[llvm-exegesis] Fix missing std::move.
[llvm-complete.git] / lib / Analysis / InstructionSimplify.cpp
blob86f5652f830648bc6e10e2700ddc9a3ffc2f357e
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).
18 //===----------------------------------------------------------------------===//
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;
45 #define DEBUG_TYPE "instsimplify"
47 enum { RecursionLimit = 3 };
49 STATISTIC(NumExpand, "Number of expansions");
50 STATISTIC(NumReassoc, "Number of reassociations");
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);
68 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;
76 CmpInst::Predicate ExpectedPred, Pred1, Pred2;
77 if (BinOpCode == BinaryOperator::Or) {
78 ExpectedPred = ICmpInst::ICMP_NE;
79 } else if (BinOpCode == BinaryOperator::And) {
80 ExpectedPred = ICmpInst::ICMP_EQ;
81 } else
82 return nullptr;
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
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 if (!match(Cond, m_c_BinOp(m_c_ICmp(Pred1, m_Specific(TrueVal),
99 m_Specific(FalseVal)),
100 m_ICmp(Pred2, m_Value(X), m_Value(Y)))) ||
101 Pred1 != Pred2 || Pred1 != ExpectedPred)
102 return nullptr;
104 if (X == TrueVal || X == FalseVal || Y == TrueVal || Y == FalseVal)
105 return BinOpCode == BinaryOperator::Or ? TrueVal : FalseVal;
107 return nullptr;
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 return ConstantInt::getFalse(Ty);
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 return ConstantInt::getTrue(Ty);
122 /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
123 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 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
131 return true;
132 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
133 CRHS == LHS;
136 /// Does the given value dominate the specified phi node?
137 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;
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 if (!I->getParent() || !P->getParent() || !I->getFunction())
147 return false;
149 // If we have a DominatorTree then do a precise test.
150 if (DT)
151 return DT->dominates(I, P);
153 // Otherwise, if the instruction is in the entry block and is not an invoke,
154 // then it obviously dominates all phi nodes.
155 if (I->getParent() == &I->getFunction()->getEntryBlock() &&
156 !isa<InvokeInst>(I))
157 return true;
159 return false;
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 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 if (!MaxRecurse--)
172 return nullptr;
174 // Check whether the expression has the form "(A op' B) op C".
175 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
176 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 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
181 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 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
185 && L == B && R == A)) {
186 ++NumExpand;
187 return LHS;
189 // Otherwise return "L op' R" if it simplifies.
190 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
191 ++NumExpand;
192 return V;
197 // Check whether the expression has the form "A op (B op' C)".
198 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
199 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 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
204 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 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
208 && L == C && R == B)) {
209 ++NumExpand;
210 return RHS;
212 // Otherwise return "L op' R" if it simplifies.
213 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
214 ++NumExpand;
215 return V;
220 return nullptr;
223 /// Generic simplifications for associative binary operations.
224 /// Returns the simpler value, or null if none was found.
225 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!");
231 // Recursion is always used, so bail out at once if we already hit the limit.
232 if (!MaxRecurse--)
233 return nullptr;
235 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
236 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
238 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
239 if (Op0 && Op0->getOpcode() == Opcode) {
240 Value *A = Op0->getOperand(0);
241 Value *B = Op0->getOperand(1);
242 Value *C = RHS;
244 // Does "B op C" simplify?
245 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 if (V == B) return LHS;
249 // Otherwise return "A op V" if it simplifies.
250 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
251 ++NumReassoc;
252 return W;
257 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
258 if (Op1 && Op1->getOpcode() == Opcode) {
259 Value *A = LHS;
260 Value *B = Op1->getOperand(0);
261 Value *C = Op1->getOperand(1);
263 // Does "A op B" simplify?
264 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 if (V == B) return RHS;
268 // Otherwise return "V op C" if it simplifies.
269 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
270 ++NumReassoc;
271 return W;
276 // The remaining transforms require commutativity as well as associativity.
277 if (!Instruction::isCommutative(Opcode))
278 return nullptr;
280 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
281 if (Op0 && Op0->getOpcode() == Opcode) {
282 Value *A = Op0->getOperand(0);
283 Value *B = Op0->getOperand(1);
284 Value *C = RHS;
286 // Does "C op A" simplify?
287 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 if (V == A) return LHS;
291 // Otherwise return "V op B" if it simplifies.
292 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
293 ++NumReassoc;
294 return W;
299 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
300 if (Op1 && Op1->getOpcode() == Opcode) {
301 Value *A = LHS;
302 Value *B = Op1->getOperand(0);
303 Value *C = Op1->getOperand(1);
305 // Does "C op A" simplify?
306 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 if (V == C) return RHS;
310 // Otherwise return "B op V" if it simplifies.
311 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
312 ++NumReassoc;
313 return W;
318 return nullptr;
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 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 if (!MaxRecurse--)
330 return nullptr;
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);
340 // Evaluate the BinOp on the true and false branches of the select.
341 Value *TV;
342 Value *FV;
343 if (SI == LHS) {
344 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
345 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
346 } else {
347 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
348 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
351 // If they simplified to the same value, then return the common value.
352 // If they both failed to simplify then return null.
353 if (TV == FV)
354 return TV;
356 // If one branch simplified to undef, return the other one.
357 if (TV && isa<UndefValue>(TV))
358 return FV;
359 if (FV && isa<UndefValue>(FV))
360 return TV;
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 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
365 return SI;
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 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 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
374 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 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
379 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
380 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
381 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
382 Simplified->getOperand(1) == UnsimplifiedRHS)
383 return Simplified;
384 if (Simplified->isCommutative() &&
385 Simplified->getOperand(1) == UnsimplifiedLHS &&
386 Simplified->getOperand(0) == UnsimplifiedRHS)
387 return Simplified;
391 return nullptr;
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 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 if (!MaxRecurse--)
402 return nullptr;
404 // Make sure the select is on the LHS.
405 if (!isa<SelectInst>(LHS)) {
406 std::swap(LHS, RHS);
407 Pred = CmpInst::getSwappedPredicate(Pred);
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();
415 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
416 // Does "cmp TV, RHS" simplify?
417 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
418 if (TCmp == Cond) {
419 // It not only simplified, it simplified to the select condition. Replace
420 // it with 'true'.
421 TCmp = getTrue(Cond->getType());
422 } 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 if (!isSameCompare(Cond, Pred, TV, RHS))
426 return nullptr;
427 TCmp = getTrue(Cond->getType());
430 // Does "cmp FV, RHS" simplify?
431 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
432 if (FCmp == Cond) {
433 // It not only simplified, it simplified to the select condition. Replace
434 // it with 'false'.
435 FCmp = getFalse(Cond->getType());
436 } 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 if (!isSameCompare(Cond, Pred, FV, RHS))
440 return nullptr;
441 FCmp = getFalse(Cond->getType());
444 // If both sides simplified to the same value, then use it as the result of
445 // the original comparison.
446 if (TCmp == FCmp)
447 return TCmp;
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 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 if (match(FCmp, m_Zero()))
457 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 if (match(TCmp, m_One()))
462 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 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
467 if (Value *V =
468 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
469 Q, MaxRecurse))
470 return V;
472 return nullptr;
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 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 if (!MaxRecurse--)
484 return nullptr;
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 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 if (!valueDominatesPHI(LHS, PI, Q.DT))
497 return nullptr;
500 // Evaluate the BinOp on the incoming phi values.
501 Value *CommonValue = nullptr;
502 for (Value *Incoming : PI->incoming_values()) {
503 // If the incoming value is the phi node itself, it can safely be skipped.
504 if (Incoming == PI) continue;
505 Value *V = PI == LHS ?
506 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
507 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 if (!V || (CommonValue && V != CommonValue))
511 return nullptr;
512 CommonValue = V;
515 return CommonValue;
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 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 if (!MaxRecurse--)
526 return nullptr;
528 // Make sure the phi is on the LHS.
529 if (!isa<PHINode>(LHS)) {
530 std::swap(LHS, RHS);
531 Pred = CmpInst::getSwappedPredicate(Pred);
533 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!");
534 PHINode *PI = cast<PHINode>(LHS);
536 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
537 if (!valueDominatesPHI(RHS, PI, Q.DT))
538 return nullptr;
540 // Evaluate the BinOp on the incoming phi values.
541 Value *CommonValue = nullptr;
542 for (Value *Incoming : PI->incoming_values()) {
543 // If the incoming value is the phi node itself, it can safely be skipped.
544 if (Incoming == PI) continue;
545 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 if (!V || (CommonValue && V != CommonValue))
549 return nullptr;
550 CommonValue = V;
553 return CommonValue;
556 static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
557 Value *&Op0, Value *&Op1,
558 const SimplifyQuery &Q) {
559 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
560 if (auto *CRHS = dyn_cast<Constant>(Op1))
561 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
563 // Canonicalize the constant to the RHS if this is a commutative operation.
564 if (Instruction::isCommutative(Opcode))
565 std::swap(Op0, Op1);
567 return nullptr;
570 /// Given operands for an Add, see if we can fold the result.
571 /// If not, this returns null.
572 static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
573 const SimplifyQuery &Q, unsigned MaxRecurse) {
574 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
575 return C;
577 // X + undef -> undef
578 if (match(Op1, m_Undef()))
579 return Op1;
581 // X + 0 -> X
582 if (match(Op1, m_Zero()))
583 return Op0;
585 // If two operands are negative, return 0.
586 if (isKnownNegation(Op0, Op1))
587 return Constant::getNullValue(Op0->getType());
589 // X + (Y - X) -> Y
590 // (Y - X) + X -> Y
591 // Eg: X + -X -> 0
592 Value *Y = nullptr;
593 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
594 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
595 return Y;
597 // X + ~X -> -1 since ~X = -X-1
598 Type *Ty = Op0->getType();
599 if (match(Op0, m_Not(m_Specific(Op1))) ||
600 match(Op1, m_Not(m_Specific(Op0))))
601 return Constant::getAllOnesValue(Ty);
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 if ((IsNSW || IsNUW) && match(Op1, m_SignMask()) &&
607 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
608 return Y;
610 // add nuw %x, -1 -> -1, because %x can only be 0.
611 if (IsNUW && match(Op1, m_AllOnes()))
612 return Op1; // Which is -1.
614 /// i1 add -> xor.
615 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
616 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
617 return V;
619 // Try some generic simplifications for associative operations.
620 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
621 MaxRecurse))
622 return V;
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.
633 return nullptr;
636 Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW,
637 const SimplifyQuery &Query) {
638 return ::SimplifyAddInst(Op0, Op1, IsNSW, IsNUW, Query, RecursionLimit);
641 /// Compute the base pointer and cumulative constant offsets for V.
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.
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 static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
652 bool AllowNonInbounds = false) {
653 assert(V->getType()->isPtrOrPtrVectorTy());
655 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
656 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
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 Visited.insert(V);
662 do {
663 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
664 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
665 !GEP->accumulateConstantOffset(DL, Offset))
666 break;
667 V = GEP->getPointerOperand();
668 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
669 V = cast<Operator>(V)->getOperand(0);
670 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
671 if (GA->isInterposable())
672 break;
673 V = GA->getAliasee();
674 } else {
675 if (auto CS = CallSite(V))
676 if (Value *RV = CS.getReturnedArgOperand()) {
677 V = RV;
678 continue;
680 break;
682 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!");
683 } while (Visited.insert(V).second);
685 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
686 if (V->getType()->isVectorTy())
687 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
688 OffsetIntPtr);
689 return OffsetIntPtr;
692 /// Compute the constant difference between two pointer values.
693 /// If the difference is not a constant, returns zero.
694 static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
695 Value *RHS) {
696 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
697 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
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 if (LHS != RHS)
702 return nullptr;
704 // Otherwise, the difference of LHS - RHS can be computed as:
705 // LHS - RHS
706 // = (LHSOffset + Base) - (RHSOffset + Base)
707 // = LHSOffset - RHSOffset
708 return ConstantExpr::getSub(LHSOffset, RHSOffset);
711 /// Given operands for a Sub, see if we can fold the result.
712 /// If not, this returns null.
713 static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
714 const SimplifyQuery &Q, unsigned MaxRecurse) {
715 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
716 return C;
718 // X - undef -> undef
719 // undef - X -> undef
720 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
721 return UndefValue::get(Op0->getType());
723 // X - 0 -> X
724 if (match(Op1, m_Zero()))
725 return Op0;
727 // X - X -> 0
728 if (Op0 == Op1)
729 return Constant::getNullValue(Op0->getType());
731 // Is this a negation?
732 if (match(Op0, m_Zero())) {
733 // 0 - X -> 0 if the sub is NUW.
734 if (isNUW)
735 return Constant::getNullValue(Op0->getType());
737 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
738 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 if (isNSW)
742 return Constant::getNullValue(Op0->getType());
744 // 0 - X -> X if X is 0 or the minimum signed value.
745 return Op1;
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 Value *X = nullptr, *Y = nullptr, *Z = Op1;
752 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
753 // See if "V === Y - Z" simplifies.
754 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
755 // It does! Now see if "X + V" simplifies.
756 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
757 // It does, we successfully reassociated!
758 ++NumReassoc;
759 return W;
761 // See if "V === X - Z" simplifies.
762 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
763 // It does! Now see if "Y + V" simplifies.
764 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
765 // It does, we successfully reassociated!
766 ++NumReassoc;
767 return W;
771 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
772 // For example, X - (X + 1) -> -1
773 X = Op0;
774 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
775 // See if "V === X - Y" simplifies.
776 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
777 // It does! Now see if "V - Z" simplifies.
778 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
779 // It does, we successfully reassociated!
780 ++NumReassoc;
781 return W;
783 // See if "V === X - Z" simplifies.
784 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
785 // It does! Now see if "V - Y" simplifies.
786 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
787 // It does, we successfully reassociated!
788 ++NumReassoc;
789 return W;
793 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
794 // For example, X - (X - Y) -> Y.
795 Z = Op0;
796 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
797 // See if "V === Z - X" simplifies.
798 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
799 // It does! Now see if "V + Y" simplifies.
800 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
801 // It does, we successfully reassociated!
802 ++NumReassoc;
803 return W;
806 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
807 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
808 match(Op1, m_Trunc(m_Value(Y))))
809 if (X->getType() == Y->getType())
810 // See if "V === X - Y" simplifies.
811 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
812 // It does! Now see if "trunc V" simplifies.
813 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
814 Q, MaxRecurse - 1))
815 // It does, return the simplified "trunc V".
816 return W;
818 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
819 if (match(Op0, m_PtrToInt(m_Value(X))) &&
820 match(Op1, m_PtrToInt(m_Value(Y))))
821 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
822 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
824 // i1 sub -> xor.
825 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
826 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
827 return V;
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.
838 return nullptr;
841 Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
842 const SimplifyQuery &Q) {
843 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
846 /// Given operands for a Mul, see if we can fold the result.
847 /// If not, this returns null.
848 static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
849 unsigned MaxRecurse) {
850 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
851 return C;
853 // X * undef -> 0
854 // X * 0 -> 0
855 if (match(Op1, m_CombineOr(m_Undef(), m_Zero())))
856 return Constant::getNullValue(Op0->getType());
858 // X * 1 -> X
859 if (match(Op1, m_One()))
860 return Op0;
862 // (X / Y) * Y -> X if the division is exact.
863 Value *X = nullptr;
864 if (Q.IIQ.UseInstrInfo &&
865 (match(Op0,
866 m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
867 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0)))))) // Y * (X / Y)
868 return X;
870 // i1 mul -> and.
871 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
872 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
873 return V;
875 // Try some generic simplifications for associative operations.
876 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
877 MaxRecurse))
878 return V;
880 // Mul distributes over Add. Try some generic simplifications based on this.
881 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
882 Q, MaxRecurse))
883 return V;
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 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
888 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
889 MaxRecurse))
890 return V;
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 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
895 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
896 MaxRecurse))
897 return V;
899 return nullptr;
902 Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
903 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
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 static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
909 Type *Ty = Op0->getType();
911 // X / undef -> undef
912 // X % undef -> undef
913 if (match(Op1, m_Undef()))
914 return Op1;
916 // X / 0 -> undef
917 // X % 0 -> undef
918 // We don't need to preserve faults!
919 if (match(Op1, m_Zero()))
920 return UndefValue::get(Ty);
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 if (Op1C && Ty->isVectorTy()) {
926 unsigned NumElts = Ty->getVectorNumElements();
927 for (unsigned i = 0; i != NumElts; ++i) {
928 Constant *Elt = Op1C->getAggregateElement(i);
929 if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
930 return UndefValue::get(Ty);
934 // undef / X -> 0
935 // undef % X -> 0
936 if (match(Op0, m_Undef()))
937 return Constant::getNullValue(Ty);
939 // 0 / X -> 0
940 // 0 % X -> 0
941 if (match(Op0, m_Zero()))
942 return Constant::getNullValue(Op0->getType());
944 // X / X -> 1
945 // X % X -> 0
946 if (Op0 == Op1)
947 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
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 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1) ||
956 (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
957 return IsDiv ? Op0 : Constant::getNullValue(Ty);
959 return nullptr;
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 static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
966 const SimplifyQuery &Q, unsigned MaxRecurse) {
967 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
968 Constant *C = dyn_cast_or_null<Constant>(V);
969 return (C && C->isAllOnesValue());
972 /// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
973 /// to simplify X % Y to X.
974 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 if (!MaxRecurse--)
978 return false;
980 if (IsSigned) {
981 // |X| / |Y| --> 0
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.
986 // Make sure that a constant is not the minimum signed value because taking
987 // the abs() of that is undefined.
988 Type *Ty = X->getType();
989 const APInt *C;
990 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 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
995 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
996 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
997 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
998 return true;
1000 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 if (C->isMinSignedValue())
1005 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
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 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
1011 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
1012 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
1013 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
1014 return true;
1016 return false;
1019 // IsSigned == false.
1020 // Is the dividend unsigned less than the divisor?
1021 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
1024 /// These are simplifications common to SDiv and UDiv.
1025 static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1026 const SimplifyQuery &Q, unsigned MaxRecurse) {
1027 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1028 return C;
1030 if (Value *V = simplifyDivRem(Op0, Op1, true))
1031 return V;
1033 bool IsSigned = Opcode == Instruction::SDiv;
1035 // (X * Y) / Y -> X if the multiplication does not overflow.
1036 Value *X;
1037 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
1038 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1039 // If the Mul does not overflow, then we are good to go.
1040 if ((IsSigned && Q.IIQ.hasNoSignedWrap(Mul)) ||
1041 (!IsSigned && Q.IIQ.hasNoUnsignedWrap(Mul)))
1042 return X;
1043 // If X has the form X = A / Y, then X * Y cannot overflow.
1044 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
1045 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
1046 return X;
1049 // (X rem Y) / Y -> 0
1050 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1051 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1052 return Constant::getNullValue(Op0->getType());
1054 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1055 ConstantInt *C1, *C2;
1056 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1057 match(Op1, m_ConstantInt(C2))) {
1058 bool Overflow;
1059 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1060 if (Overflow)
1061 return Constant::getNullValue(Op0->getType());
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 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1067 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1068 return V;
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 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1073 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1074 return V;
1076 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1077 return Constant::getNullValue(Op0->getType());
1079 return nullptr;
1082 /// These are simplifications common to SRem and URem.
1083 static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1084 const SimplifyQuery &Q, unsigned MaxRecurse) {
1085 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1086 return C;
1088 if (Value *V = simplifyDivRem(Op0, Op1, false))
1089 return V;
1091 // (X % Y) % Y -> X % Y
1092 if ((Opcode == Instruction::SRem &&
1093 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1094 (Opcode == Instruction::URem &&
1095 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1096 return Op0;
1098 // (X << Y) % X -> 0
1099 if (Q.IIQ.UseInstrInfo &&
1100 ((Opcode == Instruction::SRem &&
1101 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1102 (Opcode == Instruction::URem &&
1103 match(Op0, m_NUWShl(m_Specific(Op1), m_Value())))))
1104 return Constant::getNullValue(Op0->getType());
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 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1109 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1110 return V;
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 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1115 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1116 return V;
1118 // If X / Y == 0, then X % Y == X.
1119 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1120 return Op0;
1122 return nullptr;
1125 /// Given operands for an SDiv, see if we can fold the result.
1126 /// If not, this returns null.
1127 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 if (isKnownNegation(Op0, Op1, /*NeedNSW=*/true))
1131 return Constant::getAllOnesValue(Op0->getType());
1133 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1136 Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1137 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1140 /// Given operands for a UDiv, see if we can fold the result.
1141 /// If not, this returns null.
1142 static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1143 unsigned MaxRecurse) {
1144 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1147 Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1148 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1151 /// Given operands for an SRem, see if we can fold the result.
1152 /// If not, this returns null.
1153 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 if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
1159 return ConstantInt::getNullValue(Op0->getType());
1161 // If the two operands are negated, return 0.
1162 if (isKnownNegation(Op0, Op1))
1163 return ConstantInt::getNullValue(Op0->getType());
1165 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1168 Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1169 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1172 /// Given operands for a URem, see if we can fold the result.
1173 /// If not, this returns null.
1174 static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1175 unsigned MaxRecurse) {
1176 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1179 Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1180 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1183 /// Returns true if a shift by \c Amount always yields undef.
1184 static bool isUndefShift(Value *Amount) {
1185 Constant *C = dyn_cast<Constant>(Amount);
1186 if (!C)
1187 return false;
1189 // X shift by undef -> undef because it may shift by the bitwidth.
1190 if (isa<UndefValue>(C))
1191 return true;
1193 // Shifting by the bitwidth or more is undefined.
1194 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1195 if (CI->getValue().getLimitedValue() >=
1196 CI->getType()->getScalarSizeInBits())
1197 return true;
1199 // If all lanes of a vector shift are undefined the whole shift is.
1200 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1201 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1202 if (!isUndefShift(C->getAggregateElement(I)))
1203 return false;
1204 return true;
1207 return false;
1210 /// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1211 /// If not, this returns null.
1212 static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1213 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1214 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1215 return C;
1217 // 0 shift by X -> 0
1218 if (match(Op0, m_Zero()))
1219 return Constant::getNullValue(Op0->getType());
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 if (match(Op1, m_Zero()) ||
1226 (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1227 return Op0;
1229 // Fold undefined shifts.
1230 if (isUndefShift(Op1))
1231 return UndefValue::get(Op0->getType());
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 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1236 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1237 return V;
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 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1242 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1243 return V;
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 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1248 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1249 return UndefValue::get(Op0->getType());
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 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1255 return Op0;
1257 return nullptr;
1260 /// Given operands for an Shl, LShr or AShr, see if we can
1261 /// fold the result. If not, this returns null.
1262 static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1263 Value *Op1, bool isExact, const SimplifyQuery &Q,
1264 unsigned MaxRecurse) {
1265 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1266 return V;
1268 // X >> X -> 0
1269 if (Op0 == Op1)
1270 return Constant::getNullValue(Op0->getType());
1272 // undef >> X -> 0
1273 // undef >> X -> undef (if it's exact)
1274 if (match(Op0, m_Undef()))
1275 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1277 // The low bit cannot be shifted out of an exact shift if it is set.
1278 if (isExact) {
1279 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1280 if (Op0Known.One[0])
1281 return Op0;
1284 return nullptr;
1287 /// Given operands for an Shl, see if we can fold the result.
1288 /// If not, this returns null.
1289 static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1290 const SimplifyQuery &Q, unsigned MaxRecurse) {
1291 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1292 return V;
1294 // undef << X -> 0
1295 // undef << X -> undef if (if it's NSW/NUW)
1296 if (match(Op0, m_Undef()))
1297 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1299 // (X >> A) << A -> X
1300 Value *X;
1301 if (Q.IIQ.UseInstrInfo &&
1302 match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1303 return X;
1305 // shl nuw i8 C, %x -> C iff C has sign bit set.
1306 if (isNUW && match(Op0, m_Negative()))
1307 return Op0;
1308 // NOTE: could use computeKnownBits() / LazyValueInfo,
1309 // but the cost-benefit analysis suggests it isn't worth it.
1311 return nullptr;
1314 Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1315 const SimplifyQuery &Q) {
1316 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1319 /// Given operands for an LShr, see if we can fold the result.
1320 /// If not, this returns null.
1321 static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1322 const SimplifyQuery &Q, unsigned MaxRecurse) {
1323 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1324 MaxRecurse))
1325 return V;
1327 // (X << A) >> A -> X
1328 Value *X;
1329 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1330 return X;
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 if (match(Op1, m_APInt(ShRAmt)) &&
1340 match(Op0, m_c_Or(m_NUWShl(m_Value(X), m_APInt(ShLAmt)), m_Value(Y))) &&
1341 *ShRAmt == *ShLAmt) {
1342 const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1343 const unsigned Width = Op0->getType()->getScalarSizeInBits();
1344 const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1345 if (ShRAmt->uge(EffWidthY))
1346 return X;
1349 return nullptr;
1352 Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1353 const SimplifyQuery &Q) {
1354 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1357 /// Given operands for an AShr, see if we can fold the result.
1358 /// If not, this returns null.
1359 static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1360 const SimplifyQuery &Q, unsigned MaxRecurse) {
1361 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1362 MaxRecurse))
1363 return V;
1365 // all ones >>a X -> -1
1366 // Do not return Op0 because it may contain undef elements if it's a vector.
1367 if (match(Op0, m_AllOnes()))
1368 return Constant::getAllOnesValue(Op0->getType());
1370 // (X << A) >> A -> X
1371 Value *X;
1372 if (Q.IIQ.UseInstrInfo && match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1373 return X;
1375 // Arithmetic shifting an all-sign-bit value is a no-op.
1376 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1377 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1378 return Op0;
1380 return nullptr;
1383 Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1384 const SimplifyQuery &Q) {
1385 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1388 /// Commuted variants are assumed to be handled by calling this function again
1389 /// with the parameters swapped.
1390 static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1391 ICmpInst *UnsignedICmp, bool IsAnd) {
1392 Value *X, *Y;
1394 ICmpInst::Predicate EqPred;
1395 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
1396 !ICmpInst::isEquality(EqPred))
1397 return nullptr;
1399 ICmpInst::Predicate UnsignedPred;
1400 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
1401 ICmpInst::isUnsigned(UnsignedPred))
1403 else if (match(UnsignedICmp,
1404 m_ICmp(UnsignedPred, m_Specific(Y), m_Value(X))) &&
1405 ICmpInst::isUnsigned(UnsignedPred))
1406 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1407 else
1408 return nullptr;
1410 // X < Y && Y != 0 --> X < Y
1411 // X < Y || Y != 0 --> Y != 0
1412 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1413 return IsAnd ? UnsignedICmp : ZeroICmp;
1415 // X >= Y || Y != 0 --> true
1416 // X >= Y || Y == 0 --> X >= Y
1417 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1418 if (EqPred == ICmpInst::ICMP_NE)
1419 return getTrue(UnsignedICmp->getType());
1420 return UnsignedICmp;
1423 // X < Y && Y == 0 --> false
1424 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1425 IsAnd)
1426 return getFalse(UnsignedICmp->getType());
1428 return nullptr;
1431 /// Commuted variants are assumed to be handled by calling this function again
1432 /// with the parameters swapped.
1433 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 return nullptr;
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 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1444 return Op0;
1446 // Check for any combination of predicates that are guaranteed to be disjoint.
1447 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1448 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1449 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1450 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1451 return getFalse(Op0->getType());
1453 return nullptr;
1456 /// Commuted variants are assumed to be handled by calling this function again
1457 /// with the parameters swapped.
1458 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 return nullptr;
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 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1469 return Op1;
1471 // Check for any combination of predicates that cover the entire range of
1472 // possibilities.
1473 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1474 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1475 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1476 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1477 return getTrue(Op0->getType());
1479 return nullptr;
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 static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1486 bool IsAnd) {
1487 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1488 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1489 return nullptr;
1491 const APInt *C0, *C1;
1492 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1493 !match(Cmp1->getOperand(1), m_APInt(C1)))
1494 return nullptr;
1496 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1497 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1499 // For and-of-compares, check if the intersection is empty:
1500 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1501 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1502 return getFalse(Cmp0->getType());
1504 // For or-of-compares, check if the union is full:
1505 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1506 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1507 return getTrue(Cmp0->getType());
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 if (Range0.contains(Range1))
1515 return IsAnd ? Cmp1 : Cmp0;
1516 if (Range1.contains(Range0))
1517 return IsAnd ? Cmp0 : Cmp1;
1519 return nullptr;
1522 static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1523 bool IsAnd) {
1524 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1525 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1526 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1527 return nullptr;
1529 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1530 return nullptr;
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);
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.
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 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1545 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1546 return Cmp1;
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 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1553 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1554 return Cmp0;
1556 return nullptr;
1559 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 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1566 return nullptr;
1568 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1569 return nullptr;
1571 auto *AddInst = cast<OverflowingBinaryOperator>(Op0->getOperand(0));
1572 if (AddInst->getOperand(1) != Op1->getOperand(1))
1573 return nullptr;
1575 Type *ITy = Op0->getType();
1576 bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1577 bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1579 const APInt Delta = *C1 - *C0;
1580 if (C0->isStrictlyPositive()) {
1581 if (Delta == 2) {
1582 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1583 return getFalse(ITy);
1584 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1585 return getFalse(ITy);
1587 if (Delta == 1) {
1588 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1589 return getFalse(ITy);
1590 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1591 return getFalse(ITy);
1594 if (C0->getBoolValue() && isNUW) {
1595 if (Delta == 2)
1596 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1597 return getFalse(ITy);
1598 if (Delta == 1)
1599 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1600 return getFalse(ITy);
1603 return nullptr;
1606 static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1607 const InstrInfoQuery &IIQ) {
1608 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1609 return X;
1610 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1611 return X;
1613 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1614 return X;
1615 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1616 return X;
1618 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1619 return X;
1621 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1622 return X;
1624 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1, IIQ))
1625 return X;
1626 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0, IIQ))
1627 return X;
1629 return nullptr;
1632 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 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1639 return nullptr;
1641 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1642 return nullptr;
1644 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1645 if (AddInst->getOperand(1) != Op1->getOperand(1))
1646 return nullptr;
1648 Type *ITy = Op0->getType();
1649 bool isNSW = IIQ.hasNoSignedWrap(AddInst);
1650 bool isNUW = IIQ.hasNoUnsignedWrap(AddInst);
1652 const APInt Delta = *C1 - *C0;
1653 if (C0->isStrictlyPositive()) {
1654 if (Delta == 2) {
1655 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1656 return getTrue(ITy);
1657 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1658 return getTrue(ITy);
1660 if (Delta == 1) {
1661 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1662 return getTrue(ITy);
1663 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1664 return getTrue(ITy);
1667 if (C0->getBoolValue() && isNUW) {
1668 if (Delta == 2)
1669 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1670 return getTrue(ITy);
1671 if (Delta == 1)
1672 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1673 return getTrue(ITy);
1676 return nullptr;
1679 static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1,
1680 const InstrInfoQuery &IIQ) {
1681 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1682 return X;
1683 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1684 return X;
1686 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1687 return X;
1688 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1689 return X;
1691 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1692 return X;
1694 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1695 return X;
1697 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1, IIQ))
1698 return X;
1699 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0, IIQ))
1700 return X;
1702 return nullptr;
1705 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 if (LHS0->getType() != RHS0->getType())
1710 return nullptr;
1712 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1713 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1714 (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 if ((isKnownNeverNaN(LHS0, TLI) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1724 (isKnownNeverNaN(LHS1, TLI) && (LHS0 == RHS0 || LHS0 == RHS1)))
1725 return RHS;
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 if ((isKnownNeverNaN(RHS0, TLI) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1736 (isKnownNeverNaN(RHS1, TLI) && (RHS0 == LHS0 || RHS0 == LHS1)))
1737 return LHS;
1740 return nullptr;
1743 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 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1749 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1750 Op0 = Cast0->getOperand(0);
1751 Op1 = Cast1->getOperand(0);
1754 Value *V = nullptr;
1755 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1756 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1757 if (ICmp0 && ICmp1)
1758 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1, Q.IIQ)
1759 : simplifyOrOfICmps(ICmp0, ICmp1, Q.IIQ);
1761 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1762 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1763 if (FCmp0 && FCmp1)
1764 V = simplifyAndOrOfFCmps(Q.TLI, FCmp0, FCmp1, IsAnd);
1766 if (!V)
1767 return nullptr;
1768 if (!Cast0)
1769 return V;
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 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1776 return nullptr;
1779 /// Given operands for an And, see if we can fold the result.
1780 /// If not, this returns null.
1781 static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1782 unsigned MaxRecurse) {
1783 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1784 return C;
1786 // X & undef -> 0
1787 if (match(Op1, m_Undef()))
1788 return Constant::getNullValue(Op0->getType());
1790 // X & X = X
1791 if (Op0 == Op1)
1792 return Op0;
1794 // X & 0 = 0
1795 if (match(Op1, m_Zero()))
1796 return Constant::getNullValue(Op0->getType());
1798 // X & -1 = X
1799 if (match(Op1, m_AllOnes()))
1800 return Op0;
1802 // A & ~A = ~A & A = 0
1803 if (match(Op0, m_Not(m_Specific(Op1))) ||
1804 match(Op1, m_Not(m_Specific(Op0))))
1805 return Constant::getNullValue(Op0->getType());
1807 // (A | ?) & A = A
1808 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1809 return Op1;
1811 // A & (A | ?) = A
1812 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1813 return Op0;
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 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 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1823 (~(*Mask)).lshr(*ShAmt).isNullValue())
1824 return Op0;
1826 // If all bits in the inverted and shifted mask are clear:
1827 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1828 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1829 (~(*Mask)).shl(*ShAmt).isNullValue())
1830 return Op0;
1833 // A & (-A) = A if A is a power of two or zero.
1834 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1835 match(Op1, m_Neg(m_Specific(Op0)))) {
1836 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1837 Q.DT))
1838 return Op0;
1839 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1840 Q.DT))
1841 return Op1;
1844 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
1845 return V;
1847 // Try some generic simplifications for associative operations.
1848 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1849 MaxRecurse))
1850 return V;
1852 // And distributes over Or. Try some generic simplifications based on this.
1853 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1854 Q, MaxRecurse))
1855 return V;
1857 // And distributes over Xor. Try some generic simplifications based on this.
1858 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1859 Q, MaxRecurse))
1860 return V;
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 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1865 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1866 MaxRecurse))
1867 return V;
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 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1872 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1873 MaxRecurse))
1874 return V;
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 if (match(Op1, m_APInt(Mask)) &&
1888 match(Op0, m_c_Or(m_CombineAnd(m_NUWShl(m_Value(X), m_APInt(ShAmt)),
1889 m_Value(XShifted)),
1890 m_Value(Y)))) {
1891 const unsigned Width = Op0->getType()->getScalarSizeInBits();
1892 const unsigned ShftCnt = ShAmt->getLimitedValue(Width);
1893 const KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1894 const unsigned EffWidthY = Width - YKnown.countMinLeadingZeros();
1895 if (EffWidthY <= ShftCnt) {
1896 const KnownBits XKnown = computeKnownBits(X, Q.DL, 0, Q.AC, Q.CxtI,
1897 Q.DT);
1898 const unsigned EffWidthX = Width - XKnown.countMinLeadingZeros();
1899 const APInt EffBitsY = APInt::getLowBitsSet(Width, EffWidthY);
1900 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 if (EffBitsY.isSubsetOf(*Mask) && !EffBitsX.intersects(*Mask))
1904 return Y;
1905 if (EffBitsX.isSubsetOf(*Mask) && !EffBitsY.intersects(*Mask))
1906 return XShifted;
1910 return nullptr;
1913 Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1914 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1917 /// Given operands for an Or, see if we can fold the result.
1918 /// If not, this returns null.
1919 static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1920 unsigned MaxRecurse) {
1921 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1922 return C;
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 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1928 return Constant::getAllOnesValue(Op0->getType());
1930 // X | X = X
1931 // X | 0 = X
1932 if (Op0 == Op1 || match(Op1, m_Zero()))
1933 return Op0;
1935 // A | ~A = ~A | A = -1
1936 if (match(Op0, m_Not(m_Specific(Op1))) ||
1937 match(Op1, m_Not(m_Specific(Op0))))
1938 return Constant::getAllOnesValue(Op0->getType());
1940 // (A & ?) | A = A
1941 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1942 return Op1;
1944 // A | (A & ?) = A
1945 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1946 return Op0;
1948 // ~(A & ?) | A = -1
1949 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1950 return Constant::getAllOnesValue(Op1->getType());
1952 // A | ~(A & ?) = -1
1953 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1954 return Constant::getAllOnesValue(Op0->getType());
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 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1962 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1963 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1964 return Op1;
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 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1972 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1973 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1974 return Op0;
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 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1981 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1982 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1983 return Op1;
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 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1990 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1991 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1992 return Op0;
1994 if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
1995 return V;
1997 // Try some generic simplifications for associative operations.
1998 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1999 MaxRecurse))
2000 return V;
2002 // Or distributes over And. Try some generic simplifications based on this.
2003 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
2004 MaxRecurse))
2005 return V;
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 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
2010 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
2011 MaxRecurse))
2012 return V;
2014 // (A & C1)|(B & C2)
2015 const APInt *C1, *C2;
2016 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
2017 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
2018 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 if (C2->isMask() && // C2 == 0+1+
2025 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
2026 // Add commutes, try both ways.
2027 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2028 return A;
2030 // Or commutes, try both ways.
2031 if (C1->isMask() &&
2032 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
2033 // Add commutes, try both ways.
2034 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2035 return B;
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 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
2043 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
2044 return V;
2046 return nullptr;
2049 Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2050 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
2053 /// Given operands for a Xor, see if we can fold the result.
2054 /// If not, this returns null.
2055 static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
2056 unsigned MaxRecurse) {
2057 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
2058 return C;
2060 // A ^ undef -> undef
2061 if (match(Op1, m_Undef()))
2062 return Op1;
2064 // A ^ 0 = A
2065 if (match(Op1, m_Zero()))
2066 return Op0;
2068 // A ^ A = 0
2069 if (Op0 == Op1)
2070 return Constant::getNullValue(Op0->getType());
2072 // A ^ ~A = ~A ^ A = -1
2073 if (match(Op0, m_Not(m_Specific(Op1))) ||
2074 match(Op1, m_Not(m_Specific(Op0))))
2075 return Constant::getAllOnesValue(Op0->getType());
2077 // Try some generic simplifications for associative operations.
2078 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
2079 MaxRecurse))
2080 return V;
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.
2091 return nullptr;
2094 Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
2095 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
2099 static Type *GetCompareTy(Value *Op) {
2100 return CmpInst::makeCmpResultType(Op->getType());
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 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 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
2116 return Cmp;
2117 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
2118 LHS == CmpRHS && RHS == CmpLHS)
2119 return Cmp;
2120 return nullptr;
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.
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."
2133 // This is pretty permissive.
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.)
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.
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 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 LHS = LHS->stripPointerCasts();
2158 RHS = RHS->stripPointerCasts();
2160 // A non-null pointer is not equal to a null pointer.
2161 if (llvm::isKnownNonZero(LHS, DL, 0, nullptr, nullptr, nullptr,
2162 IIQ.UseInstrInfo) &&
2163 isa<ConstantPointerNull>(RHS) &&
2164 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2165 return ConstantInt::get(GetCompareTy(LHS),
2166 !CmpInst::isTrueWhenEqual(Pred));
2168 // We can only fold certain predicates on pointer comparisons.
2169 switch (Pred) {
2170 default:
2171 return nullptr;
2173 // Equality comaprisons are easy to fold.
2174 case CmpInst::ICMP_EQ:
2175 case CmpInst::ICMP_NE:
2176 break;
2178 // We can only handle unsigned relational comparisons because 'inbounds' on
2179 // a GEP only protects against unsigned wrapping.
2180 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 Pred = ICmpInst::getSignedPredicate(Pred);
2187 break;
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 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2197 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
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 if (LHS == RHS)
2202 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2204 // Various optimizations for (in)equality comparisons.
2205 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...
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.
2226 // So, we'll assume that two non-empty allocas have different addresses
2227 // for now.
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.
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 (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 ObjectSizeOpts Opts;
2241 Opts.NullIsUnknownSize =
2242 NullPointerIsDefined(cast<AllocaInst>(LHS)->getFunction());
2243 if (LHSOffsetCI && RHSOffsetCI &&
2244 getObjectSize(LHS, LHSSize, DL, TLI, Opts) &&
2245 getObjectSize(RHS, RHSSize, DL, TLI, Opts)) {
2246 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2247 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2248 if (!LHSOffsetValue.isNegative() &&
2249 !RHSOffsetValue.isNegative() &&
2250 LHSOffsetValue.ult(LHSSize) &&
2251 RHSOffsetValue.ult(RHSSize)) {
2252 return ConstantInt::get(GetCompareTy(LHS),
2253 !CmpInst::isTrueWhenEqual(Pred));
2257 // Repeat the above check but this time without depending on DataLayout
2258 // or being able to compute a precise size.
2259 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2260 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2261 LHSOffset->isNullValue() &&
2262 RHSOffset->isNullValue())
2263 return ConstantInt::get(GetCompareTy(LHS),
2264 !CmpInst::isTrueWhenEqual(Pred));
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 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2272 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2273 if (LHS == RHS)
2274 return ConstantExpr::getICmp(Pred,
2275 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2276 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
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 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2285 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2287 // Is the set of underlying objects all noalias calls?
2288 auto IsNAC = [](ArrayRef<Value *> Objects) {
2289 return all_of(Objects, isNoAliasCall);
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;
2312 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2313 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2314 return ConstantInt::get(GetCompareTy(LHS),
2315 !CmpInst::isTrueWhenEqual(Pred));
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 if (isAllocLikeFn(LHS, TLI) &&
2322 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2323 MI = LHS;
2324 else if (isAllocLikeFn(RHS, TLI) &&
2325 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2326 MI = RHS;
2327 // FIXME: We should also fold the compare when the pointer escapes, but the
2328 // compare dominates the pointer escape
2329 if (MI && !PointerMayBeCaptured(MI, true, true))
2330 return ConstantInt::get(GetCompareTy(LHS),
2331 CmpInst::isFalseWhenEqual(Pred));
2334 // Otherwise, fail.
2335 return nullptr;
2338 /// Fold an icmp when its operands have i1 scalar type.
2339 static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2340 Value *RHS, const SimplifyQuery &Q) {
2341 Type *ITy = GetCompareTy(LHS); // The return type.
2342 Type *OpTy = LHS->getType(); // The operand type.
2343 if (!OpTy->isIntOrIntVectorTy(1))
2344 return nullptr;
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 if (match(RHS, m_Zero())) {
2350 switch (Pred) {
2351 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 return LHS;
2356 case CmpInst::ICMP_ULT: // X <u 0 -> false
2357 case CmpInst::ICMP_SGT: // X >s 0 -> false
2358 return getFalse(ITy);
2360 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2361 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2362 return getTrue(ITy);
2364 default: break;
2366 } else if (match(RHS, m_One())) {
2367 switch (Pred) {
2368 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 return LHS;
2373 case CmpInst::ICMP_UGT: // X >u 1 -> false
2374 case CmpInst::ICMP_SLT: // X <s -1 -> false
2375 return getFalse(ITy);
2377 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2378 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2379 return getTrue(ITy);
2381 default: break;
2385 switch (Pred) {
2386 default:
2387 break;
2388 case ICmpInst::ICMP_UGE:
2389 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2390 return getTrue(ITy);
2391 break;
2392 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 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2401 return getTrue(ITy);
2402 break;
2403 case ICmpInst::ICMP_ULE:
2404 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2405 return getTrue(ITy);
2406 break;
2409 return nullptr;
2412 /// Try hard to fold icmp with zero RHS because this is a common case.
2413 static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2414 Value *RHS, const SimplifyQuery &Q) {
2415 if (!match(RHS, m_Zero()))
2416 return nullptr;
2418 Type *ITy = GetCompareTy(LHS); // The return type.
2419 switch (Pred) {
2420 default:
2421 llvm_unreachable("Unknown ICmp predicate!");
2422 case ICmpInst::ICMP_ULT:
2423 return getFalse(ITy);
2424 case ICmpInst::ICMP_UGE:
2425 return getTrue(ITy);
2426 case ICmpInst::ICMP_EQ:
2427 case ICmpInst::ICMP_ULE:
2428 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2429 return getFalse(ITy);
2430 break;
2431 case ICmpInst::ICMP_NE:
2432 case ICmpInst::ICMP_UGT:
2433 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo))
2434 return getTrue(ITy);
2435 break;
2436 case ICmpInst::ICMP_SLT: {
2437 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2438 if (LHSKnown.isNegative())
2439 return getTrue(ITy);
2440 if (LHSKnown.isNonNegative())
2441 return getFalse(ITy);
2442 break;
2444 case ICmpInst::ICMP_SLE: {
2445 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2446 if (LHSKnown.isNegative())
2447 return getTrue(ITy);
2448 if (LHSKnown.isNonNegative() &&
2449 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2450 return getFalse(ITy);
2451 break;
2453 case ICmpInst::ICMP_SGE: {
2454 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2455 if (LHSKnown.isNegative())
2456 return getFalse(ITy);
2457 if (LHSKnown.isNonNegative())
2458 return getTrue(ITy);
2459 break;
2461 case ICmpInst::ICMP_SGT: {
2462 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2463 if (LHSKnown.isNegative())
2464 return getFalse(ITy);
2465 if (LHSKnown.isNonNegative() &&
2466 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2467 return getTrue(ITy);
2468 break;
2472 return nullptr;
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 static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper,
2479 const InstrInfoQuery &IIQ) {
2480 unsigned Width = Lower.getBitWidth();
2481 const APInt *C;
2482 switch (BO.getOpcode()) {
2483 case Instruction::Add:
2484 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 if (IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
2487 // 'add nuw x, C' produces [C, UINT_MAX].
2488 Lower = *C;
2489 } else if (IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(&BO))) {
2490 if (C->isNegative()) {
2491 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2492 Lower = APInt::getSignedMinValue(Width);
2493 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2494 } else {
2495 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2496 Lower = APInt::getSignedMinValue(Width) + *C;
2497 Upper = APInt::getSignedMaxValue(Width) + 1;
2501 break;
2503 case Instruction::And:
2504 if (match(BO.getOperand(1), m_APInt(C)))
2505 // 'and x, C' produces [0, C].
2506 Upper = *C + 1;
2507 break;
2509 case Instruction::Or:
2510 if (match(BO.getOperand(1), m_APInt(C)))
2511 // 'or x, C' produces [C, UINT_MAX].
2512 Lower = *C;
2513 break;
2515 case Instruction::AShr:
2516 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2517 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2518 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2519 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2520 } else if (match(BO.getOperand(0), m_APInt(C))) {
2521 unsigned ShiftAmount = Width - 1;
2522 if (!C->isNullValue() && IIQ.isExact(&BO))
2523 ShiftAmount = C->countTrailingZeros();
2524 if (C->isNegative()) {
2525 // 'ashr C, x' produces [C, C >> (Width-1)]
2526 Lower = *C;
2527 Upper = C->ashr(ShiftAmount) + 1;
2528 } else {
2529 // 'ashr C, x' produces [C >> (Width-1), C]
2530 Lower = C->ashr(ShiftAmount);
2531 Upper = *C + 1;
2534 break;
2536 case Instruction::LShr:
2537 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2538 // 'lshr x, C' produces [0, UINT_MAX >> C].
2539 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2540 } else if (match(BO.getOperand(0), m_APInt(C))) {
2541 // 'lshr C, x' produces [C >> (Width-1), C].
2542 unsigned ShiftAmount = Width - 1;
2543 if (!C->isNullValue() && IIQ.isExact(&BO))
2544 ShiftAmount = C->countTrailingZeros();
2545 Lower = C->lshr(ShiftAmount);
2546 Upper = *C + 1;
2548 break;
2550 case Instruction::Shl:
2551 if (match(BO.getOperand(0), m_APInt(C))) {
2552 if (IIQ.hasNoUnsignedWrap(&BO)) {
2553 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2554 Lower = *C;
2555 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2556 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2557 if (C->isNegative()) {
2558 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2559 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2560 Lower = C->shl(ShiftAmount);
2561 Upper = *C + 1;
2562 } else {
2563 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2564 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2565 Lower = *C;
2566 Upper = C->shl(ShiftAmount) + 1;
2570 break;
2572 case Instruction::SDiv:
2573 if (match(BO.getOperand(1), m_APInt(C))) {
2574 APInt IntMin = APInt::getSignedMinValue(Width);
2575 APInt IntMax = APInt::getSignedMaxValue(Width);
2576 if (C->isAllOnesValue()) {
2577 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2578 // where C != -1 and C != 0 and C != 1
2579 Lower = IntMin + 1;
2580 Upper = IntMax + 1;
2581 } 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 Lower = IntMin.sdiv(*C);
2585 Upper = IntMax.sdiv(*C);
2586 if (Lower.sgt(Upper))
2587 std::swap(Lower, Upper);
2588 Upper = Upper + 1;
2589 assert(Upper != Lower && "Upper part of range has wrapped!");
2591 } else if (match(BO.getOperand(0), m_APInt(C))) {
2592 if (C->isMinSignedValue()) {
2593 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2594 Lower = *C;
2595 Upper = Lower.lshr(1) + 1;
2596 } else {
2597 // 'sdiv C, x' produces [-|C|, |C|].
2598 Upper = C->abs() + 1;
2599 Lower = (-Upper) + 1;
2602 break;
2604 case Instruction::UDiv:
2605 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2606 // 'udiv x, C' produces [0, UINT_MAX / C].
2607 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2608 } else if (match(BO.getOperand(0), m_APInt(C))) {
2609 // 'udiv C, x' produces [0, C].
2610 Upper = *C + 1;
2612 break;
2614 case Instruction::SRem:
2615 if (match(BO.getOperand(1), m_APInt(C))) {
2616 // 'srem x, C' produces (-|C|, |C|).
2617 Upper = C->abs();
2618 Lower = (-Upper) + 1;
2620 break;
2622 case Instruction::URem:
2623 if (match(BO.getOperand(1), m_APInt(C)))
2624 // 'urem x, C' produces [0, C).
2625 Upper = *C;
2626 break;
2628 default:
2629 break;
2633 static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2634 Value *RHS, const InstrInfoQuery &IIQ) {
2635 Type *ITy = GetCompareTy(RHS); // The return type.
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 if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) {
2643 if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero()))
2644 return ConstantInt::getFalse(ITy);
2645 if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()))
2646 return ConstantInt::getTrue(ITy);
2649 const APInt *C;
2650 if (!match(RHS, m_APInt(C)))
2651 return nullptr;
2653 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2654 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2655 if (RHS_CR.isEmptySet())
2656 return ConstantInt::getFalse(ITy);
2657 if (RHS_CR.isFullSet())
2658 return ConstantInt::getTrue(ITy);
2660 // Find the range of possible values for binary operators.
2661 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 setLimitsForBinOp(*BO, Lower, Upper, IIQ);
2667 ConstantRange LHS_CR =
2668 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2670 if (auto *I = dyn_cast<Instruction>(LHS))
2671 if (auto *Ranges = IIQ.getMetadata(I, LLVMContext::MD_range))
2672 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2674 if (!LHS_CR.isFullSet()) {
2675 if (RHS_CR.contains(LHS_CR))
2676 return ConstantInt::getTrue(ITy);
2677 if (RHS_CR.inverse().contains(LHS_CR))
2678 return ConstantInt::getFalse(ITy);
2681 return nullptr;
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 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.
2692 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2693 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2694 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 if (LBO && LBO->getOpcode() == Instruction::Add) {
2700 A = LBO->getOperand(0);
2701 B = LBO->getOperand(1);
2702 NoLHSWrapProblem =
2703 ICmpInst::isEquality(Pred) ||
2704 (CmpInst::isUnsigned(Pred) &&
2705 Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO))) ||
2706 (CmpInst::isSigned(Pred) &&
2707 Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)));
2709 if (RBO && RBO->getOpcode() == Instruction::Add) {
2710 C = RBO->getOperand(0);
2711 D = RBO->getOperand(1);
2712 NoRHSWrapProblem =
2713 ICmpInst::isEquality(Pred) ||
2714 (CmpInst::isUnsigned(Pred) &&
2715 Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(RBO))) ||
2716 (CmpInst::isSigned(Pred) &&
2717 Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RBO)));
2720 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2721 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2722 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2723 Constant::getNullValue(RHS->getType()), Q,
2724 MaxRecurse - 1))
2725 return V;
2727 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2728 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2729 if (Value *V =
2730 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2731 C == LHS ? D : C, Q, MaxRecurse - 1))
2732 return V;
2734 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2735 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 if (A == C) {
2740 // C + B == C + D -> B == D
2741 Y = B;
2742 Z = D;
2743 } else if (A == D) {
2744 // D + B == C + D -> B == C
2745 Y = B;
2746 Z = C;
2747 } 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;
2757 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2758 return V;
2763 Value *Y = nullptr;
2764 // icmp pred (or X, Y), X
2765 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2766 if (Pred == ICmpInst::ICMP_ULT)
2767 return getFalse(ITy);
2768 if (Pred == ICmpInst::ICMP_UGE)
2769 return getTrue(ITy);
2771 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2772 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2773 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2774 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2775 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2776 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2777 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2780 // icmp pred X, (or X, Y)
2781 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2782 if (Pred == ICmpInst::ICMP_ULE)
2783 return getTrue(ITy);
2784 if (Pred == ICmpInst::ICMP_UGT)
2785 return getFalse(ITy);
2787 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2788 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2789 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2790 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2791 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2792 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2793 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2798 // icmp pred (and X, Y), X
2799 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2800 if (Pred == ICmpInst::ICMP_UGT)
2801 return getFalse(ITy);
2802 if (Pred == ICmpInst::ICMP_ULE)
2803 return getTrue(ITy);
2805 // icmp pred X, (and X, Y)
2806 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2807 if (Pred == ICmpInst::ICMP_UGE)
2808 return getTrue(ITy);
2809 if (Pred == ICmpInst::ICMP_ULT)
2810 return getFalse(ITy);
2813 // 0 - (zext X) pred C
2814 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2815 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2816 if (RHSC->getValue().isStrictlyPositive()) {
2817 if (Pred == ICmpInst::ICMP_SLT)
2818 return ConstantInt::getTrue(RHSC->getContext());
2819 if (Pred == ICmpInst::ICMP_SGE)
2820 return ConstantInt::getFalse(RHSC->getContext());
2821 if (Pred == ICmpInst::ICMP_EQ)
2822 return ConstantInt::getFalse(RHSC->getContext());
2823 if (Pred == ICmpInst::ICMP_NE)
2824 return ConstantInt::getTrue(RHSC->getContext());
2826 if (RHSC->getValue().isNonNegative()) {
2827 if (Pred == ICmpInst::ICMP_SLE)
2828 return ConstantInt::getTrue(RHSC->getContext());
2829 if (Pred == ICmpInst::ICMP_SGT)
2830 return ConstantInt::getFalse(RHSC->getContext());
2835 // icmp pred (urem X, Y), Y
2836 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2837 switch (Pred) {
2838 default:
2839 break;
2840 case ICmpInst::ICMP_SGT:
2841 case ICmpInst::ICMP_SGE: {
2842 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2843 if (!Known.isNonNegative())
2844 break;
2845 LLVM_FALLTHROUGH;
2847 case ICmpInst::ICMP_EQ:
2848 case ICmpInst::ICMP_UGT:
2849 case ICmpInst::ICMP_UGE:
2850 return getFalse(ITy);
2851 case ICmpInst::ICMP_SLT:
2852 case ICmpInst::ICMP_SLE: {
2853 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2854 if (!Known.isNonNegative())
2855 break;
2856 LLVM_FALLTHROUGH;
2858 case ICmpInst::ICMP_NE:
2859 case ICmpInst::ICMP_ULT:
2860 case ICmpInst::ICMP_ULE:
2861 return getTrue(ITy);
2865 // icmp pred X, (urem Y, X)
2866 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2867 switch (Pred) {
2868 default:
2869 break;
2870 case ICmpInst::ICMP_SGT:
2871 case ICmpInst::ICMP_SGE: {
2872 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2873 if (!Known.isNonNegative())
2874 break;
2875 LLVM_FALLTHROUGH;
2877 case ICmpInst::ICMP_NE:
2878 case ICmpInst::ICMP_UGT:
2879 case ICmpInst::ICMP_UGE:
2880 return getTrue(ITy);
2881 case ICmpInst::ICMP_SLT:
2882 case ICmpInst::ICMP_SLE: {
2883 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2884 if (!Known.isNonNegative())
2885 break;
2886 LLVM_FALLTHROUGH;
2888 case ICmpInst::ICMP_EQ:
2889 case ICmpInst::ICMP_ULT:
2890 case ICmpInst::ICMP_ULE:
2891 return getFalse(ITy);
2895 // x >> y <=u x
2896 // x udiv y <=u x.
2897 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2898 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2899 // icmp pred (X op Y), X
2900 if (Pred == ICmpInst::ICMP_UGT)
2901 return getFalse(ITy);
2902 if (Pred == ICmpInst::ICMP_ULE)
2903 return getTrue(ITy);
2906 // x >=u x >> y
2907 // x >=u x udiv y.
2908 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2909 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2910 // icmp pred X, (X op Y)
2911 if (Pred == ICmpInst::ICMP_ULT)
2912 return getFalse(ITy);
2913 if (Pred == ICmpInst::ICMP_UGE)
2914 return getTrue(ITy);
2917 // handle:
2918 // CI2 << X == CI
2919 // CI2 << X != CI
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 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2925 CI2Val->isPowerOf2()) {
2926 if (!CIVal->isPowerOf2()) {
2927 // CI2 << X can equal zero in some circumstances,
2928 // this simplification is unsafe if CI is zero.
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 if (Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2936 Q.IIQ.hasNoUnsignedWrap(cast<OverflowingBinaryOperator>(LBO)) ||
2937 CI2Val->isOneValue() || !CI->isZero()) {
2938 if (Pred == ICmpInst::ICMP_EQ)
2939 return ConstantInt::getFalse(RHS->getContext());
2940 if (Pred == ICmpInst::ICMP_NE)
2941 return ConstantInt::getTrue(RHS->getContext());
2944 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2945 if (Pred == ICmpInst::ICMP_UGT)
2946 return ConstantInt::getFalse(RHS->getContext());
2947 if (Pred == ICmpInst::ICMP_ULE)
2948 return ConstantInt::getTrue(RHS->getContext());
2953 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2954 LBO->getOperand(1) == RBO->getOperand(1)) {
2955 switch (LBO->getOpcode()) {
2956 default:
2957 break;
2958 case Instruction::UDiv:
2959 case Instruction::LShr:
2960 if (ICmpInst::isSigned(Pred) || !Q.IIQ.isExact(LBO) ||
2961 !Q.IIQ.isExact(RBO))
2962 break;
2963 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2964 RBO->getOperand(0), Q, MaxRecurse - 1))
2965 return V;
2966 break;
2967 case Instruction::SDiv:
2968 if (!ICmpInst::isEquality(Pred) || !Q.IIQ.isExact(LBO) ||
2969 !Q.IIQ.isExact(RBO))
2970 break;
2971 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2972 RBO->getOperand(0), Q, MaxRecurse - 1))
2973 return V;
2974 break;
2975 case Instruction::AShr:
2976 if (!Q.IIQ.isExact(LBO) || !Q.IIQ.isExact(RBO))
2977 break;
2978 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2979 RBO->getOperand(0), Q, MaxRecurse - 1))
2980 return V;
2981 break;
2982 case Instruction::Shl: {
2983 bool NUW = Q.IIQ.hasNoUnsignedWrap(LBO) && Q.IIQ.hasNoUnsignedWrap(RBO);
2984 bool NSW = Q.IIQ.hasNoSignedWrap(LBO) && Q.IIQ.hasNoSignedWrap(RBO);
2985 if (!NUW && !NSW)
2986 break;
2987 if (!NSW && ICmpInst::isSigned(Pred))
2988 break;
2989 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2990 RBO->getOperand(0), Q, MaxRecurse - 1))
2991 return V;
2992 break;
2996 return nullptr;
2999 /// Simplify integer comparisons where at least one operand of the compare
3000 /// matches an integer min/max idiom.
3001 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".
3009 // Signed variants on "max(a,b)>=a -> true".
3010 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3011 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 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
3017 (A == LHS || B == LHS)) {
3018 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 P = CmpInst::getSwappedPredicate(Pred);
3023 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3024 (A == RHS || B == RHS)) {
3025 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 P = CmpInst::getSwappedPredicate(Pred);
3031 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
3032 (A == LHS || B == LHS)) {
3033 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;
3040 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3041 // Cases correspond to "max(A, B) p A".
3042 switch (P) {
3043 default:
3044 break;
3045 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 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3050 return V;
3051 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3052 return V;
3053 // Otherwise, see if "A EqP B" simplifies.
3054 if (MaxRecurse)
3055 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3056 return V;
3057 break;
3058 case CmpInst::ICMP_NE:
3059 case CmpInst::ICMP_SGT: {
3060 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 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3064 return V;
3065 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3066 return V;
3067 // Otherwise, see if "A InvEqP B" simplifies.
3068 if (MaxRecurse)
3069 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3070 return V;
3071 break;
3073 case CmpInst::ICMP_SGE:
3074 // Always true.
3075 return getTrue(ITy);
3076 case CmpInst::ICMP_SLT:
3077 // Always false.
3078 return getFalse(ITy);
3082 // Unsigned variants on "max(a,b)>=a -> true".
3083 P = CmpInst::BAD_ICMP_PREDICATE;
3084 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
3085 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 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
3091 (A == LHS || B == LHS)) {
3092 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 P = CmpInst::getSwappedPredicate(Pred);
3097 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3098 (A == RHS || B == RHS)) {
3099 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 P = CmpInst::getSwappedPredicate(Pred);
3105 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
3106 (A == LHS || B == LHS)) {
3107 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;
3114 if (P != CmpInst::BAD_ICMP_PREDICATE) {
3115 // Cases correspond to "max(A, B) p A".
3116 switch (P) {
3117 default:
3118 break;
3119 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 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
3124 return V;
3125 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
3126 return V;
3127 // Otherwise, see if "A EqP B" simplifies.
3128 if (MaxRecurse)
3129 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
3130 return V;
3131 break;
3132 case CmpInst::ICMP_NE:
3133 case CmpInst::ICMP_UGT: {
3134 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 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
3138 return V;
3139 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
3140 return V;
3141 // Otherwise, see if "A InvEqP B" simplifies.
3142 if (MaxRecurse)
3143 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
3144 return V;
3145 break;
3147 case CmpInst::ICMP_UGE:
3148 // Always true.
3149 return getTrue(ITy);
3150 case CmpInst::ICMP_ULT:
3151 // Always false.
3152 return getFalse(ITy);
3156 // Variants on "max(x,y) >= min(x,z)".
3157 Value *C, *D;
3158 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
3159 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
3160 (A == C || A == D || B == C || B == D)) {
3161 // max(x, ?) pred min(x, ?).
3162 if (Pred == CmpInst::ICMP_SGE)
3163 // Always true.
3164 return getTrue(ITy);
3165 if (Pred == CmpInst::ICMP_SLT)
3166 // Always false.
3167 return getFalse(ITy);
3168 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3169 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3170 (A == C || A == D || B == C || B == D)) {
3171 // min(x, ?) pred max(x, ?).
3172 if (Pred == CmpInst::ICMP_SLE)
3173 // Always true.
3174 return getTrue(ITy);
3175 if (Pred == CmpInst::ICMP_SGT)
3176 // Always false.
3177 return getFalse(ITy);
3178 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3179 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3180 (A == C || A == D || B == C || B == D)) {
3181 // max(x, ?) pred min(x, ?).
3182 if (Pred == CmpInst::ICMP_UGE)
3183 // Always true.
3184 return getTrue(ITy);
3185 if (Pred == CmpInst::ICMP_ULT)
3186 // Always false.
3187 return getFalse(ITy);
3188 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3189 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3190 (A == C || A == D || B == C || B == D)) {
3191 // min(x, ?) pred max(x, ?).
3192 if (Pred == CmpInst::ICMP_ULE)
3193 // Always true.
3194 return getTrue(ITy);
3195 if (Pred == CmpInst::ICMP_UGT)
3196 // Always false.
3197 return getFalse(ITy);
3200 return nullptr;
3203 /// Given operands for an ICmpInst, see if we can fold the result.
3204 /// If not, this returns null.
3205 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!");
3210 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3211 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3212 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3214 // If we have a constant, make sure it is on the RHS.
3215 std::swap(LHS, RHS);
3216 Pred = CmpInst::getSwappedPredicate(Pred);
3219 Type *ITy = GetCompareTy(LHS); // The return type.
3221 // icmp X, X -> true/false
3222 // icmp X, undef -> true/false because undef could be X.
3223 if (LHS == RHS || isa<UndefValue>(RHS))
3224 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3226 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3227 return V;
3229 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3230 return V;
3232 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS, Q.IIQ))
3233 return V;
3235 // If both operands have range metadata, use the metadata
3236 // to simplify the comparison.
3237 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3238 auto RHS_Instr = cast<Instruction>(RHS);
3239 auto LHS_Instr = cast<Instruction>(LHS);
3241 if (Q.IIQ.getMetadata(RHS_Instr, LLVMContext::MD_range) &&
3242 Q.IIQ.getMetadata(LHS_Instr, LLVMContext::MD_range)) {
3243 auto RHS_CR = getConstantRangeFromMetadata(
3244 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3245 auto LHS_CR = getConstantRangeFromMetadata(
3246 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3248 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3249 if (Satisfied_CR.contains(LHS_CR))
3250 return ConstantInt::getTrue(RHS->getContext());
3252 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3253 CmpInst::getInversePredicate(Pred), RHS_CR);
3254 if (InversedSatisfied_CR.contains(LHS_CR))
3255 return ConstantInt::getFalse(RHS->getContext());
3259 // Compare of cast, for example (zext X) != 0 -> X != 0
3260 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3261 Instruction *LI = cast<CastInst>(LHS);
3262 Value *SrcOp = LI->getOperand(0);
3263 Type *SrcTy = SrcOp->getType();
3264 Type *DstTy = LI->getType();
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 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3269 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3270 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3271 // Transfer the cast to the constant.
3272 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3273 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3274 Q, MaxRecurse-1))
3275 return V;
3276 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3277 if (RI->getOperand(0)->getType() == SrcTy)
3278 // Compare without the cast.
3279 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3280 Q, MaxRecurse-1))
3281 return V;
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 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3290 // Compare X and Y. Note that signed predicates become unsigned.
3291 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3292 SrcOp, RI->getOperand(0), Q,
3293 MaxRecurse-1))
3294 return V;
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 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3302 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3304 // If the re-extended constant didn't change then this is effectively
3305 // also a case of comparing two zero-extended values.
3306 if (RExt == CI && MaxRecurse)
3307 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3308 SrcOp, Trunc, Q, MaxRecurse-1))
3309 return V;
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 if (RExt != CI) {
3314 switch (Pred) {
3315 default: llvm_unreachable("Unknown ICmp predicate!");
3316 // LHS <u RHS.
3317 case ICmpInst::ICMP_EQ:
3318 case ICmpInst::ICMP_UGT:
3319 case ICmpInst::ICMP_UGE:
3320 return ConstantInt::getFalse(CI->getContext());
3322 case ICmpInst::ICMP_NE:
3323 case ICmpInst::ICMP_ULT:
3324 case ICmpInst::ICMP_ULE:
3325 return ConstantInt::getTrue(CI->getContext());
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 return CI->getValue().isNegative() ?
3332 ConstantInt::getTrue(CI->getContext()) :
3333 ConstantInt::getFalse(CI->getContext());
3335 case ICmpInst::ICMP_SLT:
3336 case ICmpInst::ICMP_SLE:
3337 return CI->getValue().isNegative() ?
3338 ConstantInt::getFalse(CI->getContext()) :
3339 ConstantInt::getTrue(CI->getContext());
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 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3350 // Compare X and Y. Note that the predicate does not change.
3351 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3352 Q, MaxRecurse-1))
3353 return V;
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 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3361 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3363 // If the re-extended constant didn't change then this is effectively
3364 // also a case of comparing two sign-extended values.
3365 if (RExt == CI && MaxRecurse)
3366 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3367 return V;
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 if (RExt != CI) {
3372 switch (Pred) {
3373 default: llvm_unreachable("Unknown ICmp predicate!");
3374 case ICmpInst::ICMP_EQ:
3375 return ConstantInt::getFalse(CI->getContext());
3376 case ICmpInst::ICMP_NE:
3377 return ConstantInt::getTrue(CI->getContext());
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 return CI->getValue().isNegative() ?
3384 ConstantInt::getTrue(CI->getContext()) :
3385 ConstantInt::getFalse(CI->getContext());
3386 case ICmpInst::ICMP_SLT:
3387 case ICmpInst::ICMP_SLE:
3388 return CI->getValue().isNegative() ?
3389 ConstantInt::getFalse(CI->getContext()) :
3390 ConstantInt::getTrue(CI->getContext());
3392 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3393 // LHS >u RHS.
3394 case ICmpInst::ICMP_UGT:
3395 case ICmpInst::ICMP_UGE:
3396 // Comparison is true iff the LHS <s 0.
3397 if (MaxRecurse)
3398 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3399 Constant::getNullValue(SrcTy),
3400 Q, MaxRecurse-1))
3401 return V;
3402 break;
3403 case ICmpInst::ICMP_ULT:
3404 case ICmpInst::ICMP_ULE:
3405 // Comparison is true iff the LHS >=s 0.
3406 if (MaxRecurse)
3407 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3408 Constant::getNullValue(SrcTy),
3409 Q, MaxRecurse-1))
3410 return V;
3411 break;
3418 // icmp eq|ne X, Y -> false|true if X != Y
3419 if (ICmpInst::isEquality(Pred) &&
3420 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT, Q.IIQ.UseInstrInfo)) {
3421 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3424 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3425 return V;
3427 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3428 return V;
3430 // Simplify comparisons of related pointers using a powerful, recursive
3431 // GEP-walk when we have target data available..
3432 if (LHS->getType()->isPointerTy())
3433 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 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3439 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3440 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3441 Q.DL.getTypeSizeInBits(CRHS->getType()))
3442 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;
3447 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3448 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3449 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3450 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3451 (ICmpInst::isEquality(Pred) ||
3452 (GLHS->isInBounds() && GRHS->isInBounds() &&
3453 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 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3458 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3459 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3460 GLHS->getSourceElementType(), Null, IndicesLHS);
3462 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3463 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3464 GLHS->getSourceElementType(), Null, IndicesRHS);
3465 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
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 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3474 return V;
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 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3480 return V;
3482 return nullptr;
3485 Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3486 const SimplifyQuery &Q) {
3487 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3490 /// Given operands for an FCmpInst, see if we can fold the result.
3491 /// If not, this returns null.
3492 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!");
3498 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3499 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3500 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3502 // If we have a constant, make sure it is on the RHS.
3503 std::swap(LHS, RHS);
3504 Pred = CmpInst::getSwappedPredicate(Pred);
3507 // Fold trivial predicates.
3508 Type *RetTy = GetCompareTy(LHS);
3509 if (Pred == FCmpInst::FCMP_FALSE)
3510 return getFalse(RetTy);
3511 if (Pred == FCmpInst::FCMP_TRUE)
3512 return getTrue(RetTy);
3514 // Fold (un)ordered comparison if we can determine there are no NaNs.
3515 if (Pred == FCmpInst::FCMP_UNO || Pred == FCmpInst::FCMP_ORD)
3516 if (FMF.noNaNs() ||
3517 (isKnownNeverNaN(LHS, Q.TLI) && isKnownNeverNaN(RHS, Q.TLI)))
3518 return ConstantInt::get(RetTy, Pred == FCmpInst::FCMP_ORD);
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 if (match(RHS, m_NaN()))
3524 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3526 // fcmp pred x, undef and fcmp pred undef, x
3527 // fold to true if unordered, false if ordered
3528 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 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3534 // fcmp x,x -> true/false. Not all compares are foldable.
3535 if (LHS == RHS) {
3536 if (CmpInst::isTrueWhenEqual(Pred))
3537 return getTrue(RetTy);
3538 if (CmpInst::isFalseWhenEqual(Pred))
3539 return getFalse(RetTy);
3542 // Handle fcmp with constant RHS.
3543 const APFloat *C;
3544 if (match(RHS, m_APFloat(C))) {
3545 // Check whether the constant is an infinity.
3546 if (C->isInfinity()) {
3547 if (C->isNegative()) {
3548 switch (Pred) {
3549 case FCmpInst::FCMP_OLT:
3550 // No value is ordered and less than negative infinity.
3551 return getFalse(RetTy);
3552 case FCmpInst::FCMP_UGE:
3553 // All values are unordered with or at least negative infinity.
3554 return getTrue(RetTy);
3555 default:
3556 break;
3558 } else {
3559 switch (Pred) {
3560 case FCmpInst::FCMP_OGT:
3561 // No value is ordered and greater than infinity.
3562 return getFalse(RetTy);
3563 case FCmpInst::FCMP_ULE:
3564 // All values are unordered with and at most infinity.
3565 return getTrue(RetTy);
3566 default:
3567 break;
3571 if (C->isZero()) {
3572 switch (Pred) {
3573 case FCmpInst::FCMP_UGE:
3574 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3575 return getTrue(RetTy);
3576 break;
3577 case FCmpInst::FCMP_OLT:
3578 // X < 0
3579 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3580 return getFalse(RetTy);
3581 break;
3582 default:
3583 break;
3585 } 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 case FCmpInst::FCMP_UGE:
3591 case FCmpInst::FCMP_UGT:
3592 case FCmpInst::FCMP_UNE:
3593 // (X >= 0) implies (X > C) when (C < 0)
3594 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3595 return getTrue(RetTy);
3596 break;
3597 case FCmpInst::FCMP_OEQ:
3598 case FCmpInst::FCMP_OLE:
3599 case FCmpInst::FCMP_OLT:
3600 // (X >= 0) implies !(X < C) when (C < 0)
3601 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3602 return getFalse(RetTy);
3603 break;
3604 default:
3605 break;
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 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3614 return V;
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 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3620 return V;
3622 return nullptr;
3625 Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3626 FastMathFlags FMF, const SimplifyQuery &Q) {
3627 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3630 /// See if V simplifies when its operand Op is replaced with RepOp.
3631 static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3632 const SimplifyQuery &Q,
3633 unsigned MaxRecurse) {
3634 // Trivial replacement.
3635 if (V == Op)
3636 return RepOp;
3638 // We cannot replace a constant, and shouldn't even try.
3639 if (isa<Constant>(Op))
3640 return nullptr;
3642 auto *I = dyn_cast<Instruction>(V);
3643 if (!I)
3644 return nullptr;
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
3653 // We can't replace %sel with %add unless we strip away the flags.
3654 if (isa<OverflowingBinaryOperator>(B))
3655 if (Q.IIQ.hasNoSignedWrap(B) || Q.IIQ.hasNoUnsignedWrap(B))
3656 return nullptr;
3657 if (isa<PossiblyExactOperator>(B) && Q.IIQ.isExact(B))
3658 return nullptr;
3660 if (MaxRecurse) {
3661 if (B->getOperand(0) == Op)
3662 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3663 MaxRecurse - 1);
3664 if (B->getOperand(1) == Op)
3665 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3666 MaxRecurse - 1);
3670 // Same for CmpInsts.
3671 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3672 if (MaxRecurse) {
3673 if (C->getOperand(0) == Op)
3674 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3675 MaxRecurse - 1);
3676 if (C->getOperand(1) == Op)
3677 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3678 MaxRecurse - 1);
3682 // Same for GEPs.
3683 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
3684 if (MaxRecurse) {
3685 SmallVector<Value *, 8> NewOps(GEP->getNumOperands());
3686 transform(GEP->operands(), NewOps.begin(),
3687 [&](Value *V) { return V == Op ? RepOp : V; });
3688 return SimplifyGEPInst(GEP->getSourceElementType(), NewOps, Q,
3689 MaxRecurse - 1);
3693 // TODO: We could hand off more cases to instsimplify here.
3695 // If all operands are constant after substituting Op for RepOp then we can
3696 // constant fold the instruction.
3697 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3698 // Build a list of all constant operands.
3699 SmallVector<Constant *, 8> ConstOps;
3700 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3701 if (I->getOperand(i) == Op)
3702 ConstOps.push_back(CRepOp);
3703 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3704 ConstOps.push_back(COp);
3705 else
3706 break;
3709 // All operands were constants, fold it.
3710 if (ConstOps.size() == I->getNumOperands()) {
3711 if (CmpInst *C = dyn_cast<CmpInst>(I))
3712 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3713 ConstOps[1], Q.DL, Q.TLI);
3715 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3716 if (!LI->isVolatile())
3717 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3719 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3723 return nullptr;
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 static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3729 const APInt *Y, bool TrueWhenUnset) {
3730 const APInt *C;
3732 // (X & Y) == 0 ? X & ~Y : X --> X
3733 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3734 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3735 *Y == ~*C)
3736 return TrueWhenUnset ? FalseVal : TrueVal;
3738 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3739 // (X & Y) != 0 ? X : X & ~Y --> X
3740 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3741 *Y == ~*C)
3742 return TrueWhenUnset ? FalseVal : TrueVal;
3744 if (Y->isPowerOf2()) {
3745 // (X & Y) == 0 ? X | Y : X --> X | Y
3746 // (X & Y) != 0 ? X | Y : X --> X
3747 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3748 *Y == *C)
3749 return TrueWhenUnset ? TrueVal : FalseVal;
3751 // (X & Y) == 0 ? X : X | Y --> X
3752 // (X & Y) != 0 ? X : X | Y --> X | Y
3753 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3754 *Y == *C)
3755 return TrueWhenUnset ? TrueVal : FalseVal;
3758 return nullptr;
3761 /// An alternative way to test if a bit is set or not uses sgt/slt instead of
3762 /// eq/ne.
3763 static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3764 ICmpInst::Predicate Pred,
3765 Value *TrueVal, Value *FalseVal) {
3766 Value *X;
3767 APInt Mask;
3768 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3769 return nullptr;
3771 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3772 Pred == ICmpInst::ICMP_EQ);
3775 /// Try to simplify a select instruction when its condition operand is an
3776 /// integer comparison.
3777 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 return nullptr;
3785 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3786 Value *X;
3787 const APInt *Y;
3788 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3789 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3790 Pred == ICmpInst::ICMP_EQ))
3791 return V;
3794 // Check for other compares that behave like bit test.
3795 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3796 TrueVal, FalseVal))
3797 return V;
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 if (Pred == ICmpInst::ICMP_EQ) {
3803 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3804 TrueVal ||
3805 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3806 TrueVal)
3807 return FalseVal;
3808 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3809 FalseVal ||
3810 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3811 FalseVal)
3812 return FalseVal;
3813 } else if (Pred == ICmpInst::ICMP_NE) {
3814 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3815 FalseVal ||
3816 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3817 FalseVal)
3818 return TrueVal;
3819 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3820 TrueVal ||
3821 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3822 TrueVal)
3823 return TrueVal;
3826 return nullptr;
3829 /// Given operands for a SelectInst, see if we can fold the result.
3830 /// If not, this returns null.
3831 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 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3838 // select undef, X, Y -> X or Y
3839 if (isa<UndefValue>(CondC))
3840 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3842 // TODO: Vector constants with undef elements don't simplify.
3844 // select true, X, Y -> X
3845 if (CondC->isAllOnesValue())
3846 return TrueVal;
3847 // select false, X, Y -> Y
3848 if (CondC->isNullValue())
3849 return FalseVal;
3852 // select ?, X, X -> X
3853 if (TrueVal == FalseVal)
3854 return TrueVal;
3856 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3857 return FalseVal;
3858 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3859 return TrueVal;
3861 if (Value *V =
3862 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3863 return V;
3865 if (Value *V = foldSelectWithBinaryOp(Cond, TrueVal, FalseVal))
3866 return V;
3868 return nullptr;
3871 Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3872 const SimplifyQuery &Q) {
3873 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3876 /// Given operands for an GetElementPtrInst, see if we can fold the result.
3877 /// If not, this returns null.
3878 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 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3884 // getelementptr P -> P.
3885 if (Ops.size() == 1)
3886 return Ops[0];
3888 // Compute the (pointer) type returned by the GEP instruction.
3889 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3890 Type *GEPTy = PointerType::get(LastType, AS);
3891 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3892 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3893 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3894 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3896 if (isa<UndefValue>(Ops[0]))
3897 return UndefValue::get(GEPTy);
3899 if (Ops.size() == 2) {
3900 // getelementptr P, 0 -> P.
3901 if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy)
3902 return Ops[0];
3904 Type *Ty = SrcTy;
3905 if (Ty->isSized()) {
3906 Value *P;
3907 uint64_t C;
3908 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3909 // getelementptr P, N -> P if P points to a type of zero size.
3910 if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy)
3911 return Ops[0];
3913 // The following transforms are only safe if the ptrtoint cast
3914 // doesn't truncate the pointers.
3915 if (Ops[1]->getType()->getScalarSizeInBits() ==
3916 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;
3927 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3928 if (TyAllocSize == 1 &&
3929 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3930 if (Value *R = PtrToIntOrZero(P))
3931 return R;
3933 // getelementptr V, (ashr (sub P, V), C) -> Q
3934 // if P points to a type of size 1 << C.
3935 if (match(Ops[1],
3936 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3937 m_ConstantInt(C))) &&
3938 TyAllocSize == 1ULL << C)
3939 if (Value *R = PtrToIntOrZero(P))
3940 return R;
3942 // getelementptr V, (sdiv (sub P, V), C) -> Q
3943 // if P points to a type of size C.
3944 if (match(Ops[1],
3945 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3946 m_SpecificInt(TyAllocSize))))
3947 if (Value *R = PtrToIntOrZero(P))
3948 return R;
3953 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 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3958 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3959 APInt BasePtrOffset(IdxWidth, 0);
3960 Value *StrippedBasePtr =
3961 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3962 BasePtrOffset);
3964 // gep (gep V, C), (sub 0, V) -> C
3965 if (match(Ops.back(),
3966 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3967 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3968 return ConstantExpr::getIntToPtr(CI, GEPTy);
3970 // gep (gep V, C), (xor V, -1) -> C-1
3971 if (match(Ops.back(),
3972 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3973 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3974 return ConstantExpr::getIntToPtr(CI, GEPTy);
3979 // Check to see if this is constant foldable.
3980 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3981 return nullptr;
3983 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3984 Ops.slice(1));
3985 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3986 return CEFolded;
3987 return CE;
3990 Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3991 const SimplifyQuery &Q) {
3992 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3995 /// Given operands for an InsertValueInst, see if we can fold the result.
3996 /// If not, this returns null.
3997 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 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
4004 // insertvalue x, undef, n -> x
4005 if (match(Val, m_Undef()))
4006 return Agg;
4008 // insertvalue x, (extractvalue y, n), n
4009 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
4010 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
4011 EV->getIndices() == Idxs) {
4012 // insertvalue undef, (extractvalue y, n), n -> y
4013 if (match(Agg, m_Undef()))
4014 return EV->getAggregateOperand();
4016 // insertvalue y, (extractvalue y, n), n -> y
4017 if (Agg == EV->getAggregateOperand())
4018 return Agg;
4021 return nullptr;
4024 Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
4025 ArrayRef<unsigned> Idxs,
4026 const SimplifyQuery &Q) {
4027 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
4030 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 if (VecC && ValC && IdxC)
4037 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
4039 // Fold into undef if index is out of bounds.
4040 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
4041 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
4042 if (CI->uge(NumElements))
4043 return UndefValue::get(Vec->getType());
4046 // If index is undef, it might be out of bounds (see above case)
4047 if (isa<UndefValue>(Idx))
4048 return UndefValue::get(Vec->getType());
4050 return nullptr;
4053 /// Given operands for an ExtractValueInst, see if we can fold the result.
4054 /// If not, this returns null.
4055 static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4056 const SimplifyQuery &, unsigned) {
4057 if (auto *CAgg = dyn_cast<Constant>(Agg))
4058 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
4060 // extractvalue x, (insertvalue y, elt, n), n -> elt
4061 unsigned NumIdxs = Idxs.size();
4062 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
4063 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
4064 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
4065 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
4066 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
4067 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
4068 Idxs.slice(0, NumCommonIdxs)) {
4069 if (NumIdxs == NumInsertValueIdxs)
4070 return IVI->getInsertedValueOperand();
4071 break;
4075 return nullptr;
4078 Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
4079 const SimplifyQuery &Q) {
4080 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
4083 /// Given operands for an ExtractElementInst, see if we can fold the result.
4084 /// If not, this returns null.
4085 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 return ConstantFoldExtractElementInstruction(CVec, CIdx);
4091 // The index is not relevant if our vector is a splat.
4092 if (auto *Splat = CVec->getSplatValue())
4093 return Splat;
4095 if (isa<UndefValue>(Vec))
4096 return UndefValue::get(Vec->getType()->getVectorElementType());
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 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
4103 // definitely out of bounds, thus undefined result
4104 return UndefValue::get(Vec->getType()->getVectorElementType());
4105 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
4106 return Elt;
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 if (isa<UndefValue>(Idx))
4112 return UndefValue::get(Vec->getType()->getVectorElementType());
4114 return nullptr;
4117 Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
4118 const SimplifyQuery &Q) {
4119 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
4122 /// See if we can fold the given phi. If not, returns null.
4123 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 for (Value *Incoming : PN->incoming_values()) {
4129 // If the incoming value is the phi node itself, it can safely be skipped.
4130 if (Incoming == PN) continue;
4131 if (isa<UndefValue>(Incoming)) {
4132 // Remember that we saw an undef value, but otherwise ignore them.
4133 HasUndefInput = true;
4134 continue;
4136 if (CommonValue && Incoming != CommonValue)
4137 return nullptr; // Not the same, bail out.
4138 CommonValue = Incoming;
4141 // If CommonValue is null then all of the incoming values were either undef or
4142 // equal to the phi node itself.
4143 if (!CommonValue)
4144 return UndefValue::get(PN->getType());
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 if (HasUndefInput)
4150 return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
4152 return CommonValue;
4155 static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4156 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4157 if (auto *C = dyn_cast<Constant>(Op))
4158 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4160 if (auto *CI = dyn_cast<CastInst>(Op)) {
4161 auto *Src = CI->getOperand(0);
4162 Type *SrcTy = Src->getType();
4163 Type *MidTy = CI->getType();
4164 Type *DstTy = Ty;
4165 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 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4170 Type *MidIntPtrTy =
4171 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4172 Type *DstIntPtrTy =
4173 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4174 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4175 SrcIntPtrTy, MidIntPtrTy,
4176 DstIntPtrTy) == Instruction::BitCast)
4177 return Src;
4181 // bitcast x -> x
4182 if (CastOpc == Instruction::BitCast)
4183 if (Op->getType() == Ty)
4184 return Op;
4186 return nullptr;
4189 Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4190 const SimplifyQuery &Q) {
4191 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
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 static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4198 int MaskVal, Value *RootVec,
4199 unsigned MaxRecurse) {
4200 if (!MaxRecurse--)
4201 return nullptr;
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 if (MaskVal == -1)
4206 return nullptr;
4208 // The mask value chooses which source operand we need to look at next.
4209 int InVecNumElts = Op0->getType()->getVectorNumElements();
4210 int RootElt = MaskVal;
4211 Value *SourceOp = Op0;
4212 if (MaskVal >= InVecNumElts) {
4213 RootElt = MaskVal - InVecNumElts;
4214 SourceOp = Op1;
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 return foldIdentityShuffles(
4221 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4222 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4225 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4226 // size?
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 if (!RootVec)
4231 RootVec = SourceOp;
4233 // Give up as soon as a source operand does not match the existing root value.
4234 if (RootVec != SourceOp)
4235 return nullptr;
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 if (RootElt != DestElt)
4240 return nullptr;
4242 return RootVec;
4245 static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4246 Type *RetTy, const SimplifyQuery &Q,
4247 unsigned MaxRecurse) {
4248 if (isa<UndefValue>(Mask))
4249 return UndefValue::get(RetTy);
4251 Type *InVecTy = Op0->getType();
4252 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4253 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4255 SmallVector<int, 32> Indices;
4256 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4257 assert(MaskNumElts == Indices.size() &&
4258 "Size of Indices not same as number of mask elements?");
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 for (unsigned i = 0; i != MaskNumElts; ++i) {
4264 if (Indices[i] == -1)
4265 continue;
4266 if ((unsigned)Indices[i] < InVecNumElts)
4267 MaskSelects0 = true;
4268 else
4269 MaskSelects1 = true;
4271 if (!MaskSelects0)
4272 Op0 = UndefValue::get(InVecTy);
4273 if (!MaskSelects1)
4274 Op1 = UndefValue::get(InVecTy);
4276 auto *Op0Const = dyn_cast<Constant>(Op0);
4277 auto *Op1Const = dyn_cast<Constant>(Op1);
4279 // If all operands are constant, constant fold the shuffle.
4280 if (Op0Const && Op1Const)
4281 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4283 // Canonicalization: if only one input vector is constant, it shall be the
4284 // second one.
4285 if (Op0Const && !Op1Const) {
4286 std::swap(Op0, Op1);
4287 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
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 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4294 OpShuf->getMask()->getSplatValue())
4295 return Op0;
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 if (find(Indices, -1) != Indices.end())
4301 return nullptr;
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 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 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4314 // We can't replace a widening/narrowing shuffle with one of its operands.
4315 if (!RootVec || RootVec->getType() != RetTy)
4316 return nullptr;
4318 return RootVec;
4321 /// Given operands for a ShuffleVectorInst, fold the result or return null.
4322 Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4323 Type *RetTy, const SimplifyQuery &Q) {
4324 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4327 static Constant *propagateNaN(Constant *In) {
4328 // If the input is a vector with undef elements, just return a default NaN.
4329 if (!In->isNaN())
4330 return ConstantFP::getNaN(In->getType());
4332 // Propagate the existing NaN constant when possible.
4333 // TODO: Should we quiet a signaling NaN?
4334 return In;
4337 static Constant *simplifyFPBinop(Value *Op0, Value *Op1) {
4338 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4339 return ConstantFP::getNaN(Op0->getType());
4341 if (match(Op0, m_NaN()))
4342 return propagateNaN(cast<Constant>(Op0));
4343 if (match(Op1, m_NaN()))
4344 return propagateNaN(cast<Constant>(Op1));
4346 return nullptr;
4349 /// Given operands for an FAdd, see if we can fold the result. If not, this
4350 /// returns null.
4351 static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4352 const SimplifyQuery &Q, unsigned MaxRecurse) {
4353 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4354 return C;
4356 if (Constant *C = simplifyFPBinop(Op0, Op1))
4357 return C;
4359 // fadd X, -0 ==> X
4360 if (match(Op1, m_NegZeroFP()))
4361 return Op0;
4363 // fadd X, 0 ==> X, when we know X is not -0
4364 if (match(Op1, m_PosZeroFP()) &&
4365 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4366 return Op0;
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 if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) ||
4376 match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0)))))
4377 return ConstantFP::getNullValue(Op0->getType());
4379 // (X - Y) + Y --> X
4380 // Y + (X - Y) --> X
4381 Value *X;
4382 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4383 (match(Op0, m_FSub(m_Value(X), m_Specific(Op1))) ||
4384 match(Op1, m_FSub(m_Value(X), m_Specific(Op0)))))
4385 return X;
4387 return nullptr;
4390 /// Given operands for an FSub, see if we can fold the result. If not, this
4391 /// returns null.
4392 static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4393 const SimplifyQuery &Q, unsigned MaxRecurse) {
4394 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4395 return C;
4397 if (Constant *C = simplifyFPBinop(Op0, Op1))
4398 return C;
4400 // fsub X, +0 ==> X
4401 if (match(Op1, m_PosZeroFP()))
4402 return Op0;
4404 // fsub X, -0 ==> X, when we know X is not -0
4405 if (match(Op1, m_NegZeroFP()) &&
4406 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4407 return Op0;
4409 // fsub -0.0, (fsub -0.0, X) ==> X
4410 Value *X;
4411 if (match(Op0, m_NegZeroFP()) &&
4412 match(Op1, m_FSub(m_NegZeroFP(), m_Value(X))))
4413 return X;
4415 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4416 if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
4417 match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))))
4418 return X;
4420 // fsub nnan x, x ==> 0.0
4421 if (FMF.noNaNs() && Op0 == Op1)
4422 return Constant::getNullValue(Op0->getType());
4424 // Y - (Y - X) --> X
4425 // (X + Y) - Y --> X
4426 if (FMF.noSignedZeros() && FMF.allowReassoc() &&
4427 (match(Op1, m_FSub(m_Specific(Op0), m_Value(X))) ||
4428 match(Op0, m_c_FAdd(m_Specific(Op1), m_Value(X)))))
4429 return X;
4431 return nullptr;
4434 /// Given the operands for an FMul, see if we can fold the result
4435 static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4436 const SimplifyQuery &Q, unsigned MaxRecurse) {
4437 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4438 return C;
4440 if (Constant *C = simplifyFPBinop(Op0, Op1))
4441 return C;
4443 // fmul X, 1.0 ==> X
4444 if (match(Op1, m_FPOne()))
4445 return Op0;
4447 // fmul nnan nsz X, 0 ==> 0
4448 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP()))
4449 return ConstantFP::getNullValue(Op0->getType());
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 if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) &&
4457 FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros())
4458 return X;
4460 return nullptr;
4463 Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4464 const SimplifyQuery &Q) {
4465 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4469 Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4470 const SimplifyQuery &Q) {
4471 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4474 Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4475 const SimplifyQuery &Q) {
4476 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4479 static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4480 const SimplifyQuery &Q, unsigned) {
4481 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4482 return C;
4484 if (Constant *C = simplifyFPBinop(Op0, Op1))
4485 return C;
4487 // X / 1.0 -> X
4488 if (match(Op1, m_FPOne()))
4489 return Op0;
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 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()))
4495 return ConstantFP::getNullValue(Op0->getType());
4497 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 if (Op0 == Op1)
4501 return ConstantFP::get(Op0->getType(), 1.0);
4503 // (X * Y) / Y --> X if we can reassociate to the above form.
4504 Value *X;
4505 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4506 return X;
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 if (match(Op0, m_FNegNSZ(m_Specific(Op1))) ||
4512 match(Op1, m_FNegNSZ(m_Specific(Op0))))
4513 return ConstantFP::get(Op0->getType(), -1.0);
4516 return nullptr;
4519 Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4520 const SimplifyQuery &Q) {
4521 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4524 static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4525 const SimplifyQuery &Q, unsigned) {
4526 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4527 return C;
4529 if (Constant *C = simplifyFPBinop(Op0, Op1))
4530 return C;
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 if (FMF.noNaNs()) {
4536 // +0 % X -> 0
4537 if (match(Op0, m_PosZeroFP()))
4538 return ConstantFP::getNullValue(Op0->getType());
4539 // -0 % X -> -0
4540 if (match(Op0, m_NegZeroFP()))
4541 return ConstantFP::getNegativeZero(Op0->getType());
4544 return nullptr;
4547 Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4548 const SimplifyQuery &Q) {
4549 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4552 //=== Helper functions for higher up the class hierarchy.
4554 /// Given operands for a BinaryOperator, see if we can fold the result.
4555 /// If not, this returns null.
4556 static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4557 const SimplifyQuery &Q, unsigned MaxRecurse) {
4558 switch (Opcode) {
4559 case Instruction::Add:
4560 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4561 case Instruction::Sub:
4562 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4563 case Instruction::Mul:
4564 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4565 case Instruction::SDiv:
4566 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4567 case Instruction::UDiv:
4568 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4569 case Instruction::SRem:
4570 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4571 case Instruction::URem:
4572 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4573 case Instruction::Shl:
4574 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4575 case Instruction::LShr:
4576 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4577 case Instruction::AShr:
4578 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4579 case Instruction::And:
4580 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4581 case Instruction::Or:
4582 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4583 case Instruction::Xor:
4584 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4585 case Instruction::FAdd:
4586 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4587 case Instruction::FSub:
4588 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4589 case Instruction::FMul:
4590 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4591 case Instruction::FDiv:
4592 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4593 case Instruction::FRem:
4594 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4595 default:
4596 llvm_unreachable("Unexpected opcode");
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 static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4605 const FastMathFlags &FMF, const SimplifyQuery &Q,
4606 unsigned MaxRecurse) {
4607 switch (Opcode) {
4608 case Instruction::FAdd:
4609 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4610 case Instruction::FSub:
4611 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4612 case Instruction::FMul:
4613 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4614 case Instruction::FDiv:
4615 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4616 default:
4617 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4621 Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4622 const SimplifyQuery &Q) {
4623 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4626 Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4627 FastMathFlags FMF, const SimplifyQuery &Q) {
4628 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4631 /// Given operands for a CmpInst, see if we can fold the result.
4632 static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4633 const SimplifyQuery &Q, unsigned MaxRecurse) {
4634 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4635 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4636 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4639 Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4640 const SimplifyQuery &Q) {
4641 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4644 static bool IsIdempotent(Intrinsic::ID ID) {
4645 switch (ID) {
4646 default: return false;
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;
4661 static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4662 const DataLayout &DL) {
4663 GlobalValue *PtrSym;
4664 APInt PtrOffset;
4665 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4666 return nullptr;
4668 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4669 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4670 Type *Int32PtrTy = Int32Ty->getPointerTo();
4671 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4673 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4674 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4675 return nullptr;
4677 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4678 if (OffsetInt % 4 != 0)
4679 return nullptr;
4681 Constant *C = ConstantExpr::getGetElementPtr(
4682 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4683 ConstantInt::get(Int64Ty, OffsetInt / 4));
4684 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4685 if (!Loaded)
4686 return nullptr;
4688 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4689 if (!LoadedCE)
4690 return nullptr;
4692 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4693 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4694 if (!LoadedCE)
4695 return nullptr;
4698 if (LoadedCE->getOpcode() != Instruction::Sub)
4699 return nullptr;
4701 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4702 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4703 return nullptr;
4704 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4706 Constant *LoadedRHS = LoadedCE->getOperand(1);
4707 GlobalValue *LoadedRHSSym;
4708 APInt LoadedRHSOffset;
4709 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4710 DL) ||
4711 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4712 return nullptr;
4714 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4717 static bool maskIsAllZeroOrUndef(Value *Mask) {
4718 auto *ConstMask = dyn_cast<Constant>(Mask);
4719 if (!ConstMask)
4720 return false;
4721 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4722 return true;
4723 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4724 ++I) {
4725 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4726 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4727 continue;
4728 return false;
4730 return true;
4733 static Value *simplifyUnaryIntrinsic(Function *F, Value *Op0,
4734 const SimplifyQuery &Q) {
4735 // Idempotent functions return the same result when called repeatedly.
4736 Intrinsic::ID IID = F->getIntrinsicID();
4737 if (IsIdempotent(IID))
4738 if (auto *II = dyn_cast<IntrinsicInst>(Op0))
4739 if (II->getIntrinsicID() == IID)
4740 return II;
4742 Value *X;
4743 switch (IID) {
4744 case Intrinsic::fabs:
4745 if (SignBitMustBeZero(Op0, Q.TLI)) return Op0;
4746 break;
4747 case Intrinsic::bswap:
4748 // bswap(bswap(x)) -> x
4749 if (match(Op0, m_BSwap(m_Value(X)))) return X;
4750 break;
4751 case Intrinsic::bitreverse:
4752 // bitreverse(bitreverse(x)) -> x
4753 if (match(Op0, m_BitReverse(m_Value(X)))) return X;
4754 break;
4755 case Intrinsic::exp:
4756 // exp(log(x)) -> x
4757 if (Q.CxtI->hasAllowReassoc() &&
4758 match(Op0, m_Intrinsic<Intrinsic::log>(m_Value(X)))) return X;
4759 break;
4760 case Intrinsic::exp2:
4761 // exp2(log2(x)) -> x
4762 if (Q.CxtI->hasAllowReassoc() &&
4763 match(Op0, m_Intrinsic<Intrinsic::log2>(m_Value(X)))) return X;
4764 break;
4765 case Intrinsic::log:
4766 // log(exp(x)) -> x
4767 if (Q.CxtI->hasAllowReassoc() &&
4768 match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X)))) return X;
4769 break;
4770 case Intrinsic::log2:
4771 // log2(exp2(x)) -> x
4772 if (Q.CxtI->hasAllowReassoc() &&
4773 match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) return X;
4774 break;
4775 default:
4776 break;
4779 return nullptr;
4782 static Value *simplifyBinaryIntrinsic(Function *F, Value *Op0, Value *Op1,
4783 const SimplifyQuery &Q) {
4784 Intrinsic::ID IID = F->getIntrinsicID();
4785 Type *ReturnType = F->getReturnType();
4786 switch (IID) {
4787 case Intrinsic::usub_with_overflow:
4788 case Intrinsic::ssub_with_overflow:
4789 // X - X -> { 0, false }
4790 if (Op0 == Op1)
4791 return Constant::getNullValue(ReturnType);
4792 // X - undef -> undef
4793 // undef - X -> undef
4794 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4795 return UndefValue::get(ReturnType);
4796 break;
4797 case Intrinsic::uadd_with_overflow:
4798 case Intrinsic::sadd_with_overflow:
4799 // X + undef -> undef
4800 if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1))
4801 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 if (match(Op0, m_Zero()) || match(Op1, m_Zero()))
4808 return Constant::getNullValue(ReturnType);
4809 // undef * X -> { 0, false }
4810 // X * undef -> { 0, false }
4811 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
4812 return Constant::getNullValue(ReturnType);
4813 break;
4814 case Intrinsic::load_relative:
4815 if (auto *C0 = dyn_cast<Constant>(Op0))
4816 if (auto *C1 = dyn_cast<Constant>(Op1))
4817 return SimplifyRelativeLoad(C0, C1, Q.DL);
4818 break;
4819 case Intrinsic::powi:
4820 if (auto *Power = dyn_cast<ConstantInt>(Op1)) {
4821 // powi(x, 0) -> 1.0
4822 if (Power->isZero())
4823 return ConstantFP::get(Op0->getType(), 1.0);
4824 // powi(x, 1) -> x
4825 if (Power->isOne())
4826 return Op0;
4828 break;
4829 case Intrinsic::maxnum:
4830 case Intrinsic::minnum: {
4831 // If the arguments are the same, this is a no-op.
4832 if (Op0 == Op1) return Op0;
4834 // If one argument is NaN or undef, return the other argument.
4835 if (match(Op0, m_CombineOr(m_NaN(), m_Undef()))) return Op1;
4836 if (match(Op1, m_CombineOr(m_NaN(), m_Undef()))) return Op0;
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 if (M0->getIntrinsicID() == IID &&
4842 (M0->getOperand(0) == Op1 || M0->getOperand(1) == Op1))
4843 return Op0;
4844 if (auto *M1 = dyn_cast<IntrinsicInst>(Op1))
4845 if (M1->getIntrinsicID() == IID &&
4846 (M1->getOperand(0) == Op0 || M1->getOperand(1) == Op0))
4847 return Op1;
4849 // minnum(X, -Inf) --> -Inf (and commuted variant)
4850 // maxnum(X, +Inf) --> +Inf (and commuted variant)
4851 bool UseNegInf = IID == Intrinsic::minnum;
4852 const APFloat *C;
4853 if ((match(Op0, m_APFloat(C)) && C->isInfinity() &&
4854 C->isNegative() == UseNegInf) ||
4855 (match(Op1, m_APFloat(C)) && C->isInfinity() &&
4856 C->isNegative() == UseNegInf))
4857 return ConstantFP::getInfinity(ReturnType, UseNegInf);
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 break;
4865 default:
4866 break;
4869 return nullptr;
4872 template <typename IterTy>
4873 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 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4877 if (NumOperands == 0)
4878 return nullptr;
4880 Intrinsic::ID IID = F->getIntrinsicID();
4881 if (NumOperands == 1)
4882 return simplifyUnaryIntrinsic(F, ArgBegin[0], Q);
4884 if (NumOperands == 2)
4885 return simplifyBinaryIntrinsic(F, ArgBegin[0], ArgBegin[1], Q);
4887 // Handle intrinsics with 3 or more arguments.
4888 switch (IID) {
4889 case Intrinsic::masked_load: {
4890 Value *MaskArg = ArgBegin[2];
4891 Value *PassthruArg = ArgBegin[3];
4892 // If the mask is all zeros or undef, the "passthru" argument is the result.
4893 if (maskIsAllZeroOrUndef(MaskArg))
4894 return PassthruArg;
4895 return nullptr;
4897 case Intrinsic::fshl:
4898 case Intrinsic::fshr: {
4899 Value *ShAmtArg = ArgBegin[2];
4900 const APInt *ShAmtC;
4901 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 APInt BitWidth = APInt(ShAmtC->getBitWidth(), ShAmtC->getBitWidth());
4905 if (ShAmtC->urem(BitWidth).isNullValue())
4906 return ArgBegin[IID == Intrinsic::fshl ? 0 : 1];
4908 return nullptr;
4910 default:
4911 return nullptr;
4915 template <typename IterTy>
4916 static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4917 IterTy ArgEnd, const SimplifyQuery &Q,
4918 unsigned MaxRecurse) {
4919 Type *Ty = V->getType();
4920 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4921 Ty = PTy->getElementType();
4922 FunctionType *FTy = cast<FunctionType>(Ty);
4924 // call undef -> undef
4925 // call null -> undef
4926 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4927 return UndefValue::get(FTy->getReturnType());
4929 Function *F = dyn_cast<Function>(V);
4930 if (!F)
4931 return nullptr;
4933 if (F->isIntrinsic())
4934 if (Value *Ret = simplifyIntrinsic(F, ArgBegin, ArgEnd, Q))
4935 return Ret;
4937 if (!canConstantFoldCallTo(CS, F))
4938 return nullptr;
4940 SmallVector<Constant *, 4> ConstantArgs;
4941 ConstantArgs.reserve(ArgEnd - ArgBegin);
4942 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4943 Constant *C = dyn_cast<Constant>(*I);
4944 if (!C)
4945 return nullptr;
4946 ConstantArgs.push_back(C);
4949 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4952 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4953 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4954 const SimplifyQuery &Q) {
4955 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4958 Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4959 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4960 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4963 Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4964 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4965 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4966 Q, RecursionLimit);
4969 /// See if we can compute a simplified version of this instruction.
4970 /// If not, this returns null.
4972 Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4973 OptimizationRemarkEmitter *ORE) {
4974 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4975 Value *Result;
4977 switch (I->getOpcode()) {
4978 default:
4979 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4980 break;
4981 case Instruction::FAdd:
4982 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4983 I->getFastMathFlags(), Q);
4984 break;
4985 case Instruction::Add:
4986 Result =
4987 SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4988 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 Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
4999 Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
5000 break;
5001 case Instruction::FMul:
5002 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
5003 I->getFastMathFlags(), Q);
5004 break;
5005 case Instruction::Mul:
5006 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
5007 break;
5008 case Instruction::SDiv:
5009 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 break;
5014 case Instruction::FDiv:
5015 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
5016 I->getFastMathFlags(), Q);
5017 break;
5018 case Instruction::SRem:
5019 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
5020 break;
5021 case Instruction::URem:
5022 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
5023 break;
5024 case Instruction::FRem:
5025 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
5026 I->getFastMathFlags(), Q);
5027 break;
5028 case Instruction::Shl:
5029 Result =
5030 SimplifyShlInst(I->getOperand(0), I->getOperand(1),
5031 Q.IIQ.hasNoSignedWrap(cast<BinaryOperator>(I)),
5032 Q.IIQ.hasNoUnsignedWrap(cast<BinaryOperator>(I)), Q);
5033 break;
5034 case Instruction::LShr:
5035 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
5036 Q.IIQ.isExact(cast<BinaryOperator>(I)), Q);
5037 break;
5038 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 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
5044 break;
5045 case Instruction::Or:
5046 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
5047 break;
5048 case Instruction::Xor:
5049 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
5050 break;
5051 case Instruction::ICmp:
5052 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
5053 I->getOperand(0), I->getOperand(1), Q);
5054 break;
5055 case Instruction::FCmp:
5056 Result =
5057 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
5058 I->getOperand(1), I->getFastMathFlags(), Q);
5059 break;
5060 case Instruction::Select:
5061 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
5062 I->getOperand(2), Q);
5063 break;
5064 case Instruction::GetElementPtr: {
5065 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
5066 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
5067 Ops, Q);
5068 break;
5070 case Instruction::InsertValue: {
5071 InsertValueInst *IV = cast<InsertValueInst>(I);
5072 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
5073 IV->getInsertedValueOperand(),
5074 IV->getIndices(), Q);
5075 break;
5077 case Instruction::InsertElement: {
5078 auto *IE = cast<InsertElementInst>(I);
5079 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
5080 IE->getOperand(2), Q);
5081 break;
5083 case Instruction::ExtractValue: {
5084 auto *EVI = cast<ExtractValueInst>(I);
5085 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
5086 EVI->getIndices(), Q);
5087 break;
5089 case Instruction::ExtractElement: {
5090 auto *EEI = cast<ExtractElementInst>(I);
5091 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
5092 EEI->getIndexOperand(), Q);
5093 break;
5095 case Instruction::ShuffleVector: {
5096 auto *SVI = cast<ShuffleVectorInst>(I);
5097 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5098 SVI->getMask(), SVI->getType(), Q);
5099 break;
5101 case Instruction::PHI:
5102 Result = SimplifyPHINode(cast<PHINode>(I), Q);
5103 break;
5104 case Instruction::Call: {
5105 CallSite CS(cast<CallInst>(I));
5106 Result = SimplifyCall(CS, Q);
5107 break;
5109 #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
5110 #include "llvm/IR/Instruction.def"
5111 #undef HANDLE_CAST_INST
5112 Result =
5113 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
5114 break;
5115 case Instruction::Alloca:
5116 // No simplifications for Alloca and it can't be constant folded.
5117 Result = nullptr;
5118 break;
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 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
5125 if (Known.isConstant())
5126 Result = ConstantInt::get(I->getType(), Known.getConstant());
5129 /// If called on unreachable code, the above logic may report that the
5130 /// instruction simplified to itself. Make life easier for users by
5131 /// detecting that case here, returning a safe value instead.
5132 return Result == I ? UndefValue::get(I->getType()) : Result;
5135 /// Implementation of recursive simplification through an instruction's
5136 /// uses.
5138 /// This is the common implementation of the recursive simplification routines.
5139 /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
5140 /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
5141 /// instructions to process and attempt to simplify it using
5142 /// InstructionSimplify.
5144 /// This routine returns 'true' only when *it* simplifies something. The passed
5145 /// in simplified value does not count toward this.
5146 static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
5147 const TargetLibraryInfo *TLI,
5148 const DominatorTree *DT,
5149 AssumptionCache *AC) {
5150 bool Simplified = false;
5151 SmallSetVector<Instruction *, 8> Worklist;
5152 const DataLayout &DL = I->getModule()->getDataLayout();
5154 // If we have an explicit value to collapse to, do that round of the
5155 // simplification loop by hand initially.
5156 if (SimpleV) {
5157 for (User *U : I->users())
5158 if (U != I)
5159 Worklist.insert(cast<Instruction>(U));
5161 // Replace the instruction with its simplified value.
5162 I->replaceAllUsesWith(SimpleV);
5164 // Gracefully handle edge cases where the instruction is not wired into any
5165 // parent block.
5166 if (I->getParent() && !I->isEHPad() && !I->isTerminator() &&
5167 !I->mayHaveSideEffects())
5168 I->eraseFromParent();
5169 } else {
5170 Worklist.insert(I);
5173 // Note that we must test the size on each iteration, the worklist can grow.
5174 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
5175 I = Worklist[Idx];
5177 // See if this instruction simplifies.
5178 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
5179 if (!SimpleV)
5180 continue;
5182 Simplified = true;
5184 // Stash away all the uses of the old instruction so we can check them for
5185 // recursive simplifications after a RAUW. This is cheaper than checking all
5186 // uses of To on the recursive step in most cases.
5187 for (User *U : I->users())
5188 Worklist.insert(cast<Instruction>(U));
5190 // Replace the instruction with its simplified value.
5191 I->replaceAllUsesWith(SimpleV);
5193 // Gracefully handle edge cases where the instruction is not wired into any
5194 // parent block.
5195 if (I->getParent() && !I->isEHPad() && !I->isTerminator() &&
5196 !I->mayHaveSideEffects())
5197 I->eraseFromParent();
5199 return Simplified;
5202 bool llvm::recursivelySimplifyInstruction(Instruction *I,
5203 const TargetLibraryInfo *TLI,
5204 const DominatorTree *DT,
5205 AssumptionCache *AC) {
5206 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
5209 bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
5210 const TargetLibraryInfo *TLI,
5211 const DominatorTree *DT,
5212 AssumptionCache *AC) {
5213 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!");
5214 assert(SimpleV && "Must provide a simplified value.");
5215 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
5218 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 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
5225 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
5226 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5229 const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5230 const DataLayout &DL) {
5231 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5234 template <class T, class... TArgs>
5235 const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5236 Function &F) {
5237 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5238 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5239 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5240 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5242 template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,
5243 Function &);