File: | lib/Transforms/Scalar/JumpThreading.cpp |
Location: | line 968, column 7 |
Description: | Called C++ object pointer is null |
1 | //===- JumpThreading.cpp - Thread control through conditional blocks ------===// | |||
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 the Jump Threading pass. | |||
11 | // | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #include "llvm/Transforms/Scalar.h" | |||
15 | #include "llvm/ADT/DenseMap.h" | |||
16 | #include "llvm/ADT/DenseSet.h" | |||
17 | #include "llvm/ADT/STLExtras.h" | |||
18 | #include "llvm/ADT/SmallPtrSet.h" | |||
19 | #include "llvm/ADT/SmallSet.h" | |||
20 | #include "llvm/ADT/Statistic.h" | |||
21 | #include "llvm/Analysis/CFG.h" | |||
22 | #include "llvm/Analysis/ConstantFolding.h" | |||
23 | #include "llvm/Analysis/InstructionSimplify.h" | |||
24 | #include "llvm/Analysis/LazyValueInfo.h" | |||
25 | #include "llvm/Analysis/Loads.h" | |||
26 | #include "llvm/IR/DataLayout.h" | |||
27 | #include "llvm/IR/IntrinsicInst.h" | |||
28 | #include "llvm/IR/LLVMContext.h" | |||
29 | #include "llvm/IR/Metadata.h" | |||
30 | #include "llvm/IR/ValueHandle.h" | |||
31 | #include "llvm/Pass.h" | |||
32 | #include "llvm/Support/CommandLine.h" | |||
33 | #include "llvm/Support/Debug.h" | |||
34 | #include "llvm/Support/raw_ostream.h" | |||
35 | #include "llvm/Target/TargetLibraryInfo.h" | |||
36 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
37 | #include "llvm/Transforms/Utils/Local.h" | |||
38 | #include "llvm/Transforms/Utils/SSAUpdater.h" | |||
39 | using namespace llvm; | |||
40 | ||||
41 | #define DEBUG_TYPE"jump-threading" "jump-threading" | |||
42 | ||||
43 | STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = { "jump-threading", "Number of jumps threaded" , 0, 0 }; | |||
44 | STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = { "jump-threading", "Number of terminators folded" , 0, 0 }; | |||
45 | STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi")static llvm::Statistic NumDupes = { "jump-threading", "Number of branch blocks duplicated to eliminate phi" , 0, 0 }; | |||
46 | ||||
47 | static cl::opt<unsigned> | |||
48 | BBDuplicateThreshold("jump-threading-threshold", | |||
49 | cl::desc("Max block size to duplicate for jump threading"), | |||
50 | cl::init(6), cl::Hidden); | |||
51 | ||||
52 | namespace { | |||
53 | // These are at global scope so static functions can use them too. | |||
54 | typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo; | |||
55 | typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy; | |||
56 | ||||
57 | // This is used to keep track of what kind of constant we're currently hoping | |||
58 | // to find. | |||
59 | enum ConstantPreference { | |||
60 | WantInteger, | |||
61 | WantBlockAddress | |||
62 | }; | |||
63 | ||||
64 | /// This pass performs 'jump threading', which looks at blocks that have | |||
65 | /// multiple predecessors and multiple successors. If one or more of the | |||
66 | /// predecessors of the block can be proven to always jump to one of the | |||
67 | /// successors, we forward the edge from the predecessor to the successor by | |||
68 | /// duplicating the contents of this block. | |||
69 | /// | |||
70 | /// An example of when this can occur is code like this: | |||
71 | /// | |||
72 | /// if () { ... | |||
73 | /// X = 4; | |||
74 | /// } | |||
75 | /// if (X < 3) { | |||
76 | /// | |||
77 | /// In this case, the unconditional branch at the end of the first if can be | |||
78 | /// revectored to the false side of the second if. | |||
79 | /// | |||
80 | class JumpThreading : public FunctionPass { | |||
81 | const DataLayout *DL; | |||
82 | TargetLibraryInfo *TLI; | |||
83 | LazyValueInfo *LVI; | |||
84 | #ifdef NDEBUG | |||
85 | SmallPtrSet<BasicBlock*, 16> LoopHeaders; | |||
86 | #else | |||
87 | SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; | |||
88 | #endif | |||
89 | DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet; | |||
90 | ||||
91 | unsigned BBDupThreshold; | |||
92 | ||||
93 | // RAII helper for updating the recursion stack. | |||
94 | struct RecursionSetRemover { | |||
95 | DenseSet<std::pair<Value*, BasicBlock*> > &TheSet; | |||
96 | std::pair<Value*, BasicBlock*> ThePair; | |||
97 | ||||
98 | RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S, | |||
99 | std::pair<Value*, BasicBlock*> P) | |||
100 | : TheSet(S), ThePair(P) { } | |||
101 | ||||
102 | ~RecursionSetRemover() { | |||
103 | TheSet.erase(ThePair); | |||
104 | } | |||
105 | }; | |||
106 | public: | |||
107 | static char ID; // Pass identification | |||
108 | JumpThreading(int T = -1) : FunctionPass(ID) { | |||
109 | BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T); | |||
110 | initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); | |||
111 | } | |||
112 | ||||
113 | bool runOnFunction(Function &F) override; | |||
114 | ||||
115 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
116 | AU.addRequired<LazyValueInfo>(); | |||
117 | AU.addPreserved<LazyValueInfo>(); | |||
118 | AU.addRequired<TargetLibraryInfo>(); | |||
119 | } | |||
120 | ||||
121 | void FindLoopHeaders(Function &F); | |||
122 | bool ProcessBlock(BasicBlock *BB); | |||
123 | bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs, | |||
124 | BasicBlock *SuccBB); | |||
125 | bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, | |||
126 | const SmallVectorImpl<BasicBlock *> &PredBBs); | |||
127 | ||||
128 | bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, | |||
129 | PredValueInfo &Result, | |||
130 | ConstantPreference Preference, | |||
131 | Instruction *CxtI = nullptr); | |||
132 | bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, | |||
133 | ConstantPreference Preference, | |||
134 | Instruction *CxtI = nullptr); | |||
135 | ||||
136 | bool ProcessBranchOnPHI(PHINode *PN); | |||
137 | bool ProcessBranchOnXOR(BinaryOperator *BO); | |||
138 | ||||
139 | bool SimplifyPartiallyRedundantLoad(LoadInst *LI); | |||
140 | bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB); | |||
141 | }; | |||
142 | } | |||
143 | ||||
144 | char JumpThreading::ID = 0; | |||
145 | INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void* initializeJumpThreadingPassOnce(PassRegistry & Registry) { | |||
146 | "Jump Threading", false, false)static void* initializeJumpThreadingPassOnce(PassRegistry & Registry) { | |||
147 | INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)initializeLazyValueInfoPass(Registry); | |||
148 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)initializeTargetLibraryInfoPass(Registry); | |||
149 | INITIALIZE_PASS_END(JumpThreading, "jump-threading",PassInfo *PI = new PassInfo("Jump Threading", "jump-threading" , & JumpThreading ::ID, PassInfo::NormalCtor_t(callDefaultCtor < JumpThreading >), false, false); Registry.registerPass (*PI, true); return PI; } void llvm::initializeJumpThreadingPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializeJumpThreadingPassOnce( Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150, &initialized); } | |||
150 | "Jump Threading", false, false)PassInfo *PI = new PassInfo("Jump Threading", "jump-threading" , & JumpThreading ::ID, PassInfo::NormalCtor_t(callDefaultCtor < JumpThreading >), false, false); Registry.registerPass (*PI, true); return PI; } void llvm::initializeJumpThreadingPass (PassRegistry &Registry) { static volatile sys::cas_flag initialized = 0; sys::cas_flag old_val = sys::CompareAndSwap(&initialized , 1, 0); if (old_val == 0) { initializeJumpThreadingPassOnce( Registry); sys::MemoryFence(); AnnotateIgnoreWritesBegin("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150); AnnotateHappensBefore("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150, &initialized); initialized = 2; AnnotateIgnoreWritesEnd ("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150); } else { sys::cas_flag tmp = initialized; sys::MemoryFence (); while (tmp != 2) { tmp = initialized; sys::MemoryFence(); } } AnnotateHappensAfter("/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 150, &initialized); } | |||
151 | ||||
152 | // Public interface to the Jump Threading pass | |||
153 | FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); } | |||
154 | ||||
155 | /// runOnFunction - Top level algorithm. | |||
156 | /// | |||
157 | bool JumpThreading::runOnFunction(Function &F) { | |||
158 | if (skipOptnoneFunction(F)) | |||
159 | return false; | |||
160 | ||||
161 | DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "Jump threading on function '" << F.getName() << "'\n"; } } while (0); | |||
162 | DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); | |||
163 | DL = DLP ? &DLP->getDataLayout() : nullptr; | |||
164 | TLI = &getAnalysis<TargetLibraryInfo>(); | |||
165 | LVI = &getAnalysis<LazyValueInfo>(); | |||
166 | ||||
167 | // Remove unreachable blocks from function as they may result in infinite | |||
168 | // loop. We do threading if we found something profitable. Jump threading a | |||
169 | // branch can create other opportunities. If these opportunities form a cycle | |||
170 | // i.e. if any jump treading is undoing previous threading in the path, then | |||
171 | // we will loop forever. We take care of this issue by not jump threading for | |||
172 | // back edges. This works for normal cases but not for unreachable blocks as | |||
173 | // they may have cycle with no back edge. | |||
174 | removeUnreachableBlocks(F); | |||
175 | ||||
176 | FindLoopHeaders(F); | |||
177 | ||||
178 | bool Changed, EverChanged = false; | |||
179 | do { | |||
180 | Changed = false; | |||
181 | for (Function::iterator I = F.begin(), E = F.end(); I != E;) { | |||
182 | BasicBlock *BB = I; | |||
183 | // Thread all of the branches we can over this block. | |||
184 | while (ProcessBlock(BB)) | |||
185 | Changed = true; | |||
186 | ||||
187 | ++I; | |||
188 | ||||
189 | // If the block is trivially dead, zap it. This eliminates the successor | |||
190 | // edges which simplifies the CFG. | |||
191 | if (pred_begin(BB) == pred_end(BB) && | |||
192 | BB != &BB->getParent()->getEntryBlock()) { | |||
193 | DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " JT: Deleting dead block '" << BB->getName() << "' with terminator: " << *BB->getTerminator() << '\n'; } } while (0) | |||
194 | << "' with terminator: " << *BB->getTerminator() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " JT: Deleting dead block '" << BB->getName() << "' with terminator: " << *BB->getTerminator() << '\n'; } } while (0); | |||
195 | LoopHeaders.erase(BB); | |||
196 | LVI->eraseBlock(BB); | |||
197 | DeleteDeadBlock(BB); | |||
198 | Changed = true; | |||
199 | continue; | |||
200 | } | |||
201 | ||||
202 | BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); | |||
203 | ||||
204 | // Can't thread an unconditional jump, but if the block is "almost | |||
205 | // empty", we can replace uses of it with uses of the successor and make | |||
206 | // this dead. | |||
207 | if (BI && BI->isUnconditional() && | |||
208 | BB != &BB->getParent()->getEntryBlock() && | |||
209 | // If the terminator is the only non-phi instruction, try to nuke it. | |||
210 | BB->getFirstNonPHIOrDbg()->isTerminator()) { | |||
211 | // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the | |||
212 | // block, we have to make sure it isn't in the LoopHeaders set. We | |||
213 | // reinsert afterward if needed. | |||
214 | bool ErasedFromLoopHeaders = LoopHeaders.erase(BB); | |||
215 | BasicBlock *Succ = BI->getSuccessor(0); | |||
216 | ||||
217 | // FIXME: It is always conservatively correct to drop the info | |||
218 | // for a block even if it doesn't get erased. This isn't totally | |||
219 | // awesome, but it allows us to use AssertingVH to prevent nasty | |||
220 | // dangling pointer issues within LazyValueInfo. | |||
221 | LVI->eraseBlock(BB); | |||
222 | if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) { | |||
223 | Changed = true; | |||
224 | // If we deleted BB and BB was the header of a loop, then the | |||
225 | // successor is now the header of the loop. | |||
226 | BB = Succ; | |||
227 | } | |||
228 | ||||
229 | if (ErasedFromLoopHeaders) | |||
230 | LoopHeaders.insert(BB); | |||
231 | } | |||
232 | } | |||
233 | EverChanged |= Changed; | |||
234 | } while (Changed); | |||
235 | ||||
236 | LoopHeaders.clear(); | |||
237 | return EverChanged; | |||
238 | } | |||
239 | ||||
240 | /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to | |||
241 | /// thread across it. Stop scanning the block when passing the threshold. | |||
242 | static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB, | |||
243 | unsigned Threshold) { | |||
244 | /// Ignore PHI nodes, these will be flattened when duplication happens. | |||
245 | BasicBlock::const_iterator I = BB->getFirstNonPHI(); | |||
246 | ||||
247 | // FIXME: THREADING will delete values that are just used to compute the | |||
248 | // branch, so they shouldn't count against the duplication cost. | |||
249 | ||||
250 | // Sum up the cost of each instruction until we get to the terminator. Don't | |||
251 | // include the terminator because the copy won't include it. | |||
252 | unsigned Size = 0; | |||
253 | for (; !isa<TerminatorInst>(I); ++I) { | |||
254 | ||||
255 | // Stop scanning the block if we've reached the threshold. | |||
256 | if (Size > Threshold) | |||
257 | return Size; | |||
258 | ||||
259 | // Debugger intrinsics don't incur code size. | |||
260 | if (isa<DbgInfoIntrinsic>(I)) continue; | |||
261 | ||||
262 | // If this is a pointer->pointer bitcast, it is free. | |||
263 | if (isa<BitCastInst>(I) && I->getType()->isPointerTy()) | |||
264 | continue; | |||
265 | ||||
266 | // All other instructions count for at least one unit. | |||
267 | ++Size; | |||
268 | ||||
269 | // Calls are more expensive. If they are non-intrinsic calls, we model them | |||
270 | // as having cost of 4. If they are a non-vector intrinsic, we model them | |||
271 | // as having cost of 2 total, and if they are a vector intrinsic, we model | |||
272 | // them as having cost 1. | |||
273 | if (const CallInst *CI = dyn_cast<CallInst>(I)) { | |||
274 | if (CI->cannotDuplicate()) | |||
275 | // Blocks with NoDuplicate are modelled as having infinite cost, so they | |||
276 | // are never duplicated. | |||
277 | return ~0U; | |||
278 | else if (!isa<IntrinsicInst>(CI)) | |||
279 | Size += 3; | |||
280 | else if (!CI->getType()->isVectorTy()) | |||
281 | Size += 1; | |||
282 | } | |||
283 | } | |||
284 | ||||
285 | // Threading through a switch statement is particularly profitable. If this | |||
286 | // block ends in a switch, decrease its cost to make it more likely to happen. | |||
287 | if (isa<SwitchInst>(I)) | |||
288 | Size = Size > 6 ? Size-6 : 0; | |||
289 | ||||
290 | // The same holds for indirect branches, but slightly more so. | |||
291 | if (isa<IndirectBrInst>(I)) | |||
292 | Size = Size > 8 ? Size-8 : 0; | |||
293 | ||||
294 | return Size; | |||
295 | } | |||
296 | ||||
297 | /// FindLoopHeaders - We do not want jump threading to turn proper loop | |||
298 | /// structures into irreducible loops. Doing this breaks up the loop nesting | |||
299 | /// hierarchy and pessimizes later transformations. To prevent this from | |||
300 | /// happening, we first have to find the loop headers. Here we approximate this | |||
301 | /// by finding targets of backedges in the CFG. | |||
302 | /// | |||
303 | /// Note that there definitely are cases when we want to allow threading of | |||
304 | /// edges across a loop header. For example, threading a jump from outside the | |||
305 | /// loop (the preheader) to an exit block of the loop is definitely profitable. | |||
306 | /// It is also almost always profitable to thread backedges from within the loop | |||
307 | /// to exit blocks, and is often profitable to thread backedges to other blocks | |||
308 | /// within the loop (forming a nested loop). This simple analysis is not rich | |||
309 | /// enough to track all of these properties and keep it up-to-date as the CFG | |||
310 | /// mutates, so we don't allow any of these transformations. | |||
311 | /// | |||
312 | void JumpThreading::FindLoopHeaders(Function &F) { | |||
313 | SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; | |||
314 | FindFunctionBackedges(F, Edges); | |||
315 | ||||
316 | for (unsigned i = 0, e = Edges.size(); i != e; ++i) | |||
317 | LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); | |||
318 | } | |||
319 | ||||
320 | /// getKnownConstant - Helper method to determine if we can thread over a | |||
321 | /// terminator with the given value as its condition, and if so what value to | |||
322 | /// use for that. What kind of value this is depends on whether we want an | |||
323 | /// integer or a block address, but an undef is always accepted. | |||
324 | /// Returns null if Val is null or not an appropriate constant. | |||
325 | static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { | |||
326 | if (!Val) | |||
327 | return nullptr; | |||
328 | ||||
329 | // Undef is "known" enough. | |||
330 | if (UndefValue *U = dyn_cast<UndefValue>(Val)) | |||
331 | return U; | |||
332 | ||||
333 | if (Preference == WantBlockAddress) | |||
334 | return dyn_cast<BlockAddress>(Val->stripPointerCasts()); | |||
335 | ||||
336 | return dyn_cast<ConstantInt>(Val); | |||
337 | } | |||
338 | ||||
339 | /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see | |||
340 | /// if we can infer that the value is a known ConstantInt/BlockAddress or undef | |||
341 | /// in any of our predecessors. If so, return the known list of value and pred | |||
342 | /// BB in the result vector. | |||
343 | /// | |||
344 | /// This returns true if there were any known values. | |||
345 | /// | |||
346 | bool JumpThreading:: | |||
347 | ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result, | |||
348 | ConstantPreference Preference, | |||
349 | Instruction *CxtI) { | |||
350 | // This method walks up use-def chains recursively. Because of this, we could | |||
351 | // get into an infinite loop going around loops in the use-def chain. To | |||
352 | // prevent this, keep track of what (value, block) pairs we've already visited | |||
353 | // and terminate the search if we loop back to them | |||
354 | if (!RecursionSet.insert(std::make_pair(V, BB)).second) | |||
355 | return false; | |||
356 | ||||
357 | // An RAII help to remove this pair from the recursion set once the recursion | |||
358 | // stack pops back out again. | |||
359 | RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB)); | |||
360 | ||||
361 | // If V is a constant, then it is known in all predecessors. | |||
362 | if (Constant *KC = getKnownConstant(V, Preference)) { | |||
363 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) | |||
364 | Result.push_back(std::make_pair(KC, *PI)); | |||
365 | ||||
366 | return true; | |||
367 | } | |||
368 | ||||
369 | // If V is a non-instruction value, or an instruction in a different block, | |||
370 | // then it can't be derived from a PHI. | |||
371 | Instruction *I = dyn_cast<Instruction>(V); | |||
372 | if (!I || I->getParent() != BB) { | |||
373 | ||||
374 | // Okay, if this is a live-in value, see if it has a known value at the end | |||
375 | // of any of our predecessors. | |||
376 | // | |||
377 | // FIXME: This should be an edge property, not a block end property. | |||
378 | /// TODO: Per PR2563, we could infer value range information about a | |||
379 | /// predecessor based on its terminator. | |||
380 | // | |||
381 | // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if | |||
382 | // "I" is a non-local compare-with-a-constant instruction. This would be | |||
383 | // able to handle value inequalities better, for example if the compare is | |||
384 | // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. | |||
385 | // Perhaps getConstantOnEdge should be smart enough to do this? | |||
386 | ||||
387 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { | |||
388 | BasicBlock *P = *PI; | |||
389 | // If the value is known by LazyValueInfo to be a constant in a | |||
390 | // predecessor, use that information to try to thread this block. | |||
391 | Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI); | |||
392 | if (Constant *KC = getKnownConstant(PredCst, Preference)) | |||
393 | Result.push_back(std::make_pair(KC, P)); | |||
394 | } | |||
395 | ||||
396 | return !Result.empty(); | |||
397 | } | |||
398 | ||||
399 | /// If I is a PHI node, then we know the incoming values for any constants. | |||
400 | if (PHINode *PN = dyn_cast<PHINode>(I)) { | |||
401 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |||
402 | Value *InVal = PN->getIncomingValue(i); | |||
403 | if (Constant *KC = getKnownConstant(InVal, Preference)) { | |||
404 | Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); | |||
405 | } else { | |||
406 | Constant *CI = LVI->getConstantOnEdge(InVal, | |||
407 | PN->getIncomingBlock(i), | |||
408 | BB, CxtI); | |||
409 | if (Constant *KC = getKnownConstant(CI, Preference)) | |||
410 | Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i))); | |||
411 | } | |||
412 | } | |||
413 | ||||
414 | return !Result.empty(); | |||
415 | } | |||
416 | ||||
417 | PredValueInfoTy LHSVals, RHSVals; | |||
418 | ||||
419 | // Handle some boolean conditions. | |||
420 | if (I->getType()->getPrimitiveSizeInBits() == 1) { | |||
421 | assert(Preference == WantInteger && "One-bit non-integer type?")((Preference == WantInteger && "One-bit non-integer type?" ) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"One-bit non-integer type?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 421, __PRETTY_FUNCTION__)); | |||
422 | // X | true -> true | |||
423 | // X & false -> false | |||
424 | if (I->getOpcode() == Instruction::Or || | |||
425 | I->getOpcode() == Instruction::And) { | |||
426 | ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, | |||
427 | WantInteger, CxtI); | |||
428 | ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals, | |||
429 | WantInteger, CxtI); | |||
430 | ||||
431 | if (LHSVals.empty() && RHSVals.empty()) | |||
432 | return false; | |||
433 | ||||
434 | ConstantInt *InterestingVal; | |||
435 | if (I->getOpcode() == Instruction::Or) | |||
436 | InterestingVal = ConstantInt::getTrue(I->getContext()); | |||
437 | else | |||
438 | InterestingVal = ConstantInt::getFalse(I->getContext()); | |||
439 | ||||
440 | SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; | |||
441 | ||||
442 | // Scan for the sentinel. If we find an undef, force it to the | |||
443 | // interesting value: x|undef -> true and x&undef -> false. | |||
444 | for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) | |||
445 | if (LHSVals[i].first == InterestingVal || | |||
446 | isa<UndefValue>(LHSVals[i].first)) { | |||
447 | Result.push_back(LHSVals[i]); | |||
448 | Result.back().first = InterestingVal; | |||
449 | LHSKnownBBs.insert(LHSVals[i].second); | |||
450 | } | |||
451 | for (unsigned i = 0, e = RHSVals.size(); i != e; ++i) | |||
452 | if (RHSVals[i].first == InterestingVal || | |||
453 | isa<UndefValue>(RHSVals[i].first)) { | |||
454 | // If we already inferred a value for this block on the LHS, don't | |||
455 | // re-add it. | |||
456 | if (!LHSKnownBBs.count(RHSVals[i].second)) { | |||
457 | Result.push_back(RHSVals[i]); | |||
458 | Result.back().first = InterestingVal; | |||
459 | } | |||
460 | } | |||
461 | ||||
462 | return !Result.empty(); | |||
463 | } | |||
464 | ||||
465 | // Handle the NOT form of XOR. | |||
466 | if (I->getOpcode() == Instruction::Xor && | |||
467 | isa<ConstantInt>(I->getOperand(1)) && | |||
468 | cast<ConstantInt>(I->getOperand(1))->isOne()) { | |||
469 | ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result, | |||
470 | WantInteger, CxtI); | |||
471 | if (Result.empty()) | |||
472 | return false; | |||
473 | ||||
474 | // Invert the known values. | |||
475 | for (unsigned i = 0, e = Result.size(); i != e; ++i) | |||
476 | Result[i].first = ConstantExpr::getNot(Result[i].first); | |||
477 | ||||
478 | return true; | |||
479 | } | |||
480 | ||||
481 | // Try to simplify some other binary operator values. | |||
482 | } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { | |||
483 | assert(Preference != WantBlockAddress((Preference != WantBlockAddress && "A binary operator creating a block address?" ) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 484, __PRETTY_FUNCTION__)) | |||
484 | && "A binary operator creating a block address?")((Preference != WantBlockAddress && "A binary operator creating a block address?" ) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 484, __PRETTY_FUNCTION__)); | |||
485 | if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { | |||
486 | PredValueInfoTy LHSVals; | |||
487 | ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals, | |||
488 | WantInteger, CxtI); | |||
489 | ||||
490 | // Try to use constant folding to simplify the binary operator. | |||
491 | for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { | |||
492 | Constant *V = LHSVals[i].first; | |||
493 | Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); | |||
494 | ||||
495 | if (Constant *KC = getKnownConstant(Folded, WantInteger)) | |||
496 | Result.push_back(std::make_pair(KC, LHSVals[i].second)); | |||
497 | } | |||
498 | } | |||
499 | ||||
500 | return !Result.empty(); | |||
501 | } | |||
502 | ||||
503 | // Handle compare with phi operand, where the PHI is defined in this block. | |||
504 | if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { | |||
505 | assert(Preference == WantInteger && "Compares only produce integers")((Preference == WantInteger && "Compares only produce integers" ) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"Compares only produce integers\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 505, __PRETTY_FUNCTION__)); | |||
506 | PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0)); | |||
507 | if (PN && PN->getParent() == BB) { | |||
508 | // We can do this simplification if any comparisons fold to true or false. | |||
509 | // See if any do. | |||
510 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |||
511 | BasicBlock *PredBB = PN->getIncomingBlock(i); | |||
512 | Value *LHS = PN->getIncomingValue(i); | |||
513 | Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB); | |||
514 | ||||
515 | Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL); | |||
516 | if (!Res) { | |||
517 | if (!isa<Constant>(RHS)) | |||
518 | continue; | |||
519 | ||||
520 | LazyValueInfo::Tristate | |||
521 | ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS, | |||
522 | cast<Constant>(RHS), PredBB, BB, | |||
523 | CxtI ? CxtI : Cmp); | |||
524 | if (ResT == LazyValueInfo::Unknown) | |||
525 | continue; | |||
526 | Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); | |||
527 | } | |||
528 | ||||
529 | if (Constant *KC = getKnownConstant(Res, WantInteger)) | |||
530 | Result.push_back(std::make_pair(KC, PredBB)); | |||
531 | } | |||
532 | ||||
533 | return !Result.empty(); | |||
534 | } | |||
535 | ||||
536 | // If comparing a live-in value against a constant, see if we know the | |||
537 | // live-in value on any predecessors. | |||
538 | if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) { | |||
539 | if (!isa<Instruction>(Cmp->getOperand(0)) || | |||
540 | cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) { | |||
541 | Constant *RHSCst = cast<Constant>(Cmp->getOperand(1)); | |||
542 | ||||
543 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){ | |||
544 | BasicBlock *P = *PI; | |||
545 | // If the value is known by LazyValueInfo to be a constant in a | |||
546 | // predecessor, use that information to try to thread this block. | |||
547 | LazyValueInfo::Tristate Res = | |||
548 | LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0), | |||
549 | RHSCst, P, BB, CxtI ? CxtI : Cmp); | |||
550 | if (Res == LazyValueInfo::Unknown) | |||
551 | continue; | |||
552 | ||||
553 | Constant *ResC = ConstantInt::get(Cmp->getType(), Res); | |||
554 | Result.push_back(std::make_pair(ResC, P)); | |||
555 | } | |||
556 | ||||
557 | return !Result.empty(); | |||
558 | } | |||
559 | ||||
560 | // Try to find a constant value for the LHS of a comparison, | |||
561 | // and evaluate it statically if we can. | |||
562 | if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) { | |||
563 | PredValueInfoTy LHSVals; | |||
564 | ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals, | |||
565 | WantInteger, CxtI); | |||
566 | ||||
567 | for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) { | |||
568 | Constant *V = LHSVals[i].first; | |||
569 | Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(), | |||
570 | V, CmpConst); | |||
571 | if (Constant *KC = getKnownConstant(Folded, WantInteger)) | |||
572 | Result.push_back(std::make_pair(KC, LHSVals[i].second)); | |||
573 | } | |||
574 | ||||
575 | return !Result.empty(); | |||
576 | } | |||
577 | } | |||
578 | } | |||
579 | ||||
580 | if (SelectInst *SI = dyn_cast<SelectInst>(I)) { | |||
581 | // Handle select instructions where at least one operand is a known constant | |||
582 | // and we can figure out the condition value for any predecessor block. | |||
583 | Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); | |||
584 | Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); | |||
585 | PredValueInfoTy Conds; | |||
586 | if ((TrueVal || FalseVal) && | |||
587 | ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds, | |||
588 | WantInteger, CxtI)) { | |||
589 | for (unsigned i = 0, e = Conds.size(); i != e; ++i) { | |||
590 | Constant *Cond = Conds[i].first; | |||
591 | ||||
592 | // Figure out what value to use for the condition. | |||
593 | bool KnownCond; | |||
594 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { | |||
595 | // A known boolean. | |||
596 | KnownCond = CI->isOne(); | |||
597 | } else { | |||
598 | assert(isa<UndefValue>(Cond) && "Unexpected condition value")((isa<UndefValue>(Cond) && "Unexpected condition value" ) ? static_cast<void> (0) : __assert_fail ("isa<UndefValue>(Cond) && \"Unexpected condition value\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 598, __PRETTY_FUNCTION__)); | |||
599 | // Either operand will do, so be sure to pick the one that's a known | |||
600 | // constant. | |||
601 | // FIXME: Do this more cleverly if both values are known constants? | |||
602 | KnownCond = (TrueVal != nullptr); | |||
603 | } | |||
604 | ||||
605 | // See if the select has a known constant value for this predecessor. | |||
606 | if (Constant *Val = KnownCond ? TrueVal : FalseVal) | |||
607 | Result.push_back(std::make_pair(Val, Conds[i].second)); | |||
608 | } | |||
609 | ||||
610 | return !Result.empty(); | |||
611 | } | |||
612 | } | |||
613 | ||||
614 | // If all else fails, see if LVI can figure out a constant value for us. | |||
615 | Constant *CI = LVI->getConstant(V, BB, CxtI); | |||
616 | if (Constant *KC = getKnownConstant(CI, Preference)) { | |||
617 | for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) | |||
618 | Result.push_back(std::make_pair(KC, *PI)); | |||
619 | } | |||
620 | ||||
621 | return !Result.empty(); | |||
622 | } | |||
623 | ||||
624 | ||||
625 | ||||
626 | /// GetBestDestForBranchOnUndef - If we determine that the specified block ends | |||
627 | /// in an undefined jump, decide which block is best to revector to. | |||
628 | /// | |||
629 | /// Since we can pick an arbitrary destination, we pick the successor with the | |||
630 | /// fewest predecessors. This should reduce the in-degree of the others. | |||
631 | /// | |||
632 | static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) { | |||
633 | TerminatorInst *BBTerm = BB->getTerminator(); | |||
634 | unsigned MinSucc = 0; | |||
635 | BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); | |||
636 | // Compute the successor with the minimum number of predecessors. | |||
637 | unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); | |||
638 | for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { | |||
639 | TestBB = BBTerm->getSuccessor(i); | |||
640 | unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); | |||
641 | if (NumPreds < MinNumPreds) { | |||
642 | MinSucc = i; | |||
643 | MinNumPreds = NumPreds; | |||
644 | } | |||
645 | } | |||
646 | ||||
647 | return MinSucc; | |||
648 | } | |||
649 | ||||
650 | static bool hasAddressTakenAndUsed(BasicBlock *BB) { | |||
651 | if (!BB->hasAddressTaken()) return false; | |||
652 | ||||
653 | // If the block has its address taken, it may be a tree of dead constants | |||
654 | // hanging off of it. These shouldn't keep the block alive. | |||
655 | BlockAddress *BA = BlockAddress::get(BB); | |||
656 | BA->removeDeadConstantUsers(); | |||
657 | return !BA->use_empty(); | |||
658 | } | |||
659 | ||||
660 | /// ProcessBlock - If there are any predecessors whose control can be threaded | |||
661 | /// through to a successor, transform them now. | |||
662 | bool JumpThreading::ProcessBlock(BasicBlock *BB) { | |||
663 | // If the block is trivially dead, just return and let the caller nuke it. | |||
664 | // This simplifies other transformations. | |||
665 | if (pred_begin(BB) == pred_end(BB) && | |||
666 | BB != &BB->getParent()->getEntryBlock()) | |||
667 | return false; | |||
668 | ||||
669 | // If this block has a single predecessor, and if that pred has a single | |||
670 | // successor, merge the blocks. This encourages recursive jump threading | |||
671 | // because now the condition in this block can be threaded through | |||
672 | // predecessors of our predecessor block. | |||
673 | if (BasicBlock *SinglePred = BB->getSinglePredecessor()) { | |||
674 | if (SinglePred->getTerminator()->getNumSuccessors() == 1 && | |||
675 | SinglePred != BB && !hasAddressTakenAndUsed(BB)) { | |||
676 | // If SinglePred was a loop header, BB becomes one. | |||
677 | if (LoopHeaders.erase(SinglePred)) | |||
678 | LoopHeaders.insert(BB); | |||
679 | ||||
680 | LVI->eraseBlock(SinglePred); | |||
681 | MergeBasicBlockIntoOnlyPred(BB); | |||
682 | ||||
683 | return true; | |||
684 | } | |||
685 | } | |||
686 | ||||
687 | // What kind of constant we're looking for. | |||
688 | ConstantPreference Preference = WantInteger; | |||
689 | ||||
690 | // Look to see if the terminator is a conditional branch, switch or indirect | |||
691 | // branch, if not we can't thread it. | |||
692 | Value *Condition; | |||
693 | Instruction *Terminator = BB->getTerminator(); | |||
694 | if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { | |||
695 | // Can't thread an unconditional jump. | |||
696 | if (BI->isUnconditional()) return false; | |||
697 | Condition = BI->getCondition(); | |||
698 | } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { | |||
699 | Condition = SI->getCondition(); | |||
700 | } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { | |||
701 | // Can't thread indirect branch with no successors. | |||
702 | if (IB->getNumSuccessors() == 0) return false; | |||
703 | Condition = IB->getAddress()->stripPointerCasts(); | |||
704 | Preference = WantBlockAddress; | |||
705 | } else { | |||
706 | return false; // Must be an invoke. | |||
707 | } | |||
708 | ||||
709 | // Run constant folding to see if we can reduce the condition to a simple | |||
710 | // constant. | |||
711 | if (Instruction *I = dyn_cast<Instruction>(Condition)) { | |||
712 | Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI); | |||
713 | if (SimpleVal) { | |||
714 | I->replaceAllUsesWith(SimpleVal); | |||
715 | I->eraseFromParent(); | |||
716 | Condition = SimpleVal; | |||
717 | } | |||
718 | } | |||
719 | ||||
720 | // If the terminator is branching on an undef, we can pick any of the | |||
721 | // successors to branch to. Let GetBestDestForJumpOnUndef decide. | |||
722 | if (isa<UndefValue>(Condition)) { | |||
723 | unsigned BestSucc = GetBestDestForJumpOnUndef(BB); | |||
724 | ||||
725 | // Fold the branch/switch. | |||
726 | TerminatorInst *BBTerm = BB->getTerminator(); | |||
727 | for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { | |||
728 | if (i == BestSucc) continue; | |||
729 | BBTerm->getSuccessor(i)->removePredecessor(BB, true); | |||
730 | } | |||
731 | ||||
732 | DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " In block '" << BB->getName() << "' folding undef terminator: " << *BBTerm << '\n'; } } while (0) | |||
733 | << "' folding undef terminator: " << *BBTerm << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " In block '" << BB->getName() << "' folding undef terminator: " << *BBTerm << '\n'; } } while (0); | |||
734 | BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); | |||
735 | BBTerm->eraseFromParent(); | |||
736 | return true; | |||
737 | } | |||
738 | ||||
739 | // If the terminator of this block is branching on a constant, simplify the | |||
740 | // terminator to an unconditional branch. This can occur due to threading in | |||
741 | // other blocks. | |||
742 | if (getKnownConstant(Condition, Preference)) { | |||
743 | DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " In block '" << BB->getName() << "' folding terminator: " << * BB->getTerminator() << '\n'; } } while (0) | |||
744 | << "' folding terminator: " << *BB->getTerminator() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " In block '" << BB->getName() << "' folding terminator: " << * BB->getTerminator() << '\n'; } } while (0); | |||
745 | ++NumFolds; | |||
746 | ConstantFoldTerminator(BB, true); | |||
747 | return true; | |||
748 | } | |||
749 | ||||
750 | Instruction *CondInst = dyn_cast<Instruction>(Condition); | |||
751 | ||||
752 | // All the rest of our checks depend on the condition being an instruction. | |||
753 | if (!CondInst) { | |||
754 | // FIXME: Unify this with code below. | |||
755 | if (ProcessThreadableEdges(Condition, BB, Preference, Terminator)) | |||
756 | return true; | |||
757 | return false; | |||
758 | } | |||
759 | ||||
760 | ||||
761 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { | |||
762 | // For a comparison where the LHS is outside this block, it's possible | |||
763 | // that we've branched on it before. Used LVI to see if we can simplify | |||
764 | // the branch based on that. | |||
765 | BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); | |||
766 | Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); | |||
767 | pred_iterator PI = pred_begin(BB), PE = pred_end(BB); | |||
768 | if (CondBr && CondConst && CondBr->isConditional() && PI != PE && | |||
769 | (!isa<Instruction>(CondCmp->getOperand(0)) || | |||
770 | cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) { | |||
771 | // For predecessor edge, determine if the comparison is true or false | |||
772 | // on that edge. If they're all true or all false, we can simplify the | |||
773 | // branch. | |||
774 | // FIXME: We could handle mixed true/false by duplicating code. | |||
775 | LazyValueInfo::Tristate Baseline = | |||
776 | LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0), | |||
777 | CondConst, *PI, BB, CondCmp); | |||
778 | if (Baseline != LazyValueInfo::Unknown) { | |||
779 | // Check that all remaining incoming values match the first one. | |||
780 | while (++PI != PE) { | |||
781 | LazyValueInfo::Tristate Ret = | |||
782 | LVI->getPredicateOnEdge(CondCmp->getPredicate(), | |||
783 | CondCmp->getOperand(0), CondConst, *PI, BB, | |||
784 | CondCmp); | |||
785 | if (Ret != Baseline) break; | |||
786 | } | |||
787 | ||||
788 | // If we terminated early, then one of the values didn't match. | |||
789 | if (PI == PE) { | |||
790 | unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0; | |||
791 | unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1; | |||
792 | CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); | |||
793 | BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); | |||
794 | CondBr->eraseFromParent(); | |||
795 | return true; | |||
796 | } | |||
797 | } | |||
798 | ||||
799 | } else if (CondBr && CondConst && CondBr->isConditional()) { | |||
800 | // There might be an invairant in the same block with the conditional | |||
801 | // that can determine the predicate. | |||
802 | ||||
803 | LazyValueInfo::Tristate Ret = | |||
804 | LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0), | |||
805 | CondConst, CondCmp); | |||
806 | if (Ret != LazyValueInfo::Unknown) { | |||
807 | unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0; | |||
808 | unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1; | |||
809 | CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true); | |||
810 | BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); | |||
811 | CondBr->eraseFromParent(); | |||
812 | return true; | |||
813 | } | |||
814 | } | |||
815 | ||||
816 | if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB)) | |||
817 | return true; | |||
818 | } | |||
819 | ||||
820 | // Check for some cases that are worth simplifying. Right now we want to look | |||
821 | // for loads that are used by a switch or by the condition for the branch. If | |||
822 | // we see one, check to see if it's partially redundant. If so, insert a PHI | |||
823 | // which can then be used to thread the values. | |||
824 | // | |||
825 | Value *SimplifyValue = CondInst; | |||
826 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) | |||
827 | if (isa<Constant>(CondCmp->getOperand(1))) | |||
828 | SimplifyValue = CondCmp->getOperand(0); | |||
829 | ||||
830 | // TODO: There are other places where load PRE would be profitable, such as | |||
831 | // more complex comparisons. | |||
832 | if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) | |||
833 | if (SimplifyPartiallyRedundantLoad(LI)) | |||
834 | return true; | |||
835 | ||||
836 | ||||
837 | // Handle a variety of cases where we are branching on something derived from | |||
838 | // a PHI node in the current block. If we can prove that any predecessors | |||
839 | // compute a predictable value based on a PHI node, thread those predecessors. | |||
840 | // | |||
841 | if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator)) | |||
842 | return true; | |||
843 | ||||
844 | // If this is an otherwise-unfoldable branch on a phi node in the current | |||
845 | // block, see if we can simplify. | |||
846 | if (PHINode *PN = dyn_cast<PHINode>(CondInst)) | |||
847 | if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) | |||
848 | return ProcessBranchOnPHI(PN); | |||
849 | ||||
850 | ||||
851 | // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. | |||
852 | if (CondInst->getOpcode() == Instruction::Xor && | |||
853 | CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) | |||
854 | return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst)); | |||
855 | ||||
856 | ||||
857 | // TODO: If we have: "br (X > 0)" and we have a predecessor where we know | |||
858 | // "(X == 4)", thread through this block. | |||
859 | ||||
860 | return false; | |||
861 | } | |||
862 | ||||
863 | /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant | |||
864 | /// load instruction, eliminate it by replacing it with a PHI node. This is an | |||
865 | /// important optimization that encourages jump threading, and needs to be run | |||
866 | /// interlaced with other jump threading tasks. | |||
867 | bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { | |||
868 | // Don't hack volatile/atomic loads. | |||
869 | if (!LI->isSimple()) return false; | |||
| ||||
870 | ||||
871 | // If the load is defined in a block with exactly one predecessor, it can't be | |||
872 | // partially redundant. | |||
873 | BasicBlock *LoadBB = LI->getParent(); | |||
874 | if (LoadBB->getSinglePredecessor()) | |||
875 | return false; | |||
876 | ||||
877 | // If the load is defined in a landing pad, it can't be partially redundant, | |||
878 | // because the edges between the invoke and the landing pad cannot have other | |||
879 | // instructions between them. | |||
880 | if (LoadBB->isLandingPad()) | |||
881 | return false; | |||
882 | ||||
883 | Value *LoadedPtr = LI->getOperand(0); | |||
884 | ||||
885 | // If the loaded operand is defined in the LoadBB, it can't be available. | |||
886 | // TODO: Could do simple PHI translation, that would be fun :) | |||
887 | if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) | |||
888 | if (PtrOp->getParent() == LoadBB) | |||
889 | return false; | |||
890 | ||||
891 | // Scan a few instructions up from the load, to see if it is obviously live at | |||
892 | // the entry to its block. | |||
893 | BasicBlock::iterator BBIt = LI; | |||
894 | ||||
895 | if (Value *AvailableVal = | |||
896 | FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) { | |||
897 | // If the value if the load is locally available within the block, just use | |||
898 | // it. This frequently occurs for reg2mem'd allocas. | |||
899 | //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; | |||
900 | ||||
901 | // If the returned value is the load itself, replace with an undef. This can | |||
902 | // only happen in dead loops. | |||
903 | if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType()); | |||
904 | LI->replaceAllUsesWith(AvailableVal); | |||
905 | LI->eraseFromParent(); | |||
906 | return true; | |||
907 | } | |||
908 | ||||
909 | // Otherwise, if we scanned the whole block and got to the top of the block, | |||
910 | // we know the block is locally transparent to the load. If not, something | |||
911 | // might clobber its value. | |||
912 | if (BBIt != LoadBB->begin()) | |||
913 | return false; | |||
914 | ||||
915 | // If all of the loads and stores that feed the value have the same AA tags, | |||
916 | // then we can propagate them onto any newly inserted loads. | |||
917 | AAMDNodes AATags; | |||
918 | LI->getAAMetadata(AATags); | |||
919 | ||||
920 | SmallPtrSet<BasicBlock*, 8> PredsScanned; | |||
921 | typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; | |||
922 | AvailablePredsTy AvailablePreds; | |||
923 | BasicBlock *OneUnavailablePred = nullptr; | |||
924 | ||||
925 | // If we got here, the loaded value is transparent through to the start of the | |||
926 | // block. Check to see if it is available in any of the predecessor blocks. | |||
927 | for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); | |||
928 | PI != PE; ++PI) { | |||
929 | BasicBlock *PredBB = *PI; | |||
930 | ||||
931 | // If we already scanned this predecessor, skip it. | |||
932 | if (!PredsScanned.insert(PredBB)) | |||
933 | continue; | |||
934 | ||||
935 | // Scan the predecessor to see if the value is available in the pred. | |||
936 | BBIt = PredBB->end(); | |||
937 | AAMDNodes ThisAATags; | |||
938 | Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6, | |||
939 | nullptr, &ThisAATags); | |||
940 | if (!PredAvailable) { | |||
941 | OneUnavailablePred = PredBB; | |||
942 | continue; | |||
943 | } | |||
944 | ||||
945 | // If AA tags disagree or are not present, forget about them. | |||
946 | if (AATags != ThisAATags) AATags = AAMDNodes(); | |||
947 | ||||
948 | // If so, this load is partially redundant. Remember this info so that we | |||
949 | // can create a PHI node. | |||
950 | AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); | |||
951 | } | |||
952 | ||||
953 | // If the loaded value isn't available in any predecessor, it isn't partially | |||
954 | // redundant. | |||
955 | if (AvailablePreds.empty()) return false; | |||
956 | ||||
957 | // Okay, the loaded value is available in at least one (and maybe all!) | |||
958 | // predecessors. If the value is unavailable in more than one unique | |||
959 | // predecessor, we want to insert a merge block for those common predecessors. | |||
960 | // This ensures that we only have to insert one reload, thus not increasing | |||
961 | // code size. | |||
962 | BasicBlock *UnavailablePred = nullptr; | |||
963 | ||||
964 | // If there is exactly one predecessor where the value is unavailable, the | |||
965 | // already computed 'OneUnavailablePred' block is it. If it ends in an | |||
966 | // unconditional branch, we know that it isn't a critical edge. | |||
967 | if (PredsScanned.size() == AvailablePreds.size()+1 && | |||
968 | OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { | |||
| ||||
969 | UnavailablePred = OneUnavailablePred; | |||
970 | } else if (PredsScanned.size() != AvailablePreds.size()) { | |||
971 | // Otherwise, we had multiple unavailable predecessors or we had a critical | |||
972 | // edge from the one. | |||
973 | SmallVector<BasicBlock*, 8> PredsToSplit; | |||
974 | SmallPtrSet<BasicBlock*, 8> AvailablePredSet; | |||
975 | ||||
976 | for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) | |||
977 | AvailablePredSet.insert(AvailablePreds[i].first); | |||
978 | ||||
979 | // Add all the unavailable predecessors to the PredsToSplit list. | |||
980 | for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); | |||
981 | PI != PE; ++PI) { | |||
982 | BasicBlock *P = *PI; | |||
983 | // If the predecessor is an indirect goto, we can't split the edge. | |||
984 | if (isa<IndirectBrInst>(P->getTerminator())) | |||
985 | return false; | |||
986 | ||||
987 | if (!AvailablePredSet.count(P)) | |||
988 | PredsToSplit.push_back(P); | |||
989 | } | |||
990 | ||||
991 | // Split them out to their own block. | |||
992 | UnavailablePred = | |||
993 | SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this); | |||
994 | } | |||
995 | ||||
996 | // If the value isn't available in all predecessors, then there will be | |||
997 | // exactly one where it isn't available. Insert a load on that edge and add | |||
998 | // it to the AvailablePreds list. | |||
999 | if (UnavailablePred) { | |||
1000 | assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&((UnavailablePred->getTerminator()->getNumSuccessors() == 1 && "Can't handle critical edge here!") ? static_cast <void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1001, __PRETTY_FUNCTION__)) | |||
1001 | "Can't handle critical edge here!")((UnavailablePred->getTerminator()->getNumSuccessors() == 1 && "Can't handle critical edge here!") ? static_cast <void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1001, __PRETTY_FUNCTION__)); | |||
1002 | LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false, | |||
1003 | LI->getAlignment(), | |||
1004 | UnavailablePred->getTerminator()); | |||
1005 | NewVal->setDebugLoc(LI->getDebugLoc()); | |||
1006 | if (AATags) | |||
1007 | NewVal->setAAMetadata(AATags); | |||
1008 | ||||
1009 | AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); | |||
1010 | } | |||
1011 | ||||
1012 | // Now we know that each predecessor of this block has a value in | |||
1013 | // AvailablePreds, sort them for efficient access as we're walking the preds. | |||
1014 | array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); | |||
1015 | ||||
1016 | // Create a PHI node at the start of the block for the PRE'd load value. | |||
1017 | pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); | |||
1018 | PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "", | |||
1019 | LoadBB->begin()); | |||
1020 | PN->takeName(LI); | |||
1021 | PN->setDebugLoc(LI->getDebugLoc()); | |||
1022 | ||||
1023 | // Insert new entries into the PHI for each predecessor. A single block may | |||
1024 | // have multiple entries here. | |||
1025 | for (pred_iterator PI = PB; PI != PE; ++PI) { | |||
1026 | BasicBlock *P = *PI; | |||
1027 | AvailablePredsTy::iterator I = | |||
1028 | std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), | |||
1029 | std::make_pair(P, (Value*)nullptr)); | |||
1030 | ||||
1031 | assert(I != AvailablePreds.end() && I->first == P &&((I != AvailablePreds.end() && I->first == P && "Didn't find entry for predecessor!") ? static_cast<void> (0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1032, __PRETTY_FUNCTION__)) | |||
1032 | "Didn't find entry for predecessor!")((I != AvailablePreds.end() && I->first == P && "Didn't find entry for predecessor!") ? static_cast<void> (0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1032, __PRETTY_FUNCTION__)); | |||
1033 | ||||
1034 | PN->addIncoming(I->second, I->first); | |||
1035 | } | |||
1036 | ||||
1037 | //cerr << "PRE: " << *LI << *PN << "\n"; | |||
1038 | ||||
1039 | LI->replaceAllUsesWith(PN); | |||
1040 | LI->eraseFromParent(); | |||
1041 | ||||
1042 | return true; | |||
1043 | } | |||
1044 | ||||
1045 | /// FindMostPopularDest - The specified list contains multiple possible | |||
1046 | /// threadable destinations. Pick the one that occurs the most frequently in | |||
1047 | /// the list. | |||
1048 | static BasicBlock * | |||
1049 | FindMostPopularDest(BasicBlock *BB, | |||
1050 | const SmallVectorImpl<std::pair<BasicBlock*, | |||
1051 | BasicBlock*> > &PredToDestList) { | |||
1052 | assert(!PredToDestList.empty())((!PredToDestList.empty()) ? static_cast<void> (0) : __assert_fail ("!PredToDestList.empty()", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1052, __PRETTY_FUNCTION__)); | |||
1053 | ||||
1054 | // Determine popularity. If there are multiple possible destinations, we | |||
1055 | // explicitly choose to ignore 'undef' destinations. We prefer to thread | |||
1056 | // blocks with known and real destinations to threading undef. We'll handle | |||
1057 | // them later if interesting. | |||
1058 | DenseMap<BasicBlock*, unsigned> DestPopularity; | |||
1059 | for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) | |||
1060 | if (PredToDestList[i].second) | |||
1061 | DestPopularity[PredToDestList[i].second]++; | |||
1062 | ||||
1063 | // Find the most popular dest. | |||
1064 | DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin(); | |||
1065 | BasicBlock *MostPopularDest = DPI->first; | |||
1066 | unsigned Popularity = DPI->second; | |||
1067 | SmallVector<BasicBlock*, 4> SamePopularity; | |||
1068 | ||||
1069 | for (++DPI; DPI != DestPopularity.end(); ++DPI) { | |||
1070 | // If the popularity of this entry isn't higher than the popularity we've | |||
1071 | // seen so far, ignore it. | |||
1072 | if (DPI->second < Popularity) | |||
1073 | ; // ignore. | |||
1074 | else if (DPI->second == Popularity) { | |||
1075 | // If it is the same as what we've seen so far, keep track of it. | |||
1076 | SamePopularity.push_back(DPI->first); | |||
1077 | } else { | |||
1078 | // If it is more popular, remember it. | |||
1079 | SamePopularity.clear(); | |||
1080 | MostPopularDest = DPI->first; | |||
1081 | Popularity = DPI->second; | |||
1082 | } | |||
1083 | } | |||
1084 | ||||
1085 | // Okay, now we know the most popular destination. If there is more than one | |||
1086 | // destination, we need to determine one. This is arbitrary, but we need | |||
1087 | // to make a deterministic decision. Pick the first one that appears in the | |||
1088 | // successor list. | |||
1089 | if (!SamePopularity.empty()) { | |||
1090 | SamePopularity.push_back(MostPopularDest); | |||
1091 | TerminatorInst *TI = BB->getTerminator(); | |||
1092 | for (unsigned i = 0; ; ++i) { | |||
1093 | assert(i != TI->getNumSuccessors() && "Didn't find any successor!")((i != TI->getNumSuccessors() && "Didn't find any successor!" ) ? static_cast<void> (0) : __assert_fail ("i != TI->getNumSuccessors() && \"Didn't find any successor!\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1093, __PRETTY_FUNCTION__)); | |||
1094 | ||||
1095 | if (std::find(SamePopularity.begin(), SamePopularity.end(), | |||
1096 | TI->getSuccessor(i)) == SamePopularity.end()) | |||
1097 | continue; | |||
1098 | ||||
1099 | MostPopularDest = TI->getSuccessor(i); | |||
1100 | break; | |||
1101 | } | |||
1102 | } | |||
1103 | ||||
1104 | // Okay, we have finally picked the most popular destination. | |||
1105 | return MostPopularDest; | |||
1106 | } | |||
1107 | ||||
1108 | bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB, | |||
1109 | ConstantPreference Preference, | |||
1110 | Instruction *CxtI) { | |||
1111 | // If threading this would thread across a loop header, don't even try to | |||
1112 | // thread the edge. | |||
1113 | if (LoopHeaders.count(BB)) | |||
1114 | return false; | |||
1115 | ||||
1116 | PredValueInfoTy PredValues; | |||
1117 | if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI)) | |||
1118 | return false; | |||
1119 | ||||
1120 | assert(!PredValues.empty() &&((!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values" ) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1121, __PRETTY_FUNCTION__)) | |||
1121 | "ComputeValueKnownInPredecessors returned true with no values")((!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values" ) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1121, __PRETTY_FUNCTION__)); | |||
1122 | ||||
1123 | DEBUG(dbgs() << "IN BB: " << *BB;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { dbgs () << " BB '" << BB->getName() << "': FOUND condition = " << *PredValues[i].first << " for pred '" << PredValues[i].second->getName() << "'.\n"; }; } } while (0) | |||
1124 | for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { dbgs () << " BB '" << BB->getName() << "': FOUND condition = " << *PredValues[i].first << " for pred '" << PredValues[i].second->getName() << "'.\n"; }; } } while (0) | |||
1125 | dbgs() << " BB '" << BB->getName() << "': FOUND condition = "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { dbgs () << " BB '" << BB->getName() << "': FOUND condition = " << *PredValues[i].first << " for pred '" << PredValues[i].second->getName() << "'.\n"; }; } } while (0) | |||
1126 | << *PredValues[i].firstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { dbgs () << " BB '" << BB->getName() << "': FOUND condition = " << *PredValues[i].first << " for pred '" << PredValues[i].second->getName() << "'.\n"; }; } } while (0) | |||
1127 | << " for pred '" << PredValues[i].second->getName() << "'.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { dbgs () << " BB '" << BB->getName() << "': FOUND condition = " << *PredValues[i].first << " for pred '" << PredValues[i].second->getName() << "'.\n"; }; } } while (0) | |||
1128 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { dbgs () << " BB '" << BB->getName() << "': FOUND condition = " << *PredValues[i].first << " for pred '" << PredValues[i].second->getName() << "'.\n"; }; } } while (0); | |||
1129 | ||||
1130 | // Decide what we want to thread through. Convert our list of known values to | |||
1131 | // a list of known destinations for each pred. This also discards duplicate | |||
1132 | // predecessors and keeps track of the undefined inputs (which are represented | |||
1133 | // as a null dest in the PredToDestList). | |||
1134 | SmallPtrSet<BasicBlock*, 16> SeenPreds; | |||
1135 | SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; | |||
1136 | ||||
1137 | BasicBlock *OnlyDest = nullptr; | |||
1138 | BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; | |||
1139 | ||||
1140 | for (unsigned i = 0, e = PredValues.size(); i != e; ++i) { | |||
1141 | BasicBlock *Pred = PredValues[i].second; | |||
1142 | if (!SeenPreds.insert(Pred)) | |||
1143 | continue; // Duplicate predecessor entry. | |||
1144 | ||||
1145 | // If the predecessor ends with an indirect goto, we can't change its | |||
1146 | // destination. | |||
1147 | if (isa<IndirectBrInst>(Pred->getTerminator())) | |||
1148 | continue; | |||
1149 | ||||
1150 | Constant *Val = PredValues[i].first; | |||
1151 | ||||
1152 | BasicBlock *DestBB; | |||
1153 | if (isa<UndefValue>(Val)) | |||
1154 | DestBB = nullptr; | |||
1155 | else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) | |||
1156 | DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); | |||
1157 | else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { | |||
1158 | DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor(); | |||
1159 | } else { | |||
1160 | assert(isa<IndirectBrInst>(BB->getTerminator())((isa<IndirectBrInst>(BB->getTerminator()) && "Unexpected terminator") ? static_cast<void> (0) : __assert_fail ("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1161, __PRETTY_FUNCTION__)) | |||
1161 | && "Unexpected terminator")((isa<IndirectBrInst>(BB->getTerminator()) && "Unexpected terminator") ? static_cast<void> (0) : __assert_fail ("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1161, __PRETTY_FUNCTION__)); | |||
1162 | DestBB = cast<BlockAddress>(Val)->getBasicBlock(); | |||
1163 | } | |||
1164 | ||||
1165 | // If we have exactly one destination, remember it for efficiency below. | |||
1166 | if (PredToDestList.empty()) | |||
1167 | OnlyDest = DestBB; | |||
1168 | else if (OnlyDest != DestBB) | |||
1169 | OnlyDest = MultipleDestSentinel; | |||
1170 | ||||
1171 | PredToDestList.push_back(std::make_pair(Pred, DestBB)); | |||
1172 | } | |||
1173 | ||||
1174 | // If all edges were unthreadable, we fail. | |||
1175 | if (PredToDestList.empty()) | |||
1176 | return false; | |||
1177 | ||||
1178 | // Determine which is the most common successor. If we have many inputs and | |||
1179 | // this block is a switch, we want to start by threading the batch that goes | |||
1180 | // to the most popular destination first. If we only know about one | |||
1181 | // threadable destination (the common case) we can avoid this. | |||
1182 | BasicBlock *MostPopularDest = OnlyDest; | |||
1183 | ||||
1184 | if (MostPopularDest == MultipleDestSentinel) | |||
1185 | MostPopularDest = FindMostPopularDest(BB, PredToDestList); | |||
1186 | ||||
1187 | // Now that we know what the most popular destination is, factor all | |||
1188 | // predecessors that will jump to it into a single predecessor. | |||
1189 | SmallVector<BasicBlock*, 16> PredsToFactor; | |||
1190 | for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i) | |||
1191 | if (PredToDestList[i].second == MostPopularDest) { | |||
1192 | BasicBlock *Pred = PredToDestList[i].first; | |||
1193 | ||||
1194 | // This predecessor may be a switch or something else that has multiple | |||
1195 | // edges to the block. Factor each of these edges by listing them | |||
1196 | // according to # occurrences in PredsToFactor. | |||
1197 | TerminatorInst *PredTI = Pred->getTerminator(); | |||
1198 | for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i) | |||
1199 | if (PredTI->getSuccessor(i) == BB) | |||
1200 | PredsToFactor.push_back(Pred); | |||
1201 | } | |||
1202 | ||||
1203 | // If the threadable edges are branching on an undefined value, we get to pick | |||
1204 | // the destination that these predecessors should get to. | |||
1205 | if (!MostPopularDest) | |||
1206 | MostPopularDest = BB->getTerminator()-> | |||
1207 | getSuccessor(GetBestDestForJumpOnUndef(BB)); | |||
1208 | ||||
1209 | // Ok, try to thread it! | |||
1210 | return ThreadEdge(BB, PredsToFactor, MostPopularDest); | |||
1211 | } | |||
1212 | ||||
1213 | /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on | |||
1214 | /// a PHI node in the current block. See if there are any simplifications we | |||
1215 | /// can do based on inputs to the phi node. | |||
1216 | /// | |||
1217 | bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) { | |||
1218 | BasicBlock *BB = PN->getParent(); | |||
1219 | ||||
1220 | // TODO: We could make use of this to do it once for blocks with common PHI | |||
1221 | // values. | |||
1222 | SmallVector<BasicBlock*, 1> PredBBs; | |||
1223 | PredBBs.resize(1); | |||
1224 | ||||
1225 | // If any of the predecessor blocks end in an unconditional branch, we can | |||
1226 | // *duplicate* the conditional branch into that block in order to further | |||
1227 | // encourage jump threading and to eliminate cases where we have branch on a | |||
1228 | // phi of an icmp (branch on icmp is much better). | |||
1229 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |||
1230 | BasicBlock *PredBB = PN->getIncomingBlock(i); | |||
1231 | if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) | |||
1232 | if (PredBr->isUnconditional()) { | |||
1233 | PredBBs[0] = PredBB; | |||
1234 | // Try to duplicate BB into PredBB. | |||
1235 | if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs)) | |||
1236 | return true; | |||
1237 | } | |||
1238 | } | |||
1239 | ||||
1240 | return false; | |||
1241 | } | |||
1242 | ||||
1243 | /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on | |||
1244 | /// a xor instruction in the current block. See if there are any | |||
1245 | /// simplifications we can do based on inputs to the xor. | |||
1246 | /// | |||
1247 | bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) { | |||
1248 | BasicBlock *BB = BO->getParent(); | |||
1249 | ||||
1250 | // If either the LHS or RHS of the xor is a constant, don't do this | |||
1251 | // optimization. | |||
1252 | if (isa<ConstantInt>(BO->getOperand(0)) || | |||
1253 | isa<ConstantInt>(BO->getOperand(1))) | |||
1254 | return false; | |||
1255 | ||||
1256 | // If the first instruction in BB isn't a phi, we won't be able to infer | |||
1257 | // anything special about any particular predecessor. | |||
1258 | if (!isa<PHINode>(BB->front())) | |||
1259 | return false; | |||
1260 | ||||
1261 | // If we have a xor as the branch input to this block, and we know that the | |||
1262 | // LHS or RHS of the xor in any predecessor is true/false, then we can clone | |||
1263 | // the condition into the predecessor and fix that value to true, saving some | |||
1264 | // logical ops on that path and encouraging other paths to simplify. | |||
1265 | // | |||
1266 | // This copies something like this: | |||
1267 | // | |||
1268 | // BB: | |||
1269 | // %X = phi i1 [1], [%X'] | |||
1270 | // %Y = icmp eq i32 %A, %B | |||
1271 | // %Z = xor i1 %X, %Y | |||
1272 | // br i1 %Z, ... | |||
1273 | // | |||
1274 | // Into: | |||
1275 | // BB': | |||
1276 | // %Y = icmp ne i32 %A, %B | |||
1277 | // br i1 %Z, ... | |||
1278 | ||||
1279 | PredValueInfoTy XorOpValues; | |||
1280 | bool isLHS = true; | |||
1281 | if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, | |||
1282 | WantInteger, BO)) { | |||
1283 | assert(XorOpValues.empty())((XorOpValues.empty()) ? static_cast<void> (0) : __assert_fail ("XorOpValues.empty()", "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1283, __PRETTY_FUNCTION__)); | |||
1284 | if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, | |||
1285 | WantInteger, BO)) | |||
1286 | return false; | |||
1287 | isLHS = false; | |||
1288 | } | |||
1289 | ||||
1290 | assert(!XorOpValues.empty() &&((!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values" ) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1291, __PRETTY_FUNCTION__)) | |||
1291 | "ComputeValueKnownInPredecessors returned true with no values")((!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values" ) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1291, __PRETTY_FUNCTION__)); | |||
1292 | ||||
1293 | // Scan the information to see which is most popular: true or false. The | |||
1294 | // predecessors can be of the set true, false, or undef. | |||
1295 | unsigned NumTrue = 0, NumFalse = 0; | |||
1296 | for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { | |||
1297 | if (isa<UndefValue>(XorOpValues[i].first)) | |||
1298 | // Ignore undefs for the count. | |||
1299 | continue; | |||
1300 | if (cast<ConstantInt>(XorOpValues[i].first)->isZero()) | |||
1301 | ++NumFalse; | |||
1302 | else | |||
1303 | ++NumTrue; | |||
1304 | } | |||
1305 | ||||
1306 | // Determine which value to split on, true, false, or undef if neither. | |||
1307 | ConstantInt *SplitVal = nullptr; | |||
1308 | if (NumTrue > NumFalse) | |||
1309 | SplitVal = ConstantInt::getTrue(BB->getContext()); | |||
1310 | else if (NumTrue != 0 || NumFalse != 0) | |||
1311 | SplitVal = ConstantInt::getFalse(BB->getContext()); | |||
1312 | ||||
1313 | // Collect all of the blocks that this can be folded into so that we can | |||
1314 | // factor this once and clone it once. | |||
1315 | SmallVector<BasicBlock*, 8> BlocksToFoldInto; | |||
1316 | for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) { | |||
1317 | if (XorOpValues[i].first != SplitVal && | |||
1318 | !isa<UndefValue>(XorOpValues[i].first)) | |||
1319 | continue; | |||
1320 | ||||
1321 | BlocksToFoldInto.push_back(XorOpValues[i].second); | |||
1322 | } | |||
1323 | ||||
1324 | // If we inferred a value for all of the predecessors, then duplication won't | |||
1325 | // help us. However, we can just replace the LHS or RHS with the constant. | |||
1326 | if (BlocksToFoldInto.size() == | |||
1327 | cast<PHINode>(BB->front()).getNumIncomingValues()) { | |||
1328 | if (!SplitVal) { | |||
1329 | // If all preds provide undef, just nuke the xor, because it is undef too. | |||
1330 | BO->replaceAllUsesWith(UndefValue::get(BO->getType())); | |||
1331 | BO->eraseFromParent(); | |||
1332 | } else if (SplitVal->isZero()) { | |||
1333 | // If all preds provide 0, replace the xor with the other input. | |||
1334 | BO->replaceAllUsesWith(BO->getOperand(isLHS)); | |||
1335 | BO->eraseFromParent(); | |||
1336 | } else { | |||
1337 | // If all preds provide 1, set the computed value to 1. | |||
1338 | BO->setOperand(!isLHS, SplitVal); | |||
1339 | } | |||
1340 | ||||
1341 | return true; | |||
1342 | } | |||
1343 | ||||
1344 | // Try to duplicate BB into PredBB. | |||
1345 | return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); | |||
1346 | } | |||
1347 | ||||
1348 | ||||
1349 | /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new | |||
1350 | /// predecessor to the PHIBB block. If it has PHI nodes, add entries for | |||
1351 | /// NewPred using the entries from OldPred (suitably mapped). | |||
1352 | static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, | |||
1353 | BasicBlock *OldPred, | |||
1354 | BasicBlock *NewPred, | |||
1355 | DenseMap<Instruction*, Value*> &ValueMap) { | |||
1356 | for (BasicBlock::iterator PNI = PHIBB->begin(); | |||
1357 | PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) { | |||
1358 | // Ok, we have a PHI node. Figure out what the incoming value was for the | |||
1359 | // DestBlock. | |||
1360 | Value *IV = PN->getIncomingValueForBlock(OldPred); | |||
1361 | ||||
1362 | // Remap the value if necessary. | |||
1363 | if (Instruction *Inst = dyn_cast<Instruction>(IV)) { | |||
1364 | DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); | |||
1365 | if (I != ValueMap.end()) | |||
1366 | IV = I->second; | |||
1367 | } | |||
1368 | ||||
1369 | PN->addIncoming(IV, NewPred); | |||
1370 | } | |||
1371 | } | |||
1372 | ||||
1373 | /// ThreadEdge - We have decided that it is safe and profitable to factor the | |||
1374 | /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB | |||
1375 | /// across BB. Transform the IR to reflect this change. | |||
1376 | bool JumpThreading::ThreadEdge(BasicBlock *BB, | |||
1377 | const SmallVectorImpl<BasicBlock*> &PredBBs, | |||
1378 | BasicBlock *SuccBB) { | |||
1379 | // If threading to the same block as we come from, we would infinite loop. | |||
1380 | if (SuccBB == BB) { | |||
1381 | DEBUG(dbgs() << " Not threading across BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading across BB '" << BB->getName() << "' - would thread to self!\n" ; } } while (0) | |||
1382 | << "' - would thread to self!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading across BB '" << BB->getName() << "' - would thread to self!\n" ; } } while (0); | |||
1383 | return false; | |||
1384 | } | |||
1385 | ||||
1386 | // If threading this would thread across a loop header, don't thread the edge. | |||
1387 | // See the comments above FindLoopHeaders for justifications and caveats. | |||
1388 | if (LoopHeaders.count(BB)) { | |||
1389 | DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading across loop header BB '" << BB->getName() << "' to dest BB '" << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; } } while (0) | |||
1390 | << "' to dest BB '" << SuccBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading across loop header BB '" << BB->getName() << "' to dest BB '" << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; } } while (0) | |||
1391 | << "' - it might create an irreducible loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading across loop header BB '" << BB->getName() << "' to dest BB '" << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; } } while (0); | |||
1392 | return false; | |||
1393 | } | |||
1394 | ||||
1395 | unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold); | |||
1396 | if (JumpThreadCost > BBDupThreshold) { | |||
1397 | DEBUG(dbgs() << " Not threading BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading BB '" << BB->getName() << "' - Cost is too high: " << JumpThreadCost << "\n"; } } while (0) | |||
1398 | << "' - Cost is too high: " << JumpThreadCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading BB '" << BB->getName() << "' - Cost is too high: " << JumpThreadCost << "\n"; } } while (0); | |||
1399 | return false; | |||
1400 | } | |||
1401 | ||||
1402 | // And finally, do it! Start by factoring the predecessors is needed. | |||
1403 | BasicBlock *PredBB; | |||
1404 | if (PredBBs.size() == 1) | |||
1405 | PredBB = PredBBs[0]; | |||
1406 | else { | |||
1407 | DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (0) | |||
1408 | << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (0); | |||
1409 | PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); | |||
1410 | } | |||
1411 | ||||
1412 | // And finally, do it! | |||
1413 | DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" << SuccBB ->getName() << "' with cost: " << JumpThreadCost << ", across block:\n " << *BB << "\n"; } } while (0) | |||
1414 | << SuccBB->getName() << "' with cost: " << JumpThreadCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" << SuccBB ->getName() << "' with cost: " << JumpThreadCost << ", across block:\n " << *BB << "\n"; } } while (0) | |||
1415 | << ", across block:\n "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" << SuccBB ->getName() << "' with cost: " << JumpThreadCost << ", across block:\n " << *BB << "\n"; } } while (0) | |||
1416 | << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" << SuccBB ->getName() << "' with cost: " << JumpThreadCost << ", across block:\n " << *BB << "\n"; } } while (0); | |||
1417 | ||||
1418 | LVI->threadEdge(PredBB, BB, SuccBB); | |||
1419 | ||||
1420 | // We are going to have to map operands from the original BB block to the new | |||
1421 | // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to | |||
1422 | // account for entry from PredBB. | |||
1423 | DenseMap<Instruction*, Value*> ValueMapping; | |||
1424 | ||||
1425 | BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), | |||
1426 | BB->getName()+".thread", | |||
1427 | BB->getParent(), BB); | |||
1428 | NewBB->moveAfter(PredBB); | |||
1429 | ||||
1430 | BasicBlock::iterator BI = BB->begin(); | |||
1431 | for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) | |||
1432 | ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); | |||
1433 | ||||
1434 | // Clone the non-phi instructions of BB into NewBB, keeping track of the | |||
1435 | // mapping and using it to remap operands in the cloned instructions. | |||
1436 | for (; !isa<TerminatorInst>(BI); ++BI) { | |||
1437 | Instruction *New = BI->clone(); | |||
1438 | New->setName(BI->getName()); | |||
1439 | NewBB->getInstList().push_back(New); | |||
1440 | ValueMapping[BI] = New; | |||
1441 | ||||
1442 | // Remap operands to patch up intra-block references. | |||
1443 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) | |||
1444 | if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { | |||
1445 | DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); | |||
1446 | if (I != ValueMapping.end()) | |||
1447 | New->setOperand(i, I->second); | |||
1448 | } | |||
1449 | } | |||
1450 | ||||
1451 | // We didn't copy the terminator from BB over to NewBB, because there is now | |||
1452 | // an unconditional jump to SuccBB. Insert the unconditional jump. | |||
1453 | BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB); | |||
1454 | NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); | |||
1455 | ||||
1456 | // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the | |||
1457 | // PHI nodes for NewBB now. | |||
1458 | AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); | |||
1459 | ||||
1460 | // If there were values defined in BB that are used outside the block, then we | |||
1461 | // now have to update all uses of the value to use either the original value, | |||
1462 | // the cloned value, or some PHI derived value. This can require arbitrary | |||
1463 | // PHI insertion, of which we are prepared to do, clean these up now. | |||
1464 | SSAUpdater SSAUpdate; | |||
1465 | SmallVector<Use*, 16> UsesToRename; | |||
1466 | for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { | |||
1467 | // Scan all uses of this instruction to see if it is used outside of its | |||
1468 | // block, and if so, record them in UsesToRename. | |||
1469 | for (Use &U : I->uses()) { | |||
1470 | Instruction *User = cast<Instruction>(U.getUser()); | |||
1471 | if (PHINode *UserPN = dyn_cast<PHINode>(User)) { | |||
1472 | if (UserPN->getIncomingBlock(U) == BB) | |||
1473 | continue; | |||
1474 | } else if (User->getParent() == BB) | |||
1475 | continue; | |||
1476 | ||||
1477 | UsesToRename.push_back(&U); | |||
1478 | } | |||
1479 | ||||
1480 | // If there are no uses outside the block, we're done with this instruction. | |||
1481 | if (UsesToRename.empty()) | |||
1482 | continue; | |||
1483 | ||||
1484 | DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"; } } while (0); | |||
1485 | ||||
1486 | // We found a use of I outside of BB. Rename all uses of I that are outside | |||
1487 | // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks | |||
1488 | // with the two values we know. | |||
1489 | SSAUpdate.Initialize(I->getType(), I->getName()); | |||
1490 | SSAUpdate.AddAvailableValue(BB, I); | |||
1491 | SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]); | |||
1492 | ||||
1493 | while (!UsesToRename.empty()) | |||
1494 | SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); | |||
1495 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "\n"; } } while (0); | |||
1496 | } | |||
1497 | ||||
1498 | ||||
1499 | // Ok, NewBB is good to go. Update the terminator of PredBB to jump to | |||
1500 | // NewBB instead of BB. This eliminates predecessors from BB, which requires | |||
1501 | // us to simplify any PHI nodes in BB. | |||
1502 | TerminatorInst *PredTerm = PredBB->getTerminator(); | |||
1503 | for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) | |||
1504 | if (PredTerm->getSuccessor(i) == BB) { | |||
1505 | BB->removePredecessor(PredBB, true); | |||
1506 | PredTerm->setSuccessor(i, NewBB); | |||
1507 | } | |||
1508 | ||||
1509 | // At this point, the IR is fully up to date and consistent. Do a quick scan | |||
1510 | // over the new instructions and zap any that are constants or dead. This | |||
1511 | // frequently happens because of phi translation. | |||
1512 | SimplifyInstructionsInBlock(NewBB, DL, TLI); | |||
1513 | ||||
1514 | // Threaded an edge! | |||
1515 | ++NumThreads; | |||
1516 | return true; | |||
1517 | } | |||
1518 | ||||
1519 | /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch | |||
1520 | /// to BB which contains an i1 PHI node and a conditional branch on that PHI. | |||
1521 | /// If we can duplicate the contents of BB up into PredBB do so now, this | |||
1522 | /// improves the odds that the branch will be on an analyzable instruction like | |||
1523 | /// a compare. | |||
1524 | bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, | |||
1525 | const SmallVectorImpl<BasicBlock *> &PredBBs) { | |||
1526 | assert(!PredBBs.empty() && "Can't handle an empty set")((!PredBBs.empty() && "Can't handle an empty set") ? static_cast <void> (0) : __assert_fail ("!PredBBs.empty() && \"Can't handle an empty set\"" , "/tmp/buildd/llvm-toolchain-snapshot-3.6~svn219601/lib/Transforms/Scalar/JumpThreading.cpp" , 1526, __PRETTY_FUNCTION__)); | |||
1527 | ||||
1528 | // If BB is a loop header, then duplicating this block outside the loop would | |||
1529 | // cause us to transform this into an irreducible loop, don't do this. | |||
1530 | // See the comments above FindLoopHeaders for justifications and caveats. | |||
1531 | if (LoopHeaders.count(BB)) { | |||
1532 | DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not duplicating loop header '" << BB->getName() << "' into predecessor block '" << PredBBs[0]->getName() << "' - it might create an irreducible loop!\n" ; } } while (0) | |||
1533 | << "' into predecessor block '" << PredBBs[0]->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not duplicating loop header '" << BB->getName() << "' into predecessor block '" << PredBBs[0]->getName() << "' - it might create an irreducible loop!\n" ; } } while (0) | |||
1534 | << "' - it might create an irreducible loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not duplicating loop header '" << BB->getName() << "' into predecessor block '" << PredBBs[0]->getName() << "' - it might create an irreducible loop!\n" ; } } while (0); | |||
1535 | return false; | |||
1536 | } | |||
1537 | ||||
1538 | unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold); | |||
1539 | if (DuplicationCost > BBDupThreshold) { | |||
1540 | DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not duplicating BB '" << BB->getName() << "' - Cost is too high: " << DuplicationCost << "\n"; } } while (0) | |||
1541 | << "' - Cost is too high: " << DuplicationCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not duplicating BB '" << BB->getName() << "' - Cost is too high: " << DuplicationCost << "\n"; } } while (0); | |||
1542 | return false; | |||
1543 | } | |||
1544 | ||||
1545 | // And finally, do it! Start by factoring the predecessors is needed. | |||
1546 | BasicBlock *PredBB; | |||
1547 | if (PredBBs.size() == 1) | |||
1548 | PredBB = PredBBs[0]; | |||
1549 | else { | |||
1550 | DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (0) | |||
1551 | << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (0); | |||
1552 | PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this); | |||
1553 | } | |||
1554 | ||||
1555 | // Okay, we decided to do this! Clone all the instructions in BB onto the end | |||
1556 | // of PredBB. | |||
1557 | DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" << PredBB->getName() << "' to eliminate branch on phi. Cost: " << DuplicationCost << " block is:" << *BB << "\n"; } } while (0) | |||
1558 | << PredBB->getName() << "' to eliminate branch on phi. Cost: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" << PredBB->getName() << "' to eliminate branch on phi. Cost: " << DuplicationCost << " block is:" << *BB << "\n"; } } while (0) | |||
1559 | << DuplicationCost << " block is:" << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Duplicating block '" << BB->getName() << "' into end of '" << PredBB->getName() << "' to eliminate branch on phi. Cost: " << DuplicationCost << " block is:" << *BB << "\n"; } } while (0); | |||
1560 | ||||
1561 | // Unless PredBB ends with an unconditional branch, split the edge so that we | |||
1562 | // can just clone the bits from BB into the end of the new PredBB. | |||
1563 | BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); | |||
1564 | ||||
1565 | if (!OldPredBranch || !OldPredBranch->isUnconditional()) { | |||
1566 | PredBB = SplitEdge(PredBB, BB, this); | |||
1567 | OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); | |||
1568 | } | |||
1569 | ||||
1570 | // We are going to have to map operands from the original BB block into the | |||
1571 | // PredBB block. Evaluate PHI nodes in BB. | |||
1572 | DenseMap<Instruction*, Value*> ValueMapping; | |||
1573 | ||||
1574 | BasicBlock::iterator BI = BB->begin(); | |||
1575 | for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) | |||
1576 | ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); | |||
1577 | ||||
1578 | // Clone the non-phi instructions of BB into PredBB, keeping track of the | |||
1579 | // mapping and using it to remap operands in the cloned instructions. | |||
1580 | for (; BI != BB->end(); ++BI) { | |||
1581 | Instruction *New = BI->clone(); | |||
1582 | ||||
1583 | // Remap operands to patch up intra-block references. | |||
1584 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) | |||
1585 | if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { | |||
1586 | DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); | |||
1587 | if (I != ValueMapping.end()) | |||
1588 | New->setOperand(i, I->second); | |||
1589 | } | |||
1590 | ||||
1591 | // If this instruction can be simplified after the operands are updated, | |||
1592 | // just use the simplified value instead. This frequently happens due to | |||
1593 | // phi translation. | |||
1594 | if (Value *IV = SimplifyInstruction(New, DL)) { | |||
1595 | delete New; | |||
1596 | ValueMapping[BI] = IV; | |||
1597 | } else { | |||
1598 | // Otherwise, insert the new instruction into the block. | |||
1599 | New->setName(BI->getName()); | |||
1600 | PredBB->getInstList().insert(OldPredBranch, New); | |||
1601 | ValueMapping[BI] = New; | |||
1602 | } | |||
1603 | } | |||
1604 | ||||
1605 | // Check to see if the targets of the branch had PHI nodes. If so, we need to | |||
1606 | // add entries to the PHI nodes for branch from PredBB now. | |||
1607 | BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); | |||
1608 | AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, | |||
1609 | ValueMapping); | |||
1610 | AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, | |||
1611 | ValueMapping); | |||
1612 | ||||
1613 | // If there were values defined in BB that are used outside the block, then we | |||
1614 | // now have to update all uses of the value to use either the original value, | |||
1615 | // the cloned value, or some PHI derived value. This can require arbitrary | |||
1616 | // PHI insertion, of which we are prepared to do, clean these up now. | |||
1617 | SSAUpdater SSAUpdate; | |||
1618 | SmallVector<Use*, 16> UsesToRename; | |||
1619 | for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { | |||
1620 | // Scan all uses of this instruction to see if it is used outside of its | |||
1621 | // block, and if so, record them in UsesToRename. | |||
1622 | for (Use &U : I->uses()) { | |||
1623 | Instruction *User = cast<Instruction>(U.getUser()); | |||
1624 | if (PHINode *UserPN = dyn_cast<PHINode>(User)) { | |||
1625 | if (UserPN->getIncomingBlock(U) == BB) | |||
1626 | continue; | |||
1627 | } else if (User->getParent() == BB) | |||
1628 | continue; | |||
1629 | ||||
1630 | UsesToRename.push_back(&U); | |||
1631 | } | |||
1632 | ||||
1633 | // If there are no uses outside the block, we're done with this instruction. | |||
1634 | if (UsesToRename.empty()) | |||
1635 | continue; | |||
1636 | ||||
1637 | DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "JT: Renaming non-local uses of: " << *I << "\n"; } } while (0); | |||
1638 | ||||
1639 | // We found a use of I outside of BB. Rename all uses of I that are outside | |||
1640 | // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks | |||
1641 | // with the two values we know. | |||
1642 | SSAUpdate.Initialize(I->getType(), I->getName()); | |||
1643 | SSAUpdate.AddAvailableValue(BB, I); | |||
1644 | SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]); | |||
1645 | ||||
1646 | while (!UsesToRename.empty()) | |||
1647 | SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); | |||
1648 | DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "\n"; } } while (0); | |||
1649 | } | |||
1650 | ||||
1651 | // PredBB no longer jumps to BB, remove entries in the PHI node for the edge | |||
1652 | // that we nuked. | |||
1653 | BB->removePredecessor(PredBB, true); | |||
1654 | ||||
1655 | // Remove the unconditional branch at the end of the PredBB block. | |||
1656 | OldPredBranch->eraseFromParent(); | |||
1657 | ||||
1658 | ++NumDupes; | |||
1659 | return true; | |||
1660 | } | |||
1661 | ||||
1662 | /// TryToUnfoldSelect - Look for blocks of the form | |||
1663 | /// bb1: | |||
1664 | /// %a = select | |||
1665 | /// br bb | |||
1666 | /// | |||
1667 | /// bb2: | |||
1668 | /// %p = phi [%a, %bb] ... | |||
1669 | /// %c = icmp %p | |||
1670 | /// br i1 %c | |||
1671 | /// | |||
1672 | /// And expand the select into a branch structure if one of its arms allows %c | |||
1673 | /// to be folded. This later enables threading from bb1 over bb2. | |||
1674 | bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { | |||
1675 | BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); | |||
1676 | PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0)); | |||
1677 | Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1)); | |||
1678 | ||||
1679 | if (!CondBr || !CondBr->isConditional() || !CondLHS || | |||
1680 | CondLHS->getParent() != BB) | |||
1681 | return false; | |||
1682 | ||||
1683 | for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { | |||
1684 | BasicBlock *Pred = CondLHS->getIncomingBlock(I); | |||
1685 | SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I)); | |||
1686 | ||||
1687 | // Look if one of the incoming values is a select in the corresponding | |||
1688 | // predecessor. | |||
1689 | if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) | |||
1690 | continue; | |||
1691 | ||||
1692 | BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); | |||
1693 | if (!PredTerm || !PredTerm->isUnconditional()) | |||
1694 | continue; | |||
1695 | ||||
1696 | // Now check if one of the select values would allow us to constant fold the | |||
1697 | // terminator in BB. We don't do the transform if both sides fold, those | |||
1698 | // cases will be threaded in any case. | |||
1699 | LazyValueInfo::Tristate LHSFolds = | |||
1700 | LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1), | |||
1701 | CondRHS, Pred, BB, CondCmp); | |||
1702 | LazyValueInfo::Tristate RHSFolds = | |||
1703 | LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2), | |||
1704 | CondRHS, Pred, BB, CondCmp); | |||
1705 | if ((LHSFolds != LazyValueInfo::Unknown || | |||
1706 | RHSFolds != LazyValueInfo::Unknown) && | |||
1707 | LHSFolds != RHSFolds) { | |||
1708 | // Expand the select. | |||
1709 | // | |||
1710 | // Pred -- | |||
1711 | // | v | |||
1712 | // | NewBB | |||
1713 | // | | | |||
1714 | // |----- | |||
1715 | // v | |||
1716 | // BB | |||
1717 | BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold", | |||
1718 | BB->getParent(), BB); | |||
1719 | // Move the unconditional branch to NewBB. | |||
1720 | PredTerm->removeFromParent(); | |||
1721 | NewBB->getInstList().insert(NewBB->end(), PredTerm); | |||
1722 | // Create a conditional branch and update PHI nodes. | |||
1723 | BranchInst::Create(NewBB, BB, SI->getCondition(), Pred); | |||
1724 | CondLHS->setIncomingValue(I, SI->getFalseValue()); | |||
1725 | CondLHS->addIncoming(SI->getTrueValue(), NewBB); | |||
1726 | // The select is now dead. | |||
1727 | SI->eraseFromParent(); | |||
1728 | ||||
1729 | // Update any other PHI nodes in BB. | |||
1730 | for (BasicBlock::iterator BI = BB->begin(); | |||
1731 | PHINode *Phi = dyn_cast<PHINode>(BI); ++BI) | |||
1732 | if (Phi != CondLHS) | |||
1733 | Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB); | |||
1734 | return true; | |||
1735 | } | |||
1736 | } | |||
1737 | return false; | |||
1738 | } |