Bug Summary

File:lib/Analysis/InstructionSimplify.cpp
Warning:line 1306, column 51
1st function call argument is an uninitialized value

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name InstructionSimplify.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-eagerly-assume -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-7/lib/clang/7.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-7~svn326551/build-llvm/lib/Analysis -I /build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis -I /build/llvm-toolchain-snapshot-7~svn326551/build-llvm/include -I /build/llvm-toolchain-snapshot-7~svn326551/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0/backward -internal-isystem /usr/include/clang/7.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-7/lib/clang/7.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-7~svn326551/build-llvm/lib/Analysis -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-checker optin.performance.Padding -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-03-02-155150-1477-1 -x c++ /build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp

/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp

1//===- InstructionSimplify.cpp - Fold instruction operands ----------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements routines for folding instructions into simpler forms
11// that do not require creating new instructions. This does constant folding
12// ("add i32 1, 1" -> "2") but can also handle non-constant operands, either
13// returning a constant ("and i32 %x, 0" -> "0") or an already existing value
14// ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been
15// simplified: This is usually true and assuming it simplifies the logic (if
16// they have not been simplified then results are correct but maybe suboptimal).
17//
18//===----------------------------------------------------------------------===//
19
20#include "llvm/Analysis/InstructionSimplify.h"
21#include "llvm/ADT/SetVector.h"
22#include "llvm/ADT/Statistic.h"
23#include "llvm/Analysis/AliasAnalysis.h"
24#include "llvm/Analysis/AssumptionCache.h"
25#include "llvm/Analysis/CaptureTracking.h"
26#include "llvm/Analysis/CmpInstAnalysis.h"
27#include "llvm/Analysis/ConstantFolding.h"
28#include "llvm/Analysis/LoopAnalysisManager.h"
29#include "llvm/Analysis/MemoryBuiltins.h"
30#include "llvm/Analysis/ValueTracking.h"
31#include "llvm/Analysis/VectorUtils.h"
32#include "llvm/IR/ConstantRange.h"
33#include "llvm/IR/DataLayout.h"
34#include "llvm/IR/Dominators.h"
35#include "llvm/IR/GetElementPtrTypeIterator.h"
36#include "llvm/IR/GlobalAlias.h"
37#include "llvm/IR/Operator.h"
38#include "llvm/IR/PatternMatch.h"
39#include "llvm/IR/ValueHandle.h"
40#include "llvm/Support/KnownBits.h"
41#include <algorithm>
42using namespace llvm;
43using namespace llvm::PatternMatch;
44
45#define DEBUG_TYPE"instsimplify" "instsimplify"
46
47enum { RecursionLimit = 3 };
48
49STATISTIC(NumExpand, "Number of expansions")static llvm::Statistic NumExpand = {"instsimplify", "NumExpand"
, "Number of expansions", {0}, {false}}
;
50STATISTIC(NumReassoc, "Number of reassociations")static llvm::Statistic NumReassoc = {"instsimplify", "NumReassoc"
, "Number of reassociations", {0}, {false}}
;
51
52static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned);
53static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &,
54 unsigned);
55static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &,
56 const SimplifyQuery &, unsigned);
57static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &,
58 unsigned);
59static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
60 const SimplifyQuery &Q, unsigned MaxRecurse);
61static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned);
62static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned);
63static Value *SimplifyCastInst(unsigned, Value *, Type *,
64 const SimplifyQuery &, unsigned);
65
66/// For a boolean type or a vector of boolean type, return false or a vector
67/// with every element false.
68static Constant *getFalse(Type *Ty) {
69 return ConstantInt::getFalse(Ty);
70}
71
72/// For a boolean type or a vector of boolean type, return true or a vector
73/// with every element true.
74static Constant *getTrue(Type *Ty) {
75 return ConstantInt::getTrue(Ty);
76}
77
78/// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"?
79static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS,
80 Value *RHS) {
81 CmpInst *Cmp = dyn_cast<CmpInst>(V);
82 if (!Cmp)
83 return false;
84 CmpInst::Predicate CPred = Cmp->getPredicate();
85 Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1);
86 if (CPred == Pred && CLHS == LHS && CRHS == RHS)
87 return true;
88 return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS &&
89 CRHS == LHS;
90}
91
92/// Does the given value dominate the specified phi node?
93static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) {
94 Instruction *I = dyn_cast<Instruction>(V);
95 if (!I)
96 // Arguments and constants dominate all instructions.
97 return true;
98
99 // If we are processing instructions (and/or basic blocks) that have not been
100 // fully added to a function, the parent nodes may still be null. Simply
101 // return the conservative answer in these cases.
102 if (!I->getParent() || !P->getParent() || !I->getParent()->getParent())
103 return false;
104
105 // If we have a DominatorTree then do a precise test.
106 if (DT)
107 return DT->dominates(I, P);
108
109 // Otherwise, if the instruction is in the entry block and is not an invoke,
110 // then it obviously dominates all phi nodes.
111 if (I->getParent() == &I->getParent()->getParent()->getEntryBlock() &&
112 !isa<InvokeInst>(I))
113 return true;
114
115 return false;
116}
117
118/// Simplify "A op (B op' C)" by distributing op over op', turning it into
119/// "(A op B) op' (A op C)". Here "op" is given by Opcode and "op'" is
120/// given by OpcodeToExpand, while "A" corresponds to LHS and "B op' C" to RHS.
121/// Also performs the transform "(A op' B) op C" -> "(A op C) op' (B op C)".
122/// Returns the simplified value, or null if no simplification was performed.
123static Value *ExpandBinOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS,
124 Instruction::BinaryOps OpcodeToExpand,
125 const SimplifyQuery &Q, unsigned MaxRecurse) {
126 // Recursion is always used, so bail out at once if we already hit the limit.
127 if (!MaxRecurse--)
128 return nullptr;
129
130 // Check whether the expression has the form "(A op' B) op C".
131 if (BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS))
132 if (Op0->getOpcode() == OpcodeToExpand) {
133 // It does! Try turning it into "(A op C) op' (B op C)".
134 Value *A = Op0->getOperand(0), *B = Op0->getOperand(1), *C = RHS;
135 // Do "A op C" and "B op C" both simplify?
136 if (Value *L = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse))
137 if (Value *R = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
138 // They do! Return "L op' R" if it simplifies or is already available.
139 // If "L op' R" equals "A op' B" then "L op' R" is just the LHS.
140 if ((L == A && R == B) || (Instruction::isCommutative(OpcodeToExpand)
141 && L == B && R == A)) {
142 ++NumExpand;
143 return LHS;
144 }
145 // Otherwise return "L op' R" if it simplifies.
146 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
147 ++NumExpand;
148 return V;
149 }
150 }
151 }
152
153 // Check whether the expression has the form "A op (B op' C)".
154 if (BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS))
155 if (Op1->getOpcode() == OpcodeToExpand) {
156 // It does! Try turning it into "(A op B) op' (A op C)".
157 Value *A = LHS, *B = Op1->getOperand(0), *C = Op1->getOperand(1);
158 // Do "A op B" and "A op C" both simplify?
159 if (Value *L = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse))
160 if (Value *R = SimplifyBinOp(Opcode, A, C, Q, MaxRecurse)) {
161 // They do! Return "L op' R" if it simplifies or is already available.
162 // If "L op' R" equals "B op' C" then "L op' R" is just the RHS.
163 if ((L == B && R == C) || (Instruction::isCommutative(OpcodeToExpand)
164 && L == C && R == B)) {
165 ++NumExpand;
166 return RHS;
167 }
168 // Otherwise return "L op' R" if it simplifies.
169 if (Value *V = SimplifyBinOp(OpcodeToExpand, L, R, Q, MaxRecurse)) {
170 ++NumExpand;
171 return V;
172 }
173 }
174 }
175
176 return nullptr;
177}
178
179/// Generic simplifications for associative binary operations.
180/// Returns the simpler value, or null if none was found.
181static Value *SimplifyAssociativeBinOp(Instruction::BinaryOps Opcode,
182 Value *LHS, Value *RHS,
183 const SimplifyQuery &Q,
184 unsigned MaxRecurse) {
185 assert(Instruction::isAssociative(Opcode) && "Not an associative operation!")(static_cast <bool> (Instruction::isAssociative(Opcode)
&& "Not an associative operation!") ? void (0) : __assert_fail
("Instruction::isAssociative(Opcode) && \"Not an associative operation!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 185, __extension__ __PRETTY_FUNCTION__))
;
186
187 // Recursion is always used, so bail out at once if we already hit the limit.
188 if (!MaxRecurse--)
189 return nullptr;
190
191 BinaryOperator *Op0 = dyn_cast<BinaryOperator>(LHS);
192 BinaryOperator *Op1 = dyn_cast<BinaryOperator>(RHS);
193
194 // Transform: "(A op B) op C" ==> "A op (B op C)" if it simplifies completely.
195 if (Op0 && Op0->getOpcode() == Opcode) {
196 Value *A = Op0->getOperand(0);
197 Value *B = Op0->getOperand(1);
198 Value *C = RHS;
199
200 // Does "B op C" simplify?
201 if (Value *V = SimplifyBinOp(Opcode, B, C, Q, MaxRecurse)) {
202 // It does! Return "A op V" if it simplifies or is already available.
203 // If V equals B then "A op V" is just the LHS.
204 if (V == B) return LHS;
205 // Otherwise return "A op V" if it simplifies.
206 if (Value *W = SimplifyBinOp(Opcode, A, V, Q, MaxRecurse)) {
207 ++NumReassoc;
208 return W;
209 }
210 }
211 }
212
213 // Transform: "A op (B op C)" ==> "(A op B) op C" if it simplifies completely.
214 if (Op1 && Op1->getOpcode() == Opcode) {
215 Value *A = LHS;
216 Value *B = Op1->getOperand(0);
217 Value *C = Op1->getOperand(1);
218
219 // Does "A op B" simplify?
220 if (Value *V = SimplifyBinOp(Opcode, A, B, Q, MaxRecurse)) {
221 // It does! Return "V op C" if it simplifies or is already available.
222 // If V equals B then "V op C" is just the RHS.
223 if (V == B) return RHS;
224 // Otherwise return "V op C" if it simplifies.
225 if (Value *W = SimplifyBinOp(Opcode, V, C, Q, MaxRecurse)) {
226 ++NumReassoc;
227 return W;
228 }
229 }
230 }
231
232 // The remaining transforms require commutativity as well as associativity.
233 if (!Instruction::isCommutative(Opcode))
234 return nullptr;
235
236 // Transform: "(A op B) op C" ==> "(C op A) op B" if it simplifies completely.
237 if (Op0 && Op0->getOpcode() == Opcode) {
238 Value *A = Op0->getOperand(0);
239 Value *B = Op0->getOperand(1);
240 Value *C = RHS;
241
242 // Does "C op A" simplify?
243 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
244 // It does! Return "V op B" if it simplifies or is already available.
245 // If V equals A then "V op B" is just the LHS.
246 if (V == A) return LHS;
247 // Otherwise return "V op B" if it simplifies.
248 if (Value *W = SimplifyBinOp(Opcode, V, B, Q, MaxRecurse)) {
249 ++NumReassoc;
250 return W;
251 }
252 }
253 }
254
255 // Transform: "A op (B op C)" ==> "B op (C op A)" if it simplifies completely.
256 if (Op1 && Op1->getOpcode() == Opcode) {
257 Value *A = LHS;
258 Value *B = Op1->getOperand(0);
259 Value *C = Op1->getOperand(1);
260
261 // Does "C op A" simplify?
262 if (Value *V = SimplifyBinOp(Opcode, C, A, Q, MaxRecurse)) {
263 // It does! Return "B op V" if it simplifies or is already available.
264 // If V equals C then "B op V" is just the RHS.
265 if (V == C) return RHS;
266 // Otherwise return "B op V" if it simplifies.
267 if (Value *W = SimplifyBinOp(Opcode, B, V, Q, MaxRecurse)) {
268 ++NumReassoc;
269 return W;
270 }
271 }
272 }
273
274 return nullptr;
275}
276
277/// In the case of a binary operation with a select instruction as an operand,
278/// try to simplify the binop by seeing whether evaluating it on both branches
279/// of the select results in the same value. Returns the common value if so,
280/// otherwise returns null.
281static Value *ThreadBinOpOverSelect(Instruction::BinaryOps Opcode, Value *LHS,
282 Value *RHS, const SimplifyQuery &Q,
283 unsigned MaxRecurse) {
284 // Recursion is always used, so bail out at once if we already hit the limit.
285 if (!MaxRecurse--)
286 return nullptr;
287
288 SelectInst *SI;
289 if (isa<SelectInst>(LHS)) {
290 SI = cast<SelectInst>(LHS);
291 } else {
292 assert(isa<SelectInst>(RHS) && "No select instruction operand!")(static_cast <bool> (isa<SelectInst>(RHS) &&
"No select instruction operand!") ? void (0) : __assert_fail
("isa<SelectInst>(RHS) && \"No select instruction operand!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 292, __extension__ __PRETTY_FUNCTION__))
;
293 SI = cast<SelectInst>(RHS);
294 }
295
296 // Evaluate the BinOp on the true and false branches of the select.
297 Value *TV;
298 Value *FV;
299 if (SI == LHS) {
300 TV = SimplifyBinOp(Opcode, SI->getTrueValue(), RHS, Q, MaxRecurse);
301 FV = SimplifyBinOp(Opcode, SI->getFalseValue(), RHS, Q, MaxRecurse);
302 } else {
303 TV = SimplifyBinOp(Opcode, LHS, SI->getTrueValue(), Q, MaxRecurse);
304 FV = SimplifyBinOp(Opcode, LHS, SI->getFalseValue(), Q, MaxRecurse);
305 }
306
307 // If they simplified to the same value, then return the common value.
308 // If they both failed to simplify then return null.
309 if (TV == FV)
310 return TV;
311
312 // If one branch simplified to undef, return the other one.
313 if (TV && isa<UndefValue>(TV))
314 return FV;
315 if (FV && isa<UndefValue>(FV))
316 return TV;
317
318 // If applying the operation did not change the true and false select values,
319 // then the result of the binop is the select itself.
320 if (TV == SI->getTrueValue() && FV == SI->getFalseValue())
321 return SI;
322
323 // If one branch simplified and the other did not, and the simplified
324 // value is equal to the unsimplified one, return the simplified value.
325 // For example, select (cond, X, X & Z) & Z -> X & Z.
326 if ((FV && !TV) || (TV && !FV)) {
327 // Check that the simplified value has the form "X op Y" where "op" is the
328 // same as the original operation.
329 Instruction *Simplified = dyn_cast<Instruction>(FV ? FV : TV);
330 if (Simplified && Simplified->getOpcode() == unsigned(Opcode)) {
331 // The value that didn't simplify is "UnsimplifiedLHS op UnsimplifiedRHS".
332 // We already know that "op" is the same as for the simplified value. See
333 // if the operands match too. If so, return the simplified value.
334 Value *UnsimplifiedBranch = FV ? SI->getTrueValue() : SI->getFalseValue();
335 Value *UnsimplifiedLHS = SI == LHS ? UnsimplifiedBranch : LHS;
336 Value *UnsimplifiedRHS = SI == LHS ? RHS : UnsimplifiedBranch;
337 if (Simplified->getOperand(0) == UnsimplifiedLHS &&
338 Simplified->getOperand(1) == UnsimplifiedRHS)
339 return Simplified;
340 if (Simplified->isCommutative() &&
341 Simplified->getOperand(1) == UnsimplifiedLHS &&
342 Simplified->getOperand(0) == UnsimplifiedRHS)
343 return Simplified;
344 }
345 }
346
347 return nullptr;
348}
349
350/// In the case of a comparison with a select instruction, try to simplify the
351/// comparison by seeing whether both branches of the select result in the same
352/// value. Returns the common value if so, otherwise returns null.
353static Value *ThreadCmpOverSelect(CmpInst::Predicate Pred, Value *LHS,
354 Value *RHS, const SimplifyQuery &Q,
355 unsigned MaxRecurse) {
356 // Recursion is always used, so bail out at once if we already hit the limit.
357 if (!MaxRecurse--)
358 return nullptr;
359
360 // Make sure the select is on the LHS.
361 if (!isa<SelectInst>(LHS)) {
362 std::swap(LHS, RHS);
363 Pred = CmpInst::getSwappedPredicate(Pred);
364 }
365 assert(isa<SelectInst>(LHS) && "Not comparing with a select instruction!")(static_cast <bool> (isa<SelectInst>(LHS) &&
"Not comparing with a select instruction!") ? void (0) : __assert_fail
("isa<SelectInst>(LHS) && \"Not comparing with a select instruction!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 365, __extension__ __PRETTY_FUNCTION__))
;
366 SelectInst *SI = cast<SelectInst>(LHS);
367 Value *Cond = SI->getCondition();
368 Value *TV = SI->getTrueValue();
369 Value *FV = SI->getFalseValue();
370
371 // Now that we have "cmp select(Cond, TV, FV), RHS", analyse it.
372 // Does "cmp TV, RHS" simplify?
373 Value *TCmp = SimplifyCmpInst(Pred, TV, RHS, Q, MaxRecurse);
374 if (TCmp == Cond) {
375 // It not only simplified, it simplified to the select condition. Replace
376 // it with 'true'.
377 TCmp = getTrue(Cond->getType());
378 } else if (!TCmp) {
379 // It didn't simplify. However if "cmp TV, RHS" is equal to the select
380 // condition then we can replace it with 'true'. Otherwise give up.
381 if (!isSameCompare(Cond, Pred, TV, RHS))
382 return nullptr;
383 TCmp = getTrue(Cond->getType());
384 }
385
386 // Does "cmp FV, RHS" simplify?
387 Value *FCmp = SimplifyCmpInst(Pred, FV, RHS, Q, MaxRecurse);
388 if (FCmp == Cond) {
389 // It not only simplified, it simplified to the select condition. Replace
390 // it with 'false'.
391 FCmp = getFalse(Cond->getType());
392 } else if (!FCmp) {
393 // It didn't simplify. However if "cmp FV, RHS" is equal to the select
394 // condition then we can replace it with 'false'. Otherwise give up.
395 if (!isSameCompare(Cond, Pred, FV, RHS))
396 return nullptr;
397 FCmp = getFalse(Cond->getType());
398 }
399
400 // If both sides simplified to the same value, then use it as the result of
401 // the original comparison.
402 if (TCmp == FCmp)
403 return TCmp;
404
405 // The remaining cases only make sense if the select condition has the same
406 // type as the result of the comparison, so bail out if this is not so.
407 if (Cond->getType()->isVectorTy() != RHS->getType()->isVectorTy())
408 return nullptr;
409 // If the false value simplified to false, then the result of the compare
410 // is equal to "Cond && TCmp". This also catches the case when the false
411 // value simplified to false and the true value to true, returning "Cond".
412 if (match(FCmp, m_Zero()))
413 if (Value *V = SimplifyAndInst(Cond, TCmp, Q, MaxRecurse))
414 return V;
415 // If the true value simplified to true, then the result of the compare
416 // is equal to "Cond || FCmp".
417 if (match(TCmp, m_One()))
418 if (Value *V = SimplifyOrInst(Cond, FCmp, Q, MaxRecurse))
419 return V;
420 // Finally, if the false value simplified to true and the true value to
421 // false, then the result of the compare is equal to "!Cond".
422 if (match(FCmp, m_One()) && match(TCmp, m_Zero()))
423 if (Value *V =
424 SimplifyXorInst(Cond, Constant::getAllOnesValue(Cond->getType()),
425 Q, MaxRecurse))
426 return V;
427
428 return nullptr;
429}
430
431/// In the case of a binary operation with an operand that is a PHI instruction,
432/// try to simplify the binop by seeing whether evaluating it on the incoming
433/// phi values yields the same result for every value. If so returns the common
434/// value, otherwise returns null.
435static Value *ThreadBinOpOverPHI(Instruction::BinaryOps Opcode, Value *LHS,
436 Value *RHS, const SimplifyQuery &Q,
437 unsigned MaxRecurse) {
438 // Recursion is always used, so bail out at once if we already hit the limit.
439 if (!MaxRecurse--)
440 return nullptr;
441
442 PHINode *PI;
443 if (isa<PHINode>(LHS)) {
444 PI = cast<PHINode>(LHS);
445 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
446 if (!ValueDominatesPHI(RHS, PI, Q.DT))
447 return nullptr;
448 } else {
449 assert(isa<PHINode>(RHS) && "No PHI instruction operand!")(static_cast <bool> (isa<PHINode>(RHS) &&
"No PHI instruction operand!") ? void (0) : __assert_fail ("isa<PHINode>(RHS) && \"No PHI instruction operand!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 449, __extension__ __PRETTY_FUNCTION__))
;
450 PI = cast<PHINode>(RHS);
451 // Bail out if LHS and the phi may be mutually interdependent due to a loop.
452 if (!ValueDominatesPHI(LHS, PI, Q.DT))
453 return nullptr;
454 }
455
456 // Evaluate the BinOp on the incoming phi values.
457 Value *CommonValue = nullptr;
458 for (Value *Incoming : PI->incoming_values()) {
459 // If the incoming value is the phi node itself, it can safely be skipped.
460 if (Incoming == PI) continue;
461 Value *V = PI == LHS ?
462 SimplifyBinOp(Opcode, Incoming, RHS, Q, MaxRecurse) :
463 SimplifyBinOp(Opcode, LHS, Incoming, Q, MaxRecurse);
464 // If the operation failed to simplify, or simplified to a different value
465 // to previously, then give up.
466 if (!V || (CommonValue && V != CommonValue))
467 return nullptr;
468 CommonValue = V;
469 }
470
471 return CommonValue;
472}
473
474/// In the case of a comparison with a PHI instruction, try to simplify the
475/// comparison by seeing whether comparing with all of the incoming phi values
476/// yields the same result every time. If so returns the common result,
477/// otherwise returns null.
478static Value *ThreadCmpOverPHI(CmpInst::Predicate Pred, Value *LHS, Value *RHS,
479 const SimplifyQuery &Q, unsigned MaxRecurse) {
480 // Recursion is always used, so bail out at once if we already hit the limit.
481 if (!MaxRecurse--)
482 return nullptr;
483
484 // Make sure the phi is on the LHS.
485 if (!isa<PHINode>(LHS)) {
486 std::swap(LHS, RHS);
487 Pred = CmpInst::getSwappedPredicate(Pred);
488 }
489 assert(isa<PHINode>(LHS) && "Not comparing with a phi instruction!")(static_cast <bool> (isa<PHINode>(LHS) &&
"Not comparing with a phi instruction!") ? void (0) : __assert_fail
("isa<PHINode>(LHS) && \"Not comparing with a phi instruction!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 489, __extension__ __PRETTY_FUNCTION__))
;
490 PHINode *PI = cast<PHINode>(LHS);
491
492 // Bail out if RHS and the phi may be mutually interdependent due to a loop.
493 if (!ValueDominatesPHI(RHS, PI, Q.DT))
494 return nullptr;
495
496 // Evaluate the BinOp on the incoming phi values.
497 Value *CommonValue = nullptr;
498 for (Value *Incoming : PI->incoming_values()) {
499 // If the incoming value is the phi node itself, it can safely be skipped.
500 if (Incoming == PI) continue;
501 Value *V = SimplifyCmpInst(Pred, Incoming, RHS, Q, MaxRecurse);
502 // If the operation failed to simplify, or simplified to a different value
503 // to previously, then give up.
504 if (!V || (CommonValue && V != CommonValue))
505 return nullptr;
506 CommonValue = V;
507 }
508
509 return CommonValue;
510}
511
512static Constant *foldOrCommuteConstant(Instruction::BinaryOps Opcode,
513 Value *&Op0, Value *&Op1,
514 const SimplifyQuery &Q) {
515 if (auto *CLHS = dyn_cast<Constant>(Op0)) {
516 if (auto *CRHS = dyn_cast<Constant>(Op1))
517 return ConstantFoldBinaryOpOperands(Opcode, CLHS, CRHS, Q.DL);
518
519 // Canonicalize the constant to the RHS if this is a commutative operation.
520 if (Instruction::isCommutative(Opcode))
521 std::swap(Op0, Op1);
522 }
523 return nullptr;
524}
525
526/// Given operands for an Add, see if we can fold the result.
527/// If not, this returns null.
528static Value *SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
529 const SimplifyQuery &Q, unsigned MaxRecurse) {
530 if (Constant *C = foldOrCommuteConstant(Instruction::Add, Op0, Op1, Q))
531 return C;
532
533 // X + undef -> undef
534 if (match(Op1, m_Undef()))
535 return Op1;
536
537 // X + 0 -> X
538 if (match(Op1, m_Zero()))
539 return Op0;
540
541 // X + (Y - X) -> Y
542 // (Y - X) + X -> Y
543 // Eg: X + -X -> 0
544 Value *Y = nullptr;
545 if (match(Op1, m_Sub(m_Value(Y), m_Specific(Op0))) ||
546 match(Op0, m_Sub(m_Value(Y), m_Specific(Op1))))
547 return Y;
548
549 // X + ~X -> -1 since ~X = -X-1
550 Type *Ty = Op0->getType();
551 if (match(Op0, m_Not(m_Specific(Op1))) ||
552 match(Op1, m_Not(m_Specific(Op0))))
553 return Constant::getAllOnesValue(Ty);
554
555 // add nsw/nuw (xor Y, signmask), signmask --> Y
556 // The no-wrapping add guarantees that the top bit will be set by the add.
557 // Therefore, the xor must be clearing the already set sign bit of Y.
558 if ((isNSW || isNUW) && match(Op1, m_SignMask()) &&
559 match(Op0, m_Xor(m_Value(Y), m_SignMask())))
560 return Y;
561
562 /// i1 add -> xor.
563 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
564 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
565 return V;
566
567 // Try some generic simplifications for associative operations.
568 if (Value *V = SimplifyAssociativeBinOp(Instruction::Add, Op0, Op1, Q,
569 MaxRecurse))
570 return V;
571
572 // Threading Add over selects and phi nodes is pointless, so don't bother.
573 // Threading over the select in "A + select(cond, B, C)" means evaluating
574 // "A+B" and "A+C" and seeing if they are equal; but they are equal if and
575 // only if B and C are equal. If B and C are equal then (since we assume
576 // that operands have already been simplified) "select(cond, B, C)" should
577 // have been simplified to the common value of B and C already. Analysing
578 // "A+B" and "A+C" thus gains nothing, but costs compile time. Similarly
579 // for threading over phi nodes.
580
581 return nullptr;
582}
583
584Value *llvm::SimplifyAddInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
585 const SimplifyQuery &Query) {
586 return ::SimplifyAddInst(Op0, Op1, isNSW, isNUW, Query, RecursionLimit);
587}
588
589/// \brief Compute the base pointer and cumulative constant offsets for V.
590///
591/// This strips all constant offsets off of V, leaving it the base pointer, and
592/// accumulates the total constant offset applied in the returned constant. It
593/// returns 0 if V is not a pointer, and returns the constant '0' if there are
594/// no constant offsets applied.
595///
596/// This is very similar to GetPointerBaseWithConstantOffset except it doesn't
597/// follow non-inbounds geps. This allows it to remain usable for icmp ult/etc.
598/// folding.
599static Constant *stripAndComputeConstantOffsets(const DataLayout &DL, Value *&V,
600 bool AllowNonInbounds = false) {
601 assert(V->getType()->isPtrOrPtrVectorTy())(static_cast <bool> (V->getType()->isPtrOrPtrVectorTy
()) ? void (0) : __assert_fail ("V->getType()->isPtrOrPtrVectorTy()"
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 601, __extension__ __PRETTY_FUNCTION__))
;
602
603 Type *IntPtrTy = DL.getIntPtrType(V->getType())->getScalarType();
604 APInt Offset = APInt::getNullValue(IntPtrTy->getIntegerBitWidth());
605
606 // Even though we don't look through PHI nodes, we could be called on an
607 // instruction in an unreachable block, which may be on a cycle.
608 SmallPtrSet<Value *, 4> Visited;
609 Visited.insert(V);
610 do {
611 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
612 if ((!AllowNonInbounds && !GEP->isInBounds()) ||
613 !GEP->accumulateConstantOffset(DL, Offset))
614 break;
615 V = GEP->getPointerOperand();
616 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
617 V = cast<Operator>(V)->getOperand(0);
618 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
619 if (GA->isInterposable())
620 break;
621 V = GA->getAliasee();
622 } else {
623 if (auto CS = CallSite(V))
624 if (Value *RV = CS.getReturnedArgOperand()) {
625 V = RV;
626 continue;
627 }
628 break;
629 }
630 assert(V->getType()->isPtrOrPtrVectorTy() && "Unexpected operand type!")(static_cast <bool> (V->getType()->isPtrOrPtrVectorTy
() && "Unexpected operand type!") ? void (0) : __assert_fail
("V->getType()->isPtrOrPtrVectorTy() && \"Unexpected operand type!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 630, __extension__ __PRETTY_FUNCTION__))
;
631 } while (Visited.insert(V).second);
632
633 Constant *OffsetIntPtr = ConstantInt::get(IntPtrTy, Offset);
634 if (V->getType()->isVectorTy())
635 return ConstantVector::getSplat(V->getType()->getVectorNumElements(),
636 OffsetIntPtr);
637 return OffsetIntPtr;
638}
639
640/// \brief Compute the constant difference between two pointer values.
641/// If the difference is not a constant, returns zero.
642static Constant *computePointerDifference(const DataLayout &DL, Value *LHS,
643 Value *RHS) {
644 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
645 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
646
647 // If LHS and RHS are not related via constant offsets to the same base
648 // value, there is nothing we can do here.
649 if (LHS != RHS)
650 return nullptr;
651
652 // Otherwise, the difference of LHS - RHS can be computed as:
653 // LHS - RHS
654 // = (LHSOffset + Base) - (RHSOffset + Base)
655 // = LHSOffset - RHSOffset
656 return ConstantExpr::getSub(LHSOffset, RHSOffset);
657}
658
659/// Given operands for a Sub, see if we can fold the result.
660/// If not, this returns null.
661static Value *SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
662 const SimplifyQuery &Q, unsigned MaxRecurse) {
663 if (Constant *C = foldOrCommuteConstant(Instruction::Sub, Op0, Op1, Q))
664 return C;
665
666 // X - undef -> undef
667 // undef - X -> undef
668 if (match(Op0, m_Undef()) || match(Op1, m_Undef()))
669 return UndefValue::get(Op0->getType());
670
671 // X - 0 -> X
672 if (match(Op1, m_Zero()))
673 return Op0;
674
675 // X - X -> 0
676 if (Op0 == Op1)
677 return Constant::getNullValue(Op0->getType());
678
679 // Is this a negation?
680 if (match(Op0, m_Zero())) {
681 // 0 - X -> 0 if the sub is NUW.
682 if (isNUW)
683 return Op0;
684
685 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
686 if (Known.Zero.isMaxSignedValue()) {
687 // Op1 is either 0 or the minimum signed value. If the sub is NSW, then
688 // Op1 must be 0 because negating the minimum signed value is undefined.
689 if (isNSW)
690 return Op0;
691
692 // 0 - X -> X if X is 0 or the minimum signed value.
693 return Op1;
694 }
695 }
696
697 // (X + Y) - Z -> X + (Y - Z) or Y + (X - Z) if everything simplifies.
698 // For example, (X + Y) - Y -> X; (Y + X) - Y -> X
699 Value *X = nullptr, *Y = nullptr, *Z = Op1;
700 if (MaxRecurse && match(Op0, m_Add(m_Value(X), m_Value(Y)))) { // (X + Y) - Z
701 // See if "V === Y - Z" simplifies.
702 if (Value *V = SimplifyBinOp(Instruction::Sub, Y, Z, Q, MaxRecurse-1))
703 // It does! Now see if "X + V" simplifies.
704 if (Value *W = SimplifyBinOp(Instruction::Add, X, V, Q, MaxRecurse-1)) {
705 // It does, we successfully reassociated!
706 ++NumReassoc;
707 return W;
708 }
709 // See if "V === X - Z" simplifies.
710 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
711 // It does! Now see if "Y + V" simplifies.
712 if (Value *W = SimplifyBinOp(Instruction::Add, Y, V, Q, MaxRecurse-1)) {
713 // It does, we successfully reassociated!
714 ++NumReassoc;
715 return W;
716 }
717 }
718
719 // X - (Y + Z) -> (X - Y) - Z or (X - Z) - Y if everything simplifies.
720 // For example, X - (X + 1) -> -1
721 X = Op0;
722 if (MaxRecurse && match(Op1, m_Add(m_Value(Y), m_Value(Z)))) { // X - (Y + Z)
723 // See if "V === X - Y" simplifies.
724 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
725 // It does! Now see if "V - Z" simplifies.
726 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Z, Q, MaxRecurse-1)) {
727 // It does, we successfully reassociated!
728 ++NumReassoc;
729 return W;
730 }
731 // See if "V === X - Z" simplifies.
732 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Z, Q, MaxRecurse-1))
733 // It does! Now see if "V - Y" simplifies.
734 if (Value *W = SimplifyBinOp(Instruction::Sub, V, Y, Q, MaxRecurse-1)) {
735 // It does, we successfully reassociated!
736 ++NumReassoc;
737 return W;
738 }
739 }
740
741 // Z - (X - Y) -> (Z - X) + Y if everything simplifies.
742 // For example, X - (X - Y) -> Y.
743 Z = Op0;
744 if (MaxRecurse && match(Op1, m_Sub(m_Value(X), m_Value(Y)))) // Z - (X - Y)
745 // See if "V === Z - X" simplifies.
746 if (Value *V = SimplifyBinOp(Instruction::Sub, Z, X, Q, MaxRecurse-1))
747 // It does! Now see if "V + Y" simplifies.
748 if (Value *W = SimplifyBinOp(Instruction::Add, V, Y, Q, MaxRecurse-1)) {
749 // It does, we successfully reassociated!
750 ++NumReassoc;
751 return W;
752 }
753
754 // trunc(X) - trunc(Y) -> trunc(X - Y) if everything simplifies.
755 if (MaxRecurse && match(Op0, m_Trunc(m_Value(X))) &&
756 match(Op1, m_Trunc(m_Value(Y))))
757 if (X->getType() == Y->getType())
758 // See if "V === X - Y" simplifies.
759 if (Value *V = SimplifyBinOp(Instruction::Sub, X, Y, Q, MaxRecurse-1))
760 // It does! Now see if "trunc V" simplifies.
761 if (Value *W = SimplifyCastInst(Instruction::Trunc, V, Op0->getType(),
762 Q, MaxRecurse - 1))
763 // It does, return the simplified "trunc V".
764 return W;
765
766 // Variations on GEP(base, I, ...) - GEP(base, i, ...) -> GEP(null, I-i, ...).
767 if (match(Op0, m_PtrToInt(m_Value(X))) &&
768 match(Op1, m_PtrToInt(m_Value(Y))))
769 if (Constant *Result = computePointerDifference(Q.DL, X, Y))
770 return ConstantExpr::getIntegerCast(Result, Op0->getType(), true);
771
772 // i1 sub -> xor.
773 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
774 if (Value *V = SimplifyXorInst(Op0, Op1, Q, MaxRecurse-1))
775 return V;
776
777 // Threading Sub over selects and phi nodes is pointless, so don't bother.
778 // Threading over the select in "A - select(cond, B, C)" means evaluating
779 // "A-B" and "A-C" and seeing if they are equal; but they are equal if and
780 // only if B and C are equal. If B and C are equal then (since we assume
781 // that operands have already been simplified) "select(cond, B, C)" should
782 // have been simplified to the common value of B and C already. Analysing
783 // "A-B" and "A-C" thus gains nothing, but costs compile time. Similarly
784 // for threading over phi nodes.
785
786 return nullptr;
787}
788
789Value *llvm::SimplifySubInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
790 const SimplifyQuery &Q) {
791 return ::SimplifySubInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
792}
793
794/// Given operands for a Mul, see if we can fold the result.
795/// If not, this returns null.
796static Value *SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
797 unsigned MaxRecurse) {
798 if (Constant *C = foldOrCommuteConstant(Instruction::Mul, Op0, Op1, Q))
799 return C;
800
801 // X * undef -> 0
802 if (match(Op1, m_Undef()))
803 return Constant::getNullValue(Op0->getType());
804
805 // X * 0 -> 0
806 if (match(Op1, m_Zero()))
807 return Op1;
808
809 // X * 1 -> X
810 if (match(Op1, m_One()))
811 return Op0;
812
813 // (X / Y) * Y -> X if the division is exact.
814 Value *X = nullptr;
815 if (match(Op0, m_Exact(m_IDiv(m_Value(X), m_Specific(Op1)))) || // (X / Y) * Y
816 match(Op1, m_Exact(m_IDiv(m_Value(X), m_Specific(Op0))))) // Y * (X / Y)
817 return X;
818
819 // i1 mul -> and.
820 if (MaxRecurse && Op0->getType()->isIntOrIntVectorTy(1))
821 if (Value *V = SimplifyAndInst(Op0, Op1, Q, MaxRecurse-1))
822 return V;
823
824 // Try some generic simplifications for associative operations.
825 if (Value *V = SimplifyAssociativeBinOp(Instruction::Mul, Op0, Op1, Q,
826 MaxRecurse))
827 return V;
828
829 // Mul distributes over Add. Try some generic simplifications based on this.
830 if (Value *V = ExpandBinOp(Instruction::Mul, Op0, Op1, Instruction::Add,
831 Q, MaxRecurse))
832 return V;
833
834 // If the operation is with the result of a select instruction, check whether
835 // operating on either branch of the select always yields the same value.
836 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
837 if (Value *V = ThreadBinOpOverSelect(Instruction::Mul, Op0, Op1, Q,
838 MaxRecurse))
839 return V;
840
841 // If the operation is with the result of a phi instruction, check whether
842 // operating on all incoming values of the phi always yields the same value.
843 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
844 if (Value *V = ThreadBinOpOverPHI(Instruction::Mul, Op0, Op1, Q,
845 MaxRecurse))
846 return V;
847
848 return nullptr;
849}
850
851Value *llvm::SimplifyMulInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
852 return ::SimplifyMulInst(Op0, Op1, Q, RecursionLimit);
853}
854
855/// Check for common or similar folds of integer division or integer remainder.
856/// This applies to all 4 opcodes (sdiv/udiv/srem/urem).
857static Value *simplifyDivRem(Value *Op0, Value *Op1, bool IsDiv) {
858 Type *Ty = Op0->getType();
859
860 // X / undef -> undef
861 // X % undef -> undef
862 if (match(Op1, m_Undef()))
863 return Op1;
864
865 // X / 0 -> undef
866 // X % 0 -> undef
867 // We don't need to preserve faults!
868 if (match(Op1, m_Zero()))
869 return UndefValue::get(Ty);
870
871 // If any element of a constant divisor vector is zero or undef, the whole op
872 // is undef.
873 auto *Op1C = dyn_cast<Constant>(Op1);
874 if (Op1C && Ty->isVectorTy()) {
875 unsigned NumElts = Ty->getVectorNumElements();
876 for (unsigned i = 0; i != NumElts; ++i) {
877 Constant *Elt = Op1C->getAggregateElement(i);
878 if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
879 return UndefValue::get(Ty);
880 }
881 }
882
883 // undef / X -> 0
884 // undef % X -> 0
885 if (match(Op0, m_Undef()))
886 return Constant::getNullValue(Ty);
887
888 // 0 / X -> 0
889 // 0 % X -> 0
890 if (match(Op0, m_Zero()))
891 return Op0;
892
893 // X / X -> 1
894 // X % X -> 0
895 if (Op0 == Op1)
896 return IsDiv ? ConstantInt::get(Ty, 1) : Constant::getNullValue(Ty);
897
898 // X / 1 -> X
899 // X % 1 -> 0
900 // If this is a boolean op (single-bit element type), we can't have
901 // division-by-zero or remainder-by-zero, so assume the divisor is 1.
902 if (match(Op1, m_One()) || Ty->isIntOrIntVectorTy(1))
903 return IsDiv ? Op0 : Constant::getNullValue(Ty);
904
905 return nullptr;
906}
907
908/// Given a predicate and two operands, return true if the comparison is true.
909/// This is a helper for div/rem simplification where we return some other value
910/// when we can prove a relationship between the operands.
911static bool isICmpTrue(ICmpInst::Predicate Pred, Value *LHS, Value *RHS,
912 const SimplifyQuery &Q, unsigned MaxRecurse) {
913 Value *V = SimplifyICmpInst(Pred, LHS, RHS, Q, MaxRecurse);
914 Constant *C = dyn_cast_or_null<Constant>(V);
915 return (C && C->isAllOnesValue());
916}
917
918/// Return true if we can simplify X / Y to 0. Remainder can adapt that answer
919/// to simplify X % Y to X.
920static bool isDivZero(Value *X, Value *Y, const SimplifyQuery &Q,
921 unsigned MaxRecurse, bool IsSigned) {
922 // Recursion is always used, so bail out at once if we already hit the limit.
923 if (!MaxRecurse--)
924 return false;
925
926 if (IsSigned) {
927 // |X| / |Y| --> 0
928 //
929 // We require that 1 operand is a simple constant. That could be extended to
930 // 2 variables if we computed the sign bit for each.
931 //
932 // Make sure that a constant is not the minimum signed value because taking
933 // the abs() of that is undefined.
934 Type *Ty = X->getType();
935 const APInt *C;
936 if (match(X, m_APInt(C)) && !C->isMinSignedValue()) {
937 // Is the variable divisor magnitude always greater than the constant
938 // dividend magnitude?
939 // |Y| > |C| --> Y < -abs(C) or Y > abs(C)
940 Constant *PosDividendC = ConstantInt::get(Ty, C->abs());
941 Constant *NegDividendC = ConstantInt::get(Ty, -C->abs());
942 if (isICmpTrue(CmpInst::ICMP_SLT, Y, NegDividendC, Q, MaxRecurse) ||
943 isICmpTrue(CmpInst::ICMP_SGT, Y, PosDividendC, Q, MaxRecurse))
944 return true;
945 }
946 if (match(Y, m_APInt(C))) {
947 // Special-case: we can't take the abs() of a minimum signed value. If
948 // that's the divisor, then all we have to do is prove that the dividend
949 // is also not the minimum signed value.
950 if (C->isMinSignedValue())
951 return isICmpTrue(CmpInst::ICMP_NE, X, Y, Q, MaxRecurse);
952
953 // Is the variable dividend magnitude always less than the constant
954 // divisor magnitude?
955 // |X| < |C| --> X > -abs(C) and X < abs(C)
956 Constant *PosDivisorC = ConstantInt::get(Ty, C->abs());
957 Constant *NegDivisorC = ConstantInt::get(Ty, -C->abs());
958 if (isICmpTrue(CmpInst::ICMP_SGT, X, NegDivisorC, Q, MaxRecurse) &&
959 isICmpTrue(CmpInst::ICMP_SLT, X, PosDivisorC, Q, MaxRecurse))
960 return true;
961 }
962 return false;
963 }
964
965 // IsSigned == false.
966 // Is the dividend unsigned less than the divisor?
967 return isICmpTrue(ICmpInst::ICMP_ULT, X, Y, Q, MaxRecurse);
968}
969
970/// These are simplifications common to SDiv and UDiv.
971static Value *simplifyDiv(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
972 const SimplifyQuery &Q, unsigned MaxRecurse) {
973 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
974 return C;
975
976 if (Value *V = simplifyDivRem(Op0, Op1, true))
977 return V;
978
979 bool IsSigned = Opcode == Instruction::SDiv;
980
981 // (X * Y) / Y -> X if the multiplication does not overflow.
982 Value *X;
983 if (match(Op0, m_c_Mul(m_Value(X), m_Specific(Op1)))) {
984 auto *Mul = cast<OverflowingBinaryOperator>(Op0);
985 // If the Mul does not overflow, then we are good to go.
986 if ((IsSigned && Mul->hasNoSignedWrap()) ||
987 (!IsSigned && Mul->hasNoUnsignedWrap()))
988 return X;
989 // If X has the form X = A / Y, then X * Y cannot overflow.
990 if ((IsSigned && match(X, m_SDiv(m_Value(), m_Specific(Op1)))) ||
991 (!IsSigned && match(X, m_UDiv(m_Value(), m_Specific(Op1)))))
992 return X;
993 }
994
995 // (X rem Y) / Y -> 0
996 if ((IsSigned && match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
997 (!IsSigned && match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
998 return Constant::getNullValue(Op0->getType());
999
1000 // (X /u C1) /u C2 -> 0 if C1 * C2 overflow
1001 ConstantInt *C1, *C2;
1002 if (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_ConstantInt(C1))) &&
1003 match(Op1, m_ConstantInt(C2))) {
1004 bool Overflow;
1005 (void)C1->getValue().umul_ov(C2->getValue(), Overflow);
1006 if (Overflow)
1007 return Constant::getNullValue(Op0->getType());
1008 }
1009
1010 // If the operation is with the result of a select instruction, check whether
1011 // operating on either branch of the select always yields the same value.
1012 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1013 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1014 return V;
1015
1016 // If the operation is with the result of a phi instruction, check whether
1017 // operating on all incoming values of the phi always yields the same value.
1018 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1019 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1020 return V;
1021
1022 if (isDivZero(Op0, Op1, Q, MaxRecurse, IsSigned))
1023 return Constant::getNullValue(Op0->getType());
1024
1025 return nullptr;
1026}
1027
1028/// These are simplifications common to SRem and URem.
1029static Value *simplifyRem(Instruction::BinaryOps Opcode, Value *Op0, Value *Op1,
1030 const SimplifyQuery &Q, unsigned MaxRecurse) {
1031 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1032 return C;
1033
1034 if (Value *V = simplifyDivRem(Op0, Op1, false))
1035 return V;
1036
1037 // (X % Y) % Y -> X % Y
1038 if ((Opcode == Instruction::SRem &&
1039 match(Op0, m_SRem(m_Value(), m_Specific(Op1)))) ||
1040 (Opcode == Instruction::URem &&
1041 match(Op0, m_URem(m_Value(), m_Specific(Op1)))))
1042 return Op0;
1043
1044 // (X << Y) % X -> 0
1045 if ((Opcode == Instruction::SRem &&
1046 match(Op0, m_NSWShl(m_Specific(Op1), m_Value()))) ||
1047 (Opcode == Instruction::URem &&
1048 match(Op0, m_NUWShl(m_Specific(Op1), m_Value()))))
1049 return Constant::getNullValue(Op0->getType());
1050
1051 // If the operation is with the result of a select instruction, check whether
1052 // operating on either branch of the select always yields the same value.
1053 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1054 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1055 return V;
1056
1057 // If the operation is with the result of a phi instruction, check whether
1058 // operating on all incoming values of the phi always yields the same value.
1059 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1060 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1061 return V;
1062
1063 // If X / Y == 0, then X % Y == X.
1064 if (isDivZero(Op0, Op1, Q, MaxRecurse, Opcode == Instruction::SRem))
1065 return Op0;
1066
1067 return nullptr;
1068}
1069
1070/// Given operands for an SDiv, see if we can fold the result.
1071/// If not, this returns null.
1072static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1073 unsigned MaxRecurse) {
1074 return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse);
1075}
1076
1077Value *llvm::SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1078 return ::SimplifySDivInst(Op0, Op1, Q, RecursionLimit);
1079}
1080
1081/// Given operands for a UDiv, see if we can fold the result.
1082/// If not, this returns null.
1083static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1084 unsigned MaxRecurse) {
1085 return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse);
1086}
1087
1088Value *llvm::SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1089 return ::SimplifyUDivInst(Op0, Op1, Q, RecursionLimit);
1090}
1091
1092/// Given operands for an SRem, see if we can fold the result.
1093/// If not, this returns null.
1094static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1095 unsigned MaxRecurse) {
1096 return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse);
1097}
1098
1099Value *llvm::SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1100 return ::SimplifySRemInst(Op0, Op1, Q, RecursionLimit);
1101}
1102
1103/// Given operands for a URem, see if we can fold the result.
1104/// If not, this returns null.
1105static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1106 unsigned MaxRecurse) {
1107 return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse);
1108}
1109
1110Value *llvm::SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1111 return ::SimplifyURemInst(Op0, Op1, Q, RecursionLimit);
1112}
1113
1114/// Returns true if a shift by \c Amount always yields undef.
1115static bool isUndefShift(Value *Amount) {
1116 Constant *C = dyn_cast<Constant>(Amount);
1117 if (!C)
1118 return false;
1119
1120 // X shift by undef -> undef because it may shift by the bitwidth.
1121 if (isa<UndefValue>(C))
1122 return true;
1123
1124 // Shifting by the bitwidth or more is undefined.
1125 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
1126 if (CI->getValue().getLimitedValue() >=
1127 CI->getType()->getScalarSizeInBits())
1128 return true;
1129
1130 // If all lanes of a vector shift are undefined the whole shift is.
1131 if (isa<ConstantVector>(C) || isa<ConstantDataVector>(C)) {
1132 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E; ++I)
1133 if (!isUndefShift(C->getAggregateElement(I)))
1134 return false;
1135 return true;
1136 }
1137
1138 return false;
1139}
1140
1141/// Given operands for an Shl, LShr or AShr, see if we can fold the result.
1142/// If not, this returns null.
1143static Value *SimplifyShift(Instruction::BinaryOps Opcode, Value *Op0,
1144 Value *Op1, const SimplifyQuery &Q, unsigned MaxRecurse) {
1145 if (Constant *C = foldOrCommuteConstant(Opcode, Op0, Op1, Q))
1146 return C;
1147
1148 // 0 shift by X -> 0
1149 if (match(Op0, m_Zero()))
1150 return Op0;
1151
1152 // X shift by 0 -> X
1153 if (match(Op1, m_Zero()))
1154 return Op0;
1155
1156 // Fold undefined shifts.
1157 if (isUndefShift(Op1))
1158 return UndefValue::get(Op0->getType());
1159
1160 // If the operation is with the result of a select instruction, check whether
1161 // operating on either branch of the select always yields the same value.
1162 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1163 if (Value *V = ThreadBinOpOverSelect(Opcode, Op0, Op1, Q, MaxRecurse))
1164 return V;
1165
1166 // If the operation is with the result of a phi instruction, check whether
1167 // operating on all incoming values of the phi always yields the same value.
1168 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1169 if (Value *V = ThreadBinOpOverPHI(Opcode, Op0, Op1, Q, MaxRecurse))
1170 return V;
1171
1172 // If any bits in the shift amount make that value greater than or equal to
1173 // the number of bits in the type, the shift is undefined.
1174 KnownBits Known = computeKnownBits(Op1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1175 if (Known.One.getLimitedValue() >= Known.getBitWidth())
1176 return UndefValue::get(Op0->getType());
1177
1178 // If all valid bits in the shift amount are known zero, the first operand is
1179 // unchanged.
1180 unsigned NumValidShiftBits = Log2_32_Ceil(Known.getBitWidth());
1181 if (Known.countMinTrailingZeros() >= NumValidShiftBits)
1182 return Op0;
1183
1184 return nullptr;
1185}
1186
1187/// \brief Given operands for an Shl, LShr or AShr, see if we can
1188/// fold the result. If not, this returns null.
1189static Value *SimplifyRightShift(Instruction::BinaryOps Opcode, Value *Op0,
1190 Value *Op1, bool isExact, const SimplifyQuery &Q,
1191 unsigned MaxRecurse) {
1192 if (Value *V = SimplifyShift(Opcode, Op0, Op1, Q, MaxRecurse))
1193 return V;
1194
1195 // X >> X -> 0
1196 if (Op0 == Op1)
1197 return Constant::getNullValue(Op0->getType());
1198
1199 // undef >> X -> 0
1200 // undef >> X -> undef (if it's exact)
1201 if (match(Op0, m_Undef()))
1202 return isExact ? Op0 : Constant::getNullValue(Op0->getType());
1203
1204 // The low bit cannot be shifted out of an exact shift if it is set.
1205 if (isExact) {
1206 KnownBits Op0Known = computeKnownBits(Op0, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1207 if (Op0Known.One[0])
1208 return Op0;
1209 }
1210
1211 return nullptr;
1212}
1213
1214/// Given operands for an Shl, see if we can fold the result.
1215/// If not, this returns null.
1216static Value *SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1217 const SimplifyQuery &Q, unsigned MaxRecurse) {
1218 if (Value *V = SimplifyShift(Instruction::Shl, Op0, Op1, Q, MaxRecurse))
1219 return V;
1220
1221 // undef << X -> 0
1222 // undef << X -> undef if (if it's NSW/NUW)
1223 if (match(Op0, m_Undef()))
1224 return isNSW || isNUW ? Op0 : Constant::getNullValue(Op0->getType());
1225
1226 // (X >> A) << A -> X
1227 Value *X;
1228 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_Specific(Op1)))))
1229 return X;
1230 return nullptr;
1231}
1232
1233Value *llvm::SimplifyShlInst(Value *Op0, Value *Op1, bool isNSW, bool isNUW,
1234 const SimplifyQuery &Q) {
1235 return ::SimplifyShlInst(Op0, Op1, isNSW, isNUW, Q, RecursionLimit);
1236}
1237
1238/// Given operands for an LShr, see if we can fold the result.
1239/// If not, this returns null.
1240static Value *SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1241 const SimplifyQuery &Q, unsigned MaxRecurse) {
1242 if (Value *V = SimplifyRightShift(Instruction::LShr, Op0, Op1, isExact, Q,
1243 MaxRecurse))
1244 return V;
1245
1246 // (X << A) >> A -> X
1247 Value *X;
1248 if (match(Op0, m_NUWShl(m_Value(X), m_Specific(Op1))))
1249 return X;
1250
1251 return nullptr;
1252}
1253
1254Value *llvm::SimplifyLShrInst(Value *Op0, Value *Op1, bool isExact,
1255 const SimplifyQuery &Q) {
1256 return ::SimplifyLShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1257}
1258
1259/// Given operands for an AShr, see if we can fold the result.
1260/// If not, this returns null.
1261static Value *SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1262 const SimplifyQuery &Q, unsigned MaxRecurse) {
1263 if (Value *V = SimplifyRightShift(Instruction::AShr, Op0, Op1, isExact, Q,
1264 MaxRecurse))
1265 return V;
1266
1267 // all ones >>a X -> -1
1268 // Do not return Op0 because it may contain undef elements if it's a vector.
1269 if (match(Op0, m_AllOnes()))
1270 return Constant::getAllOnesValue(Op0->getType());
1271
1272 // (X << A) >> A -> X
1273 Value *X;
1274 if (match(Op0, m_NSWShl(m_Value(X), m_Specific(Op1))))
1275 return X;
1276
1277 // Arithmetic shifting an all-sign-bit value is a no-op.
1278 unsigned NumSignBits = ComputeNumSignBits(Op0, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
1279 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
1280 return Op0;
1281
1282 return nullptr;
1283}
1284
1285Value *llvm::SimplifyAShrInst(Value *Op0, Value *Op1, bool isExact,
1286 const SimplifyQuery &Q) {
1287 return ::SimplifyAShrInst(Op0, Op1, isExact, Q, RecursionLimit);
1288}
1289
1290/// Commuted variants are assumed to be handled by calling this function again
1291/// with the parameters swapped.
1292static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp,
1293 ICmpInst *UnsignedICmp, bool IsAnd) {
1294 Value *X, *Y;
2
'X' declared without an initial value
1295
1296 ICmpInst::Predicate EqPred;
1297 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) ||
3
Taking false branch
1298 !ICmpInst::isEquality(EqPred))
1299 return nullptr;
1300
1301 ICmpInst::Predicate UnsignedPred;
1302 if (match(UnsignedICmp, m_ICmp(UnsignedPred, m_Value(X), m_Specific(Y))) &&
4
Calling 'm_Value'
9
Returning from 'm_Value'
10
Taking false branch
1303 ICmpInst::isUnsigned(UnsignedPred))
1304 ;
1305 else if (match(UnsignedICmp,
1306 m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) &&
11
1st function call argument is an uninitialized value
1307 ICmpInst::isUnsigned(UnsignedPred))
1308 UnsignedPred = ICmpInst::getSwappedPredicate(UnsignedPred);
1309 else
1310 return nullptr;
1311
1312 // X < Y && Y != 0 --> X < Y
1313 // X < Y || Y != 0 --> Y != 0
1314 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE)
1315 return IsAnd ? UnsignedICmp : ZeroICmp;
1316
1317 // X >= Y || Y != 0 --> true
1318 // X >= Y || Y == 0 --> X >= Y
1319 if (UnsignedPred == ICmpInst::ICMP_UGE && !IsAnd) {
1320 if (EqPred == ICmpInst::ICMP_NE)
1321 return getTrue(UnsignedICmp->getType());
1322 return UnsignedICmp;
1323 }
1324
1325 // X < Y && Y == 0 --> false
1326 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_EQ &&
1327 IsAnd)
1328 return getFalse(UnsignedICmp->getType());
1329
1330 return nullptr;
1331}
1332
1333/// Commuted variants are assumed to be handled by calling this function again
1334/// with the parameters swapped.
1335static Value *simplifyAndOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1336 ICmpInst::Predicate Pred0, Pred1;
1337 Value *A ,*B;
1338 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1339 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1340 return nullptr;
1341
1342 // We have (icmp Pred0, A, B) & (icmp Pred1, A, B).
1343 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1344 // can eliminate Op1 from this 'and'.
1345 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1346 return Op0;
1347
1348 // Check for any combination of predicates that are guaranteed to be disjoint.
1349 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1350 (Pred0 == ICmpInst::ICMP_EQ && ICmpInst::isFalseWhenEqual(Pred1)) ||
1351 (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT) ||
1352 (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT))
1353 return getFalse(Op0->getType());
1354
1355 return nullptr;
1356}
1357
1358/// Commuted variants are assumed to be handled by calling this function again
1359/// with the parameters swapped.
1360static Value *simplifyOrOfICmpsWithSameOperands(ICmpInst *Op0, ICmpInst *Op1) {
1361 ICmpInst::Predicate Pred0, Pred1;
1362 Value *A ,*B;
1363 if (!match(Op0, m_ICmp(Pred0, m_Value(A), m_Value(B))) ||
1364 !match(Op1, m_ICmp(Pred1, m_Specific(A), m_Specific(B))))
1365 return nullptr;
1366
1367 // We have (icmp Pred0, A, B) | (icmp Pred1, A, B).
1368 // If Op1 is always implied true by Op0, then Op0 is a subset of Op1, and we
1369 // can eliminate Op0 from this 'or'.
1370 if (ICmpInst::isImpliedTrueByMatchingCmp(Pred0, Pred1))
1371 return Op1;
1372
1373 // Check for any combination of predicates that cover the entire range of
1374 // possibilities.
1375 if ((Pred0 == ICmpInst::getInversePredicate(Pred1)) ||
1376 (Pred0 == ICmpInst::ICMP_NE && ICmpInst::isTrueWhenEqual(Pred1)) ||
1377 (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGE) ||
1378 (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGE))
1379 return getTrue(Op0->getType());
1380
1381 return nullptr;
1382}
1383
1384/// Test if a pair of compares with a shared operand and 2 constants has an
1385/// empty set intersection, full set union, or if one compare is a superset of
1386/// the other.
1387static Value *simplifyAndOrOfICmpsWithConstants(ICmpInst *Cmp0, ICmpInst *Cmp1,
1388 bool IsAnd) {
1389 // Look for this pattern: {and/or} (icmp X, C0), (icmp X, C1)).
1390 if (Cmp0->getOperand(0) != Cmp1->getOperand(0))
1391 return nullptr;
1392
1393 const APInt *C0, *C1;
1394 if (!match(Cmp0->getOperand(1), m_APInt(C0)) ||
1395 !match(Cmp1->getOperand(1), m_APInt(C1)))
1396 return nullptr;
1397
1398 auto Range0 = ConstantRange::makeExactICmpRegion(Cmp0->getPredicate(), *C0);
1399 auto Range1 = ConstantRange::makeExactICmpRegion(Cmp1->getPredicate(), *C1);
1400
1401 // For and-of-compares, check if the intersection is empty:
1402 // (icmp X, C0) && (icmp X, C1) --> empty set --> false
1403 if (IsAnd && Range0.intersectWith(Range1).isEmptySet())
1404 return getFalse(Cmp0->getType());
1405
1406 // For or-of-compares, check if the union is full:
1407 // (icmp X, C0) || (icmp X, C1) --> full set --> true
1408 if (!IsAnd && Range0.unionWith(Range1).isFullSet())
1409 return getTrue(Cmp0->getType());
1410
1411 // Is one range a superset of the other?
1412 // If this is and-of-compares, take the smaller set:
1413 // (icmp sgt X, 4) && (icmp sgt X, 42) --> icmp sgt X, 42
1414 // If this is or-of-compares, take the larger set:
1415 // (icmp sgt X, 4) || (icmp sgt X, 42) --> icmp sgt X, 4
1416 if (Range0.contains(Range1))
1417 return IsAnd ? Cmp1 : Cmp0;
1418 if (Range1.contains(Range0))
1419 return IsAnd ? Cmp0 : Cmp1;
1420
1421 return nullptr;
1422}
1423
1424static Value *simplifyAndOrOfICmpsWithZero(ICmpInst *Cmp0, ICmpInst *Cmp1,
1425 bool IsAnd) {
1426 ICmpInst::Predicate P0 = Cmp0->getPredicate(), P1 = Cmp1->getPredicate();
1427 if (!match(Cmp0->getOperand(1), m_Zero()) ||
1428 !match(Cmp1->getOperand(1), m_Zero()) || P0 != P1)
1429 return nullptr;
1430
1431 if ((IsAnd && P0 != ICmpInst::ICMP_NE) || (!IsAnd && P1 != ICmpInst::ICMP_EQ))
1432 return nullptr;
1433
1434 // We have either "(X == 0 || Y == 0)" or "(X != 0 && Y != 0)".
1435 Value *X = Cmp0->getOperand(0);
1436 Value *Y = Cmp1->getOperand(0);
1437
1438 // If one of the compares is a masked version of a (not) null check, then
1439 // that compare implies the other, so we eliminate the other. Optionally, look
1440 // through a pointer-to-int cast to match a null check of a pointer type.
1441
1442 // (X == 0) || (([ptrtoint] X & ?) == 0) --> ([ptrtoint] X & ?) == 0
1443 // (X == 0) || ((? & [ptrtoint] X) == 0) --> (? & [ptrtoint] X) == 0
1444 // (X != 0) && (([ptrtoint] X & ?) != 0) --> ([ptrtoint] X & ?) != 0
1445 // (X != 0) && ((? & [ptrtoint] X) != 0) --> (? & [ptrtoint] X) != 0
1446 if (match(Y, m_c_And(m_Specific(X), m_Value())) ||
1447 match(Y, m_c_And(m_PtrToInt(m_Specific(X)), m_Value())))
1448 return Cmp1;
1449
1450 // (([ptrtoint] Y & ?) == 0) || (Y == 0) --> ([ptrtoint] Y & ?) == 0
1451 // ((? & [ptrtoint] Y) == 0) || (Y == 0) --> (? & [ptrtoint] Y) == 0
1452 // (([ptrtoint] Y & ?) != 0) && (Y != 0) --> ([ptrtoint] Y & ?) != 0
1453 // ((? & [ptrtoint] Y) != 0) && (Y != 0) --> (? & [ptrtoint] Y) != 0
1454 if (match(X, m_c_And(m_Specific(Y), m_Value())) ||
1455 match(X, m_c_And(m_PtrToInt(m_Specific(Y)), m_Value())))
1456 return Cmp0;
1457
1458 return nullptr;
1459}
1460
1461static Value *simplifyAndOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1462 // (icmp (add V, C0), C1) & (icmp V, C0)
1463 ICmpInst::Predicate Pred0, Pred1;
1464 const APInt *C0, *C1;
1465 Value *V;
1466 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1467 return nullptr;
1468
1469 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1470 return nullptr;
1471
1472 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1473 if (AddInst->getOperand(1) != Op1->getOperand(1))
1474 return nullptr;
1475
1476 Type *ITy = Op0->getType();
1477 bool isNSW = AddInst->hasNoSignedWrap();
1478 bool isNUW = AddInst->hasNoUnsignedWrap();
1479
1480 const APInt Delta = *C1 - *C0;
1481 if (C0->isStrictlyPositive()) {
1482 if (Delta == 2) {
1483 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_SGT)
1484 return getFalse(ITy);
1485 if (Pred0 == ICmpInst::ICMP_SLT && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1486 return getFalse(ITy);
1487 }
1488 if (Delta == 1) {
1489 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_SGT)
1490 return getFalse(ITy);
1491 if (Pred0 == ICmpInst::ICMP_SLE && Pred1 == ICmpInst::ICMP_SGT && isNSW)
1492 return getFalse(ITy);
1493 }
1494 }
1495 if (C0->getBoolValue() && isNUW) {
1496 if (Delta == 2)
1497 if (Pred0 == ICmpInst::ICMP_ULT && Pred1 == ICmpInst::ICMP_UGT)
1498 return getFalse(ITy);
1499 if (Delta == 1)
1500 if (Pred0 == ICmpInst::ICMP_ULE && Pred1 == ICmpInst::ICMP_UGT)
1501 return getFalse(ITy);
1502 }
1503
1504 return nullptr;
1505}
1506
1507static Value *simplifyAndOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1508 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/true))
1509 return X;
1510 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/true))
1511 return X;
1512
1513 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op0, Op1))
1514 return X;
1515 if (Value *X = simplifyAndOfICmpsWithSameOperands(Op1, Op0))
1516 return X;
1517
1518 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, true))
1519 return X;
1520
1521 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, true))
1522 return X;
1523
1524 if (Value *X = simplifyAndOfICmpsWithAdd(Op0, Op1))
1525 return X;
1526 if (Value *X = simplifyAndOfICmpsWithAdd(Op1, Op0))
1527 return X;
1528
1529 return nullptr;
1530}
1531
1532static Value *simplifyOrOfICmpsWithAdd(ICmpInst *Op0, ICmpInst *Op1) {
1533 // (icmp (add V, C0), C1) | (icmp V, C0)
1534 ICmpInst::Predicate Pred0, Pred1;
1535 const APInt *C0, *C1;
1536 Value *V;
1537 if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
1538 return nullptr;
1539
1540 if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
1541 return nullptr;
1542
1543 auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
1544 if (AddInst->getOperand(1) != Op1->getOperand(1))
1545 return nullptr;
1546
1547 Type *ITy = Op0->getType();
1548 bool isNSW = AddInst->hasNoSignedWrap();
1549 bool isNUW = AddInst->hasNoUnsignedWrap();
1550
1551 const APInt Delta = *C1 - *C0;
1552 if (C0->isStrictlyPositive()) {
1553 if (Delta == 2) {
1554 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
1555 return getTrue(ITy);
1556 if (Pred0 == ICmpInst::ICMP_SGE && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1557 return getTrue(ITy);
1558 }
1559 if (Delta == 1) {
1560 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_SLE)
1561 return getTrue(ITy);
1562 if (Pred0 == ICmpInst::ICMP_SGT && Pred1 == ICmpInst::ICMP_SLE && isNSW)
1563 return getTrue(ITy);
1564 }
1565 }
1566 if (C0->getBoolValue() && isNUW) {
1567 if (Delta == 2)
1568 if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
1569 return getTrue(ITy);
1570 if (Delta == 1)
1571 if (Pred0 == ICmpInst::ICMP_UGT && Pred1 == ICmpInst::ICMP_ULE)
1572 return getTrue(ITy);
1573 }
1574
1575 return nullptr;
1576}
1577
1578static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) {
1579 if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
1
Calling 'simplifyUnsignedRangeCheck'
1580 return X;
1581 if (Value *X = simplifyUnsignedRangeCheck(Op1, Op0, /*IsAnd=*/false))
1582 return X;
1583
1584 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op0, Op1))
1585 return X;
1586 if (Value *X = simplifyOrOfICmpsWithSameOperands(Op1, Op0))
1587 return X;
1588
1589 if (Value *X = simplifyAndOrOfICmpsWithConstants(Op0, Op1, false))
1590 return X;
1591
1592 if (Value *X = simplifyAndOrOfICmpsWithZero(Op0, Op1, false))
1593 return X;
1594
1595 if (Value *X = simplifyOrOfICmpsWithAdd(Op0, Op1))
1596 return X;
1597 if (Value *X = simplifyOrOfICmpsWithAdd(Op1, Op0))
1598 return X;
1599
1600 return nullptr;
1601}
1602
1603static Value *simplifyAndOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1604 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1605 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1606 if (LHS0->getType() != RHS0->getType())
1607 return nullptr;
1608
1609 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1610 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1611 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1612 // (fcmp ord NNAN, X) & (fcmp ord X, Y) --> fcmp ord X, Y
1613 // (fcmp ord NNAN, X) & (fcmp ord Y, X) --> fcmp ord Y, X
1614 // (fcmp ord X, NNAN) & (fcmp ord X, Y) --> fcmp ord X, Y
1615 // (fcmp ord X, NNAN) & (fcmp ord Y, X) --> fcmp ord Y, X
1616 // (fcmp uno NNAN, X) | (fcmp uno X, Y) --> fcmp uno X, Y
1617 // (fcmp uno NNAN, X) | (fcmp uno Y, X) --> fcmp uno Y, X
1618 // (fcmp uno X, NNAN) | (fcmp uno X, Y) --> fcmp uno X, Y
1619 // (fcmp uno X, NNAN) | (fcmp uno Y, X) --> fcmp uno Y, X
1620 if ((isKnownNeverNaN(LHS0) && (LHS1 == RHS0 || LHS1 == RHS1)) ||
1621 (isKnownNeverNaN(LHS1) && (LHS0 == RHS0 || LHS0 == RHS1)))
1622 return RHS;
1623
1624 // (fcmp ord X, Y) & (fcmp ord NNAN, X) --> fcmp ord X, Y
1625 // (fcmp ord Y, X) & (fcmp ord NNAN, X) --> fcmp ord Y, X
1626 // (fcmp ord X, Y) & (fcmp ord X, NNAN) --> fcmp ord X, Y
1627 // (fcmp ord Y, X) & (fcmp ord X, NNAN) --> fcmp ord Y, X
1628 // (fcmp uno X, Y) | (fcmp uno NNAN, X) --> fcmp uno X, Y
1629 // (fcmp uno Y, X) | (fcmp uno NNAN, X) --> fcmp uno Y, X
1630 // (fcmp uno X, Y) | (fcmp uno X, NNAN) --> fcmp uno X, Y
1631 // (fcmp uno Y, X) | (fcmp uno X, NNAN) --> fcmp uno Y, X
1632 if ((isKnownNeverNaN(RHS0) && (RHS1 == LHS0 || RHS1 == LHS1)) ||
1633 (isKnownNeverNaN(RHS1) && (RHS0 == LHS0 || RHS0 == LHS1)))
1634 return LHS;
1635 }
1636
1637 return nullptr;
1638}
1639
1640static Value *simplifyAndOrOfCmps(Value *Op0, Value *Op1, bool IsAnd) {
1641 // Look through casts of the 'and' operands to find compares.
1642 auto *Cast0 = dyn_cast<CastInst>(Op0);
1643 auto *Cast1 = dyn_cast<CastInst>(Op1);
1644 if (Cast0 && Cast1 && Cast0->getOpcode() == Cast1->getOpcode() &&
1645 Cast0->getSrcTy() == Cast1->getSrcTy()) {
1646 Op0 = Cast0->getOperand(0);
1647 Op1 = Cast1->getOperand(0);
1648 }
1649
1650 Value *V = nullptr;
1651 auto *ICmp0 = dyn_cast<ICmpInst>(Op0);
1652 auto *ICmp1 = dyn_cast<ICmpInst>(Op1);
1653 if (ICmp0 && ICmp1)
1654 V = IsAnd ? simplifyAndOfICmps(ICmp0, ICmp1) :
1655 simplifyOrOfICmps(ICmp0, ICmp1);
1656
1657 auto *FCmp0 = dyn_cast<FCmpInst>(Op0);
1658 auto *FCmp1 = dyn_cast<FCmpInst>(Op1);
1659 if (FCmp0 && FCmp1)
1660 V = simplifyAndOrOfFCmps(FCmp0, FCmp1, IsAnd);
1661
1662 if (!V)
1663 return nullptr;
1664 if (!Cast0)
1665 return V;
1666
1667 // If we looked through casts, we can only handle a constant simplification
1668 // because we are not allowed to create a cast instruction here.
1669 if (auto *C = dyn_cast<Constant>(V))
1670 return ConstantExpr::getCast(Cast0->getOpcode(), C, Cast0->getType());
1671
1672 return nullptr;
1673}
1674
1675/// Given operands for an And, see if we can fold the result.
1676/// If not, this returns null.
1677static Value *SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1678 unsigned MaxRecurse) {
1679 if (Constant *C = foldOrCommuteConstant(Instruction::And, Op0, Op1, Q))
1680 return C;
1681
1682 // X & undef -> 0
1683 if (match(Op1, m_Undef()))
1684 return Constant::getNullValue(Op0->getType());
1685
1686 // X & X = X
1687 if (Op0 == Op1)
1688 return Op0;
1689
1690 // X & 0 = 0
1691 if (match(Op1, m_Zero()))
1692 return Op1;
1693
1694 // X & -1 = X
1695 if (match(Op1, m_AllOnes()))
1696 return Op0;
1697
1698 // A & ~A = ~A & A = 0
1699 if (match(Op0, m_Not(m_Specific(Op1))) ||
1700 match(Op1, m_Not(m_Specific(Op0))))
1701 return Constant::getNullValue(Op0->getType());
1702
1703 // (A | ?) & A = A
1704 if (match(Op0, m_c_Or(m_Specific(Op1), m_Value())))
1705 return Op1;
1706
1707 // A & (A | ?) = A
1708 if (match(Op1, m_c_Or(m_Specific(Op0), m_Value())))
1709 return Op0;
1710
1711 // A mask that only clears known zeros of a shifted value is a no-op.
1712 Value *X;
1713 const APInt *Mask;
1714 const APInt *ShAmt;
1715 if (match(Op1, m_APInt(Mask))) {
1716 // If all bits in the inverted and shifted mask are clear:
1717 // and (shl X, ShAmt), Mask --> shl X, ShAmt
1718 if (match(Op0, m_Shl(m_Value(X), m_APInt(ShAmt))) &&
1719 (~(*Mask)).lshr(*ShAmt).isNullValue())
1720 return Op0;
1721
1722 // If all bits in the inverted and shifted mask are clear:
1723 // and (lshr X, ShAmt), Mask --> lshr X, ShAmt
1724 if (match(Op0, m_LShr(m_Value(X), m_APInt(ShAmt))) &&
1725 (~(*Mask)).shl(*ShAmt).isNullValue())
1726 return Op0;
1727 }
1728
1729 // A & (-A) = A if A is a power of two or zero.
1730 if (match(Op0, m_Neg(m_Specific(Op1))) ||
1731 match(Op1, m_Neg(m_Specific(Op0)))) {
1732 if (isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1733 Q.DT))
1734 return Op0;
1735 if (isKnownToBeAPowerOfTwo(Op1, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI,
1736 Q.DT))
1737 return Op1;
1738 }
1739
1740 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, true))
1741 return V;
1742
1743 // Try some generic simplifications for associative operations.
1744 if (Value *V = SimplifyAssociativeBinOp(Instruction::And, Op0, Op1, Q,
1745 MaxRecurse))
1746 return V;
1747
1748 // And distributes over Or. Try some generic simplifications based on this.
1749 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Or,
1750 Q, MaxRecurse))
1751 return V;
1752
1753 // And distributes over Xor. Try some generic simplifications based on this.
1754 if (Value *V = ExpandBinOp(Instruction::And, Op0, Op1, Instruction::Xor,
1755 Q, MaxRecurse))
1756 return V;
1757
1758 // If the operation is with the result of a select instruction, check whether
1759 // operating on either branch of the select always yields the same value.
1760 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1761 if (Value *V = ThreadBinOpOverSelect(Instruction::And, Op0, Op1, Q,
1762 MaxRecurse))
1763 return V;
1764
1765 // If the operation is with the result of a phi instruction, check whether
1766 // operating on all incoming values of the phi always yields the same value.
1767 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1768 if (Value *V = ThreadBinOpOverPHI(Instruction::And, Op0, Op1, Q,
1769 MaxRecurse))
1770 return V;
1771
1772 return nullptr;
1773}
1774
1775Value *llvm::SimplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1776 return ::SimplifyAndInst(Op0, Op1, Q, RecursionLimit);
1777}
1778
1779/// Given operands for an Or, see if we can fold the result.
1780/// If not, this returns null.
1781static Value *SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1782 unsigned MaxRecurse) {
1783 if (Constant *C = foldOrCommuteConstant(Instruction::Or, Op0, Op1, Q))
1784 return C;
1785
1786 // X | undef -> -1
1787 // X | -1 = -1
1788 // Do not return Op1 because it may contain undef elements if it's a vector.
1789 if (match(Op1, m_Undef()) || match(Op1, m_AllOnes()))
1790 return Constant::getAllOnesValue(Op0->getType());
1791
1792 // X | X = X
1793 // X | 0 = X
1794 if (Op0 == Op1 || match(Op1, m_Zero()))
1795 return Op0;
1796
1797 // A | ~A = ~A | A = -1
1798 if (match(Op0, m_Not(m_Specific(Op1))) ||
1799 match(Op1, m_Not(m_Specific(Op0))))
1800 return Constant::getAllOnesValue(Op0->getType());
1801
1802 // (A & ?) | A = A
1803 if (match(Op0, m_c_And(m_Specific(Op1), m_Value())))
1804 return Op1;
1805
1806 // A | (A & ?) = A
1807 if (match(Op1, m_c_And(m_Specific(Op0), m_Value())))
1808 return Op0;
1809
1810 // ~(A & ?) | A = -1
1811 if (match(Op0, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1812 return Constant::getAllOnesValue(Op1->getType());
1813
1814 // A | ~(A & ?) = -1
1815 if (match(Op1, m_Not(m_c_And(m_Specific(Op1), m_Value()))))
1816 return Constant::getAllOnesValue(Op0->getType());
1817
1818 Value *A, *B;
1819 // (A & ~B) | (A ^ B) -> (A ^ B)
1820 // (~B & A) | (A ^ B) -> (A ^ B)
1821 // (A & ~B) | (B ^ A) -> (B ^ A)
1822 // (~B & A) | (B ^ A) -> (B ^ A)
1823 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
1824 (match(Op0, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1825 match(Op0, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1826 return Op1;
1827
1828 // Commute the 'or' operands.
1829 // (A ^ B) | (A & ~B) -> (A ^ B)
1830 // (A ^ B) | (~B & A) -> (A ^ B)
1831 // (B ^ A) | (A & ~B) -> (B ^ A)
1832 // (B ^ A) | (~B & A) -> (B ^ A)
1833 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
1834 (match(Op1, m_c_And(m_Specific(A), m_Not(m_Specific(B)))) ||
1835 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))))
1836 return Op0;
1837
1838 // (A & B) | (~A ^ B) -> (~A ^ B)
1839 // (B & A) | (~A ^ B) -> (~A ^ B)
1840 // (A & B) | (B ^ ~A) -> (B ^ ~A)
1841 // (B & A) | (B ^ ~A) -> (B ^ ~A)
1842 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1843 (match(Op1, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1844 match(Op1, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1845 return Op1;
1846
1847 // (~A ^ B) | (A & B) -> (~A ^ B)
1848 // (~A ^ B) | (B & A) -> (~A ^ B)
1849 // (B ^ ~A) | (A & B) -> (B ^ ~A)
1850 // (B ^ ~A) | (B & A) -> (B ^ ~A)
1851 if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
1852 (match(Op0, m_c_Xor(m_Specific(A), m_Not(m_Specific(B)))) ||
1853 match(Op0, m_c_Xor(m_Not(m_Specific(A)), m_Specific(B)))))
1854 return Op0;
1855
1856 if (Value *V = simplifyAndOrOfCmps(Op0, Op1, false))
1857 return V;
1858
1859 // Try some generic simplifications for associative operations.
1860 if (Value *V = SimplifyAssociativeBinOp(Instruction::Or, Op0, Op1, Q,
1861 MaxRecurse))
1862 return V;
1863
1864 // Or distributes over And. Try some generic simplifications based on this.
1865 if (Value *V = ExpandBinOp(Instruction::Or, Op0, Op1, Instruction::And, Q,
1866 MaxRecurse))
1867 return V;
1868
1869 // If the operation is with the result of a select instruction, check whether
1870 // operating on either branch of the select always yields the same value.
1871 if (isa<SelectInst>(Op0) || isa<SelectInst>(Op1))
1872 if (Value *V = ThreadBinOpOverSelect(Instruction::Or, Op0, Op1, Q,
1873 MaxRecurse))
1874 return V;
1875
1876 // (A & C1)|(B & C2)
1877 const APInt *C1, *C2;
1878 if (match(Op0, m_And(m_Value(A), m_APInt(C1))) &&
1879 match(Op1, m_And(m_Value(B), m_APInt(C2)))) {
1880 if (*C1 == ~*C2) {
1881 // (A & C1)|(B & C2)
1882 // If we have: ((V + N) & C1) | (V & C2)
1883 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1884 // replace with V+N.
1885 Value *N;
1886 if (C2->isMask() && // C2 == 0+1+
1887 match(A, m_c_Add(m_Specific(B), m_Value(N)))) {
1888 // Add commutes, try both ways.
1889 if (MaskedValueIsZero(N, *C2, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1890 return A;
1891 }
1892 // Or commutes, try both ways.
1893 if (C1->isMask() &&
1894 match(B, m_c_Add(m_Specific(A), m_Value(N)))) {
1895 // Add commutes, try both ways.
1896 if (MaskedValueIsZero(N, *C1, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
1897 return B;
1898 }
1899 }
1900 }
1901
1902 // If the operation is with the result of a phi instruction, check whether
1903 // operating on all incoming values of the phi always yields the same value.
1904 if (isa<PHINode>(Op0) || isa<PHINode>(Op1))
1905 if (Value *V = ThreadBinOpOverPHI(Instruction::Or, Op0, Op1, Q, MaxRecurse))
1906 return V;
1907
1908 return nullptr;
1909}
1910
1911Value *llvm::SimplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1912 return ::SimplifyOrInst(Op0, Op1, Q, RecursionLimit);
1913}
1914
1915/// Given operands for a Xor, see if we can fold the result.
1916/// If not, this returns null.
1917static Value *SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
1918 unsigned MaxRecurse) {
1919 if (Constant *C = foldOrCommuteConstant(Instruction::Xor, Op0, Op1, Q))
1920 return C;
1921
1922 // A ^ undef -> undef
1923 if (match(Op1, m_Undef()))
1924 return Op1;
1925
1926 // A ^ 0 = A
1927 if (match(Op1, m_Zero()))
1928 return Op0;
1929
1930 // A ^ A = 0
1931 if (Op0 == Op1)
1932 return Constant::getNullValue(Op0->getType());
1933
1934 // A ^ ~A = ~A ^ A = -1
1935 if (match(Op0, m_Not(m_Specific(Op1))) ||
1936 match(Op1, m_Not(m_Specific(Op0))))
1937 return Constant::getAllOnesValue(Op0->getType());
1938
1939 // Try some generic simplifications for associative operations.
1940 if (Value *V = SimplifyAssociativeBinOp(Instruction::Xor, Op0, Op1, Q,
1941 MaxRecurse))
1942 return V;
1943
1944 // Threading Xor over selects and phi nodes is pointless, so don't bother.
1945 // Threading over the select in "A ^ select(cond, B, C)" means evaluating
1946 // "A^B" and "A^C" and seeing if they are equal; but they are equal if and
1947 // only if B and C are equal. If B and C are equal then (since we assume
1948 // that operands have already been simplified) "select(cond, B, C)" should
1949 // have been simplified to the common value of B and C already. Analysing
1950 // "A^B" and "A^C" thus gains nothing, but costs compile time. Similarly
1951 // for threading over phi nodes.
1952
1953 return nullptr;
1954}
1955
1956Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) {
1957 return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit);
1958}
1959
1960
1961static Type *GetCompareTy(Value *Op) {
1962 return CmpInst::makeCmpResultType(Op->getType());
1963}
1964
1965/// Rummage around inside V looking for something equivalent to the comparison
1966/// "LHS Pred RHS". Return such a value if found, otherwise return null.
1967/// Helper function for analyzing max/min idioms.
1968static Value *ExtractEquivalentCondition(Value *V, CmpInst::Predicate Pred,
1969 Value *LHS, Value *RHS) {
1970 SelectInst *SI = dyn_cast<SelectInst>(V);
1971 if (!SI)
1972 return nullptr;
1973 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1974 if (!Cmp)
1975 return nullptr;
1976 Value *CmpLHS = Cmp->getOperand(0), *CmpRHS = Cmp->getOperand(1);
1977 if (Pred == Cmp->getPredicate() && LHS == CmpLHS && RHS == CmpRHS)
1978 return Cmp;
1979 if (Pred == CmpInst::getSwappedPredicate(Cmp->getPredicate()) &&
1980 LHS == CmpRHS && RHS == CmpLHS)
1981 return Cmp;
1982 return nullptr;
1983}
1984
1985// A significant optimization not implemented here is assuming that alloca
1986// addresses are not equal to incoming argument values. They don't *alias*,
1987// as we say, but that doesn't mean they aren't equal, so we take a
1988// conservative approach.
1989//
1990// This is inspired in part by C++11 5.10p1:
1991// "Two pointers of the same type compare equal if and only if they are both
1992// null, both point to the same function, or both represent the same
1993// address."
1994//
1995// This is pretty permissive.
1996//
1997// It's also partly due to C11 6.5.9p6:
1998// "Two pointers compare equal if and only if both are null pointers, both are
1999// pointers to the same object (including a pointer to an object and a
2000// subobject at its beginning) or function, both are pointers to one past the
2001// last element of the same array object, or one is a pointer to one past the
2002// end of one array object and the other is a pointer to the start of a
2003// different array object that happens to immediately follow the first array
2004// object in the address space.)
2005//
2006// C11's version is more restrictive, however there's no reason why an argument
2007// couldn't be a one-past-the-end value for a stack object in the caller and be
2008// equal to the beginning of a stack object in the callee.
2009//
2010// If the C and C++ standards are ever made sufficiently restrictive in this
2011// area, it may be possible to update LLVM's semantics accordingly and reinstate
2012// this optimization.
2013static Constant *
2014computePointerICmp(const DataLayout &DL, const TargetLibraryInfo *TLI,
2015 const DominatorTree *DT, CmpInst::Predicate Pred,
2016 AssumptionCache *AC, const Instruction *CxtI,
2017 Value *LHS, Value *RHS) {
2018 // First, skip past any trivial no-ops.
2019 LHS = LHS->stripPointerCasts();
2020 RHS = RHS->stripPointerCasts();
2021
2022 // A non-null pointer is not equal to a null pointer.
2023 if (llvm::isKnownNonZero(LHS, DL) && isa<ConstantPointerNull>(RHS) &&
2024 (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE))
2025 return ConstantInt::get(GetCompareTy(LHS),
2026 !CmpInst::isTrueWhenEqual(Pred));
2027
2028 // We can only fold certain predicates on pointer comparisons.
2029 switch (Pred) {
2030 default:
2031 return nullptr;
2032
2033 // Equality comaprisons are easy to fold.
2034 case CmpInst::ICMP_EQ:
2035 case CmpInst::ICMP_NE:
2036 break;
2037
2038 // We can only handle unsigned relational comparisons because 'inbounds' on
2039 // a GEP only protects against unsigned wrapping.
2040 case CmpInst::ICMP_UGT:
2041 case CmpInst::ICMP_UGE:
2042 case CmpInst::ICMP_ULT:
2043 case CmpInst::ICMP_ULE:
2044 // However, we have to switch them to their signed variants to handle
2045 // negative indices from the base pointer.
2046 Pred = ICmpInst::getSignedPredicate(Pred);
2047 break;
2048 }
2049
2050 // Strip off any constant offsets so that we can reason about them.
2051 // It's tempting to use getUnderlyingObject or even just stripInBoundsOffsets
2052 // here and compare base addresses like AliasAnalysis does, however there are
2053 // numerous hazards. AliasAnalysis and its utilities rely on special rules
2054 // governing loads and stores which don't apply to icmps. Also, AliasAnalysis
2055 // doesn't need to guarantee pointer inequality when it says NoAlias.
2056 Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS);
2057 Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS);
2058
2059 // If LHS and RHS are related via constant offsets to the same base
2060 // value, we can replace it with an icmp which just compares the offsets.
2061 if (LHS == RHS)
2062 return ConstantExpr::getICmp(Pred, LHSOffset, RHSOffset);
2063
2064 // Various optimizations for (in)equality comparisons.
2065 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
2066 // Different non-empty allocations that exist at the same time have
2067 // different addresses (if the program can tell). Global variables always
2068 // exist, so they always exist during the lifetime of each other and all
2069 // allocas. Two different allocas usually have different addresses...
2070 //
2071 // However, if there's an @llvm.stackrestore dynamically in between two
2072 // allocas, they may have the same address. It's tempting to reduce the
2073 // scope of the problem by only looking at *static* allocas here. That would
2074 // cover the majority of allocas while significantly reducing the likelihood
2075 // of having an @llvm.stackrestore pop up in the middle. However, it's not
2076 // actually impossible for an @llvm.stackrestore to pop up in the middle of
2077 // an entry block. Also, if we have a block that's not attached to a
2078 // function, we can't tell if it's "static" under the current definition.
2079 // Theoretically, this problem could be fixed by creating a new kind of
2080 // instruction kind specifically for static allocas. Such a new instruction
2081 // could be required to be at the top of the entry block, thus preventing it
2082 // from being subject to a @llvm.stackrestore. Instcombine could even
2083 // convert regular allocas into these special allocas. It'd be nifty.
2084 // However, until then, this problem remains open.
2085 //
2086 // So, we'll assume that two non-empty allocas have different addresses
2087 // for now.
2088 //
2089 // With all that, if the offsets are within the bounds of their allocations
2090 // (and not one-past-the-end! so we can't use inbounds!), and their
2091 // allocations aren't the same, the pointers are not equal.
2092 //
2093 // Note that it's not necessary to check for LHS being a global variable
2094 // address, due to canonicalization and constant folding.
2095 if (isa<AllocaInst>(LHS) &&
2096 (isa<AllocaInst>(RHS) || isa<GlobalVariable>(RHS))) {
2097 ConstantInt *LHSOffsetCI = dyn_cast<ConstantInt>(LHSOffset);
2098 ConstantInt *RHSOffsetCI = dyn_cast<ConstantInt>(RHSOffset);
2099 uint64_t LHSSize, RHSSize;
2100 if (LHSOffsetCI && RHSOffsetCI &&
2101 getObjectSize(LHS, LHSSize, DL, TLI) &&
2102 getObjectSize(RHS, RHSSize, DL, TLI)) {
2103 const APInt &LHSOffsetValue = LHSOffsetCI->getValue();
2104 const APInt &RHSOffsetValue = RHSOffsetCI->getValue();
2105 if (!LHSOffsetValue.isNegative() &&
2106 !RHSOffsetValue.isNegative() &&
2107 LHSOffsetValue.ult(LHSSize) &&
2108 RHSOffsetValue.ult(RHSSize)) {
2109 return ConstantInt::get(GetCompareTy(LHS),
2110 !CmpInst::isTrueWhenEqual(Pred));
2111 }
2112 }
2113
2114 // Repeat the above check but this time without depending on DataLayout
2115 // or being able to compute a precise size.
2116 if (!cast<PointerType>(LHS->getType())->isEmptyTy() &&
2117 !cast<PointerType>(RHS->getType())->isEmptyTy() &&
2118 LHSOffset->isNullValue() &&
2119 RHSOffset->isNullValue())
2120 return ConstantInt::get(GetCompareTy(LHS),
2121 !CmpInst::isTrueWhenEqual(Pred));
2122 }
2123
2124 // Even if an non-inbounds GEP occurs along the path we can still optimize
2125 // equality comparisons concerning the result. We avoid walking the whole
2126 // chain again by starting where the last calls to
2127 // stripAndComputeConstantOffsets left off and accumulate the offsets.
2128 Constant *LHSNoBound = stripAndComputeConstantOffsets(DL, LHS, true);
2129 Constant *RHSNoBound = stripAndComputeConstantOffsets(DL, RHS, true);
2130 if (LHS == RHS)
2131 return ConstantExpr::getICmp(Pred,
2132 ConstantExpr::getAdd(LHSOffset, LHSNoBound),
2133 ConstantExpr::getAdd(RHSOffset, RHSNoBound));
2134
2135 // If one side of the equality comparison must come from a noalias call
2136 // (meaning a system memory allocation function), and the other side must
2137 // come from a pointer that cannot overlap with dynamically-allocated
2138 // memory within the lifetime of the current function (allocas, byval
2139 // arguments, globals), then determine the comparison result here.
2140 SmallVector<Value *, 8> LHSUObjs, RHSUObjs;
2141 GetUnderlyingObjects(LHS, LHSUObjs, DL);
2142 GetUnderlyingObjects(RHS, RHSUObjs, DL);
2143
2144 // Is the set of underlying objects all noalias calls?
2145 auto IsNAC = [](ArrayRef<Value *> Objects) {
2146 return all_of(Objects, isNoAliasCall);
2147 };
2148
2149 // Is the set of underlying objects all things which must be disjoint from
2150 // noalias calls. For allocas, we consider only static ones (dynamic
2151 // allocas might be transformed into calls to malloc not simultaneously
2152 // live with the compared-to allocation). For globals, we exclude symbols
2153 // that might be resolve lazily to symbols in another dynamically-loaded
2154 // library (and, thus, could be malloc'ed by the implementation).
2155 auto IsAllocDisjoint = [](ArrayRef<Value *> Objects) {
2156 return all_of(Objects, [](Value *V) {
2157 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
2158 return AI->getParent() && AI->getFunction() && AI->isStaticAlloca();
2159 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V))
2160 return (GV->hasLocalLinkage() || GV->hasHiddenVisibility() ||
2161 GV->hasProtectedVisibility() || GV->hasGlobalUnnamedAddr()) &&
2162 !GV->isThreadLocal();
2163 if (const Argument *A = dyn_cast<Argument>(V))
2164 return A->hasByValAttr();
2165 return false;
2166 });
2167 };
2168
2169 if ((IsNAC(LHSUObjs) && IsAllocDisjoint(RHSUObjs)) ||
2170 (IsNAC(RHSUObjs) && IsAllocDisjoint(LHSUObjs)))
2171 return ConstantInt::get(GetCompareTy(LHS),
2172 !CmpInst::isTrueWhenEqual(Pred));
2173
2174 // Fold comparisons for non-escaping pointer even if the allocation call
2175 // cannot be elided. We cannot fold malloc comparison to null. Also, the
2176 // dynamic allocation call could be either of the operands.
2177 Value *MI = nullptr;
2178 if (isAllocLikeFn(LHS, TLI) &&
2179 llvm::isKnownNonZero(RHS, DL, 0, nullptr, CxtI, DT))
2180 MI = LHS;
2181 else if (isAllocLikeFn(RHS, TLI) &&
2182 llvm::isKnownNonZero(LHS, DL, 0, nullptr, CxtI, DT))
2183 MI = RHS;
2184 // FIXME: We should also fold the compare when the pointer escapes, but the
2185 // compare dominates the pointer escape
2186 if (MI && !PointerMayBeCaptured(MI, true, true))
2187 return ConstantInt::get(GetCompareTy(LHS),
2188 CmpInst::isFalseWhenEqual(Pred));
2189 }
2190
2191 // Otherwise, fail.
2192 return nullptr;
2193}
2194
2195/// Fold an icmp when its operands have i1 scalar type.
2196static Value *simplifyICmpOfBools(CmpInst::Predicate Pred, Value *LHS,
2197 Value *RHS, const SimplifyQuery &Q) {
2198 Type *ITy = GetCompareTy(LHS); // The return type.
2199 Type *OpTy = LHS->getType(); // The operand type.
2200 if (!OpTy->isIntOrIntVectorTy(1))
2201 return nullptr;
2202
2203 // A boolean compared to true/false can be simplified in 14 out of the 20
2204 // (10 predicates * 2 constants) possible combinations. Cases not handled here
2205 // require a 'not' of the LHS, so those must be transformed in InstCombine.
2206 if (match(RHS, m_Zero())) {
2207 switch (Pred) {
2208 case CmpInst::ICMP_NE: // X != 0 -> X
2209 case CmpInst::ICMP_UGT: // X >u 0 -> X
2210 case CmpInst::ICMP_SLT: // X <s 0 -> X
2211 return LHS;
2212
2213 case CmpInst::ICMP_ULT: // X <u 0 -> false
2214 case CmpInst::ICMP_SGT: // X >s 0 -> false
2215 return getFalse(ITy);
2216
2217 case CmpInst::ICMP_UGE: // X >=u 0 -> true
2218 case CmpInst::ICMP_SLE: // X <=s 0 -> true
2219 return getTrue(ITy);
2220
2221 default: break;
2222 }
2223 } else if (match(RHS, m_One())) {
2224 switch (Pred) {
2225 case CmpInst::ICMP_EQ: // X == 1 -> X
2226 case CmpInst::ICMP_UGE: // X >=u 1 -> X
2227 case CmpInst::ICMP_SLE: // X <=s -1 -> X
2228 return LHS;
2229
2230 case CmpInst::ICMP_UGT: // X >u 1 -> false
2231 case CmpInst::ICMP_SLT: // X <s -1 -> false
2232 return getFalse(ITy);
2233
2234 case CmpInst::ICMP_ULE: // X <=u 1 -> true
2235 case CmpInst::ICMP_SGE: // X >=s -1 -> true
2236 return getTrue(ITy);
2237
2238 default: break;
2239 }
2240 }
2241
2242 switch (Pred) {
2243 default:
2244 break;
2245 case ICmpInst::ICMP_UGE:
2246 if (isImpliedCondition(RHS, LHS, Q.DL).getValueOr(false))
2247 return getTrue(ITy);
2248 break;
2249 case ICmpInst::ICMP_SGE:
2250 /// For signed comparison, the values for an i1 are 0 and -1
2251 /// respectively. This maps into a truth table of:
2252 /// LHS | RHS | LHS >=s RHS | LHS implies RHS
2253 /// 0 | 0 | 1 (0 >= 0) | 1
2254 /// 0 | 1 | 1 (0 >= -1) | 1
2255 /// 1 | 0 | 0 (-1 >= 0) | 0
2256 /// 1 | 1 | 1 (-1 >= -1) | 1
2257 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2258 return getTrue(ITy);
2259 break;
2260 case ICmpInst::ICMP_ULE:
2261 if (isImpliedCondition(LHS, RHS, Q.DL).getValueOr(false))
2262 return getTrue(ITy);
2263 break;
2264 }
2265
2266 return nullptr;
2267}
2268
2269/// Try hard to fold icmp with zero RHS because this is a common case.
2270static Value *simplifyICmpWithZero(CmpInst::Predicate Pred, Value *LHS,
2271 Value *RHS, const SimplifyQuery &Q) {
2272 if (!match(RHS, m_Zero()))
2273 return nullptr;
2274
2275 Type *ITy = GetCompareTy(LHS); // The return type.
2276 switch (Pred) {
2277 default:
2278 llvm_unreachable("Unknown ICmp predicate!")::llvm::llvm_unreachable_internal("Unknown ICmp predicate!", "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 2278)
;
2279 case ICmpInst::ICMP_ULT:
2280 return getFalse(ITy);
2281 case ICmpInst::ICMP_UGE:
2282 return getTrue(ITy);
2283 case ICmpInst::ICMP_EQ:
2284 case ICmpInst::ICMP_ULE:
2285 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2286 return getFalse(ITy);
2287 break;
2288 case ICmpInst::ICMP_NE:
2289 case ICmpInst::ICMP_UGT:
2290 if (isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2291 return getTrue(ITy);
2292 break;
2293 case ICmpInst::ICMP_SLT: {
2294 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2295 if (LHSKnown.isNegative())
2296 return getTrue(ITy);
2297 if (LHSKnown.isNonNegative())
2298 return getFalse(ITy);
2299 break;
2300 }
2301 case ICmpInst::ICMP_SLE: {
2302 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2303 if (LHSKnown.isNegative())
2304 return getTrue(ITy);
2305 if (LHSKnown.isNonNegative() &&
2306 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2307 return getFalse(ITy);
2308 break;
2309 }
2310 case ICmpInst::ICMP_SGE: {
2311 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2312 if (LHSKnown.isNegative())
2313 return getFalse(ITy);
2314 if (LHSKnown.isNonNegative())
2315 return getTrue(ITy);
2316 break;
2317 }
2318 case ICmpInst::ICMP_SGT: {
2319 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2320 if (LHSKnown.isNegative())
2321 return getFalse(ITy);
2322 if (LHSKnown.isNonNegative() &&
2323 isKnownNonZero(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT))
2324 return getTrue(ITy);
2325 break;
2326 }
2327 }
2328
2329 return nullptr;
2330}
2331
2332/// Many binary operators with a constant operand have an easy-to-compute
2333/// range of outputs. This can be used to fold a comparison to always true or
2334/// always false.
2335static void setLimitsForBinOp(BinaryOperator &BO, APInt &Lower, APInt &Upper) {
2336 unsigned Width = Lower.getBitWidth();
2337 const APInt *C;
2338 switch (BO.getOpcode()) {
2339 case Instruction::Add:
2340 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2341 // FIXME: If we have both nuw and nsw, we should reduce the range further.
2342 if (BO.hasNoUnsignedWrap()) {
2343 // 'add nuw x, C' produces [C, UINT_MAX].
2344 Lower = *C;
2345 } else if (BO.hasNoSignedWrap()) {
2346 if (C->isNegative()) {
2347 // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C].
2348 Lower = APInt::getSignedMinValue(Width);
2349 Upper = APInt::getSignedMaxValue(Width) + *C + 1;
2350 } else {
2351 // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX].
2352 Lower = APInt::getSignedMinValue(Width) + *C;
2353 Upper = APInt::getSignedMaxValue(Width) + 1;
2354 }
2355 }
2356 }
2357 break;
2358
2359 case Instruction::And:
2360 if (match(BO.getOperand(1), m_APInt(C)))
2361 // 'and x, C' produces [0, C].
2362 Upper = *C + 1;
2363 break;
2364
2365 case Instruction::Or:
2366 if (match(BO.getOperand(1), m_APInt(C)))
2367 // 'or x, C' produces [C, UINT_MAX].
2368 Lower = *C;
2369 break;
2370
2371 case Instruction::AShr:
2372 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2373 // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C].
2374 Lower = APInt::getSignedMinValue(Width).ashr(*C);
2375 Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1;
2376 } else if (match(BO.getOperand(0), m_APInt(C))) {
2377 unsigned ShiftAmount = Width - 1;
2378 if (!C->isNullValue() && BO.isExact())
2379 ShiftAmount = C->countTrailingZeros();
2380 if (C->isNegative()) {
2381 // 'ashr C, x' produces [C, C >> (Width-1)]
2382 Lower = *C;
2383 Upper = C->ashr(ShiftAmount) + 1;
2384 } else {
2385 // 'ashr C, x' produces [C >> (Width-1), C]
2386 Lower = C->ashr(ShiftAmount);
2387 Upper = *C + 1;
2388 }
2389 }
2390 break;
2391
2392 case Instruction::LShr:
2393 if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) {
2394 // 'lshr x, C' produces [0, UINT_MAX >> C].
2395 Upper = APInt::getAllOnesValue(Width).lshr(*C) + 1;
2396 } else if (match(BO.getOperand(0), m_APInt(C))) {
2397 // 'lshr C, x' produces [C >> (Width-1), C].
2398 unsigned ShiftAmount = Width - 1;
2399 if (!C->isNullValue() && BO.isExact())
2400 ShiftAmount = C->countTrailingZeros();
2401 Lower = C->lshr(ShiftAmount);
2402 Upper = *C + 1;
2403 }
2404 break;
2405
2406 case Instruction::Shl:
2407 if (match(BO.getOperand(0), m_APInt(C))) {
2408 if (BO.hasNoUnsignedWrap()) {
2409 // 'shl nuw C, x' produces [C, C << CLZ(C)]
2410 Lower = *C;
2411 Upper = Lower.shl(Lower.countLeadingZeros()) + 1;
2412 } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw?
2413 if (C->isNegative()) {
2414 // 'shl nsw C, x' produces [C << CLO(C)-1, C]
2415 unsigned ShiftAmount = C->countLeadingOnes() - 1;
2416 Lower = C->shl(ShiftAmount);
2417 Upper = *C + 1;
2418 } else {
2419 // 'shl nsw C, x' produces [C, C << CLZ(C)-1]
2420 unsigned ShiftAmount = C->countLeadingZeros() - 1;
2421 Lower = *C;
2422 Upper = C->shl(ShiftAmount) + 1;
2423 }
2424 }
2425 }
2426 break;
2427
2428 case Instruction::SDiv:
2429 if (match(BO.getOperand(1), m_APInt(C))) {
2430 APInt IntMin = APInt::getSignedMinValue(Width);
2431 APInt IntMax = APInt::getSignedMaxValue(Width);
2432 if (C->isAllOnesValue()) {
2433 // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX]
2434 // where C != -1 and C != 0 and C != 1
2435 Lower = IntMin + 1;
2436 Upper = IntMax + 1;
2437 } else if (C->countLeadingZeros() < Width - 1) {
2438 // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C]
2439 // where C != -1 and C != 0 and C != 1
2440 Lower = IntMin.sdiv(*C);
2441 Upper = IntMax.sdiv(*C);
2442 if (Lower.sgt(Upper))
2443 std::swap(Lower, Upper);
2444 Upper = Upper + 1;
2445 assert(Upper != Lower && "Upper part of range has wrapped!")(static_cast <bool> (Upper != Lower && "Upper part of range has wrapped!"
) ? void (0) : __assert_fail ("Upper != Lower && \"Upper part of range has wrapped!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 2445, __extension__ __PRETTY_FUNCTION__))
;
2446 }
2447 } else if (match(BO.getOperand(0), m_APInt(C))) {
2448 if (C->isMinSignedValue()) {
2449 // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2].
2450 Lower = *C;
2451 Upper = Lower.lshr(1) + 1;
2452 } else {
2453 // 'sdiv C, x' produces [-|C|, |C|].
2454 Upper = C->abs() + 1;
2455 Lower = (-Upper) + 1;
2456 }
2457 }
2458 break;
2459
2460 case Instruction::UDiv:
2461 if (match(BO.getOperand(1), m_APInt(C)) && !C->isNullValue()) {
2462 // 'udiv x, C' produces [0, UINT_MAX / C].
2463 Upper = APInt::getMaxValue(Width).udiv(*C) + 1;
2464 } else if (match(BO.getOperand(0), m_APInt(C))) {
2465 // 'udiv C, x' produces [0, C].
2466 Upper = *C + 1;
2467 }
2468 break;
2469
2470 case Instruction::SRem:
2471 if (match(BO.getOperand(1), m_APInt(C))) {
2472 // 'srem x, C' produces (-|C|, |C|).
2473 Upper = C->abs();
2474 Lower = (-Upper) + 1;
2475 }
2476 break;
2477
2478 case Instruction::URem:
2479 if (match(BO.getOperand(1), m_APInt(C)))
2480 // 'urem x, C' produces [0, C).
2481 Upper = *C;
2482 break;
2483
2484 default:
2485 break;
2486 }
2487}
2488
2489static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS,
2490 Value *RHS) {
2491 const APInt *C;
2492 if (!match(RHS, m_APInt(C)))
2493 return nullptr;
2494
2495 // Rule out tautological comparisons (eg., ult 0 or uge 0).
2496 ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C);
2497 if (RHS_CR.isEmptySet())
2498 return ConstantInt::getFalse(GetCompareTy(RHS));
2499 if (RHS_CR.isFullSet())
2500 return ConstantInt::getTrue(GetCompareTy(RHS));
2501
2502 // Find the range of possible values for binary operators.
2503 unsigned Width = C->getBitWidth();
2504 APInt Lower = APInt(Width, 0);
2505 APInt Upper = APInt(Width, 0);
2506 if (auto *BO = dyn_cast<BinaryOperator>(LHS))
2507 setLimitsForBinOp(*BO, Lower, Upper);
2508
2509 ConstantRange LHS_CR =
2510 Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true);
2511
2512 if (auto *I = dyn_cast<Instruction>(LHS))
2513 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range))
2514 LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges));
2515
2516 if (!LHS_CR.isFullSet()) {
2517 if (RHS_CR.contains(LHS_CR))
2518 return ConstantInt::getTrue(GetCompareTy(RHS));
2519 if (RHS_CR.inverse().contains(LHS_CR))
2520 return ConstantInt::getFalse(GetCompareTy(RHS));
2521 }
2522
2523 return nullptr;
2524}
2525
2526/// TODO: A large part of this logic is duplicated in InstCombine's
2527/// foldICmpBinOp(). We should be able to share that and avoid the code
2528/// duplication.
2529static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS,
2530 Value *RHS, const SimplifyQuery &Q,
2531 unsigned MaxRecurse) {
2532 Type *ITy = GetCompareTy(LHS); // The return type.
2533
2534 BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS);
2535 BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS);
2536 if (MaxRecurse && (LBO || RBO)) {
2537 // Analyze the case when either LHS or RHS is an add instruction.
2538 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
2539 // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null).
2540 bool NoLHSWrapProblem = false, NoRHSWrapProblem = false;
2541 if (LBO && LBO->getOpcode() == Instruction::Add) {
2542 A = LBO->getOperand(0);
2543 B = LBO->getOperand(1);
2544 NoLHSWrapProblem =
2545 ICmpInst::isEquality(Pred) ||
2546 (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) ||
2547 (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap());
2548 }
2549 if (RBO && RBO->getOpcode() == Instruction::Add) {
2550 C = RBO->getOperand(0);
2551 D = RBO->getOperand(1);
2552 NoRHSWrapProblem =
2553 ICmpInst::isEquality(Pred) ||
2554 (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) ||
2555 (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap());
2556 }
2557
2558 // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow.
2559 if ((A == RHS || B == RHS) && NoLHSWrapProblem)
2560 if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A,
2561 Constant::getNullValue(RHS->getType()), Q,
2562 MaxRecurse - 1))
2563 return V;
2564
2565 // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow.
2566 if ((C == LHS || D == LHS) && NoRHSWrapProblem)
2567 if (Value *V =
2568 SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()),
2569 C == LHS ? D : C, Q, MaxRecurse - 1))
2570 return V;
2571
2572 // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow.
2573 if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem &&
2574 NoRHSWrapProblem) {
2575 // Determine Y and Z in the form icmp (X+Y), (X+Z).
2576 Value *Y, *Z;
2577 if (A == C) {
2578 // C + B == C + D -> B == D
2579 Y = B;
2580 Z = D;
2581 } else if (A == D) {
2582 // D + B == C + D -> B == C
2583 Y = B;
2584 Z = C;
2585 } else if (B == C) {
2586 // A + C == C + D -> A == D
2587 Y = A;
2588 Z = D;
2589 } else {
2590 assert(B == D)(static_cast <bool> (B == D) ? void (0) : __assert_fail
("B == D", "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 2590, __extension__ __PRETTY_FUNCTION__))
;
2591 // A + D == C + D -> A == C
2592 Y = A;
2593 Z = C;
2594 }
2595 if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1))
2596 return V;
2597 }
2598 }
2599
2600 {
2601 Value *Y = nullptr;
2602 // icmp pred (or X, Y), X
2603 if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) {
2604 if (Pred == ICmpInst::ICMP_ULT)
2605 return getFalse(ITy);
2606 if (Pred == ICmpInst::ICMP_UGE)
2607 return getTrue(ITy);
2608
2609 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) {
2610 KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2611 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2612 if (RHSKnown.isNonNegative() && YKnown.isNegative())
2613 return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy);
2614 if (RHSKnown.isNegative() || YKnown.isNonNegative())
2615 return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy);
2616 }
2617 }
2618 // icmp pred X, (or X, Y)
2619 if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) {
2620 if (Pred == ICmpInst::ICMP_ULE)
2621 return getTrue(ITy);
2622 if (Pred == ICmpInst::ICMP_UGT)
2623 return getFalse(ITy);
2624
2625 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) {
2626 KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2627 KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2628 if (LHSKnown.isNonNegative() && YKnown.isNegative())
2629 return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy);
2630 if (LHSKnown.isNegative() || YKnown.isNonNegative())
2631 return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy);
2632 }
2633 }
2634 }
2635
2636 // icmp pred (and X, Y), X
2637 if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) {
2638 if (Pred == ICmpInst::ICMP_UGT)
2639 return getFalse(ITy);
2640 if (Pred == ICmpInst::ICMP_ULE)
2641 return getTrue(ITy);
2642 }
2643 // icmp pred X, (and X, Y)
2644 if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) {
2645 if (Pred == ICmpInst::ICMP_UGE)
2646 return getTrue(ITy);
2647 if (Pred == ICmpInst::ICMP_ULT)
2648 return getFalse(ITy);
2649 }
2650
2651 // 0 - (zext X) pred C
2652 if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) {
2653 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2654 if (RHSC->getValue().isStrictlyPositive()) {
2655 if (Pred == ICmpInst::ICMP_SLT)
2656 return ConstantInt::getTrue(RHSC->getContext());
2657 if (Pred == ICmpInst::ICMP_SGE)
2658 return ConstantInt::getFalse(RHSC->getContext());
2659 if (Pred == ICmpInst::ICMP_EQ)
2660 return ConstantInt::getFalse(RHSC->getContext());
2661 if (Pred == ICmpInst::ICMP_NE)
2662 return ConstantInt::getTrue(RHSC->getContext());
2663 }
2664 if (RHSC->getValue().isNonNegative()) {
2665 if (Pred == ICmpInst::ICMP_SLE)
2666 return ConstantInt::getTrue(RHSC->getContext());
2667 if (Pred == ICmpInst::ICMP_SGT)
2668 return ConstantInt::getFalse(RHSC->getContext());
2669 }
2670 }
2671 }
2672
2673 // icmp pred (urem X, Y), Y
2674 if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) {
2675 switch (Pred) {
2676 default:
2677 break;
2678 case ICmpInst::ICMP_SGT:
2679 case ICmpInst::ICMP_SGE: {
2680 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2681 if (!Known.isNonNegative())
2682 break;
2683 LLVM_FALLTHROUGH[[clang::fallthrough]];
2684 }
2685 case ICmpInst::ICMP_EQ:
2686 case ICmpInst::ICMP_UGT:
2687 case ICmpInst::ICMP_UGE:
2688 return getFalse(ITy);
2689 case ICmpInst::ICMP_SLT:
2690 case ICmpInst::ICMP_SLE: {
2691 KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2692 if (!Known.isNonNegative())
2693 break;
2694 LLVM_FALLTHROUGH[[clang::fallthrough]];
2695 }
2696 case ICmpInst::ICMP_NE:
2697 case ICmpInst::ICMP_ULT:
2698 case ICmpInst::ICMP_ULE:
2699 return getTrue(ITy);
2700 }
2701 }
2702
2703 // icmp pred X, (urem Y, X)
2704 if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) {
2705 switch (Pred) {
2706 default:
2707 break;
2708 case ICmpInst::ICMP_SGT:
2709 case ICmpInst::ICMP_SGE: {
2710 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2711 if (!Known.isNonNegative())
2712 break;
2713 LLVM_FALLTHROUGH[[clang::fallthrough]];
2714 }
2715 case ICmpInst::ICMP_NE:
2716 case ICmpInst::ICMP_UGT:
2717 case ICmpInst::ICMP_UGE:
2718 return getTrue(ITy);
2719 case ICmpInst::ICMP_SLT:
2720 case ICmpInst::ICMP_SLE: {
2721 KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT);
2722 if (!Known.isNonNegative())
2723 break;
2724 LLVM_FALLTHROUGH[[clang::fallthrough]];
2725 }
2726 case ICmpInst::ICMP_EQ:
2727 case ICmpInst::ICMP_ULT:
2728 case ICmpInst::ICMP_ULE:
2729 return getFalse(ITy);
2730 }
2731 }
2732
2733 // x >> y <=u x
2734 // x udiv y <=u x.
2735 if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) ||
2736 match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) {
2737 // icmp pred (X op Y), X
2738 if (Pred == ICmpInst::ICMP_UGT)
2739 return getFalse(ITy);
2740 if (Pred == ICmpInst::ICMP_ULE)
2741 return getTrue(ITy);
2742 }
2743
2744 // x >=u x >> y
2745 // x >=u x udiv y.
2746 if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) ||
2747 match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) {
2748 // icmp pred X, (X op Y)
2749 if (Pred == ICmpInst::ICMP_ULT)
2750 return getFalse(ITy);
2751 if (Pred == ICmpInst::ICMP_UGE)
2752 return getTrue(ITy);
2753 }
2754
2755 // handle:
2756 // CI2 << X == CI
2757 // CI2 << X != CI
2758 //
2759 // where CI2 is a power of 2 and CI isn't
2760 if (auto *CI = dyn_cast<ConstantInt>(RHS)) {
2761 const APInt *CI2Val, *CIVal = &CI->getValue();
2762 if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) &&
2763 CI2Val->isPowerOf2()) {
2764 if (!CIVal->isPowerOf2()) {
2765 // CI2 << X can equal zero in some circumstances,
2766 // this simplification is unsafe if CI is zero.
2767 //
2768 // We know it is safe if:
2769 // - The shift is nsw, we can't shift out the one bit.
2770 // - The shift is nuw, we can't shift out the one bit.
2771 // - CI2 is one
2772 // - CI isn't zero
2773 if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() ||
2774 CI2Val->isOneValue() || !CI->isZero()) {
2775 if (Pred == ICmpInst::ICMP_EQ)
2776 return ConstantInt::getFalse(RHS->getContext());
2777 if (Pred == ICmpInst::ICMP_NE)
2778 return ConstantInt::getTrue(RHS->getContext());
2779 }
2780 }
2781 if (CIVal->isSignMask() && CI2Val->isOneValue()) {
2782 if (Pred == ICmpInst::ICMP_UGT)
2783 return ConstantInt::getFalse(RHS->getContext());
2784 if (Pred == ICmpInst::ICMP_ULE)
2785 return ConstantInt::getTrue(RHS->getContext());
2786 }
2787 }
2788 }
2789
2790 if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() &&
2791 LBO->getOperand(1) == RBO->getOperand(1)) {
2792 switch (LBO->getOpcode()) {
2793 default:
2794 break;
2795 case Instruction::UDiv:
2796 case Instruction::LShr:
2797 if (ICmpInst::isSigned(Pred) || !LBO->isExact() || !RBO->isExact())
2798 break;
2799 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2800 RBO->getOperand(0), Q, MaxRecurse - 1))
2801 return V;
2802 break;
2803 case Instruction::SDiv:
2804 if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact())
2805 break;
2806 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2807 RBO->getOperand(0), Q, MaxRecurse - 1))
2808 return V;
2809 break;
2810 case Instruction::AShr:
2811 if (!LBO->isExact() || !RBO->isExact())
2812 break;
2813 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2814 RBO->getOperand(0), Q, MaxRecurse - 1))
2815 return V;
2816 break;
2817 case Instruction::Shl: {
2818 bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap();
2819 bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap();
2820 if (!NUW && !NSW)
2821 break;
2822 if (!NSW && ICmpInst::isSigned(Pred))
2823 break;
2824 if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0),
2825 RBO->getOperand(0), Q, MaxRecurse - 1))
2826 return V;
2827 break;
2828 }
2829 }
2830 }
2831 return nullptr;
2832}
2833
2834/// Simplify integer comparisons where at least one operand of the compare
2835/// matches an integer min/max idiom.
2836static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS,
2837 Value *RHS, const SimplifyQuery &Q,
2838 unsigned MaxRecurse) {
2839 Type *ITy = GetCompareTy(LHS); // The return type.
2840 Value *A, *B;
2841 CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE;
2842 CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B".
2843
2844 // Signed variants on "max(a,b)>=a -> true".
2845 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2846 if (A != RHS)
2847 std::swap(A, B); // smax(A, B) pred A.
2848 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2849 // We analyze this as smax(A, B) pred A.
2850 P = Pred;
2851 } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) &&
2852 (A == LHS || B == LHS)) {
2853 if (A != LHS)
2854 std::swap(A, B); // A pred smax(A, B).
2855 EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B".
2856 // We analyze this as smax(A, B) swapped-pred A.
2857 P = CmpInst::getSwappedPredicate(Pred);
2858 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
2859 (A == RHS || B == RHS)) {
2860 if (A != RHS)
2861 std::swap(A, B); // smin(A, B) pred A.
2862 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2863 // We analyze this as smax(-A, -B) swapped-pred -A.
2864 // Note that we do not need to actually form -A or -B thanks to EqP.
2865 P = CmpInst::getSwappedPredicate(Pred);
2866 } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) &&
2867 (A == LHS || B == LHS)) {
2868 if (A != LHS)
2869 std::swap(A, B); // A pred smin(A, B).
2870 EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B".
2871 // We analyze this as smax(-A, -B) pred -A.
2872 // Note that we do not need to actually form -A or -B thanks to EqP.
2873 P = Pred;
2874 }
2875 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2876 // Cases correspond to "max(A, B) p A".
2877 switch (P) {
2878 default:
2879 break;
2880 case CmpInst::ICMP_EQ:
2881 case CmpInst::ICMP_SLE:
2882 // Equivalent to "A EqP B". This may be the same as the condition tested
2883 // in the max/min; if so, we can just return that.
2884 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2885 return V;
2886 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2887 return V;
2888 // Otherwise, see if "A EqP B" simplifies.
2889 if (MaxRecurse)
2890 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2891 return V;
2892 break;
2893 case CmpInst::ICMP_NE:
2894 case CmpInst::ICMP_SGT: {
2895 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2896 // Equivalent to "A InvEqP B". This may be the same as the condition
2897 // tested in the max/min; if so, we can just return that.
2898 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2899 return V;
2900 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2901 return V;
2902 // Otherwise, see if "A InvEqP B" simplifies.
2903 if (MaxRecurse)
2904 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2905 return V;
2906 break;
2907 }
2908 case CmpInst::ICMP_SGE:
2909 // Always true.
2910 return getTrue(ITy);
2911 case CmpInst::ICMP_SLT:
2912 // Always false.
2913 return getFalse(ITy);
2914 }
2915 }
2916
2917 // Unsigned variants on "max(a,b)>=a -> true".
2918 P = CmpInst::BAD_ICMP_PREDICATE;
2919 if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) {
2920 if (A != RHS)
2921 std::swap(A, B); // umax(A, B) pred A.
2922 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2923 // We analyze this as umax(A, B) pred A.
2924 P = Pred;
2925 } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) &&
2926 (A == LHS || B == LHS)) {
2927 if (A != LHS)
2928 std::swap(A, B); // A pred umax(A, B).
2929 EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B".
2930 // We analyze this as umax(A, B) swapped-pred A.
2931 P = CmpInst::getSwappedPredicate(Pred);
2932 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
2933 (A == RHS || B == RHS)) {
2934 if (A != RHS)
2935 std::swap(A, B); // umin(A, B) pred A.
2936 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2937 // We analyze this as umax(-A, -B) swapped-pred -A.
2938 // Note that we do not need to actually form -A or -B thanks to EqP.
2939 P = CmpInst::getSwappedPredicate(Pred);
2940 } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) &&
2941 (A == LHS || B == LHS)) {
2942 if (A != LHS)
2943 std::swap(A, B); // A pred umin(A, B).
2944 EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B".
2945 // We analyze this as umax(-A, -B) pred -A.
2946 // Note that we do not need to actually form -A or -B thanks to EqP.
2947 P = Pred;
2948 }
2949 if (P != CmpInst::BAD_ICMP_PREDICATE) {
2950 // Cases correspond to "max(A, B) p A".
2951 switch (P) {
2952 default:
2953 break;
2954 case CmpInst::ICMP_EQ:
2955 case CmpInst::ICMP_ULE:
2956 // Equivalent to "A EqP B". This may be the same as the condition tested
2957 // in the max/min; if so, we can just return that.
2958 if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B))
2959 return V;
2960 if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B))
2961 return V;
2962 // Otherwise, see if "A EqP B" simplifies.
2963 if (MaxRecurse)
2964 if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1))
2965 return V;
2966 break;
2967 case CmpInst::ICMP_NE:
2968 case CmpInst::ICMP_UGT: {
2969 CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP);
2970 // Equivalent to "A InvEqP B". This may be the same as the condition
2971 // tested in the max/min; if so, we can just return that.
2972 if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B))
2973 return V;
2974 if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B))
2975 return V;
2976 // Otherwise, see if "A InvEqP B" simplifies.
2977 if (MaxRecurse)
2978 if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1))
2979 return V;
2980 break;
2981 }
2982 case CmpInst::ICMP_UGE:
2983 // Always true.
2984 return getTrue(ITy);
2985 case CmpInst::ICMP_ULT:
2986 // Always false.
2987 return getFalse(ITy);
2988 }
2989 }
2990
2991 // Variants on "max(x,y) >= min(x,z)".
2992 Value *C, *D;
2993 if (match(LHS, m_SMax(m_Value(A), m_Value(B))) &&
2994 match(RHS, m_SMin(m_Value(C), m_Value(D))) &&
2995 (A == C || A == D || B == C || B == D)) {
2996 // max(x, ?) pred min(x, ?).
2997 if (Pred == CmpInst::ICMP_SGE)
2998 // Always true.
2999 return getTrue(ITy);
3000 if (Pred == CmpInst::ICMP_SLT)
3001 // Always false.
3002 return getFalse(ITy);
3003 } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) &&
3004 match(RHS, m_SMax(m_Value(C), m_Value(D))) &&
3005 (A == C || A == D || B == C || B == D)) {
3006 // min(x, ?) pred max(x, ?).
3007 if (Pred == CmpInst::ICMP_SLE)
3008 // Always true.
3009 return getTrue(ITy);
3010 if (Pred == CmpInst::ICMP_SGT)
3011 // Always false.
3012 return getFalse(ITy);
3013 } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) &&
3014 match(RHS, m_UMin(m_Value(C), m_Value(D))) &&
3015 (A == C || A == D || B == C || B == D)) {
3016 // max(x, ?) pred min(x, ?).
3017 if (Pred == CmpInst::ICMP_UGE)
3018 // Always true.
3019 return getTrue(ITy);
3020 if (Pred == CmpInst::ICMP_ULT)
3021 // Always false.
3022 return getFalse(ITy);
3023 } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) &&
3024 match(RHS, m_UMax(m_Value(C), m_Value(D))) &&
3025 (A == C || A == D || B == C || B == D)) {
3026 // min(x, ?) pred max(x, ?).
3027 if (Pred == CmpInst::ICMP_ULE)
3028 // Always true.
3029 return getTrue(ITy);
3030 if (Pred == CmpInst::ICMP_UGT)
3031 // Always false.
3032 return getFalse(ITy);
3033 }
3034
3035 return nullptr;
3036}
3037
3038/// Given operands for an ICmpInst, see if we can fold the result.
3039/// If not, this returns null.
3040static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3041 const SimplifyQuery &Q, unsigned MaxRecurse) {
3042 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3043 assert(CmpInst::isIntPredicate(Pred) && "Not an integer compare!")(static_cast <bool> (CmpInst::isIntPredicate(Pred) &&
"Not an integer compare!") ? void (0) : __assert_fail ("CmpInst::isIntPredicate(Pred) && \"Not an integer compare!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 3043, __extension__ __PRETTY_FUNCTION__))
;
3044
3045 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3046 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3047 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3048
3049 // If we have a constant, make sure it is on the RHS.
3050 std::swap(LHS, RHS);
3051 Pred = CmpInst::getSwappedPredicate(Pred);
3052 }
3053
3054 Type *ITy = GetCompareTy(LHS); // The return type.
3055
3056 // icmp X, X -> true/false
3057 // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false
3058 // because X could be 0.
3059 if (LHS == RHS || isa<UndefValue>(RHS))
3060 return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred));
3061
3062 if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q))
3063 return V;
3064
3065 if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q))
3066 return V;
3067
3068 if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS))
3069 return V;
3070
3071 // If both operands have range metadata, use the metadata
3072 // to simplify the comparison.
3073 if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) {
3074 auto RHS_Instr = cast<Instruction>(RHS);
3075 auto LHS_Instr = cast<Instruction>(LHS);
3076
3077 if (RHS_Instr->getMetadata(LLVMContext::MD_range) &&
3078 LHS_Instr->getMetadata(LLVMContext::MD_range)) {
3079 auto RHS_CR = getConstantRangeFromMetadata(
3080 *RHS_Instr->getMetadata(LLVMContext::MD_range));
3081 auto LHS_CR = getConstantRangeFromMetadata(
3082 *LHS_Instr->getMetadata(LLVMContext::MD_range));
3083
3084 auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR);
3085 if (Satisfied_CR.contains(LHS_CR))
3086 return ConstantInt::getTrue(RHS->getContext());
3087
3088 auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion(
3089 CmpInst::getInversePredicate(Pred), RHS_CR);
3090 if (InversedSatisfied_CR.contains(LHS_CR))
3091 return ConstantInt::getFalse(RHS->getContext());
3092 }
3093 }
3094
3095 // Compare of cast, for example (zext X) != 0 -> X != 0
3096 if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) {
3097 Instruction *LI = cast<CastInst>(LHS);
3098 Value *SrcOp = LI->getOperand(0);
3099 Type *SrcTy = SrcOp->getType();
3100 Type *DstTy = LI->getType();
3101
3102 // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input
3103 // if the integer type is the same size as the pointer type.
3104 if (MaxRecurse && isa<PtrToIntInst>(LI) &&
3105 Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) {
3106 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
3107 // Transfer the cast to the constant.
3108 if (Value *V = SimplifyICmpInst(Pred, SrcOp,
3109 ConstantExpr::getIntToPtr(RHSC, SrcTy),
3110 Q, MaxRecurse-1))
3111 return V;
3112 } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) {
3113 if (RI->getOperand(0)->getType() == SrcTy)
3114 // Compare without the cast.
3115 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3116 Q, MaxRecurse-1))
3117 return V;
3118 }
3119 }
3120
3121 if (isa<ZExtInst>(LHS)) {
3122 // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the
3123 // same type.
3124 if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) {
3125 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3126 // Compare X and Y. Note that signed predicates become unsigned.
3127 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3128 SrcOp, RI->getOperand(0), Q,
3129 MaxRecurse-1))
3130 return V;
3131 }
3132 // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended
3133 // too. If not, then try to deduce the result of the comparison.
3134 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3135 // Compute the constant that would happen if we truncated to SrcTy then
3136 // reextended to DstTy.
3137 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3138 Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy);
3139
3140 // If the re-extended constant didn't change then this is effectively
3141 // also a case of comparing two zero-extended values.
3142 if (RExt == CI && MaxRecurse)
3143 if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred),
3144 SrcOp, Trunc, Q, MaxRecurse-1))
3145 return V;
3146
3147 // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit
3148 // there. Use this to work out the result of the comparison.
3149 if (RExt != CI) {
3150 switch (Pred) {
3151 default: llvm_unreachable("Unknown ICmp predicate!")::llvm::llvm_unreachable_internal("Unknown ICmp predicate!", "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 3151)
;
3152 // LHS <u RHS.
3153 case ICmpInst::ICMP_EQ:
3154 case ICmpInst::ICMP_UGT:
3155 case ICmpInst::ICMP_UGE:
3156 return ConstantInt::getFalse(CI->getContext());
3157
3158 case ICmpInst::ICMP_NE:
3159 case ICmpInst::ICMP_ULT:
3160 case ICmpInst::ICMP_ULE:
3161 return ConstantInt::getTrue(CI->getContext());
3162
3163 // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS
3164 // is non-negative then LHS <s RHS.
3165 case ICmpInst::ICMP_SGT:
3166 case ICmpInst::ICMP_SGE:
3167 return CI->getValue().isNegative() ?
3168 ConstantInt::getTrue(CI->getContext()) :
3169 ConstantInt::getFalse(CI->getContext());
3170
3171 case ICmpInst::ICMP_SLT:
3172 case ICmpInst::ICMP_SLE:
3173 return CI->getValue().isNegative() ?
3174 ConstantInt::getFalse(CI->getContext()) :
3175 ConstantInt::getTrue(CI->getContext());
3176 }
3177 }
3178 }
3179 }
3180
3181 if (isa<SExtInst>(LHS)) {
3182 // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the
3183 // same type.
3184 if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) {
3185 if (MaxRecurse && SrcTy == RI->getOperand(0)->getType())
3186 // Compare X and Y. Note that the predicate does not change.
3187 if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0),
3188 Q, MaxRecurse-1))
3189 return V;
3190 }
3191 // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended
3192 // too. If not, then try to deduce the result of the comparison.
3193 else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
3194 // Compute the constant that would happen if we truncated to SrcTy then
3195 // reextended to DstTy.
3196 Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy);
3197 Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy);
3198
3199 // If the re-extended constant didn't change then this is effectively
3200 // also a case of comparing two sign-extended values.
3201 if (RExt == CI && MaxRecurse)
3202 if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1))
3203 return V;
3204
3205 // Otherwise the upper bits of LHS are all equal, while RHS has varying
3206 // bits there. Use this to work out the result of the comparison.
3207 if (RExt != CI) {
3208 switch (Pred) {
3209 default: llvm_unreachable("Unknown ICmp predicate!")::llvm::llvm_unreachable_internal("Unknown ICmp predicate!", "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 3209)
;
3210 case ICmpInst::ICMP_EQ:
3211 return ConstantInt::getFalse(CI->getContext());
3212 case ICmpInst::ICMP_NE:
3213 return ConstantInt::getTrue(CI->getContext());
3214
3215 // If RHS is non-negative then LHS <s RHS. If RHS is negative then
3216 // LHS >s RHS.
3217 case ICmpInst::ICMP_SGT:
3218 case ICmpInst::ICMP_SGE:
3219 return CI->getValue().isNegative() ?
3220 ConstantInt::getTrue(CI->getContext()) :
3221 ConstantInt::getFalse(CI->getContext());
3222 case ICmpInst::ICMP_SLT:
3223 case ICmpInst::ICMP_SLE:
3224 return CI->getValue().isNegative() ?
3225 ConstantInt::getFalse(CI->getContext()) :
3226 ConstantInt::getTrue(CI->getContext());
3227
3228 // If LHS is non-negative then LHS <u RHS. If LHS is negative then
3229 // LHS >u RHS.
3230 case ICmpInst::ICMP_UGT:
3231 case ICmpInst::ICMP_UGE:
3232 // Comparison is true iff the LHS <s 0.
3233 if (MaxRecurse)
3234 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp,
3235 Constant::getNullValue(SrcTy),
3236 Q, MaxRecurse-1))
3237 return V;
3238 break;
3239 case ICmpInst::ICMP_ULT:
3240 case ICmpInst::ICMP_ULE:
3241 // Comparison is true iff the LHS >=s 0.
3242 if (MaxRecurse)
3243 if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp,
3244 Constant::getNullValue(SrcTy),
3245 Q, MaxRecurse-1))
3246 return V;
3247 break;
3248 }
3249 }
3250 }
3251 }
3252 }
3253
3254 // icmp eq|ne X, Y -> false|true if X != Y
3255 if (ICmpInst::isEquality(Pred) &&
3256 isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) {
3257 return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy);
3258 }
3259
3260 if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse))
3261 return V;
3262
3263 if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse))
3264 return V;
3265
3266 // Simplify comparisons of related pointers using a powerful, recursive
3267 // GEP-walk when we have target data available..
3268 if (LHS->getType()->isPointerTy())
3269 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS,
3270 RHS))
3271 return C;
3272 if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS))
3273 if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS))
3274 if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) ==
3275 Q.DL.getTypeSizeInBits(CLHS->getType()) &&
3276 Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) ==
3277 Q.DL.getTypeSizeInBits(CRHS->getType()))
3278 if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI,
3279 CLHS->getPointerOperand(),
3280 CRHS->getPointerOperand()))
3281 return C;
3282
3283 if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) {
3284 if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) {
3285 if (GLHS->getPointerOperand() == GRHS->getPointerOperand() &&
3286 GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() &&
3287 (ICmpInst::isEquality(Pred) ||
3288 (GLHS->isInBounds() && GRHS->isInBounds() &&
3289 Pred == ICmpInst::getSignedPredicate(Pred)))) {
3290 // The bases are equal and the indices are constant. Build a constant
3291 // expression GEP with the same indices and a null base pointer to see
3292 // what constant folding can make out of it.
3293 Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType());
3294 SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end());
3295 Constant *NewLHS = ConstantExpr::getGetElementPtr(
3296 GLHS->getSourceElementType(), Null, IndicesLHS);
3297
3298 SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end());
3299 Constant *NewRHS = ConstantExpr::getGetElementPtr(
3300 GLHS->getSourceElementType(), Null, IndicesRHS);
3301 return ConstantExpr::getICmp(Pred, NewLHS, NewRHS);
3302 }
3303 }
3304 }
3305
3306 // If the comparison is with the result of a select instruction, check whether
3307 // comparing with either branch of the select always yields the same value.
3308 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3309 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3310 return V;
3311
3312 // If the comparison is with the result of a phi instruction, check whether
3313 // doing the compare with each incoming phi value yields a common result.
3314 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3315 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3316 return V;
3317
3318 return nullptr;
3319}
3320
3321Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3322 const SimplifyQuery &Q) {
3323 return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
3324}
3325
3326/// Given operands for an FCmpInst, see if we can fold the result.
3327/// If not, this returns null.
3328static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3329 FastMathFlags FMF, const SimplifyQuery &Q,
3330 unsigned MaxRecurse) {
3331 CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate;
3332 assert(CmpInst::isFPPredicate(Pred) && "Not an FP compare!")(static_cast <bool> (CmpInst::isFPPredicate(Pred) &&
"Not an FP compare!") ? void (0) : __assert_fail ("CmpInst::isFPPredicate(Pred) && \"Not an FP compare!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 3332, __extension__ __PRETTY_FUNCTION__))
;
3333
3334 if (Constant *CLHS = dyn_cast<Constant>(LHS)) {
3335 if (Constant *CRHS = dyn_cast<Constant>(RHS))
3336 return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI);
3337
3338 // If we have a constant, make sure it is on the RHS.
3339 std::swap(LHS, RHS);
3340 Pred = CmpInst::getSwappedPredicate(Pred);
3341 }
3342
3343 // Fold trivial predicates.
3344 Type *RetTy = GetCompareTy(LHS);
3345 if (Pred == FCmpInst::FCMP_FALSE)
3346 return getFalse(RetTy);
3347 if (Pred == FCmpInst::FCMP_TRUE)
3348 return getTrue(RetTy);
3349
3350 // UNO/ORD predicates can be trivially folded if NaNs are ignored.
3351 if (FMF.noNaNs()) {
3352 if (Pred == FCmpInst::FCMP_UNO)
3353 return getFalse(RetTy);
3354 if (Pred == FCmpInst::FCMP_ORD)
3355 return getTrue(RetTy);
3356 }
3357
3358 // fcmp pred x, undef and fcmp pred undef, x
3359 // fold to true if unordered, false if ordered
3360 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) {
3361 // Choosing NaN for the undef will always make unordered comparison succeed
3362 // and ordered comparison fail.
3363 return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred));
3364 }
3365
3366 // fcmp x,x -> true/false. Not all compares are foldable.
3367 if (LHS == RHS) {
3368 if (CmpInst::isTrueWhenEqual(Pred))
3369 return getTrue(RetTy);
3370 if (CmpInst::isFalseWhenEqual(Pred))
3371 return getFalse(RetTy);
3372 }
3373
3374 // Handle fcmp with constant RHS.
3375 const APFloat *C;
3376 if (match(RHS, m_APFloat(C))) {
3377 // If the constant is a nan, see if we can fold the comparison based on it.
3378 if (C->isNaN()) {
3379 if (FCmpInst::isOrdered(Pred)) // True "if ordered and foo"
3380 return getFalse(RetTy);
3381 assert(FCmpInst::isUnordered(Pred) &&(static_cast <bool> (FCmpInst::isUnordered(Pred) &&
"Comparison must be either ordered or unordered!") ? void (0
) : __assert_fail ("FCmpInst::isUnordered(Pred) && \"Comparison must be either ordered or unordered!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 3382, __extension__ __PRETTY_FUNCTION__))
3382 "Comparison must be either ordered or unordered!")(static_cast <bool> (FCmpInst::isUnordered(Pred) &&
"Comparison must be either ordered or unordered!") ? void (0
) : __assert_fail ("FCmpInst::isUnordered(Pred) && \"Comparison must be either ordered or unordered!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 3382, __extension__ __PRETTY_FUNCTION__))
;
3383 // True if unordered.
3384 return getTrue(RetTy);
3385 }
3386 // Check whether the constant is an infinity.
3387 if (C->isInfinity()) {
3388 if (C->isNegative()) {
3389 switch (Pred) {
3390 case FCmpInst::FCMP_OLT:
3391 // No value is ordered and less than negative infinity.
3392 return getFalse(RetTy);
3393 case FCmpInst::FCMP_UGE:
3394 // All values are unordered with or at least negative infinity.
3395 return getTrue(RetTy);
3396 default:
3397 break;
3398 }
3399 } else {
3400 switch (Pred) {
3401 case FCmpInst::FCMP_OGT:
3402 // No value is ordered and greater than infinity.
3403 return getFalse(RetTy);
3404 case FCmpInst::FCMP_ULE:
3405 // All values are unordered with and at most infinity.
3406 return getTrue(RetTy);
3407 default:
3408 break;
3409 }
3410 }
3411 }
3412 if (C->isZero()) {
3413 switch (Pred) {
3414 case FCmpInst::FCMP_UGE:
3415 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3416 return getTrue(RetTy);
3417 break;
3418 case FCmpInst::FCMP_OLT:
3419 // X < 0
3420 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3421 return getFalse(RetTy);
3422 break;
3423 default:
3424 break;
3425 }
3426 } else if (C->isNegative()) {
3427 assert(!C->isNaN() && "Unexpected NaN constant!")(static_cast <bool> (!C->isNaN() && "Unexpected NaN constant!"
) ? void (0) : __assert_fail ("!C->isNaN() && \"Unexpected NaN constant!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 3427, __extension__ __PRETTY_FUNCTION__))
;
3428 // TODO: We can catch more cases by using a range check rather than
3429 // relying on CannotBeOrderedLessThanZero.
3430 switch (Pred) {
3431 case FCmpInst::FCMP_UGE:
3432 case FCmpInst::FCMP_UGT:
3433 case FCmpInst::FCMP_UNE:
3434 // (X >= 0) implies (X > C) when (C < 0)
3435 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3436 return getTrue(RetTy);
3437 break;
3438 case FCmpInst::FCMP_OEQ:
3439 case FCmpInst::FCMP_OLE:
3440 case FCmpInst::FCMP_OLT:
3441 // (X >= 0) implies !(X < C) when (C < 0)
3442 if (CannotBeOrderedLessThanZero(LHS, Q.TLI))
3443 return getFalse(RetTy);
3444 break;
3445 default:
3446 break;
3447 }
3448 }
3449 }
3450
3451 // If the comparison is with the result of a select instruction, check whether
3452 // comparing with either branch of the select always yields the same value.
3453 if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS))
3454 if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse))
3455 return V;
3456
3457 // If the comparison is with the result of a phi instruction, check whether
3458 // doing the compare with each incoming phi value yields a common result.
3459 if (isa<PHINode>(LHS) || isa<PHINode>(RHS))
3460 if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse))
3461 return V;
3462
3463 return nullptr;
3464}
3465
3466Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
3467 FastMathFlags FMF, const SimplifyQuery &Q) {
3468 return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit);
3469}
3470
3471/// See if V simplifies when its operand Op is replaced with RepOp.
3472static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
3473 const SimplifyQuery &Q,
3474 unsigned MaxRecurse) {
3475 // Trivial replacement.
3476 if (V == Op)
3477 return RepOp;
3478
3479 // We cannot replace a constant, and shouldn't even try.
3480 if (isa<Constant>(Op))
3481 return nullptr;
3482
3483 auto *I = dyn_cast<Instruction>(V);
3484 if (!I)
3485 return nullptr;
3486
3487 // If this is a binary operator, try to simplify it with the replaced op.
3488 if (auto *B = dyn_cast<BinaryOperator>(I)) {
3489 // Consider:
3490 // %cmp = icmp eq i32 %x, 2147483647
3491 // %add = add nsw i32 %x, 1
3492 // %sel = select i1 %cmp, i32 -2147483648, i32 %add
3493 //
3494 // We can't replace %sel with %add unless we strip away the flags.
3495 if (isa<OverflowingBinaryOperator>(B))
3496 if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap())
3497 return nullptr;
3498 if (isa<PossiblyExactOperator>(B))
3499 if (B->isExact())
3500 return nullptr;
3501
3502 if (MaxRecurse) {
3503 if (B->getOperand(0) == Op)
3504 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q,
3505 MaxRecurse - 1);
3506 if (B->getOperand(1) == Op)
3507 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q,
3508 MaxRecurse - 1);
3509 }
3510 }
3511
3512 // Same for CmpInsts.
3513 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
3514 if (MaxRecurse) {
3515 if (C->getOperand(0) == Op)
3516 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q,
3517 MaxRecurse - 1);
3518 if (C->getOperand(1) == Op)
3519 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q,
3520 MaxRecurse - 1);
3521 }
3522 }
3523
3524 // TODO: We could hand off more cases to instsimplify here.
3525
3526 // If all operands are constant after substituting Op for RepOp then we can
3527 // constant fold the instruction.
3528 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
3529 // Build a list of all constant operands.
3530 SmallVector<Constant *, 8> ConstOps;
3531 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
3532 if (I->getOperand(i) == Op)
3533 ConstOps.push_back(CRepOp);
3534 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
3535 ConstOps.push_back(COp);
3536 else
3537 break;
3538 }
3539
3540 // All operands were constants, fold it.
3541 if (ConstOps.size() == I->getNumOperands()) {
3542 if (CmpInst *C = dyn_cast<CmpInst>(I))
3543 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
3544 ConstOps[1], Q.DL, Q.TLI);
3545
3546 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3547 if (!LI->isVolatile())
3548 return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL);
3549
3550 return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI);
3551 }
3552 }
3553
3554 return nullptr;
3555}
3556
3557/// Try to simplify a select instruction when its condition operand is an
3558/// integer comparison where one operand of the compare is a constant.
3559static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X,
3560 const APInt *Y, bool TrueWhenUnset) {
3561 const APInt *C;
3562
3563 // (X & Y) == 0 ? X & ~Y : X --> X
3564 // (X & Y) != 0 ? X & ~Y : X --> X & ~Y
3565 if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) &&
3566 *Y == ~*C)
3567 return TrueWhenUnset ? FalseVal : TrueVal;
3568
3569 // (X & Y) == 0 ? X : X & ~Y --> X & ~Y
3570 // (X & Y) != 0 ? X : X & ~Y --> X
3571 if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) &&
3572 *Y == ~*C)
3573 return TrueWhenUnset ? FalseVal : TrueVal;
3574
3575 if (Y->isPowerOf2()) {
3576 // (X & Y) == 0 ? X | Y : X --> X | Y
3577 // (X & Y) != 0 ? X | Y : X --> X
3578 if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) &&
3579 *Y == *C)
3580 return TrueWhenUnset ? TrueVal : FalseVal;
3581
3582 // (X & Y) == 0 ? X : X | Y --> X
3583 // (X & Y) != 0 ? X : X | Y --> X | Y
3584 if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) &&
3585 *Y == *C)
3586 return TrueWhenUnset ? TrueVal : FalseVal;
3587 }
3588
3589 return nullptr;
3590}
3591
3592/// An alternative way to test if a bit is set or not uses sgt/slt instead of
3593/// eq/ne.
3594static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS,
3595 ICmpInst::Predicate Pred,
3596 Value *TrueVal, Value *FalseVal) {
3597 Value *X;
3598 APInt Mask;
3599 if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask))
3600 return nullptr;
3601
3602 return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask,
3603 Pred == ICmpInst::ICMP_EQ);
3604}
3605
3606/// Try to simplify a select instruction when its condition operand is an
3607/// integer comparison.
3608static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal,
3609 Value *FalseVal, const SimplifyQuery &Q,
3610 unsigned MaxRecurse) {
3611 ICmpInst::Predicate Pred;
3612 Value *CmpLHS, *CmpRHS;
3613 if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS))))
3614 return nullptr;
3615
3616 if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) {
3617 Value *X;
3618 const APInt *Y;
3619 if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y))))
3620 if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y,
3621 Pred == ICmpInst::ICMP_EQ))
3622 return V;
3623 }
3624
3625 // Check for other compares that behave like bit test.
3626 if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred,
3627 TrueVal, FalseVal))
3628 return V;
3629
3630 if (CondVal->hasOneUse()) {
3631 const APInt *C;
3632 if (match(CmpRHS, m_APInt(C))) {
3633 // X < MIN ? T : F --> F
3634 if (Pred == ICmpInst::ICMP_SLT && C->isMinSignedValue())
3635 return FalseVal;
3636 // X < MIN ? T : F --> F
3637 if (Pred == ICmpInst::ICMP_ULT && C->isMinValue())
3638 return FalseVal;
3639 // X > MAX ? T : F --> F
3640 if (Pred == ICmpInst::ICMP_SGT && C->isMaxSignedValue())
3641 return FalseVal;
3642 // X > MAX ? T : F --> F
3643 if (Pred == ICmpInst::ICMP_UGT && C->isMaxValue())
3644 return FalseVal;
3645 }
3646 }
3647
3648 // If we have an equality comparison, then we know the value in one of the
3649 // arms of the select. See if substituting this value into the arm and
3650 // simplifying the result yields the same value as the other arm.
3651 if (Pred == ICmpInst::ICMP_EQ) {
3652 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3653 TrueVal ||
3654 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3655 TrueVal)
3656 return FalseVal;
3657 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3658 FalseVal ||
3659 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3660 FalseVal)
3661 return FalseVal;
3662 } else if (Pred == ICmpInst::ICMP_NE) {
3663 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3664 FalseVal ||
3665 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3666 FalseVal)
3667 return TrueVal;
3668 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) ==
3669 TrueVal ||
3670 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) ==
3671 TrueVal)
3672 return TrueVal;
3673 }
3674
3675 return nullptr;
3676}
3677
3678/// Given operands for a SelectInst, see if we can fold the result.
3679/// If not, this returns null.
3680static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3681 const SimplifyQuery &Q, unsigned MaxRecurse) {
3682 if (auto *CondC = dyn_cast<Constant>(Cond)) {
3683 if (auto *TrueC = dyn_cast<Constant>(TrueVal))
3684 if (auto *FalseC = dyn_cast<Constant>(FalseVal))
3685 return ConstantFoldSelectInstruction(CondC, TrueC, FalseC);
3686
3687 // select undef, X, Y -> X or Y
3688 if (isa<UndefValue>(CondC))
3689 return isa<Constant>(FalseVal) ? FalseVal : TrueVal;
3690
3691 // TODO: Vector constants with undef elements don't simplify.
3692
3693 // select true, X, Y -> X
3694 if (CondC->isAllOnesValue())
3695 return TrueVal;
3696 // select false, X, Y -> Y
3697 if (CondC->isNullValue())
3698 return FalseVal;
3699 }
3700
3701 // select ?, X, X -> X
3702 if (TrueVal == FalseVal)
3703 return TrueVal;
3704
3705 if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X
3706 return FalseVal;
3707 if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X
3708 return TrueVal;
3709
3710 if (Value *V =
3711 simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse))
3712 return V;
3713
3714 return nullptr;
3715}
3716
3717Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal,
3718 const SimplifyQuery &Q) {
3719 return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit);
3720}
3721
3722/// Given operands for an GetElementPtrInst, see if we can fold the result.
3723/// If not, this returns null.
3724static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3725 const SimplifyQuery &Q, unsigned) {
3726 // The type of the GEP pointer operand.
3727 unsigned AS =
3728 cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace();
3729
3730 // getelementptr P -> P.
3731 if (Ops.size() == 1)
3732 return Ops[0];
3733
3734 // Compute the (pointer) type returned by the GEP instruction.
3735 Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1));
3736 Type *GEPTy = PointerType::get(LastType, AS);
3737 if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType()))
3738 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3739 else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType()))
3740 GEPTy = VectorType::get(GEPTy, VT->getNumElements());
3741
3742 if (isa<UndefValue>(Ops[0]))
3743 return UndefValue::get(GEPTy);
3744
3745 if (Ops.size() == 2) {
3746 // getelementptr P, 0 -> P.
3747 if (match(Ops[1], m_Zero()))
3748 return Ops[0];
3749
3750 Type *Ty = SrcTy;
3751 if (Ty->isSized()) {
3752 Value *P;
3753 uint64_t C;
3754 uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty);
3755 // getelementptr P, N -> P if P points to a type of zero size.
3756 if (TyAllocSize == 0)
3757 return Ops[0];
3758
3759 // The following transforms are only safe if the ptrtoint cast
3760 // doesn't truncate the pointers.
3761 if (Ops[1]->getType()->getScalarSizeInBits() ==
3762 Q.DL.getIndexSizeInBits(AS)) {
3763 auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * {
3764 if (match(P, m_Zero()))
3765 return Constant::getNullValue(GEPTy);
3766 Value *Temp;
3767 if (match(P, m_PtrToInt(m_Value(Temp))))
3768 if (Temp->getType() == GEPTy)
3769 return Temp;
3770 return nullptr;
3771 };
3772
3773 // getelementptr V, (sub P, V) -> P if P points to a type of size 1.
3774 if (TyAllocSize == 1 &&
3775 match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0])))))
3776 if (Value *R = PtrToIntOrZero(P))
3777 return R;
3778
3779 // getelementptr V, (ashr (sub P, V), C) -> Q
3780 // if P points to a type of size 1 << C.
3781 if (match(Ops[1],
3782 m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3783 m_ConstantInt(C))) &&
3784 TyAllocSize == 1ULL << C)
3785 if (Value *R = PtrToIntOrZero(P))
3786 return R;
3787
3788 // getelementptr V, (sdiv (sub P, V), C) -> Q
3789 // if P points to a type of size C.
3790 if (match(Ops[1],
3791 m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))),
3792 m_SpecificInt(TyAllocSize))))
3793 if (Value *R = PtrToIntOrZero(P))
3794 return R;
3795 }
3796 }
3797 }
3798
3799 if (Q.DL.getTypeAllocSize(LastType) == 1 &&
3800 all_of(Ops.slice(1).drop_back(1),
3801 [](Value *Idx) { return match(Idx, m_Zero()); })) {
3802 unsigned IdxWidth =
3803 Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace());
3804 if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) {
3805 APInt BasePtrOffset(IdxWidth, 0);
3806 Value *StrippedBasePtr =
3807 Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL,
3808 BasePtrOffset);
3809
3810 // gep (gep V, C), (sub 0, V) -> C
3811 if (match(Ops.back(),
3812 m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) {
3813 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset);
3814 return ConstantExpr::getIntToPtr(CI, GEPTy);
3815 }
3816 // gep (gep V, C), (xor V, -1) -> C-1
3817 if (match(Ops.back(),
3818 m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) {
3819 auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1);
3820 return ConstantExpr::getIntToPtr(CI, GEPTy);
3821 }
3822 }
3823 }
3824
3825 // Check to see if this is constant foldable.
3826 if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); }))
3827 return nullptr;
3828
3829 auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]),
3830 Ops.slice(1));
3831 if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL))
3832 return CEFolded;
3833 return CE;
3834}
3835
3836Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops,
3837 const SimplifyQuery &Q) {
3838 return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit);
3839}
3840
3841/// Given operands for an InsertValueInst, see if we can fold the result.
3842/// If not, this returns null.
3843static Value *SimplifyInsertValueInst(Value *Agg, Value *Val,
3844 ArrayRef<unsigned> Idxs, const SimplifyQuery &Q,
3845 unsigned) {
3846 if (Constant *CAgg = dyn_cast<Constant>(Agg))
3847 if (Constant *CVal = dyn_cast<Constant>(Val))
3848 return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs);
3849
3850 // insertvalue x, undef, n -> x
3851 if (match(Val, m_Undef()))
3852 return Agg;
3853
3854 // insertvalue x, (extractvalue y, n), n
3855 if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val))
3856 if (EV->getAggregateOperand()->getType() == Agg->getType() &&
3857 EV->getIndices() == Idxs) {
3858 // insertvalue undef, (extractvalue y, n), n -> y
3859 if (match(Agg, m_Undef()))
3860 return EV->getAggregateOperand();
3861
3862 // insertvalue y, (extractvalue y, n), n -> y
3863 if (Agg == EV->getAggregateOperand())
3864 return Agg;
3865 }
3866
3867 return nullptr;
3868}
3869
3870Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val,
3871 ArrayRef<unsigned> Idxs,
3872 const SimplifyQuery &Q) {
3873 return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit);
3874}
3875
3876Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx,
3877 const SimplifyQuery &Q) {
3878 // Try to constant fold.
3879 auto *VecC = dyn_cast<Constant>(Vec);
3880 auto *ValC = dyn_cast<Constant>(Val);
3881 auto *IdxC = dyn_cast<Constant>(Idx);
3882 if (VecC && ValC && IdxC)
3883 return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC);
3884
3885 // Fold into undef if index is out of bounds.
3886 if (auto *CI = dyn_cast<ConstantInt>(Idx)) {
3887 uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements();
3888 if (CI->uge(NumElements))
3889 return UndefValue::get(Vec->getType());
3890 }
3891
3892 // If index is undef, it might be out of bounds (see above case)
3893 if (isa<UndefValue>(Idx))
3894 return UndefValue::get(Vec->getType());
3895
3896 return nullptr;
3897}
3898
3899/// Given operands for an ExtractValueInst, see if we can fold the result.
3900/// If not, this returns null.
3901static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3902 const SimplifyQuery &, unsigned) {
3903 if (auto *CAgg = dyn_cast<Constant>(Agg))
3904 return ConstantFoldExtractValueInstruction(CAgg, Idxs);
3905
3906 // extractvalue x, (insertvalue y, elt, n), n -> elt
3907 unsigned NumIdxs = Idxs.size();
3908 for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr;
3909 IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) {
3910 ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices();
3911 unsigned NumInsertValueIdxs = InsertValueIdxs.size();
3912 unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs);
3913 if (InsertValueIdxs.slice(0, NumCommonIdxs) ==
3914 Idxs.slice(0, NumCommonIdxs)) {
3915 if (NumIdxs == NumInsertValueIdxs)
3916 return IVI->getInsertedValueOperand();
3917 break;
3918 }
3919 }
3920
3921 return nullptr;
3922}
3923
3924Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs,
3925 const SimplifyQuery &Q) {
3926 return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit);
3927}
3928
3929/// Given operands for an ExtractElementInst, see if we can fold the result.
3930/// If not, this returns null.
3931static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &,
3932 unsigned) {
3933 if (auto *CVec = dyn_cast<Constant>(Vec)) {
3934 if (auto *CIdx = dyn_cast<Constant>(Idx))
3935 return ConstantFoldExtractElementInstruction(CVec, CIdx);
3936
3937 // The index is not relevant if our vector is a splat.
3938 if (auto *Splat = CVec->getSplatValue())
3939 return Splat;
3940
3941 if (isa<UndefValue>(Vec))
3942 return UndefValue::get(Vec->getType()->getVectorElementType());
3943 }
3944
3945 // If extracting a specified index from the vector, see if we can recursively
3946 // find a previously computed scalar that was inserted into the vector.
3947 if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) {
3948 if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements()))
3949 // definitely out of bounds, thus undefined result
3950 return UndefValue::get(Vec->getType()->getVectorElementType());
3951 if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue()))
3952 return Elt;
3953 }
3954
3955 // An undef extract index can be arbitrarily chosen to be an out-of-range
3956 // index value, which would result in the instruction being undef.
3957 if (isa<UndefValue>(Idx))
3958 return UndefValue::get(Vec->getType()->getVectorElementType());
3959
3960 return nullptr;
3961}
3962
3963Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx,
3964 const SimplifyQuery &Q) {
3965 return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit);
3966}
3967
3968/// See if we can fold the given phi. If not, returns null.
3969static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) {
3970 // If all of the PHI's incoming values are the same then replace the PHI node
3971 // with the common value.
3972 Value *CommonValue = nullptr;
3973 bool HasUndefInput = false;
3974 for (Value *Incoming : PN->incoming_values()) {
3975 // If the incoming value is the phi node itself, it can safely be skipped.
3976 if (Incoming == PN) continue;
3977 if (isa<UndefValue>(Incoming)) {
3978 // Remember that we saw an undef value, but otherwise ignore them.
3979 HasUndefInput = true;
3980 continue;
3981 }
3982 if (CommonValue && Incoming != CommonValue)
3983 return nullptr; // Not the same, bail out.
3984 CommonValue = Incoming;
3985 }
3986
3987 // If CommonValue is null then all of the incoming values were either undef or
3988 // equal to the phi node itself.
3989 if (!CommonValue)
3990 return UndefValue::get(PN->getType());
3991
3992 // If we have a PHI node like phi(X, undef, X), where X is defined by some
3993 // instruction, we cannot return X as the result of the PHI node unless it
3994 // dominates the PHI block.
3995 if (HasUndefInput)
3996 return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
3997
3998 return CommonValue;
3999}
4000
4001static Value *SimplifyCastInst(unsigned CastOpc, Value *Op,
4002 Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) {
4003 if (auto *C = dyn_cast<Constant>(Op))
4004 return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL);
4005
4006 if (auto *CI = dyn_cast<CastInst>(Op)) {
4007 auto *Src = CI->getOperand(0);
4008 Type *SrcTy = Src->getType();
4009 Type *MidTy = CI->getType();
4010 Type *DstTy = Ty;
4011 if (Src->getType() == Ty) {
4012 auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode());
4013 auto SecondOp = static_cast<Instruction::CastOps>(CastOpc);
4014 Type *SrcIntPtrTy =
4015 SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr;
4016 Type *MidIntPtrTy =
4017 MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr;
4018 Type *DstIntPtrTy =
4019 DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr;
4020 if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy,
4021 SrcIntPtrTy, MidIntPtrTy,
4022 DstIntPtrTy) == Instruction::BitCast)
4023 return Src;
4024 }
4025 }
4026
4027 // bitcast x -> x
4028 if (CastOpc == Instruction::BitCast)
4029 if (Op->getType() == Ty)
4030 return Op;
4031
4032 return nullptr;
4033}
4034
4035Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
4036 const SimplifyQuery &Q) {
4037 return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit);
4038}
4039
4040/// For the given destination element of a shuffle, peek through shuffles to
4041/// match a root vector source operand that contains that element in the same
4042/// vector lane (ie, the same mask index), so we can eliminate the shuffle(s).
4043static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1,
4044 int MaskVal, Value *RootVec,
4045 unsigned MaxRecurse) {
4046 if (!MaxRecurse--)
4047 return nullptr;
4048
4049 // Bail out if any mask value is undefined. That kind of shuffle may be
4050 // simplified further based on demanded bits or other folds.
4051 if (MaskVal == -1)
4052 return nullptr;
4053
4054 // The mask value chooses which source operand we need to look at next.
4055 int InVecNumElts = Op0->getType()->getVectorNumElements();
4056 int RootElt = MaskVal;
4057 Value *SourceOp = Op0;
4058 if (MaskVal >= InVecNumElts) {
4059 RootElt = MaskVal - InVecNumElts;
4060 SourceOp = Op1;
4061 }
4062
4063 // If the source operand is a shuffle itself, look through it to find the
4064 // matching root vector.
4065 if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) {
4066 return foldIdentityShuffles(
4067 DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1),
4068 SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse);
4069 }
4070
4071 // TODO: Look through bitcasts? What if the bitcast changes the vector element
4072 // size?
4073
4074 // The source operand is not a shuffle. Initialize the root vector value for
4075 // this shuffle if that has not been done yet.
4076 if (!RootVec)
4077 RootVec = SourceOp;
4078
4079 // Give up as soon as a source operand does not match the existing root value.
4080 if (RootVec != SourceOp)
4081 return nullptr;
4082
4083 // The element must be coming from the same lane in the source vector
4084 // (although it may have crossed lanes in intermediate shuffles).
4085 if (RootElt != DestElt)
4086 return nullptr;
4087
4088 return RootVec;
4089}
4090
4091static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4092 Type *RetTy, const SimplifyQuery &Q,
4093 unsigned MaxRecurse) {
4094 if (isa<UndefValue>(Mask))
4095 return UndefValue::get(RetTy);
4096
4097 Type *InVecTy = Op0->getType();
4098 unsigned MaskNumElts = Mask->getType()->getVectorNumElements();
4099 unsigned InVecNumElts = InVecTy->getVectorNumElements();
4100
4101 SmallVector<int, 32> Indices;
4102 ShuffleVectorInst::getShuffleMask(Mask, Indices);
4103 assert(MaskNumElts == Indices.size() &&(static_cast <bool> (MaskNumElts == Indices.size() &&
"Size of Indices not same as number of mask elements?") ? void
(0) : __assert_fail ("MaskNumElts == Indices.size() && \"Size of Indices not same as number of mask elements?\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 4104, __extension__ __PRETTY_FUNCTION__))
4104 "Size of Indices not same as number of mask elements?")(static_cast <bool> (MaskNumElts == Indices.size() &&
"Size of Indices not same as number of mask elements?") ? void
(0) : __assert_fail ("MaskNumElts == Indices.size() && \"Size of Indices not same as number of mask elements?\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 4104, __extension__ __PRETTY_FUNCTION__))
;
4105
4106 // Canonicalization: If mask does not select elements from an input vector,
4107 // replace that input vector with undef.
4108 bool MaskSelects0 = false, MaskSelects1 = false;
4109 for (unsigned i = 0; i != MaskNumElts; ++i) {
4110 if (Indices[i] == -1)
4111 continue;
4112 if ((unsigned)Indices[i] < InVecNumElts)
4113 MaskSelects0 = true;
4114 else
4115 MaskSelects1 = true;
4116 }
4117 if (!MaskSelects0)
4118 Op0 = UndefValue::get(InVecTy);
4119 if (!MaskSelects1)
4120 Op1 = UndefValue::get(InVecTy);
4121
4122 auto *Op0Const = dyn_cast<Constant>(Op0);
4123 auto *Op1Const = dyn_cast<Constant>(Op1);
4124
4125 // If all operands are constant, constant fold the shuffle.
4126 if (Op0Const && Op1Const)
4127 return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask);
4128
4129 // Canonicalization: if only one input vector is constant, it shall be the
4130 // second one.
4131 if (Op0Const && !Op1Const) {
4132 std::swap(Op0, Op1);
4133 ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts);
4134 }
4135
4136 // A shuffle of a splat is always the splat itself. Legal if the shuffle's
4137 // value type is same as the input vectors' type.
4138 if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0))
4139 if (isa<UndefValue>(Op1) && RetTy == InVecTy &&
4140 OpShuf->getMask()->getSplatValue())
4141 return Op0;
4142
4143 // Don't fold a shuffle with undef mask elements. This may get folded in a
4144 // better way using demanded bits or other analysis.
4145 // TODO: Should we allow this?
4146 if (find(Indices, -1) != Indices.end())
4147 return nullptr;
4148
4149 // Check if every element of this shuffle can be mapped back to the
4150 // corresponding element of a single root vector. If so, we don't need this
4151 // shuffle. This handles simple identity shuffles as well as chains of
4152 // shuffles that may widen/narrow and/or move elements across lanes and back.
4153 Value *RootVec = nullptr;
4154 for (unsigned i = 0; i != MaskNumElts; ++i) {
4155 // Note that recursion is limited for each vector element, so if any element
4156 // exceeds the limit, this will fail to simplify.
4157 RootVec =
4158 foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse);
4159
4160 // We can't replace a widening/narrowing shuffle with one of its operands.
4161 if (!RootVec || RootVec->getType() != RetTy)
4162 return nullptr;
4163 }
4164 return RootVec;
4165}
4166
4167/// Given operands for a ShuffleVectorInst, fold the result or return null.
4168Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask,
4169 Type *RetTy, const SimplifyQuery &Q) {
4170 return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit);
4171}
4172
4173/// Given operands for an FAdd, see if we can fold the result. If not, this
4174/// returns null.
4175static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4176 const SimplifyQuery &Q, unsigned MaxRecurse) {
4177 if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q))
4178 return C;
4179
4180 // fadd X, -0 ==> X
4181 if (match(Op1, m_NegZero()))
4182 return Op0;
4183
4184 // fadd X, 0 ==> X, when we know X is not -0
4185 if (match(Op1, m_Zero()) &&
4186 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4187 return Op0;
4188
4189 // fadd [nnan ninf] X, (fsub [nnan ninf] 0, X) ==> 0
4190 // where nnan and ninf have to occur at least once somewhere in this
4191 // expression
4192 Value *SubOp = nullptr;
4193 if (match(Op1, m_FSub(m_AnyZero(), m_Specific(Op0))))
4194 SubOp = Op1;
4195 else if (match(Op0, m_FSub(m_AnyZero(), m_Specific(Op1))))
4196 SubOp = Op0;
4197 if (SubOp) {
4198 Instruction *FSub = cast<Instruction>(SubOp);
4199 if ((FMF.noNaNs() || FSub->hasNoNaNs()) &&
4200 (FMF.noInfs() || FSub->hasNoInfs()))
4201 return Constant::getNullValue(Op0->getType());
4202 }
4203
4204 return nullptr;
4205}
4206
4207/// Given operands for an FSub, see if we can fold the result. If not, this
4208/// returns null.
4209static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4210 const SimplifyQuery &Q, unsigned MaxRecurse) {
4211 if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
4212 return C;
4213
4214 // fsub X, 0 ==> X
4215 if (match(Op1, m_Zero()))
4216 return Op0;
4217
4218 // fsub X, -0 ==> X, when we know X is not -0
4219 if (match(Op1, m_NegZero()) &&
4220 (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
4221 return Op0;
4222
4223 // fsub -0.0, (fsub -0.0, X) ==> X
4224 Value *X;
4225 if (match(Op0, m_NegZero()) && match(Op1, m_FSub(m_NegZero(), m_Value(X))))
4226 return X;
4227
4228 // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
4229 if (FMF.noSignedZeros() && match(Op0, m_AnyZero()) &&
4230 match(Op1, m_FSub(m_AnyZero(), m_Value(X))))
4231 return X;
4232
4233 // fsub nnan x, x ==> 0.0
4234 if (FMF.noNaNs() && Op0 == Op1)
4235 return Constant::getNullValue(Op0->getType());
4236
4237 return nullptr;
4238}
4239
4240/// Given the operands for an FMul, see if we can fold the result
4241static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4242 const SimplifyQuery &Q, unsigned MaxRecurse) {
4243 if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q))
4244 return C;
4245
4246 // fmul X, 1.0 ==> X
4247 if (match(Op1, m_FPOne()))
4248 return Op0;
4249
4250 // fmul nnan nsz X, 0 ==> 0
4251 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZero()))
4252 return Op1;
4253
4254 // sqrt(X) * sqrt(X) --> X
4255 Value *X;
4256 if (FMF.isFast() && Op0 == Op1 &&
4257 match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))))
4258 return X;
4259
4260 return nullptr;
4261}
4262
4263Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4264 const SimplifyQuery &Q) {
4265 return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit);
4266}
4267
4268
4269Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4270 const SimplifyQuery &Q) {
4271 return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit);
4272}
4273
4274Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4275 const SimplifyQuery &Q) {
4276 return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit);
4277}
4278
4279static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4280 const SimplifyQuery &Q, unsigned) {
4281 if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q))
4282 return C;
4283
4284 // undef / X -> undef (the undef could be a snan).
4285 if (match(Op0, m_Undef()))
4286 return Op0;
4287
4288 // X / undef -> undef
4289 if (match(Op1, m_Undef()))
4290 return Op1;
4291
4292 // X / 1.0 -> X
4293 if (match(Op1, m_FPOne()))
4294 return Op0;
4295
4296 // 0 / X -> 0
4297 // Requires that NaNs are off (X could be zero) and signed zeroes are
4298 // ignored (X could be positive or negative, so the output sign is unknown).
4299 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4300 return Op0;
4301
4302 if (FMF.noNaNs()) {
4303 // X / X -> 1.0 is legal when NaNs are ignored.
4304 // We can ignore infinities because INF/INF is NaN.
4305 if (Op0 == Op1)
4306 return ConstantFP::get(Op0->getType(), 1.0);
4307
4308 // (X * Y) / Y --> X if we can reassociate to the above form.
4309 Value *X;
4310 if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1))))
4311 return X;
4312
4313 // -X / X -> -1.0 and
4314 // X / -X -> -1.0 are legal when NaNs are ignored.
4315 // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored.
4316 if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) &&
4317 BinaryOperator::getFNegArgument(Op0) == Op1) ||
4318 (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) &&
4319 BinaryOperator::getFNegArgument(Op1) == Op0))
4320 return ConstantFP::get(Op0->getType(), -1.0);
4321 }
4322
4323 return nullptr;
4324}
4325
4326Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4327 const SimplifyQuery &Q) {
4328 return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit);
4329}
4330
4331static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4332 const SimplifyQuery &Q, unsigned) {
4333 if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q))
4334 return C;
4335
4336 // undef % X -> undef (the undef could be a snan).
4337 if (match(Op0, m_Undef()))
4338 return Op0;
4339
4340 // X % undef -> undef
4341 if (match(Op1, m_Undef()))
4342 return Op1;
4343
4344 // 0 % X -> 0
4345 // Requires that NaNs are off (X could be zero) and signed zeroes are
4346 // ignored (X could be positive or negative, so the output sign is unknown).
4347 if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZero()))
4348 return Op0;
4349
4350 return nullptr;
4351}
4352
4353Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF,
4354 const SimplifyQuery &Q) {
4355 return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit);
4356}
4357
4358//=== Helper functions for higher up the class hierarchy.
4359
4360/// Given operands for a BinaryOperator, see if we can fold the result.
4361/// If not, this returns null.
4362static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4363 const SimplifyQuery &Q, unsigned MaxRecurse) {
4364 switch (Opcode) {
4365 case Instruction::Add:
4366 return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse);
4367 case Instruction::Sub:
4368 return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse);
4369 case Instruction::Mul:
4370 return SimplifyMulInst(LHS, RHS, Q, MaxRecurse);
4371 case Instruction::SDiv:
4372 return SimplifySDivInst(LHS, RHS, Q, MaxRecurse);
4373 case Instruction::UDiv:
4374 return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse);
4375 case Instruction::SRem:
4376 return SimplifySRemInst(LHS, RHS, Q, MaxRecurse);
4377 case Instruction::URem:
4378 return SimplifyURemInst(LHS, RHS, Q, MaxRecurse);
4379 case Instruction::Shl:
4380 return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse);
4381 case Instruction::LShr:
4382 return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse);
4383 case Instruction::AShr:
4384 return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse);
4385 case Instruction::And:
4386 return SimplifyAndInst(LHS, RHS, Q, MaxRecurse);
4387 case Instruction::Or:
4388 return SimplifyOrInst(LHS, RHS, Q, MaxRecurse);
4389 case Instruction::Xor:
4390 return SimplifyXorInst(LHS, RHS, Q, MaxRecurse);
4391 case Instruction::FAdd:
4392 return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4393 case Instruction::FSub:
4394 return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4395 case Instruction::FMul:
4396 return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4397 case Instruction::FDiv:
4398 return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4399 case Instruction::FRem:
4400 return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4401 default:
4402 llvm_unreachable("Unexpected opcode")::llvm::llvm_unreachable_internal("Unexpected opcode", "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 4402)
;
4403 }
4404}
4405
4406/// Given operands for a BinaryOperator, see if we can fold the result.
4407/// If not, this returns null.
4408/// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the
4409/// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp.
4410static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4411 const FastMathFlags &FMF, const SimplifyQuery &Q,
4412 unsigned MaxRecurse) {
4413 switch (Opcode) {
4414 case Instruction::FAdd:
4415 return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse);
4416 case Instruction::FSub:
4417 return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse);
4418 case Instruction::FMul:
4419 return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse);
4420 case Instruction::FDiv:
4421 return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse);
4422 default:
4423 return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse);
4424 }
4425}
4426
4427Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4428 const SimplifyQuery &Q) {
4429 return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit);
4430}
4431
4432Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS,
4433 FastMathFlags FMF, const SimplifyQuery &Q) {
4434 return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit);
4435}
4436
4437/// Given operands for a CmpInst, see if we can fold the result.
4438static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4439 const SimplifyQuery &Q, unsigned MaxRecurse) {
4440 if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate))
4441 return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse);
4442 return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse);
4443}
4444
4445Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS,
4446 const SimplifyQuery &Q) {
4447 return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit);
4448}
4449
4450static bool IsIdempotent(Intrinsic::ID ID) {
4451 switch (ID) {
4452 default: return false;
4453
4454 // Unary idempotent: f(f(x)) = f(x)
4455 case Intrinsic::fabs:
4456 case Intrinsic::floor:
4457 case Intrinsic::ceil:
4458 case Intrinsic::trunc:
4459 case Intrinsic::rint:
4460 case Intrinsic::nearbyint:
4461 case Intrinsic::round:
4462 case Intrinsic::canonicalize:
4463 return true;
4464 }
4465}
4466
4467static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset,
4468 const DataLayout &DL) {
4469 GlobalValue *PtrSym;
4470 APInt PtrOffset;
4471 if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL))
4472 return nullptr;
4473
4474 Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext());
4475 Type *Int32Ty = Type::getInt32Ty(Ptr->getContext());
4476 Type *Int32PtrTy = Int32Ty->getPointerTo();
4477 Type *Int64Ty = Type::getInt64Ty(Ptr->getContext());
4478
4479 auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset);
4480 if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64)
4481 return nullptr;
4482
4483 uint64_t OffsetInt = OffsetConstInt->getSExtValue();
4484 if (OffsetInt % 4 != 0)
4485 return nullptr;
4486
4487 Constant *C = ConstantExpr::getGetElementPtr(
4488 Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy),
4489 ConstantInt::get(Int64Ty, OffsetInt / 4));
4490 Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL);
4491 if (!Loaded)
4492 return nullptr;
4493
4494 auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded);
4495 if (!LoadedCE)
4496 return nullptr;
4497
4498 if (LoadedCE->getOpcode() == Instruction::Trunc) {
4499 LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4500 if (!LoadedCE)
4501 return nullptr;
4502 }
4503
4504 if (LoadedCE->getOpcode() != Instruction::Sub)
4505 return nullptr;
4506
4507 auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0));
4508 if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt)
4509 return nullptr;
4510 auto *LoadedLHSPtr = LoadedLHS->getOperand(0);
4511
4512 Constant *LoadedRHS = LoadedCE->getOperand(1);
4513 GlobalValue *LoadedRHSSym;
4514 APInt LoadedRHSOffset;
4515 if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset,
4516 DL) ||
4517 PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset)
4518 return nullptr;
4519
4520 return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy);
4521}
4522
4523static bool maskIsAllZeroOrUndef(Value *Mask) {
4524 auto *ConstMask = dyn_cast<Constant>(Mask);
4525 if (!ConstMask)
4526 return false;
4527 if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask))
4528 return true;
4529 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E;
4530 ++I) {
4531 if (auto *MaskElt = ConstMask->getAggregateElement(I))
4532 if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt))
4533 continue;
4534 return false;
4535 }
4536 return true;
4537}
4538
4539template <typename IterTy>
4540static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd,
4541 const SimplifyQuery &Q, unsigned MaxRecurse) {
4542 Intrinsic::ID IID = F->getIntrinsicID();
4543 unsigned NumOperands = std::distance(ArgBegin, ArgEnd);
4544
4545 // Unary Ops
4546 if (NumOperands == 1) {
4547 // Perform idempotent optimizations
4548 if (IsIdempotent(IID)) {
4549 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) {
4550 if (II->getIntrinsicID() == IID)
4551 return II;
4552 }
4553 }
4554
4555 Value *IIOperand = *ArgBegin;
4556 Value *X;
4557 switch (IID) {
4558 case Intrinsic::fabs: {
4559 if (SignBitMustBeZero(IIOperand, Q.TLI))
4560 return IIOperand;
4561 return nullptr;
4562 }
4563 case Intrinsic::bswap: {
4564 // bswap(bswap(x)) -> x
4565 if (match(IIOperand, m_BSwap(m_Value(X))))
4566 return X;
4567 return nullptr;
4568 }
4569 case Intrinsic::bitreverse: {
4570 // bitreverse(bitreverse(x)) -> x
4571 if (match(IIOperand, m_BitReverse(m_Value(X))))
4572 return X;
4573 return nullptr;
4574 }
4575 case Intrinsic::exp: {
4576 // exp(log(x)) -> x
4577 if (Q.CxtI->hasAllowReassoc() &&
4578 match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X))))
4579 return X;
4580 return nullptr;
4581 }
4582 case Intrinsic::exp2: {
4583 // exp2(log2(x)) -> x
4584 if (Q.CxtI->hasAllowReassoc() &&
4585 match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X))))
4586 return X;
4587 return nullptr;
4588 }
4589 case Intrinsic::log: {
4590 // log(exp(x)) -> x
4591 if (Q.CxtI->hasAllowReassoc() &&
4592 match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X))))
4593 return X;
4594 return nullptr;
4595 }
4596 case Intrinsic::log2: {
4597 // log2(exp2(x)) -> x
4598 if (Q.CxtI->hasAllowReassoc() &&
4599 match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) {
4600 return X;
4601 }
4602 return nullptr;
4603 }
4604 default:
4605 return nullptr;
4606 }
4607 }
4608
4609 // Binary Ops
4610 if (NumOperands == 2) {
4611 Value *LHS = *ArgBegin;
4612 Value *RHS = *(ArgBegin + 1);
4613 Type *ReturnType = F->getReturnType();
4614
4615 switch (IID) {
4616 case Intrinsic::usub_with_overflow:
4617 case Intrinsic::ssub_with_overflow: {
4618 // X - X -> { 0, false }
4619 if (LHS == RHS)
4620 return Constant::getNullValue(ReturnType);
4621
4622 // X - undef -> undef
4623 // undef - X -> undef
4624 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4625 return UndefValue::get(ReturnType);
4626
4627 return nullptr;
4628 }
4629 case Intrinsic::uadd_with_overflow:
4630 case Intrinsic::sadd_with_overflow: {
4631 // X + undef -> undef
4632 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
4633 return UndefValue::get(ReturnType);
4634
4635 return nullptr;
4636 }
4637 case Intrinsic::umul_with_overflow:
4638 case Intrinsic::smul_with_overflow: {
4639 // 0 * X -> { 0, false }
4640 // X * 0 -> { 0, false }
4641 if (match(LHS, m_Zero()) || match(RHS, m_Zero()))
4642 return Constant::getNullValue(ReturnType);
4643
4644 // undef * X -> { 0, false }
4645 // X * undef -> { 0, false }
4646 if (match(LHS, m_Undef()) || match(RHS, m_Undef()))
4647 return Constant::getNullValue(ReturnType);
4648
4649 return nullptr;
4650 }
4651 case Intrinsic::load_relative: {
4652 Constant *C0 = dyn_cast<Constant>(LHS);
4653 Constant *C1 = dyn_cast<Constant>(RHS);
4654 if (C0 && C1)
4655 return SimplifyRelativeLoad(C0, C1, Q.DL);
4656 return nullptr;
4657 }
4658 case Intrinsic::powi:
4659 if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) {
4660 // powi(x, 0) -> 1.0
4661 if (Power->isZero())
4662 return ConstantFP::get(LHS->getType(), 1.0);
4663 // powi(x, 1) -> x
4664 if (Power->isOne())
4665 return LHS;
4666 }
4667 return nullptr;
4668 default:
4669 return nullptr;
4670 }
4671 }
4672
4673 // Simplify calls to llvm.masked.load.*
4674 switch (IID) {
4675 case Intrinsic::masked_load: {
4676 Value *MaskArg = ArgBegin[2];
4677 Value *PassthruArg = ArgBegin[3];
4678 // If the mask is all zeros or undef, the "passthru" argument is the result.
4679 if (maskIsAllZeroOrUndef(MaskArg))
4680 return PassthruArg;
4681 return nullptr;
4682 }
4683 default:
4684 return nullptr;
4685 }
4686}
4687
4688template <typename IterTy>
4689static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin,
4690 IterTy ArgEnd, const SimplifyQuery &Q,
4691 unsigned MaxRecurse) {
4692 Type *Ty = V->getType();
4693 if (PointerType *PTy = dyn_cast<PointerType>(Ty))
4694 Ty = PTy->getElementType();
4695 FunctionType *FTy = cast<FunctionType>(Ty);
4696
4697 // call undef -> undef
4698 // call null -> undef
4699 if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V))
4700 return UndefValue::get(FTy->getReturnType());
4701
4702 Function *F = dyn_cast<Function>(V);
4703 if (!F)
4704 return nullptr;
4705
4706 if (F->isIntrinsic())
4707 if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse))
4708 return Ret;
4709
4710 if (!canConstantFoldCallTo(CS, F))
4711 return nullptr;
4712
4713 SmallVector<Constant *, 4> ConstantArgs;
4714 ConstantArgs.reserve(ArgEnd - ArgBegin);
4715 for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) {
4716 Constant *C = dyn_cast<Constant>(*I);
4717 if (!C)
4718 return nullptr;
4719 ConstantArgs.push_back(C);
4720 }
4721
4722 return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI);
4723}
4724
4725Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4726 User::op_iterator ArgBegin, User::op_iterator ArgEnd,
4727 const SimplifyQuery &Q) {
4728 return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit);
4729}
4730
4731Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V,
4732 ArrayRef<Value *> Args, const SimplifyQuery &Q) {
4733 return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit);
4734}
4735
4736Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) {
4737 CallSite CS(const_cast<Instruction*>(ICS.getInstruction()));
4738 return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(),
4739 Q, RecursionLimit);
4740}
4741
4742/// See if we can compute a simplified version of this instruction.
4743/// If not, this returns null.
4744
4745Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ,
4746 OptimizationRemarkEmitter *ORE) {
4747 const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I);
4748 Value *Result;
4749
4750 switch (I->getOpcode()) {
4751 default:
4752 Result = ConstantFoldInstruction(I, Q.DL, Q.TLI);
4753 break;
4754 case Instruction::FAdd:
4755 Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1),
4756 I->getFastMathFlags(), Q);
4757 break;
4758 case Instruction::Add:
4759 Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1),
4760 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4761 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4762 break;
4763 case Instruction::FSub:
4764 Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1),
4765 I->getFastMathFlags(), Q);
4766 break;
4767 case Instruction::Sub:
4768 Result = SimplifySubInst(I->getOperand(0), I->getOperand(1),
4769 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4770 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4771 break;
4772 case Instruction::FMul:
4773 Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1),
4774 I->getFastMathFlags(), Q);
4775 break;
4776 case Instruction::Mul:
4777 Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q);
4778 break;
4779 case Instruction::SDiv:
4780 Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q);
4781 break;
4782 case Instruction::UDiv:
4783 Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q);
4784 break;
4785 case Instruction::FDiv:
4786 Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1),
4787 I->getFastMathFlags(), Q);
4788 break;
4789 case Instruction::SRem:
4790 Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q);
4791 break;
4792 case Instruction::URem:
4793 Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q);
4794 break;
4795 case Instruction::FRem:
4796 Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1),
4797 I->getFastMathFlags(), Q);
4798 break;
4799 case Instruction::Shl:
4800 Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1),
4801 cast<BinaryOperator>(I)->hasNoSignedWrap(),
4802 cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q);
4803 break;
4804 case Instruction::LShr:
4805 Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1),
4806 cast<BinaryOperator>(I)->isExact(), Q);
4807 break;
4808 case Instruction::AShr:
4809 Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1),
4810 cast<BinaryOperator>(I)->isExact(), Q);
4811 break;
4812 case Instruction::And:
4813 Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q);
4814 break;
4815 case Instruction::Or:
4816 Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q);
4817 break;
4818 case Instruction::Xor:
4819 Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q);
4820 break;
4821 case Instruction::ICmp:
4822 Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(),
4823 I->getOperand(0), I->getOperand(1), Q);
4824 break;
4825 case Instruction::FCmp:
4826 Result =
4827 SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0),
4828 I->getOperand(1), I->getFastMathFlags(), Q);
4829 break;
4830 case Instruction::Select:
4831 Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1),
4832 I->getOperand(2), Q);
4833 break;
4834 case Instruction::GetElementPtr: {
4835 SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end());
4836 Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(),
4837 Ops, Q);
4838 break;
4839 }
4840 case Instruction::InsertValue: {
4841 InsertValueInst *IV = cast<InsertValueInst>(I);
4842 Result = SimplifyInsertValueInst(IV->getAggregateOperand(),
4843 IV->getInsertedValueOperand(),
4844 IV->getIndices(), Q);
4845 break;
4846 }
4847 case Instruction::InsertElement: {
4848 auto *IE = cast<InsertElementInst>(I);
4849 Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1),
4850 IE->getOperand(2), Q);
4851 break;
4852 }
4853 case Instruction::ExtractValue: {
4854 auto *EVI = cast<ExtractValueInst>(I);
4855 Result = SimplifyExtractValueInst(EVI->getAggregateOperand(),
4856 EVI->getIndices(), Q);
4857 break;
4858 }
4859 case Instruction::ExtractElement: {
4860 auto *EEI = cast<ExtractElementInst>(I);
4861 Result = SimplifyExtractElementInst(EEI->getVectorOperand(),
4862 EEI->getIndexOperand(), Q);
4863 break;
4864 }
4865 case Instruction::ShuffleVector: {
4866 auto *SVI = cast<ShuffleVectorInst>(I);
4867 Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
4868 SVI->getMask(), SVI->getType(), Q);
4869 break;
4870 }
4871 case Instruction::PHI:
4872 Result = SimplifyPHINode(cast<PHINode>(I), Q);
4873 break;
4874 case Instruction::Call: {
4875 CallSite CS(cast<CallInst>(I));
4876 Result = SimplifyCall(CS, Q);
4877 break;
4878 }
4879#define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc:
4880#include "llvm/IR/Instruction.def"
4881#undef HANDLE_CAST_INST
4882 Result =
4883 SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q);
4884 break;
4885 case Instruction::Alloca:
4886 // No simplifications for Alloca and it can't be constant folded.
4887 Result = nullptr;
4888 break;
4889 }
4890
4891 // In general, it is possible for computeKnownBits to determine all bits in a
4892 // value even when the operands are not all constants.
4893 if (!Result && I->getType()->isIntOrIntVectorTy()) {
4894 KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE);
4895 if (Known.isConstant())
4896 Result = ConstantInt::get(I->getType(), Known.getConstant());
4897 }
4898
4899 /// If called on unreachable code, the above logic may report that the
4900 /// instruction simplified to itself. Make life easier for users by
4901 /// detecting that case here, returning a safe value instead.
4902 return Result == I ? UndefValue::get(I->getType()) : Result;
4903}
4904
4905/// \brief Implementation of recursive simplification through an instruction's
4906/// uses.
4907///
4908/// This is the common implementation of the recursive simplification routines.
4909/// If we have a pre-simplified value in 'SimpleV', that is forcibly used to
4910/// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of
4911/// instructions to process and attempt to simplify it using
4912/// InstructionSimplify.
4913///
4914/// This routine returns 'true' only when *it* simplifies something. The passed
4915/// in simplified value does not count toward this.
4916static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV,
4917 const TargetLibraryInfo *TLI,
4918 const DominatorTree *DT,
4919 AssumptionCache *AC) {
4920 bool Simplified = false;
4921 SmallSetVector<Instruction *, 8> Worklist;
4922 const DataLayout &DL = I->getModule()->getDataLayout();
4923
4924 // If we have an explicit value to collapse to, do that round of the
4925 // simplification loop by hand initially.
4926 if (SimpleV) {
4927 for (User *U : I->users())
4928 if (U != I)
4929 Worklist.insert(cast<Instruction>(U));
4930
4931 // Replace the instruction with its simplified value.
4932 I->replaceAllUsesWith(SimpleV);
4933
4934 // Gracefully handle edge cases where the instruction is not wired into any
4935 // parent block.
4936 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4937 !I->mayHaveSideEffects())
4938 I->eraseFromParent();
4939 } else {
4940 Worklist.insert(I);
4941 }
4942
4943 // Note that we must test the size on each iteration, the worklist can grow.
4944 for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) {
4945 I = Worklist[Idx];
4946
4947 // See if this instruction simplifies.
4948 SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC});
4949 if (!SimpleV)
4950 continue;
4951
4952 Simplified = true;
4953
4954 // Stash away all the uses of the old instruction so we can check them for
4955 // recursive simplifications after a RAUW. This is cheaper than checking all
4956 // uses of To on the recursive step in most cases.
4957 for (User *U : I->users())
4958 Worklist.insert(cast<Instruction>(U));
4959
4960 // Replace the instruction with its simplified value.
4961 I->replaceAllUsesWith(SimpleV);
4962
4963 // Gracefully handle edge cases where the instruction is not wired into any
4964 // parent block.
4965 if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) &&
4966 !I->mayHaveSideEffects())
4967 I->eraseFromParent();
4968 }
4969 return Simplified;
4970}
4971
4972bool llvm::recursivelySimplifyInstruction(Instruction *I,
4973 const TargetLibraryInfo *TLI,
4974 const DominatorTree *DT,
4975 AssumptionCache *AC) {
4976 return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC);
4977}
4978
4979bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV,
4980 const TargetLibraryInfo *TLI,
4981 const DominatorTree *DT,
4982 AssumptionCache *AC) {
4983 assert(I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!")(static_cast <bool> (I != SimpleV && "replaceAndRecursivelySimplify(X,X) is not valid!"
) ? void (0) : __assert_fail ("I != SimpleV && \"replaceAndRecursivelySimplify(X,X) is not valid!\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 4983, __extension__ __PRETTY_FUNCTION__))
;
4984 assert(SimpleV && "Must provide a simplified value.")(static_cast <bool> (SimpleV && "Must provide a simplified value."
) ? void (0) : __assert_fail ("SimpleV && \"Must provide a simplified value.\""
, "/build/llvm-toolchain-snapshot-7~svn326551/lib/Analysis/InstructionSimplify.cpp"
, 4984, __extension__ __PRETTY_FUNCTION__))
;
4985 return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC);
4986}
4987
4988namespace llvm {
4989const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) {
4990 auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4991 auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
4992 auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
4993 auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr;
4994 auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>();
4995 auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr;
4996 return {F.getParent()->getDataLayout(), TLI, DT, AC};
4997}
4998
4999const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR,
5000 const DataLayout &DL) {
5001 return {DL, &AR.TLI, &AR.DT, &AR.AC};
5002}
5003
5004template <class T, class... TArgs>
5005const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM,
5006 Function &F) {
5007 auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F);
5008 auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F);
5009 auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F);
5010 return {F.getParent()->getDataLayout(), TLI, DT, AC};
5011}
5012template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &,
5013 Function &);
5014}

/build/llvm-toolchain-snapshot-7~svn326551/include/llvm/IR/PatternMatch.h

1//===- PatternMatch.h - Match on the LLVM IR --------------------*- C++ -*-===//
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 provides a simple and efficient mechanism for performing general
11// tree-based pattern matches on the LLVM IR. The power of these routines is
12// that it allows you to write concise patterns that are expressive and easy to
13// understand. The other major advantage of this is that it allows you to
14// trivially capture/bind elements in the pattern to variables. For example,
15// you can do something like this:
16//
17// Value *Exp = ...
18// Value *X, *Y; ConstantInt *C1, *C2; // (X & C1) | (Y & C2)
19// if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)),
20// m_And(m_Value(Y), m_ConstantInt(C2))))) {
21// ... Pattern is matched and variables are bound ...
22// }
23//
24// This is primarily useful to things like the instruction combiner, but can
25// also be useful for static analysis tools or code generators.
26//
27//===----------------------------------------------------------------------===//
28
29#ifndef LLVM_IR_PATTERNMATCH_H
30#define LLVM_IR_PATTERNMATCH_H
31
32#include "llvm/ADT/APFloat.h"
33#include "llvm/ADT/APInt.h"
34#include "llvm/IR/CallSite.h"
35#include "llvm/IR/Constant.h"
36#include "llvm/IR/Constants.h"
37#include "llvm/IR/InstrTypes.h"
38#include "llvm/IR/Instruction.h"
39#include "llvm/IR/Instructions.h"
40#include "llvm/IR/Intrinsics.h"
41#include "llvm/IR/Operator.h"
42#include "llvm/IR/Value.h"
43#include "llvm/Support/Casting.h"
44#include <cstdint>
45
46namespace llvm {
47namespace PatternMatch {
48
49template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) {
50 return const_cast<Pattern &>(P).match(V);
51}
52
53template <typename SubPattern_t> struct OneUse_match {
54 SubPattern_t SubPattern;
55
56 OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {}
57
58 template <typename OpTy> bool match(OpTy *V) {
59 return V->hasOneUse() && SubPattern.match(V);
60 }
61};
62
63template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) {
64 return SubPattern;
65}
66
67template <typename Class> struct class_match {
68 template <typename ITy> bool match(ITy *V) { return isa<Class>(V); }
69};
70
71/// Match an arbitrary value and ignore it.
72inline class_match<Value> m_Value() { return class_match<Value>(); }
73
74/// Match an arbitrary binary operation and ignore it.
75inline class_match<BinaryOperator> m_BinOp() {
76 return class_match<BinaryOperator>();
77}
78
79/// Matches any compare instruction and ignore it.
80inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); }
81
82/// Match an arbitrary ConstantInt and ignore it.
83inline class_match<ConstantInt> m_ConstantInt() {
84 return class_match<ConstantInt>();
85}
86
87/// Match an arbitrary undef constant.
88inline class_match<UndefValue> m_Undef() { return class_match<UndefValue>(); }
89
90/// Match an arbitrary Constant and ignore it.
91inline class_match<Constant> m_Constant() { return class_match<Constant>(); }
92
93/// Matching combinators
94template <typename LTy, typename RTy> struct match_combine_or {
95 LTy L;
96 RTy R;
97
98 match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
99
100 template <typename ITy> bool match(ITy *V) {
101 if (L.match(V))
102 return true;
103 if (R.match(V))
104 return true;
105 return false;
106 }
107};
108
109template <typename LTy, typename RTy> struct match_combine_and {
110 LTy L;
111 RTy R;
112
113 match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
114
115 template <typename ITy> bool match(ITy *V) {
116 if (L.match(V))
117 if (R.match(V))
118 return true;
119 return false;
120 }
121};
122
123/// Combine two pattern matchers matching L || R
124template <typename LTy, typename RTy>
125inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) {
126 return match_combine_or<LTy, RTy>(L, R);
127}
128
129/// Combine two pattern matchers matching L && R
130template <typename LTy, typename RTy>
131inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) {
132 return match_combine_and<LTy, RTy>(L, R);
133}
134
135struct match_zero {
136 template <typename ITy> bool match(ITy *V) {
137 if (const auto *C = dyn_cast<Constant>(V))
138 return C->isNullValue();
139 return false;
140 }
141};
142
143/// Match an arbitrary zero/null constant. This includes
144/// zero_initializer for vectors and ConstantPointerNull for pointers.
145inline match_zero m_Zero() { return match_zero(); }
146
147struct match_neg_zero {
148 template <typename ITy> bool match(ITy *V) {
149 if (const auto *C = dyn_cast<Constant>(V))
150 return C->isNegativeZeroValue();
151 return false;
152 }
153};
154
155/// Match an arbitrary zero/null constant. This includes
156/// zero_initializer for vectors and ConstantPointerNull for pointers. For
157/// floating point constants, this will match negative zero but not positive
158/// zero
159inline match_neg_zero m_NegZero() { return match_neg_zero(); }
160
161struct match_any_zero {
162 template <typename ITy> bool match(ITy *V) {
163 if (const auto *C = dyn_cast<Constant>(V))
164 return C->isZeroValue();
165 return false;
166 }
167};
168
169/// Match an arbitrary zero/null constant. This includes
170/// zero_initializer for vectors and ConstantPointerNull for pointers. For
171/// floating point constants, this will match negative zero and positive zero
172inline match_any_zero m_AnyZero() { return match_any_zero(); }
173
174struct match_nan {
175 template <typename ITy> bool match(ITy *V) {
176 if (const auto *C = dyn_cast<ConstantFP>(V))
177 return C->isNaN();
178 return false;
179 }
180};
181
182/// Match an arbitrary NaN constant. This includes quiet and signalling nans.
183inline match_nan m_NaN() { return match_nan(); }
184
185struct apint_match {
186 const APInt *&Res;
187
188 apint_match(const APInt *&R) : Res(R) {}
189
190 template <typename ITy> bool match(ITy *V) {
191 if (auto *CI = dyn_cast<ConstantInt>(V)) {
192 Res = &CI->getValue();
193 return true;
194 }
195 if (V->getType()->isVectorTy())
196 if (const auto *C = dyn_cast<Constant>(V))
197 if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue())) {
198 Res = &CI->getValue();
199 return true;
200 }
201 return false;
202 }
203};
204// Either constexpr if or renaming ConstantFP::getValueAPF to
205// ConstantFP::getValue is needed to do it via single template
206// function for both apint/apfloat.
207struct apfloat_match {
208 const APFloat *&Res;
209 apfloat_match(const APFloat *&R) : Res(R) {}
210 template <typename ITy> bool match(ITy *V) {
211 if (auto *CI = dyn_cast<ConstantFP>(V)) {
212 Res = &CI->getValueAPF();
213 return true;
214 }
215 if (V->getType()->isVectorTy())
216 if (const auto *C = dyn_cast<Constant>(V))
217 if (auto *CI = dyn_cast_or_null<ConstantFP>(C->getSplatValue())) {
218 Res = &CI->getValueAPF();
219 return true;
220 }
221 return false;
222 }
223};
224
225/// Match a ConstantInt or splatted ConstantVector, binding the
226/// specified pointer to the contained APInt.
227inline apint_match m_APInt(const APInt *&Res) { return Res; }
228
229/// Match a ConstantFP or splatted ConstantVector, binding the
230/// specified pointer to the contained APFloat.
231inline apfloat_match m_APFloat(const APFloat *&Res) { return Res; }
232
233template <int64_t Val> struct constantint_match {
234 template <typename ITy> bool match(ITy *V) {
235 if (const auto *CI = dyn_cast<ConstantInt>(V)) {
236 const APInt &CIV = CI->getValue();
237 if (Val >= 0)
238 return CIV == static_cast<uint64_t>(Val);
239 // If Val is negative, and CI is shorter than it, truncate to the right
240 // number of bits. If it is larger, then we have to sign extend. Just
241 // compare their negated values.
242 return -CIV == -Val;
243 }
244 return false;
245 }
246};
247
248/// Match a ConstantInt with a specific value.
249template <int64_t Val> inline constantint_match<Val> m_ConstantInt() {
250 return constantint_match<Val>();
251}
252
253/// This helper class is used to match scalar and vector constants that satisfy
254/// a specified predicate. For vector constants, undefined elements are ignored.
255template <typename Predicate> struct cst_pred_ty : public Predicate {
256 template <typename ITy> bool match(ITy *V) {
257 if (const auto *CI = dyn_cast<ConstantInt>(V))
258 return this->isValue(CI->getValue());
259 if (V->getType()->isVectorTy()) {
260 if (const auto *C = dyn_cast<Constant>(V)) {
261 if (const auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
262 return this->isValue(CI->getValue());
263
264 // Non-splat vector constant: check each element for a match.
265 unsigned NumElts = V->getType()->getVectorNumElements();
266 assert(NumElts != 0 && "Constant vector with no elements?")(static_cast <bool> (NumElts != 0 && "Constant vector with no elements?"
) ? void (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\""
, "/build/llvm-toolchain-snapshot-7~svn326551/include/llvm/IR/PatternMatch.h"
, 266, __extension__ __PRETTY_FUNCTION__))
;
267 for (unsigned i = 0; i != NumElts; ++i) {
268 Constant *Elt = C->getAggregateElement(i);
269 if (!Elt)
270 return false;
271 if (isa<UndefValue>(Elt))
272 continue;
273 auto *CI = dyn_cast<ConstantInt>(Elt);
274 if (!CI || !this->isValue(CI->getValue()))
275 return false;
276 }
277 return true;
278 }
279 }
280 return false;
281 }
282};
283
284/// This helper class is used to match scalar and vector constants that
285/// satisfy a specified predicate, and bind them to an APInt.
286template <typename Predicate> struct api_pred_ty : public Predicate {
287 const APInt *&Res;
288
289 api_pred_ty(const APInt *&R) : Res(R) {}
290
291 template <typename ITy> bool match(ITy *V) {
292 if (const auto *CI = dyn_cast<ConstantInt>(V))
293 if (this->isValue(CI->getValue())) {
294 Res = &CI->getValue();
295 return true;
296 }
297 if (V->getType()->isVectorTy())
298 if (const auto *C = dyn_cast<Constant>(V))
299 if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
300 if (this->isValue(CI->getValue())) {
301 Res = &CI->getValue();
302 return true;
303 }
304
305 return false;
306 }
307};
308
309///////////////////////////////////////////////////////////////////////////////
310//
311// Encapsulate constant value queries for use in templated predicate matchers.
312// This allows checking if constants match using compound predicates and works
313// with vector constants, possibly with relaxed constraints. For example, ignore
314// undef values.
315//
316///////////////////////////////////////////////////////////////////////////////
317
318struct is_all_ones {
319 bool isValue(const APInt &C) { return C.isAllOnesValue(); }
320};
321/// Match an integer or vector with all bits set.
322inline cst_pred_ty<is_all_ones> m_AllOnes() {
323 return cst_pred_ty<is_all_ones>();
324}
325
326struct is_maxsignedvalue {
327 bool isValue(const APInt &C) { return C.isMaxSignedValue(); }
328};
329/// Match an integer or vector with values having all bits except for the high
330/// bit set (0x7f...).
331inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() {
332 return cst_pred_ty<is_maxsignedvalue>();
333}
334inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) {
335 return V;
336}
337
338struct is_negative {
339 bool isValue(const APInt &C) { return C.isNegative(); }
340};
341/// Match an integer or vector of negative values.
342inline cst_pred_ty<is_negative> m_Negative() {
343 return cst_pred_ty<is_negative>();
344}
345inline api_pred_ty<is_negative> m_Negative(const APInt *&V) {
346 return V;
347}
348
349struct is_nonnegative {
350 bool isValue(const APInt &C) { return C.isNonNegative(); }
351};
352/// Match an integer or vector of nonnegative values.
353inline cst_pred_ty<is_nonnegative> m_NonNegative() {
354 return cst_pred_ty<is_nonnegative>();
355}
356inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) {
357 return V;
358}
359
360struct is_one {
361 bool isValue(const APInt &C) { return C.isOneValue(); }
362};
363/// Match an integer 1 or a vector with all elements equal to 1.
364inline cst_pred_ty<is_one> m_One() {
365 return cst_pred_ty<is_one>();
366}
367
368struct is_power2 {
369 bool isValue(const APInt &C) { return C.isPowerOf2(); }
370};
371/// Match an integer or vector power-of-2.
372inline cst_pred_ty<is_power2> m_Power2() {
373 return cst_pred_ty<is_power2>();
374}
375inline api_pred_ty<is_power2> m_Power2(const APInt *&V) {
376 return V;
377}
378
379struct is_power2_or_zero {
380 bool isValue(const APInt &C) { return !C || C.isPowerOf2(); }
381};
382/// Match an integer or vector of 0 or power-of-2 values.
383inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() {
384 return cst_pred_ty<is_power2_or_zero>();
385}
386inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) {
387 return V;
388}
389
390struct is_sign_mask {
391 bool isValue(const APInt &C) { return C.isSignMask(); }
392};
393/// Match an integer or vector with only the sign bit(s) set.
394inline cst_pred_ty<is_sign_mask> m_SignMask() {
395 return cst_pred_ty<is_sign_mask>();
396}
397
398///////////////////////////////////////////////////////////////////////////////
399
400template <typename Class> struct bind_ty {
401 Class *&VR;
402
403 bind_ty(Class *&V) : VR(V) {}
6
Returning without writing to 'V'
404
405 template <typename ITy> bool match(ITy *V) {
406 if (auto *CV = dyn_cast<Class>(V)) {
407 VR = CV;
408 return true;
409 }
410 return false;
411 }
412};
413
414/// Match a value, capturing it if we match.
415inline bind_ty<Value> m_Value(Value *&V) { return V; }
5
Calling constructor for 'bind_ty'
7
Returning from constructor for 'bind_ty'
8
Returning without writing to 'V'
416inline bind_ty<const Value> m_Value(const Value *&V) { return V; }
417
418/// Match an instruction, capturing it if we match.
419inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; }
420/// Match a binary operator, capturing it if we match.
421inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; }
422
423/// Match a ConstantInt, capturing the value if we match.
424inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; }
425
426/// Match a Constant, capturing the value if we match.
427inline bind_ty<Constant> m_Constant(Constant *&C) { return C; }
428
429/// Match a ConstantFP, capturing the value if we match.
430inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; }
431
432/// Match a specified Value*.
433struct specificval_ty {
434 const Value *Val;
435
436 specificval_ty(const Value *V) : Val(V) {}
437
438 template <typename ITy> bool match(ITy *V) { return V == Val; }
439};
440
441/// Match if we have a specific specified value.
442inline specificval_ty m_Specific(const Value *V) { return V; }
443
444/// Match a specified floating point value or vector of all elements of
445/// that value.
446struct specific_fpval {
447 double Val;
448
449 specific_fpval(double V) : Val(V) {}
450
451 template <typename ITy> bool match(ITy *V) {
452 if (const auto *CFP = dyn_cast<ConstantFP>(V))
453 return CFP->isExactlyValue(Val);
454 if (V->getType()->isVectorTy())
455 if (const auto *C = dyn_cast<Constant>(V))
456 if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue()))
457 return CFP->isExactlyValue(Val);
458 return false;
459 }
460};
461
462/// Match a specific floating point value or vector with all elements
463/// equal to the value.
464inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); }
465
466/// Match a float 1.0 or vector with all elements equal to 1.0.
467inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); }
468
469struct bind_const_intval_ty {
470 uint64_t &VR;
471
472 bind_const_intval_ty(uint64_t &V) : VR(V) {}
473
474 template <typename ITy> bool match(ITy *V) {
475 if (const auto *CV = dyn_cast<ConstantInt>(V))
476 if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) {
477 VR = CV->getZExtValue();
478 return true;
479 }
480 return false;
481 }
482};
483
484/// Match a specified integer value or vector of all elements of that
485// value.
486struct specific_intval {
487 uint64_t Val;
488
489 specific_intval(uint64_t V) : Val(V) {}
490
491 template <typename ITy> bool match(ITy *V) {
492 const auto *CI = dyn_cast<ConstantInt>(V);
493 if (!CI && V->getType()->isVectorTy())
494 if (const auto *C = dyn_cast<Constant>(V))
495 CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue());
496
497 return CI && CI->getValue() == Val;
498 }
499};
500
501/// Match a specific integer value or vector with all elements equal to
502/// the value.
503inline specific_intval m_SpecificInt(uint64_t V) { return specific_intval(V); }
504
505/// Match a ConstantInt and bind to its value. This does not match
506/// ConstantInts wider than 64-bits.
507inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; }
508
509//===----------------------------------------------------------------------===//
510// Matcher for any binary operator.
511//
512template <typename LHS_t, typename RHS_t, bool Commutable = false>
513struct AnyBinaryOp_match {
514 LHS_t L;
515 RHS_t R;
516
517 AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
518
519 template <typename OpTy> bool match(OpTy *V) {
520 if (auto *I = dyn_cast<BinaryOperator>(V))
521 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
522 (Commutable && R.match(I->getOperand(0)) &&
523 L.match(I->getOperand(1)));
524 return false;
525 }
526};
527
528template <typename LHS, typename RHS>
529inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) {
530 return AnyBinaryOp_match<LHS, RHS>(L, R);
531}
532
533//===----------------------------------------------------------------------===//
534// Matchers for specific binary operators.
535//
536
537template <typename LHS_t, typename RHS_t, unsigned Opcode,
538 bool Commutable = false>
539struct BinaryOp_match {
540 LHS_t L;
541 RHS_t R;
542
543 BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
544
545 template <typename OpTy> bool match(OpTy *V) {
546 if (V->getValueID() == Value::InstructionVal + Opcode) {
547 auto *I = cast<BinaryOperator>(V);
548 return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
549 (Commutable && R.match(I->getOperand(0)) &&
550 L.match(I->getOperand(1)));
551 }
552 if (auto *CE = dyn_cast<ConstantExpr>(V))
553 return CE->getOpcode() == Opcode &&
554 ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) ||
555 (Commutable && R.match(CE->getOperand(0)) &&
556 L.match(CE->getOperand(1))));
557 return false;
558 }
559};
560
561template <typename LHS, typename RHS>
562inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L,
563 const RHS &R) {
564 return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R);
565}
566
567template <typename LHS, typename RHS>
568inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L,
569 const RHS &R) {
570 return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R);
571}
572
573template <typename LHS, typename RHS>
574inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L,
575 const RHS &R) {
576 return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R);
577}
578
579template <typename LHS, typename RHS>
580inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L,
581 const RHS &R) {
582 return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R);
583}
584
585template <typename LHS, typename RHS>
586inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L,
587 const RHS &R) {
588 return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R);
589}
590
591template <typename LHS, typename RHS>
592inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L,
593 const RHS &R) {
594 return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R);
595}
596
597template <typename LHS, typename RHS>
598inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L,
599 const RHS &R) {
600 return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R);
601}
602
603template <typename LHS, typename RHS>
604inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L,
605 const RHS &R) {
606 return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R);
607}
608
609template <typename LHS, typename RHS>
610inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L,
611 const RHS &R) {
612 return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R);
613}
614
615template <typename LHS, typename RHS>
616inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L,
617 const RHS &R) {
618 return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R);
619}
620
621template <typename LHS, typename RHS>
622inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L,
623 const RHS &R) {
624 return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R);
625}
626
627template <typename LHS, typename RHS>
628inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L,
629 const RHS &R) {
630 return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R);
631}
632
633template <typename LHS, typename RHS>
634inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L,
635 const RHS &R) {
636 return BinaryOp_match<LHS, RHS, Instruction::And>(L, R);
637}
638
639template <typename LHS, typename RHS>
640inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L,
641 const RHS &R) {
642 return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R);
643}
644
645template <typename LHS, typename RHS>
646inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L,
647 const RHS &R) {
648 return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R);
649}
650
651template <typename LHS, typename RHS>
652inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L,
653 const RHS &R) {
654 return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R);
655}
656
657template <typename LHS, typename RHS>
658inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L,
659 const RHS &R) {
660 return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R);
661}
662
663template <typename LHS, typename RHS>
664inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L,
665 const RHS &R) {
666 return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R);
667}
668
669template <typename LHS_t, typename RHS_t, unsigned Opcode,
670 unsigned WrapFlags = 0>
671struct OverflowingBinaryOp_match {
672 LHS_t L;
673 RHS_t R;
674
675 OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS)
676 : L(LHS), R(RHS) {}
677
678 template <typename OpTy> bool match(OpTy *V) {
679 if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) {
680 if (Op->getOpcode() != Opcode)
681 return false;
682 if (WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap &&
683 !Op->hasNoUnsignedWrap())
684 return false;
685 if (WrapFlags & OverflowingBinaryOperator::NoSignedWrap &&
686 !Op->hasNoSignedWrap())
687 return false;
688 return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1));
689 }
690 return false;
691 }
692};
693
694template <typename LHS, typename RHS>
695inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
696 OverflowingBinaryOperator::NoSignedWrap>
697m_NSWAdd(const LHS &L, const RHS &R) {
698 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
699 OverflowingBinaryOperator::NoSignedWrap>(
700 L, R);
701}
702template <typename LHS, typename RHS>
703inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
704 OverflowingBinaryOperator::NoSignedWrap>
705m_NSWSub(const LHS &L, const RHS &R) {
706 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
707 OverflowingBinaryOperator::NoSignedWrap>(
708 L, R);
709}
710template <typename LHS, typename RHS>
711inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
712 OverflowingBinaryOperator::NoSignedWrap>
713m_NSWMul(const LHS &L, const RHS &R) {
714 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
715 OverflowingBinaryOperator::NoSignedWrap>(
716 L, R);
717}
718template <typename LHS, typename RHS>
719inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
720 OverflowingBinaryOperator::NoSignedWrap>
721m_NSWShl(const LHS &L, const RHS &R) {
722 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
723 OverflowingBinaryOperator::NoSignedWrap>(
724 L, R);
725}
726
727template <typename LHS, typename RHS>
728inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
729 OverflowingBinaryOperator::NoUnsignedWrap>
730m_NUWAdd(const LHS &L, const RHS &R) {
731 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
732 OverflowingBinaryOperator::NoUnsignedWrap>(
733 L, R);
734}
735template <typename LHS, typename RHS>
736inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
737 OverflowingBinaryOperator::NoUnsignedWrap>
738m_NUWSub(const LHS &L, const RHS &R) {
739 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
740 OverflowingBinaryOperator::NoUnsignedWrap>(
741 L, R);
742}
743template <typename LHS, typename RHS>
744inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
745 OverflowingBinaryOperator::NoUnsignedWrap>
746m_NUWMul(const LHS &L, const RHS &R) {
747 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
748 OverflowingBinaryOperator::NoUnsignedWrap>(
749 L, R);
750}
751template <typename LHS, typename RHS>
752inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
753 OverflowingBinaryOperator::NoUnsignedWrap>
754m_NUWShl(const LHS &L, const RHS &R) {
755 return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
756 OverflowingBinaryOperator::NoUnsignedWrap>(
757 L, R);
758}
759
760//===----------------------------------------------------------------------===//
761// Class that matches a group of binary opcodes.
762//
763template <typename LHS_t, typename RHS_t, typename Predicate>
764struct BinOpPred_match : Predicate {
765 LHS_t L;
766 RHS_t R;
767
768 BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
769
770 template <typename OpTy> bool match(OpTy *V) {
771 if (auto *I = dyn_cast<Instruction>(V))
772 return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) &&
773 R.match(I->getOperand(1));
774 if (auto *CE = dyn_cast<ConstantExpr>(V))
775 return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) &&
776 R.match(CE->getOperand(1));
777 return false;
778 }
779};
780
781struct is_shift_op {
782 bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); }
783};
784
785struct is_right_shift_op {
786 bool isOpType(unsigned Opcode) {
787 return Opcode == Instruction::LShr || Opcode == Instruction::AShr;
788 }
789};
790
791struct is_logical_shift_op {
792 bool isOpType(unsigned Opcode) {
793 return Opcode == Instruction::LShr || Opcode == Instruction::Shl;
794 }
795};
796
797struct is_bitwiselogic_op {
798 bool isOpType(unsigned Opcode) {
799 return Instruction::isBitwiseLogicOp(Opcode);
800 }
801};
802
803struct is_idiv_op {
804 bool isOpType(unsigned Opcode) {
805 return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
806 }
807};
808
809/// Matches shift operations.
810template <typename LHS, typename RHS>
811inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L,
812 const RHS &R) {
813 return BinOpPred_match<LHS, RHS, is_shift_op>(L, R);
814}
815
816/// Matches logical shift operations.
817template <typename LHS, typename RHS>
818inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L,
819 const RHS &R) {
820 return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R);
821}
822
823/// Matches logical shift operations.
824template <typename LHS, typename RHS>
825inline BinOpPred_match<LHS, RHS, is_logical_shift_op>
826m_LogicalShift(const LHS &L, const RHS &R) {
827 return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R);
828}
829
830/// Matches bitwise logic operations.
831template <typename LHS, typename RHS>
832inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op>
833m_BitwiseLogic(const LHS &L, const RHS &R) {
834 return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R);
835}
836
837/// Matches integer division operations.
838template <typename LHS, typename RHS>
839inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L,
840 const RHS &R) {
841 return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R);
842}
843
844//===----------------------------------------------------------------------===//
845// Class that matches exact binary ops.
846//
847template <typename SubPattern_t> struct Exact_match {
848 SubPattern_t SubPattern;
849
850 Exact_match(const SubPattern_t &SP) : SubPattern(SP) {}
851
852 template <typename OpTy> bool match(OpTy *V) {
853 if (auto *PEO = dyn_cast<PossiblyExactOperator>(V))
854 return PEO->isExact() && SubPattern.match(V);
855 return false;
856 }
857};
858
859template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) {
860 return SubPattern;
861}
862
863//===----------------------------------------------------------------------===//
864// Matchers for CmpInst classes
865//
866
867template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy,
868 bool Commutable = false>
869struct CmpClass_match {
870 PredicateTy &Predicate;
871 LHS_t L;
872 RHS_t R;
873
874 CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS)
875 : Predicate(Pred), L(LHS), R(RHS) {}
876
877 template <typename OpTy> bool match(OpTy *V) {
878 if (auto *I = dyn_cast<Class>(V))
879 if ((L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
880 (Commutable && R.match(I->getOperand(0)) &&
881 L.match(I->getOperand(1)))) {
882 Predicate = I->getPredicate();
883 return true;
884 }
885 return false;
886 }
887};
888
889template <typename LHS, typename RHS>
890inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>
891m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
892 return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R);
893}
894
895template <typename LHS, typename RHS>
896inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>
897m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
898 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R);
899}
900
901template <typename LHS, typename RHS>
902inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>
903m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
904 return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R);
905}
906
907//===----------------------------------------------------------------------===//
908// Matchers for SelectInst classes
909//
910
911template <typename Cond_t, typename LHS_t, typename RHS_t>
912struct SelectClass_match {
913 Cond_t C;
914 LHS_t L;
915 RHS_t R;
916
917 SelectClass_match(const Cond_t &Cond, const LHS_t &LHS, const RHS_t &RHS)
918 : C(Cond), L(LHS), R(RHS) {}
919
920 template <typename OpTy> bool match(OpTy *V) {
921 if (auto *I = dyn_cast<SelectInst>(V))
922 return C.match(I->getOperand(0)) && L.match(I->getOperand(1)) &&
923 R.match(I->getOperand(2));
924 return false;
925 }
926};
927
928template <typename Cond, typename LHS, typename RHS>
929inline SelectClass_match<Cond, LHS, RHS> m_Select(const Cond &C, const LHS &L,
930 const RHS &R) {
931 return SelectClass_match<Cond, LHS, RHS>(C, L, R);
932}
933
934/// This matches a select of two constants, e.g.:
935/// m_SelectCst<-1, 0>(m_Value(V))
936template <int64_t L, int64_t R, typename Cond>
937inline SelectClass_match<Cond, constantint_match<L>, constantint_match<R>>
938m_SelectCst(const Cond &C) {
939 return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>());
940}
941
942//===----------------------------------------------------------------------===//
943// Matchers for CastInst classes
944//
945
946template <typename Op_t, unsigned Opcode> struct CastClass_match {
947 Op_t Op;
948
949 CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {}
950
951 template <typename OpTy> bool match(OpTy *V) {
952 if (auto *O = dyn_cast<Operator>(V))
953 return O->getOpcode() == Opcode && Op.match(O->getOperand(0));
954 return false;
955 }
956};
957
958/// Matches BitCast.
959template <typename OpTy>
960inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) {
961 return CastClass_match<OpTy, Instruction::BitCast>(Op);
962}
963
964/// Matches PtrToInt.
965template <typename OpTy>
966inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) {
967 return CastClass_match<OpTy, Instruction::PtrToInt>(Op);
968}
969
970/// Matches Trunc.
971template <typename OpTy>
972inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) {
973 return CastClass_match<OpTy, Instruction::Trunc>(Op);
974}
975
976/// Matches SExt.
977template <typename OpTy>
978inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) {
979 return CastClass_match<OpTy, Instruction::SExt>(Op);
980}
981
982/// Matches ZExt.
983template <typename OpTy>
984inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) {
985 return CastClass_match<OpTy, Instruction::ZExt>(Op);
986}
987
988template <typename OpTy>
989inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
990 CastClass_match<OpTy, Instruction::SExt>>
991m_ZExtOrSExt(const OpTy &Op) {
992 return m_CombineOr(m_ZExt(Op), m_SExt(Op));
993}
994
995/// Matches UIToFP.
996template <typename OpTy>
997inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) {
998 return CastClass_match<OpTy, Instruction::UIToFP>(Op);
999}
1000
1001/// Matches SIToFP.
1002template <typename OpTy>
1003inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) {
1004 return CastClass_match<OpTy, Instruction::SIToFP>(Op);
1005}
1006
1007/// Matches FPTrunc
1008template <typename OpTy>
1009inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) {
1010 return CastClass_match<OpTy, Instruction::FPTrunc>(Op);
1011}
1012
1013/// Matches FPExt
1014template <typename OpTy>
1015inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) {
1016 return CastClass_match<OpTy, Instruction::FPExt>(Op);
1017}
1018
1019//===----------------------------------------------------------------------===//
1020// Matcher for LoadInst classes
1021//
1022
1023template <typename Op_t> struct LoadClass_match {
1024 Op_t Op;
1025
1026 LoadClass_match(const Op_t &OpMatch) : Op(OpMatch) {}
1027
1028 template <typename OpTy> bool match(OpTy *V) {
1029 if (auto *LI = dyn_cast<LoadInst>(V))
1030 return Op.match(LI->getPointerOperand());
1031 return false;
1032 }
1033};
1034
1035/// Matches LoadInst.
1036template <typename OpTy> inline LoadClass_match<OpTy> m_Load(const OpTy &Op) {
1037 return LoadClass_match<OpTy>(Op);
1038}
1039//===----------------------------------------------------------------------===//
1040// Matchers for unary operators
1041//
1042
1043template <typename LHS_t> struct not_match {
1044 LHS_t L;
1045
1046 not_match(const LHS_t &LHS) : L(LHS) {}
1047
1048 template <typename OpTy> bool match(OpTy *V) {
1049 if (auto *O = dyn_cast<Operator>(V))
1050 if (O->getOpcode() == Instruction::Xor) {
1051 if (isAllOnes(O->getOperand(1)))
1052 return L.match(O->getOperand(0));
1053 if (isAllOnes(O->getOperand(0)))
1054 return L.match(O->getOperand(1));
1055 }
1056 return false;
1057 }
1058
1059private:
1060 bool isAllOnes(Value *V) {
1061 return isa<Constant>(V) && cast<Constant>(V)->isAllOnesValue();
1062 }
1063};
1064
1065template <typename LHS> inline not_match<LHS> m_Not(const LHS &L) { return L; }
1066
1067template <typename LHS_t> struct neg_match {
1068 LHS_t L;
1069
1070 neg_match(const LHS_t &LHS) : L(LHS) {}
1071
1072 template <typename OpTy> bool match(OpTy *V) {
1073 if (auto *O = dyn_cast<Operator>(V))
1074 if (O->getOpcode() == Instruction::Sub)
1075 return matchIfNeg(O->getOperand(0), O->getOperand(1));
1076 return false;
1077 }
1078
1079private:
1080 bool matchIfNeg(Value *LHS, Value *RHS) {
1081 return ((isa<ConstantInt>(LHS) && cast<ConstantInt>(LHS)->isZero()) ||
1082 isa<ConstantAggregateZero>(LHS)) &&
1083 L.match(RHS);
1084 }
1085};
1086
1087/// Match an integer negate.
1088template <typename LHS> inline neg_match<LHS> m_Neg(const LHS &L) { return L; }
1089
1090template <typename LHS_t> struct fneg_match {
1091 LHS_t L;
1092
1093 fneg_match(const LHS_t &LHS) : L(LHS) {}
1094
1095 template <typename OpTy> bool match(OpTy *V) {
1096 if (auto *O = dyn_cast<Operator>(V))
1097 if (O->getOpcode() == Instruction::FSub)
1098 return matchIfFNeg(O->getOperand(0), O->getOperand(1));
1099 return false;
1100 }
1101
1102private:
1103 bool matchIfFNeg(Value *LHS, Value *RHS) {
1104 if (const auto *C = dyn_cast<Constant>(LHS))
1105 return C->isNegativeZeroValue() && L.match(RHS);
1106 return false;
1107 }
1108};
1109
1110/// Match a floating point negate.
1111template <typename LHS> inline fneg_match<LHS> m_FNeg(const LHS &L) {
1112 return L;
1113}
1114
1115//===----------------------------------------------------------------------===//
1116// Matchers for control flow.
1117//
1118
1119struct br_match {
1120 BasicBlock *&Succ;
1121
1122 br_match(BasicBlock *&Succ) : Succ(Succ) {}
1123
1124 template <typename OpTy> bool match(OpTy *V) {
1125 if (auto *BI = dyn_cast<BranchInst>(V))
1126 if (BI->isUnconditional()) {
1127 Succ = BI->getSuccessor(0);
1128 return true;
1129 }
1130 return false;
1131 }
1132};
1133
1134inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); }
1135
1136template <typename Cond_t> struct brc_match {
1137 Cond_t Cond;
1138 BasicBlock *&T, *&F;
1139
1140 brc_match(const Cond_t &C, BasicBlock *&t, BasicBlock *&f)
1141 : Cond(C), T(t), F(f) {}
1142
1143 template <typename OpTy> bool match(OpTy *V) {
1144 if (auto *BI = dyn_cast<BranchInst>(V))
1145 if (BI->isConditional() && Cond.match(BI->getCondition())) {
1146 T = BI->getSuccessor(0);
1147 F = BI->getSuccessor(1);
1148 return true;
1149 }
1150 return false;
1151 }
1152};
1153
1154template <typename Cond_t>
1155inline brc_match<Cond_t> m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) {
1156 return brc_match<Cond_t>(C, T, F);
1157}
1158
1159//===----------------------------------------------------------------------===//
1160// Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y).
1161//
1162
1163template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t,
1164 bool Commutable = false>
1165struct MaxMin_match {
1166 LHS_t L;
1167 RHS_t R;
1168
1169 MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1170
1171 template <typename OpTy> bool match(OpTy *V) {
1172 // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x".
1173 auto *SI = dyn_cast<SelectInst>(V);
1174 if (!SI)
1175 return false;
1176 auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition());
1177 if (!Cmp)
1178 return false;
1179 // At this point we have a select conditioned on a comparison. Check that
1180 // it is the values returned by the select that are being compared.
1181 Value *TrueVal = SI->getTrueValue();
1182 Value *FalseVal = SI->getFalseValue();
1183 Value *LHS = Cmp->getOperand(0);
1184 Value *RHS = Cmp->getOperand(1);
1185 if ((TrueVal != LHS || FalseVal != RHS) &&
1186 (TrueVal != RHS || FalseVal != LHS))
1187 return false;
1188 typename CmpInst_t::Predicate Pred =
1189 LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate();
1190 // Does "(x pred y) ? x : y" represent the desired max/min operation?
1191 if (!Pred_t::match(Pred))
1192 return false;
1193 // It does! Bind the operands.
1194 return (L.match(LHS) && R.match(RHS)) ||
1195 (Commutable && R.match(LHS) && L.match(RHS));
1196 }
1197};
1198
1199/// Helper class for identifying signed max predicates.
1200struct smax_pred_ty {
1201 static bool match(ICmpInst::Predicate Pred) {
1202 return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE;
1203 }
1204};
1205
1206/// Helper class for identifying signed min predicates.
1207struct smin_pred_ty {
1208 static bool match(ICmpInst::Predicate Pred) {
1209 return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE;
1210 }
1211};
1212
1213/// Helper class for identifying unsigned max predicates.
1214struct umax_pred_ty {
1215 static bool match(ICmpInst::Predicate Pred) {
1216 return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE;
1217 }
1218};
1219
1220/// Helper class for identifying unsigned min predicates.
1221struct umin_pred_ty {
1222 static bool match(ICmpInst::Predicate Pred) {
1223 return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE;
1224 }
1225};
1226
1227/// Helper class for identifying ordered max predicates.
1228struct ofmax_pred_ty {
1229 static bool match(FCmpInst::Predicate Pred) {
1230 return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE;
1231 }
1232};
1233
1234/// Helper class for identifying ordered min predicates.
1235struct ofmin_pred_ty {
1236 static bool match(FCmpInst::Predicate Pred) {
1237 return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE;
1238 }
1239};
1240
1241/// Helper class for identifying unordered max predicates.
1242struct ufmax_pred_ty {
1243 static bool match(FCmpInst::Predicate Pred) {
1244 return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE;
1245 }
1246};
1247
1248/// Helper class for identifying unordered min predicates.
1249struct ufmin_pred_ty {
1250 static bool match(FCmpInst::Predicate Pred) {
1251 return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE;
1252 }
1253};
1254
1255template <typename LHS, typename RHS>
1256inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L,
1257 const RHS &R) {
1258 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R);
1259}
1260
1261template <typename LHS, typename RHS>
1262inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L,
1263 const RHS &R) {
1264 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R);
1265}
1266
1267template <typename LHS, typename RHS>
1268inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L,
1269 const RHS &R) {
1270 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R);
1271}
1272
1273template <typename LHS, typename RHS>
1274inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L,
1275 const RHS &R) {
1276 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R);
1277}
1278
1279/// Match an 'ordered' floating point maximum function.
1280/// Floating point has one special value 'NaN'. Therefore, there is no total
1281/// order. However, if we can ignore the 'NaN' value (for example, because of a
1282/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1283/// semantics. In the presence of 'NaN' we have to preserve the original
1284/// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate.
1285///
1286/// max(L, R) iff L and R are not NaN
1287/// m_OrdFMax(L, R) = R iff L or R are NaN
1288template <typename LHS, typename RHS>
1289inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L,
1290 const RHS &R) {
1291 return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R);
1292}
1293
1294/// Match an 'ordered' floating point minimum function.
1295/// Floating point has one special value 'NaN'. Therefore, there is no total
1296/// order. However, if we can ignore the 'NaN' value (for example, because of a
1297/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1298/// semantics. In the presence of 'NaN' we have to preserve the original
1299/// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate.
1300///
1301/// min(L, R) iff L and R are not NaN
1302/// m_OrdFMin(L, R) = R iff L or R are NaN
1303template <typename LHS, typename RHS>
1304inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L,
1305 const RHS &R) {
1306 return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R);
1307}
1308
1309/// Match an 'unordered' floating point maximum function.
1310/// Floating point has one special value 'NaN'. Therefore, there is no total
1311/// order. However, if we can ignore the 'NaN' value (for example, because of a
1312/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1313/// semantics. In the presence of 'NaN' we have to preserve the original
1314/// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate.
1315///
1316/// max(L, R) iff L and R are not NaN
1317/// m_UnordFMax(L, R) = L iff L or R are NaN
1318template <typename LHS, typename RHS>
1319inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>
1320m_UnordFMax(const LHS &L, const RHS &R) {
1321 return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R);
1322}
1323
1324/// Match an 'unordered' floating point minimum function.
1325/// Floating point has one special value 'NaN'. Therefore, there is no total
1326/// order. However, if we can ignore the 'NaN' value (for example, because of a
1327/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1328/// semantics. In the presence of 'NaN' we have to preserve the original
1329/// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate.
1330///
1331/// min(L, R) iff L and R are not NaN
1332/// m_UnordFMin(L, R) = L iff L or R are NaN
1333template <typename LHS, typename RHS>
1334inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>
1335m_UnordFMin(const LHS &L, const RHS &R) {
1336 return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R);
1337}
1338
1339//===----------------------------------------------------------------------===//
1340// Matchers for overflow check patterns: e.g. (a + b) u< a
1341//
1342
1343template <typename LHS_t, typename RHS_t, typename Sum_t>
1344struct UAddWithOverflow_match {
1345 LHS_t L;
1346 RHS_t R;
1347 Sum_t S;
1348
1349 UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S)
1350 : L(L), R(R), S(S) {}
1351
1352 template <typename OpTy> bool match(OpTy *V) {
1353 Value *ICmpLHS, *ICmpRHS;
1354 ICmpInst::Predicate Pred;
1355 if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V))
1356 return false;
1357
1358 Value *AddLHS, *AddRHS;
1359 auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS));
1360
1361 // (a + b) u< a, (a + b) u< b
1362 if (Pred == ICmpInst::ICMP_ULT)
1363 if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS))
1364 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
1365
1366 // a >u (a + b), b >u (a + b)
1367 if (Pred == ICmpInst::ICMP_UGT)
1368 if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS))
1369 return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
1370
1371 return false;
1372 }
1373};
1374
1375/// Match an icmp instruction checking for unsigned overflow on addition.
1376///
1377/// S is matched to the addition whose result is being checked for overflow, and
1378/// L and R are matched to the LHS and RHS of S.
1379template <typename LHS_t, typename RHS_t, typename Sum_t>
1380UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>
1381m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) {
1382 return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S);
1383}
1384
1385template <typename Opnd_t> struct Argument_match {
1386 unsigned OpI;
1387 Opnd_t Val;
1388
1389 Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {}
1390
1391 template <typename OpTy> bool match(OpTy *V) {
1392 CallSite CS(V);
1393 return CS.isCall() && Val.match(CS.getArgument(OpI));
1394 }
1395};
1396
1397/// Match an argument.
1398template <unsigned OpI, typename Opnd_t>
1399inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) {
1400 return Argument_match<Opnd_t>(OpI, Op);
1401}
1402
1403/// Intrinsic matchers.
1404struct IntrinsicID_match {
1405 unsigned ID;
1406
1407 IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {}
1408
1409 template <typename OpTy> bool match(OpTy *V) {
1410 if (const auto *CI = dyn_cast<CallInst>(V))
1411 if (const auto *F = CI->getCalledFunction())
1412 return F->getIntrinsicID() == ID;
1413 return false;
1414 }
1415};
1416
1417/// Intrinsic matches are combinations of ID matchers, and argument
1418/// matchers. Higher arity matcher are defined recursively in terms of and-ing
1419/// them with lower arity matchers. Here's some convenient typedefs for up to
1420/// several arguments, and more can be added as needed
1421template <typename T0 = void, typename T1 = void, typename T2 = void,
1422 typename T3 = void, typename T4 = void, typename T5 = void,
1423 typename T6 = void, typename T7 = void, typename T8 = void,
1424 typename T9 = void, typename T10 = void>
1425struct m_Intrinsic_Ty;
1426template <typename T0> struct m_Intrinsic_Ty<T0> {
1427 using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>;
1428};
1429template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> {
1430 using Ty =
1431 match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>;
1432};
1433template <typename T0, typename T1, typename T2>
1434struct m_Intrinsic_Ty<T0, T1, T2> {
1435 using Ty =
1436 match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty,
1437 Argument_match<T2>>;
1438};
1439template <typename T0, typename T1, typename T2, typename T3>
1440struct m_Intrinsic_Ty<T0, T1, T2, T3> {
1441 using Ty =
1442 match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty,
1443 Argument_match<T3>>;
1444};
1445
1446/// Match intrinsic calls like this:
1447/// m_Intrinsic<Intrinsic::fabs>(m_Value(X))
1448template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() {
1449 return IntrinsicID_match(IntrID);
1450}
1451
1452template <Intrinsic::ID IntrID, typename T0>
1453inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) {
1454 return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0));
1455}
1456
1457template <Intrinsic::ID IntrID, typename T0, typename T1>
1458inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0,
1459 const T1 &Op1) {
1460 return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1));
1461}
1462
1463template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2>
1464inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty
1465m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) {
1466 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2));
1467}
1468
1469template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
1470 typename T3>
1471inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty
1472m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) {
1473 return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3));
1474}
1475
1476// Helper intrinsic matching specializations.
1477template <typename Opnd0>
1478inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) {
1479 return m_Intrinsic<Intrinsic::bitreverse>(Op0);
1480}
1481
1482template <typename Opnd0>
1483inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) {
1484 return m_Intrinsic<Intrinsic::bswap>(Op0);
1485}
1486
1487template <typename Opnd0, typename Opnd1>
1488inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0,
1489 const Opnd1 &Op1) {
1490 return m_Intrinsic<Intrinsic::minnum>(Op0, Op1);
1491}
1492
1493template <typename Opnd0, typename Opnd1>
1494inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0,
1495 const Opnd1 &Op1) {
1496 return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1);
1497}
1498
1499template <typename Opnd_t> struct Signum_match {
1500 Opnd_t Val;
1501 Signum_match(const Opnd_t &V) : Val(V) {}
1502
1503 template <typename OpTy> bool match(OpTy *V) {
1504 unsigned TypeSize = V->getType()->getScalarSizeInBits();
1505 if (TypeSize == 0)
1506 return false;
1507
1508 unsigned ShiftWidth = TypeSize - 1;
1509 Value *OpL = nullptr, *OpR = nullptr;
1510
1511 // This is the representation of signum we match:
1512 //
1513 // signum(x) == (x >> 63) | (-x >>u 63)
1514 //
1515 // An i1 value is its own signum, so it's correct to match
1516 //
1517 // signum(x) == (x >> 0) | (-x >>u 0)
1518 //
1519 // for i1 values.
1520
1521 auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth));
1522 auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth));
1523 auto Signum = m_Or(LHS, RHS);
1524
1525 return Signum.match(V) && OpL == OpR && Val.match(OpL);
1526 }
1527};
1528
1529/// Matches a signum pattern.
1530///
1531/// signum(x) =
1532/// x > 0 -> 1
1533/// x == 0 -> 0
1534/// x < 0 -> -1
1535template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) {
1536 return Signum_match<Val_t>(V);
1537}
1538
1539//===----------------------------------------------------------------------===//
1540// Matchers for two-operands operators with the operators in either order
1541//
1542
1543/// Matches a BinaryOperator with LHS and RHS in either order.
1544template <typename LHS, typename RHS>
1545inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) {
1546 return AnyBinaryOp_match<LHS, RHS, true>(L, R);
1547}
1548
1549/// Matches an ICmp with a predicate over LHS and RHS in either order.
1550/// Does not swap the predicate.
1551template <typename LHS, typename RHS>
1552inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>
1553m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1554 return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L,
1555 R);
1556}
1557
1558/// Matches a Add with LHS and RHS in either order.
1559template <typename LHS, typename RHS>
1560inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L,
1561 const RHS &R) {
1562 return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R);
1563}
1564
1565/// Matches a Mul with LHS and RHS in either order.
1566template <typename LHS, typename RHS>
1567inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L,
1568 const RHS &R) {
1569 return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R);
1570}
1571
1572/// Matches an And with LHS and RHS in either order.
1573template <typename LHS, typename RHS>
1574inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L,
1575 const RHS &R) {
1576 return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R);
1577}
1578
1579/// Matches an Or with LHS and RHS in either order.
1580template <typename LHS, typename RHS>
1581inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L,
1582 const RHS &R) {
1583 return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R);
1584}
1585
1586/// Matches an Xor with LHS and RHS in either order.
1587template <typename LHS, typename RHS>
1588inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L,
1589 const RHS &R) {
1590 return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R);
1591}
1592
1593/// Matches an SMin with LHS and RHS in either order.
1594template <typename LHS, typename RHS>
1595inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>
1596m_c_SMin(const LHS &L, const RHS &R) {
1597 return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R);
1598}
1599/// Matches an SMax with LHS and RHS in either order.
1600template <typename LHS, typename RHS>
1601inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>
1602m_c_SMax(const LHS &L, const RHS &R) {
1603 return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R);
1604}
1605/// Matches a UMin with LHS and RHS in either order.
1606template <typename LHS, typename RHS>
1607inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>
1608m_c_UMin(const LHS &L, const RHS &R) {
1609 return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R);
1610}
1611/// Matches a UMax with LHS and RHS in either order.
1612template <typename LHS, typename RHS>
1613inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>
1614m_c_UMax(const LHS &L, const RHS &R) {
1615 return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R);
1616}
1617
1618/// Matches FAdd with LHS and RHS in either order.
1619template <typename LHS, typename RHS>
1620inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true>
1621m_c_FAdd(const LHS &L, const RHS &R) {
1622 return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R);
1623}
1624
1625/// Matches FMul with LHS and RHS in either order.
1626template <typename LHS, typename RHS>
1627inline BinaryOp_match<LHS, RHS, Instruction::FMul, true>
1628m_c_FMul(const LHS &L, const RHS &R) {
1629 return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R);
1630}
1631
1632} // end namespace PatternMatch
1633} // end namespace llvm
1634
1635#endif // LLVM_IR_PATTERNMATCH_H