2425 75 : return getTrue(ITy);
2426 545890 : case ICmpInst::ICMP_EQ:
2427 : case ICmpInst::ICMP_ULE:
2428 545890 : if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2429 3178 : return getFalse(ITy);
2430 : break;
2431 183113 : case ICmpInst::ICMP_NE:
2432 : case ICmpInst::ICMP_UGT:
2433 183113 : if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2434 223 : return getTrue(ITy);
2435 : break;
2436 9702 : case ICmpInst::ICMP_SLT: {
2437 9702 : KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2438 9702 : if (LHSKnown.isNegative())
2439 22 : return getTrue(ITy);
2440 9699 : if (LHSKnown.isNonNegative())
2441 19 : return getFalse(ITy);
2442 9680 : break;
2443 : }
2444 73 : case ICmpInst::ICMP_SLE: {
2445 73 : KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2446 73 : if (LHSKnown.isNegative())
2447 0 : return getTrue(ITy);
2448 76 : if (LHSKnown.isNonNegative() &&
2449 3 : isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2450 0 : return getFalse(ITy);
2451 73 : break;
2452 : }
2453 5294 : case ICmpInst::ICMP_SGE: {
2454 5294 : KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2455 5294 : if (LHSKnown.isNegative())
2456 3 : return getFalse(ITy);
2457 5294 : if (LHSKnown.isNonNegative())
2458 3 : return getTrue(ITy);
2459 5291 : break;
2460 : }
2461 17436 : case ICmpInst::ICMP_SGT: {
2462 17436 : KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2463 17436 : if (LHSKnown.isNegative())
2464 15 : return getFalse(ITy);
2465 17711 : if (LHSKnown.isNonNegative() &&
2466 276 : isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2467 14 : return getTrue(ITy);
2468 17421 : break;
2469 : }
2470 : }
2471 :
2472 : return nullptr;
2473 : }
2474 :
2475 : /// Many binary operators with a constant operand have an easy-to-compute
2476 : /// range of outputs. This can be used to fold a comparison to always true or
2477 : /// always false.
2478 314511 : static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper,
2479 : const InstrInfoQuery &IIQ) {
2480 314511 : unsigned Width = Lower.getBitWidth();
2481 : const APInt *C;
2482 314511 : switch (BO.getOpcode()) {
2483 : case Instruction::Add:
2484 172191 : if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2485 : // FIXME: If we have both nuw and nsw, we should reduce the range further.
2486 51933 : if (IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
2487 : // 'add nuw x, C' produces [C, UINT_MAX].
2488 15342 : Lower = *C;
2489 36554 : } else if (IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
2490 10360 : if (C->isNegative()) {
2491 : // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2492 7406 : Lower = APInt::getSignedMinValue(Width);
2493 14812 : Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2494 : } else {
2495 : // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2496 2954 : Lower = APInt::getSignedMinValue(Width) + *C;
2497 5908 : Upper = APInt::getSignedMaxValue(Width) + 1;
2498 : }
2499 : }
2500 : }
2501 : break;
2502 :
2503 : case Instruction::And:
2504 283574 : if (match(BO.getOperand(1), m_APInt(C)))
2505 : // 'and x, C' produces [0, C].
2506 237388 : Upper = *C + 1;
2507 : break;
2508 :
2509 : case Instruction::Or:
2510 2540 : if (match(BO.getOperand(1), m_APInt(C)))
2511 : // 'or x, C' produces [C, UINT_MAX].
2512 526 : Lower = *C;
2513 : break;
2514 :
2515 : case Instruction::AShr:
2516 7190 : if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2517 : // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2518 7092 : Lower = APInt::getSignedMinValue(Width).ashr(*C);
2519 7093 : Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2520 98 : } else if (match(BO.getOperand(0), m_APInt(C))) {
2521 36 : unsigned ShiftAmount = Width - 1;
2522 108 : if (!C->isNullValue() && IIQ.isExact(&BO))
2523 16 : ShiftAmount = C->countTrailingZeros();
2524 36 : if (C->isNegative()) {
2525 : // 'ashr C, x' produces [C, C >> (Width-1)]
2526 35 : Lower = *C;
2527 70 : Upper = C->ashr(ShiftAmount) + 1;
2528 : } else {
2529 : // 'ashr C, x' produces [C >> (Width-1), C]
2530 1 : Lower = C->ashr(ShiftAmount);
2531 2 : Upper = *C + 1;
2532 : }
2533 : }
2534 : break;
2535 :
2536 : case Instruction::LShr:
2537 7054 : if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2538 : // 'lshr x, C' produces [0, UINT_MAX >> C].
2539 6546 : Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2540 512 : } else if (match(BO.getOperand(0), m_APInt(C))) {
2541 : // 'lshr C, x' produces [C >> (Width-1), C].
2542 103 : unsigned ShiftAmount = Width - 1;
2543 309 : if (!C->isNullValue() && IIQ.isExact(&BO))
2544 16 : ShiftAmount = C->countTrailingZeros();
2545 103 : Lower = C->lshr(ShiftAmount);
2546 206 : Upper = *C + 1;
2547 : }
2548 : break;
2549 :
2550 : case Instruction::Shl:
2551 1790 : if (match(BO.getOperand(0), m_APInt(C))) {
2552 104 : if (IIQ.hasNoUnsignedWrap(&BO)) {
2553 : // 'shl nuw C, x' produces [C, C << CLZ(C)]
2554 3 : Lower = *C;
2555 6 : Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2556 49 : } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2557 8 : if (C->isNegative()) {
2558 : // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2559 6 : unsigned ShiftAmount = C->countLeadingOnes() - 1;
2560 6 : Lower = C->shl(ShiftAmount);
2561 12 : Upper = *C + 1;
2562 : } else {
2563 : // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2564 2 : unsigned ShiftAmount = C->countLeadingZeros() - 1;
2565 2 : Lower = *C;
2566 4 : Upper = C->shl(ShiftAmount) + 1;
2567 : }
2568 : }
2569 : }
2570 : break;
2571 :
2572 : case Instruction::SDiv:
2573 5974 : if (match(BO.getOperand(1), m_APInt(C))) {
2574 2984 : APInt IntMin = APInt::getSignedMinValue(Width);
2575 2984 : APInt IntMax = APInt::getSignedMaxValue(Width);
2576 5968 : if (C->isAllOnesValue()) {
2577 : // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2578 : // where C != -1 and C != 0 and C != 1
2579 2 : Lower = IntMin + 1;
2580 2 : Upper = IntMax + 1;
2581 2982 : } else if (C->countLeadingZeros() < Width - 1) {
2582 : // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2583 : // where C != -1 and C != 0 and C != 1
2584 2982 : Lower = IntMin.sdiv(*C);
2585 5964 : Upper = IntMax.sdiv(*C);
2586 2982 : if (Lower.sgt(Upper))
2587 : std::swap(Lower, Upper);
2588 2982 : Upper = Upper + 1;
2589 : assert(Upper != Lower && "Upper part of range has wrapped!");
2590 : }
2591 6 : } else if (match(BO.getOperand(0), m_APInt(C))) {
2592 3 : if (C->isMinSignedValue()) {
2593 : // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2594 1 : Lower = *C;
2595 2 : Upper = Lower.lshr(1) + 1;
2596 : } else {
2597 : // 'sdiv C, x' produces [-|C|, |C|].
2598 4 : Upper = C->abs() + 1;
2599 2 : Lower = (-Upper) + 1;
2600 : }
2601 : }
2602 : break;
2603 :
2604 : case Instruction::UDiv:
2605 4193 : if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2606 : // 'udiv x, C' produces [0, UINT_MAX / C].
2607 374 : Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2608 3632 : } else if (match(BO.getOperand(0), m_APInt(C))) {
2609 : // 'udiv C, x' produces [0, C].
2610 3382 : Upper = *C + 1;
2611 : }
2612 : break;
2613 :
2614 : case Instruction::SRem:
2615 634 : if (match(BO.getOperand(1), m_APInt(C))) {
2616 : // 'srem x, C' produces (-|C|, |C|).
2617 128 : Upper = C->abs();
2618 64 : Lower = (-Upper) + 1;
2619 : }
2620 : break;
2621 :
2622 : case Instruction::URem:
2623 4294 : if (match(BO.getOperand(1), m_APInt(C)))
2624 : // 'urem x, C' produces [0, C).
2625 1164 : Upper = *C;
2626 : break;
2627 :
2628 : default:
2629 : break;
2630 : }
2631 314511 : }
2632 :
2633 1563530 : static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2634 : Value *RHS, const InstrInfoQuery &IIQ) {
2635 : Type *ITy = GetCompareTy(RHS); // The return type.
2636 :
2637 : Value *X;
2638 : // Sign-bit checks can be optimized to true/false after unsigned
2639 : // floating-point casts:
2640 : // icmp slt (bitcast (uitofp X)), 0 --> false
2641 : // icmp sgt (bitcast (uitofp X)), -1 --> true
2642 1563530 : if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2643 45 : if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2644 9 : return ConstantInt::getFalse(ITy);
2645 36 : if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2646 9 : return ConstantInt::getTrue(ITy);
2647 : }
2648 :
2649 : const APInt *C;
2650 1563512 : if (!match(RHS, m_APInt(C)))
2651 : return nullptr;
2652 :
2653 : // Rule out tautological comparisons (eg., ult 0 or uge 0).
2654 1529152 : ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2655 764576 : if (RHS_CR.isEmptySet())
2656 804 : return ConstantInt::getFalse(ITy);
2657 763772 : if (RHS_CR.isFullSet())
2658 5 : return ConstantInt::getTrue(ITy);
2659 :
2660 : // Find the range of possible values for binary operators.
2661 763767 : unsigned Width = C->getBitWidth();
2662 : APInt Lower = APInt(Width, 0);
2663 : APInt Upper = APInt(Width, 0);
2664 : if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2665 314511 : setLimitsForBinOp(*BO, Lower, Upper, IIQ);
2666 :
2667 : ConstantRange LHS_CR =
2668 2001465 : Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2669 :
2670 : if (auto *I = dyn_cast<Instruction>(LHS))
2671 1118077 : if (auto *Ranges = IIQ.getMetadata(I, LLVMContext::MD_range))
2672 41198 : LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2673 :
2674 763767 : if (!LHS_CR.isFullSet()) {
2675 199175 : if (RHS_CR.contains(LHS_CR))
2676 454 : return ConstantInt::getTrue(ITy);
2677 198721 : if (RHS_CR.inverse().contains(LHS_CR))
2678 344 : return ConstantInt::getFalse(ITy);
2679 : }
2680 :
2681 : return nullptr;
2682 : }
2683 :
2684 : /// TODO: A large part of this logic is duplicated in InstCombine's
2685 : /// foldICmpBinOp(). We should be able to share that and avoid the code
2686 : /// duplication.
2687 1555246 : static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2688 : Value *RHS, const SimplifyQuery &Q,
2689 : unsigned MaxRecurse) {
2690 : Type *ITy = GetCompareTy(LHS); // The return type.
2691 :
2692 : BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2693 : BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2694 1555246 : if (MaxRecurse && (LBO || RBO)) {
2695 : // Analyze the case when either LHS or RHS is an add instruction.
2696 : Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2697 : // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2698 : bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2699 387046 : if (LBO && LBO->getOpcode() == Instruction::Add) {
2700 : A = LBO->getOperand(0);
2701 : B = LBO->getOperand(1);
2702 : NoLHSWrapProblem =
2703 62801 : ICmpInst::isEquality(Pred) ||
2704 115741 : (CmpInst::isUnsigned(Pred) &&
2705 186100 : Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
2706 58893 : (CmpInst::isSigned(Pred) &&
2707 9849 : Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
2708 : }
2709 387046 : if (RBO && RBO->getOpcode() == Instruction::Add) {
2710 : C = RBO->getOperand(0);
2711 : D = RBO->getOperand(1);
2712 : NoRHSWrapProblem =
2713 8378 : ICmpInst::isEquality(Pred) ||
2714 16394 : (CmpInst::isUnsigned(Pred) &&
2715 31173 : Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
2716 8500 : (CmpInst::isSigned(Pred) &&
2717 357 : Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
2718 : }
2719 :
2720 : // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2721 387046 : if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2722 2319 : if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2723 1164 : Constant::getNullValue(RHS->getType()), Q,
2724 : MaxRecurse - 1))
2725 : return V;
2726 :
2727 : // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2728 387025 : if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2729 1 : if (Value *V =
2730 1 : SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2731 : C == LHS ? D : C, Q, MaxRecurse - 1))
2732 : return V;
2733 :
2734 : // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2735 387024 : if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2736 : NoRHSWrapProblem) {
2737 : // Determine Y and Z in the form icmp (X+Y), (X+Z).
2738 : Value *Y, *Z;
2739 16 : if (A == C) {
2740 : // C + B == C + D -> B == D
2741 : Y = B;
2742 : Z = D;
2743 12 : } else if (A == D) {
2744 : // D + B == C + D -> B == C
2745 : Y = B;
2746 : Z = C;
2747 10 : } else if (B == C) {
2748 : // A + C == C + D -> A == D
2749 : Y = A;
2750 : Z = D;
2751 : } else {
2752 : assert(B == D);
2753 : // A + D == C + D -> A == C
2754 : Y = A;
2755 : Z = C;
2756 : }
2757 16 : if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2758 : return V;
2759 : }
2760 : }
2761 :
2762 : {
2763 1555218 : Value *Y = nullptr;
2764 : // icmp pred (or X, Y), X
2765 1555218 : if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2766 1027 : if (Pred == ICmpInst::ICMP_ULT)
2767 28 : return getFalse(ITy);
2768 1026 : if (Pred == ICmpInst::ICMP_UGE)
2769 1 : return getTrue(ITy);
2770 :
2771 1025 : if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2772 204 : KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2773 204 : KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2774 156 : if (RHSKnown.isNonNegative() && YKnown.isNegative())
2775 14 : return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2776 206 : if (RHSKnown.isNegative() || YKnown.isNonNegative())
2777 10 : return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2778 : }
2779 : }
2780 : // icmp pred X, (or X, Y)
2781 1555204 : if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2782 108 : if (Pred == ICmpInst::ICMP_ULE)
2783 1 : return getTrue(ITy);
2784 107 : if (Pred == ICmpInst::ICMP_UGT)
2785 1 : return getFalse(ITy);
2786 :
2787 106 : if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2788 200 : KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2789 200 : KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2790 152 : if (LHSKnown.isNonNegative() && YKnown.isNegative())
2791 14 : return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2792 202 : if (LHSKnown.isNegative() || YKnown.isNonNegative())
2793 10 : return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2794 : }
2795 : }
2796 : }
2797 :
2798 : // icmp pred (and X, Y), X
2799 1555190 : if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2800 2912 : if (Pred == ICmpInst::ICMP_UGT)
2801 4 : return getFalse(ITy);
2802 2908 : if (Pred == ICmpInst::ICMP_ULE)
2803 4 : return getTrue(ITy);
2804 : }
2805 : // icmp pred X, (and X, Y)
2806 1555182 : if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2807 739 : if (Pred == ICmpInst::ICMP_UGE)
2808 4 : return getTrue(ITy);
2809 735 : if (Pred == ICmpInst::ICMP_ULT)
2810 4 : return getFalse(ITy);
2811 : }
2812 :
2813 : // 0 - (zext X) pred C
2814 2774131 : if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2815 : if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2816 6 : if (RHSC->getValue().isStrictlyPositive()) {
2817 4 : if (Pred == ICmpInst::ICMP_SLT)
2818 1 : return ConstantInt::getTrue(RHSC->getContext());
2819 3 : if (Pred == ICmpInst::ICMP_SGE)
2820 1 : return ConstantInt::getFalse(RHSC->getContext());
2821 2 : if (Pred == ICmpInst::ICMP_EQ)
2822 1 : return ConstantInt::getFalse(RHSC->getContext());
2823 1 : if (Pred == ICmpInst::ICMP_NE)
2824 1 : return ConstantInt::getTrue(RHSC->getContext());
2825 : }
2826 2 : if (RHSC->getValue().isNonNegative()) {
2827 2 : if (Pred == ICmpInst::ICMP_SLE)
2828 1 : return ConstantInt::getTrue(RHSC->getContext());
2829 1 : if (Pred == ICmpInst::ICMP_SGT)
2830 1 : return ConstantInt::getFalse(RHSC->getContext());
2831 : }
2832 : }
2833 : }
2834 :
2835 : // icmp pred (urem X, Y), Y
2836 1555168 : if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2837 32 : switch (Pred) {
2838 : default:
2839 : break;
2840 1 : case ICmpInst::ICMP_SGT:
2841 : case ICmpInst::ICMP_SGE: {
2842 1 : KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2843 1 : if (!Known.isNonNegative())
2844 : break;
2845 : LLVM_FALLTHROUGH;
2846 : }
2847 : case ICmpInst::ICMP_EQ:
2848 : case ICmpInst::ICMP_UGT:
2849 : case ICmpInst::ICMP_UGE:
2850 10 : return getFalse(ITy);
2851 1 : case ICmpInst::ICMP_SLT:
2852 : case ICmpInst::ICMP_SLE: {
2853 1 : KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2854 1 : if (!Known.isNonNegative())
2855 : break;
2856 : LLVM_FALLTHROUGH;
2857 : }
2858 : case ICmpInst::ICMP_NE:
2859 : case ICmpInst::ICMP_ULT:
2860 : case ICmpInst::ICMP_ULE:
2861 20 : return getTrue(ITy);
2862 : }
2863 : }
2864 :
2865 : // icmp pred X, (urem Y, X)
2866 1555138 : if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2867 358 : switch (Pred) {
2868 : default:
2869 : break;
2870 0 : case ICmpInst::ICMP_SGT:
2871 : case ICmpInst::ICMP_SGE: {
2872 0 : KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2873 0 : if (!Known.isNonNegative())
2874 : break;
2875 : LLVM_FALLTHROUGH;
2876 : }
2877 : case ICmpInst::ICMP_NE:
2878 : case ICmpInst::ICMP_UGT:
2879 : case ICmpInst::ICMP_UGE:
2880 1 : return getTrue(ITy);
2881 0 : case ICmpInst::ICMP_SLT:
2882 : case ICmpInst::ICMP_SLE: {
2883 0 : KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2884 0 : if (!Known.isNonNegative())
2885 : break;
2886 : LLVM_FALLTHROUGH;
2887 : }
2888 : case ICmpInst::ICMP_EQ:
2889 : case ICmpInst::ICMP_ULT:
2890 : case ICmpInst::ICMP_ULE:
2891 357 : return getFalse(ITy);
2892 : }
2893 : }
2894 :
2895 : // x >> y <=u x
2896 : // x udiv y <=u x.
2897 1554780 : if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2898 1554774 : match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2899 : // icmp pred (X op Y), X
2900 10 : if (Pred == ICmpInst::ICMP_UGT)
2901 2 : return getFalse(ITy);
2902 8 : if (Pred == ICmpInst::ICMP_ULE)
2903 2 : return getTrue(ITy);
2904 : }
2905 :
2906 : // x >=u x >> y
2907 : // x >=u x udiv y.
2908 1554776 : if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2909 1554774 : match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2910 : // icmp pred X, (X op Y)
2911 4 : if (Pred == ICmpInst::ICMP_ULT)
2912 2 : return getFalse(ITy);
2913 2 : if (Pred == ICmpInst::ICMP_UGE)
2914 2 : return getTrue(ITy);
2915 : }
2916 :
2917 : // handle:
2918 : // CI2 << X == CI
2919 : // CI2 << X != CI
2920 : //
2921 : // where CI2 is a power of 2 and CI isn't
2922 : if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2923 : const APInt *CI2Val, *CIVal = &CI->getValue();
2924 753549 : if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2925 28 : CI2Val->isPowerOf2()) {
2926 22 : if (!CIVal->isPowerOf2()) {
2927 : // CI2 << X can equal zero in some circumstances,
2928 : // this simplification is unsafe if CI is zero.
2929 : //
2930 : // We know it is safe if:
2931 : // - The shift is nsw, we can't shift out the one bit.
2932 : // - The shift is nuw, we can't shift out the one bit.
2933 : // - CI2 is one
2934 : // - CI isn't zero
2935 26 : if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2936 12 : Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2937 14 : CI2Val->isOneValue() || !CI->isZero()) {
2938 12 : if (Pred == ICmpInst::ICMP_EQ)
2939 5 : return ConstantInt::getFalse(RHS->getContext());
2940 10 : if (Pred == ICmpInst::ICMP_NE)
2941 1 : return ConstantInt::getTrue(RHS->getContext());
2942 : }
2943 : }
2944 23 : if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2945 4 : if (Pred == ICmpInst::ICMP_UGT)
2946 1 : return ConstantInt::getFalse(RHS->getContext());
2947 3 : if (Pred == ICmpInst::ICMP_ULE)
2948 1 : return ConstantInt::getTrue(RHS->getContext());
2949 : }
2950 : }
2951 : }
2952 :
2953 1560267 : if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2954 : LBO->getOperand(1) == RBO->getOperand(1)) {
2955 2078 : switch (LBO->getOpcode()) {
2956 : default:
2957 : break;
2958 408 : case Instruction::UDiv:
2959 : case Instruction::LShr:
2960 817 : if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
2961 : !Q.IIQ.isExact(RBO))
2962 : break;
2963 2 : if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2964 : RBO->getOperand(0), Q, MaxRecurse - 1))
2965 1 : return V;
2966 : break;
2967 : case Instruction::SDiv:
2968 1063 : if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
2969 : !Q.IIQ.isExact(RBO))
2970 : break;
2971 600 : if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2972 : RBO->getOperand(0), Q, MaxRecurse - 1))
2973 300 : return V;
2974 : break;
2975 562 : case Instruction::AShr:
2976 1672 : if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
2977 : break;
2978 1096 : if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2979 : RBO->getOperand(0), Q, MaxRecurse - 1))
2980 548 : return V;
2981 : break;
2982 60 : case Instruction::Shl: {
2983 177 : bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
2984 120 : bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
2985 60 : if (!NUW && !NSW)
2986 : break;
2987 57 : if (!NSW && ICmpInst::isSigned(Pred))
2988 : break;
2989 0 : if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2990 : RBO->getOperand(0), Q, MaxRecurse - 1))
2991 0 : return V;
2992 : break;
2993 : }
2994 : }
2995 : }
2996 : return nullptr;
2997 : }
2998 :
2999 : /// Simplify integer comparisons where at least one operand of the compare
3000 : /// matches an integer min/max idiom.
3001 1554765 : static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
3002 : Value *RHS, const SimplifyQuery &Q,
3003 : unsigned MaxRecurse) {
3004 : Type *ITy = GetCompareTy(LHS); // The return type.
3005 : Value *A, *B;
3006 : CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
3007 : CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
3008 :
3009 : // Signed variants on "max(a,b)>=a -> true".
3010 1554765 : if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3011 101 : if (A != RHS)
3012 : std::swap(A, B); // smax(A, B) pred A.
3013 : EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3014 : // We analyze this as smax(A, B) pred A.
3015 : P = Pred;
3016 1554675 : } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
3017 128 : (A == LHS || B == LHS)) {
3018 11 : if (A != LHS)
3019 : std::swap(A, B); // A pred smax(A, B).
3020 : EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
3021 : // We analyze this as smax(A, B) swapped-pred A.
3022 11 : P = CmpInst::getSwappedPredicate(Pred);
3023 1554669 : } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3024 527 : (A == RHS || B == RHS)) {
3025 16 : if (A != RHS)
3026 : std::swap(A, B); // smin(A, B) pred A.
3027 : EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3028 : // We analyze this as smax(-A, -B) swapped-pred -A.
3029 : // Note that we do not need to actually form -A or -B thanks to EqP.
3030 16 : P = CmpInst::getSwappedPredicate(Pred);
3031 1555272 : } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
3032 635 : (A == LHS || B == LHS)) {
3033 68 : if (A != LHS)
3034 : std::swap(A, B); // A pred smin(A, B).
3035 : EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
3036 : // We analyze this as smax(-A, -B) pred -A.
3037 : // Note that we do not need to actually form -A or -B thanks to EqP.
3038 : P = Pred;
3039 : }
3040 196 : if (P != CmpInst::BAD_ICMP_PREDICATE) {
3041 : // Cases correspond to "max(A, B) p A".
3042 196 : switch (P) {
3043 : default:
3044 : break;
3045 18 : case CmpInst::ICMP_EQ:
3046 : case CmpInst::ICMP_SLE:
3047 : // Equivalent to "A EqP B". This may be the same as the condition tested
3048 : // in the max/min; if so, we can just return that.
3049 18 : if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3050 17 : return V;
3051 17 : if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3052 16 : return V;
3053 : // Otherwise, see if "A EqP B" simplifies.
3054 16 : if (MaxRecurse)
3055 16 : if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3056 : return V;
3057 : break;
3058 73 : case CmpInst::ICMP_NE:
3059 : case CmpInst::ICMP_SGT: {
3060 73 : CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3061 : // Equivalent to "A InvEqP B". This may be the same as the condition
3062 : // tested in the max/min; if so, we can just return that.
3063 73 : if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3064 : return V;
3065 69 : if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3066 : return V;
3067 : // Otherwise, see if "A InvEqP B" simplifies.
3068 47 : if (MaxRecurse)
3069 47 : if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3070 : return V;
3071 : break;
3072 : }
3073 : case CmpInst::ICMP_SGE:
3074 : // Always true.
3075 8 : return getTrue(ITy);
3076 : case CmpInst::ICMP_SLT:
3077 : // Always false.
3078 97 : return getFalse(ITy);
3079 : }
3080 : }
3081 :
3082 : // Unsigned variants on "max(a,b)>=a -> true".
3083 : P = CmpInst::BAD_ICMP_PREDICATE;
3084 1554632 : if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3085 126 : if (A != RHS)
3086 : std::swap(A, B); // umax(A, B) pred A.
3087 : EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3088 : // We analyze this as umax(A, B) pred A.
3089 : P = Pred;
3090 1554521 : } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
3091 117 : (A == LHS || B == LHS)) {
3092 15 : if (A != LHS)
3093 : std::swap(A, B); // A pred umax(A, B).
3094 : EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
3095 : // We analyze this as umax(A, B) swapped-pred A.
3096 15 : P = CmpInst::getSwappedPredicate(Pred);
3097 1554502 : } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3098 2432 : (A == RHS || B == RHS)) {
3099 11 : if (A != RHS)
3100 : std::swap(A, B); // umin(A, B) pred A.
3101 : EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3102 : // We analyze this as umax(-A, -B) swapped-pred -A.
3103 : // Note that we do not need to actually form -A or -B thanks to EqP.
3104 11 : P = CmpInst::getSwappedPredicate(Pred);
3105 1555643 : } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3106 1163 : (A == LHS || B == LHS)) {
3107 65 : if (A != LHS)
3108 : std::swap(A, B); // A pred umin(A, B).
3109 : EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
3110 : // We analyze this as umax(-A, -B) pred -A.
3111 : // Note that we do not need to actually form -A or -B thanks to EqP.
3112 : P = Pred;
3113 : }
3114 217 : if (P != CmpInst::BAD_ICMP_PREDICATE) {
3115 : // Cases correspond to "max(A, B) p A".
3116 217 : switch (P) {
3117 : default:
3118 : break;
3119 19 : case CmpInst::ICMP_EQ:
3120 : case CmpInst::ICMP_ULE:
3121 : // Equivalent to "A EqP B". This may be the same as the condition tested
3122 : // in the max/min; if so, we can just return that.
3123 19 : if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3124 18 : return V;
3125 18 : if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3126 17 : return V;
3127 : // Otherwise, see if "A EqP B" simplifies.
3128 17 : if (MaxRecurse)
3129 17 : if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3130 : return V;
3131 : break;
3132 71 : case CmpInst::ICMP_NE:
3133 : case CmpInst::ICMP_UGT: {
3134 71 : CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
3135 : // Equivalent to "A InvEqP B". This may be the same as the condition
3136 : // tested in the max/min; if so, we can just return that.
3137 71 : if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3138 : return V;
3139 67 : if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3140 : return V;
3141 : // Otherwise, see if "A InvEqP B" simplifies.
3142 45 : if (MaxRecurse)
3143 45 : if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3144 : return V;
3145 : break;
3146 : }
3147 : case CmpInst::ICMP_UGE:
3148 : // Always true.
3149 4 : return getTrue(ITy);
3150 : case CmpInst::ICMP_ULT:
3151 : // Always false.
3152 123 : return getFalse(ITy);
3153 : }
3154 : }
3155 :
3156 : // Variants on "max(x,y) >= min(x,z)".
3157 : Value *C, *D;
3158 1555107 : if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3159 1554515 : match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3160 38 : (A == C || A == D || B == C || B == D)) {
3161 : // max(x, ?) pred min(x, ?).
3162 38 : if (Pred == CmpInst::ICMP_SGE)
3163 : // Always true.
3164 1 : return getTrue(ITy);
3165 37 : if (Pred == CmpInst::ICMP_SLT)
3166 : // Always false.
3167 1 : return getFalse(ITy);
3168 1554956 : } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3169 1554441 : match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3170 2 : (A == C || A == D || B == C || B == D)) {
3171 : // min(x, ?) pred max(x, ?).
3172 2 : if (Pred == CmpInst::ICMP_SLE)
3173 : // Always true.
3174 1 : return getTrue(ITy);
3175 1 : if (Pred == CmpInst::ICMP_SGT)
3176 : // Always false.
3177 1 : return getFalse(ITy);
3178 1555352 : } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3179 1554475 : match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3180 38 : (A == C || A == D || B == C || B == D)) {
3181 : // max(x, ?) pred min(x, ?).
3182 38 : if (Pred == CmpInst::ICMP_UGE)
3183 : // Always true.
3184 1 : return getTrue(ITy);
3185 37 : if (Pred == CmpInst::ICMP_ULT)
3186 : // Always false.
3187 1 : return getFalse(ITy);
3188 1556827 : } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3189 1556827 : match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3190 2 : (A == C || A == D || B == C || B == D)) {
3191 : // min(x, ?) pred max(x, ?).
3192 2 : if (Pred == CmpInst::ICMP_ULE)
3193 : // Always true.
3194 1 : return getTrue(ITy);
3195 1 : if (Pred == CmpInst::ICMP_UGT)
3196 : // Always false.
3197 1 : return getFalse(ITy);
3198 : }
3199 :
3200 : return nullptr;
3201 : }
3202 :
3203 : /// Given operands for an ICmpInst, see if we can fold the result.
3204 : /// If not, this returns null.
3205 1621688 : static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3206 : const SimplifyQuery &Q, unsigned MaxRecurse) {
3207 : CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3208 : assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!");
3209 :
3210 : if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3211 : if (Constant *CRHS = dyn_cast<Constant>(RHS))
3212 49035 : return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3213 :
3214 : // If we have a constant, make sure it is on the RHS.
3215 : std::swap(LHS, RHS);
3216 9342 : Pred = CmpInst::getSwappedPredicate(Pred);
3217 : }
3218 :
3219 : Type *ITy = GetCompareTy(LHS); // The return type.
3220 :
3221 : // icmp X, X -> true/false
3222 : // icmp X, undef -> true/false because undef could be X.
3223 1572653 : if (LHS == RHS || isa<UndefValue>(RHS))
3224 1273 : return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3225 :
3226 1571380 : if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3227 : return V;
3228 :
3229 1567158 : if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3230 : return V;
3231 :
3232 1563530 : if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3233 : return V;
3234 :
3235 : // If both operands have range metadata, use the metadata
3236 : // to simplify the comparison.
3237 1561905 : if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3238 : auto RHS_Instr = cast<Instruction>(RHS);
3239 : auto LHS_Instr = cast<Instruction>(LHS);
3240 :
3241 541735 : if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3242 657 : Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3243 : auto RHS_CR = getConstantRangeFromMetadata(
3244 1266 : *RHS_Instr->getMetadata(LLVMContext::MD_range));
3245 : auto LHS_CR = getConstantRangeFromMetadata(
3246 1266 : *LHS_Instr->getMetadata(LLVMContext::MD_range));
3247 :
3248 1266 : auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3249 634 : if (Satisfied_CR.contains(LHS_CR))
3250 2 : return ConstantInt::getTrue(RHS->getContext());
3251 :
3252 : auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3253 1265 : CmpInst::getInversePredicate(Pred), RHS_CR);
3254 633 : if (InversedSatisfied_CR.contains(LHS_CR))
3255 1 : return ConstantInt::getFalse(RHS->getContext());
3256 : }
3257 : }
3258 :
3259 : // Compare of cast, for example (zext X) != 0 -> X != 0
3260 41693 : if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3261 : Instruction *LI = cast<CastInst>(LHS);
3262 36175 : Value *SrcOp = LI->getOperand(0);
3263 36175 : Type *SrcTy = SrcOp->getType();
3264 36175 : Type *DstTy = LI->getType();
3265 :
3266 : // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3267 : // if the integer type is the same size as the pointer type.
3268 36175 : if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3269 1642 : Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3270 : if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3271 : // Transfer the cast to the constant.
3272 972 : if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3273 972 : ConstantExpr::getIntToPtr(RHSC, SrcTy),
3274 : Q, MaxRecurse-1))
3275 : return V;
3276 : } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3277 612 : if (RI->getOperand(0)->getType() == SrcTy)
3278 : // Compare without the cast.
3279 562 : if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3280 : Q, MaxRecurse-1))
3281 : return V;
3282 : }
3283 : }
3284 :
3285 : if (isa<ZExtInst>(LHS)) {
3286 : // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3287 : // same type.
3288 : if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3289 2864 : if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3290 : // Compare X and Y. Note that signed predicates become unsigned.
3291 1262 : if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3292 : SrcOp, RI->getOperand(0), Q,
3293 : MaxRecurse-1))
3294 : return V;
3295 : }
3296 : // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3297 : // too. If not, then try to deduce the result of the comparison.
3298 : else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3299 : // Compute the constant that would happen if we truncated to SrcTy then
3300 : // reextended to DstTy.
3301 15043 : Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3302 15043 : Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3303 :
3304 : // If the re-extended constant didn't change then this is effectively
3305 : // also a case of comparing two zero-extended values.
3306 15043 : if (RExt == CI && MaxRecurse)
3307 13262 : if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3308 : SrcOp, Trunc, Q, MaxRecurse-1))
3309 : return V;
3310 :
3311 : // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3312 : // there. Use this to work out the result of the comparison.
3313 10920 : if (RExt != CI) {
3314 1655 : switch (Pred) {
3315 0 : default: llvm_unreachable("Unknown ICmp predicate!");
3316 : // LHS <u RHS.
3317 1422 : case ICmpInst::ICMP_EQ:
3318 : case ICmpInst::ICMP_UGT:
3319 : case ICmpInst::ICMP_UGE:
3320 1422 : return ConstantInt::getFalse(CI->getContext());
3321 :
3322 15 : case ICmpInst::ICMP_NE:
3323 : case ICmpInst::ICMP_ULT:
3324 : case ICmpInst::ICMP_ULE:
3325 15 : return ConstantInt::getTrue(CI->getContext());
3326 :
3327 : // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3328 : // is non-negative then LHS <s RHS.
3329 : case ICmpInst::ICMP_SGT:
3330 : case ICmpInst::ICMP_SGE:
3331 209 : return CI->getValue().isNegative() ?
3332 207 : ConstantInt::getTrue(CI->getContext()) :
3333 209 : ConstantInt::getFalse(CI->getContext());
3334 :
3335 : case ICmpInst::ICMP_SLT:
3336 : case ICmpInst::ICMP_SLE:
3337 9 : return CI->getValue().isNegative() ?
3338 3 : ConstantInt::getFalse(CI->getContext()) :
3339 9 : ConstantInt::getTrue(CI->getContext());
3340 : }
3341 : }
3342 : }
3343 : }
3344 :
3345 : if (isa<SExtInst>(LHS)) {
3346 : // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3347 : // same type.
3348 : if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3349 3720 : if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3350 : // Compare X and Y. Note that the predicate does not change.
3351 1857 : if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3352 : Q, MaxRecurse-1))
3353 : return V;
3354 : }
3355 : // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3356 : // too. If not, then try to deduce the result of the comparison.
3357 : else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3358 : // Compute the constant that would happen if we truncated to SrcTy then
3359 : // reextended to DstTy.
3360 3040 : Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3361 3040 : Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3362 :
3363 : // If the re-extended constant didn't change then this is effectively
3364 : // also a case of comparing two sign-extended values.
3365 3040 : if (RExt == CI && MaxRecurse)
3366 3001 : if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3367 : return V;
3368 :
3369 : // Otherwise the upper bits of LHS are all equal, while RHS has varying
3370 : // bits there. Use this to work out the result of the comparison.
3371 2971 : if (RExt != CI) {
3372 39 : switch (Pred) {
3373 0 : default: llvm_unreachable("Unknown ICmp predicate!");
3374 7 : case ICmpInst::ICMP_EQ:
3375 7 : return ConstantInt::getFalse(CI->getContext());
3376 8 : case ICmpInst::ICMP_NE:
3377 8 : return ConstantInt::getTrue(CI->getContext());
3378 :
3379 : // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3380 : // LHS >s RHS.
3381 : case ICmpInst::ICMP_SGT:
3382 : case ICmpInst::ICMP_SGE:
3383 9 : return CI->getValue().isNegative() ?
3384 8 : ConstantInt::getTrue(CI->getContext()) :
3385 9 : ConstantInt::getFalse(CI->getContext());
3386 : case ICmpInst::ICMP_SLT:
3387 : case ICmpInst::ICMP_SLE:
3388 9 : return CI->getValue().isNegative() ?
3389 1 : ConstantInt::getFalse(CI->getContext()) :
3390 9 : ConstantInt::getTrue(CI->getContext());
3391 :
3392 : // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3393 : // LHS >u RHS.
3394 3 : case ICmpInst::ICMP_UGT:
3395 : case ICmpInst::ICMP_UGE:
3396 : // Comparison is true iff the LHS <s 0.
3397 3 : if (MaxRecurse)
3398 3 : if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3399 3 : Constant::getNullValue(SrcTy),
3400 : Q, MaxRecurse-1))
3401 : return V;
3402 : break;
3403 3 : case ICmpInst::ICMP_ULT:
3404 : case ICmpInst::ICMP_ULE:
3405 : // Comparison is true iff the LHS >=s 0.
3406 3 : if (MaxRecurse)
3407 3 : if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3408 3 : Constant::getNullValue(SrcTy),
3409 : Q, MaxRecurse-1))
3410 : return V;
3411 : break;
3412 : }
3413 : }
3414 : }
3415 : }
3416 : }
3417 :
3418 : // icmp eq|ne X, Y -> false|true if X != Y
3419 2610181 : if (ICmpInst::isEquality(Pred) &&
3420 1054214 : isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3421 721 : return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3422 : }
3423 :
3424 1555246 : if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3425 : return V;
3426 :
3427 1554765 : if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3428 : return V;
3429 :
3430 : // Simplify comparisons of related pointers using a powerful, recursive
3431 : // GEP-walk when we have target data available..
3432 3108938 : if (LHS->getType()->isPointerTy())
3433 535681 : if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3434 : Q.IIQ, LHS, RHS))
3435 : return C;
3436 : if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3437 : if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3438 1326 : if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3439 1282 : Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3440 1238 : Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3441 619 : Q.DL.getTypeSizeInBits(CRHS->getType()))
3442 619 : if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3443 : Q.IIQ, CLHS->getPointerOperand(),
3444 : CRHS->getPointerOperand()))
3445 : return C;
3446 :
3447 : if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3448 : if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3449 565 : if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3450 7086 : GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3451 10 : (ICmpInst::isEquality(Pred) ||
3452 18 : (GLHS->isInBounds() && GRHS->isInBounds() &&
3453 4 : Pred == ICmpInst::getSignedPredicate(Pred)))) {
3454 : // The bases are equal and the indices are constant. Build a constant
3455 : // expression GEP with the same indices and a null base pointer to see
3456 : // what constant folding can make out of it.
3457 4 : Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3458 : SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3459 4 : Constant *NewLHS = ConstantExpr::getGetElementPtr(
3460 : GLHS->getSourceElementType(), Null, IndicesLHS);
3461 :
3462 : SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3463 4 : Constant *NewRHS = ConstantExpr::getGetElementPtr(
3464 : GLHS->getSourceElementType(), Null, IndicesRHS);
3465 4 : return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3466 : }
3467 : }
3468 : }
3469 :
3470 : // If the comparison is with the result of a select instruction, check whether
3471 : // comparing with either branch of the select always yields the same value.
3472 : if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3473 14001 : if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3474 : return V;
3475 :
3476 : // If the comparison is with the result of a phi instruction, check whether
3477 : // doing the compare with each incoming phi value yields a common result.
3478 : if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3479 123987 : if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3480 353 : return V;
3481 :
3482 : return nullptr;
3483 : }
3484 :
3485 1330972 : Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3486 : const SimplifyQuery &Q) {
3487 1330972 : return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3488 : }
3489 :
3490 : /// Given operands for an FCmpInst, see if we can fold the result.
3491 : /// If not, this returns null.
3492 18027 : static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3493 : FastMathFlags FMF, const SimplifyQuery &Q,
3494 : unsigned MaxRecurse) {
3495 : CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3496 : assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!");
3497 :
3498 : if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3499 : if (Constant *CRHS = dyn_cast<Constant>(RHS))
3500 170 : return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3501 :
3502 : // If we have a constant, make sure it is on the RHS.
3503 : std::swap(LHS, RHS);
3504 277 : Pred = CmpInst::getSwappedPredicate(Pred);
3505 : }
3506 :
3507 : // Fold trivial predicates.
3508 : Type *RetTy = GetCompareTy(LHS);
3509 17857 : if (Pred == FCmpInst::FCMP_FALSE)
3510 35 : return getFalse(RetTy);
3511 17822 : if (Pred == FCmpInst::FCMP_TRUE)
3512 35 : return getTrue(RetTy);
3513 :
3514 : // Fold (un)ordered comparison if we can determine there are no NaNs.
3515 17787 : if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3516 4892 : if (FMF.noNaNs() ||
3517 1653 : (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3518 36 : return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
3519 :
3520 : // NaN is unordered; NaN is not ordered.
3521 : assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&
3522 : "Comparison must be either ordered or unordered");
3523 17751 : if (match(RHS, m_NaN()))
3524 15 : return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3525 :
3526 : // fcmp pred x, undef and fcmp pred undef, x
3527 : // fold to true if unordered, false if ordered
3528 17736 : if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3529 : // Choosing NaN for the undef will always make unordered comparison succeed
3530 : // and ordered comparison fail.
3531 8 : return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3532 : }
3533 :
3534 : // fcmp x,x -> true/false. Not all compares are foldable.
3535 17728 : if (LHS == RHS) {
3536 565 : if (CmpInst::isTrueWhenEqual(Pred))
3537 2 : return getTrue(RetTy);
3538 563 : if (CmpInst::isFalseWhenEqual(Pred))
3539 7 : return getFalse(RetTy);
3540 : }
3541 :
3542 : // Handle fcmp with constant RHS.
3543 : const APFloat *C;
3544 17719 : if (match(RHS, m_APFloat(C))) {
3545 : // Check whether the constant is an infinity.
3546 9676 : if (C->isInfinity()) {
3547 1019 : if (C->isNegative()) {
3548 90 : switch (Pred) {
3549 : case FCmpInst::FCMP_OLT:
3550 : // No value is ordered and less than negative infinity.
3551 1 : return getFalse(RetTy);
3552 : case FCmpInst::FCMP_UGE:
3553 : // All values are unordered with or at least negative infinity.
3554 1 : return getTrue(RetTy);
3555 : default:
3556 : break;
3557 : }
3558 : } else {
3559 929 : switch (Pred) {
3560 : case FCmpInst::FCMP_OGT:
3561 : // No value is ordered and greater than infinity.
3562 1 : return getFalse(RetTy);
3563 : case FCmpInst::FCMP_ULE:
3564 : // All values are unordered with and at most infinity.
3565 3 : return getTrue(RetTy);
3566 : default:
3567 : break;
3568 : }
3569 : }
3570 : }
3571 4832 : if (C->isZero()) {
3572 2358 : switch (Pred) {
3573 24 : case FCmpInst::FCMP_UGE:
3574 24 : if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3575 5 : return getTrue(RetTy);
3576 : break;
3577 176 : case FCmpInst::FCMP_OLT:
3578 : // X < 0
3579 176 : if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3580 6 : return getFalse(RetTy);
3581 : break;
3582 : default:
3583 : break;
3584 : }
3585 2474 : } else if (C->isNegative()) {
3586 : assert(!C->isNaN() && "Unexpected NaN constant!");
3587 : // TODO: We can catch more cases by using a range check rather than
3588 : // relying on CannotBeOrderedLessThanZero.
3589 : switch (Pred) {
3590 27 : case FCmpInst::FCMP_UGE:
3591 : case FCmpInst::FCMP_UGT:
3592 : case FCmpInst::FCMP_UNE:
3593 : // (X >= 0) implies (X > C) when (C < 0)
3594 27 : if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3595 5 : return getTrue(RetTy);
3596 : break;
3597 97 : case FCmpInst::FCMP_OEQ:
3598 : case FCmpInst::FCMP_OLE:
3599 : case FCmpInst::FCMP_OLT:
3600 : // (X >= 0) implies !(X < C) when (C < 0)
3601 97 : if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3602 6 : return getFalse(RetTy);
3603 : break;
3604 : default:
3605 : break;
3606 : }
3607 : }
3608 : }
3609 :
3610 : // If the comparison is with the result of a select instruction, check whether
3611 : // comparing with either branch of the select always yields the same value.
3612 : if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3613 234 : if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3614 : return V;
3615 :
3616 : // If the comparison is with the result of a phi instruction, check whether
3617 : // doing the compare with each incoming phi value yields a common result.
3618 : if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3619 390 : if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3620 3 : return V;
3621 :
3622 : return nullptr;
3623 : }
3624 :
3625 17392 : Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3626 : FastMathFlags FMF, const SimplifyQuery &Q) {
3627 17392 : return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3628 : }
3629 :
3630 : /// See if V simplifies when its operand Op is replaced with RepOp.
3631 269924 : static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3632 : const SimplifyQuery &Q,
3633 : unsigned MaxRecurse) {
3634 : // Trivial replacement.
3635 269924 : if (V == Op)
3636 5236 : return RepOp;
3637 :
3638 : // We cannot replace a constant, and shouldn't even try.
3639 264688 : if (isa<Constant>(Op))
3640 : return nullptr;
3641 :
3642 : auto *I = dyn_cast<Instruction>(V);
3643 : if (!I)
3644 : return nullptr;
3645 :
3646 : // If this is a binary operator, try to simplify it with the replaced op.
3647 : if (auto *B = dyn_cast<BinaryOperator>(I)) {
3648 : // Consider:
3649 : // %cmp = icmp eq i32 %x, 2147483647
3650 : // %add = add nsw i32 %x, 1
3651 : // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3652 : //
3653 : // We can't replace %sel with %add unless we strip away the flags.
3654 : if (isa<OverflowingBinaryOperator>(B))
3655 8254 : if (Q.IIQ.hasNoSignedWrap(B) || Q.IIQ.hasNoUnsignedWrap(B))
3656 251 : return nullptr;
3657 1792 : if (isa<PossiblyExactOperator>(B) && Q.IIQ.isExact(B))
3658 : return nullptr;
3659 :
3660 3325 : if (MaxRecurse) {
3661 3325 : if (B->getOperand(0) == Op)
3662 81 : return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3663 81 : MaxRecurse - 1);
3664 3244 : if (B->getOperand(1) == Op)
3665 21 : return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3666 21 : MaxRecurse - 1);
3667 : }
3668 : }
3669 :
3670 : // Same for CmpInsts.
3671 : if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3672 214 : if (MaxRecurse) {
3673 214 : if (C->getOperand(0) == Op)
3674 17 : return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3675 17 : MaxRecurse - 1);
3676 197 : if (C->getOperand(1) == Op)
3677 0 : return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3678 0 : MaxRecurse - 1);
3679 : }
3680 : }
3681 :
3682 : // Same for GEPs.
3683 : if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3684 30089 : if (MaxRecurse) {
3685 60178 : SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3686 30089 : transform(GEP->operands(), NewOps.begin(),
3687 104914 : [&](Value *V) { return V == Op ? RepOp : V; });
3688 30089 : return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3689 : MaxRecurse - 1);
3690 : }
3691 : }
3692 :
3693 : // TODO: We could hand off more cases to instsimplify here.
3694 :
3695 : // If all operands are constant after substituting Op for RepOp then we can
3696 : // constant fold the instruction.
3697 72893 : if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3698 : // Build a list of all constant operands.
3699 : SmallVector<Constant *, 8> ConstOps;
3700 64297 : for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3701 127524 : if (I->getOperand(i) == Op)
3702 249 : ConstOps.push_back(CRepOp);
3703 63513 : else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3704 2293 : ConstOps.push_back(COp);
3705 : else
3706 : break;
3707 : }
3708 :
3709 : // All operands were constants, fold it.
3710 123510 : if (ConstOps.size() == I->getNumOperands()) {
3711 : if (CmpInst *C = dyn_cast<CmpInst>(I))
3712 0 : return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3713 0 : ConstOps[1], Q.DL, Q.TLI);
3714 :
3715 : if (LoadInst *LI = dyn_cast<LoadInst>(I))
3716 201 : if (!LI->isVolatile())
3717 378 : return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3718 :
3719 692 : return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3720 : }
3721 : }
3722 :
3723 66789 : return nullptr;
3724 : }
3725 :
3726 : /// Try to simplify a select instruction when its condition operand is an
3727 : /// integer comparison where one operand of the compare is a constant.
3728 33857 : static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3729 : const APInt *Y, bool TrueWhenUnset) {
3730 : const APInt *C;
3731 :
3732 : // (X & Y) == 0 ? X & ~Y : X --> X
3733 : // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3734 33868 : if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3735 33890 : *Y == ~*C)
3736 17 : return TrueWhenUnset ? FalseVal : TrueVal;
3737 :
3738 : // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3739 : // (X & Y) != 0 ? X : X & ~Y --> X
3740 33857 : if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3741 33879 : *Y == ~*C)
3742 16 : return TrueWhenUnset ? FalseVal : TrueVal;
3743 :
3744 33835 : if (Y->isPowerOf2()) {
3745 : // (X & Y) == 0 ? X | Y : X --> X | Y
3746 : // (X & Y) != 0 ? X | Y : X --> X
3747 28485 : if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3748 8 : *Y == *C)
3749 11 : return TrueWhenUnset ? TrueVal : FalseVal;
3750 :
3751 : // (X & Y) == 0 ? X : X | Y --> X
3752 : // (X & Y) != 0 ? X : X | Y --> X | Y
3753 28476 : if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3754 7 : *Y == *C)
3755 11 : return TrueWhenUnset ? TrueVal : FalseVal;
3756 : }
3757 :
3758 : return nullptr;
3759 : }
3760 :
3761 : /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3762 : /// eq/ne.
3763 99009 : static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3764 : ICmpInst::Predicate Pred,
3765 : Value *TrueVal, Value *FalseVal) {
3766 : Value *X;
3767 : APInt Mask;
3768 99009 : if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3769 : return nullptr;
3770 :
3771 6826 : return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3772 6826 : Pred == ICmpInst::ICMP_EQ);
3773 : }
3774 :
3775 : /// Try to simplify a select instruction when its condition operand is an
3776 : /// integer comparison.
3777 118609 : static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3778 : Value *FalseVal, const SimplifyQuery &Q,
3779 : unsigned MaxRecurse) {
3780 : ICmpInst::Predicate Pred;
3781 : Value *CmpLHS, *CmpRHS;
3782 : if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3783 19588 : return nullptr;
3784 :
3785 166543 : if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3786 : Value *X;
3787 : const APInt *Y;
3788 35370 : if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3789 27031 : if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3790 : Pred == ICmpInst::ICMP_EQ))
3791 12 : return V;
3792 : }
3793 :
3794 : // Check for other compares that behave like bit test.
3795 99009 : if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3796 : TrueVal, FalseVal))
3797 : return V;
3798 :
3799 : // If we have an equality comparison, then we know the value in one of the
3800 : // arms of the select. See if substituting this value into the arm and
3801 : // simplifying the result yields the same value as the other arm.
3802 98984 : if (Pred == ICmpInst::ICMP_EQ) {
3803 65853 : if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3804 131683 : TrueVal ||
3805 65830 : SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3806 : TrueVal)
3807 25 : return FalseVal;
3808 65828 : if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3809 131654 : FalseVal ||
3810 65826 : SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3811 : FalseVal)
3812 2 : return FalseVal;
3813 33131 : } else if (Pred == ICmpInst::ICMP_NE) {
3814 1657 : if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3815 3301 : FalseVal ||
3816 1644 : SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3817 : FalseVal)
3818 14 : return TrueVal;
3819 1643 : if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3820 3286 : TrueVal ||
3821 1643 : SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3822 : TrueVal)
3823 0 : return TrueVal;
3824 : }
3825 :
3826 : return nullptr;
3827 : }
3828 :
3829 : /// Given operands for a SelectInst, see if we can fold the result.
3830 : /// If not, this returns null.
3831 119228 : static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3832 : const SimplifyQuery &Q, unsigned MaxRecurse) {
3833 : if (auto *CondC = dyn_cast<Constant>(Cond)) {
3834 : if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3835 : if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3836 79 : return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3837 :
3838 : // select undef, X, Y -> X or Y
3839 526 : if (isa<UndefValue>(CondC))
3840 49 : return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3841 :
3842 : // TODO: Vector constants with undef elements don't simplify.
3843 :
3844 : // select true, X, Y -> X
3845 477 : if (CondC->isAllOnesValue())
3846 : return TrueVal;
3847 : // select false, X, Y -> Y
3848 436 : if (CondC->isNullValue())
3849 : return FalseVal;
3850 : }
3851 :
3852 : // select ?, X, X -> X
3853 118655 : if (TrueVal == FalseVal)
3854 : return TrueVal;
3855 :
3856 118626 : if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3857 : return FalseVal;
3858 118614 : if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3859 : return TrueVal;
3860 :
3861 118609 : if (Value *V =
3862 118609 : simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3863 : return V;
3864 :
3865 118531 : if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
3866 20 : return V;
3867 :
3868 : return nullptr;
3869 : }
3870 :
3871 119228 : Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3872 : const SimplifyQuery &Q) {
3873 119228 : return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3874 : }
3875 :
3876 : /// Given operands for an GetElementPtrInst, see if we can fold the result.
3877 : /// If not, this returns null.
3878 0 : static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3879 : const SimplifyQuery &Q, unsigned) {
3880 : // The type of the GEP pointer operand.
3881 : unsigned AS =
3882 0 : cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3883 :
3884 : // getelementptr P -> P.
3885 0 : if (Ops.size() == 1)
3886 : return Ops[0];
3887 :
3888 : // Compute the (pointer) type returned by the GEP instruction.
3889 0 : Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3890 0 : Type *GEPTy = PointerType::get(LastType, AS);
3891 0 : if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3892 0 : GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3893 0 : else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3894 0 : GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3895 :
3896 0 : if (isa<UndefValue>(Ops[0]))
3897 0 : return UndefValue::get(GEPTy);
3898 :
3899 0 : if (Ops.size() == 2) {
3900 : // getelementptr P, 0 -> P.
3901 0 : if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3902 : return Ops[0];
3903 :
3904 : Type *Ty = SrcTy;
3905 0 : if (Ty->isSized()) {
3906 : Value *P;
3907 : uint64_t C;
3908 0 : uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3909 : // getelementptr P, N -> P if P points to a type of zero size.
3910 0 : if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3911 0 : return Ops[0];
3912 :
3913 : // The following transforms are only safe if the ptrtoint cast
3914 : // doesn't truncate the pointers.
3915 0 : if (Ops[1]->getType()->getScalarSizeInBits() ==
3916 0 : Q.DL.getIndexSizeInBits(AS)) {
3917 : auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3918 : if (match(P, m_Zero()))
3919 : return Constant::getNullValue(GEPTy);
3920 : Value *Temp;
3921 : if (match(P, m_PtrToInt(m_Value(Temp))))
3922 : if (Temp->getType() == GEPTy)
3923 : return Temp;
3924 : return nullptr;
3925 0 : };
3926 :
3927 : // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3928 0 : if (TyAllocSize == 1 &&
3929 0 : match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3930 0 : if (Value *R = PtrToIntOrZero(P))
3931 0 : return R;
3932 :
3933 : // getelementptr V, (ashr (sub P, V), C) -> Q
3934 : // if P points to a type of size 1 << C.
3935 0 : if (match(Ops[1],
3936 0 : m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3937 0 : m_ConstantInt(C))) &&
3938 0 : TyAllocSize == 1ULL << C)
3939 0 : if (Value *R = PtrToIntOrZero(P))
3940 0 : return R;
3941 :
3942 : // getelementptr V, (sdiv (sub P, V), C) -> Q
3943 : // if P points to a type of size C.
3944 0 : if (match(Ops[1],
3945 0 : m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3946 : m_SpecificInt(TyAllocSize))))
3947 0 : if (Value *R = PtrToIntOrZero(P))
3948 0 : return R;
3949 : }
3950 : }
3951 : }
3952 :
3953 0 : if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3954 : all_of(Ops.slice(1).drop_back(1),
3955 : [](Value *Idx) { return match(Idx, m_Zero()); })) {
3956 : unsigned IdxWidth =
3957 0 : Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3958 0 : if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3959 : APInt BasePtrOffset(IdxWidth, 0);
3960 : Value *StrippedBasePtr =
3961 0 : Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3962 : BasePtrOffset);
3963 :
3964 : // gep (gep V, C), (sub 0, V) -> C
3965 0 : if (match(Ops.back(),
3966 0 : m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3967 0 : auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3968 0 : return ConstantExpr::getIntToPtr(CI, GEPTy);
3969 : }
3970 : // gep (gep V, C), (xor V, -1) -> C-1
3971 0 : if (match(Ops.back(),
3972 0 : m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3973 0 : auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3974 0 : return ConstantExpr::getIntToPtr(CI, GEPTy);
3975 : }
3976 : }
3977 : }
3978 :
3979 : // Check to see if this is constant foldable.
3980 0 : if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3981 0 : return nullptr;
3982 :
3983 0 : auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3984 : Ops.slice(1));
3985 0 : if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3986 0 : return CEFolded;
3987 : return CE;
3988 : }
3989 :
3990 6214690 : Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3991 : const SimplifyQuery &Q) {
3992 6214690 : return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3993 : }
3994 :
3995 : /// Given operands for an InsertValueInst, see if we can fold the result.
3996 : /// If not, this returns null.
3997 0 : static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3998 : ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3999 : unsigned) {
4000 : if (Constant *CAgg = dyn_cast<Constant>(Agg))
4001 : if (Constant *CVal = dyn_cast<Constant>(Val))
4002 0 : return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
4003 :
4004 : // insertvalue x, undef, n -> x
4005 0 : if (match(Val, m_Undef()))
4006 0 : return Agg;
4007 :
4008 : // insertvalue x, (extractvalue y, n), n
4009 : if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
4010 0 : if (EV->getAggregateOperand()->getType() == Agg->getType() &&
4011 0 : EV->getIndices() == Idxs) {
4012 : // insertvalue undef, (extractvalue y, n), n -> y
4013 0 : if (match(Agg, m_Undef()))
4014 : return EV->getAggregateOperand();
4015 :
4016 : // insertvalue y, (extractvalue y, n), n -> y
4017 0 : if (Agg == EV->getAggregateOperand())
4018 0 : return Agg;
4019 : }
4020 :
4021 : return nullptr;
4022 : }
4023 :
4024 79973 : Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
4025 : ArrayRef<unsigned> Idxs,
4026 : const SimplifyQuery &Q) {
4027 79973 : return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
4028 : }
4029 :
4030 30052 : Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
4031 : const SimplifyQuery &Q) {
4032 : // Try to constant fold.
4033 : auto *VecC = dyn_cast<Constant>(Vec);
4034 : auto *ValC = dyn_cast<Constant>(Val);
4035 : auto *IdxC = dyn_cast<Constant>(Idx);
4036 30052 : if (VecC && ValC && IdxC)
4037 463 : return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
4038 :
4039 : // Fold into undef if index is out of bounds.
4040 : if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
4041 28375 : uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
4042 28375 : if (CI->uge(NumElements))
4043 9 : return UndefValue::get(Vec->getType());
4044 : }
4045 :
4046 : // If index is undef, it might be out of bounds (see above case)
4047 29580 : if (isa<UndefValue>(Idx))
4048 14 : return UndefValue::get(Vec->getType());
4049 :
4050 : return nullptr;
4051 : }
4052 :
4053 : /// Given operands for an ExtractValueInst, see if we can fold the result.
4054 : /// If not, this returns null.
4055 0 : static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4056 : const SimplifyQuery &, unsigned) {
4057 : if (auto *CAgg = dyn_cast<Constant>(Agg))
4058 0 : return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4059 :
4060 : // extractvalue x, (insertvalue y, elt, n), n -> elt
4061 0 : unsigned NumIdxs = Idxs.size();
4062 0 : for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4063 : IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4064 : ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4065 0 : unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4066 0 : unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4067 0 : if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4068 : Idxs.slice(0, NumCommonIdxs)) {
4069 0 : if (NumIdxs == NumInsertValueIdxs)
4070 0 : return IVI->getInsertedValueOperand();
4071 0 : break;
4072 : }
4073 : }
4074 :
4075 : return nullptr;
4076 : }
4077 :
4078 1116833 : Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4079 : const SimplifyQuery &Q) {
4080 1116833 : return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
4081 : }
4082 :
4083 : /// Given operands for an ExtractElementInst, see if we can fold the result.
4084 : /// If not, this returns null.
4085 0 : static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
4086 : unsigned) {
4087 : if (auto *CVec = dyn_cast<Constant>(Vec)) {
4088 : if (auto *CIdx = dyn_cast<Constant>(Idx))
4089 0 : return ConstantFoldExtractElementInstruction(CVec, CIdx);
4090 :
4091 : // The index is not relevant if our vector is a splat.
4092 0 : if (auto *Splat = CVec->getSplatValue())
4093 0 : return Splat;
4094 :
4095 0 : if (isa<UndefValue>(Vec))
4096 0 : return UndefValue::get(Vec->getType()->getVectorElementType());
4097 : }
4098 :
4099 : // If extracting a specified index from the vector, see if we can recursively
4100 : // find a previously computed scalar that was inserted into the vector.
4101 : if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
4102 0 : if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4103 : // definitely out of bounds, thus undefined result
4104 0 : return UndefValue::get(Vec->getType()->getVectorElementType());
4105 0 : if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4106 0 : return Elt;
4107 : }
4108 :
4109 : // An undef extract index can be arbitrarily chosen to be an out-of-range
4110 : // index value, which would result in the instruction being undef.
4111 0 : if (isa<UndefValue>(Idx))
4112 0 : return UndefValue::get(Vec->getType()->getVectorElementType());
4113 :
4114 : return nullptr;
4115 : }
4116 :
4117 28944 : Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
4118 : const SimplifyQuery &Q) {
4119 28944 : return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
4120 : }
4121 :
4122 : /// See if we can fold the given phi. If not, returns null.
4123 0 : static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
4124 : // If all of the PHI's incoming values are the same then replace the PHI node
4125 : // with the common value.
4126 : Value *CommonValue = nullptr;
4127 : bool HasUndefInput = false;
4128 0 : for (Value *Incoming : PN->incoming_values()) {
4129 : // If the incoming value is the phi node itself, it can safely be skipped.
4130 0 : if (Incoming == PN) continue;
4131 0 : if (isa<UndefValue>(Incoming)) {
4132 : // Remember that we saw an undef value, but otherwise ignore them.
4133 : HasUndefInput = true;
4134 0 : continue;
4135 : }
4136 0 : if (CommonValue && Incoming != CommonValue)
4137 0 : return nullptr; // Not the same, bail out.
4138 : CommonValue = Incoming;
4139 : }
4140 :
4141 : // If CommonValue is null then all of the incoming values were either undef or
4142 : // equal to the phi node itself.
4143 0 : if (!CommonValue)
4144 0 : return UndefValue::get(PN->getType());
4145 :
4146 : // If we have a PHI node like phi(X, undef, X), where X is defined by some
4147 : // instruction, we cannot return X as the result of the PHI node unless it
4148 : // dominates the PHI block.
4149 0 : if (HasUndefInput)
4150 0 : return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4151 :
4152 : return CommonValue;
4153 : }
4154 :
4155 0 : static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4156 : Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4157 : if (auto *C = dyn_cast<Constant>(Op))
4158 0 : return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4159 :
4160 : if (auto *CI = dyn_cast<CastInst>(Op)) {
4161 : auto *Src = CI->getOperand(0);
4162 0 : Type *SrcTy = Src->getType();
4163 0 : Type *MidTy = CI->getType();
4164 : Type *DstTy = Ty;
4165 0 : if (Src->getType() == Ty) {
4166 : auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4167 : auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4168 : Type *SrcIntPtrTy =
4169 0 : SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4170 : Type *MidIntPtrTy =
4171 0 : MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4172 : Type *DstIntPtrTy =
4173 0 : DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4174 0 : if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4175 : SrcIntPtrTy, MidIntPtrTy,
4176 : DstIntPtrTy) == Instruction::BitCast)
4177 0 : return Src;
4178 : }
4179 : }
4180 :
4181 : // bitcast x -> x
4182 0 : if (CastOpc == Instruction::BitCast)
4183 0 : if (Op->getType() == Ty)
4184 0 : return Op;
4185 :
4186 : return nullptr;
4187 : }
4188 :
4189 4704445 : Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4190 : const SimplifyQuery &Q) {
4191 4704445 : return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4192 : }
4193 :
4194 : /// For the given destination element of a shuffle, peek through shuffles to
4195 : /// match a root vector source operand that contains that element in the same
4196 : /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4197 20883 : static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4198 : int MaskVal, Value *RootVec,
4199 : unsigned MaxRecurse) {
4200 21548 : if (!MaxRecurse--)
4201 : return nullptr;
4202 :
4203 : // Bail out if any mask value is undefined. That kind of shuffle may be
4204 : // simplified further based on demanded bits or other folds.
4205 21544 : if (MaskVal == -1)
4206 : return nullptr;
4207 :
4208 : // The mask value chooses which source operand we need to look at next.
4209 21538 : int InVecNumElts = Op0->getType()->getVectorNumElements();
4210 : int RootElt = MaskVal;
4211 : Value *SourceOp = Op0;
4212 21538 : if (MaskVal >= InVecNumElts) {
4213 2910 : RootElt = MaskVal - InVecNumElts;
4214 : SourceOp = Op1;
4215 : }
4216 :
4217 : // If the source operand is a shuffle itself, look through it to find the
4218 : // matching root vector.
4219 : if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4220 665 : return foldIdentityShuffles(
4221 : DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4222 665 : SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4223 : }
4224 :
4225 : // TODO: Look through bitcasts? What if the bitcast changes the vector element
4226 : // size?
4227 :
4228 : // The source operand is not a shuffle. Initialize the root vector value for
4229 : // this shuffle if that has not been done yet.
4230 20873 : if (!RootVec)
4231 : RootVec = SourceOp;
4232 :
4233 : // Give up as soon as a source operand does not match the existing root value.
4234 20873 : if (RootVec != SourceOp)
4235 : return nullptr;
4236 :
4237 : // The element must be coming from the same lane in the source vector
4238 : // (although it may have crossed lanes in intermediate shuffles).
4239 18628 : if (RootElt != DestElt)
4240 5953 : return nullptr;
4241 :
4242 : return RootVec;
4243 : }
4244 :
4245 0 : static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4246 : Type *RetTy, const SimplifyQuery &Q,
4247 : unsigned MaxRecurse) {
4248 0 : if (isa<UndefValue>(Mask))
4249 0 : return UndefValue::get(RetTy);
4250 :
4251 0 : Type *InVecTy = Op0->getType();
4252 0 : unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4253 : unsigned InVecNumElts = InVecTy->getVectorNumElements();
4254 :
4255 : SmallVector<int, 32> Indices;
4256 0 : ShuffleVectorInst::getShuffleMask(Mask, Indices);
4257 : assert(MaskNumElts == Indices.size() &&
4258 : "Size of Indices not same as number of mask elements?");
4259 :
4260 : // Canonicalization: If mask does not select elements from an input vector,
4261 : // replace that input vector with undef.
4262 : bool MaskSelects0 = false, MaskSelects1 = false;
4263 0 : for (unsigned i = 0; i != MaskNumElts; ++i) {
4264 0 : if (Indices[i] == -1)
4265 0 : continue;
4266 0 : if ((unsigned)Indices[i] < InVecNumElts)
4267 : MaskSelects0 = true;
4268 : else
4269 : MaskSelects1 = true;
4270 : }
4271 0 : if (!MaskSelects0)
4272 0 : Op0 = UndefValue::get(InVecTy);
4273 0 : if (!MaskSelects1)
4274 0 : Op1 = UndefValue::get(InVecTy);
4275 :
4276 : auto *Op0Const = dyn_cast<Constant>(Op0);
4277 : auto *Op1Const = dyn_cast<Constant>(Op1);
4278 :
4279 : // If all operands are constant, constant fold the shuffle.
4280 0 : if (Op0Const && Op1Const)
4281 0 : return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4282 :
4283 : // Canonicalization: if only one input vector is constant, it shall be the
4284 : // second one.
4285 0 : if (Op0Const && !Op1Const) {
4286 : std::swap(Op0, Op1);
4287 : ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4288 : }
4289 :
4290 : // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4291 : // value type is same as the input vectors' type.
4292 : if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4293 0 : if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4294 0 : OpShuf->getMask()->getSplatValue())
4295 0 : return Op0;
4296 :
4297 : // Don't fold a shuffle with undef mask elements. This may get folded in a
4298 : // better way using demanded bits or other analysis.
4299 : // TODO: Should we allow this?
4300 0 : if (find(Indices, -1) != Indices.end())
4301 0 : return nullptr;
4302 :
4303 : // Check if every element of this shuffle can be mapped back to the
4304 : // corresponding element of a single root vector. If so, we don't need this
4305 : // shuffle. This handles simple identity shuffles as well as chains of
4306 : // shuffles that may widen/narrow and/or move elements across lanes and back.
4307 : Value *RootVec = nullptr;
4308 0 : for (unsigned i = 0; i != MaskNumElts; ++i) {
4309 : // Note that recursion is limited for each vector element, so if any element
4310 : // exceeds the limit, this will fail to simplify.
4311 : RootVec =
4312 0 : foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4313 :
4314 : // We can't replace a widening/narrowing shuffle with one of its operands.
4315 0 : if (!RootVec || RootVec->getType() != RetTy)
4316 0 : return nullptr;
4317 : }
4318 : return RootVec;
4319 : }
4320 :
4321 : /// Given operands for a ShuffleVectorInst, fold the result or return null.
4322 12836 : Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4323 : Type *RetTy, const SimplifyQuery &Q) {
4324 12836 : return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4325 : }
4326 :
4327 98 : static Constant *propagateNaN(Constant *In) {
4328 : // If the input is a vector with undef elements, just return a default NaN.
4329 98 : if (!In->isNaN())
4330 1 : return ConstantFP::getNaN(In->getType());
4331 :
4332 : // Propagate the existing NaN constant when possible.
4333 : // TODO: Should we quiet a signaling NaN?
4334 : return In;
4335 : }
4336 :
4337 49803 : static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4338 49803 : if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4339 224 : return ConstantFP::getNaN(Op0->getType());
4340 :
4341 49579 : if (match(Op0, m_NaN()))
4342 14 : return propagateNaN(cast<Constant>(Op0));
4343 49565 : if (match(Op1, m_NaN()))
4344 84 : return propagateNaN(cast<Constant>(Op1));
4345 :
4346 : return nullptr;
4347 : }
4348 :
4349 : /// Given operands for an FAdd, see if we can fold the result. If not, this
4350 : /// returns null.
4351 0 : static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4352 : const SimplifyQuery &Q, unsigned MaxRecurse) {
4353 0 : if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4354 0 : return C;
4355 :
4356 0 : if (Constant *C = simplifyFPBinop(Op0, Op1))
4357 0 : return C;
4358 :
4359 : // fadd X, -0 ==> X
4360 0 : if (match(Op1, m_NegZeroFP()))
4361 0 : return Op0;
4362 :
4363 : // fadd X, 0 ==> X, when we know X is not -0
4364 0 : if (match(Op1, m_PosZeroFP()) &&
4365 0 : (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4366 0 : return Op0;
4367 :
4368 : // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant)
4369 : // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN.
4370 : // Negative zeros are allowed because we always end up with positive zero:
4371 : // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4372 : // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0
4373 : // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0
4374 : // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0
4375 0 : if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4376 0 : match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4377 0 : return ConstantFP::getNullValue(Op0->getType());
4378 :
4379 : // (X - Y) + Y --> X
4380 : // Y + (X - Y) --> X
4381 : Value *X;
4382 0 : if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4383 0 : (match(Op0, m_FSub(m_Value(X), m_Specific(Op1))) ||
4384 0 : match(Op1, m_FSub(m_Value(X), m_Specific(Op0)))))
4385 0 : return X;
4386 :
4387 : return nullptr;
4388 : }
4389 :
4390 : /// Given operands for an FSub, see if we can fold the result. If not, this
4391 : /// returns null.
4392 0 : static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4393 : const SimplifyQuery &Q, unsigned MaxRecurse) {
4394 0 : if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4395 0 : return C;
4396 :
4397 0 : if (Constant *C = simplifyFPBinop(Op0, Op1))
4398 0 : return C;
4399 :
4400 : // fsub X, +0 ==> X
4401 0 : if (match(Op1, m_PosZeroFP()))
4402 0 : return Op0;
4403 :
4404 : // fsub X, -0 ==> X, when we know X is not -0
4405 0 : if (match(Op1, m_NegZeroFP()) &&
4406 0 : (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4407 0 : return Op0;
4408 :
4409 : // fsub -0.0, (fsub -0.0, X) ==> X
4410 : Value *X;
4411 0 : if (match(Op0, m_NegZeroFP()) &&
4412 0 : match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4413 0 : return X;
4414 :
4415 : // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4416 0 : if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4417 0 : match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4418 0 : return X;
4419 :
4420 : // fsub nnan x, x ==> 0.0
4421 0 : if (FMF.noNaNs() && Op0 == Op1)
4422 0 : return Constant::getNullValue(Op0->getType());
4423 :
4424 : // Y - (Y - X) --> X
4425 : // (X + Y) - Y --> X
4426 0 : if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4427 0 : (match(Op1, m_FSub(m_Specific(Op0), m_Value(X))) ||
4428 0 : match(Op0, m_c_FAdd(m_Specific(Op1), m_Value(X)))))
4429 0 : return X;
4430 :
4431 : return nullptr;
4432 : }
4433 :
4434 : /// Given the operands for an FMul, see if we can fold the result
4435 0 : static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4436 : const SimplifyQuery &Q, unsigned MaxRecurse) {
4437 0 : if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4438 0 : return C;
4439 :
4440 0 : if (Constant *C = simplifyFPBinop(Op0, Op1))
4441 0 : return C;
4442 :
4443 : // fmul X, 1.0 ==> X
4444 0 : if (match(Op1, m_FPOne()))
4445 0 : return Op0;
4446 :
4447 : // fmul nnan nsz X, 0 ==> 0
4448 0 : if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4449 0 : return ConstantFP::getNullValue(Op0->getType());
4450 :
4451 : // sqrt(X) * sqrt(X) --> X, if we can:
4452 : // 1. Remove the intermediate rounding (reassociate).
4453 : // 2. Ignore non-zero negative numbers because sqrt would produce NAN.
4454 : // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0.
4455 : Value *X;
4456 0 : if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4457 0 : FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4458 0 : return X;
4459 :
4460 : return nullptr;
4461 : }
4462 :
4463 15645 : Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4464 : const SimplifyQuery &Q) {
4465 15645 : return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4466 : }
4467 :
4468 :
4469 9307 : Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4470 : const SimplifyQuery &Q) {
4471 9307 : return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4472 : }
4473 :
4474 16771 : Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4475 : const SimplifyQuery &Q) {
4476 16771 : return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4477 : }
4478 :
4479 0 : static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4480 : const SimplifyQuery &Q, unsigned) {
4481 0 : if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4482 0 : return C;
4483 :
4484 0 : if (Constant *C = simplifyFPBinop(Op0, Op1))
4485 0 : return C;
4486 :
4487 : // X / 1.0 -> X
4488 0 : if (match(Op1, m_FPOne()))
4489 0 : return Op0;
4490 :
4491 : // 0 / X -> 0
4492 : // Requires that NaNs are off (X could be zero) and signed zeroes are
4493 : // ignored (X could be positive or negative, so the output sign is unknown).
4494 0 : if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4495 0 : return ConstantFP::getNullValue(Op0->getType());
4496 :
4497 0 : if (FMF.noNaNs()) {
4498 : // X / X -> 1.0 is legal when NaNs are ignored.
4499 : // We can ignore infinities because INF/INF is NaN.
4500 0 : if (Op0 == Op1)
4501 0 : return ConstantFP::get(Op0->getType(), 1.0);
4502 :
4503 : // (X * Y) / Y --> X if we can reassociate to the above form.
4504 : Value *X;
4505 0 : if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4506 0 : return X;
4507 :
4508 : // -X / X -> -1.0 and
4509 : // X / -X -> -1.0 are legal when NaNs are ignored.
4510 : // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4511 0 : if (match(Op0, m_FNegNSZ(m_Specific(Op1))) ||
4512 0 : match(Op1, m_FNegNSZ(m_Specific(Op0))))
4513 0 : return ConstantFP::get(Op0->getType(), -1.0);
4514 : }
4515 :
4516 : return nullptr;
4517 : }
4518 :
4519 5940 : Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4520 : const SimplifyQuery &Q) {
4521 5940 : return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4522 : }
4523 :
4524 0 : static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4525 : const SimplifyQuery &Q, unsigned) {
4526 0 : if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4527 0 : return C;
4528 :
4529 0 : if (Constant *C = simplifyFPBinop(Op0, Op1))
4530 0 : return C;
4531 :
4532 : // Unlike fdiv, the result of frem always matches the sign of the dividend.
4533 : // The constant match may include undef elements in a vector, so return a full
4534 : // zero constant as the result.
4535 0 : if (FMF.noNaNs()) {
4536 : // +0 % X -> 0
4537 0 : if (match(Op0, m_PosZeroFP()))
4538 0 : return ConstantFP::getNullValue(Op0->getType());
4539 : // -0 % X -> -0
4540 0 : if (match(Op0, m_NegZeroFP()))
4541 0 : return ConstantFP::getNegativeZero(Op0->getType());
4542 : }
4543 :
4544 : return nullptr;
4545 : }
4546 :
4547 146 : Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4548 : const SimplifyQuery &Q) {
4549 146 : return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4550 : }
4551 :
4552 : //=== Helper functions for higher up the class hierarchy.
4553 :
4554 : /// Given operands for a BinaryOperator, see if we can fold the result.
4555 : /// If not, this returns null.
4556 4178233 : static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4557 : const SimplifyQuery &Q, unsigned MaxRecurse) {
4558 4178233 : switch (Opcode) {
4559 3718130 : case Instruction::Add:
4560 3718130 : return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4561 74343 : case Instruction::Sub:
4562 74343 : return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4563 66715 : case Instruction::Mul:
4564 66715 : return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4565 4254 : case Instruction::SDiv:
4566 4254 : return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4567 4527 : case Instruction::UDiv:
4568 4527 : return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4569 135 : case Instruction::SRem:
4570 135 : return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4571 3708 : case Instruction::URem:
4572 3708 : return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4573 37959 : case Instruction::Shl:
4574 37959 : return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4575 16520 : case Instruction::LShr:
4576 16520 : return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4577 13577 : case Instruction::AShr:
4578 13577 : return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4579 84783 : case Instruction::And:
4580 84783 : return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4581 117212 : case Instruction::Or:
4582 117212 : return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4583 34199 : case Instruction::Xor:
4584 34199 : return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4585 527 : case Instruction::FAdd:
4586 527 : return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4587 14 : case Instruction::FSub:
4588 14 : return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4589 1621 : case Instruction::FMul:
4590 1621 : return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4591 9 : case Instruction::FDiv:
4592 9 : return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4593 0 : case Instruction::FRem:
4594 0 : return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4595 0 : default:
4596 0 : llvm_unreachable("Unexpected opcode");
4597 : }
4598 : }
4599 :
4600 : /// Given operands for a BinaryOperator, see if we can fold the result.
4601 : /// If not, this returns null.
4602 : /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4603 : /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4604 0 : static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4605 : const FastMathFlags &FMF, const SimplifyQuery &Q,
4606 : unsigned MaxRecurse) {
4607 0 : switch (Opcode) {
4608 0 : case Instruction::FAdd:
4609 0 : return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4610 0 : case Instruction::FSub:
4611 0 : return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4612 0 : case Instruction::FMul:
4613 0 : return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4614 0 : case Instruction::FDiv:
4615 0 : return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4616 0 : default:
4617 0 : return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4618 : }
4619 : }
4620 :
4621 3561250 : Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4622 : const SimplifyQuery &Q) {
4623 3561250 : return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4624 : }
4625 :
4626 398 : Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4627 : FastMathFlags FMF, const SimplifyQuery &Q) {
4628 398 : return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4629 : }
4630 :
4631 : /// Given operands for a CmpInst, see if we can fold the result.
4632 146149 : static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4633 : const SimplifyQuery &Q, unsigned MaxRecurse) {
4634 146149 : if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4635 145514 : return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4636 635 : return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4637 : }
4638 :
4639 17746 : Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4640 : const SimplifyQuery &Q) {
4641 17746 : return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4642 : }
4643 :
4644 : static bool IsIdempotent(Intrinsic::ID ID) {
4645 0 : switch (ID) {
4646 : default: return false;
4647 :
4648 : // Unary idempotent: f(f(x)) = f(x)
4649 : case Intrinsic::fabs:
4650 : case Intrinsic::floor:
4651 : case Intrinsic::ceil:
4652 : case Intrinsic::trunc:
4653 : case Intrinsic::rint:
4654 : case Intrinsic::nearbyint:
4655 : case Intrinsic::round:
4656 : case Intrinsic::canonicalize:
4657 : return true;
4658 : }
4659 : }
4660 :
4661 9 : static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4662 : const DataLayout &DL) {
4663 : GlobalValue *PtrSym;
4664 : APInt PtrOffset;
4665 9 : if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4666 : return nullptr;
4667 :
4668 8 : Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4669 8 : Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4670 8 : Type *Int32PtrTy = Int32Ty->getPointerTo();
4671 8 : Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4672 :
4673 : auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4674 8 : if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4675 : return nullptr;
4676 :
4677 8 : uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4678 8 : if (OffsetInt % 4 != 0)
4679 : return nullptr;
4680 :
4681 7 : Constant *C = ConstantExpr::getGetElementPtr(
4682 : Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4683 : ConstantInt::get(Int64Ty, OffsetInt / 4));
4684 7 : Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4685 7 : if (!Loaded)
4686 : return nullptr;
4687 :
4688 : auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4689 : if (!LoadedCE)
4690 : return nullptr;
4691 :
4692 7 : if (LoadedCE->getOpcode() == Instruction::Trunc) {
4693 : LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4694 : if (!LoadedCE)
4695 : return nullptr;
4696 : }
4697 :
4698 7 : if (LoadedCE->getOpcode() != Instruction::Sub)
4699 : return nullptr;
4700 :
4701 : auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4702 5 : if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4703 : return nullptr;
4704 : auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4705 :
4706 : Constant *LoadedRHS = LoadedCE->getOperand(1);
4707 : GlobalValue *LoadedRHSSym;
4708 : APInt LoadedRHSOffset;
4709 5 : if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4710 4 : DL) ||
4711 9 : PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4712 : return nullptr;
4713 :
4714 4 : return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4715 : }
4716 :
4717 746 : static bool maskIsAllZeroOrUndef(Value *Mask) {
4718 : auto *ConstMask = dyn_cast<Constant>(Mask);
4719 : if (!ConstMask)
4720 : return false;
4721 17 : if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4722 : return true;
4723 29 : for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4724 : ++I) {
4725 29 : if (auto *MaskElt = ConstMask->getAggregateElement(I))
4726 29 : if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4727 : continue;
4728 : return false;
4729 : }
4730 : return true;
4731 : }
4732 :
4733 0 : static Value *simplifyUnaryIntrinsic(Function *F, Value *Op0,
4734 : const SimplifyQuery &Q) {
4735 : // Idempotent functions return the same result when called repeatedly.
4736 0 : Intrinsic::ID IID = F->getIntrinsicID();
4737 : if (IsIdempotent(IID))
4738 : if (auto *II = dyn_cast<IntrinsicInst>(Op0))
4739 0 : if (II->getIntrinsicID() == IID)
4740 0 : return II;
4741 :
4742 : Value *X;
4743 0 : switch (IID) {
4744 0 : case Intrinsic::fabs:
4745 0 : if (SignBitMustBeZero(Op0, Q.TLI)) return Op0;
4746 : break;
4747 : case Intrinsic::bswap:
4748 : // bswap(bswap(x)) -> x
4749 0 : if (match(Op0, m_BSwap(m_Value(X)))) return X;
4750 : break;
4751 : case Intrinsic::bitreverse:
4752 : // bitreverse(bitreverse(x)) -> x
4753 0 : if (match(Op0, m_BitReverse(m_Value(X)))) return X;
4754 : break;
4755 0 : case Intrinsic::exp:
4756 : // exp(log(x)) -> x
4757 0 : if (Q.CxtI->hasAllowReassoc() &&
4758 0 : match(Op0, m_Intrinsic<Intrinsic::log>(m_Value(X)))) return X;
4759 : break;
4760 0 : case Intrinsic::exp2:
4761 : // exp2(log2(x)) -> x
4762 0 : if (Q.CxtI->hasAllowReassoc() &&
4763 0 : match(Op0, m_Intrinsic<Intrinsic::log2>(m_Value(X)))) return X;
4764 : break;
4765 0 : case Intrinsic::log:
4766 : // log(exp(x)) -> x
4767 0 : if (Q.CxtI->hasAllowReassoc() &&
4768 0 : match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X)))) return X;
4769 : break;
4770 0 : case Intrinsic::log2:
4771 : // log2(exp2(x)) -> x
4772 0 : if (Q.CxtI->hasAllowReassoc() &&
4773 0 : match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) return X;
4774 : break;
4775 : default:
4776 : break;
4777 : }
4778 :
4779 : return nullptr;
4780 : }
4781 :
4782 0 : static Value *simplifyBinaryIntrinsic(Function *F, Value *Op0, Value *Op1,
4783 : const SimplifyQuery &Q) {
4784 0 : Intrinsic::ID IID = F->getIntrinsicID();
4785 : Type *ReturnType = F->getReturnType();
4786 0 : switch (IID) {
4787 0 : case Intrinsic::usub_with_overflow:
4788 : case Intrinsic::ssub_with_overflow:
4789 : // X - X -> { 0, false }
4790 0 : if (Op0 == Op1)
4791 0 : return Constant::getNullValue(ReturnType);
4792 : // X - undef -> undef
4793 : // undef - X -> undef
4794 0 : if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4795 0 : return UndefValue::get(ReturnType);
4796 : break;
4797 0 : case Intrinsic::uadd_with_overflow:
4798 : case Intrinsic::sadd_with_overflow:
4799 : // X + undef -> undef
4800 0 : if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4801 0 : return UndefValue::get(ReturnType);
4802 : break;
4803 : case Intrinsic::umul_with_overflow:
4804 : case Intrinsic::smul_with_overflow:
4805 : // 0 * X -> { 0, false }
4806 : // X * 0 -> { 0, false }
4807 0 : if (match(Op0, m_Zero()) || match(Op1, m_Zero()))
4808 0 : return Constant::getNullValue(ReturnType);
4809 : // undef * X -> { 0, false }
4810 : // X * undef -> { 0, false }
4811 0 : if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
4812 0 : return Constant::getNullValue(ReturnType);
4813 : break;
4814 0 : case Intrinsic::load_relative:
4815 : if (auto *C0 = dyn_cast<Constant>(Op0))
4816 : if (auto *C1 = dyn_cast<Constant>(Op1))
4817 0 : return SimplifyRelativeLoad(C0, C1, Q.DL);
4818 : break;
4819 0 : case Intrinsic::powi:
4820 : if (auto *Power = dyn_cast<ConstantInt>(Op1)) {
4821 : // powi(x, 0) -> 1.0
4822 0 : if (Power->isZero())
4823 0 : return ConstantFP::get(Op0->getType(), 1.0);
4824 : // powi(x, 1) -> x
4825 0 : if (Power->isOne())
4826 0 : return Op0;
4827 : }
4828 : break;
4829 0 : case Intrinsic::maxnum:
4830 : case Intrinsic::minnum: {
4831 : // If the arguments are the same, this is a no-op.
4832 0 : if (Op0 == Op1) return Op0;
4833 :
4834 : // If one argument is NaN or undef, return the other argument.
4835 0 : if (match(Op0, m_CombineOr(m_NaN(), m_Undef()))) return Op1;
4836 0 : if (match(Op1, m_CombineOr(m_NaN(), m_Undef()))) return Op0;
4837 :
4838 : // Min/max of the same operation with common operand:
4839 : // m(m(X, Y)), X --> m(X, Y) (4 commuted variants)
4840 : if (auto *M0 = dyn_cast<IntrinsicInst>(Op0))
4841 0 : if (M0->getIntrinsicID() == IID &&
4842 0 : (M0->getOperand(0) == Op1 || M0->getOperand(1) == Op1))
4843 0 : return Op0;
4844 : if (auto *M1 = dyn_cast<IntrinsicInst>(Op1))
4845 0 : if (M1->getIntrinsicID() == IID &&
4846 0 : (M1->getOperand(0) == Op0 || M1->getOperand(1) == Op0))
4847 0 : return Op1;
4848 :
4849 : // minnum(X, -Inf) --> -Inf (and commuted variant)
4850 : // maxnum(X, +Inf) --> +Inf (and commuted variant)
4851 0 : bool UseNegInf = IID == Intrinsic::minnum;
4852 : const APFloat *C;
4853 0 : if ((match(Op0, m_APFloat(C)) && C->isInfinity() &&
4854 0 : C->isNegative() == UseNegInf) ||
4855 0 : (match(Op1, m_APFloat(C)) && C->isInfinity() &&
4856 : C->isNegative() == UseNegInf))
4857 0 : return ConstantFP::getInfinity(ReturnType, UseNegInf);
4858 :
4859 : // TODO: minnum(nnan x, inf) -> x
4860 : // TODO: minnum(nnan ninf x, flt_max) -> x
4861 : // TODO: maxnum(nnan x, -inf) -> x
4862 : // TODO: maxnum(nnan ninf x, -flt_max) -> x
4863 0 : break;
4864 : }
4865 : default:
4866 : break;
4867 : }
4868 :
4869 : return nullptr;
4870 : }
4871 :
4872 : template <typename IterTy>
4873 2438249 : static Value *simplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4874 : const SimplifyQuery &Q) {
4875 : // Intrinsics with no operands have some kind of side effect. Don't simplify.
4876 2438249 : unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4877 2438249 : if (NumOperands == 0)
4878 : return nullptr;
4879 :
4880 2292207 : Intrinsic::ID IID = F->getIntrinsicID();
4881 2292207 : if (NumOperands == 1)
4882 32915 : return simplifyUnaryIntrinsic(F, ArgBegin[0], Q);
4883 :
4884 2259292 : if (NumOperands == 2)
4885 1066394 : return simplifyBinaryIntrinsic(F, ArgBegin[0], ArgBegin[1], Q);
4886 :
4887 : // Handle intrinsics with 3 or more arguments.
4888 1192898 : switch (IID) {
4889 746 : case Intrinsic::masked_load: {
4890 746 : Value *MaskArg = ArgBegin[2];
4891 746 : Value *PassthruArg = ArgBegin[3];
4892 : // If the mask is all zeros or undef, the "passthru" argument is the result.
4893 746 : if (maskIsAllZeroOrUndef(MaskArg))
4894 4 : return PassthruArg;
4895 : return nullptr;
4896 : }
4897 48 : case Intrinsic::fshl:
4898 : case Intrinsic::fshr: {
4899 48 : Value *ShAmtArg = ArgBegin[2];
4900 : const APInt *ShAmtC;
4901 48 : if (match(ShAmtArg, m_APInt(ShAmtC))) {
4902 : // If there's effectively no shift, return the 1st arg or 2nd arg.
4903 : // TODO: For vectors, we could check each element of a non-splat constant.
4904 36 : APInt BitWidth = APInt(ShAmtC->getBitWidth(), ShAmtC->getBitWidth());
4905 72 : if (ShAmtC->urem(BitWidth).isNullValue())
4906 20 : return ArgBegin[IID == Intrinsic::fshl ? 0 : 1];
4907 : }
4908 : return nullptr;
4909 : }
4910 : default:
4911 : return nullptr;
4912 : }
4913 : }
4914 0 :
4915 : template <typename IterTy>
4916 : static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4917 0 : IterTy ArgEnd, const SimplifyQuery &Q,
4918 0 : unsigned MaxRecurse) {
4919 : Type *Ty = V->getType();
4920 : if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4921 0 : Ty = PTy->getElementType();
4922 0 : FunctionType *FTy = cast<FunctionType>(Ty);
4923 0 :
4924 : // call undef -> undef
4925 0 : // call null -> undef
4926 0 : if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4927 : return UndefValue::get(FTy->getReturnType());
4928 :
4929 0 : Function *F = dyn_cast<Function>(V);
4930 0 : if (!F)
4931 0 : return nullptr;
4932 0 :
4933 : if (F->isIntrinsic())
4934 0 : if (Value *Ret = simplifyIntrinsic(F, ArgBegin, ArgEnd, Q))
4935 0 : return Ret;
4936 :
4937 : if (!canConstantFoldCallTo(CS, F))
4938 0 : return nullptr;
4939 :
4940 0 : SmallVector<Constant *, 4> ConstantArgs;
4941 : ConstantArgs.reserve(ArgEnd - ArgBegin);
4942 0 : for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4943 : Constant *C = dyn_cast<Constant>(*I);
4944 : if (!C)
4945 0 : return nullptr;
4946 0 : ConstantArgs.push_back(C);
4947 0 : }
4948 :
4949 : return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4950 : }
4951 :
4952 : Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4953 : User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4954 : const SimplifyQuery &Q) {
4955 2438249 : return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4956 : }
4957 :
4958 2438249 : Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4959 2438249 : ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4960 : return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4961 : }
4962 2292207 :
4963 2292207 : Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4964 32915 : CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4965 : return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4966 2259292 : Q, RecursionLimit);
4967 1066394 : }
4968 :
4969 : /// See if we can compute a simplified version of this instruction.
4970 1192898 : /// If not, this returns null.
4971 746 :
4972 746 : Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4973 746 : OptimizationRemarkEmitter *ORE) {
4974 : const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4975 746 : Value *Result;
4976 4 :
4977 : switch (I->getOpcode()) {
4978 : default:
4979 48 : Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4980 : break;
4981 48 : case Instruction::FAdd:
4982 : Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4983 48 : I->getFastMathFlags(), Q);
4984 : break;
4985 : case Instruction::Add:
4986 36 : Result =
4987 72 : SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4988 20 : Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
4989 : Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
4990 : break;
4991 : case Instruction::FSub:
4992 : Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4993 : I->getFastMathFlags(), Q);
4994 : break;
4995 : case Instruction::Sub:
4996 : Result =
4997 : SimplifySubInst(I->getOperand(0), I->getOperand(1),
4998 0 : Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
4999 : Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
5000 : break;
5001 0 : case Instruction::FMul:
5002 : Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
5003 0 : I->getFastMathFlags(), Q);
5004 : break;
5005 : case Instruction::Mul:
5006 : Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
5007 : break;
5008 0 : case Instruction::SDiv:
5009 0 : Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
5010 : break;
5011 : case Instruction::UDiv:
5012 : Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
5013 0 : break;
5014 : case Instruction::FDiv:
5015 0 : Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
5016 0 : I->getFastMathFlags(), Q);
5017 0 : break;
5018 : case Instruction::SRem:
5019 0 : Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
5020 0 : break;
5021 : case Instruction::URem:
5022 : Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
5023 0 : break;
5024 0 : case Instruction::FRem:
5025 0 : Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
5026 0 : I->getFastMathFlags(), Q);
5027 0 : break;
5028 0 : case Instruction::Shl:
5029 : Result =
5030 : SimplifyShlInst(I->getOperand(0), I->getOperand(1),
5031 0 : Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
5032 : Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
5033 0 : break;
5034 : case Instruction::LShr:
5035 : Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
5036 0 : Q.IIQ.isExact(cast<BinaryOperator>(I)), Q);
5037 : break;
5038 0 : case Instruction::AShr:
5039 : Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
5040 : Q.IIQ.isExact(cast<BinaryOperator>(I)), Q);
5041 : break;
5042 : case Instruction::And:
5043 0 : Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
5044 0 : break;
5045 : case Instruction::Or:
5046 : Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
5047 : break;
5048 0 : case Instruction::Xor:
5049 : Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
5050 0 : break;
5051 0 : case Instruction::ICmp:
5052 0 : Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
5053 : I->getOperand(0), I->getOperand(1), Q);
5054 0 : break;
5055 0 : case Instruction::FCmp:
5056 : Result =
5057 : SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
5058 0 : I->getOperand(1), I->getFastMathFlags(), Q);
5059 0 : break;
5060 0 : case Instruction::Select:
5061 0 : Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
5062 0 : I->getOperand(2), Q);
5063 0 : break;
5064 : case Instruction::GetElementPtr: {
5065 : SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
5066 0 : Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
5067 : Ops, Q);
5068 0 : break;
5069 : }
5070 : case Instruction::InsertValue: {
5071 0 : InsertValueInst *IV = cast<InsertValueInst>(I);
5072 : Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
5073 0 : IV->getInsertedValueOperand(),
5074 : IV->getIndices(), Q);
5075 : break;
5076 : }
5077 : case Instruction::InsertElement: {
5078 0 : auto *IE = cast<InsertElementInst>(I);
5079 0 : Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
5080 : IE->getOperand(2), Q);
5081 : break;
5082 : }
5083 0 : case Instruction::ExtractValue: {
5084 : auto *EVI = cast<ExtractValueInst>(I);
5085 0 : Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
5086 0 : EVI->getIndices(), Q);
5087 0 : break;
5088 : }
5089 0 : case Instruction::ExtractElement: {
5090 0 : auto *EEI = cast<ExtractElementInst>(I);
5091 : Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
5092 : EEI->getIndexOperand(), Q);
5093 0 : break;
5094 0 : }
5095 0 : case Instruction::ShuffleVector: {
5096 0 : auto *SVI = cast<ShuffleVectorInst>(I);
5097 0 : Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5098 0 : SVI->getMask(), SVI->getType(), Q);
5099 : break;
5100 : }
5101 0 : case Instruction::PHI:
5102 : Result = SimplifyPHINode(cast<PHINode>(I), Q);
5103 : break;
5104 0 : case Instruction::Call: {
5105 : CallSite CS(cast<CallInst>(I));
5106 : Result = SimplifyCall(CS, Q);
5107 0 : break;
5108 : }
5109 : #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
5110 0 : #include "llvm/IR/Instruction.def"
5111 : #undef HANDLE_CAST_INST
5112 0 : Result =
5113 : SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
5114 : break;
5115 5220588 : case Instruction::Alloca:
5116 : // No simplifications for Alloca and it can't be constant folded.
5117 10441176 : Result = nullptr;
5118 5220588 : break;
5119 : }
5120 :
5121 : // In general, it is possible for computeKnownBits to determine all bits in a
5122 : // value even when the operands are not all constants.
5123 : if (!Result && I->getType()->isIntOrIntVectorTy()) {
5124 65856975 : KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
5125 : if (Known.isConstant())
5126 65856975 : Result = ConstantInt::get(I->getType(), Known.getConstant());
5127 : }
5128 :
5129 65856975 : /// If called on unreachable code, the above logic may report that the
5130 27584677 : /// instruction simplified to itself. Make life easier for users by
5131 27584677 : /// detecting that case here, returning a safe value instead.
5132 27584677 : return Result == I ? UndefValue::get(I->getType()) : Result;
5133 11332 : }
5134 22664 :
5135 : /// Implementation of recursive simplification through an instruction's
5136 11332 : /// uses.
5137 : ///
5138 : /// This is the common implementation of the recursive simplification routines.
5139 4781666 : /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
5140 2390833 : /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
5141 2390833 : /// instructions to process and attempt to simplify it using
5142 2390833 : /// InstructionSimplify.
5143 6815 : ///
5144 13630 : /// This routine returns 'true' only when *it* simplifies something. The passed
5145 : /// in simplified value does not count toward this.
5146 6815 : static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
5147 : const TargetLibraryInfo *TLI,
5148 : const DominatorTree *DT,
5149 296974 : AssumptionCache *AC) {
5150 148487 : bool Simplified = false;
5151 148487 : SmallSetVector<Instruction *, 8> Worklist;
5152 148487 : const DataLayout &DL = I->getModule()->getDataLayout();
5153 11859 :
5154 23718 : // If we have an explicit value to collapse to, do that round of the
5155 : // simplification loop by hand initially.
5156 11859 : if (SimpleV) {
5157 34816 : for (User *U : I->users())
5158 69632 : if (U != I)
5159 34816 : Worklist.insert(cast<Instruction>(U));
5160 66415 :
5161 132830 : // Replace the instruction with its simplified value.
5162 66415 : I->replaceAllUsesWith(SimpleV);
5163 8721 :
5164 17442 : // Gracefully handle edge cases where the instruction is not wired into any
5165 8721 : // parent block.
5166 4429 : if (I->getParent() && !I->isEHPad() && !I->isTerminator() &&
5167 8858 : !I->mayHaveSideEffects())
5168 : I->eraseFromParent();
5169 4429 : } else {
5170 2701 : Worklist.insert(I);
5171 5402 : }
5172 2701 :
5173 18370 : // Note that we must test the size on each iteration, the worklist can grow.
5174 36740 : for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
5175 18370 : I = Worklist[Idx];
5176 96 :
5177 192 : // See if this instruction simplifies.
5178 : SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
5179 96 : if (!SimpleV)
5180 : continue;
5181 :
5182 54700 : Simplified = true;
5183 27350 :
5184 27350 : // Stash away all the uses of the old instruction so we can check them for
5185 27350 : // recursive simplifications after a RAUW. This is cheaper than checking all
5186 : // uses of To on the recursive step in most cases.
5187 35138 : for (User *U : I->users())
5188 17569 : Worklist.insert(cast<Instruction>(U));
5189 17569 :
5190 : // Replace the instruction with its simplified value.
5191 70588 : I->replaceAllUsesWith(SimpleV);
5192 35294 :
5193 35294 : // Gracefully handle edge cases where the instruction is not wired into any
5194 144544 : // parent block.
5195 289088 : if (I->getParent() && !I->isEHPad() && !I->isTerminator() &&
5196 144544 : !I->mayHaveSideEffects())
5197 29596 : I->eraseFromParent();
5198 59192 : }
5199 29596 : return Simplified;
5200 50298 : }
5201 100596 :
5202 50298 : bool llvm::recursivelySimplifyInstruction(Instruction *I,
5203 689208 : const TargetLibraryInfo *TLI,
5204 1378416 : const DominatorTree *DT,
5205 : AssumptionCache *AC) {
5206 689208 : return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
5207 10957 : }
5208 :
5209 21914 : bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
5210 : const TargetLibraryInfo *TLI,
5211 10957 : const DominatorTree *DT,
5212 89353 : AssumptionCache *AC) {
5213 178706 : assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5214 : assert(SimpleV && "Must provide a simplified value.");
5215 89353 : return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
5216 4941799 : }
5217 4941799 :
5218 4941799 : namespace llvm {
5219 : const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
5220 : auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
5221 : auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
5222 : auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
5223 : auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
5224 79973 : auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5225 : auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5226 : return {F.getParent()->getDataLayout(), TLI, DT, AC};
5227 79973 : }
5228 :
5229 : const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5230 : const DataLayout &DL) {
5231 20282 : return {DL, &AR.TLI, &AR.DT, &AR.AC};
5232 : }
5233 20282 :
5234 : template <class T, class... TArgs>
5235 : const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5236 : Function &F) {
5237 472784 : auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5238 : auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5239 472784 : auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5240 : return {F.getParent()->getDataLayout(), TLI, DT, AC};
5241 : }
5242 : template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,
5243 19010 : Function &);
5244 : }
|
