Mercurial > hg > Members > tobaru > cbc > CbC_llvm
annotate lib/Transforms/Scalar/TailRecursionElimination.cpp @ 82:e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
author | Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp> |
---|---|
date | Mon, 27 Oct 2014 20:20:15 +0900 |
parents | 67baa08a3894 |
children | 5e5d649e25d2 |
rev | line source |
---|---|
0 | 1 //===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===// |
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 transforms calls of the current function (self recursion) followed | |
11 // by a return instruction with a branch to the entry of the function, creating | |
12 // a loop. This pass also implements the following extensions to the basic | |
13 // algorithm: | |
14 // | |
15 // 1. Trivial instructions between the call and return do not prevent the | |
16 // transformation from taking place, though currently the analysis cannot | |
17 // support moving any really useful instructions (only dead ones). | |
18 // 2. This pass transforms functions that are prevented from being tail | |
19 // recursive by an associative and commutative expression to use an | |
20 // accumulator variable, thus compiling the typical naive factorial or | |
21 // 'fib' implementation into efficient code. | |
22 // 3. TRE is performed if the function returns void, if the return | |
23 // returns the result returned by the call, or if the function returns a | |
24 // run-time constant on all exits from the function. It is possible, though | |
25 // unlikely, that the return returns something else (like constant 0), and | |
26 // can still be TRE'd. It can be TRE'd if ALL OTHER return instructions in | |
27 // the function return the exact same value. | |
28 // 4. If it can prove that callees do not access their caller stack frame, | |
29 // they are marked as eligible for tail call elimination (by the code | |
30 // generator). | |
31 // | |
32 // There are several improvements that could be made: | |
33 // | |
34 // 1. If the function has any alloca instructions, these instructions will be | |
35 // moved out of the entry block of the function, causing them to be | |
36 // evaluated each time through the tail recursion. Safely keeping allocas | |
37 // in the entry block requires analysis to proves that the tail-called | |
38 // function does not read or write the stack object. | |
39 // 2. Tail recursion is only performed if the call immediately precedes the | |
40 // return instruction. It's possible that there could be a jump between | |
41 // the call and the return. | |
42 // 3. There can be intervening operations between the call and the return that | |
43 // prevent the TRE from occurring. For example, there could be GEP's and | |
44 // stores to memory that will not be read or written by the call. This | |
45 // requires some substantial analysis (such as with DSA) to prove safe to | |
46 // move ahead of the call, but doing so could allow many more TREs to be | |
47 // performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark. | |
48 // 4. The algorithm we use to detect if callees access their caller stack | |
49 // frames is very primitive. | |
50 // | |
51 //===----------------------------------------------------------------------===// | |
52 | |
53 #include "llvm/Transforms/Scalar.h" | |
54 #include "llvm/ADT/STLExtras.h" | |
55 #include "llvm/ADT/SmallPtrSet.h" | |
56 #include "llvm/ADT/Statistic.h" | |
57 #include "llvm/Analysis/CaptureTracking.h" | |
77 | 58 #include "llvm/Analysis/CFG.h" |
0 | 59 #include "llvm/Analysis/InlineCost.h" |
60 #include "llvm/Analysis/InstructionSimplify.h" | |
61 #include "llvm/Analysis/Loads.h" | |
62 #include "llvm/Analysis/TargetTransformInfo.h" | |
77 | 63 #include "llvm/IR/CFG.h" |
64 #include "llvm/IR/CallSite.h" | |
0 | 65 #include "llvm/IR/Constants.h" |
66 #include "llvm/IR/DerivedTypes.h" | |
77 | 67 #include "llvm/IR/DiagnosticInfo.h" |
0 | 68 #include "llvm/IR/Function.h" |
69 #include "llvm/IR/Instructions.h" | |
70 #include "llvm/IR/IntrinsicInst.h" | |
71 #include "llvm/IR/Module.h" | |
77 | 72 #include "llvm/IR/ValueHandle.h" |
0 | 73 #include "llvm/Pass.h" |
74 #include "llvm/Support/Debug.h" | |
75 #include "llvm/Support/raw_ostream.h" | |
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |
77 #include "llvm/Transforms/Utils/Local.h" | |
78 using namespace llvm; | |
79 | |
77 | 80 #define DEBUG_TYPE "tailcallelim" |
81 | |
0 | 82 STATISTIC(NumEliminated, "Number of tail calls removed"); |
83 STATISTIC(NumRetDuped, "Number of return duplicated"); | |
84 STATISTIC(NumAccumAdded, "Number of accumulators introduced"); | |
85 | |
86 namespace { | |
87 struct TailCallElim : public FunctionPass { | |
88 const TargetTransformInfo *TTI; | |
89 | |
90 static char ID; // Pass identification, replacement for typeid | |
91 TailCallElim() : FunctionPass(ID) { | |
92 initializeTailCallElimPass(*PassRegistry::getPassRegistry()); | |
93 } | |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
94 #ifndef noCbC |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
95 TailCallElim(bool f) : FunctionPass(ID) { |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
96 initializeTailCallElimPass(*PassRegistry::getPassRegistry()); |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
97 onlyForCbC = f; |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
98 } |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
99 #endif |
0 | 100 |
77 | 101 void getAnalysisUsage(AnalysisUsage &AU) const override; |
0 | 102 |
77 | 103 bool runOnFunction(Function &F) override; |
0 | 104 |
105 private: | |
77 | 106 bool runTRE(Function &F); |
107 bool markTails(Function &F, bool &AllCallsAreTailCalls); | |
108 | |
0 | 109 CallInst *FindTRECandidate(Instruction *I, |
110 bool CannotTailCallElimCallsMarkedTail); | |
111 bool EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret, | |
112 BasicBlock *&OldEntry, | |
113 bool &TailCallsAreMarkedTail, | |
114 SmallVectorImpl<PHINode *> &ArgumentPHIs, | |
115 bool CannotTailCallElimCallsMarkedTail); | |
116 bool FoldReturnAndProcessPred(BasicBlock *BB, | |
117 ReturnInst *Ret, BasicBlock *&OldEntry, | |
118 bool &TailCallsAreMarkedTail, | |
119 SmallVectorImpl<PHINode *> &ArgumentPHIs, | |
120 bool CannotTailCallElimCallsMarkedTail); | |
121 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry, | |
122 bool &TailCallsAreMarkedTail, | |
123 SmallVectorImpl<PHINode *> &ArgumentPHIs, | |
124 bool CannotTailCallElimCallsMarkedTail); | |
125 bool CanMoveAboveCall(Instruction *I, CallInst *CI); | |
126 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI); | |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
127 #ifndef noCbC |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
128 bool onlyForCbC; |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
129 public: |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
130 bool isOnlyForCbC(); |
82
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
131 bool markTailToCodeSegments(Function &F); |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
132 #endif |
0 | 133 }; |
134 } | |
135 | |
136 char TailCallElim::ID = 0; | |
137 INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim", | |
138 "Tail Call Elimination", false, false) | |
139 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) | |
140 INITIALIZE_PASS_END(TailCallElim, "tailcallelim", | |
141 "Tail Call Elimination", false, false) | |
142 | |
143 // Public interface to the TailCallElimination pass | |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
144 #ifndef noCbC |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
145 // Public interface to the TailCallElimination pass |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
146 FunctionPass *llvm::createTailCallEliminationPass(bool isOnlyForCbC) { |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
147 return new TailCallElim(isOnlyForCbC); |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
148 } |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
149 #else |
0 | 150 FunctionPass *llvm::createTailCallEliminationPass() { |
151 return new TailCallElim(); | |
152 } | |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
153 #endif |
0 | 154 |
155 void TailCallElim::getAnalysisUsage(AnalysisUsage &AU) const { | |
156 AU.addRequired<TargetTransformInfo>(); | |
157 } | |
158 | |
77 | 159 /// \brief Scan the specified function for alloca instructions. |
160 /// If it contains any dynamic allocas, returns false. | |
161 static bool CanTRE(Function &F) { | |
162 // Because of PR962, we don't TRE dynamic allocas. | |
163 for (auto &BB : F) { | |
164 for (auto &I : BB) { | |
165 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) { | |
166 if (!AI->isStaticAlloca()) | |
167 return false; | |
168 } | |
169 } | |
170 } | |
0 | 171 |
77 | 172 return true; |
173 } | |
174 | |
175 bool TailCallElim::runOnFunction(Function &F) { | |
176 if (skipOptnoneFunction(F)) | |
177 return false; | |
178 | |
179 bool AllCallsAreTailCalls = false; | |
180 bool Modified = markTails(F, AllCallsAreTailCalls); | |
181 if (AllCallsAreTailCalls) | |
182 Modified |= runTRE(F); | |
183 return Modified; | |
0 | 184 } |
185 | |
186 namespace { | |
77 | 187 struct AllocaDerivedValueTracker { |
188 // Start at a root value and walk its use-def chain to mark calls that use the | |
189 // value or a derived value in AllocaUsers, and places where it may escape in | |
190 // EscapePoints. | |
191 void walk(Value *Root) { | |
192 SmallVector<Use *, 32> Worklist; | |
193 SmallPtrSet<Use *, 32> Visited; | |
0 | 194 |
77 | 195 auto AddUsesToWorklist = [&](Value *V) { |
196 for (auto &U : V->uses()) { | |
197 if (!Visited.insert(&U)) | |
198 continue; | |
199 Worklist.push_back(&U); | |
200 } | |
201 }; | |
202 | |
203 AddUsesToWorklist(Root); | |
204 | |
205 while (!Worklist.empty()) { | |
206 Use *U = Worklist.pop_back_val(); | |
207 Instruction *I = cast<Instruction>(U->getUser()); | |
0 | 208 |
77 | 209 switch (I->getOpcode()) { |
210 case Instruction::Call: | |
211 case Instruction::Invoke: { | |
212 CallSite CS(I); | |
213 bool IsNocapture = !CS.isCallee(U) && | |
214 CS.doesNotCapture(CS.getArgumentNo(U)); | |
215 callUsesLocalStack(CS, IsNocapture); | |
216 if (IsNocapture) { | |
217 // If the alloca-derived argument is passed in as nocapture, then it | |
218 // can't propagate to the call's return. That would be capturing. | |
219 continue; | |
220 } | |
221 break; | |
222 } | |
223 case Instruction::Load: { | |
224 // The result of a load is not alloca-derived (unless an alloca has | |
225 // otherwise escaped, but this is a local analysis). | |
226 continue; | |
227 } | |
228 case Instruction::Store: { | |
229 if (U->getOperandNo() == 0) | |
230 EscapePoints.insert(I); | |
231 continue; // Stores have no users to analyze. | |
232 } | |
233 case Instruction::BitCast: | |
234 case Instruction::GetElementPtr: | |
235 case Instruction::PHI: | |
236 case Instruction::Select: | |
237 case Instruction::AddrSpaceCast: | |
238 break; | |
239 default: | |
240 EscapePoints.insert(I); | |
241 break; | |
242 } | |
243 | |
244 AddUsesToWorklist(I); | |
245 } | |
246 } | |
247 | |
248 void callUsesLocalStack(CallSite CS, bool IsNocapture) { | |
249 // Add it to the list of alloca users. | |
250 AllocaUsers.insert(CS.getInstruction()); | |
251 | |
252 // If it's nocapture then it can't capture this alloca. | |
253 if (IsNocapture) | |
254 return; | |
255 | |
256 // If it can write to memory, it can leak the alloca value. | |
257 if (!CS.onlyReadsMemory()) | |
258 EscapePoints.insert(CS.getInstruction()); | |
259 } | |
260 | |
261 SmallPtrSet<Instruction *, 32> AllocaUsers; | |
262 SmallPtrSet<Instruction *, 32> EscapePoints; | |
263 }; | |
264 } | |
265 | |
266 bool TailCallElim::markTails(Function &F, bool &AllCallsAreTailCalls) { | |
267 if (F.callsFunctionThatReturnsTwice()) | |
268 return false; | |
269 AllCallsAreTailCalls = true; | |
270 | |
271 // The local stack holds all alloca instructions and all byval arguments. | |
272 AllocaDerivedValueTracker Tracker; | |
273 for (Argument &Arg : F.args()) { | |
274 if (Arg.hasByValAttr()) | |
275 Tracker.walk(&Arg); | |
276 } | |
277 for (auto &BB : F) { | |
278 for (auto &I : BB) | |
279 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) | |
280 Tracker.walk(AI); | |
0 | 281 } |
282 | |
77 | 283 bool Modified = false; |
284 | |
82
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
285 #ifndef noCbC |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
286 if (F.getReturnType()->is__CodeTy()) |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
287 Modified = markTailToCodeSegments(F); |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
288 #endif |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
289 |
77 | 290 // Track whether a block is reachable after an alloca has escaped. Blocks that |
291 // contain the escaping instruction will be marked as being visited without an | |
292 // escaped alloca, since that is how the block began. | |
293 enum VisitType { | |
294 UNVISITED, | |
295 UNESCAPED, | |
296 ESCAPED | |
297 }; | |
298 DenseMap<BasicBlock *, VisitType> Visited; | |
299 | |
300 // We propagate the fact that an alloca has escaped from block to successor. | |
301 // Visit the blocks that are propagating the escapedness first. To do this, we | |
302 // maintain two worklists. | |
303 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped; | |
304 | |
305 // We may enter a block and visit it thinking that no alloca has escaped yet, | |
306 // then see an escape point and go back around a loop edge and come back to | |
307 // the same block twice. Because of this, we defer setting tail on calls when | |
308 // we first encounter them in a block. Every entry in this list does not | |
309 // statically use an alloca via use-def chain analysis, but may find an alloca | |
310 // through other means if the block turns out to be reachable after an escape | |
311 // point. | |
312 SmallVector<CallInst *, 32> DeferredTails; | |
313 | |
314 BasicBlock *BB = &F.getEntryBlock(); | |
315 VisitType Escaped = UNESCAPED; | |
316 do { | |
317 for (auto &I : *BB) { | |
318 if (Tracker.EscapePoints.count(&I)) | |
319 Escaped = ESCAPED; | |
320 | |
321 CallInst *CI = dyn_cast<CallInst>(&I); | |
322 if (!CI || CI->isTailCall()) | |
323 continue; | |
324 | |
325 if (CI->doesNotAccessMemory()) { | |
326 // A call to a readnone function whose arguments are all things computed | |
327 // outside this function can be marked tail. Even if you stored the | |
328 // alloca address into a global, a readnone function can't load the | |
329 // global anyhow. | |
330 // | |
331 // Note that this runs whether we know an alloca has escaped or not. If | |
332 // it has, then we can't trust Tracker.AllocaUsers to be accurate. | |
333 bool SafeToTail = true; | |
334 for (auto &Arg : CI->arg_operands()) { | |
335 if (isa<Constant>(Arg.getUser())) | |
336 continue; | |
337 if (Argument *A = dyn_cast<Argument>(Arg.getUser())) | |
338 if (!A->hasByValAttr()) | |
339 continue; | |
340 SafeToTail = false; | |
341 break; | |
342 } | |
343 if (SafeToTail) { | |
344 emitOptimizationRemark( | |
345 F.getContext(), "tailcallelim", F, CI->getDebugLoc(), | |
346 "marked this readnone call a tail call candidate"); | |
347 CI->setTailCall(); | |
348 Modified = true; | |
349 continue; | |
350 } | |
351 } | |
352 | |
353 if (Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) { | |
354 DeferredTails.push_back(CI); | |
355 } else { | |
356 AllCallsAreTailCalls = false; | |
357 } | |
358 } | |
359 | |
360 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) { | |
361 auto &State = Visited[SuccBB]; | |
362 if (State < Escaped) { | |
363 State = Escaped; | |
364 if (State == ESCAPED) | |
365 WorklistEscaped.push_back(SuccBB); | |
366 else | |
367 WorklistUnescaped.push_back(SuccBB); | |
368 } | |
369 } | |
370 | |
371 if (!WorklistEscaped.empty()) { | |
372 BB = WorklistEscaped.pop_back_val(); | |
373 Escaped = ESCAPED; | |
374 } else { | |
375 BB = nullptr; | |
376 while (!WorklistUnescaped.empty()) { | |
377 auto *NextBB = WorklistUnescaped.pop_back_val(); | |
378 if (Visited[NextBB] == UNESCAPED) { | |
379 BB = NextBB; | |
380 Escaped = UNESCAPED; | |
381 break; | |
382 } | |
383 } | |
384 } | |
385 } while (BB); | |
386 | |
387 for (CallInst *CI : DeferredTails) { | |
388 if (Visited[CI->getParent()] != ESCAPED) { | |
389 // If the escape point was part way through the block, calls after the | |
390 // escape point wouldn't have been put into DeferredTails. | |
391 emitOptimizationRemark(F.getContext(), "tailcallelim", F, | |
392 CI->getDebugLoc(), | |
393 "marked this call a tail call candidate"); | |
394 CI->setTailCall(); | |
395 Modified = true; | |
396 } else { | |
397 AllCallsAreTailCalls = false; | |
398 } | |
0 | 399 } |
400 | |
77 | 401 return Modified; |
402 } | |
0 | 403 |
77 | 404 bool TailCallElim::runTRE(Function &F) { |
0 | 405 // If this function is a varargs function, we won't be able to PHI the args |
406 // right, so don't even try to convert it... | |
407 if (F.getFunctionType()->isVarArg()) return false; | |
408 | |
409 TTI = &getAnalysis<TargetTransformInfo>(); | |
77 | 410 BasicBlock *OldEntry = nullptr; |
0 | 411 bool TailCallsAreMarkedTail = false; |
412 SmallVector<PHINode*, 8> ArgumentPHIs; | |
413 bool MadeChange = false; | |
414 | |
415 // CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls | |
416 // marked with the 'tail' attribute, because doing so would cause the stack | |
417 // size to increase (real TRE would deallocate variable sized allocas, TRE | |
418 // doesn't). | |
77 | 419 bool CanTRETailMarkedCall = CanTRE(F); |
0 | 420 |
77 | 421 // Change any tail recursive calls to loops. |
0 | 422 // |
423 // FIXME: The code generator produces really bad code when an 'escaping | |
424 // alloca' is changed from being a static alloca to being a dynamic alloca. | |
425 // Until this is resolved, disable this transformation if that would ever | |
426 // happen. This bug is PR962. | |
77 | 427 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { |
428 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) { | |
429 bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail, | |
430 ArgumentPHIs, !CanTRETailMarkedCall); | |
431 if (!Change && BB->getFirstNonPHIOrDbg() == Ret) | |
432 Change = FoldReturnAndProcessPred(BB, Ret, OldEntry, | |
433 TailCallsAreMarkedTail, ArgumentPHIs, | |
434 !CanTRETailMarkedCall); | |
435 MadeChange |= Change; | |
0 | 436 } |
437 } | |
438 | |
439 // If we eliminated any tail recursions, it's possible that we inserted some | |
440 // silly PHI nodes which just merge an initial value (the incoming operand) | |
441 // with themselves. Check to see if we did and clean up our mess if so. This | |
442 // occurs when a function passes an argument straight through to its tail | |
443 // call. | |
77 | 444 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) { |
445 PHINode *PN = ArgumentPHIs[i]; | |
0 | 446 |
77 | 447 // If the PHI Node is a dynamic constant, replace it with the value it is. |
448 if (Value *PNV = SimplifyInstruction(PN)) { | |
449 PN->replaceAllUsesWith(PNV); | |
450 PN->eraseFromParent(); | |
0 | 451 } |
452 } | |
453 | |
454 return MadeChange; | |
455 } | |
456 | |
457 | |
458 /// CanMoveAboveCall - Return true if it is safe to move the specified | |
459 /// instruction from after the call to before the call, assuming that all | |
460 /// instructions between the call and this instruction are movable. | |
461 /// | |
462 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { | |
463 // FIXME: We can move load/store/call/free instructions above the call if the | |
464 // call does not mod/ref the memory location being processed. | |
465 if (I->mayHaveSideEffects()) // This also handles volatile loads. | |
466 return false; | |
467 | |
468 if (LoadInst *L = dyn_cast<LoadInst>(I)) { | |
469 // Loads may always be moved above calls without side effects. | |
470 if (CI->mayHaveSideEffects()) { | |
471 // Non-volatile loads may be moved above a call with side effects if it | |
472 // does not write to memory and the load provably won't trap. | |
473 // FIXME: Writes to memory only matter if they may alias the pointer | |
474 // being loaded from. | |
475 if (CI->mayWriteToMemory() || | |
476 !isSafeToLoadUnconditionally(L->getPointerOperand(), L, | |
477 L->getAlignment())) | |
478 return false; | |
479 } | |
480 } | |
481 | |
482 // Otherwise, if this is a side-effect free instruction, check to make sure | |
483 // that it does not use the return value of the call. If it doesn't use the | |
484 // return value of the call, it must only use things that are defined before | |
485 // the call, or movable instructions between the call and the instruction | |
486 // itself. | |
487 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) | |
488 if (I->getOperand(i) == CI) | |
489 return false; | |
490 return true; | |
491 } | |
492 | |
493 // isDynamicConstant - Return true if the specified value is the same when the | |
494 // return would exit as it was when the initial iteration of the recursive | |
495 // function was executed. | |
496 // | |
497 // We currently handle static constants and arguments that are not modified as | |
498 // part of the recursion. | |
499 // | |
500 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) { | |
501 if (isa<Constant>(V)) return true; // Static constants are always dyn consts | |
502 | |
503 // Check to see if this is an immutable argument, if so, the value | |
504 // will be available to initialize the accumulator. | |
505 if (Argument *Arg = dyn_cast<Argument>(V)) { | |
506 // Figure out which argument number this is... | |
507 unsigned ArgNo = 0; | |
508 Function *F = CI->getParent()->getParent(); | |
509 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI) | |
510 ++ArgNo; | |
511 | |
512 // If we are passing this argument into call as the corresponding | |
513 // argument operand, then the argument is dynamically constant. | |
514 // Otherwise, we cannot transform this function safely. | |
515 if (CI->getArgOperand(ArgNo) == Arg) | |
516 return true; | |
517 } | |
518 | |
519 // Switch cases are always constant integers. If the value is being switched | |
520 // on and the return is only reachable from one of its cases, it's | |
521 // effectively constant. | |
522 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor()) | |
523 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator())) | |
524 if (SI->getCondition() == V) | |
525 return SI->getDefaultDest() != RI->getParent(); | |
526 | |
527 // Not a constant or immutable argument, we can't safely transform. | |
528 return false; | |
529 } | |
530 | |
531 // getCommonReturnValue - Check to see if the function containing the specified | |
532 // tail call consistently returns the same runtime-constant value at all exit | |
533 // points except for IgnoreRI. If so, return the returned value. | |
534 // | |
535 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) { | |
536 Function *F = CI->getParent()->getParent(); | |
77 | 537 Value *ReturnedValue = nullptr; |
0 | 538 |
539 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) { | |
540 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()); | |
77 | 541 if (RI == nullptr || RI == IgnoreRI) continue; |
0 | 542 |
543 // We can only perform this transformation if the value returned is | |
544 // evaluatable at the start of the initial invocation of the function, | |
545 // instead of at the end of the evaluation. | |
546 // | |
547 Value *RetOp = RI->getOperand(0); | |
548 if (!isDynamicConstant(RetOp, CI, RI)) | |
77 | 549 return nullptr; |
0 | 550 |
551 if (ReturnedValue && RetOp != ReturnedValue) | |
77 | 552 return nullptr; // Cannot transform if differing values are returned. |
0 | 553 ReturnedValue = RetOp; |
554 } | |
555 return ReturnedValue; | |
556 } | |
557 | |
558 /// CanTransformAccumulatorRecursion - If the specified instruction can be | |
559 /// transformed using accumulator recursion elimination, return the constant | |
560 /// which is the start of the accumulator value. Otherwise return null. | |
561 /// | |
562 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, | |
563 CallInst *CI) { | |
77 | 564 if (!I->isAssociative() || !I->isCommutative()) return nullptr; |
0 | 565 assert(I->getNumOperands() == 2 && |
566 "Associative/commutative operations should have 2 args!"); | |
567 | |
568 // Exactly one operand should be the result of the call instruction. | |
569 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) || | |
570 (I->getOperand(0) != CI && I->getOperand(1) != CI)) | |
77 | 571 return nullptr; |
0 | 572 |
573 // The only user of this instruction we allow is a single return instruction. | |
77 | 574 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back())) |
575 return nullptr; | |
0 | 576 |
577 // Ok, now we have to check all of the other return instructions in this | |
578 // function. If they return non-constants or differing values, then we cannot | |
579 // transform the function safely. | |
77 | 580 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI); |
0 | 581 } |
582 | |
583 static Instruction *FirstNonDbg(BasicBlock::iterator I) { | |
584 while (isa<DbgInfoIntrinsic>(I)) | |
585 ++I; | |
586 return &*I; | |
587 } | |
588 | |
589 CallInst* | |
590 TailCallElim::FindTRECandidate(Instruction *TI, | |
591 bool CannotTailCallElimCallsMarkedTail) { | |
592 BasicBlock *BB = TI->getParent(); | |
593 Function *F = BB->getParent(); | |
594 | |
595 if (&BB->front() == TI) // Make sure there is something before the terminator. | |
77 | 596 return nullptr; |
0 | 597 |
598 // Scan backwards from the return, checking to see if there is a tail call in | |
599 // this block. If so, set CI to it. | |
77 | 600 CallInst *CI = nullptr; |
0 | 601 BasicBlock::iterator BBI = TI; |
602 while (true) { | |
603 CI = dyn_cast<CallInst>(BBI); | |
604 if (CI && CI->getCalledFunction() == F) | |
605 break; | |
606 | |
607 if (BBI == BB->begin()) | |
77 | 608 return nullptr; // Didn't find a potential tail call. |
0 | 609 --BBI; |
610 } | |
611 | |
612 // If this call is marked as a tail call, and if there are dynamic allocas in | |
613 // the function, we cannot perform this optimization. | |
614 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) | |
77 | 615 return nullptr; |
0 | 616 |
617 // As a special case, detect code like this: | |
618 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call | |
619 // and disable this xform in this case, because the code generator will | |
620 // lower the call to fabs into inline code. | |
621 if (BB == &F->getEntryBlock() && | |
622 FirstNonDbg(BB->front()) == CI && | |
77 | 623 FirstNonDbg(std::next(BB->begin())) == TI && |
0 | 624 CI->getCalledFunction() && |
625 !TTI->isLoweredToCall(CI->getCalledFunction())) { | |
626 // A single-block function with just a call and a return. Check that | |
627 // the arguments match. | |
628 CallSite::arg_iterator I = CallSite(CI).arg_begin(), | |
629 E = CallSite(CI).arg_end(); | |
630 Function::arg_iterator FI = F->arg_begin(), | |
631 FE = F->arg_end(); | |
632 for (; I != E && FI != FE; ++I, ++FI) | |
633 if (*I != &*FI) break; | |
634 if (I == E && FI == FE) | |
77 | 635 return nullptr; |
0 | 636 } |
637 | |
638 return CI; | |
639 } | |
640 | |
641 bool TailCallElim::EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret, | |
642 BasicBlock *&OldEntry, | |
643 bool &TailCallsAreMarkedTail, | |
644 SmallVectorImpl<PHINode *> &ArgumentPHIs, | |
645 bool CannotTailCallElimCallsMarkedTail) { | |
646 // If we are introducing accumulator recursion to eliminate operations after | |
647 // the call instruction that are both associative and commutative, the initial | |
648 // value for the accumulator is placed in this variable. If this value is set | |
649 // then we actually perform accumulator recursion elimination instead of | |
650 // simple tail recursion elimination. If the operation is an LLVM instruction | |
651 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then | |
652 // we are handling the case when the return instruction returns a constant C | |
653 // which is different to the constant returned by other return instructions | |
654 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a | |
655 // special case of accumulator recursion, the operation being "return C". | |
77 | 656 Value *AccumulatorRecursionEliminationInitVal = nullptr; |
657 Instruction *AccumulatorRecursionInstr = nullptr; | |
0 | 658 |
659 // Ok, we found a potential tail call. We can currently only transform the | |
660 // tail call if all of the instructions between the call and the return are | |
661 // movable to above the call itself, leaving the call next to the return. | |
662 // Check that this is the case now. | |
663 BasicBlock::iterator BBI = CI; | |
664 for (++BBI; &*BBI != Ret; ++BBI) { | |
665 if (CanMoveAboveCall(BBI, CI)) continue; | |
666 | |
667 // If we can't move the instruction above the call, it might be because it | |
668 // is an associative and commutative operation that could be transformed | |
669 // using accumulator recursion elimination. Check to see if this is the | |
670 // case, and if so, remember the initial accumulator value for later. | |
671 if ((AccumulatorRecursionEliminationInitVal = | |
672 CanTransformAccumulatorRecursion(BBI, CI))) { | |
673 // Yes, this is accumulator recursion. Remember which instruction | |
674 // accumulates. | |
675 AccumulatorRecursionInstr = BBI; | |
676 } else { | |
677 return false; // Otherwise, we cannot eliminate the tail recursion! | |
678 } | |
679 } | |
680 | |
681 // We can only transform call/return pairs that either ignore the return value | |
682 // of the call and return void, ignore the value of the call and return a | |
683 // constant, return the value returned by the tail call, or that are being | |
684 // accumulator recursion variable eliminated. | |
685 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI && | |
686 !isa<UndefValue>(Ret->getReturnValue()) && | |
77 | 687 AccumulatorRecursionEliminationInitVal == nullptr && |
688 !getCommonReturnValue(nullptr, CI)) { | |
0 | 689 // One case remains that we are able to handle: the current return |
690 // instruction returns a constant, and all other return instructions | |
691 // return a different constant. | |
692 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret)) | |
693 return false; // Current return instruction does not return a constant. | |
694 // Check that all other return instructions return a common constant. If | |
695 // so, record it in AccumulatorRecursionEliminationInitVal. | |
696 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI); | |
697 if (!AccumulatorRecursionEliminationInitVal) | |
698 return false; | |
699 } | |
700 | |
701 BasicBlock *BB = Ret->getParent(); | |
702 Function *F = BB->getParent(); | |
703 | |
77 | 704 emitOptimizationRemark(F->getContext(), "tailcallelim", *F, CI->getDebugLoc(), |
705 "transforming tail recursion to loop"); | |
706 | |
0 | 707 // OK! We can transform this tail call. If this is the first one found, |
708 // create the new entry block, allowing us to branch back to the old entry. | |
77 | 709 if (!OldEntry) { |
0 | 710 OldEntry = &F->getEntryBlock(); |
711 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry); | |
712 NewEntry->takeName(OldEntry); | |
713 OldEntry->setName("tailrecurse"); | |
714 BranchInst::Create(OldEntry, NewEntry); | |
715 | |
716 // If this tail call is marked 'tail' and if there are any allocas in the | |
717 // entry block, move them up to the new entry block. | |
718 TailCallsAreMarkedTail = CI->isTailCall(); | |
719 if (TailCallsAreMarkedTail) | |
720 // Move all fixed sized allocas from OldEntry to NewEntry. | |
721 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), | |
722 NEBI = NewEntry->begin(); OEBI != E; ) | |
723 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) | |
724 if (isa<ConstantInt>(AI->getArraySize())) | |
725 AI->moveBefore(NEBI); | |
726 | |
727 // Now that we have created a new block, which jumps to the entry | |
728 // block, insert a PHI node for each argument of the function. | |
729 // For now, we initialize each PHI to only have the real arguments | |
730 // which are passed in. | |
731 Instruction *InsertPos = OldEntry->begin(); | |
732 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); | |
733 I != E; ++I) { | |
734 PHINode *PN = PHINode::Create(I->getType(), 2, | |
735 I->getName() + ".tr", InsertPos); | |
736 I->replaceAllUsesWith(PN); // Everyone use the PHI node now! | |
737 PN->addIncoming(I, NewEntry); | |
738 ArgumentPHIs.push_back(PN); | |
739 } | |
740 } | |
741 | |
742 // If this function has self recursive calls in the tail position where some | |
743 // are marked tail and some are not, only transform one flavor or another. We | |
744 // have to choose whether we move allocas in the entry block to the new entry | |
745 // block or not, so we can't make a good choice for both. NOTE: We could do | |
746 // slightly better here in the case that the function has no entry block | |
747 // allocas. | |
748 if (TailCallsAreMarkedTail && !CI->isTailCall()) | |
749 return false; | |
750 | |
751 // Ok, now that we know we have a pseudo-entry block WITH all of the | |
752 // required PHI nodes, add entries into the PHI node for the actual | |
753 // parameters passed into the tail-recursive call. | |
754 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) | |
755 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB); | |
756 | |
757 // If we are introducing an accumulator variable to eliminate the recursion, | |
758 // do so now. Note that we _know_ that no subsequent tail recursion | |
759 // eliminations will happen on this function because of the way the | |
760 // accumulator recursion predicate is set up. | |
761 // | |
762 if (AccumulatorRecursionEliminationInitVal) { | |
763 Instruction *AccRecInstr = AccumulatorRecursionInstr; | |
764 // Start by inserting a new PHI node for the accumulator. | |
765 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry); | |
766 PHINode *AccPN = | |
767 PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(), | |
768 std::distance(PB, PE) + 1, | |
769 "accumulator.tr", OldEntry->begin()); | |
770 | |
771 // Loop over all of the predecessors of the tail recursion block. For the | |
772 // real entry into the function we seed the PHI with the initial value, | |
773 // computed earlier. For any other existing branches to this block (due to | |
774 // other tail recursions eliminated) the accumulator is not modified. | |
775 // Because we haven't added the branch in the current block to OldEntry yet, | |
776 // it will not show up as a predecessor. | |
777 for (pred_iterator PI = PB; PI != PE; ++PI) { | |
778 BasicBlock *P = *PI; | |
779 if (P == &F->getEntryBlock()) | |
780 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P); | |
781 else | |
782 AccPN->addIncoming(AccPN, P); | |
783 } | |
784 | |
785 if (AccRecInstr) { | |
786 // Add an incoming argument for the current block, which is computed by | |
787 // our associative and commutative accumulator instruction. | |
788 AccPN->addIncoming(AccRecInstr, BB); | |
789 | |
790 // Next, rewrite the accumulator recursion instruction so that it does not | |
791 // use the result of the call anymore, instead, use the PHI node we just | |
792 // inserted. | |
793 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); | |
794 } else { | |
795 // Add an incoming argument for the current block, which is just the | |
796 // constant returned by the current return instruction. | |
797 AccPN->addIncoming(Ret->getReturnValue(), BB); | |
798 } | |
799 | |
800 // Finally, rewrite any return instructions in the program to return the PHI | |
801 // node instead of the "initval" that they do currently. This loop will | |
802 // actually rewrite the return value we are destroying, but that's ok. | |
803 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) | |
804 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) | |
805 RI->setOperand(0, AccPN); | |
806 ++NumAccumAdded; | |
807 } | |
808 | |
809 // Now that all of the PHI nodes are in place, remove the call and | |
810 // ret instructions, replacing them with an unconditional branch. | |
811 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret); | |
812 NewBI->setDebugLoc(CI->getDebugLoc()); | |
813 | |
814 BB->getInstList().erase(Ret); // Remove return. | |
815 BB->getInstList().erase(CI); // Remove call. | |
816 ++NumEliminated; | |
817 return true; | |
818 } | |
819 | |
820 bool TailCallElim::FoldReturnAndProcessPred(BasicBlock *BB, | |
821 ReturnInst *Ret, BasicBlock *&OldEntry, | |
822 bool &TailCallsAreMarkedTail, | |
823 SmallVectorImpl<PHINode *> &ArgumentPHIs, | |
824 bool CannotTailCallElimCallsMarkedTail) { | |
825 bool Change = false; | |
826 | |
827 // If the return block contains nothing but the return and PHI's, | |
828 // there might be an opportunity to duplicate the return in its | |
829 // predecessors and perform TRC there. Look for predecessors that end | |
830 // in unconditional branch and recursive call(s). | |
831 SmallVector<BranchInst*, 8> UncondBranchPreds; | |
832 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { | |
833 BasicBlock *Pred = *PI; | |
834 TerminatorInst *PTI = Pred->getTerminator(); | |
835 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) | |
836 if (BI->isUnconditional()) | |
837 UncondBranchPreds.push_back(BI); | |
838 } | |
839 | |
840 while (!UncondBranchPreds.empty()) { | |
841 BranchInst *BI = UncondBranchPreds.pop_back_val(); | |
842 BasicBlock *Pred = BI->getParent(); | |
843 if (CallInst *CI = FindTRECandidate(BI, CannotTailCallElimCallsMarkedTail)){ | |
844 DEBUG(dbgs() << "FOLDING: " << *BB | |
845 << "INTO UNCOND BRANCH PRED: " << *Pred); | |
846 EliminateRecursiveTailCall(CI, FoldReturnIntoUncondBranch(Ret, BB, Pred), | |
847 OldEntry, TailCallsAreMarkedTail, ArgumentPHIs, | |
848 CannotTailCallElimCallsMarkedTail); | |
849 ++NumRetDuped; | |
850 Change = true; | |
851 } | |
852 } | |
853 | |
854 return Change; | |
855 } | |
856 | |
857 bool | |
858 TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, | |
859 bool &TailCallsAreMarkedTail, | |
860 SmallVectorImpl<PHINode *> &ArgumentPHIs, | |
861 bool CannotTailCallElimCallsMarkedTail) { | |
862 CallInst *CI = FindTRECandidate(Ret, CannotTailCallElimCallsMarkedTail); | |
863 if (!CI) | |
864 return false; | |
865 | |
866 return EliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail, | |
867 ArgumentPHIs, | |
868 CannotTailCallElimCallsMarkedTail); | |
869 } | |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
870 |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
871 #ifndef noCbC |
82
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
872 bool TailCallElim::isOnlyForCbC(){ |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
873 return onlyForCbC; |
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
874 } |
82
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
875 |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
876 bool TailCallElim::markTailToCodeSegments(Function &F){ |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
877 bool Modified = false; |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
878 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
879 for (auto &I : *BB) { |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
880 CallInst *CI = dyn_cast<CallInst>(&I); |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
881 Function* Called; |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
882 if (CI) |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
883 Called = CI->getCalledFunction(); |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
884 else |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
885 continue; |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
886 // We should touch only code segment call. |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
887 if (Called && Called->getReturnType()->is__CodeTy()) { |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
888 CI->setTailCall(); |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
889 Modified = true; |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
890 } |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
891 } |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
892 } |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
893 return Modified; |
e218c19a8176
markTailToCodeSegments ensure code segments are marked tail
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
80
diff
changeset
|
894 } |
35
503e14e069e4
update to LLVM 3.5
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
895 #endif |