File: | lib/Analysis/InstructionSimplify.cpp |
Warning: | line 1197, column 35 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- InstructionSimplify.cpp - Fold instruction operands ----------------===// | |||
2 | // | |||
3 | // The LLVM Compiler Infrastructure | |||
4 | // | |||
5 | // This file is distributed under the University of Illinois Open Source | |||
6 | // License. See LICENSE.TXT for details. | |||
7 | // | |||
8 | //===----------------------------------------------------------------------===// | |||
9 | // | |||
10 | // This file implements routines for folding instructions into simpler forms | |||
11 | // that do not require creating new instructions. This does constant folding | |||
12 | // ("add i32 1, 1" -> "2") but can also handle non-constant operands, either | |||
13 | // returning a constant ("and i32 %x, 0" -> "0") or an already existing value | |||
14 | // ("and i32 %x, %x" -> "%x"). All operands are assumed to have already been | |||
15 | // simplified: This is usually true and assuming it simplifies the logic (if | |||
16 | // they have not been simplified then results are correct but maybe suboptimal). | |||
17 | // | |||
18 | //===----------------------------------------------------------------------===// | |||
19 | ||||
20 | #include "llvm/Analysis/InstructionSimplify.h" | |||
21 | #include "llvm/ADT/SetVector.h" | |||
22 | #include "llvm/ADT/Statistic.h" | |||
23 | #include "llvm/Analysis/AliasAnalysis.h" | |||
24 | #include "llvm/Analysis/AssumptionCache.h" | |||
25 | #include "llvm/Analysis/CaptureTracking.h" | |||
26 | #include "llvm/Analysis/CmpInstAnalysis.h" | |||
27 | #include "llvm/Analysis/ConstantFolding.h" | |||
28 | #include "llvm/Analysis/LoopAnalysisManager.h" | |||
29 | #include "llvm/Analysis/MemoryBuiltins.h" | |||
30 | #include "llvm/Analysis/ValueTracking.h" | |||
31 | #include "llvm/Analysis/VectorUtils.h" | |||
32 | #include "llvm/IR/ConstantRange.h" | |||
33 | #include "llvm/IR/DataLayout.h" | |||
34 | #include "llvm/IR/Dominators.h" | |||
35 | #include "llvm/IR/GetElementPtrTypeIterator.h" | |||
36 | #include "llvm/IR/GlobalAlias.h" | |||
37 | #include "llvm/IR/Operator.h" | |||
38 | #include "llvm/IR/PatternMatch.h" | |||
39 | #include "llvm/IR/ValueHandle.h" | |||
40 | #include "llvm/Support/KnownBits.h" | |||
41 | #include <algorithm> | |||
42 | using namespace llvm; | |||
43 | using namespace llvm::PatternMatch; | |||
44 | ||||
45 | #define DEBUG_TYPE"instsimplify" "instsimplify" | |||
46 | ||||
47 | enum { RecursionLimit = 3 }; | |||
48 | ||||
49 | STATISTIC(NumExpand, "Number of expansions")static llvm::Statistic NumExpand = {"instsimplify", "NumExpand" , "Number of expansions", {0}, {false}}; | |||
50 | STATISTIC(NumReassoc, "Number of reassociations")static llvm::Statistic NumReassoc = {"instsimplify", "NumReassoc" , "Number of reassociations", {0}, {false}}; | |||
51 | ||||
52 | static Value *SimplifyAndInst(Value *, Value *, const SimplifyQuery &, unsigned); | |||
53 | static Value *SimplifyBinOp(unsigned, Value *, Value *, const SimplifyQuery &, | |||
54 | unsigned); | |||
55 | static Value *SimplifyFPBinOp(unsigned, Value *, Value *, const FastMathFlags &, | |||
56 | const SimplifyQuery &, unsigned); | |||
57 | static Value *SimplifyCmpInst(unsigned, Value *, Value *, const SimplifyQuery &, | |||
58 | unsigned); | |||
59 | static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, | |||
60 | const SimplifyQuery &Q, unsigned MaxRecurse); | |||
61 | static Value *SimplifyOrInst(Value *, Value *, const SimplifyQuery &, unsigned); | |||
62 | static Value *SimplifyXorInst(Value *, Value *, const SimplifyQuery &, unsigned); | |||
63 | static Value *SimplifyCastInst(unsigned, Value *, Type *, | |||
64 | const SimplifyQuery &, unsigned); | |||
65 | ||||
66 | /// For a boolean type or a vector of boolean type, return false or a vector | |||
67 | /// with every element false. | |||
68 | static 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. | |||
74 | static Constant *getTrue(Type *Ty) { | |||
75 | return ConstantInt::getTrue(Ty); | |||
76 | } | |||
77 | ||||
78 | /// isSameCompare - Is V equivalent to the comparison "LHS Pred RHS"? | |||
79 | static bool isSameCompare(Value *V, CmpInst::Predicate Pred, Value *LHS, | |||
80 | Value *RHS) { | |||
81 | CmpInst *Cmp = dyn_cast<CmpInst>(V); | |||
82 | if (!Cmp) | |||
83 | return false; | |||
84 | CmpInst::Predicate CPred = Cmp->getPredicate(); | |||
85 | Value *CLHS = Cmp->getOperand(0), *CRHS = Cmp->getOperand(1); | |||
86 | if (CPred == Pred && CLHS == LHS && CRHS == RHS) | |||
87 | return true; | |||
88 | return CPred == CmpInst::getSwappedPredicate(Pred) && CLHS == RHS && | |||
89 | CRHS == LHS; | |||
90 | } | |||
91 | ||||
92 | /// Does the given value dominate the specified phi node? | |||
93 | static bool ValueDominatesPHI(Value *V, PHINode *P, const DominatorTree *DT) { | |||
94 | Instruction *I = dyn_cast<Instruction>(V); | |||
95 | if (!I) | |||
96 | // Arguments and constants dominate all instructions. | |||
97 | return true; | |||
98 | ||||
99 | // If we are processing instructions (and/or basic blocks) that have not been | |||
100 | // fully added to a function, the parent nodes may still be null. Simply | |||
101 | // return the conservative answer in these cases. | |||
102 | if (!I->getParent() || !P->getParent() || !I->getParent()->getParent()) | |||
103 | return false; | |||
104 | ||||
105 | // If we have a DominatorTree then do a precise test. | |||
106 | if (DT) | |||
107 | 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. | |||
123 | static 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. | |||
181 | static 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~svn329677/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. | |||
281 | static 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~svn329677/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. | |||
353 | static 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~svn329677/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. | |||
435 | static 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~svn329677/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. | |||
478 | static 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~svn329677/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 | ||||
512 | static 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. | |||
528 | static 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 | ||||
584 | Value *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. | |||
599 | static 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~svn329677/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~svn329677/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. | |||
642 | static Constant *computePointerDifference(const DataLayout &DL, Value *LHS, | |||
643 | Value *RHS) { | |||
644 | Constant *LHSOffset = stripAndComputeConstantOffsets(DL, LHS); | |||
645 | Constant *RHSOffset = stripAndComputeConstantOffsets(DL, RHS); | |||
646 | ||||
647 | // If LHS and RHS are not related via constant offsets to the same base | |||
648 | // value, there is nothing we can do here. | |||
649 | if (LHS != RHS) | |||
650 | return nullptr; | |||
651 | ||||
652 | // Otherwise, the difference of LHS - RHS can be computed as: | |||
653 | // LHS - RHS | |||
654 | // = (LHSOffset + Base) - (RHSOffset + Base) | |||
655 | // = LHSOffset - RHSOffset | |||
656 | return ConstantExpr::getSub(LHSOffset, RHSOffset); | |||
657 | } | |||
658 | ||||
659 | /// Given operands for a Sub, see if we can fold the result. | |||
660 | /// If not, this returns null. | |||
661 | static 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 | ||||
789 | Value *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. | |||
796 | static 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 | ||||
851 | Value *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). | |||
857 | static 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. | |||
911 | static 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. | |||
920 | static 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. | |||
971 | static 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. | |||
1029 | static 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. | |||
1072 | static Value *SimplifySDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, | |||
1073 | unsigned MaxRecurse) { | |||
1074 | return simplifyDiv(Instruction::SDiv, Op0, Op1, Q, MaxRecurse); | |||
1075 | } | |||
1076 | ||||
1077 | Value *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. | |||
1083 | static Value *SimplifyUDivInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, | |||
1084 | unsigned MaxRecurse) { | |||
1085 | return simplifyDiv(Instruction::UDiv, Op0, Op1, Q, MaxRecurse); | |||
1086 | } | |||
1087 | ||||
1088 | Value *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. | |||
1094 | static Value *SimplifySRemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, | |||
1095 | unsigned MaxRecurse) { | |||
1096 | return simplifyRem(Instruction::SRem, Op0, Op1, Q, MaxRecurse); | |||
1097 | } | |||
1098 | ||||
1099 | Value *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. | |||
1105 | static Value *SimplifyURemInst(Value *Op0, Value *Op1, const SimplifyQuery &Q, | |||
1106 | unsigned MaxRecurse) { | |||
1107 | return simplifyRem(Instruction::URem, Op0, Op1, Q, MaxRecurse); | |||
1108 | } | |||
1109 | ||||
1110 | Value *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. | |||
1115 | static 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. | |||
1143 | static 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. | |||
1189 | static 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. | |||
1216 | static 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 | ||||
1233 | Value *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. | |||
1240 | static 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 | ||||
1254 | Value *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. | |||
1261 | static 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 | ||||
1285 | Value *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. | |||
1292 | static Value *simplifyUnsignedRangeCheck(ICmpInst *ZeroICmp, | |||
1293 | ICmpInst *UnsignedICmp, bool IsAnd) { | |||
1294 | Value *X, *Y; | |||
1295 | ||||
1296 | ICmpInst::Predicate EqPred; | |||
1297 | if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(Y), m_Zero())) || | |||
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))) && | |||
1303 | ICmpInst::isUnsigned(UnsignedPred)) | |||
1304 | ; | |||
1305 | else if (match(UnsignedICmp, | |||
1306 | m_ICmp(UnsignedPred, m_Value(Y), m_Specific(X))) && | |||
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. | |||
1335 | static 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. | |||
1360 | static 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. | |||
1387 | static 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 | ||||
1424 | static 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 | ||||
1461 | static 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 | ||||
1507 | static 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 | ||||
1532 | static 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 | ||||
1578 | static Value *simplifyOrOfICmps(ICmpInst *Op0, ICmpInst *Op1) { | |||
1579 | if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false)) | |||
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 | ||||
1603 | static 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 | ||||
1640 | static 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. | |||
1677 | static 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 | ||||
1775 | Value *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. | |||
1781 | static 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 | ||||
1911 | Value *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. | |||
1917 | static 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 | ||||
1956 | Value *llvm::SimplifyXorInst(Value *Op0, Value *Op1, const SimplifyQuery &Q) { | |||
1957 | return ::SimplifyXorInst(Op0, Op1, Q, RecursionLimit); | |||
1958 | } | |||
1959 | ||||
1960 | ||||
1961 | static 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. | |||
1968 | static 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. | |||
2013 | static Constant * | |||
2014 | computePointerICmp(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. | |||
2196 | static 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. | |||
2270 | static 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~svn329677/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. | |||
2335 | static 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~svn329677/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 | ||||
2489 | static Value *simplifyICmpWithConstant(CmpInst::Predicate Pred, Value *LHS, | |||
2490 | Value *RHS) { | |||
2491 | Type *ITy = GetCompareTy(RHS); // The return type. | |||
2492 | ||||
2493 | Value *X; | |||
2494 | // Sign-bit checks can be optimized to true/false after unsigned | |||
2495 | // floating-point casts: | |||
2496 | // icmp slt (bitcast (uitofp X)), 0 --> false | |||
2497 | // icmp sgt (bitcast (uitofp X)), -1 --> true | |||
2498 | if (match(LHS, m_BitCast(m_UIToFP(m_Value(X))))) { | |||
2499 | if (Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero())) | |||
2500 | return ConstantInt::getFalse(ITy); | |||
2501 | if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes())) | |||
2502 | return ConstantInt::getTrue(ITy); | |||
2503 | } | |||
2504 | ||||
2505 | const APInt *C; | |||
2506 | if (!match(RHS, m_APInt(C))) | |||
2507 | return nullptr; | |||
2508 | ||||
2509 | // Rule out tautological comparisons (eg., ult 0 or uge 0). | |||
2510 | ConstantRange RHS_CR = ConstantRange::makeExactICmpRegion(Pred, *C); | |||
2511 | if (RHS_CR.isEmptySet()) | |||
2512 | return ConstantInt::getFalse(ITy); | |||
2513 | if (RHS_CR.isFullSet()) | |||
2514 | return ConstantInt::getTrue(ITy); | |||
2515 | ||||
2516 | // Find the range of possible values for binary operators. | |||
2517 | unsigned Width = C->getBitWidth(); | |||
2518 | APInt Lower = APInt(Width, 0); | |||
2519 | APInt Upper = APInt(Width, 0); | |||
2520 | if (auto *BO = dyn_cast<BinaryOperator>(LHS)) | |||
2521 | setLimitsForBinOp(*BO, Lower, Upper); | |||
2522 | ||||
2523 | ConstantRange LHS_CR = | |||
2524 | Lower != Upper ? ConstantRange(Lower, Upper) : ConstantRange(Width, true); | |||
2525 | ||||
2526 | if (auto *I = dyn_cast<Instruction>(LHS)) | |||
2527 | if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) | |||
2528 | LHS_CR = LHS_CR.intersectWith(getConstantRangeFromMetadata(*Ranges)); | |||
2529 | ||||
2530 | if (!LHS_CR.isFullSet()) { | |||
2531 | if (RHS_CR.contains(LHS_CR)) | |||
2532 | return ConstantInt::getTrue(ITy); | |||
2533 | if (RHS_CR.inverse().contains(LHS_CR)) | |||
2534 | return ConstantInt::getFalse(ITy); | |||
2535 | } | |||
2536 | ||||
2537 | return nullptr; | |||
2538 | } | |||
2539 | ||||
2540 | /// TODO: A large part of this logic is duplicated in InstCombine's | |||
2541 | /// foldICmpBinOp(). We should be able to share that and avoid the code | |||
2542 | /// duplication. | |||
2543 | static Value *simplifyICmpWithBinOp(CmpInst::Predicate Pred, Value *LHS, | |||
2544 | Value *RHS, const SimplifyQuery &Q, | |||
2545 | unsigned MaxRecurse) { | |||
2546 | Type *ITy = GetCompareTy(LHS); // The return type. | |||
2547 | ||||
2548 | BinaryOperator *LBO = dyn_cast<BinaryOperator>(LHS); | |||
2549 | BinaryOperator *RBO = dyn_cast<BinaryOperator>(RHS); | |||
2550 | if (MaxRecurse && (LBO || RBO)) { | |||
2551 | // Analyze the case when either LHS or RHS is an add instruction. | |||
2552 | Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; | |||
2553 | // LHS = A + B (or A and B are null); RHS = C + D (or C and D are null). | |||
2554 | bool NoLHSWrapProblem = false, NoRHSWrapProblem = false; | |||
2555 | if (LBO && LBO->getOpcode() == Instruction::Add) { | |||
2556 | A = LBO->getOperand(0); | |||
2557 | B = LBO->getOperand(1); | |||
2558 | NoLHSWrapProblem = | |||
2559 | ICmpInst::isEquality(Pred) || | |||
2560 | (CmpInst::isUnsigned(Pred) && LBO->hasNoUnsignedWrap()) || | |||
2561 | (CmpInst::isSigned(Pred) && LBO->hasNoSignedWrap()); | |||
2562 | } | |||
2563 | if (RBO && RBO->getOpcode() == Instruction::Add) { | |||
2564 | C = RBO->getOperand(0); | |||
2565 | D = RBO->getOperand(1); | |||
2566 | NoRHSWrapProblem = | |||
2567 | ICmpInst::isEquality(Pred) || | |||
2568 | (CmpInst::isUnsigned(Pred) && RBO->hasNoUnsignedWrap()) || | |||
2569 | (CmpInst::isSigned(Pred) && RBO->hasNoSignedWrap()); | |||
2570 | } | |||
2571 | ||||
2572 | // icmp (X+Y), X -> icmp Y, 0 for equalities or if there is no overflow. | |||
2573 | if ((A == RHS || B == RHS) && NoLHSWrapProblem) | |||
2574 | if (Value *V = SimplifyICmpInst(Pred, A == RHS ? B : A, | |||
2575 | Constant::getNullValue(RHS->getType()), Q, | |||
2576 | MaxRecurse - 1)) | |||
2577 | return V; | |||
2578 | ||||
2579 | // icmp X, (X+Y) -> icmp 0, Y for equalities or if there is no overflow. | |||
2580 | if ((C == LHS || D == LHS) && NoRHSWrapProblem) | |||
2581 | if (Value *V = | |||
2582 | SimplifyICmpInst(Pred, Constant::getNullValue(LHS->getType()), | |||
2583 | C == LHS ? D : C, Q, MaxRecurse - 1)) | |||
2584 | return V; | |||
2585 | ||||
2586 | // icmp (X+Y), (X+Z) -> icmp Y,Z for equalities or if there is no overflow. | |||
2587 | if (A && C && (A == C || A == D || B == C || B == D) && NoLHSWrapProblem && | |||
2588 | NoRHSWrapProblem) { | |||
2589 | // Determine Y and Z in the form icmp (X+Y), (X+Z). | |||
2590 | Value *Y, *Z; | |||
2591 | if (A == C) { | |||
2592 | // C + B == C + D -> B == D | |||
2593 | Y = B; | |||
2594 | Z = D; | |||
2595 | } else if (A == D) { | |||
2596 | // D + B == C + D -> B == C | |||
2597 | Y = B; | |||
2598 | Z = C; | |||
2599 | } else if (B == C) { | |||
2600 | // A + C == C + D -> A == D | |||
2601 | Y = A; | |||
2602 | Z = D; | |||
2603 | } else { | |||
2604 | assert(B == D)(static_cast <bool> (B == D) ? void (0) : __assert_fail ("B == D", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Analysis/InstructionSimplify.cpp" , 2604, __extension__ __PRETTY_FUNCTION__)); | |||
2605 | // A + D == C + D -> A == C | |||
2606 | Y = A; | |||
2607 | Z = C; | |||
2608 | } | |||
2609 | if (Value *V = SimplifyICmpInst(Pred, Y, Z, Q, MaxRecurse - 1)) | |||
2610 | return V; | |||
2611 | } | |||
2612 | } | |||
2613 | ||||
2614 | { | |||
2615 | Value *Y = nullptr; | |||
2616 | // icmp pred (or X, Y), X | |||
2617 | if (LBO && match(LBO, m_c_Or(m_Value(Y), m_Specific(RHS)))) { | |||
2618 | if (Pred == ICmpInst::ICMP_ULT) | |||
2619 | return getFalse(ITy); | |||
2620 | if (Pred == ICmpInst::ICMP_UGE) | |||
2621 | return getTrue(ITy); | |||
2622 | ||||
2623 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) { | |||
2624 | KnownBits RHSKnown = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2625 | KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2626 | if (RHSKnown.isNonNegative() && YKnown.isNegative()) | |||
2627 | return Pred == ICmpInst::ICMP_SLT ? getTrue(ITy) : getFalse(ITy); | |||
2628 | if (RHSKnown.isNegative() || YKnown.isNonNegative()) | |||
2629 | return Pred == ICmpInst::ICMP_SLT ? getFalse(ITy) : getTrue(ITy); | |||
2630 | } | |||
2631 | } | |||
2632 | // icmp pred X, (or X, Y) | |||
2633 | if (RBO && match(RBO, m_c_Or(m_Value(Y), m_Specific(LHS)))) { | |||
2634 | if (Pred == ICmpInst::ICMP_ULE) | |||
2635 | return getTrue(ITy); | |||
2636 | if (Pred == ICmpInst::ICMP_UGT) | |||
2637 | return getFalse(ITy); | |||
2638 | ||||
2639 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE) { | |||
2640 | KnownBits LHSKnown = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2641 | KnownBits YKnown = computeKnownBits(Y, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2642 | if (LHSKnown.isNonNegative() && YKnown.isNegative()) | |||
2643 | return Pred == ICmpInst::ICMP_SGT ? getTrue(ITy) : getFalse(ITy); | |||
2644 | if (LHSKnown.isNegative() || YKnown.isNonNegative()) | |||
2645 | return Pred == ICmpInst::ICMP_SGT ? getFalse(ITy) : getTrue(ITy); | |||
2646 | } | |||
2647 | } | |||
2648 | } | |||
2649 | ||||
2650 | // icmp pred (and X, Y), X | |||
2651 | if (LBO && match(LBO, m_c_And(m_Value(), m_Specific(RHS)))) { | |||
2652 | if (Pred == ICmpInst::ICMP_UGT) | |||
2653 | return getFalse(ITy); | |||
2654 | if (Pred == ICmpInst::ICMP_ULE) | |||
2655 | return getTrue(ITy); | |||
2656 | } | |||
2657 | // icmp pred X, (and X, Y) | |||
2658 | if (RBO && match(RBO, m_c_And(m_Value(), m_Specific(LHS)))) { | |||
2659 | if (Pred == ICmpInst::ICMP_UGE) | |||
2660 | return getTrue(ITy); | |||
2661 | if (Pred == ICmpInst::ICMP_ULT) | |||
2662 | return getFalse(ITy); | |||
2663 | } | |||
2664 | ||||
2665 | // 0 - (zext X) pred C | |||
2666 | if (!CmpInst::isUnsigned(Pred) && match(LHS, m_Neg(m_ZExt(m_Value())))) { | |||
2667 | if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) { | |||
2668 | if (RHSC->getValue().isStrictlyPositive()) { | |||
2669 | if (Pred == ICmpInst::ICMP_SLT) | |||
2670 | return ConstantInt::getTrue(RHSC->getContext()); | |||
2671 | if (Pred == ICmpInst::ICMP_SGE) | |||
2672 | return ConstantInt::getFalse(RHSC->getContext()); | |||
2673 | if (Pred == ICmpInst::ICMP_EQ) | |||
2674 | return ConstantInt::getFalse(RHSC->getContext()); | |||
2675 | if (Pred == ICmpInst::ICMP_NE) | |||
2676 | return ConstantInt::getTrue(RHSC->getContext()); | |||
2677 | } | |||
2678 | if (RHSC->getValue().isNonNegative()) { | |||
2679 | if (Pred == ICmpInst::ICMP_SLE) | |||
2680 | return ConstantInt::getTrue(RHSC->getContext()); | |||
2681 | if (Pred == ICmpInst::ICMP_SGT) | |||
2682 | return ConstantInt::getFalse(RHSC->getContext()); | |||
2683 | } | |||
2684 | } | |||
2685 | } | |||
2686 | ||||
2687 | // icmp pred (urem X, Y), Y | |||
2688 | if (LBO && match(LBO, m_URem(m_Value(), m_Specific(RHS)))) { | |||
2689 | switch (Pred) { | |||
2690 | default: | |||
2691 | break; | |||
2692 | case ICmpInst::ICMP_SGT: | |||
2693 | case ICmpInst::ICMP_SGE: { | |||
2694 | KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2695 | if (!Known.isNonNegative()) | |||
2696 | break; | |||
2697 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
2698 | } | |||
2699 | case ICmpInst::ICMP_EQ: | |||
2700 | case ICmpInst::ICMP_UGT: | |||
2701 | case ICmpInst::ICMP_UGE: | |||
2702 | return getFalse(ITy); | |||
2703 | case ICmpInst::ICMP_SLT: | |||
2704 | case ICmpInst::ICMP_SLE: { | |||
2705 | KnownBits Known = computeKnownBits(RHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2706 | if (!Known.isNonNegative()) | |||
2707 | break; | |||
2708 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
2709 | } | |||
2710 | case ICmpInst::ICMP_NE: | |||
2711 | case ICmpInst::ICMP_ULT: | |||
2712 | case ICmpInst::ICMP_ULE: | |||
2713 | return getTrue(ITy); | |||
2714 | } | |||
2715 | } | |||
2716 | ||||
2717 | // icmp pred X, (urem Y, X) | |||
2718 | if (RBO && match(RBO, m_URem(m_Value(), m_Specific(LHS)))) { | |||
2719 | switch (Pred) { | |||
2720 | default: | |||
2721 | break; | |||
2722 | case ICmpInst::ICMP_SGT: | |||
2723 | case ICmpInst::ICMP_SGE: { | |||
2724 | KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2725 | if (!Known.isNonNegative()) | |||
2726 | break; | |||
2727 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
2728 | } | |||
2729 | case ICmpInst::ICMP_NE: | |||
2730 | case ICmpInst::ICMP_UGT: | |||
2731 | case ICmpInst::ICMP_UGE: | |||
2732 | return getTrue(ITy); | |||
2733 | case ICmpInst::ICMP_SLT: | |||
2734 | case ICmpInst::ICMP_SLE: { | |||
2735 | KnownBits Known = computeKnownBits(LHS, Q.DL, 0, Q.AC, Q.CxtI, Q.DT); | |||
2736 | if (!Known.isNonNegative()) | |||
2737 | break; | |||
2738 | LLVM_FALLTHROUGH[[clang::fallthrough]]; | |||
2739 | } | |||
2740 | case ICmpInst::ICMP_EQ: | |||
2741 | case ICmpInst::ICMP_ULT: | |||
2742 | case ICmpInst::ICMP_ULE: | |||
2743 | return getFalse(ITy); | |||
2744 | } | |||
2745 | } | |||
2746 | ||||
2747 | // x >> y <=u x | |||
2748 | // x udiv y <=u x. | |||
2749 | if (LBO && (match(LBO, m_LShr(m_Specific(RHS), m_Value())) || | |||
2750 | match(LBO, m_UDiv(m_Specific(RHS), m_Value())))) { | |||
2751 | // icmp pred (X op Y), X | |||
2752 | if (Pred == ICmpInst::ICMP_UGT) | |||
2753 | return getFalse(ITy); | |||
2754 | if (Pred == ICmpInst::ICMP_ULE) | |||
2755 | return getTrue(ITy); | |||
2756 | } | |||
2757 | ||||
2758 | // x >=u x >> y | |||
2759 | // x >=u x udiv y. | |||
2760 | if (RBO && (match(RBO, m_LShr(m_Specific(LHS), m_Value())) || | |||
2761 | match(RBO, m_UDiv(m_Specific(LHS), m_Value())))) { | |||
2762 | // icmp pred X, (X op Y) | |||
2763 | if (Pred == ICmpInst::ICMP_ULT) | |||
2764 | return getFalse(ITy); | |||
2765 | if (Pred == ICmpInst::ICMP_UGE) | |||
2766 | return getTrue(ITy); | |||
2767 | } | |||
2768 | ||||
2769 | // handle: | |||
2770 | // CI2 << X == CI | |||
2771 | // CI2 << X != CI | |||
2772 | // | |||
2773 | // where CI2 is a power of 2 and CI isn't | |||
2774 | if (auto *CI = dyn_cast<ConstantInt>(RHS)) { | |||
2775 | const APInt *CI2Val, *CIVal = &CI->getValue(); | |||
2776 | if (LBO && match(LBO, m_Shl(m_APInt(CI2Val), m_Value())) && | |||
2777 | CI2Val->isPowerOf2()) { | |||
2778 | if (!CIVal->isPowerOf2()) { | |||
2779 | // CI2 << X can equal zero in some circumstances, | |||
2780 | // this simplification is unsafe if CI is zero. | |||
2781 | // | |||
2782 | // We know it is safe if: | |||
2783 | // - The shift is nsw, we can't shift out the one bit. | |||
2784 | // - The shift is nuw, we can't shift out the one bit. | |||
2785 | // - CI2 is one | |||
2786 | // - CI isn't zero | |||
2787 | if (LBO->hasNoSignedWrap() || LBO->hasNoUnsignedWrap() || | |||
2788 | CI2Val->isOneValue() || !CI->isZero()) { | |||
2789 | if (Pred == ICmpInst::ICMP_EQ) | |||
2790 | return ConstantInt::getFalse(RHS->getContext()); | |||
2791 | if (Pred == ICmpInst::ICMP_NE) | |||
2792 | return ConstantInt::getTrue(RHS->getContext()); | |||
2793 | } | |||
2794 | } | |||
2795 | if (CIVal->isSignMask() && CI2Val->isOneValue()) { | |||
2796 | if (Pred == ICmpInst::ICMP_UGT) | |||
2797 | return ConstantInt::getFalse(RHS->getContext()); | |||
2798 | if (Pred == ICmpInst::ICMP_ULE) | |||
2799 | return ConstantInt::getTrue(RHS->getContext()); | |||
2800 | } | |||
2801 | } | |||
2802 | } | |||
2803 | ||||
2804 | if (MaxRecurse && LBO && RBO && LBO->getOpcode() == RBO->getOpcode() && | |||
2805 | LBO->getOperand(1) == RBO->getOperand(1)) { | |||
2806 | switch (LBO->getOpcode()) { | |||
2807 | default: | |||
2808 | break; | |||
2809 | case Instruction::UDiv: | |||
2810 | case Instruction::LShr: | |||
2811 | if (ICmpInst::isSigned(Pred) || !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::SDiv: | |||
2818 | if (!ICmpInst::isEquality(Pred) || !LBO->isExact() || !RBO->isExact()) | |||
2819 | break; | |||
2820 | if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), | |||
2821 | RBO->getOperand(0), Q, MaxRecurse - 1)) | |||
2822 | return V; | |||
2823 | break; | |||
2824 | case Instruction::AShr: | |||
2825 | if (!LBO->isExact() || !RBO->isExact()) | |||
2826 | break; | |||
2827 | if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), | |||
2828 | RBO->getOperand(0), Q, MaxRecurse - 1)) | |||
2829 | return V; | |||
2830 | break; | |||
2831 | case Instruction::Shl: { | |||
2832 | bool NUW = LBO->hasNoUnsignedWrap() && RBO->hasNoUnsignedWrap(); | |||
2833 | bool NSW = LBO->hasNoSignedWrap() && RBO->hasNoSignedWrap(); | |||
2834 | if (!NUW && !NSW) | |||
2835 | break; | |||
2836 | if (!NSW && ICmpInst::isSigned(Pred)) | |||
2837 | break; | |||
2838 | if (Value *V = SimplifyICmpInst(Pred, LBO->getOperand(0), | |||
2839 | RBO->getOperand(0), Q, MaxRecurse - 1)) | |||
2840 | return V; | |||
2841 | break; | |||
2842 | } | |||
2843 | } | |||
2844 | } | |||
2845 | return nullptr; | |||
2846 | } | |||
2847 | ||||
2848 | /// Simplify integer comparisons where at least one operand of the compare | |||
2849 | /// matches an integer min/max idiom. | |||
2850 | static Value *simplifyICmpWithMinMax(CmpInst::Predicate Pred, Value *LHS, | |||
2851 | Value *RHS, const SimplifyQuery &Q, | |||
2852 | unsigned MaxRecurse) { | |||
2853 | Type *ITy = GetCompareTy(LHS); // The return type. | |||
2854 | Value *A, *B; | |||
2855 | CmpInst::Predicate P = CmpInst::BAD_ICMP_PREDICATE; | |||
2856 | CmpInst::Predicate EqP; // Chosen so that "A == max/min(A,B)" iff "A EqP B". | |||
2857 | ||||
2858 | // Signed variants on "max(a,b)>=a -> true". | |||
2859 | if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { | |||
2860 | if (A != RHS) | |||
2861 | std::swap(A, B); // smax(A, B) pred A. | |||
2862 | EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". | |||
2863 | // We analyze this as smax(A, B) pred A. | |||
2864 | P = Pred; | |||
2865 | } else if (match(RHS, m_SMax(m_Value(A), m_Value(B))) && | |||
2866 | (A == LHS || B == LHS)) { | |||
2867 | if (A != LHS) | |||
2868 | std::swap(A, B); // A pred smax(A, B). | |||
2869 | EqP = CmpInst::ICMP_SGE; // "A == smax(A, B)" iff "A sge B". | |||
2870 | // We analyze this as smax(A, B) swapped-pred A. | |||
2871 | P = CmpInst::getSwappedPredicate(Pred); | |||
2872 | } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && | |||
2873 | (A == RHS || B == RHS)) { | |||
2874 | if (A != RHS) | |||
2875 | std::swap(A, B); // smin(A, B) pred A. | |||
2876 | EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". | |||
2877 | // We analyze this as smax(-A, -B) swapped-pred -A. | |||
2878 | // Note that we do not need to actually form -A or -B thanks to EqP. | |||
2879 | P = CmpInst::getSwappedPredicate(Pred); | |||
2880 | } else if (match(RHS, m_SMin(m_Value(A), m_Value(B))) && | |||
2881 | (A == LHS || B == LHS)) { | |||
2882 | if (A != LHS) | |||
2883 | std::swap(A, B); // A pred smin(A, B). | |||
2884 | EqP = CmpInst::ICMP_SLE; // "A == smin(A, B)" iff "A sle B". | |||
2885 | // We analyze this as smax(-A, -B) pred -A. | |||
2886 | // Note that we do not need to actually form -A or -B thanks to EqP. | |||
2887 | P = Pred; | |||
2888 | } | |||
2889 | if (P != CmpInst::BAD_ICMP_PREDICATE) { | |||
2890 | // Cases correspond to "max(A, B) p A". | |||
2891 | switch (P) { | |||
2892 | default: | |||
2893 | break; | |||
2894 | case CmpInst::ICMP_EQ: | |||
2895 | case CmpInst::ICMP_SLE: | |||
2896 | // Equivalent to "A EqP B". This may be the same as the condition tested | |||
2897 | // in the max/min; if so, we can just return that. | |||
2898 | if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) | |||
2899 | return V; | |||
2900 | if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) | |||
2901 | return V; | |||
2902 | // Otherwise, see if "A EqP B" simplifies. | |||
2903 | if (MaxRecurse) | |||
2904 | if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1)) | |||
2905 | return V; | |||
2906 | break; | |||
2907 | case CmpInst::ICMP_NE: | |||
2908 | case CmpInst::ICMP_SGT: { | |||
2909 | CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); | |||
2910 | // Equivalent to "A InvEqP B". This may be the same as the condition | |||
2911 | // tested in the max/min; if so, we can just return that. | |||
2912 | if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) | |||
2913 | return V; | |||
2914 | if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) | |||
2915 | return V; | |||
2916 | // Otherwise, see if "A InvEqP B" simplifies. | |||
2917 | if (MaxRecurse) | |||
2918 | if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1)) | |||
2919 | return V; | |||
2920 | break; | |||
2921 | } | |||
2922 | case CmpInst::ICMP_SGE: | |||
2923 | // Always true. | |||
2924 | return getTrue(ITy); | |||
2925 | case CmpInst::ICMP_SLT: | |||
2926 | // Always false. | |||
2927 | return getFalse(ITy); | |||
2928 | } | |||
2929 | } | |||
2930 | ||||
2931 | // Unsigned variants on "max(a,b)>=a -> true". | |||
2932 | P = CmpInst::BAD_ICMP_PREDICATE; | |||
2933 | if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && (A == RHS || B == RHS)) { | |||
2934 | if (A != RHS) | |||
2935 | std::swap(A, B); // umax(A, B) pred A. | |||
2936 | EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". | |||
2937 | // We analyze this as umax(A, B) pred A. | |||
2938 | P = Pred; | |||
2939 | } else if (match(RHS, m_UMax(m_Value(A), m_Value(B))) && | |||
2940 | (A == LHS || B == LHS)) { | |||
2941 | if (A != LHS) | |||
2942 | std::swap(A, B); // A pred umax(A, B). | |||
2943 | EqP = CmpInst::ICMP_UGE; // "A == umax(A, B)" iff "A uge B". | |||
2944 | // We analyze this as umax(A, B) swapped-pred A. | |||
2945 | P = CmpInst::getSwappedPredicate(Pred); | |||
2946 | } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && | |||
2947 | (A == RHS || B == RHS)) { | |||
2948 | if (A != RHS) | |||
2949 | std::swap(A, B); // umin(A, B) pred A. | |||
2950 | EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". | |||
2951 | // We analyze this as umax(-A, -B) swapped-pred -A. | |||
2952 | // Note that we do not need to actually form -A or -B thanks to EqP. | |||
2953 | P = CmpInst::getSwappedPredicate(Pred); | |||
2954 | } else if (match(RHS, m_UMin(m_Value(A), m_Value(B))) && | |||
2955 | (A == LHS || B == LHS)) { | |||
2956 | if (A != LHS) | |||
2957 | std::swap(A, B); // A pred umin(A, B). | |||
2958 | EqP = CmpInst::ICMP_ULE; // "A == umin(A, B)" iff "A ule B". | |||
2959 | // We analyze this as umax(-A, -B) pred -A. | |||
2960 | // Note that we do not need to actually form -A or -B thanks to EqP. | |||
2961 | P = Pred; | |||
2962 | } | |||
2963 | if (P != CmpInst::BAD_ICMP_PREDICATE) { | |||
2964 | // Cases correspond to "max(A, B) p A". | |||
2965 | switch (P) { | |||
2966 | default: | |||
2967 | break; | |||
2968 | case CmpInst::ICMP_EQ: | |||
2969 | case CmpInst::ICMP_ULE: | |||
2970 | // Equivalent to "A EqP B". This may be the same as the condition tested | |||
2971 | // in the max/min; if so, we can just return that. | |||
2972 | if (Value *V = ExtractEquivalentCondition(LHS, EqP, A, B)) | |||
2973 | return V; | |||
2974 | if (Value *V = ExtractEquivalentCondition(RHS, EqP, A, B)) | |||
2975 | return V; | |||
2976 | // Otherwise, see if "A EqP B" simplifies. | |||
2977 | if (MaxRecurse) | |||
2978 | if (Value *V = SimplifyICmpInst(EqP, A, B, Q, MaxRecurse - 1)) | |||
2979 | return V; | |||
2980 | break; | |||
2981 | case CmpInst::ICMP_NE: | |||
2982 | case CmpInst::ICMP_UGT: { | |||
2983 | CmpInst::Predicate InvEqP = CmpInst::getInversePredicate(EqP); | |||
2984 | // Equivalent to "A InvEqP B". This may be the same as the condition | |||
2985 | // tested in the max/min; if so, we can just return that. | |||
2986 | if (Value *V = ExtractEquivalentCondition(LHS, InvEqP, A, B)) | |||
2987 | return V; | |||
2988 | if (Value *V = ExtractEquivalentCondition(RHS, InvEqP, A, B)) | |||
2989 | return V; | |||
2990 | // Otherwise, see if "A InvEqP B" simplifies. | |||
2991 | if (MaxRecurse) | |||
2992 | if (Value *V = SimplifyICmpInst(InvEqP, A, B, Q, MaxRecurse - 1)) | |||
2993 | return V; | |||
2994 | break; | |||
2995 | } | |||
2996 | case CmpInst::ICMP_UGE: | |||
2997 | // Always true. | |||
2998 | return getTrue(ITy); | |||
2999 | case CmpInst::ICMP_ULT: | |||
3000 | // Always false. | |||
3001 | return getFalse(ITy); | |||
3002 | } | |||
3003 | } | |||
3004 | ||||
3005 | // Variants on "max(x,y) >= min(x,z)". | |||
3006 | Value *C, *D; | |||
3007 | if (match(LHS, m_SMax(m_Value(A), m_Value(B))) && | |||
3008 | match(RHS, m_SMin(m_Value(C), m_Value(D))) && | |||
3009 | (A == C || A == D || B == C || B == D)) { | |||
3010 | // max(x, ?) pred min(x, ?). | |||
3011 | if (Pred == CmpInst::ICMP_SGE) | |||
3012 | // Always true. | |||
3013 | return getTrue(ITy); | |||
3014 | if (Pred == CmpInst::ICMP_SLT) | |||
3015 | // Always false. | |||
3016 | return getFalse(ITy); | |||
3017 | } else if (match(LHS, m_SMin(m_Value(A), m_Value(B))) && | |||
3018 | match(RHS, m_SMax(m_Value(C), m_Value(D))) && | |||
3019 | (A == C || A == D || B == C || B == D)) { | |||
3020 | // min(x, ?) pred max(x, ?). | |||
3021 | if (Pred == CmpInst::ICMP_SLE) | |||
3022 | // Always true. | |||
3023 | return getTrue(ITy); | |||
3024 | if (Pred == CmpInst::ICMP_SGT) | |||
3025 | // Always false. | |||
3026 | return getFalse(ITy); | |||
3027 | } else if (match(LHS, m_UMax(m_Value(A), m_Value(B))) && | |||
3028 | match(RHS, m_UMin(m_Value(C), m_Value(D))) && | |||
3029 | (A == C || A == D || B == C || B == D)) { | |||
3030 | // max(x, ?) pred min(x, ?). | |||
3031 | if (Pred == CmpInst::ICMP_UGE) | |||
3032 | // Always true. | |||
3033 | return getTrue(ITy); | |||
3034 | if (Pred == CmpInst::ICMP_ULT) | |||
3035 | // Always false. | |||
3036 | return getFalse(ITy); | |||
3037 | } else if (match(LHS, m_UMin(m_Value(A), m_Value(B))) && | |||
3038 | match(RHS, m_UMax(m_Value(C), m_Value(D))) && | |||
3039 | (A == C || A == D || B == C || B == D)) { | |||
3040 | // min(x, ?) pred max(x, ?). | |||
3041 | if (Pred == CmpInst::ICMP_ULE) | |||
3042 | // Always true. | |||
3043 | return getTrue(ITy); | |||
3044 | if (Pred == CmpInst::ICMP_UGT) | |||
3045 | // Always false. | |||
3046 | return getFalse(ITy); | |||
3047 | } | |||
3048 | ||||
3049 | return nullptr; | |||
3050 | } | |||
3051 | ||||
3052 | /// Given operands for an ICmpInst, see if we can fold the result. | |||
3053 | /// If not, this returns null. | |||
3054 | static Value *SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, | |||
3055 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
3056 | CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; | |||
3057 | 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~svn329677/lib/Analysis/InstructionSimplify.cpp" , 3057, __extension__ __PRETTY_FUNCTION__)); | |||
3058 | ||||
3059 | if (Constant *CLHS = dyn_cast<Constant>(LHS)) { | |||
3060 | if (Constant *CRHS = dyn_cast<Constant>(RHS)) | |||
3061 | return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); | |||
3062 | ||||
3063 | // If we have a constant, make sure it is on the RHS. | |||
3064 | std::swap(LHS, RHS); | |||
3065 | Pred = CmpInst::getSwappedPredicate(Pred); | |||
3066 | } | |||
3067 | ||||
3068 | Type *ITy = GetCompareTy(LHS); // The return type. | |||
3069 | ||||
3070 | // icmp X, X -> true/false | |||
3071 | // X icmp undef -> true/false. For example, icmp ugt %X, undef -> false | |||
3072 | // because X could be 0. | |||
3073 | if (LHS == RHS || isa<UndefValue>(RHS)) | |||
3074 | return ConstantInt::get(ITy, CmpInst::isTrueWhenEqual(Pred)); | |||
3075 | ||||
3076 | if (Value *V = simplifyICmpOfBools(Pred, LHS, RHS, Q)) | |||
3077 | return V; | |||
3078 | ||||
3079 | if (Value *V = simplifyICmpWithZero(Pred, LHS, RHS, Q)) | |||
3080 | return V; | |||
3081 | ||||
3082 | if (Value *V = simplifyICmpWithConstant(Pred, LHS, RHS)) | |||
3083 | return V; | |||
3084 | ||||
3085 | // If both operands have range metadata, use the metadata | |||
3086 | // to simplify the comparison. | |||
3087 | if (isa<Instruction>(RHS) && isa<Instruction>(LHS)) { | |||
3088 | auto RHS_Instr = cast<Instruction>(RHS); | |||
3089 | auto LHS_Instr = cast<Instruction>(LHS); | |||
3090 | ||||
3091 | if (RHS_Instr->getMetadata(LLVMContext::MD_range) && | |||
3092 | LHS_Instr->getMetadata(LLVMContext::MD_range)) { | |||
3093 | auto RHS_CR = getConstantRangeFromMetadata( | |||
3094 | *RHS_Instr->getMetadata(LLVMContext::MD_range)); | |||
3095 | auto LHS_CR = getConstantRangeFromMetadata( | |||
3096 | *LHS_Instr->getMetadata(LLVMContext::MD_range)); | |||
3097 | ||||
3098 | auto Satisfied_CR = ConstantRange::makeSatisfyingICmpRegion(Pred, RHS_CR); | |||
3099 | if (Satisfied_CR.contains(LHS_CR)) | |||
3100 | return ConstantInt::getTrue(RHS->getContext()); | |||
3101 | ||||
3102 | auto InversedSatisfied_CR = ConstantRange::makeSatisfyingICmpRegion( | |||
3103 | CmpInst::getInversePredicate(Pred), RHS_CR); | |||
3104 | if (InversedSatisfied_CR.contains(LHS_CR)) | |||
3105 | return ConstantInt::getFalse(RHS->getContext()); | |||
3106 | } | |||
3107 | } | |||
3108 | ||||
3109 | // Compare of cast, for example (zext X) != 0 -> X != 0 | |||
3110 | if (isa<CastInst>(LHS) && (isa<Constant>(RHS) || isa<CastInst>(RHS))) { | |||
3111 | Instruction *LI = cast<CastInst>(LHS); | |||
3112 | Value *SrcOp = LI->getOperand(0); | |||
3113 | Type *SrcTy = SrcOp->getType(); | |||
3114 | Type *DstTy = LI->getType(); | |||
3115 | ||||
3116 | // Turn icmp (ptrtoint x), (ptrtoint/constant) into a compare of the input | |||
3117 | // if the integer type is the same size as the pointer type. | |||
3118 | if (MaxRecurse && isa<PtrToIntInst>(LI) && | |||
3119 | Q.DL.getTypeSizeInBits(SrcTy) == DstTy->getPrimitiveSizeInBits()) { | |||
3120 | if (Constant *RHSC = dyn_cast<Constant>(RHS)) { | |||
3121 | // Transfer the cast to the constant. | |||
3122 | if (Value *V = SimplifyICmpInst(Pred, SrcOp, | |||
3123 | ConstantExpr::getIntToPtr(RHSC, SrcTy), | |||
3124 | Q, MaxRecurse-1)) | |||
3125 | return V; | |||
3126 | } else if (PtrToIntInst *RI = dyn_cast<PtrToIntInst>(RHS)) { | |||
3127 | if (RI->getOperand(0)->getType() == SrcTy) | |||
3128 | // Compare without the cast. | |||
3129 | if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), | |||
3130 | Q, MaxRecurse-1)) | |||
3131 | return V; | |||
3132 | } | |||
3133 | } | |||
3134 | ||||
3135 | if (isa<ZExtInst>(LHS)) { | |||
3136 | // Turn icmp (zext X), (zext Y) into a compare of X and Y if they have the | |||
3137 | // same type. | |||
3138 | if (ZExtInst *RI = dyn_cast<ZExtInst>(RHS)) { | |||
3139 | if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) | |||
3140 | // Compare X and Y. Note that signed predicates become unsigned. | |||
3141 | if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), | |||
3142 | SrcOp, RI->getOperand(0), Q, | |||
3143 | MaxRecurse-1)) | |||
3144 | return V; | |||
3145 | } | |||
3146 | // Turn icmp (zext X), Cst into a compare of X and Cst if Cst is extended | |||
3147 | // too. If not, then try to deduce the result of the comparison. | |||
3148 | else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { | |||
3149 | // Compute the constant that would happen if we truncated to SrcTy then | |||
3150 | // reextended to DstTy. | |||
3151 | Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); | |||
3152 | Constant *RExt = ConstantExpr::getCast(CastInst::ZExt, Trunc, DstTy); | |||
3153 | ||||
3154 | // If the re-extended constant didn't change then this is effectively | |||
3155 | // also a case of comparing two zero-extended values. | |||
3156 | if (RExt == CI && MaxRecurse) | |||
3157 | if (Value *V = SimplifyICmpInst(ICmpInst::getUnsignedPredicate(Pred), | |||
3158 | SrcOp, Trunc, Q, MaxRecurse-1)) | |||
3159 | return V; | |||
3160 | ||||
3161 | // Otherwise the upper bits of LHS are zero while RHS has a non-zero bit | |||
3162 | // there. Use this to work out the result of the comparison. | |||
3163 | if (RExt != CI) { | |||
3164 | switch (Pred) { | |||
3165 | default: llvm_unreachable("Unknown ICmp predicate!")::llvm::llvm_unreachable_internal("Unknown ICmp predicate!", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Analysis/InstructionSimplify.cpp" , 3165); | |||
3166 | // LHS <u RHS. | |||
3167 | case ICmpInst::ICMP_EQ: | |||
3168 | case ICmpInst::ICMP_UGT: | |||
3169 | case ICmpInst::ICMP_UGE: | |||
3170 | return ConstantInt::getFalse(CI->getContext()); | |||
3171 | ||||
3172 | case ICmpInst::ICMP_NE: | |||
3173 | case ICmpInst::ICMP_ULT: | |||
3174 | case ICmpInst::ICMP_ULE: | |||
3175 | return ConstantInt::getTrue(CI->getContext()); | |||
3176 | ||||
3177 | // LHS is non-negative. If RHS is negative then LHS >s LHS. If RHS | |||
3178 | // is non-negative then LHS <s RHS. | |||
3179 | case ICmpInst::ICMP_SGT: | |||
3180 | case ICmpInst::ICMP_SGE: | |||
3181 | return CI->getValue().isNegative() ? | |||
3182 | ConstantInt::getTrue(CI->getContext()) : | |||
3183 | ConstantInt::getFalse(CI->getContext()); | |||
3184 | ||||
3185 | case ICmpInst::ICMP_SLT: | |||
3186 | case ICmpInst::ICMP_SLE: | |||
3187 | return CI->getValue().isNegative() ? | |||
3188 | ConstantInt::getFalse(CI->getContext()) : | |||
3189 | ConstantInt::getTrue(CI->getContext()); | |||
3190 | } | |||
3191 | } | |||
3192 | } | |||
3193 | } | |||
3194 | ||||
3195 | if (isa<SExtInst>(LHS)) { | |||
3196 | // Turn icmp (sext X), (sext Y) into a compare of X and Y if they have the | |||
3197 | // same type. | |||
3198 | if (SExtInst *RI = dyn_cast<SExtInst>(RHS)) { | |||
3199 | if (MaxRecurse && SrcTy == RI->getOperand(0)->getType()) | |||
3200 | // Compare X and Y. Note that the predicate does not change. | |||
3201 | if (Value *V = SimplifyICmpInst(Pred, SrcOp, RI->getOperand(0), | |||
3202 | Q, MaxRecurse-1)) | |||
3203 | return V; | |||
3204 | } | |||
3205 | // Turn icmp (sext X), Cst into a compare of X and Cst if Cst is extended | |||
3206 | // too. If not, then try to deduce the result of the comparison. | |||
3207 | else if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) { | |||
3208 | // Compute the constant that would happen if we truncated to SrcTy then | |||
3209 | // reextended to DstTy. | |||
3210 | Constant *Trunc = ConstantExpr::getTrunc(CI, SrcTy); | |||
3211 | Constant *RExt = ConstantExpr::getCast(CastInst::SExt, Trunc, DstTy); | |||
3212 | ||||
3213 | // If the re-extended constant didn't change then this is effectively | |||
3214 | // also a case of comparing two sign-extended values. | |||
3215 | if (RExt == CI && MaxRecurse) | |||
3216 | if (Value *V = SimplifyICmpInst(Pred, SrcOp, Trunc, Q, MaxRecurse-1)) | |||
3217 | return V; | |||
3218 | ||||
3219 | // Otherwise the upper bits of LHS are all equal, while RHS has varying | |||
3220 | // bits there. Use this to work out the result of the comparison. | |||
3221 | if (RExt != CI) { | |||
3222 | switch (Pred) { | |||
3223 | default: llvm_unreachable("Unknown ICmp predicate!")::llvm::llvm_unreachable_internal("Unknown ICmp predicate!", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Analysis/InstructionSimplify.cpp" , 3223); | |||
3224 | case ICmpInst::ICMP_EQ: | |||
3225 | return ConstantInt::getFalse(CI->getContext()); | |||
3226 | case ICmpInst::ICMP_NE: | |||
3227 | return ConstantInt::getTrue(CI->getContext()); | |||
3228 | ||||
3229 | // If RHS is non-negative then LHS <s RHS. If RHS is negative then | |||
3230 | // LHS >s RHS. | |||
3231 | case ICmpInst::ICMP_SGT: | |||
3232 | case ICmpInst::ICMP_SGE: | |||
3233 | return CI->getValue().isNegative() ? | |||
3234 | ConstantInt::getTrue(CI->getContext()) : | |||
3235 | ConstantInt::getFalse(CI->getContext()); | |||
3236 | case ICmpInst::ICMP_SLT: | |||
3237 | case ICmpInst::ICMP_SLE: | |||
3238 | return CI->getValue().isNegative() ? | |||
3239 | ConstantInt::getFalse(CI->getContext()) : | |||
3240 | ConstantInt::getTrue(CI->getContext()); | |||
3241 | ||||
3242 | // If LHS is non-negative then LHS <u RHS. If LHS is negative then | |||
3243 | // LHS >u RHS. | |||
3244 | case ICmpInst::ICMP_UGT: | |||
3245 | case ICmpInst::ICMP_UGE: | |||
3246 | // Comparison is true iff the LHS <s 0. | |||
3247 | if (MaxRecurse) | |||
3248 | if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SLT, SrcOp, | |||
3249 | Constant::getNullValue(SrcTy), | |||
3250 | Q, MaxRecurse-1)) | |||
3251 | return V; | |||
3252 | break; | |||
3253 | case ICmpInst::ICMP_ULT: | |||
3254 | case ICmpInst::ICMP_ULE: | |||
3255 | // Comparison is true iff the LHS >=s 0. | |||
3256 | if (MaxRecurse) | |||
3257 | if (Value *V = SimplifyICmpInst(ICmpInst::ICMP_SGE, SrcOp, | |||
3258 | Constant::getNullValue(SrcTy), | |||
3259 | Q, MaxRecurse-1)) | |||
3260 | return V; | |||
3261 | break; | |||
3262 | } | |||
3263 | } | |||
3264 | } | |||
3265 | } | |||
3266 | } | |||
3267 | ||||
3268 | // icmp eq|ne X, Y -> false|true if X != Y | |||
3269 | if (ICmpInst::isEquality(Pred) && | |||
3270 | isKnownNonEqual(LHS, RHS, Q.DL, Q.AC, Q.CxtI, Q.DT)) { | |||
3271 | return Pred == ICmpInst::ICMP_NE ? getTrue(ITy) : getFalse(ITy); | |||
3272 | } | |||
3273 | ||||
3274 | if (Value *V = simplifyICmpWithBinOp(Pred, LHS, RHS, Q, MaxRecurse)) | |||
3275 | return V; | |||
3276 | ||||
3277 | if (Value *V = simplifyICmpWithMinMax(Pred, LHS, RHS, Q, MaxRecurse)) | |||
3278 | return V; | |||
3279 | ||||
3280 | // Simplify comparisons of related pointers using a powerful, recursive | |||
3281 | // GEP-walk when we have target data available.. | |||
3282 | if (LHS->getType()->isPointerTy()) | |||
3283 | if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, LHS, | |||
3284 | RHS)) | |||
3285 | return C; | |||
3286 | if (auto *CLHS = dyn_cast<PtrToIntOperator>(LHS)) | |||
3287 | if (auto *CRHS = dyn_cast<PtrToIntOperator>(RHS)) | |||
3288 | if (Q.DL.getTypeSizeInBits(CLHS->getPointerOperandType()) == | |||
3289 | Q.DL.getTypeSizeInBits(CLHS->getType()) && | |||
3290 | Q.DL.getTypeSizeInBits(CRHS->getPointerOperandType()) == | |||
3291 | Q.DL.getTypeSizeInBits(CRHS->getType())) | |||
3292 | if (auto *C = computePointerICmp(Q.DL, Q.TLI, Q.DT, Pred, Q.AC, Q.CxtI, | |||
3293 | CLHS->getPointerOperand(), | |||
3294 | CRHS->getPointerOperand())) | |||
3295 | return C; | |||
3296 | ||||
3297 | if (GetElementPtrInst *GLHS = dyn_cast<GetElementPtrInst>(LHS)) { | |||
3298 | if (GEPOperator *GRHS = dyn_cast<GEPOperator>(RHS)) { | |||
3299 | if (GLHS->getPointerOperand() == GRHS->getPointerOperand() && | |||
3300 | GLHS->hasAllConstantIndices() && GRHS->hasAllConstantIndices() && | |||
3301 | (ICmpInst::isEquality(Pred) || | |||
3302 | (GLHS->isInBounds() && GRHS->isInBounds() && | |||
3303 | Pred == ICmpInst::getSignedPredicate(Pred)))) { | |||
3304 | // The bases are equal and the indices are constant. Build a constant | |||
3305 | // expression GEP with the same indices and a null base pointer to see | |||
3306 | // what constant folding can make out of it. | |||
3307 | Constant *Null = Constant::getNullValue(GLHS->getPointerOperandType()); | |||
3308 | SmallVector<Value *, 4> IndicesLHS(GLHS->idx_begin(), GLHS->idx_end()); | |||
3309 | Constant *NewLHS = ConstantExpr::getGetElementPtr( | |||
3310 | GLHS->getSourceElementType(), Null, IndicesLHS); | |||
3311 | ||||
3312 | SmallVector<Value *, 4> IndicesRHS(GRHS->idx_begin(), GRHS->idx_end()); | |||
3313 | Constant *NewRHS = ConstantExpr::getGetElementPtr( | |||
3314 | GLHS->getSourceElementType(), Null, IndicesRHS); | |||
3315 | return ConstantExpr::getICmp(Pred, NewLHS, NewRHS); | |||
3316 | } | |||
3317 | } | |||
3318 | } | |||
3319 | ||||
3320 | // If the comparison is with the result of a select instruction, check whether | |||
3321 | // comparing with either branch of the select always yields the same value. | |||
3322 | if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) | |||
3323 | if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) | |||
3324 | return V; | |||
3325 | ||||
3326 | // If the comparison is with the result of a phi instruction, check whether | |||
3327 | // doing the compare with each incoming phi value yields a common result. | |||
3328 | if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) | |||
3329 | if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) | |||
3330 | return V; | |||
3331 | ||||
3332 | return nullptr; | |||
3333 | } | |||
3334 | ||||
3335 | Value *llvm::SimplifyICmpInst(unsigned Predicate, Value *LHS, Value *RHS, | |||
3336 | const SimplifyQuery &Q) { | |||
3337 | return ::SimplifyICmpInst(Predicate, LHS, RHS, Q, RecursionLimit); | |||
3338 | } | |||
3339 | ||||
3340 | /// Given operands for an FCmpInst, see if we can fold the result. | |||
3341 | /// If not, this returns null. | |||
3342 | static Value *SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, | |||
3343 | FastMathFlags FMF, const SimplifyQuery &Q, | |||
3344 | unsigned MaxRecurse) { | |||
3345 | CmpInst::Predicate Pred = (CmpInst::Predicate)Predicate; | |||
3346 | 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~svn329677/lib/Analysis/InstructionSimplify.cpp" , 3346, __extension__ __PRETTY_FUNCTION__)); | |||
3347 | ||||
3348 | if (Constant *CLHS = dyn_cast<Constant>(LHS)) { | |||
3349 | if (Constant *CRHS = dyn_cast<Constant>(RHS)) | |||
3350 | return ConstantFoldCompareInstOperands(Pred, CLHS, CRHS, Q.DL, Q.TLI); | |||
3351 | ||||
3352 | // If we have a constant, make sure it is on the RHS. | |||
3353 | std::swap(LHS, RHS); | |||
3354 | Pred = CmpInst::getSwappedPredicate(Pred); | |||
3355 | } | |||
3356 | ||||
3357 | // Fold trivial predicates. | |||
3358 | Type *RetTy = GetCompareTy(LHS); | |||
3359 | if (Pred == FCmpInst::FCMP_FALSE) | |||
3360 | return getFalse(RetTy); | |||
3361 | if (Pred == FCmpInst::FCMP_TRUE) | |||
3362 | return getTrue(RetTy); | |||
3363 | ||||
3364 | // UNO/ORD predicates can be trivially folded if NaNs are ignored. | |||
3365 | if (FMF.noNaNs()) { | |||
3366 | if (Pred == FCmpInst::FCMP_UNO) | |||
3367 | return getFalse(RetTy); | |||
3368 | if (Pred == FCmpInst::FCMP_ORD) | |||
3369 | return getTrue(RetTy); | |||
3370 | } | |||
3371 | ||||
3372 | // NaN is unordered; NaN is not ordered. | |||
3373 | assert((FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) &&(static_cast <bool> ((FCmpInst::isOrdered(Pred) || FCmpInst ::isUnordered(Pred)) && "Comparison must be either ordered or unordered" ) ? void (0) : __assert_fail ("(FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) && \"Comparison must be either ordered or unordered\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Analysis/InstructionSimplify.cpp" , 3374, __extension__ __PRETTY_FUNCTION__)) | |||
3374 | "Comparison must be either ordered or unordered")(static_cast <bool> ((FCmpInst::isOrdered(Pred) || FCmpInst ::isUnordered(Pred)) && "Comparison must be either ordered or unordered" ) ? void (0) : __assert_fail ("(FCmpInst::isOrdered(Pred) || FCmpInst::isUnordered(Pred)) && \"Comparison must be either ordered or unordered\"" , "/build/llvm-toolchain-snapshot-7~svn329677/lib/Analysis/InstructionSimplify.cpp" , 3374, __extension__ __PRETTY_FUNCTION__)); | |||
3375 | if (match(RHS, m_NaN())) | |||
3376 | return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred)); | |||
3377 | ||||
3378 | // fcmp pred x, undef and fcmp pred undef, x | |||
3379 | // fold to true if unordered, false if ordered | |||
3380 | if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) { | |||
3381 | // Choosing NaN for the undef will always make unordered comparison succeed | |||
3382 | // and ordered comparison fail. | |||
3383 | return ConstantInt::get(RetTy, CmpInst::isUnordered(Pred)); | |||
3384 | } | |||
3385 | ||||
3386 | // fcmp x,x -> true/false. Not all compares are foldable. | |||
3387 | if (LHS == RHS) { | |||
3388 | if (CmpInst::isTrueWhenEqual(Pred)) | |||
3389 | return getTrue(RetTy); | |||
3390 | if (CmpInst::isFalseWhenEqual(Pred)) | |||
3391 | return getFalse(RetTy); | |||
3392 | } | |||
3393 | ||||
3394 | // Handle fcmp with constant RHS. | |||
3395 | const APFloat *C; | |||
3396 | if (match(RHS, m_APFloat(C))) { | |||
3397 | // Check whether the constant is an infinity. | |||
3398 | if (C->isInfinity()) { | |||
3399 | if (C->isNegative()) { | |||
3400 | switch (Pred) { | |||
3401 | case FCmpInst::FCMP_OLT: | |||
3402 | // No value is ordered and less than negative infinity. | |||
3403 | return getFalse(RetTy); | |||
3404 | case FCmpInst::FCMP_UGE: | |||
3405 | // All values are unordered with or at least negative infinity. | |||
3406 | return getTrue(RetTy); | |||
3407 | default: | |||
3408 | break; | |||
3409 | } | |||
3410 | } else { | |||
3411 | switch (Pred) { | |||
3412 | case FCmpInst::FCMP_OGT: | |||
3413 | // No value is ordered and greater than infinity. | |||
3414 | return getFalse(RetTy); | |||
3415 | case FCmpInst::FCMP_ULE: | |||
3416 | // All values are unordered with and at most infinity. | |||
3417 | return getTrue(RetTy); | |||
3418 | default: | |||
3419 | break; | |||
3420 | } | |||
3421 | } | |||
3422 | } | |||
3423 | if (C->isZero()) { | |||
3424 | switch (Pred) { | |||
3425 | case FCmpInst::FCMP_UGE: | |||
3426 | if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) | |||
3427 | return getTrue(RetTy); | |||
3428 | break; | |||
3429 | case FCmpInst::FCMP_OLT: | |||
3430 | // X < 0 | |||
3431 | if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) | |||
3432 | return getFalse(RetTy); | |||
3433 | break; | |||
3434 | default: | |||
3435 | break; | |||
3436 | } | |||
3437 | } else if (C->isNegative()) { | |||
3438 | 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~svn329677/lib/Analysis/InstructionSimplify.cpp" , 3438, __extension__ __PRETTY_FUNCTION__)); | |||
3439 | // TODO: We can catch more cases by using a range check rather than | |||
3440 | // relying on CannotBeOrderedLessThanZero. | |||
3441 | switch (Pred) { | |||
3442 | case FCmpInst::FCMP_UGE: | |||
3443 | case FCmpInst::FCMP_UGT: | |||
3444 | case FCmpInst::FCMP_UNE: | |||
3445 | // (X >= 0) implies (X > C) when (C < 0) | |||
3446 | if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) | |||
3447 | return getTrue(RetTy); | |||
3448 | break; | |||
3449 | case FCmpInst::FCMP_OEQ: | |||
3450 | case FCmpInst::FCMP_OLE: | |||
3451 | case FCmpInst::FCMP_OLT: | |||
3452 | // (X >= 0) implies !(X < C) when (C < 0) | |||
3453 | if (CannotBeOrderedLessThanZero(LHS, Q.TLI)) | |||
3454 | return getFalse(RetTy); | |||
3455 | break; | |||
3456 | default: | |||
3457 | break; | |||
3458 | } | |||
3459 | } | |||
3460 | } | |||
3461 | ||||
3462 | // If the comparison is with the result of a select instruction, check whether | |||
3463 | // comparing with either branch of the select always yields the same value. | |||
3464 | if (isa<SelectInst>(LHS) || isa<SelectInst>(RHS)) | |||
3465 | if (Value *V = ThreadCmpOverSelect(Pred, LHS, RHS, Q, MaxRecurse)) | |||
3466 | return V; | |||
3467 | ||||
3468 | // If the comparison is with the result of a phi instruction, check whether | |||
3469 | // doing the compare with each incoming phi value yields a common result. | |||
3470 | if (isa<PHINode>(LHS) || isa<PHINode>(RHS)) | |||
3471 | if (Value *V = ThreadCmpOverPHI(Pred, LHS, RHS, Q, MaxRecurse)) | |||
3472 | return V; | |||
3473 | ||||
3474 | return nullptr; | |||
3475 | } | |||
3476 | ||||
3477 | Value *llvm::SimplifyFCmpInst(unsigned Predicate, Value *LHS, Value *RHS, | |||
3478 | FastMathFlags FMF, const SimplifyQuery &Q) { | |||
3479 | return ::SimplifyFCmpInst(Predicate, LHS, RHS, FMF, Q, RecursionLimit); | |||
3480 | } | |||
3481 | ||||
3482 | /// See if V simplifies when its operand Op is replaced with RepOp. | |||
3483 | static const Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp, | |||
3484 | const SimplifyQuery &Q, | |||
3485 | unsigned MaxRecurse) { | |||
3486 | // Trivial replacement. | |||
3487 | if (V == Op) | |||
3488 | return RepOp; | |||
3489 | ||||
3490 | // We cannot replace a constant, and shouldn't even try. | |||
3491 | if (isa<Constant>(Op)) | |||
3492 | return nullptr; | |||
3493 | ||||
3494 | auto *I = dyn_cast<Instruction>(V); | |||
3495 | if (!I) | |||
3496 | return nullptr; | |||
3497 | ||||
3498 | // If this is a binary operator, try to simplify it with the replaced op. | |||
3499 | if (auto *B = dyn_cast<BinaryOperator>(I)) { | |||
3500 | // Consider: | |||
3501 | // %cmp = icmp eq i32 %x, 2147483647 | |||
3502 | // %add = add nsw i32 %x, 1 | |||
3503 | // %sel = select i1 %cmp, i32 -2147483648, i32 %add | |||
3504 | // | |||
3505 | // We can't replace %sel with %add unless we strip away the flags. | |||
3506 | if (isa<OverflowingBinaryOperator>(B)) | |||
3507 | if (B->hasNoSignedWrap() || B->hasNoUnsignedWrap()) | |||
3508 | return nullptr; | |||
3509 | if (isa<PossiblyExactOperator>(B)) | |||
3510 | if (B->isExact()) | |||
3511 | return nullptr; | |||
3512 | ||||
3513 | if (MaxRecurse) { | |||
3514 | if (B->getOperand(0) == Op) | |||
3515 | return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), Q, | |||
3516 | MaxRecurse - 1); | |||
3517 | if (B->getOperand(1) == Op) | |||
3518 | return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, Q, | |||
3519 | MaxRecurse - 1); | |||
3520 | } | |||
3521 | } | |||
3522 | ||||
3523 | // Same for CmpInsts. | |||
3524 | if (CmpInst *C = dyn_cast<CmpInst>(I)) { | |||
3525 | if (MaxRecurse) { | |||
3526 | if (C->getOperand(0) == Op) | |||
3527 | return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), Q, | |||
3528 | MaxRecurse - 1); | |||
3529 | if (C->getOperand(1) == Op) | |||
3530 | return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, Q, | |||
3531 | MaxRecurse - 1); | |||
3532 | } | |||
3533 | } | |||
3534 | ||||
3535 | // TODO: We could hand off more cases to instsimplify here. | |||
3536 | ||||
3537 | // If all operands are constant after substituting Op for RepOp then we can | |||
3538 | // constant fold the instruction. | |||
3539 | if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) { | |||
3540 | // Build a list of all constant operands. | |||
3541 | SmallVector<Constant *, 8> ConstOps; | |||
3542 | for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { | |||
3543 | if (I->getOperand(i) == Op) | |||
3544 | ConstOps.push_back(CRepOp); | |||
3545 | else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i))) | |||
3546 | ConstOps.push_back(COp); | |||
3547 | else | |||
3548 | break; | |||
3549 | } | |||
3550 | ||||
3551 | // All operands were constants, fold it. | |||
3552 | if (ConstOps.size() == I->getNumOperands()) { | |||
3553 | if (CmpInst *C = dyn_cast<CmpInst>(I)) | |||
3554 | return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0], | |||
3555 | ConstOps[1], Q.DL, Q.TLI); | |||
3556 | ||||
3557 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | |||
3558 | if (!LI->isVolatile()) | |||
3559 | return ConstantFoldLoadFromConstPtr(ConstOps[0], LI->getType(), Q.DL); | |||
3560 | ||||
3561 | return ConstantFoldInstOperands(I, ConstOps, Q.DL, Q.TLI); | |||
3562 | } | |||
3563 | } | |||
3564 | ||||
3565 | return nullptr; | |||
3566 | } | |||
3567 | ||||
3568 | /// Try to simplify a select instruction when its condition operand is an | |||
3569 | /// integer comparison where one operand of the compare is a constant. | |||
3570 | static Value *simplifySelectBitTest(Value *TrueVal, Value *FalseVal, Value *X, | |||
3571 | const APInt *Y, bool TrueWhenUnset) { | |||
3572 | const APInt *C; | |||
3573 | ||||
3574 | // (X & Y) == 0 ? X & ~Y : X --> X | |||
3575 | // (X & Y) != 0 ? X & ~Y : X --> X & ~Y | |||
3576 | if (FalseVal == X && match(TrueVal, m_And(m_Specific(X), m_APInt(C))) && | |||
3577 | *Y == ~*C) | |||
3578 | return TrueWhenUnset ? FalseVal : TrueVal; | |||
3579 | ||||
3580 | // (X & Y) == 0 ? X : X & ~Y --> X & ~Y | |||
3581 | // (X & Y) != 0 ? X : X & ~Y --> X | |||
3582 | if (TrueVal == X && match(FalseVal, m_And(m_Specific(X), m_APInt(C))) && | |||
3583 | *Y == ~*C) | |||
3584 | return TrueWhenUnset ? FalseVal : TrueVal; | |||
3585 | ||||
3586 | if (Y->isPowerOf2()) { | |||
3587 | // (X & Y) == 0 ? X | Y : X --> X | Y | |||
3588 | // (X & Y) != 0 ? X | Y : X --> X | |||
3589 | if (FalseVal == X && match(TrueVal, m_Or(m_Specific(X), m_APInt(C))) && | |||
3590 | *Y == *C) | |||
3591 | return TrueWhenUnset ? TrueVal : FalseVal; | |||
3592 | ||||
3593 | // (X & Y) == 0 ? X : X | Y --> X | |||
3594 | // (X & Y) != 0 ? X : X | Y --> X | Y | |||
3595 | if (TrueVal == X && match(FalseVal, m_Or(m_Specific(X), m_APInt(C))) && | |||
3596 | *Y == *C) | |||
3597 | return TrueWhenUnset ? TrueVal : FalseVal; | |||
3598 | } | |||
3599 | ||||
3600 | return nullptr; | |||
3601 | } | |||
3602 | ||||
3603 | /// An alternative way to test if a bit is set or not uses sgt/slt instead of | |||
3604 | /// eq/ne. | |||
3605 | static Value *simplifySelectWithFakeICmpEq(Value *CmpLHS, Value *CmpRHS, | |||
3606 | ICmpInst::Predicate Pred, | |||
3607 | Value *TrueVal, Value *FalseVal) { | |||
3608 | Value *X; | |||
3609 | APInt Mask; | |||
3610 | if (!decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, X, Mask)) | |||
3611 | return nullptr; | |||
3612 | ||||
3613 | return simplifySelectBitTest(TrueVal, FalseVal, X, &Mask, | |||
3614 | Pred == ICmpInst::ICMP_EQ); | |||
3615 | } | |||
3616 | ||||
3617 | /// Try to simplify a select instruction when its condition operand is an | |||
3618 | /// integer comparison. | |||
3619 | static Value *simplifySelectWithICmpCond(Value *CondVal, Value *TrueVal, | |||
3620 | Value *FalseVal, const SimplifyQuery &Q, | |||
3621 | unsigned MaxRecurse) { | |||
3622 | ICmpInst::Predicate Pred; | |||
3623 | Value *CmpLHS, *CmpRHS; | |||
3624 | if (!match(CondVal, m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)))) | |||
3625 | return nullptr; | |||
3626 | ||||
3627 | if (ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero())) { | |||
3628 | Value *X; | |||
3629 | const APInt *Y; | |||
3630 | if (match(CmpLHS, m_And(m_Value(X), m_APInt(Y)))) | |||
3631 | if (Value *V = simplifySelectBitTest(TrueVal, FalseVal, X, Y, | |||
3632 | Pred == ICmpInst::ICMP_EQ)) | |||
3633 | return V; | |||
3634 | } | |||
3635 | ||||
3636 | // Check for other compares that behave like bit test. | |||
3637 | if (Value *V = simplifySelectWithFakeICmpEq(CmpLHS, CmpRHS, Pred, | |||
3638 | TrueVal, FalseVal)) | |||
3639 | return V; | |||
3640 | ||||
3641 | // If we have an equality comparison, then we know the value in one of the | |||
3642 | // arms of the select. See if substituting this value into the arm and | |||
3643 | // simplifying the result yields the same value as the other arm. | |||
3644 | if (Pred == ICmpInst::ICMP_EQ) { | |||
3645 | if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == | |||
3646 | TrueVal || | |||
3647 | SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == | |||
3648 | TrueVal) | |||
3649 | return FalseVal; | |||
3650 | if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == | |||
3651 | FalseVal || | |||
3652 | SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == | |||
3653 | FalseVal) | |||
3654 | return FalseVal; | |||
3655 | } else if (Pred == ICmpInst::ICMP_NE) { | |||
3656 | if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q, MaxRecurse) == | |||
3657 | FalseVal || | |||
3658 | SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q, MaxRecurse) == | |||
3659 | FalseVal) | |||
3660 | return TrueVal; | |||
3661 | if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q, MaxRecurse) == | |||
3662 | TrueVal || | |||
3663 | SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q, MaxRecurse) == | |||
3664 | TrueVal) | |||
3665 | return TrueVal; | |||
3666 | } | |||
3667 | ||||
3668 | return nullptr; | |||
3669 | } | |||
3670 | ||||
3671 | /// Given operands for a SelectInst, see if we can fold the result. | |||
3672 | /// If not, this returns null. | |||
3673 | static Value *SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, | |||
3674 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
3675 | if (auto *CondC = dyn_cast<Constant>(Cond)) { | |||
3676 | if (auto *TrueC = dyn_cast<Constant>(TrueVal)) | |||
3677 | if (auto *FalseC = dyn_cast<Constant>(FalseVal)) | |||
3678 | return ConstantFoldSelectInstruction(CondC, TrueC, FalseC); | |||
3679 | ||||
3680 | // select undef, X, Y -> X or Y | |||
3681 | if (isa<UndefValue>(CondC)) | |||
3682 | return isa<Constant>(FalseVal) ? FalseVal : TrueVal; | |||
3683 | ||||
3684 | // TODO: Vector constants with undef elements don't simplify. | |||
3685 | ||||
3686 | // select true, X, Y -> X | |||
3687 | if (CondC->isAllOnesValue()) | |||
3688 | return TrueVal; | |||
3689 | // select false, X, Y -> Y | |||
3690 | if (CondC->isNullValue()) | |||
3691 | return FalseVal; | |||
3692 | } | |||
3693 | ||||
3694 | // select ?, X, X -> X | |||
3695 | if (TrueVal == FalseVal) | |||
3696 | return TrueVal; | |||
3697 | ||||
3698 | if (isa<UndefValue>(TrueVal)) // select ?, undef, X -> X | |||
3699 | return FalseVal; | |||
3700 | if (isa<UndefValue>(FalseVal)) // select ?, X, undef -> X | |||
3701 | return TrueVal; | |||
3702 | ||||
3703 | if (Value *V = | |||
3704 | simplifySelectWithICmpCond(Cond, TrueVal, FalseVal, Q, MaxRecurse)) | |||
3705 | return V; | |||
3706 | ||||
3707 | return nullptr; | |||
3708 | } | |||
3709 | ||||
3710 | Value *llvm::SimplifySelectInst(Value *Cond, Value *TrueVal, Value *FalseVal, | |||
3711 | const SimplifyQuery &Q) { | |||
3712 | return ::SimplifySelectInst(Cond, TrueVal, FalseVal, Q, RecursionLimit); | |||
3713 | } | |||
3714 | ||||
3715 | /// Given operands for an GetElementPtrInst, see if we can fold the result. | |||
3716 | /// If not, this returns null. | |||
3717 | static Value *SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, | |||
3718 | const SimplifyQuery &Q, unsigned) { | |||
3719 | // The type of the GEP pointer operand. | |||
3720 | unsigned AS = | |||
3721 | cast<PointerType>(Ops[0]->getType()->getScalarType())->getAddressSpace(); | |||
3722 | ||||
3723 | // getelementptr P -> P. | |||
3724 | if (Ops.size() == 1) | |||
3725 | return Ops[0]; | |||
3726 | ||||
3727 | // Compute the (pointer) type returned by the GEP instruction. | |||
3728 | Type *LastType = GetElementPtrInst::getIndexedType(SrcTy, Ops.slice(1)); | |||
3729 | Type *GEPTy = PointerType::get(LastType, AS); | |||
3730 | if (VectorType *VT = dyn_cast<VectorType>(Ops[0]->getType())) | |||
3731 | GEPTy = VectorType::get(GEPTy, VT->getNumElements()); | |||
3732 | else if (VectorType *VT = dyn_cast<VectorType>(Ops[1]->getType())) | |||
3733 | GEPTy = VectorType::get(GEPTy, VT->getNumElements()); | |||
3734 | ||||
3735 | if (isa<UndefValue>(Ops[0])) | |||
3736 | return UndefValue::get(GEPTy); | |||
3737 | ||||
3738 | if (Ops.size() == 2) { | |||
3739 | // getelementptr P, 0 -> P. | |||
3740 | if (match(Ops[1], m_Zero()) && Ops[0]->getType() == GEPTy) | |||
3741 | return Ops[0]; | |||
3742 | ||||
3743 | Type *Ty = SrcTy; | |||
3744 | if (Ty->isSized()) { | |||
3745 | Value *P; | |||
3746 | uint64_t C; | |||
3747 | uint64_t TyAllocSize = Q.DL.getTypeAllocSize(Ty); | |||
3748 | // getelementptr P, N -> P if P points to a type of zero size. | |||
3749 | if (TyAllocSize == 0 && Ops[0]->getType() == GEPTy) | |||
3750 | return Ops[0]; | |||
3751 | ||||
3752 | // The following transforms are only safe if the ptrtoint cast | |||
3753 | // doesn't truncate the pointers. | |||
3754 | if (Ops[1]->getType()->getScalarSizeInBits() == | |||
3755 | Q.DL.getIndexSizeInBits(AS)) { | |||
3756 | auto PtrToIntOrZero = [GEPTy](Value *P) -> Value * { | |||
3757 | if (match(P, m_Zero())) | |||
3758 | return Constant::getNullValue(GEPTy); | |||
3759 | Value *Temp; | |||
3760 | if (match(P, m_PtrToInt(m_Value(Temp)))) | |||
3761 | if (Temp->getType() == GEPTy) | |||
3762 | return Temp; | |||
3763 | return nullptr; | |||
3764 | }; | |||
3765 | ||||
3766 | // getelementptr V, (sub P, V) -> P if P points to a type of size 1. | |||
3767 | if (TyAllocSize == 1 && | |||
3768 | match(Ops[1], m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))))) | |||
3769 | if (Value *R = PtrToIntOrZero(P)) | |||
3770 | return R; | |||
3771 | ||||
3772 | // getelementptr V, (ashr (sub P, V), C) -> Q | |||
3773 | // if P points to a type of size 1 << C. | |||
3774 | if (match(Ops[1], | |||
3775 | m_AShr(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), | |||
3776 | m_ConstantInt(C))) && | |||
3777 | TyAllocSize == 1ULL << C) | |||
3778 | if (Value *R = PtrToIntOrZero(P)) | |||
3779 | return R; | |||
3780 | ||||
3781 | // getelementptr V, (sdiv (sub P, V), C) -> Q | |||
3782 | // if P points to a type of size C. | |||
3783 | if (match(Ops[1], | |||
3784 | m_SDiv(m_Sub(m_Value(P), m_PtrToInt(m_Specific(Ops[0]))), | |||
3785 | m_SpecificInt(TyAllocSize)))) | |||
3786 | if (Value *R = PtrToIntOrZero(P)) | |||
3787 | return R; | |||
3788 | } | |||
3789 | } | |||
3790 | } | |||
3791 | ||||
3792 | if (Q.DL.getTypeAllocSize(LastType) == 1 && | |||
3793 | all_of(Ops.slice(1).drop_back(1), | |||
3794 | [](Value *Idx) { return match(Idx, m_Zero()); })) { | |||
3795 | unsigned IdxWidth = | |||
3796 | Q.DL.getIndexSizeInBits(Ops[0]->getType()->getPointerAddressSpace()); | |||
3797 | if (Q.DL.getTypeSizeInBits(Ops.back()->getType()) == IdxWidth) { | |||
3798 | APInt BasePtrOffset(IdxWidth, 0); | |||
3799 | Value *StrippedBasePtr = | |||
3800 | Ops[0]->stripAndAccumulateInBoundsConstantOffsets(Q.DL, | |||
3801 | BasePtrOffset); | |||
3802 | ||||
3803 | // gep (gep V, C), (sub 0, V) -> C | |||
3804 | if (match(Ops.back(), | |||
3805 | m_Sub(m_Zero(), m_PtrToInt(m_Specific(StrippedBasePtr))))) { | |||
3806 | auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset); | |||
3807 | return ConstantExpr::getIntToPtr(CI, GEPTy); | |||
3808 | } | |||
3809 | // gep (gep V, C), (xor V, -1) -> C-1 | |||
3810 | if (match(Ops.back(), | |||
3811 | m_Xor(m_PtrToInt(m_Specific(StrippedBasePtr)), m_AllOnes()))) { | |||
3812 | auto *CI = ConstantInt::get(GEPTy->getContext(), BasePtrOffset - 1); | |||
3813 | return ConstantExpr::getIntToPtr(CI, GEPTy); | |||
3814 | } | |||
3815 | } | |||
3816 | } | |||
3817 | ||||
3818 | // Check to see if this is constant foldable. | |||
3819 | if (!all_of(Ops, [](Value *V) { return isa<Constant>(V); })) | |||
3820 | return nullptr; | |||
3821 | ||||
3822 | auto *CE = ConstantExpr::getGetElementPtr(SrcTy, cast<Constant>(Ops[0]), | |||
3823 | Ops.slice(1)); | |||
3824 | if (auto *CEFolded = ConstantFoldConstant(CE, Q.DL)) | |||
3825 | return CEFolded; | |||
3826 | return CE; | |||
3827 | } | |||
3828 | ||||
3829 | Value *llvm::SimplifyGEPInst(Type *SrcTy, ArrayRef<Value *> Ops, | |||
3830 | const SimplifyQuery &Q) { | |||
3831 | return ::SimplifyGEPInst(SrcTy, Ops, Q, RecursionLimit); | |||
3832 | } | |||
3833 | ||||
3834 | /// Given operands for an InsertValueInst, see if we can fold the result. | |||
3835 | /// If not, this returns null. | |||
3836 | static Value *SimplifyInsertValueInst(Value *Agg, Value *Val, | |||
3837 | ArrayRef<unsigned> Idxs, const SimplifyQuery &Q, | |||
3838 | unsigned) { | |||
3839 | if (Constant *CAgg = dyn_cast<Constant>(Agg)) | |||
3840 | if (Constant *CVal = dyn_cast<Constant>(Val)) | |||
3841 | return ConstantFoldInsertValueInstruction(CAgg, CVal, Idxs); | |||
3842 | ||||
3843 | // insertvalue x, undef, n -> x | |||
3844 | if (match(Val, m_Undef())) | |||
3845 | return Agg; | |||
3846 | ||||
3847 | // insertvalue x, (extractvalue y, n), n | |||
3848 | if (ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val)) | |||
3849 | if (EV->getAggregateOperand()->getType() == Agg->getType() && | |||
3850 | EV->getIndices() == Idxs) { | |||
3851 | // insertvalue undef, (extractvalue y, n), n -> y | |||
3852 | if (match(Agg, m_Undef())) | |||
3853 | return EV->getAggregateOperand(); | |||
3854 | ||||
3855 | // insertvalue y, (extractvalue y, n), n -> y | |||
3856 | if (Agg == EV->getAggregateOperand()) | |||
3857 | return Agg; | |||
3858 | } | |||
3859 | ||||
3860 | return nullptr; | |||
3861 | } | |||
3862 | ||||
3863 | Value *llvm::SimplifyInsertValueInst(Value *Agg, Value *Val, | |||
3864 | ArrayRef<unsigned> Idxs, | |||
3865 | const SimplifyQuery &Q) { | |||
3866 | return ::SimplifyInsertValueInst(Agg, Val, Idxs, Q, RecursionLimit); | |||
3867 | } | |||
3868 | ||||
3869 | Value *llvm::SimplifyInsertElementInst(Value *Vec, Value *Val, Value *Idx, | |||
3870 | const SimplifyQuery &Q) { | |||
3871 | // Try to constant fold. | |||
3872 | auto *VecC = dyn_cast<Constant>(Vec); | |||
3873 | auto *ValC = dyn_cast<Constant>(Val); | |||
3874 | auto *IdxC = dyn_cast<Constant>(Idx); | |||
3875 | if (VecC && ValC && IdxC) | |||
3876 | return ConstantFoldInsertElementInstruction(VecC, ValC, IdxC); | |||
3877 | ||||
3878 | // Fold into undef if index is out of bounds. | |||
3879 | if (auto *CI = dyn_cast<ConstantInt>(Idx)) { | |||
3880 | uint64_t NumElements = cast<VectorType>(Vec->getType())->getNumElements(); | |||
3881 | if (CI->uge(NumElements)) | |||
3882 | return UndefValue::get(Vec->getType()); | |||
3883 | } | |||
3884 | ||||
3885 | // If index is undef, it might be out of bounds (see above case) | |||
3886 | if (isa<UndefValue>(Idx)) | |||
3887 | return UndefValue::get(Vec->getType()); | |||
3888 | ||||
3889 | return nullptr; | |||
3890 | } | |||
3891 | ||||
3892 | /// Given operands for an ExtractValueInst, see if we can fold the result. | |||
3893 | /// If not, this returns null. | |||
3894 | static Value *SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, | |||
3895 | const SimplifyQuery &, unsigned) { | |||
3896 | if (auto *CAgg = dyn_cast<Constant>(Agg)) | |||
3897 | return ConstantFoldExtractValueInstruction(CAgg, Idxs); | |||
3898 | ||||
3899 | // extractvalue x, (insertvalue y, elt, n), n -> elt | |||
3900 | unsigned NumIdxs = Idxs.size(); | |||
3901 | for (auto *IVI = dyn_cast<InsertValueInst>(Agg); IVI != nullptr; | |||
3902 | IVI = dyn_cast<InsertValueInst>(IVI->getAggregateOperand())) { | |||
3903 | ArrayRef<unsigned> InsertValueIdxs = IVI->getIndices(); | |||
3904 | unsigned NumInsertValueIdxs = InsertValueIdxs.size(); | |||
3905 | unsigned NumCommonIdxs = std::min(NumInsertValueIdxs, NumIdxs); | |||
3906 | if (InsertValueIdxs.slice(0, NumCommonIdxs) == | |||
3907 | Idxs.slice(0, NumCommonIdxs)) { | |||
3908 | if (NumIdxs == NumInsertValueIdxs) | |||
3909 | return IVI->getInsertedValueOperand(); | |||
3910 | break; | |||
3911 | } | |||
3912 | } | |||
3913 | ||||
3914 | return nullptr; | |||
3915 | } | |||
3916 | ||||
3917 | Value *llvm::SimplifyExtractValueInst(Value *Agg, ArrayRef<unsigned> Idxs, | |||
3918 | const SimplifyQuery &Q) { | |||
3919 | return ::SimplifyExtractValueInst(Agg, Idxs, Q, RecursionLimit); | |||
3920 | } | |||
3921 | ||||
3922 | /// Given operands for an ExtractElementInst, see if we can fold the result. | |||
3923 | /// If not, this returns null. | |||
3924 | static Value *SimplifyExtractElementInst(Value *Vec, Value *Idx, const SimplifyQuery &, | |||
3925 | unsigned) { | |||
3926 | if (auto *CVec = dyn_cast<Constant>(Vec)) { | |||
3927 | if (auto *CIdx = dyn_cast<Constant>(Idx)) | |||
3928 | return ConstantFoldExtractElementInstruction(CVec, CIdx); | |||
3929 | ||||
3930 | // The index is not relevant if our vector is a splat. | |||
3931 | if (auto *Splat = CVec->getSplatValue()) | |||
3932 | return Splat; | |||
3933 | ||||
3934 | if (isa<UndefValue>(Vec)) | |||
3935 | return UndefValue::get(Vec->getType()->getVectorElementType()); | |||
3936 | } | |||
3937 | ||||
3938 | // If extracting a specified index from the vector, see if we can recursively | |||
3939 | // find a previously computed scalar that was inserted into the vector. | |||
3940 | if (auto *IdxC = dyn_cast<ConstantInt>(Idx)) { | |||
3941 | if (IdxC->getValue().uge(Vec->getType()->getVectorNumElements())) | |||
3942 | // definitely out of bounds, thus undefined result | |||
3943 | return UndefValue::get(Vec->getType()->getVectorElementType()); | |||
3944 | if (Value *Elt = findScalarElement(Vec, IdxC->getZExtValue())) | |||
3945 | return Elt; | |||
3946 | } | |||
3947 | ||||
3948 | // An undef extract index can be arbitrarily chosen to be an out-of-range | |||
3949 | // index value, which would result in the instruction being undef. | |||
3950 | if (isa<UndefValue>(Idx)) | |||
3951 | return UndefValue::get(Vec->getType()->getVectorElementType()); | |||
3952 | ||||
3953 | return nullptr; | |||
3954 | } | |||
3955 | ||||
3956 | Value *llvm::SimplifyExtractElementInst(Value *Vec, Value *Idx, | |||
3957 | const SimplifyQuery &Q) { | |||
3958 | return ::SimplifyExtractElementInst(Vec, Idx, Q, RecursionLimit); | |||
3959 | } | |||
3960 | ||||
3961 | /// See if we can fold the given phi. If not, returns null. | |||
3962 | static Value *SimplifyPHINode(PHINode *PN, const SimplifyQuery &Q) { | |||
3963 | // If all of the PHI's incoming values are the same then replace the PHI node | |||
3964 | // with the common value. | |||
3965 | Value *CommonValue = nullptr; | |||
3966 | bool HasUndefInput = false; | |||
3967 | for (Value *Incoming : PN->incoming_values()) { | |||
3968 | // If the incoming value is the phi node itself, it can safely be skipped. | |||
3969 | if (Incoming == PN) continue; | |||
3970 | if (isa<UndefValue>(Incoming)) { | |||
3971 | // Remember that we saw an undef value, but otherwise ignore them. | |||
3972 | HasUndefInput = true; | |||
3973 | continue; | |||
3974 | } | |||
3975 | if (CommonValue && Incoming != CommonValue) | |||
3976 | return nullptr; // Not the same, bail out. | |||
3977 | CommonValue = Incoming; | |||
3978 | } | |||
3979 | ||||
3980 | // If CommonValue is null then all of the incoming values were either undef or | |||
3981 | // equal to the phi node itself. | |||
3982 | if (!CommonValue) | |||
3983 | return UndefValue::get(PN->getType()); | |||
3984 | ||||
3985 | // If we have a PHI node like phi(X, undef, X), where X is defined by some | |||
3986 | // instruction, we cannot return X as the result of the PHI node unless it | |||
3987 | // dominates the PHI block. | |||
3988 | if (HasUndefInput) | |||
3989 | return ValueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr; | |||
3990 | ||||
3991 | return CommonValue; | |||
3992 | } | |||
3993 | ||||
3994 | static Value *SimplifyCastInst(unsigned CastOpc, Value *Op, | |||
3995 | Type *Ty, const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
3996 | if (auto *C = dyn_cast<Constant>(Op)) | |||
3997 | return ConstantFoldCastOperand(CastOpc, C, Ty, Q.DL); | |||
3998 | ||||
3999 | if (auto *CI = dyn_cast<CastInst>(Op)) { | |||
4000 | auto *Src = CI->getOperand(0); | |||
4001 | Type *SrcTy = Src->getType(); | |||
4002 | Type *MidTy = CI->getType(); | |||
4003 | Type *DstTy = Ty; | |||
4004 | if (Src->getType() == Ty) { | |||
4005 | auto FirstOp = static_cast<Instruction::CastOps>(CI->getOpcode()); | |||
4006 | auto SecondOp = static_cast<Instruction::CastOps>(CastOpc); | |||
4007 | Type *SrcIntPtrTy = | |||
4008 | SrcTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(SrcTy) : nullptr; | |||
4009 | Type *MidIntPtrTy = | |||
4010 | MidTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(MidTy) : nullptr; | |||
4011 | Type *DstIntPtrTy = | |||
4012 | DstTy->isPtrOrPtrVectorTy() ? Q.DL.getIntPtrType(DstTy) : nullptr; | |||
4013 | if (CastInst::isEliminableCastPair(FirstOp, SecondOp, SrcTy, MidTy, DstTy, | |||
4014 | SrcIntPtrTy, MidIntPtrTy, | |||
4015 | DstIntPtrTy) == Instruction::BitCast) | |||
4016 | return Src; | |||
4017 | } | |||
4018 | } | |||
4019 | ||||
4020 | // bitcast x -> x | |||
4021 | if (CastOpc == Instruction::BitCast) | |||
4022 | if (Op->getType() == Ty) | |||
4023 | return Op; | |||
4024 | ||||
4025 | return nullptr; | |||
4026 | } | |||
4027 | ||||
4028 | Value *llvm::SimplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty, | |||
4029 | const SimplifyQuery &Q) { | |||
4030 | return ::SimplifyCastInst(CastOpc, Op, Ty, Q, RecursionLimit); | |||
4031 | } | |||
4032 | ||||
4033 | /// For the given destination element of a shuffle, peek through shuffles to | |||
4034 | /// match a root vector source operand that contains that element in the same | |||
4035 | /// vector lane (ie, the same mask index), so we can eliminate the shuffle(s). | |||
4036 | static Value *foldIdentityShuffles(int DestElt, Value *Op0, Value *Op1, | |||
4037 | int MaskVal, Value *RootVec, | |||
4038 | unsigned MaxRecurse) { | |||
4039 | if (!MaxRecurse--) | |||
4040 | return nullptr; | |||
4041 | ||||
4042 | // Bail out if any mask value is undefined. That kind of shuffle may be | |||
4043 | // simplified further based on demanded bits or other folds. | |||
4044 | if (MaskVal == -1) | |||
4045 | return nullptr; | |||
4046 | ||||
4047 | // The mask value chooses which source operand we need to look at next. | |||
4048 | int InVecNumElts = Op0->getType()->getVectorNumElements(); | |||
4049 | int RootElt = MaskVal; | |||
4050 | Value *SourceOp = Op0; | |||
4051 | if (MaskVal >= InVecNumElts) { | |||
4052 | RootElt = MaskVal - InVecNumElts; | |||
4053 | SourceOp = Op1; | |||
4054 | } | |||
4055 | ||||
4056 | // If the source operand is a shuffle itself, look through it to find the | |||
4057 | // matching root vector. | |||
4058 | if (auto *SourceShuf = dyn_cast<ShuffleVectorInst>(SourceOp)) { | |||
4059 | return foldIdentityShuffles( | |||
4060 | DestElt, SourceShuf->getOperand(0), SourceShuf->getOperand(1), | |||
4061 | SourceShuf->getMaskValue(RootElt), RootVec, MaxRecurse); | |||
4062 | } | |||
4063 | ||||
4064 | // TODO: Look through bitcasts? What if the bitcast changes the vector element | |||
4065 | // size? | |||
4066 | ||||
4067 | // The source operand is not a shuffle. Initialize the root vector value for | |||
4068 | // this shuffle if that has not been done yet. | |||
4069 | if (!RootVec) | |||
4070 | RootVec = SourceOp; | |||
4071 | ||||
4072 | // Give up as soon as a source operand does not match the existing root value. | |||
4073 | if (RootVec != SourceOp) | |||
4074 | return nullptr; | |||
4075 | ||||
4076 | // The element must be coming from the same lane in the source vector | |||
4077 | // (although it may have crossed lanes in intermediate shuffles). | |||
4078 | if (RootElt != DestElt) | |||
4079 | return nullptr; | |||
4080 | ||||
4081 | return RootVec; | |||
4082 | } | |||
4083 | ||||
4084 | static Value *SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask, | |||
4085 | Type *RetTy, const SimplifyQuery &Q, | |||
4086 | unsigned MaxRecurse) { | |||
4087 | if (isa<UndefValue>(Mask)) | |||
4088 | return UndefValue::get(RetTy); | |||
4089 | ||||
4090 | Type *InVecTy = Op0->getType(); | |||
4091 | unsigned MaskNumElts = Mask->getType()->getVectorNumElements(); | |||
4092 | unsigned InVecNumElts = InVecTy->getVectorNumElements(); | |||
4093 | ||||
4094 | SmallVector<int, 32> Indices; | |||
4095 | ShuffleVectorInst::getShuffleMask(Mask, Indices); | |||
4096 | 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~svn329677/lib/Analysis/InstructionSimplify.cpp" , 4097, __extension__ __PRETTY_FUNCTION__)) | |||
4097 | "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~svn329677/lib/Analysis/InstructionSimplify.cpp" , 4097, __extension__ __PRETTY_FUNCTION__)); | |||
4098 | ||||
4099 | // Canonicalization: If mask does not select elements from an input vector, | |||
4100 | // replace that input vector with undef. | |||
4101 | bool MaskSelects0 = false, MaskSelects1 = false; | |||
4102 | for (unsigned i = 0; i != MaskNumElts; ++i) { | |||
4103 | if (Indices[i] == -1) | |||
4104 | continue; | |||
4105 | if ((unsigned)Indices[i] < InVecNumElts) | |||
4106 | MaskSelects0 = true; | |||
4107 | else | |||
4108 | MaskSelects1 = true; | |||
4109 | } | |||
4110 | if (!MaskSelects0) | |||
4111 | Op0 = UndefValue::get(InVecTy); | |||
4112 | if (!MaskSelects1) | |||
4113 | Op1 = UndefValue::get(InVecTy); | |||
4114 | ||||
4115 | auto *Op0Const = dyn_cast<Constant>(Op0); | |||
4116 | auto *Op1Const = dyn_cast<Constant>(Op1); | |||
4117 | ||||
4118 | // If all operands are constant, constant fold the shuffle. | |||
4119 | if (Op0Const && Op1Const) | |||
4120 | return ConstantFoldShuffleVectorInstruction(Op0Const, Op1Const, Mask); | |||
4121 | ||||
4122 | // Canonicalization: if only one input vector is constant, it shall be the | |||
4123 | // second one. | |||
4124 | if (Op0Const && !Op1Const) { | |||
4125 | std::swap(Op0, Op1); | |||
4126 | ShuffleVectorInst::commuteShuffleMask(Indices, InVecNumElts); | |||
4127 | } | |||
4128 | ||||
4129 | // A shuffle of a splat is always the splat itself. Legal if the shuffle's | |||
4130 | // value type is same as the input vectors' type. | |||
4131 | if (auto *OpShuf = dyn_cast<ShuffleVectorInst>(Op0)) | |||
4132 | if (isa<UndefValue>(Op1) && RetTy == InVecTy && | |||
4133 | OpShuf->getMask()->getSplatValue()) | |||
4134 | return Op0; | |||
4135 | ||||
4136 | // Don't fold a shuffle with undef mask elements. This may get folded in a | |||
4137 | // better way using demanded bits or other analysis. | |||
4138 | // TODO: Should we allow this? | |||
4139 | if (find(Indices, -1) != Indices.end()) | |||
4140 | return nullptr; | |||
4141 | ||||
4142 | // Check if every element of this shuffle can be mapped back to the | |||
4143 | // corresponding element of a single root vector. If so, we don't need this | |||
4144 | // shuffle. This handles simple identity shuffles as well as chains of | |||
4145 | // shuffles that may widen/narrow and/or move elements across lanes and back. | |||
4146 | Value *RootVec = nullptr; | |||
4147 | for (unsigned i = 0; i != MaskNumElts; ++i) { | |||
4148 | // Note that recursion is limited for each vector element, so if any element | |||
4149 | // exceeds the limit, this will fail to simplify. | |||
4150 | RootVec = | |||
4151 | foldIdentityShuffles(i, Op0, Op1, Indices[i], RootVec, MaxRecurse); | |||
4152 | ||||
4153 | // We can't replace a widening/narrowing shuffle with one of its operands. | |||
4154 | if (!RootVec || RootVec->getType() != RetTy) | |||
4155 | return nullptr; | |||
4156 | } | |||
4157 | return RootVec; | |||
4158 | } | |||
4159 | ||||
4160 | /// Given operands for a ShuffleVectorInst, fold the result or return null. | |||
4161 | Value *llvm::SimplifyShuffleVectorInst(Value *Op0, Value *Op1, Constant *Mask, | |||
4162 | Type *RetTy, const SimplifyQuery &Q) { | |||
4163 | return ::SimplifyShuffleVectorInst(Op0, Op1, Mask, RetTy, Q, RecursionLimit); | |||
4164 | } | |||
4165 | ||||
4166 | static Constant *propagateNaN(Constant *In) { | |||
4167 | // If the input is a vector with undef elements, just return a default NaN. | |||
4168 | if (!In->isNaN()) | |||
4169 | return ConstantFP::getNaN(In->getType()); | |||
4170 | ||||
4171 | // Propagate the existing NaN constant when possible. | |||
4172 | // TODO: Should we quiet a signaling NaN? | |||
4173 | return In; | |||
4174 | } | |||
4175 | ||||
4176 | static Constant *simplifyFPBinop(Value *Op0, Value *Op1) { | |||
4177 | if (isa<UndefValue>(Op0) || isa<UndefValue>(Op1)) | |||
4178 | return ConstantFP::getNaN(Op0->getType()); | |||
4179 | ||||
4180 | if (match(Op0, m_NaN())) | |||
4181 | return propagateNaN(cast<Constant>(Op0)); | |||
4182 | if (match(Op1, m_NaN())) | |||
4183 | return propagateNaN(cast<Constant>(Op1)); | |||
4184 | ||||
4185 | return nullptr; | |||
4186 | } | |||
4187 | ||||
4188 | /// Given operands for an FAdd, see if we can fold the result. If not, this | |||
4189 | /// returns null. | |||
4190 | static Value *SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4191 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
4192 | if (Constant *C = foldOrCommuteConstant(Instruction::FAdd, Op0, Op1, Q)) | |||
4193 | return C; | |||
4194 | ||||
4195 | if (Constant *C = simplifyFPBinop(Op0, Op1)) | |||
4196 | return C; | |||
4197 | ||||
4198 | // fadd X, -0 ==> X | |||
4199 | if (match(Op1, m_NegZeroFP())) | |||
4200 | return Op0; | |||
4201 | ||||
4202 | // fadd X, 0 ==> X, when we know X is not -0 | |||
4203 | if (match(Op1, m_PosZeroFP()) && | |||
4204 | (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI))) | |||
4205 | return Op0; | |||
4206 | ||||
4207 | // With nnan: (+/-0.0 - X) + X --> 0.0 (and commuted variant) | |||
4208 | // We don't have to explicitly exclude infinities (ninf): INF + -INF == NaN. | |||
4209 | // Negative zeros are allowed because we always end up with positive zero: | |||
4210 | // X = -0.0: (-0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0 | |||
4211 | // X = -0.0: ( 0.0 - (-0.0)) + (-0.0) == ( 0.0) + (-0.0) == 0.0 | |||
4212 | // X = 0.0: (-0.0 - ( 0.0)) + ( 0.0) == (-0.0) + ( 0.0) == 0.0 | |||
4213 | // X = 0.0: ( 0.0 - ( 0.0)) + ( 0.0) == ( 0.0) + ( 0.0) == 0.0 | |||
4214 | if (FMF.noNaNs() && (match(Op0, m_FSub(m_AnyZeroFP(), m_Specific(Op1))) || | |||
4215 | match(Op1, m_FSub(m_AnyZeroFP(), m_Specific(Op0))))) | |||
4216 | return ConstantFP::getNullValue(Op0->getType()); | |||
4217 | ||||
4218 | return nullptr; | |||
4219 | } | |||
4220 | ||||
4221 | /// Given operands for an FSub, see if we can fold the result. If not, this | |||
4222 | /// returns null. | |||
4223 | static Value *SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4224 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
4225 | if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q)) | |||
4226 | return C; | |||
4227 | ||||
4228 | if (Constant *C = simplifyFPBinop(Op0, Op1)) | |||
4229 | return C; | |||
4230 | ||||
4231 | // fsub X, +0 ==> X | |||
4232 | if (match(Op1, m_PosZeroFP())) | |||
4233 | return Op0; | |||
4234 | ||||
4235 | // fsub X, -0 ==> X, when we know X is not -0 | |||
4236 | if (match(Op1, m_NegZeroFP()) && | |||
4237 | (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI))) | |||
4238 | return Op0; | |||
4239 | ||||
4240 | // fsub -0.0, (fsub -0.0, X) ==> X | |||
4241 | Value *X; | |||
4242 | if (match(Op0, m_NegZeroFP()) && | |||
4243 | match(Op1, m_FSub(m_NegZeroFP(), m_Value(X)))) | |||
4244 | return X; | |||
4245 | ||||
4246 | // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored. | |||
4247 | if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) && | |||
4248 | match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X)))) | |||
4249 | return X; | |||
4250 | ||||
4251 | // fsub nnan x, x ==> 0.0 | |||
4252 | if (FMF.noNaNs() && Op0 == Op1) | |||
4253 | return Constant::getNullValue(Op0->getType()); | |||
4254 | ||||
4255 | return nullptr; | |||
4256 | } | |||
4257 | ||||
4258 | /// Given the operands for an FMul, see if we can fold the result | |||
4259 | static Value *SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4260 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
4261 | if (Constant *C = foldOrCommuteConstant(Instruction::FMul, Op0, Op1, Q)) | |||
4262 | return C; | |||
4263 | ||||
4264 | if (Constant *C = simplifyFPBinop(Op0, Op1)) | |||
4265 | return C; | |||
4266 | ||||
4267 | // fmul X, 1.0 ==> X | |||
4268 | if (match(Op1, m_FPOne())) | |||
4269 | return Op0; | |||
4270 | ||||
4271 | // fmul nnan nsz X, 0 ==> 0 | |||
4272 | if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op1, m_AnyZeroFP())) | |||
4273 | return ConstantFP::getNullValue(Op0->getType()); | |||
4274 | ||||
4275 | // sqrt(X) * sqrt(X) --> X, if we can: | |||
4276 | // 1. Remove the intermediate rounding (reassociate). | |||
4277 | // 2. Ignore non-zero negative numbers because sqrt would produce NAN. | |||
4278 | // 3. Ignore -0.0 because sqrt(-0.0) == -0.0, but -0.0 * -0.0 == 0.0. | |||
4279 | Value *X; | |||
4280 | if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::sqrt>(m_Value(X))) && | |||
4281 | FMF.allowReassoc() && FMF.noNaNs() && FMF.noSignedZeros()) | |||
4282 | return X; | |||
4283 | ||||
4284 | return nullptr; | |||
4285 | } | |||
4286 | ||||
4287 | Value *llvm::SimplifyFAddInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4288 | const SimplifyQuery &Q) { | |||
4289 | return ::SimplifyFAddInst(Op0, Op1, FMF, Q, RecursionLimit); | |||
4290 | } | |||
4291 | ||||
4292 | ||||
4293 | Value *llvm::SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4294 | const SimplifyQuery &Q) { | |||
4295 | return ::SimplifyFSubInst(Op0, Op1, FMF, Q, RecursionLimit); | |||
4296 | } | |||
4297 | ||||
4298 | Value *llvm::SimplifyFMulInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4299 | const SimplifyQuery &Q) { | |||
4300 | return ::SimplifyFMulInst(Op0, Op1, FMF, Q, RecursionLimit); | |||
4301 | } | |||
4302 | ||||
4303 | static Value *SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4304 | const SimplifyQuery &Q, unsigned) { | |||
4305 | if (Constant *C = foldOrCommuteConstant(Instruction::FDiv, Op0, Op1, Q)) | |||
4306 | return C; | |||
4307 | ||||
4308 | if (Constant *C = simplifyFPBinop(Op0, Op1)) | |||
4309 | return C; | |||
4310 | ||||
4311 | // X / 1.0 -> X | |||
4312 | if (match(Op1, m_FPOne())) | |||
4313 | return Op0; | |||
4314 | ||||
4315 | // 0 / X -> 0 | |||
4316 | // Requires that NaNs are off (X could be zero) and signed zeroes are | |||
4317 | // ignored (X could be positive or negative, so the output sign is unknown). | |||
4318 | if (FMF.noNaNs() && FMF.noSignedZeros() && match(Op0, m_AnyZeroFP())) | |||
4319 | return ConstantFP::getNullValue(Op0->getType()); | |||
4320 | ||||
4321 | if (FMF.noNaNs()) { | |||
4322 | // X / X -> 1.0 is legal when NaNs are ignored. | |||
4323 | // We can ignore infinities because INF/INF is NaN. | |||
4324 | if (Op0 == Op1) | |||
4325 | return ConstantFP::get(Op0->getType(), 1.0); | |||
4326 | ||||
4327 | // (X * Y) / Y --> X if we can reassociate to the above form. | |||
4328 | Value *X; | |||
4329 | if (FMF.allowReassoc() && match(Op0, m_c_FMul(m_Value(X), m_Specific(Op1)))) | |||
4330 | return X; | |||
4331 | ||||
4332 | // -X / X -> -1.0 and | |||
4333 | // X / -X -> -1.0 are legal when NaNs are ignored. | |||
4334 | // We can ignore signed zeros because +-0.0/+-0.0 is NaN and ignored. | |||
4335 | if ((BinaryOperator::isFNeg(Op0, /*IgnoreZeroSign=*/true) && | |||
4336 | BinaryOperator::getFNegArgument(Op0) == Op1) || | |||
4337 | (BinaryOperator::isFNeg(Op1, /*IgnoreZeroSign=*/true) && | |||
4338 | BinaryOperator::getFNegArgument(Op1) == Op0)) | |||
4339 | return ConstantFP::get(Op0->getType(), -1.0); | |||
4340 | } | |||
4341 | ||||
4342 | return nullptr; | |||
4343 | } | |||
4344 | ||||
4345 | Value *llvm::SimplifyFDivInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4346 | const SimplifyQuery &Q) { | |||
4347 | return ::SimplifyFDivInst(Op0, Op1, FMF, Q, RecursionLimit); | |||
4348 | } | |||
4349 | ||||
4350 | static Value *SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4351 | const SimplifyQuery &Q, unsigned) { | |||
4352 | if (Constant *C = foldOrCommuteConstant(Instruction::FRem, Op0, Op1, Q)) | |||
4353 | return C; | |||
4354 | ||||
4355 | if (Constant *C = simplifyFPBinop(Op0, Op1)) | |||
4356 | return C; | |||
4357 | ||||
4358 | // Unlike fdiv, the result of frem always matches the sign of the dividend. | |||
4359 | // The constant match may include undef elements in a vector, so return a full | |||
4360 | // zero constant as the result. | |||
4361 | if (FMF.noNaNs()) { | |||
4362 | // +0 % X -> 0 | |||
4363 | if (match(Op0, m_PosZeroFP())) | |||
4364 | return ConstantFP::getNullValue(Op0->getType()); | |||
4365 | // -0 % X -> -0 | |||
4366 | if (match(Op0, m_NegZeroFP())) | |||
4367 | return ConstantFP::getNegativeZero(Op0->getType()); | |||
4368 | } | |||
4369 | ||||
4370 | return nullptr; | |||
4371 | } | |||
4372 | ||||
4373 | Value *llvm::SimplifyFRemInst(Value *Op0, Value *Op1, FastMathFlags FMF, | |||
4374 | const SimplifyQuery &Q) { | |||
4375 | return ::SimplifyFRemInst(Op0, Op1, FMF, Q, RecursionLimit); | |||
4376 | } | |||
4377 | ||||
4378 | //=== Helper functions for higher up the class hierarchy. | |||
4379 | ||||
4380 | /// Given operands for a BinaryOperator, see if we can fold the result. | |||
4381 | /// If not, this returns null. | |||
4382 | static Value *SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, | |||
4383 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
4384 | switch (Opcode) { | |||
4385 | case Instruction::Add: | |||
4386 | return SimplifyAddInst(LHS, RHS, false, false, Q, MaxRecurse); | |||
4387 | case Instruction::Sub: | |||
4388 | return SimplifySubInst(LHS, RHS, false, false, Q, MaxRecurse); | |||
4389 | case Instruction::Mul: | |||
4390 | return SimplifyMulInst(LHS, RHS, Q, MaxRecurse); | |||
4391 | case Instruction::SDiv: | |||
4392 | return SimplifySDivInst(LHS, RHS, Q, MaxRecurse); | |||
4393 | case Instruction::UDiv: | |||
4394 | return SimplifyUDivInst(LHS, RHS, Q, MaxRecurse); | |||
4395 | case Instruction::SRem: | |||
4396 | return SimplifySRemInst(LHS, RHS, Q, MaxRecurse); | |||
4397 | case Instruction::URem: | |||
4398 | return SimplifyURemInst(LHS, RHS, Q, MaxRecurse); | |||
4399 | case Instruction::Shl: | |||
4400 | return SimplifyShlInst(LHS, RHS, false, false, Q, MaxRecurse); | |||
4401 | case Instruction::LShr: | |||
4402 | return SimplifyLShrInst(LHS, RHS, false, Q, MaxRecurse); | |||
4403 | case Instruction::AShr: | |||
4404 | return SimplifyAShrInst(LHS, RHS, false, Q, MaxRecurse); | |||
4405 | case Instruction::And: | |||
4406 | return SimplifyAndInst(LHS, RHS, Q, MaxRecurse); | |||
4407 | case Instruction::Or: | |||
4408 | return SimplifyOrInst(LHS, RHS, Q, MaxRecurse); | |||
4409 | case Instruction::Xor: | |||
4410 | return SimplifyXorInst(LHS, RHS, Q, MaxRecurse); | |||
4411 | case Instruction::FAdd: | |||
4412 | return SimplifyFAddInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); | |||
4413 | case Instruction::FSub: | |||
4414 | return SimplifyFSubInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); | |||
4415 | case Instruction::FMul: | |||
4416 | return SimplifyFMulInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); | |||
4417 | case Instruction::FDiv: | |||
4418 | return SimplifyFDivInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); | |||
4419 | case Instruction::FRem: | |||
4420 | return SimplifyFRemInst(LHS, RHS, FastMathFlags(), Q, MaxRecurse); | |||
4421 | default: | |||
4422 | llvm_unreachable("Unexpected opcode")::llvm::llvm_unreachable_internal("Unexpected opcode", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Analysis/InstructionSimplify.cpp" , 4422); | |||
4423 | } | |||
4424 | } | |||
4425 | ||||
4426 | /// Given operands for a BinaryOperator, see if we can fold the result. | |||
4427 | /// If not, this returns null. | |||
4428 | /// In contrast to SimplifyBinOp, try to use FastMathFlag when folding the | |||
4429 | /// result. In case we don't need FastMathFlags, simply fall to SimplifyBinOp. | |||
4430 | static Value *SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, | |||
4431 | const FastMathFlags &FMF, const SimplifyQuery &Q, | |||
4432 | unsigned MaxRecurse) { | |||
4433 | switch (Opcode) { | |||
4434 | case Instruction::FAdd: | |||
4435 | return SimplifyFAddInst(LHS, RHS, FMF, Q, MaxRecurse); | |||
4436 | case Instruction::FSub: | |||
4437 | return SimplifyFSubInst(LHS, RHS, FMF, Q, MaxRecurse); | |||
4438 | case Instruction::FMul: | |||
4439 | return SimplifyFMulInst(LHS, RHS, FMF, Q, MaxRecurse); | |||
4440 | case Instruction::FDiv: | |||
4441 | return SimplifyFDivInst(LHS, RHS, FMF, Q, MaxRecurse); | |||
4442 | default: | |||
4443 | return SimplifyBinOp(Opcode, LHS, RHS, Q, MaxRecurse); | |||
4444 | } | |||
4445 | } | |||
4446 | ||||
4447 | Value *llvm::SimplifyBinOp(unsigned Opcode, Value *LHS, Value *RHS, | |||
4448 | const SimplifyQuery &Q) { | |||
4449 | return ::SimplifyBinOp(Opcode, LHS, RHS, Q, RecursionLimit); | |||
| ||||
4450 | } | |||
4451 | ||||
4452 | Value *llvm::SimplifyFPBinOp(unsigned Opcode, Value *LHS, Value *RHS, | |||
4453 | FastMathFlags FMF, const SimplifyQuery &Q) { | |||
4454 | return ::SimplifyFPBinOp(Opcode, LHS, RHS, FMF, Q, RecursionLimit); | |||
4455 | } | |||
4456 | ||||
4457 | /// Given operands for a CmpInst, see if we can fold the result. | |||
4458 | static Value *SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, | |||
4459 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
4460 | if (CmpInst::isIntPredicate((CmpInst::Predicate)Predicate)) | |||
4461 | return SimplifyICmpInst(Predicate, LHS, RHS, Q, MaxRecurse); | |||
4462 | return SimplifyFCmpInst(Predicate, LHS, RHS, FastMathFlags(), Q, MaxRecurse); | |||
4463 | } | |||
4464 | ||||
4465 | Value *llvm::SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, | |||
4466 | const SimplifyQuery &Q) { | |||
4467 | return ::SimplifyCmpInst(Predicate, LHS, RHS, Q, RecursionLimit); | |||
4468 | } | |||
4469 | ||||
4470 | static bool IsIdempotent(Intrinsic::ID ID) { | |||
4471 | switch (ID) { | |||
4472 | default: return false; | |||
4473 | ||||
4474 | // Unary idempotent: f(f(x)) = f(x) | |||
4475 | case Intrinsic::fabs: | |||
4476 | case Intrinsic::floor: | |||
4477 | case Intrinsic::ceil: | |||
4478 | case Intrinsic::trunc: | |||
4479 | case Intrinsic::rint: | |||
4480 | case Intrinsic::nearbyint: | |||
4481 | case Intrinsic::round: | |||
4482 | case Intrinsic::canonicalize: | |||
4483 | return true; | |||
4484 | } | |||
4485 | } | |||
4486 | ||||
4487 | static Value *SimplifyRelativeLoad(Constant *Ptr, Constant *Offset, | |||
4488 | const DataLayout &DL) { | |||
4489 | GlobalValue *PtrSym; | |||
4490 | APInt PtrOffset; | |||
4491 | if (!IsConstantOffsetFromGlobal(Ptr, PtrSym, PtrOffset, DL)) | |||
4492 | return nullptr; | |||
4493 | ||||
4494 | Type *Int8PtrTy = Type::getInt8PtrTy(Ptr->getContext()); | |||
4495 | Type *Int32Ty = Type::getInt32Ty(Ptr->getContext()); | |||
4496 | Type *Int32PtrTy = Int32Ty->getPointerTo(); | |||
4497 | Type *Int64Ty = Type::getInt64Ty(Ptr->getContext()); | |||
4498 | ||||
4499 | auto *OffsetConstInt = dyn_cast<ConstantInt>(Offset); | |||
4500 | if (!OffsetConstInt || OffsetConstInt->getType()->getBitWidth() > 64) | |||
4501 | return nullptr; | |||
4502 | ||||
4503 | uint64_t OffsetInt = OffsetConstInt->getSExtValue(); | |||
4504 | if (OffsetInt % 4 != 0) | |||
4505 | return nullptr; | |||
4506 | ||||
4507 | Constant *C = ConstantExpr::getGetElementPtr( | |||
4508 | Int32Ty, ConstantExpr::getBitCast(Ptr, Int32PtrTy), | |||
4509 | ConstantInt::get(Int64Ty, OffsetInt / 4)); | |||
4510 | Constant *Loaded = ConstantFoldLoadFromConstPtr(C, Int32Ty, DL); | |||
4511 | if (!Loaded) | |||
4512 | return nullptr; | |||
4513 | ||||
4514 | auto *LoadedCE = dyn_cast<ConstantExpr>(Loaded); | |||
4515 | if (!LoadedCE) | |||
4516 | return nullptr; | |||
4517 | ||||
4518 | if (LoadedCE->getOpcode() == Instruction::Trunc) { | |||
4519 | LoadedCE = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0)); | |||
4520 | if (!LoadedCE) | |||
4521 | return nullptr; | |||
4522 | } | |||
4523 | ||||
4524 | if (LoadedCE->getOpcode() != Instruction::Sub) | |||
4525 | return nullptr; | |||
4526 | ||||
4527 | auto *LoadedLHS = dyn_cast<ConstantExpr>(LoadedCE->getOperand(0)); | |||
4528 | if (!LoadedLHS || LoadedLHS->getOpcode() != Instruction::PtrToInt) | |||
4529 | return nullptr; | |||
4530 | auto *LoadedLHSPtr = LoadedLHS->getOperand(0); | |||
4531 | ||||
4532 | Constant *LoadedRHS = LoadedCE->getOperand(1); | |||
4533 | GlobalValue *LoadedRHSSym; | |||
4534 | APInt LoadedRHSOffset; | |||
4535 | if (!IsConstantOffsetFromGlobal(LoadedRHS, LoadedRHSSym, LoadedRHSOffset, | |||
4536 | DL) || | |||
4537 | PtrSym != LoadedRHSSym || PtrOffset != LoadedRHSOffset) | |||
4538 | return nullptr; | |||
4539 | ||||
4540 | return ConstantExpr::getBitCast(LoadedLHSPtr, Int8PtrTy); | |||
4541 | } | |||
4542 | ||||
4543 | static bool maskIsAllZeroOrUndef(Value *Mask) { | |||
4544 | auto *ConstMask = dyn_cast<Constant>(Mask); | |||
4545 | if (!ConstMask) | |||
4546 | return false; | |||
4547 | if (ConstMask->isNullValue() || isa<UndefValue>(ConstMask)) | |||
4548 | return true; | |||
4549 | for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E; | |||
4550 | ++I) { | |||
4551 | if (auto *MaskElt = ConstMask->getAggregateElement(I)) | |||
4552 | if (MaskElt->isNullValue() || isa<UndefValue>(MaskElt)) | |||
4553 | continue; | |||
4554 | return false; | |||
4555 | } | |||
4556 | return true; | |||
4557 | } | |||
4558 | ||||
4559 | template <typename IterTy> | |||
4560 | static Value *SimplifyIntrinsic(Function *F, IterTy ArgBegin, IterTy ArgEnd, | |||
4561 | const SimplifyQuery &Q, unsigned MaxRecurse) { | |||
4562 | Intrinsic::ID IID = F->getIntrinsicID(); | |||
4563 | unsigned NumOperands = std::distance(ArgBegin, ArgEnd); | |||
4564 | ||||
4565 | // Unary Ops | |||
4566 | if (NumOperands == 1) { | |||
4567 | // Perform idempotent optimizations | |||
4568 | if (IsIdempotent(IID)) { | |||
4569 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*ArgBegin)) { | |||
4570 | if (II->getIntrinsicID() == IID) | |||
4571 | return II; | |||
4572 | } | |||
4573 | } | |||
4574 | ||||
4575 | Value *IIOperand = *ArgBegin; | |||
4576 | Value *X; | |||
4577 | switch (IID) { | |||
4578 | case Intrinsic::fabs: { | |||
4579 | if (SignBitMustBeZero(IIOperand, Q.TLI)) | |||
4580 | return IIOperand; | |||
4581 | return nullptr; | |||
4582 | } | |||
4583 | case Intrinsic::bswap: { | |||
4584 | // bswap(bswap(x)) -> x | |||
4585 | if (match(IIOperand, m_BSwap(m_Value(X)))) | |||
4586 | return X; | |||
4587 | return nullptr; | |||
4588 | } | |||
4589 | case Intrinsic::bitreverse: { | |||
4590 | // bitreverse(bitreverse(x)) -> x | |||
4591 | if (match(IIOperand, m_BitReverse(m_Value(X)))) | |||
4592 | return X; | |||
4593 | return nullptr; | |||
4594 | } | |||
4595 | case Intrinsic::exp: { | |||
4596 | // exp(log(x)) -> x | |||
4597 | if (Q.CxtI->hasAllowReassoc() && | |||
4598 | match(IIOperand, m_Intrinsic<Intrinsic::log>(m_Value(X)))) | |||
4599 | return X; | |||
4600 | return nullptr; | |||
4601 | } | |||
4602 | case Intrinsic::exp2: { | |||
4603 | // exp2(log2(x)) -> x | |||
4604 | if (Q.CxtI->hasAllowReassoc() && | |||
4605 | match(IIOperand, m_Intrinsic<Intrinsic::log2>(m_Value(X)))) | |||
4606 | return X; | |||
4607 | return nullptr; | |||
4608 | } | |||
4609 | case Intrinsic::log: { | |||
4610 | // log(exp(x)) -> x | |||
4611 | if (Q.CxtI->hasAllowReassoc() && | |||
4612 | match(IIOperand, m_Intrinsic<Intrinsic::exp>(m_Value(X)))) | |||
4613 | return X; | |||
4614 | return nullptr; | |||
4615 | } | |||
4616 | case Intrinsic::log2: { | |||
4617 | // log2(exp2(x)) -> x | |||
4618 | if (Q.CxtI->hasAllowReassoc() && | |||
4619 | match(IIOperand, m_Intrinsic<Intrinsic::exp2>(m_Value(X)))) { | |||
4620 | return X; | |||
4621 | } | |||
4622 | return nullptr; | |||
4623 | } | |||
4624 | default: | |||
4625 | return nullptr; | |||
4626 | } | |||
4627 | } | |||
4628 | ||||
4629 | // Binary Ops | |||
4630 | if (NumOperands == 2) { | |||
4631 | Value *LHS = *ArgBegin; | |||
4632 | Value *RHS = *(ArgBegin + 1); | |||
4633 | Type *ReturnType = F->getReturnType(); | |||
4634 | ||||
4635 | switch (IID) { | |||
4636 | case Intrinsic::usub_with_overflow: | |||
4637 | case Intrinsic::ssub_with_overflow: { | |||
4638 | // X - X -> { 0, false } | |||
4639 | if (LHS == RHS) | |||
4640 | return Constant::getNullValue(ReturnType); | |||
4641 | ||||
4642 | // X - undef -> undef | |||
4643 | // undef - X -> undef | |||
4644 | if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) | |||
4645 | return UndefValue::get(ReturnType); | |||
4646 | ||||
4647 | return nullptr; | |||
4648 | } | |||
4649 | case Intrinsic::uadd_with_overflow: | |||
4650 | case Intrinsic::sadd_with_overflow: { | |||
4651 | // X + undef -> undef | |||
4652 | if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS)) | |||
4653 | return UndefValue::get(ReturnType); | |||
4654 | ||||
4655 | return nullptr; | |||
4656 | } | |||
4657 | case Intrinsic::umul_with_overflow: | |||
4658 | case Intrinsic::smul_with_overflow: { | |||
4659 | // 0 * X -> { 0, false } | |||
4660 | // X * 0 -> { 0, false } | |||
4661 | if (match(LHS, m_Zero()) || match(RHS, m_Zero())) | |||
4662 | return Constant::getNullValue(ReturnType); | |||
4663 | ||||
4664 | // undef * X -> { 0, false } | |||
4665 | // X * undef -> { 0, false } | |||
4666 | if (match(LHS, m_Undef()) || match(RHS, m_Undef())) | |||
4667 | return Constant::getNullValue(ReturnType); | |||
4668 | ||||
4669 | return nullptr; | |||
4670 | } | |||
4671 | case Intrinsic::load_relative: { | |||
4672 | Constant *C0 = dyn_cast<Constant>(LHS); | |||
4673 | Constant *C1 = dyn_cast<Constant>(RHS); | |||
4674 | if (C0 && C1) | |||
4675 | return SimplifyRelativeLoad(C0, C1, Q.DL); | |||
4676 | return nullptr; | |||
4677 | } | |||
4678 | case Intrinsic::powi: | |||
4679 | if (ConstantInt *Power = dyn_cast<ConstantInt>(RHS)) { | |||
4680 | // powi(x, 0) -> 1.0 | |||
4681 | if (Power->isZero()) | |||
4682 | return ConstantFP::get(LHS->getType(), 1.0); | |||
4683 | // powi(x, 1) -> x | |||
4684 | if (Power->isOne()) | |||
4685 | return LHS; | |||
4686 | } | |||
4687 | return nullptr; | |||
4688 | default: | |||
4689 | return nullptr; | |||
4690 | } | |||
4691 | } | |||
4692 | ||||
4693 | // Simplify calls to llvm.masked.load.* | |||
4694 | switch (IID) { | |||
4695 | case Intrinsic::masked_load: { | |||
4696 | Value *MaskArg = ArgBegin[2]; | |||
4697 | Value *PassthruArg = ArgBegin[3]; | |||
4698 | // If the mask is all zeros or undef, the "passthru" argument is the result. | |||
4699 | if (maskIsAllZeroOrUndef(MaskArg)) | |||
4700 | return PassthruArg; | |||
4701 | return nullptr; | |||
4702 | } | |||
4703 | default: | |||
4704 | return nullptr; | |||
4705 | } | |||
4706 | } | |||
4707 | ||||
4708 | template <typename IterTy> | |||
4709 | static Value *SimplifyCall(ImmutableCallSite CS, Value *V, IterTy ArgBegin, | |||
4710 | IterTy ArgEnd, const SimplifyQuery &Q, | |||
4711 | unsigned MaxRecurse) { | |||
4712 | Type *Ty = V->getType(); | |||
4713 | if (PointerType *PTy = dyn_cast<PointerType>(Ty)) | |||
4714 | Ty = PTy->getElementType(); | |||
4715 | FunctionType *FTy = cast<FunctionType>(Ty); | |||
4716 | ||||
4717 | // call undef -> undef | |||
4718 | // call null -> undef | |||
4719 | if (isa<UndefValue>(V) || isa<ConstantPointerNull>(V)) | |||
4720 | return UndefValue::get(FTy->getReturnType()); | |||
4721 | ||||
4722 | Function *F = dyn_cast<Function>(V); | |||
4723 | if (!F) | |||
4724 | return nullptr; | |||
4725 | ||||
4726 | if (F->isIntrinsic()) | |||
4727 | if (Value *Ret = SimplifyIntrinsic(F, ArgBegin, ArgEnd, Q, MaxRecurse)) | |||
4728 | return Ret; | |||
4729 | ||||
4730 | if (!canConstantFoldCallTo(CS, F)) | |||
4731 | return nullptr; | |||
4732 | ||||
4733 | SmallVector<Constant *, 4> ConstantArgs; | |||
4734 | ConstantArgs.reserve(ArgEnd - ArgBegin); | |||
4735 | for (IterTy I = ArgBegin, E = ArgEnd; I != E; ++I) { | |||
4736 | Constant *C = dyn_cast<Constant>(*I); | |||
4737 | if (!C) | |||
4738 | return nullptr; | |||
4739 | ConstantArgs.push_back(C); | |||
4740 | } | |||
4741 | ||||
4742 | return ConstantFoldCall(CS, F, ConstantArgs, Q.TLI); | |||
4743 | } | |||
4744 | ||||
4745 | Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V, | |||
4746 | User::op_iterator ArgBegin, User::op_iterator ArgEnd, | |||
4747 | const SimplifyQuery &Q) { | |||
4748 | return ::SimplifyCall(CS, V, ArgBegin, ArgEnd, Q, RecursionLimit); | |||
4749 | } | |||
4750 | ||||
4751 | Value *llvm::SimplifyCall(ImmutableCallSite CS, Value *V, | |||
4752 | ArrayRef<Value *> Args, const SimplifyQuery &Q) { | |||
4753 | return ::SimplifyCall(CS, V, Args.begin(), Args.end(), Q, RecursionLimit); | |||
4754 | } | |||
4755 | ||||
4756 | Value *llvm::SimplifyCall(ImmutableCallSite ICS, const SimplifyQuery &Q) { | |||
4757 | CallSite CS(const_cast<Instruction*>(ICS.getInstruction())); | |||
4758 | return ::SimplifyCall(CS, CS.getCalledValue(), CS.arg_begin(), CS.arg_end(), | |||
4759 | Q, RecursionLimit); | |||
4760 | } | |||
4761 | ||||
4762 | /// See if we can compute a simplified version of this instruction. | |||
4763 | /// If not, this returns null. | |||
4764 | ||||
4765 | Value *llvm::SimplifyInstruction(Instruction *I, const SimplifyQuery &SQ, | |||
4766 | OptimizationRemarkEmitter *ORE) { | |||
4767 | const SimplifyQuery Q = SQ.CxtI ? SQ : SQ.getWithInstruction(I); | |||
4768 | Value *Result; | |||
4769 | ||||
4770 | switch (I->getOpcode()) { | |||
4771 | default: | |||
4772 | Result = ConstantFoldInstruction(I, Q.DL, Q.TLI); | |||
4773 | break; | |||
4774 | case Instruction::FAdd: | |||
4775 | Result = SimplifyFAddInst(I->getOperand(0), I->getOperand(1), | |||
4776 | I->getFastMathFlags(), Q); | |||
4777 | break; | |||
4778 | case Instruction::Add: | |||
4779 | Result = SimplifyAddInst(I->getOperand(0), I->getOperand(1), | |||
4780 | cast<BinaryOperator>(I)->hasNoSignedWrap(), | |||
4781 | cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q); | |||
4782 | break; | |||
4783 | case Instruction::FSub: | |||
4784 | Result = SimplifyFSubInst(I->getOperand(0), I->getOperand(1), | |||
4785 | I->getFastMathFlags(), Q); | |||
4786 | break; | |||
4787 | case Instruction::Sub: | |||
4788 | Result = SimplifySubInst(I->getOperand(0), I->getOperand(1), | |||
4789 | cast<BinaryOperator>(I)->hasNoSignedWrap(), | |||
4790 | cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q); | |||
4791 | break; | |||
4792 | case Instruction::FMul: | |||
4793 | Result = SimplifyFMulInst(I->getOperand(0), I->getOperand(1), | |||
4794 | I->getFastMathFlags(), Q); | |||
4795 | break; | |||
4796 | case Instruction::Mul: | |||
4797 | Result = SimplifyMulInst(I->getOperand(0), I->getOperand(1), Q); | |||
4798 | break; | |||
4799 | case Instruction::SDiv: | |||
4800 | Result = SimplifySDivInst(I->getOperand(0), I->getOperand(1), Q); | |||
4801 | break; | |||
4802 | case Instruction::UDiv: | |||
4803 | Result = SimplifyUDivInst(I->getOperand(0), I->getOperand(1), Q); | |||
4804 | break; | |||
4805 | case Instruction::FDiv: | |||
4806 | Result = SimplifyFDivInst(I->getOperand(0), I->getOperand(1), | |||
4807 | I->getFastMathFlags(), Q); | |||
4808 | break; | |||
4809 | case Instruction::SRem: | |||
4810 | Result = SimplifySRemInst(I->getOperand(0), I->getOperand(1), Q); | |||
4811 | break; | |||
4812 | case Instruction::URem: | |||
4813 | Result = SimplifyURemInst(I->getOperand(0), I->getOperand(1), Q); | |||
4814 | break; | |||
4815 | case Instruction::FRem: | |||
4816 | Result = SimplifyFRemInst(I->getOperand(0), I->getOperand(1), | |||
4817 | I->getFastMathFlags(), Q); | |||
4818 | break; | |||
4819 | case Instruction::Shl: | |||
4820 | Result = SimplifyShlInst(I->getOperand(0), I->getOperand(1), | |||
4821 | cast<BinaryOperator>(I)->hasNoSignedWrap(), | |||
4822 | cast<BinaryOperator>(I)->hasNoUnsignedWrap(), Q); | |||
4823 | break; | |||
4824 | case Instruction::LShr: | |||
4825 | Result = SimplifyLShrInst(I->getOperand(0), I->getOperand(1), | |||
4826 | cast<BinaryOperator>(I)->isExact(), Q); | |||
4827 | break; | |||
4828 | case Instruction::AShr: | |||
4829 | Result = SimplifyAShrInst(I->getOperand(0), I->getOperand(1), | |||
4830 | cast<BinaryOperator>(I)->isExact(), Q); | |||
4831 | break; | |||
4832 | case Instruction::And: | |||
4833 | Result = SimplifyAndInst(I->getOperand(0), I->getOperand(1), Q); | |||
4834 | break; | |||
4835 | case Instruction::Or: | |||
4836 | Result = SimplifyOrInst(I->getOperand(0), I->getOperand(1), Q); | |||
4837 | break; | |||
4838 | case Instruction::Xor: | |||
4839 | Result = SimplifyXorInst(I->getOperand(0), I->getOperand(1), Q); | |||
4840 | break; | |||
4841 | case Instruction::ICmp: | |||
4842 | Result = SimplifyICmpInst(cast<ICmpInst>(I)->getPredicate(), | |||
4843 | I->getOperand(0), I->getOperand(1), Q); | |||
4844 | break; | |||
4845 | case Instruction::FCmp: | |||
4846 | Result = | |||
4847 | SimplifyFCmpInst(cast<FCmpInst>(I)->getPredicate(), I->getOperand(0), | |||
4848 | I->getOperand(1), I->getFastMathFlags(), Q); | |||
4849 | break; | |||
4850 | case Instruction::Select: | |||
4851 | Result = SimplifySelectInst(I->getOperand(0), I->getOperand(1), | |||
4852 | I->getOperand(2), Q); | |||
4853 | break; | |||
4854 | case Instruction::GetElementPtr: { | |||
4855 | SmallVector<Value *, 8> Ops(I->op_begin(), I->op_end()); | |||
4856 | Result = SimplifyGEPInst(cast<GetElementPtrInst>(I)->getSourceElementType(), | |||
4857 | Ops, Q); | |||
4858 | break; | |||
4859 | } | |||
4860 | case Instruction::InsertValue: { | |||
4861 | InsertValueInst *IV = cast<InsertValueInst>(I); | |||
4862 | Result = SimplifyInsertValueInst(IV->getAggregateOperand(), | |||
4863 | IV->getInsertedValueOperand(), | |||
4864 | IV->getIndices(), Q); | |||
4865 | break; | |||
4866 | } | |||
4867 | case Instruction::InsertElement: { | |||
4868 | auto *IE = cast<InsertElementInst>(I); | |||
4869 | Result = SimplifyInsertElementInst(IE->getOperand(0), IE->getOperand(1), | |||
4870 | IE->getOperand(2), Q); | |||
4871 | break; | |||
4872 | } | |||
4873 | case Instruction::ExtractValue: { | |||
4874 | auto *EVI = cast<ExtractValueInst>(I); | |||
4875 | Result = SimplifyExtractValueInst(EVI->getAggregateOperand(), | |||
4876 | EVI->getIndices(), Q); | |||
4877 | break; | |||
4878 | } | |||
4879 | case Instruction::ExtractElement: { | |||
4880 | auto *EEI = cast<ExtractElementInst>(I); | |||
4881 | Result = SimplifyExtractElementInst(EEI->getVectorOperand(), | |||
4882 | EEI->getIndexOperand(), Q); | |||
4883 | break; | |||
4884 | } | |||
4885 | case Instruction::ShuffleVector: { | |||
4886 | auto *SVI = cast<ShuffleVectorInst>(I); | |||
4887 | Result = SimplifyShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1), | |||
4888 | SVI->getMask(), SVI->getType(), Q); | |||
4889 | break; | |||
4890 | } | |||
4891 | case Instruction::PHI: | |||
4892 | Result = SimplifyPHINode(cast<PHINode>(I), Q); | |||
4893 | break; | |||
4894 | case Instruction::Call: { | |||
4895 | CallSite CS(cast<CallInst>(I)); | |||
4896 | Result = SimplifyCall(CS, Q); | |||
4897 | break; | |||
4898 | } | |||
4899 | #define HANDLE_CAST_INST(num, opc, clas) case Instruction::opc: | |||
4900 | #include "llvm/IR/Instruction.def" | |||
4901 | #undef HANDLE_CAST_INST | |||
4902 | Result = | |||
4903 | SimplifyCastInst(I->getOpcode(), I->getOperand(0), I->getType(), Q); | |||
4904 | break; | |||
4905 | case Instruction::Alloca: | |||
4906 | // No simplifications for Alloca and it can't be constant folded. | |||
4907 | Result = nullptr; | |||
4908 | break; | |||
4909 | } | |||
4910 | ||||
4911 | // In general, it is possible for computeKnownBits to determine all bits in a | |||
4912 | // value even when the operands are not all constants. | |||
4913 | if (!Result && I->getType()->isIntOrIntVectorTy()) { | |||
4914 | KnownBits Known = computeKnownBits(I, Q.DL, /*Depth*/ 0, Q.AC, I, Q.DT, ORE); | |||
4915 | if (Known.isConstant()) | |||
4916 | Result = ConstantInt::get(I->getType(), Known.getConstant()); | |||
4917 | } | |||
4918 | ||||
4919 | /// If called on unreachable code, the above logic may report that the | |||
4920 | /// instruction simplified to itself. Make life easier for users by | |||
4921 | /// detecting that case here, returning a safe value instead. | |||
4922 | return Result == I ? UndefValue::get(I->getType()) : Result; | |||
4923 | } | |||
4924 | ||||
4925 | /// \brief Implementation of recursive simplification through an instruction's | |||
4926 | /// uses. | |||
4927 | /// | |||
4928 | /// This is the common implementation of the recursive simplification routines. | |||
4929 | /// If we have a pre-simplified value in 'SimpleV', that is forcibly used to | |||
4930 | /// replace the instruction 'I'. Otherwise, we simply add 'I' to the list of | |||
4931 | /// instructions to process and attempt to simplify it using | |||
4932 | /// InstructionSimplify. | |||
4933 | /// | |||
4934 | /// This routine returns 'true' only when *it* simplifies something. The passed | |||
4935 | /// in simplified value does not count toward this. | |||
4936 | static bool replaceAndRecursivelySimplifyImpl(Instruction *I, Value *SimpleV, | |||
4937 | const TargetLibraryInfo *TLI, | |||
4938 | const DominatorTree *DT, | |||
4939 | AssumptionCache *AC) { | |||
4940 | bool Simplified = false; | |||
4941 | SmallSetVector<Instruction *, 8> Worklist; | |||
4942 | const DataLayout &DL = I->getModule()->getDataLayout(); | |||
4943 | ||||
4944 | // If we have an explicit value to collapse to, do that round of the | |||
4945 | // simplification loop by hand initially. | |||
4946 | if (SimpleV) { | |||
4947 | for (User *U : I->users()) | |||
4948 | if (U != I) | |||
4949 | Worklist.insert(cast<Instruction>(U)); | |||
4950 | ||||
4951 | // Replace the instruction with its simplified value. | |||
4952 | I->replaceAllUsesWith(SimpleV); | |||
4953 | ||||
4954 | // Gracefully handle edge cases where the instruction is not wired into any | |||
4955 | // parent block. | |||
4956 | if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) && | |||
4957 | !I->mayHaveSideEffects()) | |||
4958 | I->eraseFromParent(); | |||
4959 | } else { | |||
4960 | Worklist.insert(I); | |||
4961 | } | |||
4962 | ||||
4963 | // Note that we must test the size on each iteration, the worklist can grow. | |||
4964 | for (unsigned Idx = 0; Idx != Worklist.size(); ++Idx) { | |||
4965 | I = Worklist[Idx]; | |||
4966 | ||||
4967 | // See if this instruction simplifies. | |||
4968 | SimpleV = SimplifyInstruction(I, {DL, TLI, DT, AC}); | |||
4969 | if (!SimpleV) | |||
4970 | continue; | |||
4971 | ||||
4972 | Simplified = true; | |||
4973 | ||||
4974 | // Stash away all the uses of the old instruction so we can check them for | |||
4975 | // recursive simplifications after a RAUW. This is cheaper than checking all | |||
4976 | // uses of To on the recursive step in most cases. | |||
4977 | for (User *U : I->users()) | |||
4978 | Worklist.insert(cast<Instruction>(U)); | |||
4979 | ||||
4980 | // Replace the instruction with its simplified value. | |||
4981 | I->replaceAllUsesWith(SimpleV); | |||
4982 | ||||
4983 | // Gracefully handle edge cases where the instruction is not wired into any | |||
4984 | // parent block. | |||
4985 | if (I->getParent() && !I->isEHPad() && !isa<TerminatorInst>(I) && | |||
4986 | !I->mayHaveSideEffects()) | |||
4987 | I->eraseFromParent(); | |||
4988 | } | |||
4989 | return Simplified; | |||
4990 | } | |||
4991 | ||||
4992 | bool llvm::recursivelySimplifyInstruction(Instruction *I, | |||
4993 | const TargetLibraryInfo *TLI, | |||
4994 | const DominatorTree *DT, | |||
4995 | AssumptionCache *AC) { | |||
4996 | return replaceAndRecursivelySimplifyImpl(I, nullptr, TLI, DT, AC); | |||
4997 | } | |||
4998 | ||||
4999 | bool llvm::replaceAndRecursivelySimplify(Instruction *I, Value *SimpleV, | |||
5000 | const TargetLibraryInfo *TLI, | |||
5001 | const DominatorTree *DT, | |||
5002 | AssumptionCache *AC) { | |||
5003 | 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~svn329677/lib/Analysis/InstructionSimplify.cpp" , 5003, __extension__ __PRETTY_FUNCTION__)); | |||
5004 | 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~svn329677/lib/Analysis/InstructionSimplify.cpp" , 5004, __extension__ __PRETTY_FUNCTION__)); | |||
5005 | return replaceAndRecursivelySimplifyImpl(I, SimpleV, TLI, DT, AC); | |||
5006 | } | |||
5007 | ||||
5008 | namespace llvm { | |||
5009 | const SimplifyQuery getBestSimplifyQuery(Pass &P, Function &F) { | |||
5010 | auto *DTWP = P.getAnalysisIfAvailable<DominatorTreeWrapperPass>(); | |||
5011 | auto *DT = DTWP ? &DTWP->getDomTree() : nullptr; | |||
5012 | auto *TLIWP = P.getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | |||
5013 | auto *TLI = TLIWP ? &TLIWP->getTLI() : nullptr; | |||
5014 | auto *ACWP = P.getAnalysisIfAvailable<AssumptionCacheTracker>(); | |||
5015 | auto *AC = ACWP ? &ACWP->getAssumptionCache(F) : nullptr; | |||
5016 | return {F.getParent()->getDataLayout(), TLI, DT, AC}; | |||
5017 | } | |||
5018 | ||||
5019 | const SimplifyQuery getBestSimplifyQuery(LoopStandardAnalysisResults &AR, | |||
5020 | const DataLayout &DL) { | |||
5021 | return {DL, &AR.TLI, &AR.DT, &AR.AC}; | |||
5022 | } | |||
5023 | ||||
5024 | template <class T, class... TArgs> | |||
5025 | const SimplifyQuery getBestSimplifyQuery(AnalysisManager<T, TArgs...> &AM, | |||
5026 | Function &F) { | |||
5027 | auto *DT = AM.template getCachedResult<DominatorTreeAnalysis>(F); | |||
5028 | auto *TLI = AM.template getCachedResult<TargetLibraryAnalysis>(F); | |||
5029 | auto *AC = AM.template getCachedResult<AssumptionAnalysis>(F); | |||
5030 | return {F.getParent()->getDataLayout(), TLI, DT, AC}; | |||
5031 | } | |||
5032 | template const SimplifyQuery getBestSimplifyQuery(AnalysisManager<Function> &, | |||
5033 | Function &); | |||
5034 | } |