File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/llvm/lib/Transforms/Scalar/JumpThreading.cpp |
Warning: | line 1451, column 7 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- JumpThreading.cpp - Thread control through conditional blocks ------===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This file implements the Jump Threading pass. | |||
10 | // | |||
11 | //===----------------------------------------------------------------------===// | |||
12 | ||||
13 | #include "llvm/Transforms/Scalar/JumpThreading.h" | |||
14 | #include "llvm/ADT/DenseMap.h" | |||
15 | #include "llvm/ADT/DenseSet.h" | |||
16 | #include "llvm/ADT/MapVector.h" | |||
17 | #include "llvm/ADT/Optional.h" | |||
18 | #include "llvm/ADT/STLExtras.h" | |||
19 | #include "llvm/ADT/SmallPtrSet.h" | |||
20 | #include "llvm/ADT/SmallVector.h" | |||
21 | #include "llvm/ADT/Statistic.h" | |||
22 | #include "llvm/Analysis/AliasAnalysis.h" | |||
23 | #include "llvm/Analysis/BlockFrequencyInfo.h" | |||
24 | #include "llvm/Analysis/BranchProbabilityInfo.h" | |||
25 | #include "llvm/Analysis/CFG.h" | |||
26 | #include "llvm/Analysis/ConstantFolding.h" | |||
27 | #include "llvm/Analysis/DomTreeUpdater.h" | |||
28 | #include "llvm/Analysis/GlobalsModRef.h" | |||
29 | #include "llvm/Analysis/GuardUtils.h" | |||
30 | #include "llvm/Analysis/InstructionSimplify.h" | |||
31 | #include "llvm/Analysis/LazyValueInfo.h" | |||
32 | #include "llvm/Analysis/Loads.h" | |||
33 | #include "llvm/Analysis/LoopInfo.h" | |||
34 | #include "llvm/Analysis/MemoryLocation.h" | |||
35 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
36 | #include "llvm/Analysis/TargetTransformInfo.h" | |||
37 | #include "llvm/Analysis/ValueTracking.h" | |||
38 | #include "llvm/IR/BasicBlock.h" | |||
39 | #include "llvm/IR/CFG.h" | |||
40 | #include "llvm/IR/Constant.h" | |||
41 | #include "llvm/IR/ConstantRange.h" | |||
42 | #include "llvm/IR/Constants.h" | |||
43 | #include "llvm/IR/DataLayout.h" | |||
44 | #include "llvm/IR/Dominators.h" | |||
45 | #include "llvm/IR/Function.h" | |||
46 | #include "llvm/IR/InstrTypes.h" | |||
47 | #include "llvm/IR/Instruction.h" | |||
48 | #include "llvm/IR/Instructions.h" | |||
49 | #include "llvm/IR/IntrinsicInst.h" | |||
50 | #include "llvm/IR/Intrinsics.h" | |||
51 | #include "llvm/IR/LLVMContext.h" | |||
52 | #include "llvm/IR/MDBuilder.h" | |||
53 | #include "llvm/IR/Metadata.h" | |||
54 | #include "llvm/IR/Module.h" | |||
55 | #include "llvm/IR/PassManager.h" | |||
56 | #include "llvm/IR/PatternMatch.h" | |||
57 | #include "llvm/IR/Type.h" | |||
58 | #include "llvm/IR/Use.h" | |||
59 | #include "llvm/IR/Value.h" | |||
60 | #include "llvm/InitializePasses.h" | |||
61 | #include "llvm/Pass.h" | |||
62 | #include "llvm/Support/BlockFrequency.h" | |||
63 | #include "llvm/Support/BranchProbability.h" | |||
64 | #include "llvm/Support/Casting.h" | |||
65 | #include "llvm/Support/CommandLine.h" | |||
66 | #include "llvm/Support/Debug.h" | |||
67 | #include "llvm/Support/raw_ostream.h" | |||
68 | #include "llvm/Transforms/Scalar.h" | |||
69 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
70 | #include "llvm/Transforms/Utils/Cloning.h" | |||
71 | #include "llvm/Transforms/Utils/Local.h" | |||
72 | #include "llvm/Transforms/Utils/SSAUpdater.h" | |||
73 | #include "llvm/Transforms/Utils/ValueMapper.h" | |||
74 | #include <algorithm> | |||
75 | #include <cassert> | |||
76 | #include <cstdint> | |||
77 | #include <iterator> | |||
78 | #include <memory> | |||
79 | #include <utility> | |||
80 | ||||
81 | using namespace llvm; | |||
82 | using namespace jumpthreading; | |||
83 | ||||
84 | #define DEBUG_TYPE"jump-threading" "jump-threading" | |||
85 | ||||
86 | STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = {"jump-threading", "NumThreads" , "Number of jumps threaded"}; | |||
87 | STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = {"jump-threading", "NumFolds" , "Number of terminators folded"}; | |||
88 | STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi")static llvm::Statistic NumDupes = {"jump-threading", "NumDupes" , "Number of branch blocks duplicated to eliminate phi"}; | |||
89 | ||||
90 | static cl::opt<unsigned> | |||
91 | BBDuplicateThreshold("jump-threading-threshold", | |||
92 | cl::desc("Max block size to duplicate for jump threading"), | |||
93 | cl::init(6), cl::Hidden); | |||
94 | ||||
95 | static cl::opt<unsigned> | |||
96 | ImplicationSearchThreshold( | |||
97 | "jump-threading-implication-search-threshold", | |||
98 | cl::desc("The number of predecessors to search for a stronger " | |||
99 | "condition to use to thread over a weaker condition"), | |||
100 | cl::init(3), cl::Hidden); | |||
101 | ||||
102 | static cl::opt<bool> PrintLVIAfterJumpThreading( | |||
103 | "print-lvi-after-jump-threading", | |||
104 | cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false), | |||
105 | cl::Hidden); | |||
106 | ||||
107 | static cl::opt<bool> JumpThreadingFreezeSelectCond( | |||
108 | "jump-threading-freeze-select-cond", | |||
109 | cl::desc("Freeze the condition when unfolding select"), cl::init(false), | |||
110 | cl::Hidden); | |||
111 | ||||
112 | static cl::opt<bool> ThreadAcrossLoopHeaders( | |||
113 | "jump-threading-across-loop-headers", | |||
114 | cl::desc("Allow JumpThreading to thread across loop headers, for testing"), | |||
115 | cl::init(false), cl::Hidden); | |||
116 | ||||
117 | ||||
118 | namespace { | |||
119 | ||||
120 | /// This pass performs 'jump threading', which looks at blocks that have | |||
121 | /// multiple predecessors and multiple successors. If one or more of the | |||
122 | /// predecessors of the block can be proven to always jump to one of the | |||
123 | /// successors, we forward the edge from the predecessor to the successor by | |||
124 | /// duplicating the contents of this block. | |||
125 | /// | |||
126 | /// An example of when this can occur is code like this: | |||
127 | /// | |||
128 | /// if () { ... | |||
129 | /// X = 4; | |||
130 | /// } | |||
131 | /// if (X < 3) { | |||
132 | /// | |||
133 | /// In this case, the unconditional branch at the end of the first if can be | |||
134 | /// revectored to the false side of the second if. | |||
135 | class JumpThreading : public FunctionPass { | |||
136 | JumpThreadingPass Impl; | |||
137 | ||||
138 | public: | |||
139 | static char ID; // Pass identification | |||
140 | ||||
141 | JumpThreading(bool InsertFreezeWhenUnfoldingSelect = false, int T = -1) | |||
142 | : FunctionPass(ID), Impl(InsertFreezeWhenUnfoldingSelect, T) { | |||
143 | initializeJumpThreadingPass(*PassRegistry::getPassRegistry()); | |||
144 | } | |||
145 | ||||
146 | bool runOnFunction(Function &F) override; | |||
147 | ||||
148 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
149 | AU.addRequired<DominatorTreeWrapperPass>(); | |||
150 | AU.addPreserved<DominatorTreeWrapperPass>(); | |||
151 | AU.addRequired<AAResultsWrapperPass>(); | |||
152 | AU.addRequired<LazyValueInfoWrapperPass>(); | |||
153 | AU.addPreserved<LazyValueInfoWrapperPass>(); | |||
154 | AU.addPreserved<GlobalsAAWrapperPass>(); | |||
155 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | |||
156 | AU.addRequired<TargetTransformInfoWrapperPass>(); | |||
157 | } | |||
158 | ||||
159 | void releaseMemory() override { Impl.releaseMemory(); } | |||
160 | }; | |||
161 | ||||
162 | } // end anonymous namespace | |||
163 | ||||
164 | char JumpThreading::ID = 0; | |||
165 | ||||
166 | INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void *initializeJumpThreadingPassOnce(PassRegistry & Registry) { | |||
167 | "Jump Threading", false, false)static void *initializeJumpThreadingPassOnce(PassRegistry & Registry) { | |||
168 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | |||
169 | INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry); | |||
170 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | |||
171 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | |||
172 | 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; } static llvm::once_flag InitializeJumpThreadingPassFlag ; void llvm::initializeJumpThreadingPass(PassRegistry &Registry ) { llvm::call_once(InitializeJumpThreadingPassFlag, initializeJumpThreadingPassOnce , std::ref(Registry)); } | |||
173 | "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; } static llvm::once_flag InitializeJumpThreadingPassFlag ; void llvm::initializeJumpThreadingPass(PassRegistry &Registry ) { llvm::call_once(InitializeJumpThreadingPassFlag, initializeJumpThreadingPassOnce , std::ref(Registry)); } | |||
174 | ||||
175 | // Public interface to the Jump Threading pass | |||
176 | FunctionPass *llvm::createJumpThreadingPass(bool InsertFr, int Threshold) { | |||
177 | return new JumpThreading(InsertFr, Threshold); | |||
178 | } | |||
179 | ||||
180 | JumpThreadingPass::JumpThreadingPass(bool InsertFr, int T) { | |||
181 | InsertFreezeWhenUnfoldingSelect = JumpThreadingFreezeSelectCond | InsertFr; | |||
182 | DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T); | |||
183 | } | |||
184 | ||||
185 | // Update branch probability information according to conditional | |||
186 | // branch probability. This is usually made possible for cloned branches | |||
187 | // in inline instances by the context specific profile in the caller. | |||
188 | // For instance, | |||
189 | // | |||
190 | // [Block PredBB] | |||
191 | // [Branch PredBr] | |||
192 | // if (t) { | |||
193 | // Block A; | |||
194 | // } else { | |||
195 | // Block B; | |||
196 | // } | |||
197 | // | |||
198 | // [Block BB] | |||
199 | // cond = PN([true, %A], [..., %B]); // PHI node | |||
200 | // [Branch CondBr] | |||
201 | // if (cond) { | |||
202 | // ... // P(cond == true) = 1% | |||
203 | // } | |||
204 | // | |||
205 | // Here we know that when block A is taken, cond must be true, which means | |||
206 | // P(cond == true | A) = 1 | |||
207 | // | |||
208 | // Given that P(cond == true) = P(cond == true | A) * P(A) + | |||
209 | // P(cond == true | B) * P(B) | |||
210 | // we get: | |||
211 | // P(cond == true ) = P(A) + P(cond == true | B) * P(B) | |||
212 | // | |||
213 | // which gives us: | |||
214 | // P(A) is less than P(cond == true), i.e. | |||
215 | // P(t == true) <= P(cond == true) | |||
216 | // | |||
217 | // In other words, if we know P(cond == true) is unlikely, we know | |||
218 | // that P(t == true) is also unlikely. | |||
219 | // | |||
220 | static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) { | |||
221 | BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); | |||
222 | if (!CondBr) | |||
223 | return; | |||
224 | ||||
225 | uint64_t TrueWeight, FalseWeight; | |||
226 | if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight)) | |||
227 | return; | |||
228 | ||||
229 | if (TrueWeight + FalseWeight == 0) | |||
230 | // Zero branch_weights do not give a hint for getting branch probabilities. | |||
231 | // Technically it would result in division by zero denominator, which is | |||
232 | // TrueWeight + FalseWeight. | |||
233 | return; | |||
234 | ||||
235 | // Returns the outgoing edge of the dominating predecessor block | |||
236 | // that leads to the PhiNode's incoming block: | |||
237 | auto GetPredOutEdge = | |||
238 | [](BasicBlock *IncomingBB, | |||
239 | BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> { | |||
240 | auto *PredBB = IncomingBB; | |||
241 | auto *SuccBB = PhiBB; | |||
242 | SmallPtrSet<BasicBlock *, 16> Visited; | |||
243 | while (true) { | |||
244 | BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()); | |||
245 | if (PredBr && PredBr->isConditional()) | |||
246 | return {PredBB, SuccBB}; | |||
247 | Visited.insert(PredBB); | |||
248 | auto *SinglePredBB = PredBB->getSinglePredecessor(); | |||
249 | if (!SinglePredBB) | |||
250 | return {nullptr, nullptr}; | |||
251 | ||||
252 | // Stop searching when SinglePredBB has been visited. It means we see | |||
253 | // an unreachable loop. | |||
254 | if (Visited.count(SinglePredBB)) | |||
255 | return {nullptr, nullptr}; | |||
256 | ||||
257 | SuccBB = PredBB; | |||
258 | PredBB = SinglePredBB; | |||
259 | } | |||
260 | }; | |||
261 | ||||
262 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |||
263 | Value *PhiOpnd = PN->getIncomingValue(i); | |||
264 | ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd); | |||
265 | ||||
266 | if (!CI || !CI->getType()->isIntegerTy(1)) | |||
267 | continue; | |||
268 | ||||
269 | BranchProbability BP = | |||
270 | (CI->isOne() ? BranchProbability::getBranchProbability( | |||
271 | TrueWeight, TrueWeight + FalseWeight) | |||
272 | : BranchProbability::getBranchProbability( | |||
273 | FalseWeight, TrueWeight + FalseWeight)); | |||
274 | ||||
275 | auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB); | |||
276 | if (!PredOutEdge.first) | |||
277 | return; | |||
278 | ||||
279 | BasicBlock *PredBB = PredOutEdge.first; | |||
280 | BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()); | |||
281 | if (!PredBr) | |||
282 | return; | |||
283 | ||||
284 | uint64_t PredTrueWeight, PredFalseWeight; | |||
285 | // FIXME: We currently only set the profile data when it is missing. | |||
286 | // With PGO, this can be used to refine even existing profile data with | |||
287 | // context information. This needs to be done after more performance | |||
288 | // testing. | |||
289 | if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight)) | |||
290 | continue; | |||
291 | ||||
292 | // We can not infer anything useful when BP >= 50%, because BP is the | |||
293 | // upper bound probability value. | |||
294 | if (BP >= BranchProbability(50, 100)) | |||
295 | continue; | |||
296 | ||||
297 | SmallVector<uint32_t, 2> Weights; | |||
298 | if (PredBr->getSuccessor(0) == PredOutEdge.second) { | |||
299 | Weights.push_back(BP.getNumerator()); | |||
300 | Weights.push_back(BP.getCompl().getNumerator()); | |||
301 | } else { | |||
302 | Weights.push_back(BP.getCompl().getNumerator()); | |||
303 | Weights.push_back(BP.getNumerator()); | |||
304 | } | |||
305 | PredBr->setMetadata(LLVMContext::MD_prof, | |||
306 | MDBuilder(PredBr->getParent()->getContext()) | |||
307 | .createBranchWeights(Weights)); | |||
308 | } | |||
309 | } | |||
310 | ||||
311 | /// runOnFunction - Toplevel algorithm. | |||
312 | bool JumpThreading::runOnFunction(Function &F) { | |||
313 | if (skipFunction(F)) | |||
314 | return false; | |||
315 | auto TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | |||
316 | // Jump Threading has no sense for the targets with divergent CF | |||
317 | if (TTI->hasBranchDivergence()) | |||
318 | return false; | |||
319 | auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | |||
320 | auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | |||
321 | auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI(); | |||
322 | auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | |||
323 | DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy); | |||
324 | std::unique_ptr<BlockFrequencyInfo> BFI; | |||
325 | std::unique_ptr<BranchProbabilityInfo> BPI; | |||
326 | if (F.hasProfileData()) { | |||
327 | LoopInfo LI{DominatorTree(F)}; | |||
328 | BPI.reset(new BranchProbabilityInfo(F, LI, TLI)); | |||
329 | BFI.reset(new BlockFrequencyInfo(F, *BPI, LI)); | |||
330 | } | |||
331 | ||||
332 | bool Changed = Impl.runImpl(F, TLI, TTI, LVI, AA, &DTU, F.hasProfileData(), | |||
333 | std::move(BFI), std::move(BPI)); | |||
334 | if (PrintLVIAfterJumpThreading) { | |||
335 | dbgs() << "LVI for function '" << F.getName() << "':\n"; | |||
336 | LVI->printLVI(F, DTU.getDomTree(), dbgs()); | |||
337 | } | |||
338 | return Changed; | |||
339 | } | |||
340 | ||||
341 | PreservedAnalyses JumpThreadingPass::run(Function &F, | |||
342 | FunctionAnalysisManager &AM) { | |||
343 | auto &TTI = AM.getResult<TargetIRAnalysis>(F); | |||
344 | // Jump Threading has no sense for the targets with divergent CF | |||
345 | if (TTI.hasBranchDivergence()) | |||
346 | return PreservedAnalyses::all(); | |||
347 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); | |||
348 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); | |||
349 | auto &LVI = AM.getResult<LazyValueAnalysis>(F); | |||
350 | auto &AA = AM.getResult<AAManager>(F); | |||
351 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); | |||
352 | ||||
353 | std::unique_ptr<BlockFrequencyInfo> BFI; | |||
354 | std::unique_ptr<BranchProbabilityInfo> BPI; | |||
355 | if (F.hasProfileData()) { | |||
356 | LoopInfo LI{DominatorTree(F)}; | |||
357 | BPI.reset(new BranchProbabilityInfo(F, LI, &TLI)); | |||
358 | BFI.reset(new BlockFrequencyInfo(F, *BPI, LI)); | |||
359 | } | |||
360 | ||||
361 | bool Changed = runImpl(F, &TLI, &TTI, &LVI, &AA, &DTU, F.hasProfileData(), | |||
362 | std::move(BFI), std::move(BPI)); | |||
363 | ||||
364 | if (PrintLVIAfterJumpThreading) { | |||
365 | dbgs() << "LVI for function '" << F.getName() << "':\n"; | |||
366 | LVI.printLVI(F, DTU.getDomTree(), dbgs()); | |||
367 | } | |||
368 | ||||
369 | if (!Changed) | |||
370 | return PreservedAnalyses::all(); | |||
371 | PreservedAnalyses PA; | |||
372 | PA.preserve<DominatorTreeAnalysis>(); | |||
373 | PA.preserve<LazyValueAnalysis>(); | |||
374 | return PA; | |||
375 | } | |||
376 | ||||
377 | bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_, | |||
378 | TargetTransformInfo *TTI_, LazyValueInfo *LVI_, | |||
379 | AliasAnalysis *AA_, DomTreeUpdater *DTU_, | |||
380 | bool HasProfileData_, | |||
381 | std::unique_ptr<BlockFrequencyInfo> BFI_, | |||
382 | std::unique_ptr<BranchProbabilityInfo> BPI_) { | |||
383 | LLVM_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 (false); | |||
384 | TLI = TLI_; | |||
385 | TTI = TTI_; | |||
386 | LVI = LVI_; | |||
387 | AA = AA_; | |||
388 | DTU = DTU_; | |||
389 | BFI.reset(); | |||
390 | BPI.reset(); | |||
391 | // When profile data is available, we need to update edge weights after | |||
392 | // successful jump threading, which requires both BPI and BFI being available. | |||
393 | HasProfileData = HasProfileData_; | |||
394 | auto *GuardDecl = F.getParent()->getFunction( | |||
395 | Intrinsic::getName(Intrinsic::experimental_guard)); | |||
396 | HasGuards = GuardDecl && !GuardDecl->use_empty(); | |||
397 | if (HasProfileData) { | |||
398 | BPI = std::move(BPI_); | |||
399 | BFI = std::move(BFI_); | |||
400 | } | |||
401 | ||||
402 | // Reduce the number of instructions duplicated when optimizing strictly for | |||
403 | // size. | |||
404 | if (BBDuplicateThreshold.getNumOccurrences()) | |||
405 | BBDupThreshold = BBDuplicateThreshold; | |||
406 | else if (F.hasFnAttribute(Attribute::MinSize)) | |||
407 | BBDupThreshold = 3; | |||
408 | else | |||
409 | BBDupThreshold = DefaultBBDupThreshold; | |||
410 | ||||
411 | // JumpThreading must not processes blocks unreachable from entry. It's a | |||
412 | // waste of compute time and can potentially lead to hangs. | |||
413 | SmallPtrSet<BasicBlock *, 16> Unreachable; | |||
414 | assert(DTU && "DTU isn't passed into JumpThreading before using it.")(static_cast <bool> (DTU && "DTU isn't passed into JumpThreading before using it." ) ? void (0) : __assert_fail ("DTU && \"DTU isn't passed into JumpThreading before using it.\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 414, __extension__ __PRETTY_FUNCTION__)); | |||
415 | assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.")(static_cast <bool> (DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed." ) ? void (0) : __assert_fail ("DTU->hasDomTree() && \"JumpThreading relies on DomTree to proceed.\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 415, __extension__ __PRETTY_FUNCTION__)); | |||
416 | DominatorTree &DT = DTU->getDomTree(); | |||
417 | for (auto &BB : F) | |||
418 | if (!DT.isReachableFromEntry(&BB)) | |||
419 | Unreachable.insert(&BB); | |||
420 | ||||
421 | if (!ThreadAcrossLoopHeaders) | |||
422 | findLoopHeaders(F); | |||
423 | ||||
424 | bool EverChanged = false; | |||
425 | bool Changed; | |||
426 | do { | |||
427 | Changed = false; | |||
428 | for (auto &BB : F) { | |||
429 | if (Unreachable.count(&BB)) | |||
430 | continue; | |||
431 | while (processBlock(&BB)) // Thread all of the branches we can over BB. | |||
432 | Changed = true; | |||
433 | ||||
434 | // Jump threading may have introduced redundant debug values into BB | |||
435 | // which should be removed. | |||
436 | if (Changed) | |||
437 | RemoveRedundantDbgInstrs(&BB); | |||
438 | ||||
439 | // Stop processing BB if it's the entry or is now deleted. The following | |||
440 | // routines attempt to eliminate BB and locating a suitable replacement | |||
441 | // for the entry is non-trivial. | |||
442 | if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB)) | |||
443 | continue; | |||
444 | ||||
445 | if (pred_empty(&BB)) { | |||
446 | // When processBlock makes BB unreachable it doesn't bother to fix up | |||
447 | // the instructions in it. We must remove BB to prevent invalid IR. | |||
448 | LLVM_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 (false) | |||
449 | << "' with terminator: " << *BB.getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " JT: Deleting dead block '" << BB.getName() << "' with terminator: " << *BB.getTerminator() << '\n'; } } while (false) | |||
450 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " JT: Deleting dead block '" << BB.getName() << "' with terminator: " << *BB.getTerminator() << '\n'; } } while (false); | |||
451 | LoopHeaders.erase(&BB); | |||
452 | LVI->eraseBlock(&BB); | |||
453 | DeleteDeadBlock(&BB, DTU); | |||
454 | Changed = true; | |||
455 | continue; | |||
456 | } | |||
457 | ||||
458 | // processBlock doesn't thread BBs with unconditional TIs. However, if BB | |||
459 | // is "almost empty", we attempt to merge BB with its sole successor. | |||
460 | auto *BI = dyn_cast<BranchInst>(BB.getTerminator()); | |||
461 | if (BI && BI->isUnconditional()) { | |||
462 | BasicBlock *Succ = BI->getSuccessor(0); | |||
463 | if ( | |||
464 | // The terminator must be the only non-phi instruction in BB. | |||
465 | BB.getFirstNonPHIOrDbg(true)->isTerminator() && | |||
466 | // Don't alter Loop headers and latches to ensure another pass can | |||
467 | // detect and transform nested loops later. | |||
468 | !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) && | |||
469 | TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) { | |||
470 | RemoveRedundantDbgInstrs(Succ); | |||
471 | // BB is valid for cleanup here because we passed in DTU. F remains | |||
472 | // BB's parent until a DTU->getDomTree() event. | |||
473 | LVI->eraseBlock(&BB); | |||
474 | Changed = true; | |||
475 | } | |||
476 | } | |||
477 | } | |||
478 | EverChanged |= Changed; | |||
479 | } while (Changed); | |||
480 | ||||
481 | LoopHeaders.clear(); | |||
482 | return EverChanged; | |||
483 | } | |||
484 | ||||
485 | // Replace uses of Cond with ToVal when safe to do so. If all uses are | |||
486 | // replaced, we can remove Cond. We cannot blindly replace all uses of Cond | |||
487 | // because we may incorrectly replace uses when guards/assumes are uses of | |||
488 | // of `Cond` and we used the guards/assume to reason about the `Cond` value | |||
489 | // at the end of block. RAUW unconditionally replaces all uses | |||
490 | // including the guards/assumes themselves and the uses before the | |||
491 | // guard/assume. | |||
492 | static void replaceFoldableUses(Instruction *Cond, Value *ToVal) { | |||
493 | assert(Cond->getType() == ToVal->getType())(static_cast <bool> (Cond->getType() == ToVal->getType ()) ? void (0) : __assert_fail ("Cond->getType() == ToVal->getType()" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 493, __extension__ __PRETTY_FUNCTION__)); | |||
494 | auto *BB = Cond->getParent(); | |||
495 | // We can unconditionally replace all uses in non-local blocks (i.e. uses | |||
496 | // strictly dominated by BB), since LVI information is true from the | |||
497 | // terminator of BB. | |||
498 | replaceNonLocalUsesWith(Cond, ToVal); | |||
499 | for (Instruction &I : reverse(*BB)) { | |||
500 | // Reached the Cond whose uses we are trying to replace, so there are no | |||
501 | // more uses. | |||
502 | if (&I == Cond) | |||
503 | break; | |||
504 | // We only replace uses in instructions that are guaranteed to reach the end | |||
505 | // of BB, where we know Cond is ToVal. | |||
506 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
507 | break; | |||
508 | I.replaceUsesOfWith(Cond, ToVal); | |||
509 | } | |||
510 | if (Cond->use_empty() && !Cond->mayHaveSideEffects()) | |||
511 | Cond->eraseFromParent(); | |||
512 | } | |||
513 | ||||
514 | /// Return the cost of duplicating a piece of this block from first non-phi | |||
515 | /// and before StopAt instruction to thread across it. Stop scanning the block | |||
516 | /// when exceeding the threshold. If duplication is impossible, returns ~0U. | |||
517 | static unsigned getJumpThreadDuplicationCost(const TargetTransformInfo *TTI, | |||
518 | BasicBlock *BB, | |||
519 | Instruction *StopAt, | |||
520 | unsigned Threshold) { | |||
521 | assert(StopAt->getParent() == BB && "Not an instruction from proper BB?")(static_cast <bool> (StopAt->getParent() == BB && "Not an instruction from proper BB?") ? void (0) : __assert_fail ("StopAt->getParent() == BB && \"Not an instruction from proper BB?\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 521, __extension__ __PRETTY_FUNCTION__)); | |||
522 | /// Ignore PHI nodes, these will be flattened when duplication happens. | |||
523 | BasicBlock::const_iterator I(BB->getFirstNonPHI()); | |||
524 | ||||
525 | // FIXME: THREADING will delete values that are just used to compute the | |||
526 | // branch, so they shouldn't count against the duplication cost. | |||
527 | ||||
528 | unsigned Bonus = 0; | |||
529 | if (BB->getTerminator() == StopAt) { | |||
530 | // Threading through a switch statement is particularly profitable. If this | |||
531 | // block ends in a switch, decrease its cost to make it more likely to | |||
532 | // happen. | |||
533 | if (isa<SwitchInst>(StopAt)) | |||
534 | Bonus = 6; | |||
535 | ||||
536 | // The same holds for indirect branches, but slightly more so. | |||
537 | if (isa<IndirectBrInst>(StopAt)) | |||
538 | Bonus = 8; | |||
539 | } | |||
540 | ||||
541 | // Bump the threshold up so the early exit from the loop doesn't skip the | |||
542 | // terminator-based Size adjustment at the end. | |||
543 | Threshold += Bonus; | |||
544 | ||||
545 | // Sum up the cost of each instruction until we get to the terminator. Don't | |||
546 | // include the terminator because the copy won't include it. | |||
547 | unsigned Size = 0; | |||
548 | for (; &*I != StopAt; ++I) { | |||
549 | ||||
550 | // Stop scanning the block if we've reached the threshold. | |||
551 | if (Size > Threshold) | |||
552 | return Size; | |||
553 | ||||
554 | // Bail out if this instruction gives back a token type, it is not possible | |||
555 | // to duplicate it if it is used outside this BB. | |||
556 | if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB)) | |||
557 | return ~0U; | |||
558 | ||||
559 | // Blocks with NoDuplicate are modelled as having infinite cost, so they | |||
560 | // are never duplicated. | |||
561 | if (const CallInst *CI = dyn_cast<CallInst>(I)) | |||
562 | if (CI->cannotDuplicate() || CI->isConvergent()) | |||
563 | return ~0U; | |||
564 | ||||
565 | if (TTI->getUserCost(&*I, TargetTransformInfo::TCK_SizeAndLatency) | |||
566 | == TargetTransformInfo::TCC_Free) | |||
567 | continue; | |||
568 | ||||
569 | // All other instructions count for at least one unit. | |||
570 | ++Size; | |||
571 | ||||
572 | // Calls are more expensive. If they are non-intrinsic calls, we model them | |||
573 | // as having cost of 4. If they are a non-vector intrinsic, we model them | |||
574 | // as having cost of 2 total, and if they are a vector intrinsic, we model | |||
575 | // them as having cost 1. | |||
576 | if (const CallInst *CI = dyn_cast<CallInst>(I)) { | |||
577 | if (!isa<IntrinsicInst>(CI)) | |||
578 | Size += 3; | |||
579 | else if (!CI->getType()->isVectorTy()) | |||
580 | Size += 1; | |||
581 | } | |||
582 | } | |||
583 | ||||
584 | return Size > Bonus ? Size - Bonus : 0; | |||
585 | } | |||
586 | ||||
587 | /// findLoopHeaders - We do not want jump threading to turn proper loop | |||
588 | /// structures into irreducible loops. Doing this breaks up the loop nesting | |||
589 | /// hierarchy and pessimizes later transformations. To prevent this from | |||
590 | /// happening, we first have to find the loop headers. Here we approximate this | |||
591 | /// by finding targets of backedges in the CFG. | |||
592 | /// | |||
593 | /// Note that there definitely are cases when we want to allow threading of | |||
594 | /// edges across a loop header. For example, threading a jump from outside the | |||
595 | /// loop (the preheader) to an exit block of the loop is definitely profitable. | |||
596 | /// It is also almost always profitable to thread backedges from within the loop | |||
597 | /// to exit blocks, and is often profitable to thread backedges to other blocks | |||
598 | /// within the loop (forming a nested loop). This simple analysis is not rich | |||
599 | /// enough to track all of these properties and keep it up-to-date as the CFG | |||
600 | /// mutates, so we don't allow any of these transformations. | |||
601 | void JumpThreadingPass::findLoopHeaders(Function &F) { | |||
602 | SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; | |||
603 | FindFunctionBackedges(F, Edges); | |||
604 | ||||
605 | for (const auto &Edge : Edges) | |||
606 | LoopHeaders.insert(Edge.second); | |||
607 | } | |||
608 | ||||
609 | /// getKnownConstant - Helper method to determine if we can thread over a | |||
610 | /// terminator with the given value as its condition, and if so what value to | |||
611 | /// use for that. What kind of value this is depends on whether we want an | |||
612 | /// integer or a block address, but an undef is always accepted. | |||
613 | /// Returns null if Val is null or not an appropriate constant. | |||
614 | static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) { | |||
615 | if (!Val) | |||
616 | return nullptr; | |||
617 | ||||
618 | // Undef is "known" enough. | |||
619 | if (UndefValue *U = dyn_cast<UndefValue>(Val)) | |||
620 | return U; | |||
621 | ||||
622 | if (Preference == WantBlockAddress) | |||
623 | return dyn_cast<BlockAddress>(Val->stripPointerCasts()); | |||
624 | ||||
625 | return dyn_cast<ConstantInt>(Val); | |||
626 | } | |||
627 | ||||
628 | /// computeValueKnownInPredecessors - Given a basic block BB and a value V, see | |||
629 | /// if we can infer that the value is a known ConstantInt/BlockAddress or undef | |||
630 | /// in any of our predecessors. If so, return the known list of value and pred | |||
631 | /// BB in the result vector. | |||
632 | /// | |||
633 | /// This returns true if there were any known values. | |||
634 | bool JumpThreadingPass::computeValueKnownInPredecessorsImpl( | |||
635 | Value *V, BasicBlock *BB, PredValueInfo &Result, | |||
636 | ConstantPreference Preference, DenseSet<Value *> &RecursionSet, | |||
637 | Instruction *CxtI) { | |||
638 | // This method walks up use-def chains recursively. Because of this, we could | |||
639 | // get into an infinite loop going around loops in the use-def chain. To | |||
640 | // prevent this, keep track of what (value, block) pairs we've already visited | |||
641 | // and terminate the search if we loop back to them | |||
642 | if (!RecursionSet.insert(V).second) | |||
643 | return false; | |||
644 | ||||
645 | // If V is a constant, then it is known in all predecessors. | |||
646 | if (Constant *KC = getKnownConstant(V, Preference)) { | |||
647 | for (BasicBlock *Pred : predecessors(BB)) | |||
648 | Result.emplace_back(KC, Pred); | |||
649 | ||||
650 | return !Result.empty(); | |||
651 | } | |||
652 | ||||
653 | // If V is a non-instruction value, or an instruction in a different block, | |||
654 | // then it can't be derived from a PHI. | |||
655 | Instruction *I = dyn_cast<Instruction>(V); | |||
656 | if (!I || I->getParent() != BB) { | |||
657 | ||||
658 | // Okay, if this is a live-in value, see if it has a known value at the end | |||
659 | // of any of our predecessors. | |||
660 | // | |||
661 | // FIXME: This should be an edge property, not a block end property. | |||
662 | /// TODO: Per PR2563, we could infer value range information about a | |||
663 | /// predecessor based on its terminator. | |||
664 | // | |||
665 | // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if | |||
666 | // "I" is a non-local compare-with-a-constant instruction. This would be | |||
667 | // able to handle value inequalities better, for example if the compare is | |||
668 | // "X < 4" and "X < 3" is known true but "X < 4" itself is not available. | |||
669 | // Perhaps getConstantOnEdge should be smart enough to do this? | |||
670 | for (BasicBlock *P : predecessors(BB)) { | |||
671 | // If the value is known by LazyValueInfo to be a constant in a | |||
672 | // predecessor, use that information to try to thread this block. | |||
673 | Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI); | |||
674 | if (Constant *KC = getKnownConstant(PredCst, Preference)) | |||
675 | Result.emplace_back(KC, P); | |||
676 | } | |||
677 | ||||
678 | return !Result.empty(); | |||
679 | } | |||
680 | ||||
681 | /// If I is a PHI node, then we know the incoming values for any constants. | |||
682 | if (PHINode *PN = dyn_cast<PHINode>(I)) { | |||
683 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |||
684 | Value *InVal = PN->getIncomingValue(i); | |||
685 | if (Constant *KC = getKnownConstant(InVal, Preference)) { | |||
686 | Result.emplace_back(KC, PN->getIncomingBlock(i)); | |||
687 | } else { | |||
688 | Constant *CI = LVI->getConstantOnEdge(InVal, | |||
689 | PN->getIncomingBlock(i), | |||
690 | BB, CxtI); | |||
691 | if (Constant *KC = getKnownConstant(CI, Preference)) | |||
692 | Result.emplace_back(KC, PN->getIncomingBlock(i)); | |||
693 | } | |||
694 | } | |||
695 | ||||
696 | return !Result.empty(); | |||
697 | } | |||
698 | ||||
699 | // Handle Cast instructions. | |||
700 | if (CastInst *CI = dyn_cast<CastInst>(I)) { | |||
701 | Value *Source = CI->getOperand(0); | |||
702 | computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference, | |||
703 | RecursionSet, CxtI); | |||
704 | if (Result.empty()) | |||
705 | return false; | |||
706 | ||||
707 | // Convert the known values. | |||
708 | for (auto &R : Result) | |||
709 | R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType()); | |||
710 | ||||
711 | return true; | |||
712 | } | |||
713 | ||||
714 | if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) { | |||
715 | Value *Source = FI->getOperand(0); | |||
716 | computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference, | |||
717 | RecursionSet, CxtI); | |||
718 | ||||
719 | erase_if(Result, [](auto &Pair) { | |||
720 | return !isGuaranteedNotToBeUndefOrPoison(Pair.first); | |||
721 | }); | |||
722 | ||||
723 | return !Result.empty(); | |||
724 | } | |||
725 | ||||
726 | // Handle some boolean conditions. | |||
727 | if (I->getType()->getPrimitiveSizeInBits() == 1) { | |||
728 | using namespace PatternMatch; | |||
729 | if (Preference != WantInteger) | |||
730 | return false; | |||
731 | // X | true -> true | |||
732 | // X & false -> false | |||
733 | Value *Op0, *Op1; | |||
734 | if (match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))) || | |||
735 | match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) { | |||
736 | PredValueInfoTy LHSVals, RHSVals; | |||
737 | ||||
738 | computeValueKnownInPredecessorsImpl(Op0, BB, LHSVals, WantInteger, | |||
739 | RecursionSet, CxtI); | |||
740 | computeValueKnownInPredecessorsImpl(Op1, BB, RHSVals, WantInteger, | |||
741 | RecursionSet, CxtI); | |||
742 | ||||
743 | if (LHSVals.empty() && RHSVals.empty()) | |||
744 | return false; | |||
745 | ||||
746 | ConstantInt *InterestingVal; | |||
747 | if (match(I, m_LogicalOr())) | |||
748 | InterestingVal = ConstantInt::getTrue(I->getContext()); | |||
749 | else | |||
750 | InterestingVal = ConstantInt::getFalse(I->getContext()); | |||
751 | ||||
752 | SmallPtrSet<BasicBlock*, 4> LHSKnownBBs; | |||
753 | ||||
754 | // Scan for the sentinel. If we find an undef, force it to the | |||
755 | // interesting value: x|undef -> true and x&undef -> false. | |||
756 | for (const auto &LHSVal : LHSVals) | |||
757 | if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) { | |||
758 | Result.emplace_back(InterestingVal, LHSVal.second); | |||
759 | LHSKnownBBs.insert(LHSVal.second); | |||
760 | } | |||
761 | for (const auto &RHSVal : RHSVals) | |||
762 | if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) { | |||
763 | // If we already inferred a value for this block on the LHS, don't | |||
764 | // re-add it. | |||
765 | if (!LHSKnownBBs.count(RHSVal.second)) | |||
766 | Result.emplace_back(InterestingVal, RHSVal.second); | |||
767 | } | |||
768 | ||||
769 | return !Result.empty(); | |||
770 | } | |||
771 | ||||
772 | // Handle the NOT form of XOR. | |||
773 | if (I->getOpcode() == Instruction::Xor && | |||
774 | isa<ConstantInt>(I->getOperand(1)) && | |||
775 | cast<ConstantInt>(I->getOperand(1))->isOne()) { | |||
776 | computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result, | |||
777 | WantInteger, RecursionSet, CxtI); | |||
778 | if (Result.empty()) | |||
779 | return false; | |||
780 | ||||
781 | // Invert the known values. | |||
782 | for (auto &R : Result) | |||
783 | R.first = ConstantExpr::getNot(R.first); | |||
784 | ||||
785 | return true; | |||
786 | } | |||
787 | ||||
788 | // Try to simplify some other binary operator values. | |||
789 | } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { | |||
790 | if (Preference != WantInteger) | |||
791 | return false; | |||
792 | if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) { | |||
793 | PredValueInfoTy LHSVals; | |||
794 | computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals, | |||
795 | WantInteger, RecursionSet, CxtI); | |||
796 | ||||
797 | // Try to use constant folding to simplify the binary operator. | |||
798 | for (const auto &LHSVal : LHSVals) { | |||
799 | Constant *V = LHSVal.first; | |||
800 | Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI); | |||
801 | ||||
802 | if (Constant *KC = getKnownConstant(Folded, WantInteger)) | |||
803 | Result.emplace_back(KC, LHSVal.second); | |||
804 | } | |||
805 | } | |||
806 | ||||
807 | return !Result.empty(); | |||
808 | } | |||
809 | ||||
810 | // Handle compare with phi operand, where the PHI is defined in this block. | |||
811 | if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) { | |||
812 | if (Preference != WantInteger) | |||
813 | return false; | |||
814 | Type *CmpType = Cmp->getType(); | |||
815 | Value *CmpLHS = Cmp->getOperand(0); | |||
816 | Value *CmpRHS = Cmp->getOperand(1); | |||
817 | CmpInst::Predicate Pred = Cmp->getPredicate(); | |||
818 | ||||
819 | PHINode *PN = dyn_cast<PHINode>(CmpLHS); | |||
820 | if (!PN) | |||
821 | PN = dyn_cast<PHINode>(CmpRHS); | |||
822 | if (PN && PN->getParent() == BB) { | |||
823 | const DataLayout &DL = PN->getModule()->getDataLayout(); | |||
824 | // We can do this simplification if any comparisons fold to true or false. | |||
825 | // See if any do. | |||
826 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |||
827 | BasicBlock *PredBB = PN->getIncomingBlock(i); | |||
828 | Value *LHS, *RHS; | |||
829 | if (PN == CmpLHS) { | |||
830 | LHS = PN->getIncomingValue(i); | |||
831 | RHS = CmpRHS->DoPHITranslation(BB, PredBB); | |||
832 | } else { | |||
833 | LHS = CmpLHS->DoPHITranslation(BB, PredBB); | |||
834 | RHS = PN->getIncomingValue(i); | |||
835 | } | |||
836 | Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL}); | |||
837 | if (!Res) { | |||
838 | if (!isa<Constant>(RHS)) | |||
839 | continue; | |||
840 | ||||
841 | // getPredicateOnEdge call will make no sense if LHS is defined in BB. | |||
842 | auto LHSInst = dyn_cast<Instruction>(LHS); | |||
843 | if (LHSInst && LHSInst->getParent() == BB) | |||
844 | continue; | |||
845 | ||||
846 | LazyValueInfo::Tristate | |||
847 | ResT = LVI->getPredicateOnEdge(Pred, LHS, | |||
848 | cast<Constant>(RHS), PredBB, BB, | |||
849 | CxtI ? CxtI : Cmp); | |||
850 | if (ResT == LazyValueInfo::Unknown) | |||
851 | continue; | |||
852 | Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT); | |||
853 | } | |||
854 | ||||
855 | if (Constant *KC = getKnownConstant(Res, WantInteger)) | |||
856 | Result.emplace_back(KC, PredBB); | |||
857 | } | |||
858 | ||||
859 | return !Result.empty(); | |||
860 | } | |||
861 | ||||
862 | // If comparing a live-in value against a constant, see if we know the | |||
863 | // live-in value on any predecessors. | |||
864 | if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) { | |||
865 | Constant *CmpConst = cast<Constant>(CmpRHS); | |||
866 | ||||
867 | if (!isa<Instruction>(CmpLHS) || | |||
868 | cast<Instruction>(CmpLHS)->getParent() != BB) { | |||
869 | for (BasicBlock *P : predecessors(BB)) { | |||
870 | // If the value is known by LazyValueInfo to be a constant in a | |||
871 | // predecessor, use that information to try to thread this block. | |||
872 | LazyValueInfo::Tristate Res = | |||
873 | LVI->getPredicateOnEdge(Pred, CmpLHS, | |||
874 | CmpConst, P, BB, CxtI ? CxtI : Cmp); | |||
875 | if (Res == LazyValueInfo::Unknown) | |||
876 | continue; | |||
877 | ||||
878 | Constant *ResC = ConstantInt::get(CmpType, Res); | |||
879 | Result.emplace_back(ResC, P); | |||
880 | } | |||
881 | ||||
882 | return !Result.empty(); | |||
883 | } | |||
884 | ||||
885 | // InstCombine can fold some forms of constant range checks into | |||
886 | // (icmp (add (x, C1)), C2). See if we have we have such a thing with | |||
887 | // x as a live-in. | |||
888 | { | |||
889 | using namespace PatternMatch; | |||
890 | ||||
891 | Value *AddLHS; | |||
892 | ConstantInt *AddConst; | |||
893 | if (isa<ConstantInt>(CmpConst) && | |||
894 | match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) { | |||
895 | if (!isa<Instruction>(AddLHS) || | |||
896 | cast<Instruction>(AddLHS)->getParent() != BB) { | |||
897 | for (BasicBlock *P : predecessors(BB)) { | |||
898 | // If the value is known by LazyValueInfo to be a ConstantRange in | |||
899 | // a predecessor, use that information to try to thread this | |||
900 | // block. | |||
901 | ConstantRange CR = LVI->getConstantRangeOnEdge( | |||
902 | AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS)); | |||
903 | // Propagate the range through the addition. | |||
904 | CR = CR.add(AddConst->getValue()); | |||
905 | ||||
906 | // Get the range where the compare returns true. | |||
907 | ConstantRange CmpRange = ConstantRange::makeExactICmpRegion( | |||
908 | Pred, cast<ConstantInt>(CmpConst)->getValue()); | |||
909 | ||||
910 | Constant *ResC; | |||
911 | if (CmpRange.contains(CR)) | |||
912 | ResC = ConstantInt::getTrue(CmpType); | |||
913 | else if (CmpRange.inverse().contains(CR)) | |||
914 | ResC = ConstantInt::getFalse(CmpType); | |||
915 | else | |||
916 | continue; | |||
917 | ||||
918 | Result.emplace_back(ResC, P); | |||
919 | } | |||
920 | ||||
921 | return !Result.empty(); | |||
922 | } | |||
923 | } | |||
924 | } | |||
925 | ||||
926 | // Try to find a constant value for the LHS of a comparison, | |||
927 | // and evaluate it statically if we can. | |||
928 | PredValueInfoTy LHSVals; | |||
929 | computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals, | |||
930 | WantInteger, RecursionSet, CxtI); | |||
931 | ||||
932 | for (const auto &LHSVal : LHSVals) { | |||
933 | Constant *V = LHSVal.first; | |||
934 | Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst); | |||
935 | if (Constant *KC = getKnownConstant(Folded, WantInteger)) | |||
936 | Result.emplace_back(KC, LHSVal.second); | |||
937 | } | |||
938 | ||||
939 | return !Result.empty(); | |||
940 | } | |||
941 | } | |||
942 | ||||
943 | if (SelectInst *SI = dyn_cast<SelectInst>(I)) { | |||
944 | // Handle select instructions where at least one operand is a known constant | |||
945 | // and we can figure out the condition value for any predecessor block. | |||
946 | Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference); | |||
947 | Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference); | |||
948 | PredValueInfoTy Conds; | |||
949 | if ((TrueVal || FalseVal) && | |||
950 | computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds, | |||
951 | WantInteger, RecursionSet, CxtI)) { | |||
952 | for (auto &C : Conds) { | |||
953 | Constant *Cond = C.first; | |||
954 | ||||
955 | // Figure out what value to use for the condition. | |||
956 | bool KnownCond; | |||
957 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) { | |||
958 | // A known boolean. | |||
959 | KnownCond = CI->isOne(); | |||
960 | } else { | |||
961 | assert(isa<UndefValue>(Cond) && "Unexpected condition value")(static_cast <bool> (isa<UndefValue>(Cond) && "Unexpected condition value") ? void (0) : __assert_fail ("isa<UndefValue>(Cond) && \"Unexpected condition value\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 961, __extension__ __PRETTY_FUNCTION__)); | |||
962 | // Either operand will do, so be sure to pick the one that's a known | |||
963 | // constant. | |||
964 | // FIXME: Do this more cleverly if both values are known constants? | |||
965 | KnownCond = (TrueVal != nullptr); | |||
966 | } | |||
967 | ||||
968 | // See if the select has a known constant value for this predecessor. | |||
969 | if (Constant *Val = KnownCond ? TrueVal : FalseVal) | |||
970 | Result.emplace_back(Val, C.second); | |||
971 | } | |||
972 | ||||
973 | return !Result.empty(); | |||
974 | } | |||
975 | } | |||
976 | ||||
977 | // If all else fails, see if LVI can figure out a constant value for us. | |||
978 | assert(CxtI->getParent() == BB && "CxtI should be in BB")(static_cast <bool> (CxtI->getParent() == BB && "CxtI should be in BB") ? void (0) : __assert_fail ("CxtI->getParent() == BB && \"CxtI should be in BB\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 978, __extension__ __PRETTY_FUNCTION__)); | |||
979 | Constant *CI = LVI->getConstant(V, CxtI); | |||
980 | if (Constant *KC = getKnownConstant(CI, Preference)) { | |||
981 | for (BasicBlock *Pred : predecessors(BB)) | |||
982 | Result.emplace_back(KC, Pred); | |||
983 | } | |||
984 | ||||
985 | return !Result.empty(); | |||
986 | } | |||
987 | ||||
988 | /// GetBestDestForBranchOnUndef - If we determine that the specified block ends | |||
989 | /// in an undefined jump, decide which block is best to revector to. | |||
990 | /// | |||
991 | /// Since we can pick an arbitrary destination, we pick the successor with the | |||
992 | /// fewest predecessors. This should reduce the in-degree of the others. | |||
993 | static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) { | |||
994 | Instruction *BBTerm = BB->getTerminator(); | |||
995 | unsigned MinSucc = 0; | |||
996 | BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); | |||
997 | // Compute the successor with the minimum number of predecessors. | |||
998 | unsigned MinNumPreds = pred_size(TestBB); | |||
999 | for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { | |||
1000 | TestBB = BBTerm->getSuccessor(i); | |||
1001 | unsigned NumPreds = pred_size(TestBB); | |||
1002 | if (NumPreds < MinNumPreds) { | |||
1003 | MinSucc = i; | |||
1004 | MinNumPreds = NumPreds; | |||
1005 | } | |||
1006 | } | |||
1007 | ||||
1008 | return MinSucc; | |||
1009 | } | |||
1010 | ||||
1011 | static bool hasAddressTakenAndUsed(BasicBlock *BB) { | |||
1012 | if (!BB->hasAddressTaken()) return false; | |||
1013 | ||||
1014 | // If the block has its address taken, it may be a tree of dead constants | |||
1015 | // hanging off of it. These shouldn't keep the block alive. | |||
1016 | BlockAddress *BA = BlockAddress::get(BB); | |||
1017 | BA->removeDeadConstantUsers(); | |||
1018 | return !BA->use_empty(); | |||
1019 | } | |||
1020 | ||||
1021 | /// processBlock - If there are any predecessors whose control can be threaded | |||
1022 | /// through to a successor, transform them now. | |||
1023 | bool JumpThreadingPass::processBlock(BasicBlock *BB) { | |||
1024 | // If the block is trivially dead, just return and let the caller nuke it. | |||
1025 | // This simplifies other transformations. | |||
1026 | if (DTU->isBBPendingDeletion(BB) || | |||
1027 | (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock())) | |||
1028 | return false; | |||
1029 | ||||
1030 | // If this block has a single predecessor, and if that pred has a single | |||
1031 | // successor, merge the blocks. This encourages recursive jump threading | |||
1032 | // because now the condition in this block can be threaded through | |||
1033 | // predecessors of our predecessor block. | |||
1034 | if (maybeMergeBasicBlockIntoOnlyPred(BB)) | |||
1035 | return true; | |||
1036 | ||||
1037 | if (tryToUnfoldSelectInCurrBB(BB)) | |||
1038 | return true; | |||
1039 | ||||
1040 | // Look if we can propagate guards to predecessors. | |||
1041 | if (HasGuards && processGuards(BB)) | |||
1042 | return true; | |||
1043 | ||||
1044 | // What kind of constant we're looking for. | |||
1045 | ConstantPreference Preference = WantInteger; | |||
1046 | ||||
1047 | // Look to see if the terminator is a conditional branch, switch or indirect | |||
1048 | // branch, if not we can't thread it. | |||
1049 | Value *Condition; | |||
1050 | Instruction *Terminator = BB->getTerminator(); | |||
1051 | if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) { | |||
1052 | // Can't thread an unconditional jump. | |||
1053 | if (BI->isUnconditional()) return false; | |||
1054 | Condition = BI->getCondition(); | |||
1055 | } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) { | |||
1056 | Condition = SI->getCondition(); | |||
1057 | } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) { | |||
1058 | // Can't thread indirect branch with no successors. | |||
1059 | if (IB->getNumSuccessors() == 0) return false; | |||
1060 | Condition = IB->getAddress()->stripPointerCasts(); | |||
1061 | Preference = WantBlockAddress; | |||
1062 | } else { | |||
1063 | return false; // Must be an invoke or callbr. | |||
1064 | } | |||
1065 | ||||
1066 | // Keep track if we constant folded the condition in this invocation. | |||
1067 | bool ConstantFolded = false; | |||
1068 | ||||
1069 | // Run constant folding to see if we can reduce the condition to a simple | |||
1070 | // constant. | |||
1071 | if (Instruction *I = dyn_cast<Instruction>(Condition)) { | |||
1072 | Value *SimpleVal = | |||
1073 | ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI); | |||
1074 | if (SimpleVal) { | |||
1075 | I->replaceAllUsesWith(SimpleVal); | |||
1076 | if (isInstructionTriviallyDead(I, TLI)) | |||
1077 | I->eraseFromParent(); | |||
1078 | Condition = SimpleVal; | |||
1079 | ConstantFolded = true; | |||
1080 | } | |||
1081 | } | |||
1082 | ||||
1083 | // If the terminator is branching on an undef or freeze undef, we can pick any | |||
1084 | // of the successors to branch to. Let getBestDestForJumpOnUndef decide. | |||
1085 | auto *FI = dyn_cast<FreezeInst>(Condition); | |||
1086 | if (isa<UndefValue>(Condition) || | |||
1087 | (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) { | |||
1088 | unsigned BestSucc = getBestDestForJumpOnUndef(BB); | |||
1089 | std::vector<DominatorTree::UpdateType> Updates; | |||
1090 | ||||
1091 | // Fold the branch/switch. | |||
1092 | Instruction *BBTerm = BB->getTerminator(); | |||
1093 | Updates.reserve(BBTerm->getNumSuccessors()); | |||
1094 | for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { | |||
1095 | if (i == BestSucc) continue; | |||
1096 | BasicBlock *Succ = BBTerm->getSuccessor(i); | |||
1097 | Succ->removePredecessor(BB, true); | |||
1098 | Updates.push_back({DominatorTree::Delete, BB, Succ}); | |||
1099 | } | |||
1100 | ||||
1101 | LLVM_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 (false) | |||
1102 | << "' folding undef terminator: " << *BBTerm << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " In block '" << BB->getName() << "' folding undef terminator: " << *BBTerm << '\n'; } } while (false); | |||
1103 | BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm); | |||
1104 | ++NumFolds; | |||
1105 | BBTerm->eraseFromParent(); | |||
1106 | DTU->applyUpdatesPermissive(Updates); | |||
1107 | if (FI) | |||
1108 | FI->eraseFromParent(); | |||
1109 | return true; | |||
1110 | } | |||
1111 | ||||
1112 | // If the terminator of this block is branching on a constant, simplify the | |||
1113 | // terminator to an unconditional branch. This can occur due to threading in | |||
1114 | // other blocks. | |||
1115 | if (getKnownConstant(Condition, Preference)) { | |||
1116 | LLVM_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 (false) | |||
1117 | << "' folding terminator: " << *BB->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " In block '" << BB->getName() << "' folding terminator: " << * BB->getTerminator() << '\n'; } } while (false) | |||
1118 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " In block '" << BB->getName() << "' folding terminator: " << * BB->getTerminator() << '\n'; } } while (false); | |||
1119 | ++NumFolds; | |||
1120 | ConstantFoldTerminator(BB, true, nullptr, DTU); | |||
1121 | if (HasProfileData) | |||
1122 | BPI->eraseBlock(BB); | |||
1123 | return true; | |||
1124 | } | |||
1125 | ||||
1126 | Instruction *CondInst = dyn_cast<Instruction>(Condition); | |||
1127 | ||||
1128 | // All the rest of our checks depend on the condition being an instruction. | |||
1129 | if (!CondInst) { | |||
1130 | // FIXME: Unify this with code below. | |||
1131 | if (processThreadableEdges(Condition, BB, Preference, Terminator)) | |||
1132 | return true; | |||
1133 | return ConstantFolded; | |||
1134 | } | |||
1135 | ||||
1136 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { | |||
1137 | // If we're branching on a conditional, LVI might be able to determine | |||
1138 | // it's value at the branch instruction. We only handle comparisons | |||
1139 | // against a constant at this time. | |||
1140 | // TODO: This should be extended to handle switches as well. | |||
1141 | BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); | |||
1142 | Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1)); | |||
1143 | if (CondBr && CondConst) { | |||
1144 | // We should have returned as soon as we turn a conditional branch to | |||
1145 | // unconditional. Because its no longer interesting as far as jump | |||
1146 | // threading is concerned. | |||
1147 | assert(CondBr->isConditional() && "Threading on unconditional terminator")(static_cast <bool> (CondBr->isConditional() && "Threading on unconditional terminator") ? void (0) : __assert_fail ("CondBr->isConditional() && \"Threading on unconditional terminator\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1147, __extension__ __PRETTY_FUNCTION__)); | |||
1148 | ||||
1149 | LazyValueInfo::Tristate Ret = | |||
1150 | LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0), | |||
1151 | CondConst, CondBr, /*UseBlockValue=*/false); | |||
1152 | if (Ret != LazyValueInfo::Unknown) { | |||
1153 | unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0; | |||
1154 | unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1; | |||
1155 | BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove); | |||
1156 | ToRemoveSucc->removePredecessor(BB, true); | |||
1157 | BranchInst *UncondBr = | |||
1158 | BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr); | |||
1159 | UncondBr->setDebugLoc(CondBr->getDebugLoc()); | |||
1160 | ++NumFolds; | |||
1161 | CondBr->eraseFromParent(); | |||
1162 | if (CondCmp->use_empty()) | |||
1163 | CondCmp->eraseFromParent(); | |||
1164 | // We can safely replace *some* uses of the CondInst if it has | |||
1165 | // exactly one value as returned by LVI. RAUW is incorrect in the | |||
1166 | // presence of guards and assumes, that have the `Cond` as the use. This | |||
1167 | // is because we use the guards/assume to reason about the `Cond` value | |||
1168 | // at the end of block, but RAUW unconditionally replaces all uses | |||
1169 | // including the guards/assumes themselves and the uses before the | |||
1170 | // guard/assume. | |||
1171 | else if (CondCmp->getParent() == BB) { | |||
1172 | auto *CI = Ret == LazyValueInfo::True ? | |||
1173 | ConstantInt::getTrue(CondCmp->getType()) : | |||
1174 | ConstantInt::getFalse(CondCmp->getType()); | |||
1175 | replaceFoldableUses(CondCmp, CI); | |||
1176 | } | |||
1177 | DTU->applyUpdatesPermissive( | |||
1178 | {{DominatorTree::Delete, BB, ToRemoveSucc}}); | |||
1179 | if (HasProfileData) | |||
1180 | BPI->eraseBlock(BB); | |||
1181 | return true; | |||
1182 | } | |||
1183 | ||||
1184 | // We did not manage to simplify this branch, try to see whether | |||
1185 | // CondCmp depends on a known phi-select pattern. | |||
1186 | if (tryToUnfoldSelect(CondCmp, BB)) | |||
1187 | return true; | |||
1188 | } | |||
1189 | } | |||
1190 | ||||
1191 | if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) | |||
1192 | if (tryToUnfoldSelect(SI, BB)) | |||
1193 | return true; | |||
1194 | ||||
1195 | // Check for some cases that are worth simplifying. Right now we want to look | |||
1196 | // for loads that are used by a switch or by the condition for the branch. If | |||
1197 | // we see one, check to see if it's partially redundant. If so, insert a PHI | |||
1198 | // which can then be used to thread the values. | |||
1199 | Value *SimplifyValue = CondInst; | |||
1200 | ||||
1201 | if (auto *FI = dyn_cast<FreezeInst>(SimplifyValue)) | |||
1202 | // Look into freeze's operand | |||
1203 | SimplifyValue = FI->getOperand(0); | |||
1204 | ||||
1205 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) | |||
1206 | if (isa<Constant>(CondCmp->getOperand(1))) | |||
1207 | SimplifyValue = CondCmp->getOperand(0); | |||
1208 | ||||
1209 | // TODO: There are other places where load PRE would be profitable, such as | |||
1210 | // more complex comparisons. | |||
1211 | if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue)) | |||
1212 | if (simplifyPartiallyRedundantLoad(LoadI)) | |||
1213 | return true; | |||
1214 | ||||
1215 | // Before threading, try to propagate profile data backwards: | |||
1216 | if (PHINode *PN = dyn_cast<PHINode>(CondInst)) | |||
1217 | if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) | |||
1218 | updatePredecessorProfileMetadata(PN, BB); | |||
1219 | ||||
1220 | // Handle a variety of cases where we are branching on something derived from | |||
1221 | // a PHI node in the current block. If we can prove that any predecessors | |||
1222 | // compute a predictable value based on a PHI node, thread those predecessors. | |||
1223 | if (processThreadableEdges(CondInst, BB, Preference, Terminator)) | |||
1224 | return true; | |||
1225 | ||||
1226 | // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in | |||
1227 | // the current block, see if we can simplify. | |||
1228 | PHINode *PN = dyn_cast<PHINode>( | |||
1229 | isa<FreezeInst>(CondInst) ? cast<FreezeInst>(CondInst)->getOperand(0) | |||
1230 | : CondInst); | |||
1231 | ||||
1232 | if (PN && PN->getParent() == BB && isa<BranchInst>(BB->getTerminator())) | |||
1233 | return processBranchOnPHI(PN); | |||
1234 | ||||
1235 | // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify. | |||
1236 | if (CondInst->getOpcode() == Instruction::Xor && | |||
1237 | CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator())) | |||
1238 | return processBranchOnXOR(cast<BinaryOperator>(CondInst)); | |||
1239 | ||||
1240 | // Search for a stronger dominating condition that can be used to simplify a | |||
1241 | // conditional branch leaving BB. | |||
1242 | if (processImpliedCondition(BB)) | |||
1243 | return true; | |||
1244 | ||||
1245 | return false; | |||
1246 | } | |||
1247 | ||||
1248 | bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) { | |||
1249 | auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); | |||
1250 | if (!BI || !BI->isConditional()) | |||
1251 | return false; | |||
1252 | ||||
1253 | Value *Cond = BI->getCondition(); | |||
1254 | BasicBlock *CurrentBB = BB; | |||
1255 | BasicBlock *CurrentPred = BB->getSinglePredecessor(); | |||
1256 | unsigned Iter = 0; | |||
1257 | ||||
1258 | auto &DL = BB->getModule()->getDataLayout(); | |||
1259 | ||||
1260 | while (CurrentPred && Iter++ < ImplicationSearchThreshold) { | |||
1261 | auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator()); | |||
1262 | if (!PBI || !PBI->isConditional()) | |||
1263 | return false; | |||
1264 | if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB) | |||
1265 | return false; | |||
1266 | ||||
1267 | bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB; | |||
1268 | Optional<bool> Implication = | |||
1269 | isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue); | |||
1270 | if (Implication) { | |||
1271 | BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1); | |||
1272 | BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0); | |||
1273 | RemoveSucc->removePredecessor(BB); | |||
1274 | BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI); | |||
1275 | UncondBI->setDebugLoc(BI->getDebugLoc()); | |||
1276 | ++NumFolds; | |||
1277 | BI->eraseFromParent(); | |||
1278 | DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}}); | |||
1279 | if (HasProfileData) | |||
1280 | BPI->eraseBlock(BB); | |||
1281 | return true; | |||
1282 | } | |||
1283 | CurrentBB = CurrentPred; | |||
1284 | CurrentPred = CurrentBB->getSinglePredecessor(); | |||
1285 | } | |||
1286 | ||||
1287 | return false; | |||
1288 | } | |||
1289 | ||||
1290 | /// Return true if Op is an instruction defined in the given block. | |||
1291 | static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) { | |||
1292 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) | |||
1293 | if (OpInst->getParent() == BB) | |||
1294 | return true; | |||
1295 | return false; | |||
1296 | } | |||
1297 | ||||
1298 | /// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially | |||
1299 | /// redundant load instruction, eliminate it by replacing it with a PHI node. | |||
1300 | /// This is an important optimization that encourages jump threading, and needs | |||
1301 | /// to be run interlaced with other jump threading tasks. | |||
1302 | bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) { | |||
1303 | // Don't hack volatile and ordered loads. | |||
1304 | if (!LoadI->isUnordered()) return false; | |||
| ||||
1305 | ||||
1306 | // If the load is defined in a block with exactly one predecessor, it can't be | |||
1307 | // partially redundant. | |||
1308 | BasicBlock *LoadBB = LoadI->getParent(); | |||
1309 | if (LoadBB->getSinglePredecessor()) | |||
1310 | return false; | |||
1311 | ||||
1312 | // If the load is defined in an EH pad, it can't be partially redundant, | |||
1313 | // because the edges between the invoke and the EH pad cannot have other | |||
1314 | // instructions between them. | |||
1315 | if (LoadBB->isEHPad()) | |||
1316 | return false; | |||
1317 | ||||
1318 | Value *LoadedPtr = LoadI->getOperand(0); | |||
1319 | ||||
1320 | // If the loaded operand is defined in the LoadBB and its not a phi, | |||
1321 | // it can't be available in predecessors. | |||
1322 | if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr)) | |||
1323 | return false; | |||
1324 | ||||
1325 | // Scan a few instructions up from the load, to see if it is obviously live at | |||
1326 | // the entry to its block. | |||
1327 | BasicBlock::iterator BBIt(LoadI); | |||
1328 | bool IsLoadCSE; | |||
1329 | if (Value *AvailableVal = FindAvailableLoadedValue( | |||
1330 | LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) { | |||
1331 | // If the value of the load is locally available within the block, just use | |||
1332 | // it. This frequently occurs for reg2mem'd allocas. | |||
1333 | ||||
1334 | if (IsLoadCSE) { | |||
1335 | LoadInst *NLoadI = cast<LoadInst>(AvailableVal); | |||
1336 | combineMetadataForCSE(NLoadI, LoadI, false); | |||
1337 | }; | |||
1338 | ||||
1339 | // If the returned value is the load itself, replace with an undef. This can | |||
1340 | // only happen in dead loops. | |||
1341 | if (AvailableVal == LoadI) | |||
1342 | AvailableVal = UndefValue::get(LoadI->getType()); | |||
1343 | if (AvailableVal->getType() != LoadI->getType()) | |||
1344 | AvailableVal = CastInst::CreateBitOrPointerCast( | |||
1345 | AvailableVal, LoadI->getType(), "", LoadI); | |||
1346 | LoadI->replaceAllUsesWith(AvailableVal); | |||
1347 | LoadI->eraseFromParent(); | |||
1348 | return true; | |||
1349 | } | |||
1350 | ||||
1351 | // Otherwise, if we scanned the whole block and got to the top of the block, | |||
1352 | // we know the block is locally transparent to the load. If not, something | |||
1353 | // might clobber its value. | |||
1354 | if (BBIt != LoadBB->begin()) | |||
1355 | return false; | |||
1356 | ||||
1357 | // If all of the loads and stores that feed the value have the same AA tags, | |||
1358 | // then we can propagate them onto any newly inserted loads. | |||
1359 | AAMDNodes AATags = LoadI->getAAMetadata(); | |||
1360 | ||||
1361 | SmallPtrSet<BasicBlock*, 8> PredsScanned; | |||
1362 | ||||
1363 | using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>; | |||
1364 | ||||
1365 | AvailablePredsTy AvailablePreds; | |||
1366 | BasicBlock *OneUnavailablePred = nullptr; | |||
1367 | SmallVector<LoadInst*, 8> CSELoads; | |||
1368 | ||||
1369 | // If we got here, the loaded value is transparent through to the start of the | |||
1370 | // block. Check to see if it is available in any of the predecessor blocks. | |||
1371 | for (BasicBlock *PredBB : predecessors(LoadBB)) { | |||
1372 | // If we already scanned this predecessor, skip it. | |||
1373 | if (!PredsScanned.insert(PredBB).second) | |||
1374 | continue; | |||
1375 | ||||
1376 | BBIt = PredBB->end(); | |||
1377 | unsigned NumScanedInst = 0; | |||
1378 | Value *PredAvailable = nullptr; | |||
1379 | // NOTE: We don't CSE load that is volatile or anything stronger than | |||
1380 | // unordered, that should have been checked when we entered the function. | |||
1381 | assert(LoadI->isUnordered() &&(static_cast <bool> (LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads") ? void (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1382, __extension__ __PRETTY_FUNCTION__)) | |||
1382 | "Attempting to CSE volatile or atomic loads")(static_cast <bool> (LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads") ? void (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1382, __extension__ __PRETTY_FUNCTION__)); | |||
1383 | // If this is a load on a phi pointer, phi-translate it and search | |||
1384 | // for available load/store to the pointer in predecessors. | |||
1385 | Type *AccessTy = LoadI->getType(); | |||
1386 | const auto &DL = LoadI->getModule()->getDataLayout(); | |||
1387 | MemoryLocation Loc(LoadedPtr->DoPHITranslation(LoadBB, PredBB), | |||
1388 | LocationSize::precise(DL.getTypeStoreSize(AccessTy)), | |||
1389 | AATags); | |||
1390 | PredAvailable = findAvailablePtrLoadStore(Loc, AccessTy, LoadI->isAtomic(), | |||
1391 | PredBB, BBIt, DefMaxInstsToScan, | |||
1392 | AA, &IsLoadCSE, &NumScanedInst); | |||
1393 | ||||
1394 | // If PredBB has a single predecessor, continue scanning through the | |||
1395 | // single predecessor. | |||
1396 | BasicBlock *SinglePredBB = PredBB; | |||
1397 | while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() && | |||
1398 | NumScanedInst < DefMaxInstsToScan) { | |||
1399 | SinglePredBB = SinglePredBB->getSinglePredecessor(); | |||
1400 | if (SinglePredBB) { | |||
1401 | BBIt = SinglePredBB->end(); | |||
1402 | PredAvailable = findAvailablePtrLoadStore( | |||
1403 | Loc, AccessTy, LoadI->isAtomic(), SinglePredBB, BBIt, | |||
1404 | (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE, | |||
1405 | &NumScanedInst); | |||
1406 | } | |||
1407 | } | |||
1408 | ||||
1409 | if (!PredAvailable) { | |||
1410 | OneUnavailablePred = PredBB; | |||
1411 | continue; | |||
1412 | } | |||
1413 | ||||
1414 | if (IsLoadCSE) | |||
1415 | CSELoads.push_back(cast<LoadInst>(PredAvailable)); | |||
1416 | ||||
1417 | // If so, this load is partially redundant. Remember this info so that we | |||
1418 | // can create a PHI node. | |||
1419 | AvailablePreds.emplace_back(PredBB, PredAvailable); | |||
1420 | } | |||
1421 | ||||
1422 | // If the loaded value isn't available in any predecessor, it isn't partially | |||
1423 | // redundant. | |||
1424 | if (AvailablePreds.empty()) return false; | |||
1425 | ||||
1426 | // Okay, the loaded value is available in at least one (and maybe all!) | |||
1427 | // predecessors. If the value is unavailable in more than one unique | |||
1428 | // predecessor, we want to insert a merge block for those common predecessors. | |||
1429 | // This ensures that we only have to insert one reload, thus not increasing | |||
1430 | // code size. | |||
1431 | BasicBlock *UnavailablePred = nullptr; | |||
1432 | ||||
1433 | // If the value is unavailable in one of predecessors, we will end up | |||
1434 | // inserting a new instruction into them. It is only valid if all the | |||
1435 | // instructions before LoadI are guaranteed to pass execution to its | |||
1436 | // successor, or if LoadI is safe to speculate. | |||
1437 | // TODO: If this logic becomes more complex, and we will perform PRE insertion | |||
1438 | // farther than to a predecessor, we need to reuse the code from GVN's PRE. | |||
1439 | // It requires domination tree analysis, so for this simple case it is an | |||
1440 | // overkill. | |||
1441 | if (PredsScanned.size() != AvailablePreds.size() && | |||
1442 | !isSafeToSpeculativelyExecute(LoadI)) | |||
1443 | for (auto I = LoadBB->begin(); &*I != LoadI; ++I) | |||
1444 | if (!isGuaranteedToTransferExecutionToSuccessor(&*I)) | |||
1445 | return false; | |||
1446 | ||||
1447 | // If there is exactly one predecessor where the value is unavailable, the | |||
1448 | // already computed 'OneUnavailablePred' block is it. If it ends in an | |||
1449 | // unconditional branch, we know that it isn't a critical edge. | |||
1450 | if (PredsScanned.size() == AvailablePreds.size()+1 && | |||
1451 | OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { | |||
| ||||
1452 | UnavailablePred = OneUnavailablePred; | |||
1453 | } else if (PredsScanned.size() != AvailablePreds.size()) { | |||
1454 | // Otherwise, we had multiple unavailable predecessors or we had a critical | |||
1455 | // edge from the one. | |||
1456 | SmallVector<BasicBlock*, 8> PredsToSplit; | |||
1457 | SmallPtrSet<BasicBlock*, 8> AvailablePredSet; | |||
1458 | ||||
1459 | for (const auto &AvailablePred : AvailablePreds) | |||
1460 | AvailablePredSet.insert(AvailablePred.first); | |||
1461 | ||||
1462 | // Add all the unavailable predecessors to the PredsToSplit list. | |||
1463 | for (BasicBlock *P : predecessors(LoadBB)) { | |||
1464 | // If the predecessor is an indirect goto, we can't split the edge. | |||
1465 | // Same for CallBr. | |||
1466 | if (isa<IndirectBrInst>(P->getTerminator()) || | |||
1467 | isa<CallBrInst>(P->getTerminator())) | |||
1468 | return false; | |||
1469 | ||||
1470 | if (!AvailablePredSet.count(P)) | |||
1471 | PredsToSplit.push_back(P); | |||
1472 | } | |||
1473 | ||||
1474 | // Split them out to their own block. | |||
1475 | UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split"); | |||
1476 | } | |||
1477 | ||||
1478 | // If the value isn't available in all predecessors, then there will be | |||
1479 | // exactly one where it isn't available. Insert a load on that edge and add | |||
1480 | // it to the AvailablePreds list. | |||
1481 | if (UnavailablePred) { | |||
1482 | assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&(static_cast <bool> (UnavailablePred->getTerminator( )->getNumSuccessors() == 1 && "Can't handle critical edge here!" ) ? void (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1483, __extension__ __PRETTY_FUNCTION__)) | |||
1483 | "Can't handle critical edge here!")(static_cast <bool> (UnavailablePred->getTerminator( )->getNumSuccessors() == 1 && "Can't handle critical edge here!" ) ? void (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1483, __extension__ __PRETTY_FUNCTION__)); | |||
1484 | LoadInst *NewVal = new LoadInst( | |||
1485 | LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred), | |||
1486 | LoadI->getName() + ".pr", false, LoadI->getAlign(), | |||
1487 | LoadI->getOrdering(), LoadI->getSyncScopeID(), | |||
1488 | UnavailablePred->getTerminator()); | |||
1489 | NewVal->setDebugLoc(LoadI->getDebugLoc()); | |||
1490 | if (AATags) | |||
1491 | NewVal->setAAMetadata(AATags); | |||
1492 | ||||
1493 | AvailablePreds.emplace_back(UnavailablePred, NewVal); | |||
1494 | } | |||
1495 | ||||
1496 | // Now we know that each predecessor of this block has a value in | |||
1497 | // AvailablePreds, sort them for efficient access as we're walking the preds. | |||
1498 | array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); | |||
1499 | ||||
1500 | // Create a PHI node at the start of the block for the PRE'd load value. | |||
1501 | pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB); | |||
1502 | PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "", | |||
1503 | &LoadBB->front()); | |||
1504 | PN->takeName(LoadI); | |||
1505 | PN->setDebugLoc(LoadI->getDebugLoc()); | |||
1506 | ||||
1507 | // Insert new entries into the PHI for each predecessor. A single block may | |||
1508 | // have multiple entries here. | |||
1509 | for (pred_iterator PI = PB; PI != PE; ++PI) { | |||
1510 | BasicBlock *P = *PI; | |||
1511 | AvailablePredsTy::iterator I = | |||
1512 | llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr)); | |||
1513 | ||||
1514 | assert(I != AvailablePreds.end() && I->first == P &&(static_cast <bool> (I != AvailablePreds.end() && I->first == P && "Didn't find entry for predecessor!" ) ? void (0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1515, __extension__ __PRETTY_FUNCTION__)) | |||
1515 | "Didn't find entry for predecessor!")(static_cast <bool> (I != AvailablePreds.end() && I->first == P && "Didn't find entry for predecessor!" ) ? void (0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1515, __extension__ __PRETTY_FUNCTION__)); | |||
1516 | ||||
1517 | // If we have an available predecessor but it requires casting, insert the | |||
1518 | // cast in the predecessor and use the cast. Note that we have to update the | |||
1519 | // AvailablePreds vector as we go so that all of the PHI entries for this | |||
1520 | // predecessor use the same bitcast. | |||
1521 | Value *&PredV = I->second; | |||
1522 | if (PredV->getType() != LoadI->getType()) | |||
1523 | PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "", | |||
1524 | P->getTerminator()); | |||
1525 | ||||
1526 | PN->addIncoming(PredV, I->first); | |||
1527 | } | |||
1528 | ||||
1529 | for (LoadInst *PredLoadI : CSELoads) { | |||
1530 | combineMetadataForCSE(PredLoadI, LoadI, true); | |||
1531 | } | |||
1532 | ||||
1533 | LoadI->replaceAllUsesWith(PN); | |||
1534 | LoadI->eraseFromParent(); | |||
1535 | ||||
1536 | return true; | |||
1537 | } | |||
1538 | ||||
1539 | /// findMostPopularDest - The specified list contains multiple possible | |||
1540 | /// threadable destinations. Pick the one that occurs the most frequently in | |||
1541 | /// the list. | |||
1542 | static BasicBlock * | |||
1543 | findMostPopularDest(BasicBlock *BB, | |||
1544 | const SmallVectorImpl<std::pair<BasicBlock *, | |||
1545 | BasicBlock *>> &PredToDestList) { | |||
1546 | assert(!PredToDestList.empty())(static_cast <bool> (!PredToDestList.empty()) ? void (0 ) : __assert_fail ("!PredToDestList.empty()", "llvm/lib/Transforms/Scalar/JumpThreading.cpp" , 1546, __extension__ __PRETTY_FUNCTION__)); | |||
1547 | ||||
1548 | // Determine popularity. If there are multiple possible destinations, we | |||
1549 | // explicitly choose to ignore 'undef' destinations. We prefer to thread | |||
1550 | // blocks with known and real destinations to threading undef. We'll handle | |||
1551 | // them later if interesting. | |||
1552 | MapVector<BasicBlock *, unsigned> DestPopularity; | |||
1553 | ||||
1554 | // Populate DestPopularity with the successors in the order they appear in the | |||
1555 | // successor list. This way, we ensure determinism by iterating it in the | |||
1556 | // same order in std::max_element below. We map nullptr to 0 so that we can | |||
1557 | // return nullptr when PredToDestList contains nullptr only. | |||
1558 | DestPopularity[nullptr] = 0; | |||
1559 | for (auto *SuccBB : successors(BB)) | |||
1560 | DestPopularity[SuccBB] = 0; | |||
1561 | ||||
1562 | for (const auto &PredToDest : PredToDestList) | |||
1563 | if (PredToDest.second) | |||
1564 | DestPopularity[PredToDest.second]++; | |||
1565 | ||||
1566 | // Find the most popular dest. | |||
1567 | using VT = decltype(DestPopularity)::value_type; | |||
1568 | auto MostPopular = std::max_element( | |||
1569 | DestPopularity.begin(), DestPopularity.end(), | |||
1570 | [](const VT &L, const VT &R) { return L.second < R.second; }); | |||
1571 | ||||
1572 | // Okay, we have finally picked the most popular destination. | |||
1573 | return MostPopular->first; | |||
1574 | } | |||
1575 | ||||
1576 | // Try to evaluate the value of V when the control flows from PredPredBB to | |||
1577 | // BB->getSinglePredecessor() and then on to BB. | |||
1578 | Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB, | |||
1579 | BasicBlock *PredPredBB, | |||
1580 | Value *V) { | |||
1581 | BasicBlock *PredBB = BB->getSinglePredecessor(); | |||
1582 | assert(PredBB && "Expected a single predecessor")(static_cast <bool> (PredBB && "Expected a single predecessor" ) ? void (0) : __assert_fail ("PredBB && \"Expected a single predecessor\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1582, __extension__ __PRETTY_FUNCTION__)); | |||
1583 | ||||
1584 | if (Constant *Cst = dyn_cast<Constant>(V)) { | |||
1585 | return Cst; | |||
1586 | } | |||
1587 | ||||
1588 | // Consult LVI if V is not an instruction in BB or PredBB. | |||
1589 | Instruction *I = dyn_cast<Instruction>(V); | |||
1590 | if (!I || (I->getParent() != BB && I->getParent() != PredBB)) { | |||
1591 | return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr); | |||
1592 | } | |||
1593 | ||||
1594 | // Look into a PHI argument. | |||
1595 | if (PHINode *PHI = dyn_cast<PHINode>(V)) { | |||
1596 | if (PHI->getParent() == PredBB) | |||
1597 | return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB)); | |||
1598 | return nullptr; | |||
1599 | } | |||
1600 | ||||
1601 | // If we have a CmpInst, try to fold it for each incoming edge into PredBB. | |||
1602 | if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) { | |||
1603 | if (CondCmp->getParent() == BB) { | |||
1604 | Constant *Op0 = | |||
1605 | evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(0)); | |||
1606 | Constant *Op1 = | |||
1607 | evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(1)); | |||
1608 | if (Op0 && Op1) { | |||
1609 | return ConstantExpr::getCompare(CondCmp->getPredicate(), Op0, Op1); | |||
1610 | } | |||
1611 | } | |||
1612 | return nullptr; | |||
1613 | } | |||
1614 | ||||
1615 | return nullptr; | |||
1616 | } | |||
1617 | ||||
1618 | bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB, | |||
1619 | ConstantPreference Preference, | |||
1620 | Instruction *CxtI) { | |||
1621 | // If threading this would thread across a loop header, don't even try to | |||
1622 | // thread the edge. | |||
1623 | if (LoopHeaders.count(BB)) | |||
1624 | return false; | |||
1625 | ||||
1626 | PredValueInfoTy PredValues; | |||
1627 | if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference, | |||
1628 | CxtI)) { | |||
1629 | // We don't have known values in predecessors. See if we can thread through | |||
1630 | // BB and its sole predecessor. | |||
1631 | return maybethreadThroughTwoBasicBlocks(BB, Cond); | |||
1632 | } | |||
1633 | ||||
1634 | assert(!PredValues.empty() &&(static_cast <bool> (!PredValues.empty() && "computeValueKnownInPredecessors returned true with no values" ) ? void (0) : __assert_fail ("!PredValues.empty() && \"computeValueKnownInPredecessors returned true with no values\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1635, __extension__ __PRETTY_FUNCTION__)) | |||
1635 | "computeValueKnownInPredecessors returned true with no values")(static_cast <bool> (!PredValues.empty() && "computeValueKnownInPredecessors returned true with no values" ) ? void (0) : __assert_fail ("!PredValues.empty() && \"computeValueKnownInPredecessors returned true with no values\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1635, __extension__ __PRETTY_FUNCTION__)); | |||
1636 | ||||
1637 | LLVM_DEBUG(dbgs() << "IN BB: " << *BB;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (const auto &PredValue : PredValues) { dbgs() << " BB '" << BB->getName() << "': FOUND condition = " << *PredValue.first << " for pred '" << PredValue .second->getName() << "'.\n"; }; } } while (false) | |||
1638 | for (const auto &PredValue : PredValues) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (const auto &PredValue : PredValues) { dbgs() << " BB '" << BB->getName() << "': FOUND condition = " << *PredValue.first << " for pred '" << PredValue .second->getName() << "'.\n"; }; } } while (false) | |||
1639 | dbgs() << " BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (const auto &PredValue : PredValues) { dbgs() << " BB '" << BB->getName() << "': FOUND condition = " << *PredValue.first << " for pred '" << PredValue .second->getName() << "'.\n"; }; } } while (false) | |||
1640 | << "': FOUND condition = " << *PredValue.firstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (const auto &PredValue : PredValues) { dbgs() << " BB '" << BB->getName() << "': FOUND condition = " << *PredValue.first << " for pred '" << PredValue .second->getName() << "'.\n"; }; } } while (false) | |||
1641 | << " for pred '" << PredValue.second->getName() << "'.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (const auto &PredValue : PredValues) { dbgs() << " BB '" << BB->getName() << "': FOUND condition = " << *PredValue.first << " for pred '" << PredValue .second->getName() << "'.\n"; }; } } while (false) | |||
1642 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "IN BB: " << *BB; for (const auto &PredValue : PredValues) { dbgs() << " BB '" << BB->getName() << "': FOUND condition = " << *PredValue.first << " for pred '" << PredValue .second->getName() << "'.\n"; }; } } while (false); | |||
1643 | ||||
1644 | // Decide what we want to thread through. Convert our list of known values to | |||
1645 | // a list of known destinations for each pred. This also discards duplicate | |||
1646 | // predecessors and keeps track of the undefined inputs (which are represented | |||
1647 | // as a null dest in the PredToDestList). | |||
1648 | SmallPtrSet<BasicBlock*, 16> SeenPreds; | |||
1649 | SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList; | |||
1650 | ||||
1651 | BasicBlock *OnlyDest = nullptr; | |||
1652 | BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL; | |||
1653 | Constant *OnlyVal = nullptr; | |||
1654 | Constant *MultipleVal = (Constant *)(intptr_t)~0ULL; | |||
1655 | ||||
1656 | for (const auto &PredValue : PredValues) { | |||
1657 | BasicBlock *Pred = PredValue.second; | |||
1658 | if (!SeenPreds.insert(Pred).second) | |||
1659 | continue; // Duplicate predecessor entry. | |||
1660 | ||||
1661 | Constant *Val = PredValue.first; | |||
1662 | ||||
1663 | BasicBlock *DestBB; | |||
1664 | if (isa<UndefValue>(Val)) | |||
1665 | DestBB = nullptr; | |||
1666 | else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { | |||
1667 | assert(isa<ConstantInt>(Val) && "Expecting a constant integer")(static_cast <bool> (isa<ConstantInt>(Val) && "Expecting a constant integer") ? void (0) : __assert_fail ( "isa<ConstantInt>(Val) && \"Expecting a constant integer\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1667, __extension__ __PRETTY_FUNCTION__)); | |||
1668 | DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero()); | |||
1669 | } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { | |||
1670 | assert(isa<ConstantInt>(Val) && "Expecting a constant integer")(static_cast <bool> (isa<ConstantInt>(Val) && "Expecting a constant integer") ? void (0) : __assert_fail ( "isa<ConstantInt>(Val) && \"Expecting a constant integer\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1670, __extension__ __PRETTY_FUNCTION__)); | |||
1671 | DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor(); | |||
1672 | } else { | |||
1673 | assert(isa<IndirectBrInst>(BB->getTerminator())(static_cast <bool> (isa<IndirectBrInst>(BB->getTerminator ()) && "Unexpected terminator") ? void (0) : __assert_fail ("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1674, __extension__ __PRETTY_FUNCTION__)) | |||
1674 | && "Unexpected terminator")(static_cast <bool> (isa<IndirectBrInst>(BB->getTerminator ()) && "Unexpected terminator") ? void (0) : __assert_fail ("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1674, __extension__ __PRETTY_FUNCTION__)); | |||
1675 | assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress")(static_cast <bool> (isa<BlockAddress>(Val) && "Expecting a constant blockaddress") ? void (0) : __assert_fail ("isa<BlockAddress>(Val) && \"Expecting a constant blockaddress\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1675, __extension__ __PRETTY_FUNCTION__)); | |||
1676 | DestBB = cast<BlockAddress>(Val)->getBasicBlock(); | |||
1677 | } | |||
1678 | ||||
1679 | // If we have exactly one destination, remember it for efficiency below. | |||
1680 | if (PredToDestList.empty()) { | |||
1681 | OnlyDest = DestBB; | |||
1682 | OnlyVal = Val; | |||
1683 | } else { | |||
1684 | if (OnlyDest != DestBB) | |||
1685 | OnlyDest = MultipleDestSentinel; | |||
1686 | // It possible we have same destination, but different value, e.g. default | |||
1687 | // case in switchinst. | |||
1688 | if (Val != OnlyVal) | |||
1689 | OnlyVal = MultipleVal; | |||
1690 | } | |||
1691 | ||||
1692 | // If the predecessor ends with an indirect goto, we can't change its | |||
1693 | // destination. Same for CallBr. | |||
1694 | if (isa<IndirectBrInst>(Pred->getTerminator()) || | |||
1695 | isa<CallBrInst>(Pred->getTerminator())) | |||
1696 | continue; | |||
1697 | ||||
1698 | PredToDestList.emplace_back(Pred, DestBB); | |||
1699 | } | |||
1700 | ||||
1701 | // If all edges were unthreadable, we fail. | |||
1702 | if (PredToDestList.empty()) | |||
1703 | return false; | |||
1704 | ||||
1705 | // If all the predecessors go to a single known successor, we want to fold, | |||
1706 | // not thread. By doing so, we do not need to duplicate the current block and | |||
1707 | // also miss potential opportunities in case we dont/cant duplicate. | |||
1708 | if (OnlyDest && OnlyDest != MultipleDestSentinel) { | |||
1709 | if (BB->hasNPredecessors(PredToDestList.size())) { | |||
1710 | bool SeenFirstBranchToOnlyDest = false; | |||
1711 | std::vector <DominatorTree::UpdateType> Updates; | |||
1712 | Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1); | |||
1713 | for (BasicBlock *SuccBB : successors(BB)) { | |||
1714 | if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) { | |||
1715 | SeenFirstBranchToOnlyDest = true; // Don't modify the first branch. | |||
1716 | } else { | |||
1717 | SuccBB->removePredecessor(BB, true); // This is unreachable successor. | |||
1718 | Updates.push_back({DominatorTree::Delete, BB, SuccBB}); | |||
1719 | } | |||
1720 | } | |||
1721 | ||||
1722 | // Finally update the terminator. | |||
1723 | Instruction *Term = BB->getTerminator(); | |||
1724 | BranchInst::Create(OnlyDest, Term); | |||
1725 | ++NumFolds; | |||
1726 | Term->eraseFromParent(); | |||
1727 | DTU->applyUpdatesPermissive(Updates); | |||
1728 | if (HasProfileData) | |||
1729 | BPI->eraseBlock(BB); | |||
1730 | ||||
1731 | // If the condition is now dead due to the removal of the old terminator, | |||
1732 | // erase it. | |||
1733 | if (auto *CondInst = dyn_cast<Instruction>(Cond)) { | |||
1734 | if (CondInst->use_empty() && !CondInst->mayHaveSideEffects()) | |||
1735 | CondInst->eraseFromParent(); | |||
1736 | // We can safely replace *some* uses of the CondInst if it has | |||
1737 | // exactly one value as returned by LVI. RAUW is incorrect in the | |||
1738 | // presence of guards and assumes, that have the `Cond` as the use. This | |||
1739 | // is because we use the guards/assume to reason about the `Cond` value | |||
1740 | // at the end of block, but RAUW unconditionally replaces all uses | |||
1741 | // including the guards/assumes themselves and the uses before the | |||
1742 | // guard/assume. | |||
1743 | else if (OnlyVal && OnlyVal != MultipleVal && | |||
1744 | CondInst->getParent() == BB) | |||
1745 | replaceFoldableUses(CondInst, OnlyVal); | |||
1746 | } | |||
1747 | return true; | |||
1748 | } | |||
1749 | } | |||
1750 | ||||
1751 | // Determine which is the most common successor. If we have many inputs and | |||
1752 | // this block is a switch, we want to start by threading the batch that goes | |||
1753 | // to the most popular destination first. If we only know about one | |||
1754 | // threadable destination (the common case) we can avoid this. | |||
1755 | BasicBlock *MostPopularDest = OnlyDest; | |||
1756 | ||||
1757 | if (MostPopularDest == MultipleDestSentinel) { | |||
1758 | // Remove any loop headers from the Dest list, threadEdge conservatively | |||
1759 | // won't process them, but we might have other destination that are eligible | |||
1760 | // and we still want to process. | |||
1761 | erase_if(PredToDestList, | |||
1762 | [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) { | |||
1763 | return LoopHeaders.contains(PredToDest.second); | |||
1764 | }); | |||
1765 | ||||
1766 | if (PredToDestList.empty()) | |||
1767 | return false; | |||
1768 | ||||
1769 | MostPopularDest = findMostPopularDest(BB, PredToDestList); | |||
1770 | } | |||
1771 | ||||
1772 | // Now that we know what the most popular destination is, factor all | |||
1773 | // predecessors that will jump to it into a single predecessor. | |||
1774 | SmallVector<BasicBlock*, 16> PredsToFactor; | |||
1775 | for (const auto &PredToDest : PredToDestList) | |||
1776 | if (PredToDest.second == MostPopularDest) { | |||
1777 | BasicBlock *Pred = PredToDest.first; | |||
1778 | ||||
1779 | // This predecessor may be a switch or something else that has multiple | |||
1780 | // edges to the block. Factor each of these edges by listing them | |||
1781 | // according to # occurrences in PredsToFactor. | |||
1782 | for (BasicBlock *Succ : successors(Pred)) | |||
1783 | if (Succ == BB) | |||
1784 | PredsToFactor.push_back(Pred); | |||
1785 | } | |||
1786 | ||||
1787 | // If the threadable edges are branching on an undefined value, we get to pick | |||
1788 | // the destination that these predecessors should get to. | |||
1789 | if (!MostPopularDest) | |||
1790 | MostPopularDest = BB->getTerminator()-> | |||
1791 | getSuccessor(getBestDestForJumpOnUndef(BB)); | |||
1792 | ||||
1793 | // Ok, try to thread it! | |||
1794 | return tryThreadEdge(BB, PredsToFactor, MostPopularDest); | |||
1795 | } | |||
1796 | ||||
1797 | /// processBranchOnPHI - We have an otherwise unthreadable conditional branch on | |||
1798 | /// a PHI node (or freeze PHI) in the current block. See if there are any | |||
1799 | /// simplifications we can do based on inputs to the phi node. | |||
1800 | bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) { | |||
1801 | BasicBlock *BB = PN->getParent(); | |||
1802 | ||||
1803 | // TODO: We could make use of this to do it once for blocks with common PHI | |||
1804 | // values. | |||
1805 | SmallVector<BasicBlock*, 1> PredBBs; | |||
1806 | PredBBs.resize(1); | |||
1807 | ||||
1808 | // If any of the predecessor blocks end in an unconditional branch, we can | |||
1809 | // *duplicate* the conditional branch into that block in order to further | |||
1810 | // encourage jump threading and to eliminate cases where we have branch on a | |||
1811 | // phi of an icmp (branch on icmp is much better). | |||
1812 | // This is still beneficial when a frozen phi is used as the branch condition | |||
1813 | // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp)) | |||
1814 | // to br(icmp(freeze ...)). | |||
1815 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { | |||
1816 | BasicBlock *PredBB = PN->getIncomingBlock(i); | |||
1817 | if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator())) | |||
1818 | if (PredBr->isUnconditional()) { | |||
1819 | PredBBs[0] = PredBB; | |||
1820 | // Try to duplicate BB into PredBB. | |||
1821 | if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs)) | |||
1822 | return true; | |||
1823 | } | |||
1824 | } | |||
1825 | ||||
1826 | return false; | |||
1827 | } | |||
1828 | ||||
1829 | /// processBranchOnXOR - We have an otherwise unthreadable conditional branch on | |||
1830 | /// a xor instruction in the current block. See if there are any | |||
1831 | /// simplifications we can do based on inputs to the xor. | |||
1832 | bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) { | |||
1833 | BasicBlock *BB = BO->getParent(); | |||
1834 | ||||
1835 | // If either the LHS or RHS of the xor is a constant, don't do this | |||
1836 | // optimization. | |||
1837 | if (isa<ConstantInt>(BO->getOperand(0)) || | |||
1838 | isa<ConstantInt>(BO->getOperand(1))) | |||
1839 | return false; | |||
1840 | ||||
1841 | // If the first instruction in BB isn't a phi, we won't be able to infer | |||
1842 | // anything special about any particular predecessor. | |||
1843 | if (!isa<PHINode>(BB->front())) | |||
1844 | return false; | |||
1845 | ||||
1846 | // If this BB is a landing pad, we won't be able to split the edge into it. | |||
1847 | if (BB->isEHPad()) | |||
1848 | return false; | |||
1849 | ||||
1850 | // If we have a xor as the branch input to this block, and we know that the | |||
1851 | // LHS or RHS of the xor in any predecessor is true/false, then we can clone | |||
1852 | // the condition into the predecessor and fix that value to true, saving some | |||
1853 | // logical ops on that path and encouraging other paths to simplify. | |||
1854 | // | |||
1855 | // This copies something like this: | |||
1856 | // | |||
1857 | // BB: | |||
1858 | // %X = phi i1 [1], [%X'] | |||
1859 | // %Y = icmp eq i32 %A, %B | |||
1860 | // %Z = xor i1 %X, %Y | |||
1861 | // br i1 %Z, ... | |||
1862 | // | |||
1863 | // Into: | |||
1864 | // BB': | |||
1865 | // %Y = icmp ne i32 %A, %B | |||
1866 | // br i1 %Y, ... | |||
1867 | ||||
1868 | PredValueInfoTy XorOpValues; | |||
1869 | bool isLHS = true; | |||
1870 | if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues, | |||
1871 | WantInteger, BO)) { | |||
1872 | assert(XorOpValues.empty())(static_cast <bool> (XorOpValues.empty()) ? void (0) : __assert_fail ("XorOpValues.empty()", "llvm/lib/Transforms/Scalar/JumpThreading.cpp" , 1872, __extension__ __PRETTY_FUNCTION__)); | |||
1873 | if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues, | |||
1874 | WantInteger, BO)) | |||
1875 | return false; | |||
1876 | isLHS = false; | |||
1877 | } | |||
1878 | ||||
1879 | assert(!XorOpValues.empty() &&(static_cast <bool> (!XorOpValues.empty() && "computeValueKnownInPredecessors returned true with no values" ) ? void (0) : __assert_fail ("!XorOpValues.empty() && \"computeValueKnownInPredecessors returned true with no values\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1880, __extension__ __PRETTY_FUNCTION__)) | |||
1880 | "computeValueKnownInPredecessors returned true with no values")(static_cast <bool> (!XorOpValues.empty() && "computeValueKnownInPredecessors returned true with no values" ) ? void (0) : __assert_fail ("!XorOpValues.empty() && \"computeValueKnownInPredecessors returned true with no values\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 1880, __extension__ __PRETTY_FUNCTION__)); | |||
1881 | ||||
1882 | // Scan the information to see which is most popular: true or false. The | |||
1883 | // predecessors can be of the set true, false, or undef. | |||
1884 | unsigned NumTrue = 0, NumFalse = 0; | |||
1885 | for (const auto &XorOpValue : XorOpValues) { | |||
1886 | if (isa<UndefValue>(XorOpValue.first)) | |||
1887 | // Ignore undefs for the count. | |||
1888 | continue; | |||
1889 | if (cast<ConstantInt>(XorOpValue.first)->isZero()) | |||
1890 | ++NumFalse; | |||
1891 | else | |||
1892 | ++NumTrue; | |||
1893 | } | |||
1894 | ||||
1895 | // Determine which value to split on, true, false, or undef if neither. | |||
1896 | ConstantInt *SplitVal = nullptr; | |||
1897 | if (NumTrue > NumFalse) | |||
1898 | SplitVal = ConstantInt::getTrue(BB->getContext()); | |||
1899 | else if (NumTrue != 0 || NumFalse != 0) | |||
1900 | SplitVal = ConstantInt::getFalse(BB->getContext()); | |||
1901 | ||||
1902 | // Collect all of the blocks that this can be folded into so that we can | |||
1903 | // factor this once and clone it once. | |||
1904 | SmallVector<BasicBlock*, 8> BlocksToFoldInto; | |||
1905 | for (const auto &XorOpValue : XorOpValues) { | |||
1906 | if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first)) | |||
1907 | continue; | |||
1908 | ||||
1909 | BlocksToFoldInto.push_back(XorOpValue.second); | |||
1910 | } | |||
1911 | ||||
1912 | // If we inferred a value for all of the predecessors, then duplication won't | |||
1913 | // help us. However, we can just replace the LHS or RHS with the constant. | |||
1914 | if (BlocksToFoldInto.size() == | |||
1915 | cast<PHINode>(BB->front()).getNumIncomingValues()) { | |||
1916 | if (!SplitVal) { | |||
1917 | // If all preds provide undef, just nuke the xor, because it is undef too. | |||
1918 | BO->replaceAllUsesWith(UndefValue::get(BO->getType())); | |||
1919 | BO->eraseFromParent(); | |||
1920 | } else if (SplitVal->isZero()) { | |||
1921 | // If all preds provide 0, replace the xor with the other input. | |||
1922 | BO->replaceAllUsesWith(BO->getOperand(isLHS)); | |||
1923 | BO->eraseFromParent(); | |||
1924 | } else { | |||
1925 | // If all preds provide 1, set the computed value to 1. | |||
1926 | BO->setOperand(!isLHS, SplitVal); | |||
1927 | } | |||
1928 | ||||
1929 | return true; | |||
1930 | } | |||
1931 | ||||
1932 | // If any of predecessors end with an indirect goto, we can't change its | |||
1933 | // destination. Same for CallBr. | |||
1934 | if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) { | |||
1935 | return isa<IndirectBrInst>(Pred->getTerminator()) || | |||
1936 | isa<CallBrInst>(Pred->getTerminator()); | |||
1937 | })) | |||
1938 | return false; | |||
1939 | ||||
1940 | // Try to duplicate BB into PredBB. | |||
1941 | return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto); | |||
1942 | } | |||
1943 | ||||
1944 | /// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new | |||
1945 | /// predecessor to the PHIBB block. If it has PHI nodes, add entries for | |||
1946 | /// NewPred using the entries from OldPred (suitably mapped). | |||
1947 | static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, | |||
1948 | BasicBlock *OldPred, | |||
1949 | BasicBlock *NewPred, | |||
1950 | DenseMap<Instruction*, Value*> &ValueMap) { | |||
1951 | for (PHINode &PN : PHIBB->phis()) { | |||
1952 | // Ok, we have a PHI node. Figure out what the incoming value was for the | |||
1953 | // DestBlock. | |||
1954 | Value *IV = PN.getIncomingValueForBlock(OldPred); | |||
1955 | ||||
1956 | // Remap the value if necessary. | |||
1957 | if (Instruction *Inst = dyn_cast<Instruction>(IV)) { | |||
1958 | DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst); | |||
1959 | if (I != ValueMap.end()) | |||
1960 | IV = I->second; | |||
1961 | } | |||
1962 | ||||
1963 | PN.addIncoming(IV, NewPred); | |||
1964 | } | |||
1965 | } | |||
1966 | ||||
1967 | /// Merge basic block BB into its sole predecessor if possible. | |||
1968 | bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) { | |||
1969 | BasicBlock *SinglePred = BB->getSinglePredecessor(); | |||
1970 | if (!SinglePred) | |||
1971 | return false; | |||
1972 | ||||
1973 | const Instruction *TI = SinglePred->getTerminator(); | |||
1974 | if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 || | |||
1975 | SinglePred == BB || hasAddressTakenAndUsed(BB)) | |||
1976 | return false; | |||
1977 | ||||
1978 | // If SinglePred was a loop header, BB becomes one. | |||
1979 | if (LoopHeaders.erase(SinglePred)) | |||
1980 | LoopHeaders.insert(BB); | |||
1981 | ||||
1982 | LVI->eraseBlock(SinglePred); | |||
1983 | MergeBasicBlockIntoOnlyPred(BB, DTU); | |||
1984 | ||||
1985 | // Now that BB is merged into SinglePred (i.e. SinglePred code followed by | |||
1986 | // BB code within one basic block `BB`), we need to invalidate the LVI | |||
1987 | // information associated with BB, because the LVI information need not be | |||
1988 | // true for all of BB after the merge. For example, | |||
1989 | // Before the merge, LVI info and code is as follows: | |||
1990 | // SinglePred: <LVI info1 for %p val> | |||
1991 | // %y = use of %p | |||
1992 | // call @exit() // need not transfer execution to successor. | |||
1993 | // assume(%p) // from this point on %p is true | |||
1994 | // br label %BB | |||
1995 | // BB: <LVI info2 for %p val, i.e. %p is true> | |||
1996 | // %x = use of %p | |||
1997 | // br label exit | |||
1998 | // | |||
1999 | // Note that this LVI info for blocks BB and SinglPred is correct for %p | |||
2000 | // (info2 and info1 respectively). After the merge and the deletion of the | |||
2001 | // LVI info1 for SinglePred. We have the following code: | |||
2002 | // BB: <LVI info2 for %p val> | |||
2003 | // %y = use of %p | |||
2004 | // call @exit() | |||
2005 | // assume(%p) | |||
2006 | // %x = use of %p <-- LVI info2 is correct from here onwards. | |||
2007 | // br label exit | |||
2008 | // LVI info2 for BB is incorrect at the beginning of BB. | |||
2009 | ||||
2010 | // Invalidate LVI information for BB if the LVI is not provably true for | |||
2011 | // all of BB. | |||
2012 | if (!isGuaranteedToTransferExecutionToSuccessor(BB)) | |||
2013 | LVI->eraseBlock(BB); | |||
2014 | return true; | |||
2015 | } | |||
2016 | ||||
2017 | /// Update the SSA form. NewBB contains instructions that are copied from BB. | |||
2018 | /// ValueMapping maps old values in BB to new ones in NewBB. | |||
2019 | void JumpThreadingPass::updateSSA( | |||
2020 | BasicBlock *BB, BasicBlock *NewBB, | |||
2021 | DenseMap<Instruction *, Value *> &ValueMapping) { | |||
2022 | // If there were values defined in BB that are used outside the block, then we | |||
2023 | // now have to update all uses of the value to use either the original value, | |||
2024 | // the cloned value, or some PHI derived value. This can require arbitrary | |||
2025 | // PHI insertion, of which we are prepared to do, clean these up now. | |||
2026 | SSAUpdater SSAUpdate; | |||
2027 | SmallVector<Use *, 16> UsesToRename; | |||
2028 | ||||
2029 | for (Instruction &I : *BB) { | |||
2030 | // Scan all uses of this instruction to see if it is used outside of its | |||
2031 | // block, and if so, record them in UsesToRename. | |||
2032 | for (Use &U : I.uses()) { | |||
2033 | Instruction *User = cast<Instruction>(U.getUser()); | |||
2034 | if (PHINode *UserPN = dyn_cast<PHINode>(User)) { | |||
2035 | if (UserPN->getIncomingBlock(U) == BB) | |||
2036 | continue; | |||
2037 | } else if (User->getParent() == BB) | |||
2038 | continue; | |||
2039 | ||||
2040 | UsesToRename.push_back(&U); | |||
2041 | } | |||
2042 | ||||
2043 | // If there are no uses outside the block, we're done with this instruction. | |||
2044 | if (UsesToRename.empty()) | |||
2045 | continue; | |||
2046 | LLVM_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 (false); | |||
2047 | ||||
2048 | // We found a use of I outside of BB. Rename all uses of I that are outside | |||
2049 | // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks | |||
2050 | // with the two values we know. | |||
2051 | SSAUpdate.Initialize(I.getType(), I.getName()); | |||
2052 | SSAUpdate.AddAvailableValue(BB, &I); | |||
2053 | SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]); | |||
2054 | ||||
2055 | while (!UsesToRename.empty()) | |||
2056 | SSAUpdate.RewriteUse(*UsesToRename.pop_back_val()); | |||
2057 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "\n"; } } while (false); | |||
2058 | } | |||
2059 | } | |||
2060 | ||||
2061 | /// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone | |||
2062 | /// arguments that come from PredBB. Return the map from the variables in the | |||
2063 | /// source basic block to the variables in the newly created basic block. | |||
2064 | DenseMap<Instruction *, Value *> | |||
2065 | JumpThreadingPass::cloneInstructions(BasicBlock::iterator BI, | |||
2066 | BasicBlock::iterator BE, BasicBlock *NewBB, | |||
2067 | BasicBlock *PredBB) { | |||
2068 | // We are going to have to map operands from the source basic block to the new | |||
2069 | // copy of the block 'NewBB'. If there are PHI nodes in the source basic | |||
2070 | // block, evaluate them to account for entry from PredBB. | |||
2071 | DenseMap<Instruction *, Value *> ValueMapping; | |||
2072 | ||||
2073 | // Clone the phi nodes of the source basic block into NewBB. The resulting | |||
2074 | // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater | |||
2075 | // might need to rewrite the operand of the cloned phi. | |||
2076 | for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) { | |||
2077 | PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB); | |||
2078 | NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB); | |||
2079 | ValueMapping[PN] = NewPN; | |||
2080 | } | |||
2081 | ||||
2082 | // Clone noalias scope declarations in the threaded block. When threading a | |||
2083 | // loop exit, we would otherwise end up with two idential scope declarations | |||
2084 | // visible at the same time. | |||
2085 | SmallVector<MDNode *> NoAliasScopes; | |||
2086 | DenseMap<MDNode *, MDNode *> ClonedScopes; | |||
2087 | LLVMContext &Context = PredBB->getContext(); | |||
2088 | identifyNoAliasScopesToClone(BI, BE, NoAliasScopes); | |||
2089 | cloneNoAliasScopes(NoAliasScopes, ClonedScopes, "thread", Context); | |||
2090 | ||||
2091 | // Clone the non-phi instructions of the source basic block into NewBB, | |||
2092 | // keeping track of the mapping and using it to remap operands in the cloned | |||
2093 | // instructions. | |||
2094 | for (; BI != BE; ++BI) { | |||
2095 | Instruction *New = BI->clone(); | |||
2096 | New->setName(BI->getName()); | |||
2097 | NewBB->getInstList().push_back(New); | |||
2098 | ValueMapping[&*BI] = New; | |||
2099 | adaptNoAliasScopes(New, ClonedScopes, Context); | |||
2100 | ||||
2101 | // Remap operands to patch up intra-block references. | |||
2102 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) | |||
2103 | if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { | |||
2104 | DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst); | |||
2105 | if (I != ValueMapping.end()) | |||
2106 | New->setOperand(i, I->second); | |||
2107 | } | |||
2108 | } | |||
2109 | ||||
2110 | return ValueMapping; | |||
2111 | } | |||
2112 | ||||
2113 | /// Attempt to thread through two successive basic blocks. | |||
2114 | bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB, | |||
2115 | Value *Cond) { | |||
2116 | // Consider: | |||
2117 | // | |||
2118 | // PredBB: | |||
2119 | // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ] | |||
2120 | // %tobool = icmp eq i32 %cond, 0 | |||
2121 | // br i1 %tobool, label %BB, label ... | |||
2122 | // | |||
2123 | // BB: | |||
2124 | // %cmp = icmp eq i32* %var, null | |||
2125 | // br i1 %cmp, label ..., label ... | |||
2126 | // | |||
2127 | // We don't know the value of %var at BB even if we know which incoming edge | |||
2128 | // we take to BB. However, once we duplicate PredBB for each of its incoming | |||
2129 | // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of | |||
2130 | // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB. | |||
2131 | ||||
2132 | // Require that BB end with a Branch for simplicity. | |||
2133 | BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); | |||
2134 | if (!CondBr) | |||
2135 | return false; | |||
2136 | ||||
2137 | // BB must have exactly one predecessor. | |||
2138 | BasicBlock *PredBB = BB->getSinglePredecessor(); | |||
2139 | if (!PredBB) | |||
2140 | return false; | |||
2141 | ||||
2142 | // Require that PredBB end with a conditional Branch. If PredBB ends with an | |||
2143 | // unconditional branch, we should be merging PredBB and BB instead. For | |||
2144 | // simplicity, we don't deal with a switch. | |||
2145 | BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); | |||
2146 | if (!PredBBBranch || PredBBBranch->isUnconditional()) | |||
2147 | return false; | |||
2148 | ||||
2149 | // If PredBB has exactly one incoming edge, we don't gain anything by copying | |||
2150 | // PredBB. | |||
2151 | if (PredBB->getSinglePredecessor()) | |||
2152 | return false; | |||
2153 | ||||
2154 | // Don't thread through PredBB if it contains a successor edge to itself, in | |||
2155 | // which case we would infinite loop. Suppose we are threading an edge from | |||
2156 | // PredPredBB through PredBB and BB to SuccBB with PredBB containing a | |||
2157 | // successor edge to itself. If we allowed jump threading in this case, we | |||
2158 | // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since | |||
2159 | // PredBB.thread has a successor edge to PredBB, we would immediately come up | |||
2160 | // with another jump threading opportunity from PredBB.thread through PredBB | |||
2161 | // and BB to SuccBB. This jump threading would repeatedly occur. That is, we | |||
2162 | // would keep peeling one iteration from PredBB. | |||
2163 | if (llvm::is_contained(successors(PredBB), PredBB)) | |||
2164 | return false; | |||
2165 | ||||
2166 | // Don't thread across a loop header. | |||
2167 | if (LoopHeaders.count(PredBB)) | |||
2168 | return false; | |||
2169 | ||||
2170 | // Avoid complication with duplicating EH pads. | |||
2171 | if (PredBB->isEHPad()) | |||
2172 | return false; | |||
2173 | ||||
2174 | // Find a predecessor that we can thread. For simplicity, we only consider a | |||
2175 | // successor edge out of BB to which we thread exactly one incoming edge into | |||
2176 | // PredBB. | |||
2177 | unsigned ZeroCount = 0; | |||
2178 | unsigned OneCount = 0; | |||
2179 | BasicBlock *ZeroPred = nullptr; | |||
2180 | BasicBlock *OnePred = nullptr; | |||
2181 | for (BasicBlock *P : predecessors(PredBB)) { | |||
2182 | if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>( | |||
2183 | evaluateOnPredecessorEdge(BB, P, Cond))) { | |||
2184 | if (CI->isZero()) { | |||
2185 | ZeroCount++; | |||
2186 | ZeroPred = P; | |||
2187 | } else if (CI->isOne()) { | |||
2188 | OneCount++; | |||
2189 | OnePred = P; | |||
2190 | } | |||
2191 | } | |||
2192 | } | |||
2193 | ||||
2194 | // Disregard complicated cases where we have to thread multiple edges. | |||
2195 | BasicBlock *PredPredBB; | |||
2196 | if (ZeroCount == 1) { | |||
2197 | PredPredBB = ZeroPred; | |||
2198 | } else if (OneCount == 1) { | |||
2199 | PredPredBB = OnePred; | |||
2200 | } else { | |||
2201 | return false; | |||
2202 | } | |||
2203 | ||||
2204 | BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred); | |||
2205 | ||||
2206 | // If threading to the same block as we come from, we would infinite loop. | |||
2207 | if (SuccBB == BB) { | |||
2208 | LLVM_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 (false) | |||
2209 | << "' - 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 (false); | |||
2210 | return false; | |||
2211 | } | |||
2212 | ||||
2213 | // If threading this would thread across a loop header, don't thread the edge. | |||
2214 | // See the comments above findLoopHeaders for justifications and caveats. | |||
2215 | if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) { | |||
2216 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2217 | bool BBIsHeader = LoopHeaders.count(BB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2218 | bool SuccIsHeader = LoopHeaders.count(SuccBB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2219 | dbgs() << " Not threading across "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2220 | << (BBIsHeader ? "loop header BB '" : "block BB '")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2221 | << BB->getName() << "' to dest "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2222 | << (SuccIsHeader ? "loop header BB '" : "block BB '")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2223 | << SuccBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2224 | << "' - it might create an irreducible loop!\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2225 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false); | |||
2226 | return false; | |||
2227 | } | |||
2228 | ||||
2229 | // Compute the cost of duplicating BB and PredBB. | |||
2230 | unsigned BBCost = getJumpThreadDuplicationCost( | |||
2231 | TTI, BB, BB->getTerminator(), BBDupThreshold); | |||
2232 | unsigned PredBBCost = getJumpThreadDuplicationCost( | |||
2233 | TTI, PredBB, PredBB->getTerminator(), BBDupThreshold); | |||
2234 | ||||
2235 | // Give up if costs are too high. We need to check BBCost and PredBBCost | |||
2236 | // individually before checking their sum because getJumpThreadDuplicationCost | |||
2237 | // return (unsigned)~0 for those basic blocks that cannot be duplicated. | |||
2238 | if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold || | |||
2239 | BBCost + PredBBCost > BBDupThreshold) { | |||
2240 | LLVM_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: " << PredBBCost << " for PredBB, " << BBCost << "for BB\n" ; } } while (false) | |||
2241 | << "' - Cost is too high: " << PredBBCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading BB '" << BB->getName() << "' - Cost is too high: " << PredBBCost << " for PredBB, " << BBCost << "for BB\n" ; } } while (false) | |||
2242 | << " for PredBB, " << BBCost << "for BB\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Not threading BB '" << BB->getName() << "' - Cost is too high: " << PredBBCost << " for PredBB, " << BBCost << "for BB\n" ; } } while (false); | |||
2243 | return false; | |||
2244 | } | |||
2245 | ||||
2246 | // Now we are ready to duplicate PredBB. | |||
2247 | threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB); | |||
2248 | return true; | |||
2249 | } | |||
2250 | ||||
2251 | void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB, | |||
2252 | BasicBlock *PredBB, | |||
2253 | BasicBlock *BB, | |||
2254 | BasicBlock *SuccBB) { | |||
2255 | LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading through '" << PredBB->getName() << "' and '" << BB ->getName() << "'\n"; } } while (false) | |||
2256 | << BB->getName() << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading through '" << PredBB->getName() << "' and '" << BB ->getName() << "'\n"; } } while (false); | |||
2257 | ||||
2258 | BranchInst *CondBr = cast<BranchInst>(BB->getTerminator()); | |||
2259 | BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator()); | |||
2260 | ||||
2261 | BasicBlock *NewBB = | |||
2262 | BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread", | |||
2263 | PredBB->getParent(), PredBB); | |||
2264 | NewBB->moveAfter(PredBB); | |||
2265 | ||||
2266 | // Set the block frequency of NewBB. | |||
2267 | if (HasProfileData) { | |||
2268 | auto NewBBFreq = BFI->getBlockFreq(PredPredBB) * | |||
2269 | BPI->getEdgeProbability(PredPredBB, PredBB); | |||
2270 | BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency()); | |||
2271 | } | |||
2272 | ||||
2273 | // We are going to have to map operands from the original BB block to the new | |||
2274 | // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them | |||
2275 | // to account for entry from PredPredBB. | |||
2276 | DenseMap<Instruction *, Value *> ValueMapping = | |||
2277 | cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB); | |||
2278 | ||||
2279 | // Copy the edge probabilities from PredBB to NewBB. | |||
2280 | if (HasProfileData) | |||
2281 | BPI->copyEdgeProbabilities(PredBB, NewBB); | |||
2282 | ||||
2283 | // Update the terminator of PredPredBB to jump to NewBB instead of PredBB. | |||
2284 | // This eliminates predecessors from PredPredBB, which requires us to simplify | |||
2285 | // any PHI nodes in PredBB. | |||
2286 | Instruction *PredPredTerm = PredPredBB->getTerminator(); | |||
2287 | for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i) | |||
2288 | if (PredPredTerm->getSuccessor(i) == PredBB) { | |||
2289 | PredBB->removePredecessor(PredPredBB, true); | |||
2290 | PredPredTerm->setSuccessor(i, NewBB); | |||
2291 | } | |||
2292 | ||||
2293 | addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB, | |||
2294 | ValueMapping); | |||
2295 | addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB, | |||
2296 | ValueMapping); | |||
2297 | ||||
2298 | DTU->applyUpdatesPermissive( | |||
2299 | {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)}, | |||
2300 | {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)}, | |||
2301 | {DominatorTree::Insert, PredPredBB, NewBB}, | |||
2302 | {DominatorTree::Delete, PredPredBB, PredBB}}); | |||
2303 | ||||
2304 | updateSSA(PredBB, NewBB, ValueMapping); | |||
2305 | ||||
2306 | // Clean up things like PHI nodes with single operands, dead instructions, | |||
2307 | // etc. | |||
2308 | SimplifyInstructionsInBlock(NewBB, TLI); | |||
2309 | SimplifyInstructionsInBlock(PredBB, TLI); | |||
2310 | ||||
2311 | SmallVector<BasicBlock *, 1> PredsToFactor; | |||
2312 | PredsToFactor.push_back(NewBB); | |||
2313 | threadEdge(BB, PredsToFactor, SuccBB); | |||
2314 | } | |||
2315 | ||||
2316 | /// tryThreadEdge - Thread an edge if it's safe and profitable to do so. | |||
2317 | bool JumpThreadingPass::tryThreadEdge( | |||
2318 | BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs, | |||
2319 | BasicBlock *SuccBB) { | |||
2320 | // If threading to the same block as we come from, we would infinite loop. | |||
2321 | if (SuccBB == BB) { | |||
2322 | LLVM_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 (false) | |||
2323 | << "' - 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 (false); | |||
2324 | return false; | |||
2325 | } | |||
2326 | ||||
2327 | // If threading this would thread across a loop header, don't thread the edge. | |||
2328 | // See the comments above findLoopHeaders for justifications and caveats. | |||
2329 | if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) { | |||
2330 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2331 | bool BBIsHeader = LoopHeaders.count(BB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2332 | bool SuccIsHeader = LoopHeaders.count(SuccBB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2333 | dbgs() << " Not threading across "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2334 | << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2335 | << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2336 | << SuccBB->getName() << "' - it might create an irreducible loop!\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false) | |||
2337 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB ); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() << " Not threading across " << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName() << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '") << SuccBB->getName() << "' - it might create an irreducible loop!\n" ; }; } } while (false); | |||
2338 | return false; | |||
2339 | } | |||
2340 | ||||
2341 | unsigned JumpThreadCost = getJumpThreadDuplicationCost( | |||
2342 | TTI, BB, BB->getTerminator(), BBDupThreshold); | |||
2343 | if (JumpThreadCost > BBDupThreshold) { | |||
2344 | LLVM_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 (false) | |||
2345 | << "' - 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 (false); | |||
2346 | return false; | |||
2347 | } | |||
2348 | ||||
2349 | threadEdge(BB, PredBBs, SuccBB); | |||
2350 | return true; | |||
2351 | } | |||
2352 | ||||
2353 | /// threadEdge - We have decided that it is safe and profitable to factor the | |||
2354 | /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB | |||
2355 | /// across BB. Transform the IR to reflect this change. | |||
2356 | void JumpThreadingPass::threadEdge(BasicBlock *BB, | |||
2357 | const SmallVectorImpl<BasicBlock *> &PredBBs, | |||
2358 | BasicBlock *SuccBB) { | |||
2359 | assert(SuccBB != BB && "Don't create an infinite loop")(static_cast <bool> (SuccBB != BB && "Don't create an infinite loop" ) ? void (0) : __assert_fail ("SuccBB != BB && \"Don't create an infinite loop\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2359, __extension__ __PRETTY_FUNCTION__)); | |||
2360 | ||||
2361 | assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&(static_cast <bool> (!LoopHeaders.count(BB) && ! LoopHeaders.count(SuccBB) && "Don't thread across loop headers" ) ? void (0) : __assert_fail ("!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) && \"Don't thread across loop headers\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2362, __extension__ __PRETTY_FUNCTION__)) | |||
2362 | "Don't thread across loop headers")(static_cast <bool> (!LoopHeaders.count(BB) && ! LoopHeaders.count(SuccBB) && "Don't thread across loop headers" ) ? void (0) : __assert_fail ("!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) && \"Don't thread across loop headers\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2362, __extension__ __PRETTY_FUNCTION__)); | |||
2363 | ||||
2364 | // And finally, do it! Start by factoring the predecessors if needed. | |||
2365 | BasicBlock *PredBB; | |||
2366 | if (PredBBs.size() == 1) | |||
2367 | PredBB = PredBBs[0]; | |||
2368 | else { | |||
2369 | LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (false) | |||
2370 | << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (false); | |||
2371 | PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm"); | |||
2372 | } | |||
2373 | ||||
2374 | // And finally, do it! | |||
2375 | LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" << SuccBB ->getName() << ", across block:\n " << *BB << "\n"; } } while (false) | |||
2376 | << "' to '" << SuccBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" << SuccBB ->getName() << ", across block:\n " << *BB << "\n"; } } while (false) | |||
2377 | << ", across block:\n " << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Threading edge from '" << PredBB->getName() << "' to '" << SuccBB ->getName() << ", across block:\n " << *BB << "\n"; } } while (false); | |||
2378 | ||||
2379 | LVI->threadEdge(PredBB, BB, SuccBB); | |||
2380 | ||||
2381 | BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), | |||
2382 | BB->getName()+".thread", | |||
2383 | BB->getParent(), BB); | |||
2384 | NewBB->moveAfter(PredBB); | |||
2385 | ||||
2386 | // Set the block frequency of NewBB. | |||
2387 | if (HasProfileData) { | |||
2388 | auto NewBBFreq = | |||
2389 | BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB); | |||
2390 | BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency()); | |||
2391 | } | |||
2392 | ||||
2393 | // Copy all the instructions from BB to NewBB except the terminator. | |||
2394 | DenseMap<Instruction *, Value *> ValueMapping = | |||
2395 | cloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB); | |||
2396 | ||||
2397 | // We didn't copy the terminator from BB over to NewBB, because there is now | |||
2398 | // an unconditional jump to SuccBB. Insert the unconditional jump. | |||
2399 | BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB); | |||
2400 | NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc()); | |||
2401 | ||||
2402 | // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the | |||
2403 | // PHI nodes for NewBB now. | |||
2404 | addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping); | |||
2405 | ||||
2406 | // Update the terminator of PredBB to jump to NewBB instead of BB. This | |||
2407 | // eliminates predecessors from BB, which requires us to simplify any PHI | |||
2408 | // nodes in BB. | |||
2409 | Instruction *PredTerm = PredBB->getTerminator(); | |||
2410 | for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) | |||
2411 | if (PredTerm->getSuccessor(i) == BB) { | |||
2412 | BB->removePredecessor(PredBB, true); | |||
2413 | PredTerm->setSuccessor(i, NewBB); | |||
2414 | } | |||
2415 | ||||
2416 | // Enqueue required DT updates. | |||
2417 | DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB}, | |||
2418 | {DominatorTree::Insert, PredBB, NewBB}, | |||
2419 | {DominatorTree::Delete, PredBB, BB}}); | |||
2420 | ||||
2421 | updateSSA(BB, NewBB, ValueMapping); | |||
2422 | ||||
2423 | // At this point, the IR is fully up to date and consistent. Do a quick scan | |||
2424 | // over the new instructions and zap any that are constants or dead. This | |||
2425 | // frequently happens because of phi translation. | |||
2426 | SimplifyInstructionsInBlock(NewBB, TLI); | |||
2427 | ||||
2428 | // Update the edge weight from BB to SuccBB, which should be less than before. | |||
2429 | updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB); | |||
2430 | ||||
2431 | // Threaded an edge! | |||
2432 | ++NumThreads; | |||
2433 | } | |||
2434 | ||||
2435 | /// Create a new basic block that will be the predecessor of BB and successor of | |||
2436 | /// all blocks in Preds. When profile data is available, update the frequency of | |||
2437 | /// this new block. | |||
2438 | BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB, | |||
2439 | ArrayRef<BasicBlock *> Preds, | |||
2440 | const char *Suffix) { | |||
2441 | SmallVector<BasicBlock *, 2> NewBBs; | |||
2442 | ||||
2443 | // Collect the frequencies of all predecessors of BB, which will be used to | |||
2444 | // update the edge weight of the result of splitting predecessors. | |||
2445 | DenseMap<BasicBlock *, BlockFrequency> FreqMap; | |||
2446 | if (HasProfileData) | |||
2447 | for (auto Pred : Preds) | |||
2448 | FreqMap.insert(std::make_pair( | |||
2449 | Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB))); | |||
2450 | ||||
2451 | // In the case when BB is a LandingPad block we create 2 new predecessors | |||
2452 | // instead of just one. | |||
2453 | if (BB->isLandingPad()) { | |||
2454 | std::string NewName = std::string(Suffix) + ".split-lp"; | |||
2455 | SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs); | |||
2456 | } else { | |||
2457 | NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix)); | |||
2458 | } | |||
2459 | ||||
2460 | std::vector<DominatorTree::UpdateType> Updates; | |||
2461 | Updates.reserve((2 * Preds.size()) + NewBBs.size()); | |||
2462 | for (auto NewBB : NewBBs) { | |||
2463 | BlockFrequency NewBBFreq(0); | |||
2464 | Updates.push_back({DominatorTree::Insert, NewBB, BB}); | |||
2465 | for (auto Pred : predecessors(NewBB)) { | |||
2466 | Updates.push_back({DominatorTree::Delete, Pred, BB}); | |||
2467 | Updates.push_back({DominatorTree::Insert, Pred, NewBB}); | |||
2468 | if (HasProfileData) // Update frequencies between Pred -> NewBB. | |||
2469 | NewBBFreq += FreqMap.lookup(Pred); | |||
2470 | } | |||
2471 | if (HasProfileData) // Apply the summed frequency to NewBB. | |||
2472 | BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency()); | |||
2473 | } | |||
2474 | ||||
2475 | DTU->applyUpdatesPermissive(Updates); | |||
2476 | return NewBBs[0]; | |||
2477 | } | |||
2478 | ||||
2479 | bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) { | |||
2480 | const Instruction *TI = BB->getTerminator(); | |||
2481 | assert(TI->getNumSuccessors() > 1 && "not a split")(static_cast <bool> (TI->getNumSuccessors() > 1 && "not a split") ? void (0) : __assert_fail ("TI->getNumSuccessors() > 1 && \"not a split\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2481, __extension__ __PRETTY_FUNCTION__)); | |||
2482 | ||||
2483 | MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof); | |||
2484 | if (!WeightsNode) | |||
2485 | return false; | |||
2486 | ||||
2487 | MDString *MDName = cast<MDString>(WeightsNode->getOperand(0)); | |||
2488 | if (MDName->getString() != "branch_weights") | |||
2489 | return false; | |||
2490 | ||||
2491 | // Ensure there are weights for all of the successors. Note that the first | |||
2492 | // operand to the metadata node is a name, not a weight. | |||
2493 | return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1; | |||
2494 | } | |||
2495 | ||||
2496 | /// Update the block frequency of BB and branch weight and the metadata on the | |||
2497 | /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 - | |||
2498 | /// Freq(PredBB->BB) / Freq(BB->SuccBB). | |||
2499 | void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB, | |||
2500 | BasicBlock *BB, | |||
2501 | BasicBlock *NewBB, | |||
2502 | BasicBlock *SuccBB) { | |||
2503 | if (!HasProfileData) | |||
2504 | return; | |||
2505 | ||||
2506 | assert(BFI && BPI && "BFI & BPI should have been created here")(static_cast <bool> (BFI && BPI && "BFI & BPI should have been created here" ) ? void (0) : __assert_fail ("BFI && BPI && \"BFI & BPI should have been created here\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2506, __extension__ __PRETTY_FUNCTION__)); | |||
2507 | ||||
2508 | // As the edge from PredBB to BB is deleted, we have to update the block | |||
2509 | // frequency of BB. | |||
2510 | auto BBOrigFreq = BFI->getBlockFreq(BB); | |||
2511 | auto NewBBFreq = BFI->getBlockFreq(NewBB); | |||
2512 | auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB); | |||
2513 | auto BBNewFreq = BBOrigFreq - NewBBFreq; | |||
2514 | BFI->setBlockFreq(BB, BBNewFreq.getFrequency()); | |||
2515 | ||||
2516 | // Collect updated outgoing edges' frequencies from BB and use them to update | |||
2517 | // edge probabilities. | |||
2518 | SmallVector<uint64_t, 4> BBSuccFreq; | |||
2519 | for (BasicBlock *Succ : successors(BB)) { | |||
2520 | auto SuccFreq = (Succ == SuccBB) | |||
2521 | ? BB2SuccBBFreq - NewBBFreq | |||
2522 | : BBOrigFreq * BPI->getEdgeProbability(BB, Succ); | |||
2523 | BBSuccFreq.push_back(SuccFreq.getFrequency()); | |||
2524 | } | |||
2525 | ||||
2526 | uint64_t MaxBBSuccFreq = | |||
2527 | *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end()); | |||
2528 | ||||
2529 | SmallVector<BranchProbability, 4> BBSuccProbs; | |||
2530 | if (MaxBBSuccFreq == 0) | |||
2531 | BBSuccProbs.assign(BBSuccFreq.size(), | |||
2532 | {1, static_cast<unsigned>(BBSuccFreq.size())}); | |||
2533 | else { | |||
2534 | for (uint64_t Freq : BBSuccFreq) | |||
2535 | BBSuccProbs.push_back( | |||
2536 | BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq)); | |||
2537 | // Normalize edge probabilities so that they sum up to one. | |||
2538 | BranchProbability::normalizeProbabilities(BBSuccProbs.begin(), | |||
2539 | BBSuccProbs.end()); | |||
2540 | } | |||
2541 | ||||
2542 | // Update edge probabilities in BPI. | |||
2543 | BPI->setEdgeProbability(BB, BBSuccProbs); | |||
2544 | ||||
2545 | // Update the profile metadata as well. | |||
2546 | // | |||
2547 | // Don't do this if the profile of the transformed blocks was statically | |||
2548 | // estimated. (This could occur despite the function having an entry | |||
2549 | // frequency in completely cold parts of the CFG.) | |||
2550 | // | |||
2551 | // In this case we don't want to suggest to subsequent passes that the | |||
2552 | // calculated weights are fully consistent. Consider this graph: | |||
2553 | // | |||
2554 | // check_1 | |||
2555 | // 50% / | | |||
2556 | // eq_1 | 50% | |||
2557 | // \ | | |||
2558 | // check_2 | |||
2559 | // 50% / | | |||
2560 | // eq_2 | 50% | |||
2561 | // \ | | |||
2562 | // check_3 | |||
2563 | // 50% / | | |||
2564 | // eq_3 | 50% | |||
2565 | // \ | | |||
2566 | // | |||
2567 | // Assuming the blocks check_* all compare the same value against 1, 2 and 3, | |||
2568 | // the overall probabilities are inconsistent; the total probability that the | |||
2569 | // value is either 1, 2 or 3 is 150%. | |||
2570 | // | |||
2571 | // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3 | |||
2572 | // becomes 0%. This is even worse if the edge whose probability becomes 0% is | |||
2573 | // the loop exit edge. Then based solely on static estimation we would assume | |||
2574 | // the loop was extremely hot. | |||
2575 | // | |||
2576 | // FIXME this locally as well so that BPI and BFI are consistent as well. We | |||
2577 | // shouldn't make edges extremely likely or unlikely based solely on static | |||
2578 | // estimation. | |||
2579 | if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) { | |||
2580 | SmallVector<uint32_t, 4> Weights; | |||
2581 | for (auto Prob : BBSuccProbs) | |||
2582 | Weights.push_back(Prob.getNumerator()); | |||
2583 | ||||
2584 | auto TI = BB->getTerminator(); | |||
2585 | TI->setMetadata( | |||
2586 | LLVMContext::MD_prof, | |||
2587 | MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights)); | |||
2588 | } | |||
2589 | } | |||
2590 | ||||
2591 | /// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch | |||
2592 | /// to BB which contains an i1 PHI node and a conditional branch on that PHI. | |||
2593 | /// If we can duplicate the contents of BB up into PredBB do so now, this | |||
2594 | /// improves the odds that the branch will be on an analyzable instruction like | |||
2595 | /// a compare. | |||
2596 | bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred( | |||
2597 | BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) { | |||
2598 | assert(!PredBBs.empty() && "Can't handle an empty set")(static_cast <bool> (!PredBBs.empty() && "Can't handle an empty set" ) ? void (0) : __assert_fail ("!PredBBs.empty() && \"Can't handle an empty set\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2598, __extension__ __PRETTY_FUNCTION__)); | |||
2599 | ||||
2600 | // If BB is a loop header, then duplicating this block outside the loop would | |||
2601 | // cause us to transform this into an irreducible loop, don't do this. | |||
2602 | // See the comments above findLoopHeaders for justifications and caveats. | |||
2603 | if (LoopHeaders.count(BB)) { | |||
2604 | LLVM_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 (false) | |||
2605 | << "' 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 (false) | |||
2606 | << "' - 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 (false); | |||
2607 | return false; | |||
2608 | } | |||
2609 | ||||
2610 | unsigned DuplicationCost = getJumpThreadDuplicationCost( | |||
2611 | TTI, BB, BB->getTerminator(), BBDupThreshold); | |||
2612 | if (DuplicationCost > BBDupThreshold) { | |||
2613 | LLVM_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 (false) | |||
2614 | << "' - 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 (false); | |||
2615 | return false; | |||
2616 | } | |||
2617 | ||||
2618 | // And finally, do it! Start by factoring the predecessors if needed. | |||
2619 | std::vector<DominatorTree::UpdateType> Updates; | |||
2620 | BasicBlock *PredBB; | |||
2621 | if (PredBBs.size() == 1) | |||
2622 | PredBB = PredBBs[0]; | |||
2623 | else { | |||
2624 | LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (false) | |||
2625 | << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << " Factoring out " << PredBBs.size() << " common predecessors.\n"; } } while (false); | |||
2626 | PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm"); | |||
2627 | } | |||
2628 | Updates.push_back({DominatorTree::Delete, PredBB, BB}); | |||
2629 | ||||
2630 | // Okay, we decided to do this! Clone all the instructions in BB onto the end | |||
2631 | // of PredBB. | |||
2632 | LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()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 (false) | |||
2633 | << "' into end of '" << PredBB->getName()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 (false) | |||
2634 | << "' 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 (false) | |||
2635 | << 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 (false); | |||
2636 | ||||
2637 | // Unless PredBB ends with an unconditional branch, split the edge so that we | |||
2638 | // can just clone the bits from BB into the end of the new PredBB. | |||
2639 | BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator()); | |||
2640 | ||||
2641 | if (!OldPredBranch || !OldPredBranch->isUnconditional()) { | |||
2642 | BasicBlock *OldPredBB = PredBB; | |||
2643 | PredBB = SplitEdge(OldPredBB, BB); | |||
2644 | Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB}); | |||
2645 | Updates.push_back({DominatorTree::Insert, PredBB, BB}); | |||
2646 | Updates.push_back({DominatorTree::Delete, OldPredBB, BB}); | |||
2647 | OldPredBranch = cast<BranchInst>(PredBB->getTerminator()); | |||
2648 | } | |||
2649 | ||||
2650 | // We are going to have to map operands from the original BB block into the | |||
2651 | // PredBB block. Evaluate PHI nodes in BB. | |||
2652 | DenseMap<Instruction*, Value*> ValueMapping; | |||
2653 | ||||
2654 | BasicBlock::iterator BI = BB->begin(); | |||
2655 | for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) | |||
2656 | ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); | |||
2657 | // Clone the non-phi instructions of BB into PredBB, keeping track of the | |||
2658 | // mapping and using it to remap operands in the cloned instructions. | |||
2659 | for (; BI != BB->end(); ++BI) { | |||
2660 | Instruction *New = BI->clone(); | |||
2661 | ||||
2662 | // Remap operands to patch up intra-block references. | |||
2663 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) | |||
2664 | if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { | |||
2665 | DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); | |||
2666 | if (I != ValueMapping.end()) | |||
2667 | New->setOperand(i, I->second); | |||
2668 | } | |||
2669 | ||||
2670 | // If this instruction can be simplified after the operands are updated, | |||
2671 | // just use the simplified value instead. This frequently happens due to | |||
2672 | // phi translation. | |||
2673 | if (Value *IV = SimplifyInstruction( | |||
2674 | New, | |||
2675 | {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) { | |||
2676 | ValueMapping[&*BI] = IV; | |||
2677 | if (!New->mayHaveSideEffects()) { | |||
2678 | New->deleteValue(); | |||
2679 | New = nullptr; | |||
2680 | } | |||
2681 | } else { | |||
2682 | ValueMapping[&*BI] = New; | |||
2683 | } | |||
2684 | if (New) { | |||
2685 | // Otherwise, insert the new instruction into the block. | |||
2686 | New->setName(BI->getName()); | |||
2687 | PredBB->getInstList().insert(OldPredBranch->getIterator(), New); | |||
2688 | // Update Dominance from simplified New instruction operands. | |||
2689 | for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) | |||
2690 | if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i))) | |||
2691 | Updates.push_back({DominatorTree::Insert, PredBB, SuccBB}); | |||
2692 | } | |||
2693 | } | |||
2694 | ||||
2695 | // Check to see if the targets of the branch had PHI nodes. If so, we need to | |||
2696 | // add entries to the PHI nodes for branch from PredBB now. | |||
2697 | BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator()); | |||
2698 | addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB, | |||
2699 | ValueMapping); | |||
2700 | addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB, | |||
2701 | ValueMapping); | |||
2702 | ||||
2703 | updateSSA(BB, PredBB, ValueMapping); | |||
2704 | ||||
2705 | // PredBB no longer jumps to BB, remove entries in the PHI node for the edge | |||
2706 | // that we nuked. | |||
2707 | BB->removePredecessor(PredBB, true); | |||
2708 | ||||
2709 | // Remove the unconditional branch at the end of the PredBB block. | |||
2710 | OldPredBranch->eraseFromParent(); | |||
2711 | if (HasProfileData) | |||
2712 | BPI->copyEdgeProbabilities(BB, PredBB); | |||
2713 | DTU->applyUpdatesPermissive(Updates); | |||
2714 | ||||
2715 | ++NumDupes; | |||
2716 | return true; | |||
2717 | } | |||
2718 | ||||
2719 | // Pred is a predecessor of BB with an unconditional branch to BB. SI is | |||
2720 | // a Select instruction in Pred. BB has other predecessors and SI is used in | |||
2721 | // a PHI node in BB. SI has no other use. | |||
2722 | // A new basic block, NewBB, is created and SI is converted to compare and | |||
2723 | // conditional branch. SI is erased from parent. | |||
2724 | void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB, | |||
2725 | SelectInst *SI, PHINode *SIUse, | |||
2726 | unsigned Idx) { | |||
2727 | // Expand the select. | |||
2728 | // | |||
2729 | // Pred -- | |||
2730 | // | v | |||
2731 | // | NewBB | |||
2732 | // | | | |||
2733 | // |----- | |||
2734 | // v | |||
2735 | // BB | |||
2736 | BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator()); | |||
2737 | BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold", | |||
2738 | BB->getParent(), BB); | |||
2739 | // Move the unconditional branch to NewBB. | |||
2740 | PredTerm->removeFromParent(); | |||
2741 | NewBB->getInstList().insert(NewBB->end(), PredTerm); | |||
2742 | // Create a conditional branch and update PHI nodes. | |||
2743 | auto *BI = BranchInst::Create(NewBB, BB, SI->getCondition(), Pred); | |||
2744 | BI->applyMergedLocation(PredTerm->getDebugLoc(), SI->getDebugLoc()); | |||
2745 | SIUse->setIncomingValue(Idx, SI->getFalseValue()); | |||
2746 | SIUse->addIncoming(SI->getTrueValue(), NewBB); | |||
2747 | ||||
2748 | // The select is now dead. | |||
2749 | SI->eraseFromParent(); | |||
2750 | DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB}, | |||
2751 | {DominatorTree::Insert, Pred, NewBB}}); | |||
2752 | ||||
2753 | // Update any other PHI nodes in BB. | |||
2754 | for (BasicBlock::iterator BI = BB->begin(); | |||
2755 | PHINode *Phi = dyn_cast<PHINode>(BI); ++BI) | |||
2756 | if (Phi != SIUse) | |||
2757 | Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB); | |||
2758 | } | |||
2759 | ||||
2760 | bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) { | |||
2761 | PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition()); | |||
2762 | ||||
2763 | if (!CondPHI || CondPHI->getParent() != BB) | |||
2764 | return false; | |||
2765 | ||||
2766 | for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) { | |||
2767 | BasicBlock *Pred = CondPHI->getIncomingBlock(I); | |||
2768 | SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I)); | |||
2769 | ||||
2770 | // The second and third condition can be potentially relaxed. Currently | |||
2771 | // the conditions help to simplify the code and allow us to reuse existing | |||
2772 | // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *) | |||
2773 | if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse()) | |||
2774 | continue; | |||
2775 | ||||
2776 | BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); | |||
2777 | if (!PredTerm || !PredTerm->isUnconditional()) | |||
2778 | continue; | |||
2779 | ||||
2780 | unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I); | |||
2781 | return true; | |||
2782 | } | |||
2783 | return false; | |||
2784 | } | |||
2785 | ||||
2786 | /// tryToUnfoldSelect - Look for blocks of the form | |||
2787 | /// bb1: | |||
2788 | /// %a = select | |||
2789 | /// br bb2 | |||
2790 | /// | |||
2791 | /// bb2: | |||
2792 | /// %p = phi [%a, %bb1] ... | |||
2793 | /// %c = icmp %p | |||
2794 | /// br i1 %c | |||
2795 | /// | |||
2796 | /// And expand the select into a branch structure if one of its arms allows %c | |||
2797 | /// to be folded. This later enables threading from bb1 over bb2. | |||
2798 | bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) { | |||
2799 | BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator()); | |||
2800 | PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0)); | |||
2801 | Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1)); | |||
2802 | ||||
2803 | if (!CondBr || !CondBr->isConditional() || !CondLHS || | |||
2804 | CondLHS->getParent() != BB) | |||
2805 | return false; | |||
2806 | ||||
2807 | for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) { | |||
2808 | BasicBlock *Pred = CondLHS->getIncomingBlock(I); | |||
2809 | SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I)); | |||
2810 | ||||
2811 | // Look if one of the incoming values is a select in the corresponding | |||
2812 | // predecessor. | |||
2813 | if (!SI || SI->getParent() != Pred || !SI->hasOneUse()) | |||
2814 | continue; | |||
2815 | ||||
2816 | BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator()); | |||
2817 | if (!PredTerm || !PredTerm->isUnconditional()) | |||
2818 | continue; | |||
2819 | ||||
2820 | // Now check if one of the select values would allow us to constant fold the | |||
2821 | // terminator in BB. We don't do the transform if both sides fold, those | |||
2822 | // cases will be threaded in any case. | |||
2823 | LazyValueInfo::Tristate LHSFolds = | |||
2824 | LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1), | |||
2825 | CondRHS, Pred, BB, CondCmp); | |||
2826 | LazyValueInfo::Tristate RHSFolds = | |||
2827 | LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2), | |||
2828 | CondRHS, Pred, BB, CondCmp); | |||
2829 | if ((LHSFolds != LazyValueInfo::Unknown || | |||
2830 | RHSFolds != LazyValueInfo::Unknown) && | |||
2831 | LHSFolds != RHSFolds) { | |||
2832 | unfoldSelectInstr(Pred, BB, SI, CondLHS, I); | |||
2833 | return true; | |||
2834 | } | |||
2835 | } | |||
2836 | return false; | |||
2837 | } | |||
2838 | ||||
2839 | /// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the | |||
2840 | /// same BB in the form | |||
2841 | /// bb: | |||
2842 | /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ... | |||
2843 | /// %s = select %p, trueval, falseval | |||
2844 | /// | |||
2845 | /// or | |||
2846 | /// | |||
2847 | /// bb: | |||
2848 | /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ... | |||
2849 | /// %c = cmp %p, 0 | |||
2850 | /// %s = select %c, trueval, falseval | |||
2851 | /// | |||
2852 | /// And expand the select into a branch structure. This later enables | |||
2853 | /// jump-threading over bb in this pass. | |||
2854 | /// | |||
2855 | /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold | |||
2856 | /// select if the associated PHI has at least one constant. If the unfolded | |||
2857 | /// select is not jump-threaded, it will be folded again in the later | |||
2858 | /// optimizations. | |||
2859 | bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) { | |||
2860 | // This transform would reduce the quality of msan diagnostics. | |||
2861 | // Disable this transform under MemorySanitizer. | |||
2862 | if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory)) | |||
2863 | return false; | |||
2864 | ||||
2865 | // If threading this would thread across a loop header, don't thread the edge. | |||
2866 | // See the comments above findLoopHeaders for justifications and caveats. | |||
2867 | if (LoopHeaders.count(BB)) | |||
2868 | return false; | |||
2869 | ||||
2870 | for (BasicBlock::iterator BI = BB->begin(); | |||
2871 | PHINode *PN = dyn_cast<PHINode>(BI); ++BI) { | |||
2872 | // Look for a Phi having at least one constant incoming value. | |||
2873 | if (llvm::all_of(PN->incoming_values(), | |||
2874 | [](Value *V) { return !isa<ConstantInt>(V); })) | |||
2875 | continue; | |||
2876 | ||||
2877 | auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) { | |||
2878 | using namespace PatternMatch; | |||
2879 | ||||
2880 | // Check if SI is in BB and use V as condition. | |||
2881 | if (SI->getParent() != BB) | |||
2882 | return false; | |||
2883 | Value *Cond = SI->getCondition(); | |||
2884 | bool IsAndOr = match(SI, m_CombineOr(m_LogicalAnd(), m_LogicalOr())); | |||
2885 | return Cond && Cond == V && Cond->getType()->isIntegerTy(1) && !IsAndOr; | |||
2886 | }; | |||
2887 | ||||
2888 | SelectInst *SI = nullptr; | |||
2889 | for (Use &U : PN->uses()) { | |||
2890 | if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) { | |||
2891 | // Look for a ICmp in BB that compares PN with a constant and is the | |||
2892 | // condition of a Select. | |||
2893 | if (Cmp->getParent() == BB && Cmp->hasOneUse() && | |||
2894 | isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo()))) | |||
2895 | if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back())) | |||
2896 | if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) { | |||
2897 | SI = SelectI; | |||
2898 | break; | |||
2899 | } | |||
2900 | } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) { | |||
2901 | // Look for a Select in BB that uses PN as condition. | |||
2902 | if (isUnfoldCandidate(SelectI, U.get())) { | |||
2903 | SI = SelectI; | |||
2904 | break; | |||
2905 | } | |||
2906 | } | |||
2907 | } | |||
2908 | ||||
2909 | if (!SI) | |||
2910 | continue; | |||
2911 | // Expand the select. | |||
2912 | Value *Cond = SI->getCondition(); | |||
2913 | if (InsertFreezeWhenUnfoldingSelect && | |||
2914 | !isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI, | |||
2915 | &DTU->getDomTree())) | |||
2916 | Cond = new FreezeInst(Cond, "cond.fr", SI); | |||
2917 | Instruction *Term = SplitBlockAndInsertIfThen(Cond, SI, false); | |||
2918 | BasicBlock *SplitBB = SI->getParent(); | |||
2919 | BasicBlock *NewBB = Term->getParent(); | |||
2920 | PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI); | |||
2921 | NewPN->addIncoming(SI->getTrueValue(), Term->getParent()); | |||
2922 | NewPN->addIncoming(SI->getFalseValue(), BB); | |||
2923 | SI->replaceAllUsesWith(NewPN); | |||
2924 | SI->eraseFromParent(); | |||
2925 | // NewBB and SplitBB are newly created blocks which require insertion. | |||
2926 | std::vector<DominatorTree::UpdateType> Updates; | |||
2927 | Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3); | |||
2928 | Updates.push_back({DominatorTree::Insert, BB, SplitBB}); | |||
2929 | Updates.push_back({DominatorTree::Insert, BB, NewBB}); | |||
2930 | Updates.push_back({DominatorTree::Insert, NewBB, SplitBB}); | |||
2931 | // BB's successors were moved to SplitBB, update DTU accordingly. | |||
2932 | for (auto *Succ : successors(SplitBB)) { | |||
2933 | Updates.push_back({DominatorTree::Delete, BB, Succ}); | |||
2934 | Updates.push_back({DominatorTree::Insert, SplitBB, Succ}); | |||
2935 | } | |||
2936 | DTU->applyUpdatesPermissive(Updates); | |||
2937 | return true; | |||
2938 | } | |||
2939 | return false; | |||
2940 | } | |||
2941 | ||||
2942 | /// Try to propagate a guard from the current BB into one of its predecessors | |||
2943 | /// in case if another branch of execution implies that the condition of this | |||
2944 | /// guard is always true. Currently we only process the simplest case that | |||
2945 | /// looks like: | |||
2946 | /// | |||
2947 | /// Start: | |||
2948 | /// %cond = ... | |||
2949 | /// br i1 %cond, label %T1, label %F1 | |||
2950 | /// T1: | |||
2951 | /// br label %Merge | |||
2952 | /// F1: | |||
2953 | /// br label %Merge | |||
2954 | /// Merge: | |||
2955 | /// %condGuard = ... | |||
2956 | /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ] | |||
2957 | /// | |||
2958 | /// And cond either implies condGuard or !condGuard. In this case all the | |||
2959 | /// instructions before the guard can be duplicated in both branches, and the | |||
2960 | /// guard is then threaded to one of them. | |||
2961 | bool JumpThreadingPass::processGuards(BasicBlock *BB) { | |||
2962 | using namespace PatternMatch; | |||
2963 | ||||
2964 | // We only want to deal with two predecessors. | |||
2965 | BasicBlock *Pred1, *Pred2; | |||
2966 | auto PI = pred_begin(BB), PE = pred_end(BB); | |||
2967 | if (PI == PE) | |||
2968 | return false; | |||
2969 | Pred1 = *PI++; | |||
2970 | if (PI == PE) | |||
2971 | return false; | |||
2972 | Pred2 = *PI++; | |||
2973 | if (PI != PE) | |||
2974 | return false; | |||
2975 | if (Pred1 == Pred2) | |||
2976 | return false; | |||
2977 | ||||
2978 | // Try to thread one of the guards of the block. | |||
2979 | // TODO: Look up deeper than to immediate predecessor? | |||
2980 | auto *Parent = Pred1->getSinglePredecessor(); | |||
2981 | if (!Parent || Parent != Pred2->getSinglePredecessor()) | |||
2982 | return false; | |||
2983 | ||||
2984 | if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator())) | |||
2985 | for (auto &I : *BB) | |||
2986 | if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI)) | |||
2987 | return true; | |||
2988 | ||||
2989 | return false; | |||
2990 | } | |||
2991 | ||||
2992 | /// Try to propagate the guard from BB which is the lower block of a diamond | |||
2993 | /// to one of its branches, in case if diamond's condition implies guard's | |||
2994 | /// condition. | |||
2995 | bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard, | |||
2996 | BranchInst *BI) { | |||
2997 | assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?")(static_cast <bool> (BI->getNumSuccessors() == 2 && "Wrong number of successors?") ? void (0) : __assert_fail ("BI->getNumSuccessors() == 2 && \"Wrong number of successors?\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2997, __extension__ __PRETTY_FUNCTION__)); | |||
2998 | assert(BI->isConditional() && "Unconditional branch has 2 successors?")(static_cast <bool> (BI->isConditional() && "Unconditional branch has 2 successors?" ) ? void (0) : __assert_fail ("BI->isConditional() && \"Unconditional branch has 2 successors?\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 2998, __extension__ __PRETTY_FUNCTION__)); | |||
2999 | Value *GuardCond = Guard->getArgOperand(0); | |||
3000 | Value *BranchCond = BI->getCondition(); | |||
3001 | BasicBlock *TrueDest = BI->getSuccessor(0); | |||
3002 | BasicBlock *FalseDest = BI->getSuccessor(1); | |||
3003 | ||||
3004 | auto &DL = BB->getModule()->getDataLayout(); | |||
3005 | bool TrueDestIsSafe = false; | |||
3006 | bool FalseDestIsSafe = false; | |||
3007 | ||||
3008 | // True dest is safe if BranchCond => GuardCond. | |||
3009 | auto Impl = isImpliedCondition(BranchCond, GuardCond, DL); | |||
3010 | if (Impl && *Impl) | |||
3011 | TrueDestIsSafe = true; | |||
3012 | else { | |||
3013 | // False dest is safe if !BranchCond => GuardCond. | |||
3014 | Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false); | |||
3015 | if (Impl && *Impl) | |||
3016 | FalseDestIsSafe = true; | |||
3017 | } | |||
3018 | ||||
3019 | if (!TrueDestIsSafe && !FalseDestIsSafe) | |||
3020 | return false; | |||
3021 | ||||
3022 | BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest; | |||
3023 | BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest; | |||
3024 | ||||
3025 | ValueToValueMapTy UnguardedMapping, GuardedMapping; | |||
3026 | Instruction *AfterGuard = Guard->getNextNode(); | |||
3027 | unsigned Cost = | |||
3028 | getJumpThreadDuplicationCost(TTI, BB, AfterGuard, BBDupThreshold); | |||
3029 | if (Cost > BBDupThreshold) | |||
3030 | return false; | |||
3031 | // Duplicate all instructions before the guard and the guard itself to the | |||
3032 | // branch where implication is not proved. | |||
3033 | BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween( | |||
3034 | BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU); | |||
3035 | assert(GuardedBlock && "Could not create the guarded block?")(static_cast <bool> (GuardedBlock && "Could not create the guarded block?" ) ? void (0) : __assert_fail ("GuardedBlock && \"Could not create the guarded block?\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 3035, __extension__ __PRETTY_FUNCTION__)); | |||
3036 | // Duplicate all instructions before the guard in the unguarded branch. | |||
3037 | // Since we have successfully duplicated the guarded block and this block | |||
3038 | // has fewer instructions, we expect it to succeed. | |||
3039 | BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween( | |||
3040 | BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU); | |||
3041 | assert(UnguardedBlock && "Could not create the unguarded block?")(static_cast <bool> (UnguardedBlock && "Could not create the unguarded block?" ) ? void (0) : __assert_fail ("UnguardedBlock && \"Could not create the unguarded block?\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 3041, __extension__ __PRETTY_FUNCTION__)); | |||
3042 | LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "Moved guard " << *Guard << " to block " << GuardedBlock->getName () << "\n"; } } while (false) | |||
3043 | << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("jump-threading")) { dbgs() << "Moved guard " << *Guard << " to block " << GuardedBlock->getName () << "\n"; } } while (false); | |||
3044 | // Some instructions before the guard may still have uses. For them, we need | |||
3045 | // to create Phi nodes merging their copies in both guarded and unguarded | |||
3046 | // branches. Those instructions that have no uses can be just removed. | |||
3047 | SmallVector<Instruction *, 4> ToRemove; | |||
3048 | for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI) | |||
3049 | if (!isa<PHINode>(&*BI)) | |||
3050 | ToRemove.push_back(&*BI); | |||
3051 | ||||
3052 | Instruction *InsertionPoint = &*BB->getFirstInsertionPt(); | |||
3053 | assert(InsertionPoint && "Empty block?")(static_cast <bool> (InsertionPoint && "Empty block?" ) ? void (0) : __assert_fail ("InsertionPoint && \"Empty block?\"" , "llvm/lib/Transforms/Scalar/JumpThreading.cpp", 3053, __extension__ __PRETTY_FUNCTION__)); | |||
3054 | // Substitute with Phis & remove. | |||
3055 | for (auto *Inst : reverse(ToRemove)) { | |||
3056 | if (!Inst->use_empty()) { | |||
3057 | PHINode *NewPN = PHINode::Create(Inst->getType(), 2); | |||
3058 | NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock); | |||
3059 | NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock); | |||
3060 | NewPN->insertBefore(InsertionPoint); | |||
3061 | Inst->replaceAllUsesWith(NewPN); | |||
3062 | } | |||
3063 | Inst->eraseFromParent(); | |||
3064 | } | |||
3065 | return true; | |||
3066 | } |