File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Instrumentation/ControlHeightReduction.cpp |
Warning: | line 939, column 42 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- ControlHeightReduction.cpp - Control Height Reduction -------------===// | ||||
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 pass merges conditional blocks of code and reduces the number of | ||||
10 | // conditional branches in the hot paths based on profiles. | ||||
11 | // | ||||
12 | //===----------------------------------------------------------------------===// | ||||
13 | |||||
14 | #include "llvm/Transforms/Instrumentation/ControlHeightReduction.h" | ||||
15 | #include "llvm/ADT/DenseMap.h" | ||||
16 | #include "llvm/ADT/DenseSet.h" | ||||
17 | #include "llvm/ADT/SmallVector.h" | ||||
18 | #include "llvm/ADT/StringSet.h" | ||||
19 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||
20 | #include "llvm/Analysis/GlobalsModRef.h" | ||||
21 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||
22 | #include "llvm/Analysis/ProfileSummaryInfo.h" | ||||
23 | #include "llvm/Analysis/RegionInfo.h" | ||||
24 | #include "llvm/Analysis/RegionIterator.h" | ||||
25 | #include "llvm/Analysis/ValueTracking.h" | ||||
26 | #include "llvm/IR/CFG.h" | ||||
27 | #include "llvm/IR/Dominators.h" | ||||
28 | #include "llvm/IR/IRBuilder.h" | ||||
29 | #include "llvm/IR/MDBuilder.h" | ||||
30 | #include "llvm/IR/PassManager.h" | ||||
31 | #include "llvm/InitializePasses.h" | ||||
32 | #include "llvm/Support/BranchProbability.h" | ||||
33 | #include "llvm/Support/CommandLine.h" | ||||
34 | #include "llvm/Support/MemoryBuffer.h" | ||||
35 | #include "llvm/Transforms/Utils.h" | ||||
36 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||
37 | #include "llvm/Transforms/Utils/Cloning.h" | ||||
38 | #include "llvm/Transforms/Utils/ValueMapper.h" | ||||
39 | |||||
40 | #include <set> | ||||
41 | #include <sstream> | ||||
42 | |||||
43 | using namespace llvm; | ||||
44 | |||||
45 | #define DEBUG_TYPE"chr" "chr" | ||||
46 | |||||
47 | #define CHR_DEBUG(X)do { } while (false) LLVM_DEBUG(X)do { } while (false) | ||||
48 | |||||
49 | static cl::opt<bool> ForceCHR("force-chr", cl::init(false), cl::Hidden, | ||||
50 | cl::desc("Apply CHR for all functions")); | ||||
51 | |||||
52 | static cl::opt<double> CHRBiasThreshold( | ||||
53 | "chr-bias-threshold", cl::init(0.99), cl::Hidden, | ||||
54 | cl::desc("CHR considers a branch bias greater than this ratio as biased")); | ||||
55 | |||||
56 | static cl::opt<unsigned> CHRMergeThreshold( | ||||
57 | "chr-merge-threshold", cl::init(2), cl::Hidden, | ||||
58 | cl::desc("CHR merges a group of N branches/selects where N >= this value")); | ||||
59 | |||||
60 | static cl::opt<std::string> CHRModuleList( | ||||
61 | "chr-module-list", cl::init(""), cl::Hidden, | ||||
62 | cl::desc("Specify file to retrieve the list of modules to apply CHR to")); | ||||
63 | |||||
64 | static cl::opt<std::string> CHRFunctionList( | ||||
65 | "chr-function-list", cl::init(""), cl::Hidden, | ||||
66 | cl::desc("Specify file to retrieve the list of functions to apply CHR to")); | ||||
67 | |||||
68 | static StringSet<> CHRModules; | ||||
69 | static StringSet<> CHRFunctions; | ||||
70 | |||||
71 | static void parseCHRFilterFiles() { | ||||
72 | if (!CHRModuleList.empty()) { | ||||
73 | auto FileOrErr = MemoryBuffer::getFile(CHRModuleList); | ||||
74 | if (!FileOrErr) { | ||||
75 | errs() << "Error: Couldn't read the chr-module-list file " << CHRModuleList << "\n"; | ||||
76 | std::exit(1); | ||||
77 | } | ||||
78 | StringRef Buf = FileOrErr->get()->getBuffer(); | ||||
79 | SmallVector<StringRef, 0> Lines; | ||||
80 | Buf.split(Lines, '\n'); | ||||
81 | for (StringRef Line : Lines) { | ||||
82 | Line = Line.trim(); | ||||
83 | if (!Line.empty()) | ||||
84 | CHRModules.insert(Line); | ||||
85 | } | ||||
86 | } | ||||
87 | if (!CHRFunctionList.empty()) { | ||||
88 | auto FileOrErr = MemoryBuffer::getFile(CHRFunctionList); | ||||
89 | if (!FileOrErr) { | ||||
90 | errs() << "Error: Couldn't read the chr-function-list file " << CHRFunctionList << "\n"; | ||||
91 | std::exit(1); | ||||
92 | } | ||||
93 | StringRef Buf = FileOrErr->get()->getBuffer(); | ||||
94 | SmallVector<StringRef, 0> Lines; | ||||
95 | Buf.split(Lines, '\n'); | ||||
96 | for (StringRef Line : Lines) { | ||||
97 | Line = Line.trim(); | ||||
98 | if (!Line.empty()) | ||||
99 | CHRFunctions.insert(Line); | ||||
100 | } | ||||
101 | } | ||||
102 | } | ||||
103 | |||||
104 | namespace { | ||||
105 | class ControlHeightReductionLegacyPass : public FunctionPass { | ||||
106 | public: | ||||
107 | static char ID; | ||||
108 | |||||
109 | ControlHeightReductionLegacyPass() : FunctionPass(ID) { | ||||
110 | initializeControlHeightReductionLegacyPassPass( | ||||
111 | *PassRegistry::getPassRegistry()); | ||||
112 | parseCHRFilterFiles(); | ||||
113 | } | ||||
114 | |||||
115 | bool runOnFunction(Function &F) override; | ||||
116 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||
117 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | ||||
118 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||
119 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | ||||
120 | AU.addRequired<RegionInfoPass>(); | ||||
121 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||
122 | } | ||||
123 | }; | ||||
124 | } // end anonymous namespace | ||||
125 | |||||
126 | char ControlHeightReductionLegacyPass::ID = 0; | ||||
127 | |||||
128 | INITIALIZE_PASS_BEGIN(ControlHeightReductionLegacyPass,static void *initializeControlHeightReductionLegacyPassPassOnce (PassRegistry &Registry) { | ||||
129 | "chr",static void *initializeControlHeightReductionLegacyPassPassOnce (PassRegistry &Registry) { | ||||
130 | "Reduce control height in the hot paths",static void *initializeControlHeightReductionLegacyPassPassOnce (PassRegistry &Registry) { | ||||
131 | false, false)static void *initializeControlHeightReductionLegacyPassPassOnce (PassRegistry &Registry) { | ||||
132 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | ||||
133 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||
134 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | ||||
135 | INITIALIZE_PASS_DEPENDENCY(RegionInfoPass)initializeRegionInfoPassPass(Registry); | ||||
136 | INITIALIZE_PASS_END(ControlHeightReductionLegacyPass,PassInfo *PI = new PassInfo( "Reduce control height in the hot paths" , "chr", &ControlHeightReductionLegacyPass::ID, PassInfo:: NormalCtor_t(callDefaultCtor<ControlHeightReductionLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeControlHeightReductionLegacyPassPassFlag ; void llvm::initializeControlHeightReductionLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeControlHeightReductionLegacyPassPassFlag , initializeControlHeightReductionLegacyPassPassOnce, std::ref (Registry)); } | ||||
137 | "chr",PassInfo *PI = new PassInfo( "Reduce control height in the hot paths" , "chr", &ControlHeightReductionLegacyPass::ID, PassInfo:: NormalCtor_t(callDefaultCtor<ControlHeightReductionLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeControlHeightReductionLegacyPassPassFlag ; void llvm::initializeControlHeightReductionLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeControlHeightReductionLegacyPassPassFlag , initializeControlHeightReductionLegacyPassPassOnce, std::ref (Registry)); } | ||||
138 | "Reduce control height in the hot paths",PassInfo *PI = new PassInfo( "Reduce control height in the hot paths" , "chr", &ControlHeightReductionLegacyPass::ID, PassInfo:: NormalCtor_t(callDefaultCtor<ControlHeightReductionLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeControlHeightReductionLegacyPassPassFlag ; void llvm::initializeControlHeightReductionLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeControlHeightReductionLegacyPassPassFlag , initializeControlHeightReductionLegacyPassPassOnce, std::ref (Registry)); } | ||||
139 | false, false)PassInfo *PI = new PassInfo( "Reduce control height in the hot paths" , "chr", &ControlHeightReductionLegacyPass::ID, PassInfo:: NormalCtor_t(callDefaultCtor<ControlHeightReductionLegacyPass >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeControlHeightReductionLegacyPassPassFlag ; void llvm::initializeControlHeightReductionLegacyPassPass(PassRegistry &Registry) { llvm::call_once(InitializeControlHeightReductionLegacyPassPassFlag , initializeControlHeightReductionLegacyPassPassOnce, std::ref (Registry)); } | ||||
140 | |||||
141 | FunctionPass *llvm::createControlHeightReductionLegacyPass() { | ||||
142 | return new ControlHeightReductionLegacyPass(); | ||||
143 | } | ||||
144 | |||||
145 | namespace { | ||||
146 | |||||
147 | struct CHRStats { | ||||
148 | CHRStats() : NumBranches(0), NumBranchesDelta(0), | ||||
149 | WeightedNumBranchesDelta(0) {} | ||||
150 | void print(raw_ostream &OS) const { | ||||
151 | OS << "CHRStats: NumBranches " << NumBranches | ||||
152 | << " NumBranchesDelta " << NumBranchesDelta | ||||
153 | << " WeightedNumBranchesDelta " << WeightedNumBranchesDelta; | ||||
154 | } | ||||
155 | uint64_t NumBranches; // The original number of conditional branches / | ||||
156 | // selects | ||||
157 | uint64_t NumBranchesDelta; // The decrease of the number of conditional | ||||
158 | // branches / selects in the hot paths due to CHR. | ||||
159 | uint64_t WeightedNumBranchesDelta; // NumBranchesDelta weighted by the profile | ||||
160 | // count at the scope entry. | ||||
161 | }; | ||||
162 | |||||
163 | // RegInfo - some properties of a Region. | ||||
164 | struct RegInfo { | ||||
165 | RegInfo() : R(nullptr), HasBranch(false) {} | ||||
166 | RegInfo(Region *RegionIn) : R(RegionIn), HasBranch(false) {} | ||||
167 | Region *R; | ||||
168 | bool HasBranch; | ||||
169 | SmallVector<SelectInst *, 8> Selects; | ||||
170 | }; | ||||
171 | |||||
172 | typedef DenseMap<Region *, DenseSet<Instruction *>> HoistStopMapTy; | ||||
173 | |||||
174 | // CHRScope - a sequence of regions to CHR together. It corresponds to a | ||||
175 | // sequence of conditional blocks. It can have subscopes which correspond to | ||||
176 | // nested conditional blocks. Nested CHRScopes form a tree. | ||||
177 | class CHRScope { | ||||
178 | public: | ||||
179 | CHRScope(RegInfo RI) : BranchInsertPoint(nullptr) { | ||||
180 | assert(RI.R && "Null RegionIn")((void)0); | ||||
181 | RegInfos.push_back(RI); | ||||
182 | } | ||||
183 | |||||
184 | Region *getParentRegion() { | ||||
185 | assert(RegInfos.size() > 0 && "Empty CHRScope")((void)0); | ||||
186 | Region *Parent = RegInfos[0].R->getParent(); | ||||
187 | assert(Parent && "Unexpected to call this on the top-level region")((void)0); | ||||
188 | return Parent; | ||||
189 | } | ||||
190 | |||||
191 | BasicBlock *getEntryBlock() { | ||||
192 | assert(RegInfos.size() > 0 && "Empty CHRScope")((void)0); | ||||
193 | return RegInfos.front().R->getEntry(); | ||||
194 | } | ||||
195 | |||||
196 | BasicBlock *getExitBlock() { | ||||
197 | assert(RegInfos.size() > 0 && "Empty CHRScope")((void)0); | ||||
198 | return RegInfos.back().R->getExit(); | ||||
199 | } | ||||
200 | |||||
201 | bool appendable(CHRScope *Next) { | ||||
202 | // The next scope is appendable only if this scope is directly connected to | ||||
203 | // it (which implies it post-dominates this scope) and this scope dominates | ||||
204 | // it (no edge to the next scope outside this scope). | ||||
205 | BasicBlock *NextEntry = Next->getEntryBlock(); | ||||
206 | if (getExitBlock() != NextEntry) | ||||
207 | // Not directly connected. | ||||
208 | return false; | ||||
209 | Region *LastRegion = RegInfos.back().R; | ||||
210 | for (BasicBlock *Pred : predecessors(NextEntry)) | ||||
211 | if (!LastRegion->contains(Pred)) | ||||
212 | // There's an edge going into the entry of the next scope from outside | ||||
213 | // of this scope. | ||||
214 | return false; | ||||
215 | return true; | ||||
216 | } | ||||
217 | |||||
218 | void append(CHRScope *Next) { | ||||
219 | assert(RegInfos.size() > 0 && "Empty CHRScope")((void)0); | ||||
220 | assert(Next->RegInfos.size() > 0 && "Empty CHRScope")((void)0); | ||||
221 | assert(getParentRegion() == Next->getParentRegion() &&((void)0) | ||||
222 | "Must be siblings")((void)0); | ||||
223 | assert(getExitBlock() == Next->getEntryBlock() &&((void)0) | ||||
224 | "Must be adjacent")((void)0); | ||||
225 | RegInfos.append(Next->RegInfos.begin(), Next->RegInfos.end()); | ||||
226 | Subs.append(Next->Subs.begin(), Next->Subs.end()); | ||||
227 | } | ||||
228 | |||||
229 | void addSub(CHRScope *SubIn) { | ||||
230 | #ifndef NDEBUG1 | ||||
231 | bool IsChild = false; | ||||
232 | for (RegInfo &RI : RegInfos) | ||||
233 | if (RI.R == SubIn->getParentRegion()) { | ||||
234 | IsChild = true; | ||||
235 | break; | ||||
236 | } | ||||
237 | assert(IsChild && "Must be a child")((void)0); | ||||
238 | #endif | ||||
239 | Subs.push_back(SubIn); | ||||
240 | } | ||||
241 | |||||
242 | // Split this scope at the boundary region into two, which will belong to the | ||||
243 | // tail and returns the tail. | ||||
244 | CHRScope *split(Region *Boundary) { | ||||
245 | assert(Boundary && "Boundary null")((void)0); | ||||
246 | assert(RegInfos.begin()->R != Boundary &&((void)0) | ||||
247 | "Can't be split at beginning")((void)0); | ||||
248 | auto BoundaryIt = llvm::find_if( | ||||
249 | RegInfos, [&Boundary](const RegInfo &RI) { return Boundary == RI.R; }); | ||||
250 | if (BoundaryIt == RegInfos.end()) | ||||
251 | return nullptr; | ||||
252 | ArrayRef<RegInfo> TailRegInfos(BoundaryIt, RegInfos.end()); | ||||
253 | DenseSet<Region *> TailRegionSet; | ||||
254 | for (const RegInfo &RI : TailRegInfos) | ||||
255 | TailRegionSet.insert(RI.R); | ||||
256 | |||||
257 | auto TailIt = | ||||
258 | std::stable_partition(Subs.begin(), Subs.end(), [&](CHRScope *Sub) { | ||||
259 | assert(Sub && "null Sub")((void)0); | ||||
260 | Region *Parent = Sub->getParentRegion(); | ||||
261 | if (TailRegionSet.count(Parent)) | ||||
262 | return false; | ||||
263 | |||||
264 | assert(llvm::any_of(((void)0) | ||||
265 | RegInfos,((void)0) | ||||
266 | [&Parent](const RegInfo &RI) { return Parent == RI.R; }) &&((void)0) | ||||
267 | "Must be in head")((void)0); | ||||
268 | return true; | ||||
269 | }); | ||||
270 | ArrayRef<CHRScope *> TailSubs(TailIt, Subs.end()); | ||||
271 | |||||
272 | assert(HoistStopMap.empty() && "MapHoistStops must be empty")((void)0); | ||||
273 | auto *Scope = new CHRScope(TailRegInfos, TailSubs); | ||||
274 | RegInfos.erase(BoundaryIt, RegInfos.end()); | ||||
275 | Subs.erase(TailIt, Subs.end()); | ||||
276 | return Scope; | ||||
277 | } | ||||
278 | |||||
279 | bool contains(Instruction *I) const { | ||||
280 | BasicBlock *Parent = I->getParent(); | ||||
281 | for (const RegInfo &RI : RegInfos) | ||||
282 | if (RI.R->contains(Parent)) | ||||
283 | return true; | ||||
284 | return false; | ||||
285 | } | ||||
286 | |||||
287 | void print(raw_ostream &OS) const; | ||||
288 | |||||
289 | SmallVector<RegInfo, 8> RegInfos; // Regions that belong to this scope | ||||
290 | SmallVector<CHRScope *, 8> Subs; // Subscopes. | ||||
291 | |||||
292 | // The instruction at which to insert the CHR conditional branch (and hoist | ||||
293 | // the dependent condition values). | ||||
294 | Instruction *BranchInsertPoint; | ||||
295 | |||||
296 | // True-biased and false-biased regions (conditional blocks), | ||||
297 | // respectively. Used only for the outermost scope and includes regions in | ||||
298 | // subscopes. The rest are unbiased. | ||||
299 | DenseSet<Region *> TrueBiasedRegions; | ||||
300 | DenseSet<Region *> FalseBiasedRegions; | ||||
301 | // Among the biased regions, the regions that get CHRed. | ||||
302 | SmallVector<RegInfo, 8> CHRRegions; | ||||
303 | |||||
304 | // True-biased and false-biased selects, respectively. Used only for the | ||||
305 | // outermost scope and includes ones in subscopes. | ||||
306 | DenseSet<SelectInst *> TrueBiasedSelects; | ||||
307 | DenseSet<SelectInst *> FalseBiasedSelects; | ||||
308 | |||||
309 | // Map from one of the above regions to the instructions to stop | ||||
310 | // hoisting instructions at through use-def chains. | ||||
311 | HoistStopMapTy HoistStopMap; | ||||
312 | |||||
313 | private: | ||||
314 | CHRScope(ArrayRef<RegInfo> RegInfosIn, ArrayRef<CHRScope *> SubsIn) | ||||
315 | : RegInfos(RegInfosIn.begin(), RegInfosIn.end()), | ||||
316 | Subs(SubsIn.begin(), SubsIn.end()), BranchInsertPoint(nullptr) {} | ||||
317 | }; | ||||
318 | |||||
319 | class CHR { | ||||
320 | public: | ||||
321 | CHR(Function &Fin, BlockFrequencyInfo &BFIin, DominatorTree &DTin, | ||||
322 | ProfileSummaryInfo &PSIin, RegionInfo &RIin, | ||||
323 | OptimizationRemarkEmitter &OREin) | ||||
324 | : F(Fin), BFI(BFIin), DT(DTin), PSI(PSIin), RI(RIin), ORE(OREin) {} | ||||
325 | |||||
326 | ~CHR() { | ||||
327 | for (CHRScope *Scope : Scopes) { | ||||
328 | delete Scope; | ||||
329 | } | ||||
330 | } | ||||
331 | |||||
332 | bool run(); | ||||
333 | |||||
334 | private: | ||||
335 | // See the comments in CHR::run() for the high level flow of the algorithm and | ||||
336 | // what the following functions do. | ||||
337 | |||||
338 | void findScopes(SmallVectorImpl<CHRScope *> &Output) { | ||||
339 | Region *R = RI.getTopLevelRegion(); | ||||
340 | if (CHRScope *Scope = findScopes(R, nullptr, nullptr, Output)) { | ||||
341 | Output.push_back(Scope); | ||||
342 | } | ||||
343 | } | ||||
344 | CHRScope *findScopes(Region *R, Region *NextRegion, Region *ParentRegion, | ||||
345 | SmallVectorImpl<CHRScope *> &Scopes); | ||||
346 | CHRScope *findScope(Region *R); | ||||
347 | void checkScopeHoistable(CHRScope *Scope); | ||||
348 | |||||
349 | void splitScopes(SmallVectorImpl<CHRScope *> &Input, | ||||
350 | SmallVectorImpl<CHRScope *> &Output); | ||||
351 | SmallVector<CHRScope *, 8> splitScope(CHRScope *Scope, | ||||
352 | CHRScope *Outer, | ||||
353 | DenseSet<Value *> *OuterConditionValues, | ||||
354 | Instruction *OuterInsertPoint, | ||||
355 | SmallVectorImpl<CHRScope *> &Output, | ||||
356 | DenseSet<Instruction *> &Unhoistables); | ||||
357 | |||||
358 | void classifyBiasedScopes(SmallVectorImpl<CHRScope *> &Scopes); | ||||
359 | void classifyBiasedScopes(CHRScope *Scope, CHRScope *OutermostScope); | ||||
360 | |||||
361 | void filterScopes(SmallVectorImpl<CHRScope *> &Input, | ||||
362 | SmallVectorImpl<CHRScope *> &Output); | ||||
363 | |||||
364 | void setCHRRegions(SmallVectorImpl<CHRScope *> &Input, | ||||
365 | SmallVectorImpl<CHRScope *> &Output); | ||||
366 | void setCHRRegions(CHRScope *Scope, CHRScope *OutermostScope); | ||||
367 | |||||
368 | void sortScopes(SmallVectorImpl<CHRScope *> &Input, | ||||
369 | SmallVectorImpl<CHRScope *> &Output); | ||||
370 | |||||
371 | void transformScopes(SmallVectorImpl<CHRScope *> &CHRScopes); | ||||
372 | void transformScopes(CHRScope *Scope, DenseSet<PHINode *> &TrivialPHIs); | ||||
373 | void cloneScopeBlocks(CHRScope *Scope, | ||||
374 | BasicBlock *PreEntryBlock, | ||||
375 | BasicBlock *ExitBlock, | ||||
376 | Region *LastRegion, | ||||
377 | ValueToValueMapTy &VMap); | ||||
378 | BranchInst *createMergedBranch(BasicBlock *PreEntryBlock, | ||||
379 | BasicBlock *EntryBlock, | ||||
380 | BasicBlock *NewEntryBlock, | ||||
381 | ValueToValueMapTy &VMap); | ||||
382 | void fixupBranchesAndSelects(CHRScope *Scope, | ||||
383 | BasicBlock *PreEntryBlock, | ||||
384 | BranchInst *MergedBR, | ||||
385 | uint64_t ProfileCount); | ||||
386 | void fixupBranch(Region *R, | ||||
387 | CHRScope *Scope, | ||||
388 | IRBuilder<> &IRB, | ||||
389 | Value *&MergedCondition, BranchProbability &CHRBranchBias); | ||||
390 | void fixupSelect(SelectInst* SI, | ||||
391 | CHRScope *Scope, | ||||
392 | IRBuilder<> &IRB, | ||||
393 | Value *&MergedCondition, BranchProbability &CHRBranchBias); | ||||
394 | void addToMergedCondition(bool IsTrueBiased, Value *Cond, | ||||
395 | Instruction *BranchOrSelect, | ||||
396 | CHRScope *Scope, | ||||
397 | IRBuilder<> &IRB, | ||||
398 | Value *&MergedCondition); | ||||
399 | |||||
400 | Function &F; | ||||
401 | BlockFrequencyInfo &BFI; | ||||
402 | DominatorTree &DT; | ||||
403 | ProfileSummaryInfo &PSI; | ||||
404 | RegionInfo &RI; | ||||
405 | OptimizationRemarkEmitter &ORE; | ||||
406 | CHRStats Stats; | ||||
407 | |||||
408 | // All the true-biased regions in the function | ||||
409 | DenseSet<Region *> TrueBiasedRegionsGlobal; | ||||
410 | // All the false-biased regions in the function | ||||
411 | DenseSet<Region *> FalseBiasedRegionsGlobal; | ||||
412 | // All the true-biased selects in the function | ||||
413 | DenseSet<SelectInst *> TrueBiasedSelectsGlobal; | ||||
414 | // All the false-biased selects in the function | ||||
415 | DenseSet<SelectInst *> FalseBiasedSelectsGlobal; | ||||
416 | // A map from biased regions to their branch bias | ||||
417 | DenseMap<Region *, BranchProbability> BranchBiasMap; | ||||
418 | // A map from biased selects to their branch bias | ||||
419 | DenseMap<SelectInst *, BranchProbability> SelectBiasMap; | ||||
420 | // All the scopes. | ||||
421 | DenseSet<CHRScope *> Scopes; | ||||
422 | }; | ||||
423 | |||||
424 | } // end anonymous namespace | ||||
425 | |||||
426 | static inline | ||||
427 | raw_ostream LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) &operator<<(raw_ostream &OS, | ||||
428 | const CHRStats &Stats) { | ||||
429 | Stats.print(OS); | ||||
430 | return OS; | ||||
431 | } | ||||
432 | |||||
433 | static inline | ||||
434 | raw_ostream &operator<<(raw_ostream &OS, const CHRScope &Scope) { | ||||
435 | Scope.print(OS); | ||||
436 | return OS; | ||||
437 | } | ||||
438 | |||||
439 | static bool shouldApply(Function &F, ProfileSummaryInfo& PSI) { | ||||
440 | if (ForceCHR) | ||||
441 | return true; | ||||
442 | |||||
443 | if (!CHRModuleList.empty() || !CHRFunctionList.empty()) { | ||||
444 | if (CHRModules.count(F.getParent()->getName())) | ||||
445 | return true; | ||||
446 | return CHRFunctions.count(F.getName()); | ||||
447 | } | ||||
448 | |||||
449 | assert(PSI.hasProfileSummary() && "Empty PSI?")((void)0); | ||||
450 | return PSI.isFunctionEntryHot(&F); | ||||
451 | } | ||||
452 | |||||
453 | static void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) dumpIR(Function &F, const char *Label, | ||||
454 | CHRStats *Stats) { | ||||
455 | StringRef FuncName = F.getName(); | ||||
456 | StringRef ModuleName = F.getParent()->getName(); | ||||
457 | (void)(FuncName); // Unused in release build. | ||||
458 | (void)(ModuleName); // Unused in release build. | ||||
459 | CHR_DEBUG(dbgs() << "CHR IR dump " << Label << " " << ModuleName << " "do { } while (false) | ||||
460 | << FuncName)do { } while (false); | ||||
461 | if (Stats) | ||||
462 | CHR_DEBUG(dbgs() << " " << *Stats)do { } while (false); | ||||
463 | CHR_DEBUG(dbgs() << "\n")do { } while (false); | ||||
464 | CHR_DEBUG(F.dump())do { } while (false); | ||||
465 | } | ||||
466 | |||||
467 | void CHRScope::print(raw_ostream &OS) const { | ||||
468 | assert(RegInfos.size() > 0 && "Empty CHRScope")((void)0); | ||||
469 | OS << "CHRScope["; | ||||
470 | OS << RegInfos.size() << ", Regions["; | ||||
471 | for (const RegInfo &RI : RegInfos) { | ||||
472 | OS << RI.R->getNameStr(); | ||||
473 | if (RI.HasBranch) | ||||
474 | OS << " B"; | ||||
475 | if (RI.Selects.size() > 0) | ||||
476 | OS << " S" << RI.Selects.size(); | ||||
477 | OS << ", "; | ||||
478 | } | ||||
479 | if (RegInfos[0].R->getParent()) { | ||||
480 | OS << "], Parent " << RegInfos[0].R->getParent()->getNameStr(); | ||||
481 | } else { | ||||
482 | // top level region | ||||
483 | OS << "]"; | ||||
484 | } | ||||
485 | OS << ", Subs["; | ||||
486 | for (CHRScope *Sub : Subs) { | ||||
487 | OS << *Sub << ", "; | ||||
488 | } | ||||
489 | OS << "]]"; | ||||
490 | } | ||||
491 | |||||
492 | // Return true if the given instruction type can be hoisted by CHR. | ||||
493 | static bool isHoistableInstructionType(Instruction *I) { | ||||
494 | return isa<BinaryOperator>(I) || isa<CastInst>(I) || isa<SelectInst>(I) || | ||||
495 | isa<GetElementPtrInst>(I) || isa<CmpInst>(I) || | ||||
496 | isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||
497 | isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) || | ||||
498 | isa<InsertValueInst>(I); | ||||
499 | } | ||||
500 | |||||
501 | // Return true if the given instruction can be hoisted by CHR. | ||||
502 | static bool isHoistable(Instruction *I, DominatorTree &DT) { | ||||
503 | if (!isHoistableInstructionType(I)) | ||||
504 | return false; | ||||
505 | return isSafeToSpeculativelyExecute(I, nullptr, &DT); | ||||
506 | } | ||||
507 | |||||
508 | // Recursively traverse the use-def chains of the given value and return a set | ||||
509 | // of the unhoistable base values defined within the scope (excluding the | ||||
510 | // first-region entry block) or the (hoistable or unhoistable) base values that | ||||
511 | // are defined outside (including the first-region entry block) of the | ||||
512 | // scope. The returned set doesn't include constants. | ||||
513 | static const std::set<Value *> & | ||||
514 | getBaseValues(Value *V, DominatorTree &DT, | ||||
515 | DenseMap<Value *, std::set<Value *>> &Visited) { | ||||
516 | auto It = Visited.find(V); | ||||
517 | if (It != Visited.end()) { | ||||
518 | return It->second; | ||||
519 | } | ||||
520 | std::set<Value *> Result; | ||||
521 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
522 | // We don't stop at a block that's not in the Scope because we would miss | ||||
523 | // some instructions that are based on the same base values if we stop | ||||
524 | // there. | ||||
525 | if (!isHoistable(I, DT)) { | ||||
526 | Result.insert(I); | ||||
527 | return Visited.insert(std::make_pair(V, std::move(Result))).first->second; | ||||
528 | } | ||||
529 | // I is hoistable above the Scope. | ||||
530 | for (Value *Op : I->operands()) { | ||||
531 | const std::set<Value *> &OpResult = getBaseValues(Op, DT, Visited); | ||||
532 | Result.insert(OpResult.begin(), OpResult.end()); | ||||
533 | } | ||||
534 | return Visited.insert(std::make_pair(V, std::move(Result))).first->second; | ||||
535 | } | ||||
536 | if (isa<Argument>(V)) { | ||||
537 | Result.insert(V); | ||||
538 | } | ||||
539 | // We don't include others like constants because those won't lead to any | ||||
540 | // chance of folding of conditions (eg two bit checks merged into one check) | ||||
541 | // after CHR. | ||||
542 | return Visited.insert(std::make_pair(V, std::move(Result))).first->second; | ||||
543 | } | ||||
544 | |||||
545 | // Return true if V is already hoisted or can be hoisted (along with its | ||||
546 | // operands) above the insert point. When it returns true and HoistStops is | ||||
547 | // non-null, the instructions to stop hoisting at through the use-def chains are | ||||
548 | // inserted into HoistStops. | ||||
549 | static bool | ||||
550 | checkHoistValue(Value *V, Instruction *InsertPoint, DominatorTree &DT, | ||||
551 | DenseSet<Instruction *> &Unhoistables, | ||||
552 | DenseSet<Instruction *> *HoistStops, | ||||
553 | DenseMap<Instruction *, bool> &Visited) { | ||||
554 | assert(InsertPoint && "Null InsertPoint")((void)0); | ||||
555 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
556 | auto It = Visited.find(I); | ||||
557 | if (It != Visited.end()) { | ||||
558 | return It->second; | ||||
559 | } | ||||
560 | assert(DT.getNode(I->getParent()) && "DT must contain I's parent block")((void)0); | ||||
561 | assert(DT.getNode(InsertPoint->getParent()) && "DT must contain Destination")((void)0); | ||||
562 | if (Unhoistables.count(I)) { | ||||
563 | // Don't hoist if they are not to be hoisted. | ||||
564 | Visited[I] = false; | ||||
565 | return false; | ||||
566 | } | ||||
567 | if (DT.dominates(I, InsertPoint)) { | ||||
568 | // We are already above the insert point. Stop here. | ||||
569 | if (HoistStops) | ||||
570 | HoistStops->insert(I); | ||||
571 | Visited[I] = true; | ||||
572 | return true; | ||||
573 | } | ||||
574 | // We aren't not above the insert point, check if we can hoist it above the | ||||
575 | // insert point. | ||||
576 | if (isHoistable(I, DT)) { | ||||
577 | // Check operands first. | ||||
578 | DenseSet<Instruction *> OpsHoistStops; | ||||
579 | bool AllOpsHoisted = true; | ||||
580 | for (Value *Op : I->operands()) { | ||||
581 | if (!checkHoistValue(Op, InsertPoint, DT, Unhoistables, &OpsHoistStops, | ||||
582 | Visited)) { | ||||
583 | AllOpsHoisted = false; | ||||
584 | break; | ||||
585 | } | ||||
586 | } | ||||
587 | if (AllOpsHoisted) { | ||||
588 | CHR_DEBUG(dbgs() << "checkHoistValue " << *I << "\n")do { } while (false); | ||||
589 | if (HoistStops) | ||||
590 | HoistStops->insert(OpsHoistStops.begin(), OpsHoistStops.end()); | ||||
591 | Visited[I] = true; | ||||
592 | return true; | ||||
593 | } | ||||
594 | } | ||||
595 | Visited[I] = false; | ||||
596 | return false; | ||||
597 | } | ||||
598 | // Non-instructions are considered hoistable. | ||||
599 | return true; | ||||
600 | } | ||||
601 | |||||
602 | // Returns true and sets the true probability and false probability of an | ||||
603 | // MD_prof metadata if it's well-formed. | ||||
604 | static bool checkMDProf(MDNode *MD, BranchProbability &TrueProb, | ||||
605 | BranchProbability &FalseProb) { | ||||
606 | if (!MD) return false; | ||||
607 | MDString *MDName = cast<MDString>(MD->getOperand(0)); | ||||
608 | if (MDName->getString() != "branch_weights" || | ||||
609 | MD->getNumOperands() != 3) | ||||
610 | return false; | ||||
611 | ConstantInt *TrueWeight = mdconst::extract<ConstantInt>(MD->getOperand(1)); | ||||
612 | ConstantInt *FalseWeight = mdconst::extract<ConstantInt>(MD->getOperand(2)); | ||||
613 | if (!TrueWeight || !FalseWeight) | ||||
614 | return false; | ||||
615 | uint64_t TrueWt = TrueWeight->getValue().getZExtValue(); | ||||
616 | uint64_t FalseWt = FalseWeight->getValue().getZExtValue(); | ||||
617 | uint64_t SumWt = TrueWt + FalseWt; | ||||
618 | |||||
619 | assert(SumWt >= TrueWt && SumWt >= FalseWt &&((void)0) | ||||
620 | "Overflow calculating branch probabilities.")((void)0); | ||||
621 | |||||
622 | // Guard against 0-to-0 branch weights to avoid a division-by-zero crash. | ||||
623 | if (SumWt == 0) | ||||
624 | return false; | ||||
625 | |||||
626 | TrueProb = BranchProbability::getBranchProbability(TrueWt, SumWt); | ||||
627 | FalseProb = BranchProbability::getBranchProbability(FalseWt, SumWt); | ||||
628 | return true; | ||||
629 | } | ||||
630 | |||||
631 | static BranchProbability getCHRBiasThreshold() { | ||||
632 | return BranchProbability::getBranchProbability( | ||||
633 | static_cast<uint64_t>(CHRBiasThreshold * 1000000), 1000000); | ||||
634 | } | ||||
635 | |||||
636 | // A helper for CheckBiasedBranch and CheckBiasedSelect. If TrueProb >= | ||||
637 | // CHRBiasThreshold, put Key into TrueSet and return true. If FalseProb >= | ||||
638 | // CHRBiasThreshold, put Key into FalseSet and return true. Otherwise, return | ||||
639 | // false. | ||||
640 | template <typename K, typename S, typename M> | ||||
641 | static bool checkBias(K *Key, BranchProbability TrueProb, | ||||
642 | BranchProbability FalseProb, S &TrueSet, S &FalseSet, | ||||
643 | M &BiasMap) { | ||||
644 | BranchProbability Threshold = getCHRBiasThreshold(); | ||||
645 | if (TrueProb >= Threshold) { | ||||
646 | TrueSet.insert(Key); | ||||
647 | BiasMap[Key] = TrueProb; | ||||
648 | return true; | ||||
649 | } else if (FalseProb >= Threshold) { | ||||
650 | FalseSet.insert(Key); | ||||
651 | BiasMap[Key] = FalseProb; | ||||
652 | return true; | ||||
653 | } | ||||
654 | return false; | ||||
655 | } | ||||
656 | |||||
657 | // Returns true and insert a region into the right biased set and the map if the | ||||
658 | // branch of the region is biased. | ||||
659 | static bool checkBiasedBranch(BranchInst *BI, Region *R, | ||||
660 | DenseSet<Region *> &TrueBiasedRegionsGlobal, | ||||
661 | DenseSet<Region *> &FalseBiasedRegionsGlobal, | ||||
662 | DenseMap<Region *, BranchProbability> &BranchBiasMap) { | ||||
663 | if (!BI->isConditional()) | ||||
664 | return false; | ||||
665 | BranchProbability ThenProb, ElseProb; | ||||
666 | if (!checkMDProf(BI->getMetadata(LLVMContext::MD_prof), | ||||
667 | ThenProb, ElseProb)) | ||||
668 | return false; | ||||
669 | BasicBlock *IfThen = BI->getSuccessor(0); | ||||
670 | BasicBlock *IfElse = BI->getSuccessor(1); | ||||
671 | assert((IfThen == R->getExit() || IfElse == R->getExit()) &&((void)0) | ||||
672 | IfThen != IfElse &&((void)0) | ||||
673 | "Invariant from findScopes")((void)0); | ||||
674 | if (IfThen == R->getExit()) { | ||||
675 | // Swap them so that IfThen/ThenProb means going into the conditional code | ||||
676 | // and IfElse/ElseProb means skipping it. | ||||
677 | std::swap(IfThen, IfElse); | ||||
678 | std::swap(ThenProb, ElseProb); | ||||
679 | } | ||||
680 | CHR_DEBUG(dbgs() << "BI " << *BI << " ")do { } while (false); | ||||
681 | CHR_DEBUG(dbgs() << "ThenProb " << ThenProb << " ")do { } while (false); | ||||
682 | CHR_DEBUG(dbgs() << "ElseProb " << ElseProb << "\n")do { } while (false); | ||||
683 | return checkBias(R, ThenProb, ElseProb, | ||||
684 | TrueBiasedRegionsGlobal, FalseBiasedRegionsGlobal, | ||||
685 | BranchBiasMap); | ||||
686 | } | ||||
687 | |||||
688 | // Returns true and insert a select into the right biased set and the map if the | ||||
689 | // select is biased. | ||||
690 | static bool checkBiasedSelect( | ||||
691 | SelectInst *SI, Region *R, | ||||
692 | DenseSet<SelectInst *> &TrueBiasedSelectsGlobal, | ||||
693 | DenseSet<SelectInst *> &FalseBiasedSelectsGlobal, | ||||
694 | DenseMap<SelectInst *, BranchProbability> &SelectBiasMap) { | ||||
695 | BranchProbability TrueProb, FalseProb; | ||||
696 | if (!checkMDProf(SI->getMetadata(LLVMContext::MD_prof), | ||||
697 | TrueProb, FalseProb)) | ||||
698 | return false; | ||||
699 | CHR_DEBUG(dbgs() << "SI " << *SI << " ")do { } while (false); | ||||
700 | CHR_DEBUG(dbgs() << "TrueProb " << TrueProb << " ")do { } while (false); | ||||
701 | CHR_DEBUG(dbgs() << "FalseProb " << FalseProb << "\n")do { } while (false); | ||||
702 | return checkBias(SI, TrueProb, FalseProb, | ||||
703 | TrueBiasedSelectsGlobal, FalseBiasedSelectsGlobal, | ||||
704 | SelectBiasMap); | ||||
705 | } | ||||
706 | |||||
707 | // Returns the instruction at which to hoist the dependent condition values and | ||||
708 | // insert the CHR branch for a region. This is the terminator branch in the | ||||
709 | // entry block or the first select in the entry block, if any. | ||||
710 | static Instruction* getBranchInsertPoint(RegInfo &RI) { | ||||
711 | Region *R = RI.R; | ||||
712 | BasicBlock *EntryBB = R->getEntry(); | ||||
713 | // The hoist point is by default the terminator of the entry block, which is | ||||
714 | // the same as the branch instruction if RI.HasBranch is true. | ||||
715 | Instruction *HoistPoint = EntryBB->getTerminator(); | ||||
716 | for (SelectInst *SI : RI.Selects) { | ||||
717 | if (SI->getParent() == EntryBB) { | ||||
718 | // Pick the first select in Selects in the entry block. Note Selects is | ||||
719 | // sorted in the instruction order within a block (asserted below). | ||||
720 | HoistPoint = SI; | ||||
721 | break; | ||||
722 | } | ||||
723 | } | ||||
724 | assert(HoistPoint && "Null HoistPoint")((void)0); | ||||
725 | #ifndef NDEBUG1 | ||||
726 | // Check that HoistPoint is the first one in Selects in the entry block, | ||||
727 | // if any. | ||||
728 | DenseSet<Instruction *> EntryBlockSelectSet; | ||||
729 | for (SelectInst *SI : RI.Selects) { | ||||
730 | if (SI->getParent() == EntryBB) { | ||||
731 | EntryBlockSelectSet.insert(SI); | ||||
732 | } | ||||
733 | } | ||||
734 | for (Instruction &I : *EntryBB) { | ||||
735 | if (EntryBlockSelectSet.contains(&I)) { | ||||
736 | assert(&I == HoistPoint &&((void)0) | ||||
737 | "HoistPoint must be the first one in Selects")((void)0); | ||||
738 | break; | ||||
739 | } | ||||
740 | } | ||||
741 | #endif | ||||
742 | return HoistPoint; | ||||
743 | } | ||||
744 | |||||
745 | // Find a CHR scope in the given region. | ||||
746 | CHRScope * CHR::findScope(Region *R) { | ||||
747 | CHRScope *Result = nullptr; | ||||
748 | BasicBlock *Entry = R->getEntry(); | ||||
749 | BasicBlock *Exit = R->getExit(); // null if top level. | ||||
750 | assert(Entry && "Entry must not be null")((void)0); | ||||
751 | assert((Exit == nullptr) == (R->isTopLevelRegion()) &&((void)0) | ||||
752 | "Only top level region has a null exit")((void)0); | ||||
753 | if (Entry) | ||||
754 | CHR_DEBUG(dbgs() << "Entry " << Entry->getName() << "\n")do { } while (false); | ||||
755 | else | ||||
756 | CHR_DEBUG(dbgs() << "Entry null\n")do { } while (false); | ||||
757 | if (Exit) | ||||
758 | CHR_DEBUG(dbgs() << "Exit " << Exit->getName() << "\n")do { } while (false); | ||||
759 | else | ||||
760 | CHR_DEBUG(dbgs() << "Exit null\n")do { } while (false); | ||||
761 | // Exclude cases where Entry is part of a subregion (hence it doesn't belong | ||||
762 | // to this region). | ||||
763 | bool EntryInSubregion = RI.getRegionFor(Entry) != R; | ||||
764 | if (EntryInSubregion) | ||||
765 | return nullptr; | ||||
766 | // Exclude loops | ||||
767 | for (BasicBlock *Pred : predecessors(Entry)) | ||||
768 | if (R->contains(Pred)) | ||||
769 | return nullptr; | ||||
770 | // If any of the basic blocks have address taken, we must skip this region | ||||
771 | // because we cannot clone basic blocks that have address taken. | ||||
772 | for (BasicBlock *BB : R->blocks()) | ||||
773 | if (BB->hasAddressTaken()) | ||||
774 | return nullptr; | ||||
775 | if (Exit) { | ||||
776 | // Try to find an if-then block (check if R is an if-then). | ||||
777 | // if (cond) { | ||||
778 | // ... | ||||
779 | // } | ||||
780 | auto *BI = dyn_cast<BranchInst>(Entry->getTerminator()); | ||||
781 | if (BI) | ||||
782 | CHR_DEBUG(dbgs() << "BI.isConditional " << BI->isConditional() << "\n")do { } while (false); | ||||
783 | else | ||||
784 | CHR_DEBUG(dbgs() << "BI null\n")do { } while (false); | ||||
785 | if (BI && BI->isConditional()) { | ||||
786 | BasicBlock *S0 = BI->getSuccessor(0); | ||||
787 | BasicBlock *S1 = BI->getSuccessor(1); | ||||
788 | CHR_DEBUG(dbgs() << "S0 " << S0->getName() << "\n")do { } while (false); | ||||
789 | CHR_DEBUG(dbgs() << "S1 " << S1->getName() << "\n")do { } while (false); | ||||
790 | if (S0 != S1 && (S0 == Exit || S1 == Exit)) { | ||||
791 | RegInfo RI(R); | ||||
792 | RI.HasBranch = checkBiasedBranch( | ||||
793 | BI, R, TrueBiasedRegionsGlobal, FalseBiasedRegionsGlobal, | ||||
794 | BranchBiasMap); | ||||
795 | Result = new CHRScope(RI); | ||||
796 | Scopes.insert(Result); | ||||
797 | CHR_DEBUG(dbgs() << "Found a region with a branch\n")do { } while (false); | ||||
798 | ++Stats.NumBranches; | ||||
799 | if (!RI.HasBranch) { | ||||
800 | ORE.emit([&]() { | ||||
801 | return OptimizationRemarkMissed(DEBUG_TYPE"chr", "BranchNotBiased", BI) | ||||
802 | << "Branch not biased"; | ||||
803 | }); | ||||
804 | } | ||||
805 | } | ||||
806 | } | ||||
807 | } | ||||
808 | { | ||||
809 | // Try to look for selects in the direct child blocks (as opposed to in | ||||
810 | // subregions) of R. | ||||
811 | // ... | ||||
812 | // if (..) { // Some subregion | ||||
813 | // ... | ||||
814 | // } | ||||
815 | // if (..) { // Some subregion | ||||
816 | // ... | ||||
817 | // } | ||||
818 | // ... | ||||
819 | // a = cond ? b : c; | ||||
820 | // ... | ||||
821 | SmallVector<SelectInst *, 8> Selects; | ||||
822 | for (RegionNode *E : R->elements()) { | ||||
823 | if (E->isSubRegion()) | ||||
824 | continue; | ||||
825 | // This returns the basic block of E if E is a direct child of R (not a | ||||
826 | // subregion.) | ||||
827 | BasicBlock *BB = E->getEntry(); | ||||
828 | // Need to push in the order to make it easier to find the first Select | ||||
829 | // later. | ||||
830 | for (Instruction &I : *BB) { | ||||
831 | if (auto *SI = dyn_cast<SelectInst>(&I)) { | ||||
832 | Selects.push_back(SI); | ||||
833 | ++Stats.NumBranches; | ||||
834 | } | ||||
835 | } | ||||
836 | } | ||||
837 | if (Selects.size() > 0) { | ||||
838 | auto AddSelects = [&](RegInfo &RI) { | ||||
839 | for (auto *SI : Selects) | ||||
840 | if (checkBiasedSelect(SI, RI.R, | ||||
841 | TrueBiasedSelectsGlobal, | ||||
842 | FalseBiasedSelectsGlobal, | ||||
843 | SelectBiasMap)) | ||||
844 | RI.Selects.push_back(SI); | ||||
845 | else | ||||
846 | ORE.emit([&]() { | ||||
847 | return OptimizationRemarkMissed(DEBUG_TYPE"chr", "SelectNotBiased", SI) | ||||
848 | << "Select not biased"; | ||||
849 | }); | ||||
850 | }; | ||||
851 | if (!Result) { | ||||
852 | CHR_DEBUG(dbgs() << "Found a select-only region\n")do { } while (false); | ||||
853 | RegInfo RI(R); | ||||
854 | AddSelects(RI); | ||||
855 | Result = new CHRScope(RI); | ||||
856 | Scopes.insert(Result); | ||||
857 | } else { | ||||
858 | CHR_DEBUG(dbgs() << "Found select(s) in a region with a branch\n")do { } while (false); | ||||
859 | AddSelects(Result->RegInfos[0]); | ||||
860 | } | ||||
861 | } | ||||
862 | } | ||||
863 | |||||
864 | if (Result) { | ||||
865 | checkScopeHoistable(Result); | ||||
866 | } | ||||
867 | return Result; | ||||
868 | } | ||||
869 | |||||
870 | // Check that any of the branch and the selects in the region could be | ||||
871 | // hoisted above the the CHR branch insert point (the most dominating of | ||||
872 | // them, either the branch (at the end of the first block) or the first | ||||
873 | // select in the first block). If the branch can't be hoisted, drop the | ||||
874 | // selects in the first blocks. | ||||
875 | // | ||||
876 | // For example, for the following scope/region with selects, we want to insert | ||||
877 | // the merged branch right before the first select in the first/entry block by | ||||
878 | // hoisting c1, c2, c3, and c4. | ||||
879 | // | ||||
880 | // // Branch insert point here. | ||||
881 | // a = c1 ? b : c; // Select 1 | ||||
882 | // d = c2 ? e : f; // Select 2 | ||||
883 | // if (c3) { // Branch | ||||
884 | // ... | ||||
885 | // c4 = foo() // A call. | ||||
886 | // g = c4 ? h : i; // Select 3 | ||||
887 | // } | ||||
888 | // | ||||
889 | // But suppose we can't hoist c4 because it's dependent on the preceding | ||||
890 | // call. Then, we drop Select 3. Furthermore, if we can't hoist c2, we also drop | ||||
891 | // Select 2. If we can't hoist c3, we drop Selects 1 & 2. | ||||
892 | void CHR::checkScopeHoistable(CHRScope *Scope) { | ||||
893 | RegInfo &RI = Scope->RegInfos[0]; | ||||
894 | Region *R = RI.R; | ||||
895 | BasicBlock *EntryBB = R->getEntry(); | ||||
896 | auto *Branch = RI.HasBranch ? | ||||
| |||||
897 | cast<BranchInst>(EntryBB->getTerminator()) : nullptr; | ||||
898 | SmallVector<SelectInst *, 8> &Selects = RI.Selects; | ||||
899 | if (RI.HasBranch
| ||||
900 | Instruction *InsertPoint = getBranchInsertPoint(RI); | ||||
901 | CHR_DEBUG(dbgs() << "InsertPoint " << *InsertPoint << "\n")do { } while (false); | ||||
902 | // Avoid a data dependence from a select or a branch to a(nother) | ||||
903 | // select. Note no instruction can't data-depend on a branch (a branch | ||||
904 | // instruction doesn't produce a value). | ||||
905 | DenseSet<Instruction *> Unhoistables; | ||||
906 | // Initialize Unhoistables with the selects. | ||||
907 | for (SelectInst *SI : Selects) { | ||||
908 | Unhoistables.insert(SI); | ||||
909 | } | ||||
910 | // Remove Selects that can't be hoisted. | ||||
911 | for (auto it = Selects.begin(); it != Selects.end(); ) { | ||||
912 | SelectInst *SI = *it; | ||||
913 | if (SI == InsertPoint) { | ||||
914 | ++it; | ||||
915 | continue; | ||||
916 | } | ||||
917 | DenseMap<Instruction *, bool> Visited; | ||||
918 | bool IsHoistable = checkHoistValue(SI->getCondition(), InsertPoint, | ||||
919 | DT, Unhoistables, nullptr, Visited); | ||||
920 | if (!IsHoistable) { | ||||
921 | CHR_DEBUG(dbgs() << "Dropping select " << *SI << "\n")do { } while (false); | ||||
922 | ORE.emit([&]() { | ||||
923 | return OptimizationRemarkMissed(DEBUG_TYPE"chr", | ||||
924 | "DropUnhoistableSelect", SI) | ||||
925 | << "Dropped unhoistable select"; | ||||
926 | }); | ||||
927 | it = Selects.erase(it); | ||||
928 | // Since we are dropping the select here, we also drop it from | ||||
929 | // Unhoistables. | ||||
930 | Unhoistables.erase(SI); | ||||
931 | } else | ||||
932 | ++it; | ||||
933 | } | ||||
934 | // Update InsertPoint after potentially removing selects. | ||||
935 | InsertPoint = getBranchInsertPoint(RI); | ||||
936 | CHR_DEBUG(dbgs() << "InsertPoint " << *InsertPoint << "\n")do { } while (false); | ||||
937 | if (RI.HasBranch && InsertPoint != Branch) { | ||||
938 | DenseMap<Instruction *, bool> Visited; | ||||
939 | bool IsHoistable = checkHoistValue(Branch->getCondition(), InsertPoint, | ||||
| |||||
940 | DT, Unhoistables, nullptr, Visited); | ||||
941 | if (!IsHoistable) { | ||||
942 | // If the branch isn't hoistable, drop the selects in the entry | ||||
943 | // block, preferring the branch, which makes the branch the hoist | ||||
944 | // point. | ||||
945 | assert(InsertPoint != Branch && "Branch must not be the hoist point")((void)0); | ||||
946 | CHR_DEBUG(dbgs() << "Dropping selects in entry block \n")do { } while (false); | ||||
947 | CHR_DEBUG(do { } while (false) | ||||
948 | for (SelectInst *SI : Selects) {do { } while (false) | ||||
949 | dbgs() << "SI " << *SI << "\n";do { } while (false) | ||||
950 | })do { } while (false); | ||||
951 | for (SelectInst *SI : Selects) { | ||||
952 | ORE.emit([&]() { | ||||
953 | return OptimizationRemarkMissed(DEBUG_TYPE"chr", | ||||
954 | "DropSelectUnhoistableBranch", SI) | ||||
955 | << "Dropped select due to unhoistable branch"; | ||||
956 | }); | ||||
957 | } | ||||
958 | llvm::erase_if(Selects, [EntryBB](SelectInst *SI) { | ||||
959 | return SI->getParent() == EntryBB; | ||||
960 | }); | ||||
961 | Unhoistables.clear(); | ||||
962 | InsertPoint = Branch; | ||||
963 | } | ||||
964 | } | ||||
965 | CHR_DEBUG(dbgs() << "InsertPoint " << *InsertPoint << "\n")do { } while (false); | ||||
966 | #ifndef NDEBUG1 | ||||
967 | if (RI.HasBranch) { | ||||
968 | assert(!DT.dominates(Branch, InsertPoint) &&((void)0) | ||||
969 | "Branch can't be already above the hoist point")((void)0); | ||||
970 | DenseMap<Instruction *, bool> Visited; | ||||
971 | assert(checkHoistValue(Branch->getCondition(), InsertPoint,((void)0) | ||||
972 | DT, Unhoistables, nullptr, Visited) &&((void)0) | ||||
973 | "checkHoistValue for branch")((void)0); | ||||
974 | } | ||||
975 | for (auto *SI : Selects) { | ||||
976 | assert(!DT.dominates(SI, InsertPoint) &&((void)0) | ||||
977 | "SI can't be already above the hoist point")((void)0); | ||||
978 | DenseMap<Instruction *, bool> Visited; | ||||
979 | assert(checkHoistValue(SI->getCondition(), InsertPoint, DT,((void)0) | ||||
980 | Unhoistables, nullptr, Visited) &&((void)0) | ||||
981 | "checkHoistValue for selects")((void)0); | ||||
982 | } | ||||
983 | CHR_DEBUG(dbgs() << "Result\n")do { } while (false); | ||||
984 | if (RI.HasBranch) { | ||||
985 | CHR_DEBUG(dbgs() << "BI " << *Branch << "\n")do { } while (false); | ||||
986 | } | ||||
987 | for (auto *SI : Selects) { | ||||
988 | CHR_DEBUG(dbgs() << "SI " << *SI << "\n")do { } while (false); | ||||
989 | } | ||||
990 | #endif | ||||
991 | } | ||||
992 | } | ||||
993 | |||||
994 | // Traverse the region tree, find all nested scopes and merge them if possible. | ||||
995 | CHRScope * CHR::findScopes(Region *R, Region *NextRegion, Region *ParentRegion, | ||||
996 | SmallVectorImpl<CHRScope *> &Scopes) { | ||||
997 | CHR_DEBUG(dbgs() << "findScopes " << R->getNameStr() << "\n")do { } while (false); | ||||
998 | CHRScope *Result = findScope(R); | ||||
999 | // Visit subscopes. | ||||
1000 | CHRScope *ConsecutiveSubscope = nullptr; | ||||
1001 | SmallVector<CHRScope *, 8> Subscopes; | ||||
1002 | for (auto It = R->begin(); It != R->end(); ++It) { | ||||
1003 | const std::unique_ptr<Region> &SubR = *It; | ||||
1004 | auto NextIt = std::next(It); | ||||
1005 | Region *NextSubR = NextIt != R->end() ? NextIt->get() : nullptr; | ||||
1006 | CHR_DEBUG(dbgs() << "Looking at subregion " << SubR.get()->getNameStr()do { } while (false) | ||||
1007 | << "\n")do { } while (false); | ||||
1008 | CHRScope *SubCHRScope = findScopes(SubR.get(), NextSubR, R, Scopes); | ||||
1009 | if (SubCHRScope) { | ||||
1010 | CHR_DEBUG(dbgs() << "Subregion Scope " << *SubCHRScope << "\n")do { } while (false); | ||||
1011 | } else { | ||||
1012 | CHR_DEBUG(dbgs() << "Subregion Scope null\n")do { } while (false); | ||||
1013 | } | ||||
1014 | if (SubCHRScope) { | ||||
1015 | if (!ConsecutiveSubscope) | ||||
1016 | ConsecutiveSubscope = SubCHRScope; | ||||
1017 | else if (!ConsecutiveSubscope->appendable(SubCHRScope)) { | ||||
1018 | Subscopes.push_back(ConsecutiveSubscope); | ||||
1019 | ConsecutiveSubscope = SubCHRScope; | ||||
1020 | } else | ||||
1021 | ConsecutiveSubscope->append(SubCHRScope); | ||||
1022 | } else { | ||||
1023 | if (ConsecutiveSubscope) { | ||||
1024 | Subscopes.push_back(ConsecutiveSubscope); | ||||
1025 | } | ||||
1026 | ConsecutiveSubscope = nullptr; | ||||
1027 | } | ||||
1028 | } | ||||
1029 | if (ConsecutiveSubscope) { | ||||
1030 | Subscopes.push_back(ConsecutiveSubscope); | ||||
1031 | } | ||||
1032 | for (CHRScope *Sub : Subscopes) { | ||||
1033 | if (Result) { | ||||
1034 | // Combine it with the parent. | ||||
1035 | Result->addSub(Sub); | ||||
1036 | } else { | ||||
1037 | // Push Subscopes as they won't be combined with the parent. | ||||
1038 | Scopes.push_back(Sub); | ||||
1039 | } | ||||
1040 | } | ||||
1041 | return Result; | ||||
1042 | } | ||||
1043 | |||||
1044 | static DenseSet<Value *> getCHRConditionValuesForRegion(RegInfo &RI) { | ||||
1045 | DenseSet<Value *> ConditionValues; | ||||
1046 | if (RI.HasBranch) { | ||||
1047 | auto *BI = cast<BranchInst>(RI.R->getEntry()->getTerminator()); | ||||
1048 | ConditionValues.insert(BI->getCondition()); | ||||
1049 | } | ||||
1050 | for (SelectInst *SI : RI.Selects) { | ||||
1051 | ConditionValues.insert(SI->getCondition()); | ||||
1052 | } | ||||
1053 | return ConditionValues; | ||||
1054 | } | ||||
1055 | |||||
1056 | |||||
1057 | // Determine whether to split a scope depending on the sets of the branch | ||||
1058 | // condition values of the previous region and the current region. We split | ||||
1059 | // (return true) it if 1) the condition values of the inner/lower scope can't be | ||||
1060 | // hoisted up to the outer/upper scope, or 2) the two sets of the condition | ||||
1061 | // values have an empty intersection (because the combined branch conditions | ||||
1062 | // won't probably lead to a simpler combined condition). | ||||
1063 | static bool shouldSplit(Instruction *InsertPoint, | ||||
1064 | DenseSet<Value *> &PrevConditionValues, | ||||
1065 | DenseSet<Value *> &ConditionValues, | ||||
1066 | DominatorTree &DT, | ||||
1067 | DenseSet<Instruction *> &Unhoistables) { | ||||
1068 | assert(InsertPoint && "Null InsertPoint")((void)0); | ||||
1069 | CHR_DEBUG(do { } while (false) | ||||
1070 | dbgs() << "shouldSplit " << *InsertPoint << " PrevConditionValues ";do { } while (false) | ||||
1071 | for (Value *V : PrevConditionValues) {do { } while (false) | ||||
1072 | dbgs() << *V << ", ";do { } while (false) | ||||
1073 | }do { } while (false) | ||||
1074 | dbgs() << " ConditionValues ";do { } while (false) | ||||
1075 | for (Value *V : ConditionValues) {do { } while (false) | ||||
1076 | dbgs() << *V << ", ";do { } while (false) | ||||
1077 | }do { } while (false) | ||||
1078 | dbgs() << "\n")do { } while (false); | ||||
1079 | // If any of Bases isn't hoistable to the hoist point, split. | ||||
1080 | for (Value *V : ConditionValues) { | ||||
1081 | DenseMap<Instruction *, bool> Visited; | ||||
1082 | if (!checkHoistValue(V, InsertPoint, DT, Unhoistables, nullptr, Visited)) { | ||||
1083 | CHR_DEBUG(dbgs() << "Split. checkHoistValue false " << *V << "\n")do { } while (false); | ||||
1084 | return true; // Not hoistable, split. | ||||
1085 | } | ||||
1086 | } | ||||
1087 | // If PrevConditionValues or ConditionValues is empty, don't split to avoid | ||||
1088 | // unnecessary splits at scopes with no branch/selects. If | ||||
1089 | // PrevConditionValues and ConditionValues don't intersect at all, split. | ||||
1090 | if (!PrevConditionValues.empty() && !ConditionValues.empty()) { | ||||
1091 | // Use std::set as DenseSet doesn't work with set_intersection. | ||||
1092 | std::set<Value *> PrevBases, Bases; | ||||
1093 | DenseMap<Value *, std::set<Value *>> Visited; | ||||
1094 | for (Value *V : PrevConditionValues) { | ||||
1095 | const std::set<Value *> &BaseValues = getBaseValues(V, DT, Visited); | ||||
1096 | PrevBases.insert(BaseValues.begin(), BaseValues.end()); | ||||
1097 | } | ||||
1098 | for (Value *V : ConditionValues) { | ||||
1099 | const std::set<Value *> &BaseValues = getBaseValues(V, DT, Visited); | ||||
1100 | Bases.insert(BaseValues.begin(), BaseValues.end()); | ||||
1101 | } | ||||
1102 | CHR_DEBUG(do { } while (false) | ||||
1103 | dbgs() << "PrevBases ";do { } while (false) | ||||
1104 | for (Value *V : PrevBases) {do { } while (false) | ||||
1105 | dbgs() << *V << ", ";do { } while (false) | ||||
1106 | }do { } while (false) | ||||
1107 | dbgs() << " Bases ";do { } while (false) | ||||
1108 | for (Value *V : Bases) {do { } while (false) | ||||
1109 | dbgs() << *V << ", ";do { } while (false) | ||||
1110 | }do { } while (false) | ||||
1111 | dbgs() << "\n")do { } while (false); | ||||
1112 | std::vector<Value *> Intersection; | ||||
1113 | std::set_intersection(PrevBases.begin(), PrevBases.end(), Bases.begin(), | ||||
1114 | Bases.end(), std::back_inserter(Intersection)); | ||||
1115 | if (Intersection.empty()) { | ||||
1116 | // Empty intersection, split. | ||||
1117 | CHR_DEBUG(dbgs() << "Split. Intersection empty\n")do { } while (false); | ||||
1118 | return true; | ||||
1119 | } | ||||
1120 | } | ||||
1121 | CHR_DEBUG(dbgs() << "No split\n")do { } while (false); | ||||
1122 | return false; // Don't split. | ||||
1123 | } | ||||
1124 | |||||
1125 | static void getSelectsInScope(CHRScope *Scope, | ||||
1126 | DenseSet<Instruction *> &Output) { | ||||
1127 | for (RegInfo &RI : Scope->RegInfos) | ||||
1128 | for (SelectInst *SI : RI.Selects) | ||||
1129 | Output.insert(SI); | ||||
1130 | for (CHRScope *Sub : Scope->Subs) | ||||
1131 | getSelectsInScope(Sub, Output); | ||||
1132 | } | ||||
1133 | |||||
1134 | void CHR::splitScopes(SmallVectorImpl<CHRScope *> &Input, | ||||
1135 | SmallVectorImpl<CHRScope *> &Output) { | ||||
1136 | for (CHRScope *Scope : Input) { | ||||
1137 | assert(!Scope->BranchInsertPoint &&((void)0) | ||||
1138 | "BranchInsertPoint must not be set")((void)0); | ||||
1139 | DenseSet<Instruction *> Unhoistables; | ||||
1140 | getSelectsInScope(Scope, Unhoistables); | ||||
1141 | splitScope(Scope, nullptr, nullptr, nullptr, Output, Unhoistables); | ||||
1142 | } | ||||
1143 | #ifndef NDEBUG1 | ||||
1144 | for (CHRScope *Scope : Output) { | ||||
1145 | assert(Scope->BranchInsertPoint && "BranchInsertPoint must be set")((void)0); | ||||
1146 | } | ||||
1147 | #endif | ||||
1148 | } | ||||
1149 | |||||
1150 | SmallVector<CHRScope *, 8> CHR::splitScope( | ||||
1151 | CHRScope *Scope, | ||||
1152 | CHRScope *Outer, | ||||
1153 | DenseSet<Value *> *OuterConditionValues, | ||||
1154 | Instruction *OuterInsertPoint, | ||||
1155 | SmallVectorImpl<CHRScope *> &Output, | ||||
1156 | DenseSet<Instruction *> &Unhoistables) { | ||||
1157 | if (Outer) { | ||||
1158 | assert(OuterConditionValues && "Null OuterConditionValues")((void)0); | ||||
1159 | assert(OuterInsertPoint && "Null OuterInsertPoint")((void)0); | ||||
1160 | } | ||||
1161 | bool PrevSplitFromOuter = true; | ||||
1162 | DenseSet<Value *> PrevConditionValues; | ||||
1163 | Instruction *PrevInsertPoint = nullptr; | ||||
1164 | SmallVector<CHRScope *, 8> Splits; | ||||
1165 | SmallVector<bool, 8> SplitsSplitFromOuter; | ||||
1166 | SmallVector<DenseSet<Value *>, 8> SplitsConditionValues; | ||||
1167 | SmallVector<Instruction *, 8> SplitsInsertPoints; | ||||
1168 | SmallVector<RegInfo, 8> RegInfos(Scope->RegInfos); // Copy | ||||
1169 | for (RegInfo &RI : RegInfos) { | ||||
1170 | Instruction *InsertPoint = getBranchInsertPoint(RI); | ||||
1171 | DenseSet<Value *> ConditionValues = getCHRConditionValuesForRegion(RI); | ||||
1172 | CHR_DEBUG(do { } while (false) | ||||
1173 | dbgs() << "ConditionValues ";do { } while (false) | ||||
1174 | for (Value *V : ConditionValues) {do { } while (false) | ||||
1175 | dbgs() << *V << ", ";do { } while (false) | ||||
1176 | }do { } while (false) | ||||
1177 | dbgs() << "\n")do { } while (false); | ||||
1178 | if (RI.R == RegInfos[0].R) { | ||||
1179 | // First iteration. Check to see if we should split from the outer. | ||||
1180 | if (Outer) { | ||||
1181 | CHR_DEBUG(dbgs() << "Outer " << *Outer << "\n")do { } while (false); | ||||
1182 | CHR_DEBUG(dbgs() << "Should split from outer at "do { } while (false) | ||||
1183 | << RI.R->getNameStr() << "\n")do { } while (false); | ||||
1184 | if (shouldSplit(OuterInsertPoint, *OuterConditionValues, | ||||
1185 | ConditionValues, DT, Unhoistables)) { | ||||
1186 | PrevConditionValues = ConditionValues; | ||||
1187 | PrevInsertPoint = InsertPoint; | ||||
1188 | ORE.emit([&]() { | ||||
1189 | return OptimizationRemarkMissed(DEBUG_TYPE"chr", | ||||
1190 | "SplitScopeFromOuter", | ||||
1191 | RI.R->getEntry()->getTerminator()) | ||||
1192 | << "Split scope from outer due to unhoistable branch/select " | ||||
1193 | << "and/or lack of common condition values"; | ||||
1194 | }); | ||||
1195 | } else { | ||||
1196 | // Not splitting from the outer. Use the outer bases and insert | ||||
1197 | // point. Union the bases. | ||||
1198 | PrevSplitFromOuter = false; | ||||
1199 | PrevConditionValues = *OuterConditionValues; | ||||
1200 | PrevConditionValues.insert(ConditionValues.begin(), | ||||
1201 | ConditionValues.end()); | ||||
1202 | PrevInsertPoint = OuterInsertPoint; | ||||
1203 | } | ||||
1204 | } else { | ||||
1205 | CHR_DEBUG(dbgs() << "Outer null\n")do { } while (false); | ||||
1206 | PrevConditionValues = ConditionValues; | ||||
1207 | PrevInsertPoint = InsertPoint; | ||||
1208 | } | ||||
1209 | } else { | ||||
1210 | CHR_DEBUG(dbgs() << "Should split from prev at "do { } while (false) | ||||
1211 | << RI.R->getNameStr() << "\n")do { } while (false); | ||||
1212 | if (shouldSplit(PrevInsertPoint, PrevConditionValues, ConditionValues, | ||||
1213 | DT, Unhoistables)) { | ||||
1214 | CHRScope *Tail = Scope->split(RI.R); | ||||
1215 | Scopes.insert(Tail); | ||||
1216 | Splits.push_back(Scope); | ||||
1217 | SplitsSplitFromOuter.push_back(PrevSplitFromOuter); | ||||
1218 | SplitsConditionValues.push_back(PrevConditionValues); | ||||
1219 | SplitsInsertPoints.push_back(PrevInsertPoint); | ||||
1220 | Scope = Tail; | ||||
1221 | PrevConditionValues = ConditionValues; | ||||
1222 | PrevInsertPoint = InsertPoint; | ||||
1223 | PrevSplitFromOuter = true; | ||||
1224 | ORE.emit([&]() { | ||||
1225 | return OptimizationRemarkMissed(DEBUG_TYPE"chr", | ||||
1226 | "SplitScopeFromPrev", | ||||
1227 | RI.R->getEntry()->getTerminator()) | ||||
1228 | << "Split scope from previous due to unhoistable branch/select " | ||||
1229 | << "and/or lack of common condition values"; | ||||
1230 | }); | ||||
1231 | } else { | ||||
1232 | // Not splitting. Union the bases. Keep the hoist point. | ||||
1233 | PrevConditionValues.insert(ConditionValues.begin(), ConditionValues.end()); | ||||
1234 | } | ||||
1235 | } | ||||
1236 | } | ||||
1237 | Splits.push_back(Scope); | ||||
1238 | SplitsSplitFromOuter.push_back(PrevSplitFromOuter); | ||||
1239 | SplitsConditionValues.push_back(PrevConditionValues); | ||||
1240 | assert(PrevInsertPoint && "Null PrevInsertPoint")((void)0); | ||||
1241 | SplitsInsertPoints.push_back(PrevInsertPoint); | ||||
1242 | assert(Splits.size() == SplitsConditionValues.size() &&((void)0) | ||||
1243 | Splits.size() == SplitsSplitFromOuter.size() &&((void)0) | ||||
1244 | Splits.size() == SplitsInsertPoints.size() && "Mismatching sizes")((void)0); | ||||
1245 | for (size_t I = 0; I < Splits.size(); ++I) { | ||||
1246 | CHRScope *Split = Splits[I]; | ||||
1247 | DenseSet<Value *> &SplitConditionValues = SplitsConditionValues[I]; | ||||
1248 | Instruction *SplitInsertPoint = SplitsInsertPoints[I]; | ||||
1249 | SmallVector<CHRScope *, 8> NewSubs; | ||||
1250 | DenseSet<Instruction *> SplitUnhoistables; | ||||
1251 | getSelectsInScope(Split, SplitUnhoistables); | ||||
1252 | for (CHRScope *Sub : Split->Subs) { | ||||
1253 | SmallVector<CHRScope *, 8> SubSplits = splitScope( | ||||
1254 | Sub, Split, &SplitConditionValues, SplitInsertPoint, Output, | ||||
1255 | SplitUnhoistables); | ||||
1256 | llvm::append_range(NewSubs, SubSplits); | ||||
1257 | } | ||||
1258 | Split->Subs = NewSubs; | ||||
1259 | } | ||||
1260 | SmallVector<CHRScope *, 8> Result; | ||||
1261 | for (size_t I = 0; I < Splits.size(); ++I) { | ||||
1262 | CHRScope *Split = Splits[I]; | ||||
1263 | if (SplitsSplitFromOuter[I]) { | ||||
1264 | // Split from the outer. | ||||
1265 | Output.push_back(Split); | ||||
1266 | Split->BranchInsertPoint = SplitsInsertPoints[I]; | ||||
1267 | CHR_DEBUG(dbgs() << "BranchInsertPoint " << *SplitsInsertPoints[I]do { } while (false) | ||||
1268 | << "\n")do { } while (false); | ||||
1269 | } else { | ||||
1270 | // Connected to the outer. | ||||
1271 | Result.push_back(Split); | ||||
1272 | } | ||||
1273 | } | ||||
1274 | if (!Outer) | ||||
1275 | assert(Result.empty() &&((void)0) | ||||
1276 | "If no outer (top-level), must return no nested ones")((void)0); | ||||
1277 | return Result; | ||||
1278 | } | ||||
1279 | |||||
1280 | void CHR::classifyBiasedScopes(SmallVectorImpl<CHRScope *> &Scopes) { | ||||
1281 | for (CHRScope *Scope : Scopes) { | ||||
1282 | assert(Scope->TrueBiasedRegions.empty() && Scope->FalseBiasedRegions.empty() && "Empty")((void)0); | ||||
1283 | classifyBiasedScopes(Scope, Scope); | ||||
1284 | CHR_DEBUG(do { } while (false) | ||||
1285 | dbgs() << "classifyBiasedScopes " << *Scope << "\n";do { } while (false) | ||||
1286 | dbgs() << "TrueBiasedRegions ";do { } while (false) | ||||
1287 | for (Region *R : Scope->TrueBiasedRegions) {do { } while (false) | ||||
1288 | dbgs() << R->getNameStr() << ", ";do { } while (false) | ||||
1289 | }do { } while (false) | ||||
1290 | dbgs() << "\n";do { } while (false) | ||||
1291 | dbgs() << "FalseBiasedRegions ";do { } while (false) | ||||
1292 | for (Region *R : Scope->FalseBiasedRegions) {do { } while (false) | ||||
1293 | dbgs() << R->getNameStr() << ", ";do { } while (false) | ||||
1294 | }do { } while (false) | ||||
1295 | dbgs() << "\n";do { } while (false) | ||||
1296 | dbgs() << "TrueBiasedSelects ";do { } while (false) | ||||
1297 | for (SelectInst *SI : Scope->TrueBiasedSelects) {do { } while (false) | ||||
1298 | dbgs() << *SI << ", ";do { } while (false) | ||||
1299 | }do { } while (false) | ||||
1300 | dbgs() << "\n";do { } while (false) | ||||
1301 | dbgs() << "FalseBiasedSelects ";do { } while (false) | ||||
1302 | for (SelectInst *SI : Scope->FalseBiasedSelects) {do { } while (false) | ||||
1303 | dbgs() << *SI << ", ";do { } while (false) | ||||
1304 | }do { } while (false) | ||||
1305 | dbgs() << "\n";)do { } while (false); | ||||
1306 | } | ||||
1307 | } | ||||
1308 | |||||
1309 | void CHR::classifyBiasedScopes(CHRScope *Scope, CHRScope *OutermostScope) { | ||||
1310 | for (RegInfo &RI : Scope->RegInfos) { | ||||
1311 | if (RI.HasBranch) { | ||||
1312 | Region *R = RI.R; | ||||
1313 | if (TrueBiasedRegionsGlobal.contains(R)) | ||||
1314 | OutermostScope->TrueBiasedRegions.insert(R); | ||||
1315 | else if (FalseBiasedRegionsGlobal.contains(R)) | ||||
1316 | OutermostScope->FalseBiasedRegions.insert(R); | ||||
1317 | else | ||||
1318 | llvm_unreachable("Must be biased")__builtin_unreachable(); | ||||
1319 | } | ||||
1320 | for (SelectInst *SI : RI.Selects) { | ||||
1321 | if (TrueBiasedSelectsGlobal.contains(SI)) | ||||
1322 | OutermostScope->TrueBiasedSelects.insert(SI); | ||||
1323 | else if (FalseBiasedSelectsGlobal.contains(SI)) | ||||
1324 | OutermostScope->FalseBiasedSelects.insert(SI); | ||||
1325 | else | ||||
1326 | llvm_unreachable("Must be biased")__builtin_unreachable(); | ||||
1327 | } | ||||
1328 | } | ||||
1329 | for (CHRScope *Sub : Scope->Subs) { | ||||
1330 | classifyBiasedScopes(Sub, OutermostScope); | ||||
1331 | } | ||||
1332 | } | ||||
1333 | |||||
1334 | static bool hasAtLeastTwoBiasedBranches(CHRScope *Scope) { | ||||
1335 | unsigned NumBiased = Scope->TrueBiasedRegions.size() + | ||||
1336 | Scope->FalseBiasedRegions.size() + | ||||
1337 | Scope->TrueBiasedSelects.size() + | ||||
1338 | Scope->FalseBiasedSelects.size(); | ||||
1339 | return NumBiased >= CHRMergeThreshold; | ||||
1340 | } | ||||
1341 | |||||
1342 | void CHR::filterScopes(SmallVectorImpl<CHRScope *> &Input, | ||||
1343 | SmallVectorImpl<CHRScope *> &Output) { | ||||
1344 | for (CHRScope *Scope : Input) { | ||||
1345 | // Filter out the ones with only one region and no subs. | ||||
1346 | if (!hasAtLeastTwoBiasedBranches(Scope)) { | ||||
1347 | CHR_DEBUG(dbgs() << "Filtered out by biased branches truthy-regions "do { } while (false) | ||||
1348 | << Scope->TrueBiasedRegions.size()do { } while (false) | ||||
1349 | << " falsy-regions " << Scope->FalseBiasedRegions.size()do { } while (false) | ||||
1350 | << " true-selects " << Scope->TrueBiasedSelects.size()do { } while (false) | ||||
1351 | << " false-selects " << Scope->FalseBiasedSelects.size() << "\n")do { } while (false); | ||||
1352 | ORE.emit([&]() { | ||||
1353 | return OptimizationRemarkMissed( | ||||
1354 | DEBUG_TYPE"chr", | ||||
1355 | "DropScopeWithOneBranchOrSelect", | ||||
1356 | Scope->RegInfos[0].R->getEntry()->getTerminator()) | ||||
1357 | << "Drop scope with < " | ||||
1358 | << ore::NV("CHRMergeThreshold", CHRMergeThreshold) | ||||
1359 | << " biased branch(es) or select(s)"; | ||||
1360 | }); | ||||
1361 | continue; | ||||
1362 | } | ||||
1363 | Output.push_back(Scope); | ||||
1364 | } | ||||
1365 | } | ||||
1366 | |||||
1367 | void CHR::setCHRRegions(SmallVectorImpl<CHRScope *> &Input, | ||||
1368 | SmallVectorImpl<CHRScope *> &Output) { | ||||
1369 | for (CHRScope *Scope : Input) { | ||||
1370 | assert(Scope->HoistStopMap.empty() && Scope->CHRRegions.empty() &&((void)0) | ||||
1371 | "Empty")((void)0); | ||||
1372 | setCHRRegions(Scope, Scope); | ||||
1373 | Output.push_back(Scope); | ||||
1374 | CHR_DEBUG(do { } while (false) | ||||
1375 | dbgs() << "setCHRRegions HoistStopMap " << *Scope << "\n";do { } while (false) | ||||
1376 | for (auto pair : Scope->HoistStopMap) {do { } while (false) | ||||
1377 | Region *R = pair.first;do { } while (false) | ||||
1378 | dbgs() << "Region " << R->getNameStr() << "\n";do { } while (false) | ||||
1379 | for (Instruction *I : pair.second) {do { } while (false) | ||||
1380 | dbgs() << "HoistStop " << *I << "\n";do { } while (false) | ||||
1381 | }do { } while (false) | ||||
1382 | }do { } while (false) | ||||
1383 | dbgs() << "CHRRegions" << "\n";do { } while (false) | ||||
1384 | for (RegInfo &RI : Scope->CHRRegions) {do { } while (false) | ||||
1385 | dbgs() << RI.R->getNameStr() << "\n";do { } while (false) | ||||
1386 | })do { } while (false); | ||||
1387 | } | ||||
1388 | } | ||||
1389 | |||||
1390 | void CHR::setCHRRegions(CHRScope *Scope, CHRScope *OutermostScope) { | ||||
1391 | DenseSet<Instruction *> Unhoistables; | ||||
1392 | // Put the biased selects in Unhoistables because they should stay where they | ||||
1393 | // are and constant-folded after CHR (in case one biased select or a branch | ||||
1394 | // can depend on another biased select.) | ||||
1395 | for (RegInfo &RI : Scope->RegInfos) { | ||||
1396 | for (SelectInst *SI : RI.Selects) { | ||||
1397 | Unhoistables.insert(SI); | ||||
1398 | } | ||||
1399 | } | ||||
1400 | Instruction *InsertPoint = OutermostScope->BranchInsertPoint; | ||||
1401 | for (RegInfo &RI : Scope->RegInfos) { | ||||
1402 | Region *R = RI.R; | ||||
1403 | DenseSet<Instruction *> HoistStops; | ||||
1404 | bool IsHoisted = false; | ||||
1405 | if (RI.HasBranch) { | ||||
1406 | assert((OutermostScope->TrueBiasedRegions.contains(R) ||((void)0) | ||||
1407 | OutermostScope->FalseBiasedRegions.contains(R)) &&((void)0) | ||||
1408 | "Must be truthy or falsy")((void)0); | ||||
1409 | auto *BI = cast<BranchInst>(R->getEntry()->getTerminator()); | ||||
1410 | // Note checkHoistValue fills in HoistStops. | ||||
1411 | DenseMap<Instruction *, bool> Visited; | ||||
1412 | bool IsHoistable = checkHoistValue(BI->getCondition(), InsertPoint, DT, | ||||
1413 | Unhoistables, &HoistStops, Visited); | ||||
1414 | assert(IsHoistable && "Must be hoistable")((void)0); | ||||
1415 | (void)(IsHoistable); // Unused in release build | ||||
1416 | IsHoisted = true; | ||||
1417 | } | ||||
1418 | for (SelectInst *SI : RI.Selects) { | ||||
1419 | assert((OutermostScope->TrueBiasedSelects.contains(SI) ||((void)0) | ||||
1420 | OutermostScope->FalseBiasedSelects.contains(SI)) &&((void)0) | ||||
1421 | "Must be true or false biased")((void)0); | ||||
1422 | // Note checkHoistValue fills in HoistStops. | ||||
1423 | DenseMap<Instruction *, bool> Visited; | ||||
1424 | bool IsHoistable = checkHoistValue(SI->getCondition(), InsertPoint, DT, | ||||
1425 | Unhoistables, &HoistStops, Visited); | ||||
1426 | assert(IsHoistable && "Must be hoistable")((void)0); | ||||
1427 | (void)(IsHoistable); // Unused in release build | ||||
1428 | IsHoisted = true; | ||||
1429 | } | ||||
1430 | if (IsHoisted) { | ||||
1431 | OutermostScope->CHRRegions.push_back(RI); | ||||
1432 | OutermostScope->HoistStopMap[R] = HoistStops; | ||||
1433 | } | ||||
1434 | } | ||||
1435 | for (CHRScope *Sub : Scope->Subs) | ||||
1436 | setCHRRegions(Sub, OutermostScope); | ||||
1437 | } | ||||
1438 | |||||
1439 | static bool CHRScopeSorter(CHRScope *Scope1, CHRScope *Scope2) { | ||||
1440 | return Scope1->RegInfos[0].R->getDepth() < Scope2->RegInfos[0].R->getDepth(); | ||||
1441 | } | ||||
1442 | |||||
1443 | void CHR::sortScopes(SmallVectorImpl<CHRScope *> &Input, | ||||
1444 | SmallVectorImpl<CHRScope *> &Output) { | ||||
1445 | Output.resize(Input.size()); | ||||
1446 | llvm::copy(Input, Output.begin()); | ||||
1447 | llvm::stable_sort(Output, CHRScopeSorter); | ||||
1448 | } | ||||
1449 | |||||
1450 | // Return true if V is already hoisted or was hoisted (along with its operands) | ||||
1451 | // to the insert point. | ||||
1452 | static void hoistValue(Value *V, Instruction *HoistPoint, Region *R, | ||||
1453 | HoistStopMapTy &HoistStopMap, | ||||
1454 | DenseSet<Instruction *> &HoistedSet, | ||||
1455 | DenseSet<PHINode *> &TrivialPHIs, | ||||
1456 | DominatorTree &DT) { | ||||
1457 | auto IT = HoistStopMap.find(R); | ||||
1458 | assert(IT != HoistStopMap.end() && "Region must be in hoist stop map")((void)0); | ||||
1459 | DenseSet<Instruction *> &HoistStops = IT->second; | ||||
1460 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
1461 | if (I == HoistPoint) | ||||
1462 | return; | ||||
1463 | if (HoistStops.count(I)) | ||||
1464 | return; | ||||
1465 | if (auto *PN = dyn_cast<PHINode>(I)) | ||||
1466 | if (TrivialPHIs.count(PN)) | ||||
1467 | // The trivial phi inserted by the previous CHR scope could replace a | ||||
1468 | // non-phi in HoistStops. Note that since this phi is at the exit of a | ||||
1469 | // previous CHR scope, which dominates this scope, it's safe to stop | ||||
1470 | // hoisting there. | ||||
1471 | return; | ||||
1472 | if (HoistedSet.count(I)) | ||||
1473 | // Already hoisted, return. | ||||
1474 | return; | ||||
1475 | assert(isHoistableInstructionType(I) && "Unhoistable instruction type")((void)0); | ||||
1476 | assert(DT.getNode(I->getParent()) && "DT must contain I's block")((void)0); | ||||
1477 | assert(DT.getNode(HoistPoint->getParent()) &&((void)0) | ||||
1478 | "DT must contain HoistPoint block")((void)0); | ||||
1479 | if (DT.dominates(I, HoistPoint)) | ||||
1480 | // We are already above the hoist point. Stop here. This may be necessary | ||||
1481 | // when multiple scopes would independently hoist the same | ||||
1482 | // instruction. Since an outer (dominating) scope would hoist it to its | ||||
1483 | // entry before an inner (dominated) scope would to its entry, the inner | ||||
1484 | // scope may see the instruction already hoisted, in which case it | ||||
1485 | // potentially wrong for the inner scope to hoist it and could cause bad | ||||
1486 | // IR (non-dominating def), but safe to skip hoisting it instead because | ||||
1487 | // it's already in a block that dominates the inner scope. | ||||
1488 | return; | ||||
1489 | for (Value *Op : I->operands()) { | ||||
1490 | hoistValue(Op, HoistPoint, R, HoistStopMap, HoistedSet, TrivialPHIs, DT); | ||||
1491 | } | ||||
1492 | I->moveBefore(HoistPoint); | ||||
1493 | HoistedSet.insert(I); | ||||
1494 | CHR_DEBUG(dbgs() << "hoistValue " << *I << "\n")do { } while (false); | ||||
1495 | } | ||||
1496 | } | ||||
1497 | |||||
1498 | // Hoist the dependent condition values of the branches and the selects in the | ||||
1499 | // scope to the insert point. | ||||
1500 | static void hoistScopeConditions(CHRScope *Scope, Instruction *HoistPoint, | ||||
1501 | DenseSet<PHINode *> &TrivialPHIs, | ||||
1502 | DominatorTree &DT) { | ||||
1503 | DenseSet<Instruction *> HoistedSet; | ||||
1504 | for (const RegInfo &RI : Scope->CHRRegions) { | ||||
1505 | Region *R = RI.R; | ||||
1506 | bool IsTrueBiased = Scope->TrueBiasedRegions.count(R); | ||||
1507 | bool IsFalseBiased = Scope->FalseBiasedRegions.count(R); | ||||
1508 | if (RI.HasBranch && (IsTrueBiased || IsFalseBiased)) { | ||||
1509 | auto *BI = cast<BranchInst>(R->getEntry()->getTerminator()); | ||||
1510 | hoistValue(BI->getCondition(), HoistPoint, R, Scope->HoistStopMap, | ||||
1511 | HoistedSet, TrivialPHIs, DT); | ||||
1512 | } | ||||
1513 | for (SelectInst *SI : RI.Selects) { | ||||
1514 | bool IsTrueBiased = Scope->TrueBiasedSelects.count(SI); | ||||
1515 | bool IsFalseBiased = Scope->FalseBiasedSelects.count(SI); | ||||
1516 | if (!(IsTrueBiased || IsFalseBiased)) | ||||
1517 | continue; | ||||
1518 | hoistValue(SI->getCondition(), HoistPoint, R, Scope->HoistStopMap, | ||||
1519 | HoistedSet, TrivialPHIs, DT); | ||||
1520 | } | ||||
1521 | } | ||||
1522 | } | ||||
1523 | |||||
1524 | // Negate the predicate if an ICmp if it's used only by branches or selects by | ||||
1525 | // swapping the operands of the branches or the selects. Returns true if success. | ||||
1526 | static bool negateICmpIfUsedByBranchOrSelectOnly(ICmpInst *ICmp, | ||||
1527 | Instruction *ExcludedUser, | ||||
1528 | CHRScope *Scope) { | ||||
1529 | for (User *U : ICmp->users()) { | ||||
1530 | if (U == ExcludedUser) | ||||
1531 | continue; | ||||
1532 | if (isa<BranchInst>(U) && cast<BranchInst>(U)->isConditional()) | ||||
1533 | continue; | ||||
1534 | if (isa<SelectInst>(U) && cast<SelectInst>(U)->getCondition() == ICmp) | ||||
1535 | continue; | ||||
1536 | return false; | ||||
1537 | } | ||||
1538 | for (User *U : ICmp->users()) { | ||||
1539 | if (U == ExcludedUser) | ||||
1540 | continue; | ||||
1541 | if (auto *BI = dyn_cast<BranchInst>(U)) { | ||||
1542 | assert(BI->isConditional() && "Must be conditional")((void)0); | ||||
1543 | BI->swapSuccessors(); | ||||
1544 | // Don't need to swap this in terms of | ||||
1545 | // TrueBiasedRegions/FalseBiasedRegions because true-based/false-based | ||||
1546 | // mean whehter the branch is likely go into the if-then rather than | ||||
1547 | // successor0/successor1 and because we can tell which edge is the then or | ||||
1548 | // the else one by comparing the destination to the region exit block. | ||||
1549 | continue; | ||||
1550 | } | ||||
1551 | if (auto *SI = dyn_cast<SelectInst>(U)) { | ||||
1552 | // Swap operands | ||||
1553 | SI->swapValues(); | ||||
1554 | SI->swapProfMetadata(); | ||||
1555 | if (Scope->TrueBiasedSelects.count(SI)) { | ||||
1556 | assert(Scope->FalseBiasedSelects.count(SI) == 0 &&((void)0) | ||||
1557 | "Must not be already in")((void)0); | ||||
1558 | Scope->FalseBiasedSelects.insert(SI); | ||||
1559 | } else if (Scope->FalseBiasedSelects.count(SI)) { | ||||
1560 | assert(Scope->TrueBiasedSelects.count(SI) == 0 &&((void)0) | ||||
1561 | "Must not be already in")((void)0); | ||||
1562 | Scope->TrueBiasedSelects.insert(SI); | ||||
1563 | } | ||||
1564 | continue; | ||||
1565 | } | ||||
1566 | llvm_unreachable("Must be a branch or a select")__builtin_unreachable(); | ||||
1567 | } | ||||
1568 | ICmp->setPredicate(CmpInst::getInversePredicate(ICmp->getPredicate())); | ||||
1569 | return true; | ||||
1570 | } | ||||
1571 | |||||
1572 | // A helper for transformScopes. Insert a trivial phi at the scope exit block | ||||
1573 | // for a value that's defined in the scope but used outside it (meaning it's | ||||
1574 | // alive at the exit block). | ||||
1575 | static void insertTrivialPHIs(CHRScope *Scope, | ||||
1576 | BasicBlock *EntryBlock, BasicBlock *ExitBlock, | ||||
1577 | DenseSet<PHINode *> &TrivialPHIs) { | ||||
1578 | SmallSetVector<BasicBlock *, 8> BlocksInScope; | ||||
1579 | for (RegInfo &RI : Scope->RegInfos) { | ||||
1580 | for (BasicBlock *BB : RI.R->blocks()) { // This includes the blocks in the | ||||
1581 | // sub-Scopes. | ||||
1582 | BlocksInScope.insert(BB); | ||||
1583 | } | ||||
1584 | } | ||||
1585 | CHR_DEBUG({do { } while (false) | ||||
1586 | dbgs() << "Inserting redundant phis\n";do { } while (false) | ||||
1587 | for (BasicBlock *BB : BlocksInScope)do { } while (false) | ||||
1588 | dbgs() << "BlockInScope " << BB->getName() << "\n";do { } while (false) | ||||
1589 | })do { } while (false); | ||||
1590 | for (BasicBlock *BB : BlocksInScope) { | ||||
1591 | for (Instruction &I : *BB) { | ||||
1592 | SmallVector<Instruction *, 8> Users; | ||||
1593 | for (User *U : I.users()) { | ||||
1594 | if (auto *UI = dyn_cast<Instruction>(U)) { | ||||
1595 | if (BlocksInScope.count(UI->getParent()) == 0 && | ||||
1596 | // Unless there's already a phi for I at the exit block. | ||||
1597 | !(isa<PHINode>(UI) && UI->getParent() == ExitBlock)) { | ||||
1598 | CHR_DEBUG(dbgs() << "V " << I << "\n")do { } while (false); | ||||
1599 | CHR_DEBUG(dbgs() << "Used outside scope by user " << *UI << "\n")do { } while (false); | ||||
1600 | Users.push_back(UI); | ||||
1601 | } else if (UI->getParent() == EntryBlock && isa<PHINode>(UI)) { | ||||
1602 | // There's a loop backedge from a block that's dominated by this | ||||
1603 | // scope to the entry block. | ||||
1604 | CHR_DEBUG(dbgs() << "V " << I << "\n")do { } while (false); | ||||
1605 | CHR_DEBUG(dbgs()do { } while (false) | ||||
1606 | << "Used at entry block (for a back edge) by a phi user "do { } while (false) | ||||
1607 | << *UI << "\n")do { } while (false); | ||||
1608 | Users.push_back(UI); | ||||
1609 | } | ||||
1610 | } | ||||
1611 | } | ||||
1612 | if (Users.size() > 0) { | ||||
1613 | // Insert a trivial phi for I (phi [&I, P0], [&I, P1], ...) at | ||||
1614 | // ExitBlock. Replace I with the new phi in UI unless UI is another | ||||
1615 | // phi at ExitBlock. | ||||
1616 | PHINode *PN = PHINode::Create(I.getType(), pred_size(ExitBlock), "", | ||||
1617 | &ExitBlock->front()); | ||||
1618 | for (BasicBlock *Pred : predecessors(ExitBlock)) { | ||||
1619 | PN->addIncoming(&I, Pred); | ||||
1620 | } | ||||
1621 | TrivialPHIs.insert(PN); | ||||
1622 | CHR_DEBUG(dbgs() << "Insert phi " << *PN << "\n")do { } while (false); | ||||
1623 | for (Instruction *UI : Users) { | ||||
1624 | for (unsigned J = 0, NumOps = UI->getNumOperands(); J < NumOps; ++J) { | ||||
1625 | if (UI->getOperand(J) == &I) { | ||||
1626 | UI->setOperand(J, PN); | ||||
1627 | } | ||||
1628 | } | ||||
1629 | CHR_DEBUG(dbgs() << "Updated user " << *UI << "\n")do { } while (false); | ||||
1630 | } | ||||
1631 | } | ||||
1632 | } | ||||
1633 | } | ||||
1634 | } | ||||
1635 | |||||
1636 | // Assert that all the CHR regions of the scope have a biased branch or select. | ||||
1637 | static void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) | ||||
1638 | assertCHRRegionsHaveBiasedBranchOrSelect(CHRScope *Scope) { | ||||
1639 | #ifndef NDEBUG1 | ||||
1640 | auto HasBiasedBranchOrSelect = [](RegInfo &RI, CHRScope *Scope) { | ||||
1641 | if (Scope->TrueBiasedRegions.count(RI.R) || | ||||
1642 | Scope->FalseBiasedRegions.count(RI.R)) | ||||
1643 | return true; | ||||
1644 | for (SelectInst *SI : RI.Selects) | ||||
1645 | if (Scope->TrueBiasedSelects.count(SI) || | ||||
1646 | Scope->FalseBiasedSelects.count(SI)) | ||||
1647 | return true; | ||||
1648 | return false; | ||||
1649 | }; | ||||
1650 | for (RegInfo &RI : Scope->CHRRegions) { | ||||
1651 | assert(HasBiasedBranchOrSelect(RI, Scope) &&((void)0) | ||||
1652 | "Must have biased branch or select")((void)0); | ||||
1653 | } | ||||
1654 | #endif | ||||
1655 | } | ||||
1656 | |||||
1657 | // Assert that all the condition values of the biased branches and selects have | ||||
1658 | // been hoisted to the pre-entry block or outside of the scope. | ||||
1659 | static void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) assertBranchOrSelectConditionHoisted( | ||||
1660 | CHRScope *Scope, BasicBlock *PreEntryBlock) { | ||||
1661 | CHR_DEBUG(dbgs() << "Biased regions condition values \n")do { } while (false); | ||||
1662 | for (RegInfo &RI : Scope->CHRRegions) { | ||||
1663 | Region *R = RI.R; | ||||
1664 | bool IsTrueBiased = Scope->TrueBiasedRegions.count(R); | ||||
1665 | bool IsFalseBiased = Scope->FalseBiasedRegions.count(R); | ||||
1666 | if (RI.HasBranch && (IsTrueBiased || IsFalseBiased)) { | ||||
1667 | auto *BI = cast<BranchInst>(R->getEntry()->getTerminator()); | ||||
1668 | Value *V = BI->getCondition(); | ||||
1669 | CHR_DEBUG(dbgs() << *V << "\n")do { } while (false); | ||||
1670 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
1671 | (void)(I); // Unused in release build. | ||||
1672 | assert((I->getParent() == PreEntryBlock ||((void)0) | ||||
1673 | !Scope->contains(I)) &&((void)0) | ||||
1674 | "Must have been hoisted to PreEntryBlock or outside the scope")((void)0); | ||||
1675 | } | ||||
1676 | } | ||||
1677 | for (SelectInst *SI : RI.Selects) { | ||||
1678 | bool IsTrueBiased = Scope->TrueBiasedSelects.count(SI); | ||||
1679 | bool IsFalseBiased = Scope->FalseBiasedSelects.count(SI); | ||||
1680 | if (!(IsTrueBiased || IsFalseBiased)) | ||||
1681 | continue; | ||||
1682 | Value *V = SI->getCondition(); | ||||
1683 | CHR_DEBUG(dbgs() << *V << "\n")do { } while (false); | ||||
1684 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
1685 | (void)(I); // Unused in release build. | ||||
1686 | assert((I->getParent() == PreEntryBlock ||((void)0) | ||||
1687 | !Scope->contains(I)) &&((void)0) | ||||
1688 | "Must have been hoisted to PreEntryBlock or outside the scope")((void)0); | ||||
1689 | } | ||||
1690 | } | ||||
1691 | } | ||||
1692 | } | ||||
1693 | |||||
1694 | void CHR::transformScopes(CHRScope *Scope, DenseSet<PHINode *> &TrivialPHIs) { | ||||
1695 | CHR_DEBUG(dbgs() << "transformScopes " << *Scope << "\n")do { } while (false); | ||||
1696 | |||||
1697 | assert(Scope->RegInfos.size() >= 1 && "Should have at least one Region")((void)0); | ||||
1698 | Region *FirstRegion = Scope->RegInfos[0].R; | ||||
1699 | BasicBlock *EntryBlock = FirstRegion->getEntry(); | ||||
1700 | Region *LastRegion = Scope->RegInfos[Scope->RegInfos.size() - 1].R; | ||||
1701 | BasicBlock *ExitBlock = LastRegion->getExit(); | ||||
1702 | Optional<uint64_t> ProfileCount = BFI.getBlockProfileCount(EntryBlock); | ||||
1703 | |||||
1704 | if (ExitBlock) { | ||||
1705 | // Insert a trivial phi at the exit block (where the CHR hot path and the | ||||
1706 | // cold path merges) for a value that's defined in the scope but used | ||||
1707 | // outside it (meaning it's alive at the exit block). We will add the | ||||
1708 | // incoming values for the CHR cold paths to it below. Without this, we'd | ||||
1709 | // miss updating phi's for such values unless there happens to already be a | ||||
1710 | // phi for that value there. | ||||
1711 | insertTrivialPHIs(Scope, EntryBlock, ExitBlock, TrivialPHIs); | ||||
1712 | } | ||||
1713 | |||||
1714 | // Split the entry block of the first region. The new block becomes the new | ||||
1715 | // entry block of the first region. The old entry block becomes the block to | ||||
1716 | // insert the CHR branch into. Note DT gets updated. Since DT gets updated | ||||
1717 | // through the split, we update the entry of the first region after the split, | ||||
1718 | // and Region only points to the entry and the exit blocks, rather than | ||||
1719 | // keeping everything in a list or set, the blocks membership and the | ||||
1720 | // entry/exit blocks of the region are still valid after the split. | ||||
1721 | CHR_DEBUG(dbgs() << "Splitting entry block " << EntryBlock->getName()do { } while (false) | ||||
1722 | << " at " << *Scope->BranchInsertPoint << "\n")do { } while (false); | ||||
1723 | BasicBlock *NewEntryBlock = | ||||
1724 | SplitBlock(EntryBlock, Scope->BranchInsertPoint, &DT); | ||||
1725 | assert(NewEntryBlock->getSinglePredecessor() == EntryBlock &&((void)0) | ||||
1726 | "NewEntryBlock's only pred must be EntryBlock")((void)0); | ||||
1727 | FirstRegion->replaceEntryRecursive(NewEntryBlock); | ||||
1728 | BasicBlock *PreEntryBlock = EntryBlock; | ||||
1729 | |||||
1730 | ValueToValueMapTy VMap; | ||||
1731 | // Clone the blocks in the scope (excluding the PreEntryBlock) to split into a | ||||
1732 | // hot path (originals) and a cold path (clones) and update the PHIs at the | ||||
1733 | // exit block. | ||||
1734 | cloneScopeBlocks(Scope, PreEntryBlock, ExitBlock, LastRegion, VMap); | ||||
1735 | |||||
1736 | // Replace the old (placeholder) branch with the new (merged) conditional | ||||
1737 | // branch. | ||||
1738 | BranchInst *MergedBr = createMergedBranch(PreEntryBlock, EntryBlock, | ||||
1739 | NewEntryBlock, VMap); | ||||
1740 | |||||
1741 | #ifndef NDEBUG1 | ||||
1742 | assertCHRRegionsHaveBiasedBranchOrSelect(Scope); | ||||
1743 | #endif | ||||
1744 | |||||
1745 | // Hoist the conditional values of the branches/selects. | ||||
1746 | hoistScopeConditions(Scope, PreEntryBlock->getTerminator(), TrivialPHIs, DT); | ||||
1747 | |||||
1748 | #ifndef NDEBUG1 | ||||
1749 | assertBranchOrSelectConditionHoisted(Scope, PreEntryBlock); | ||||
1750 | #endif | ||||
1751 | |||||
1752 | // Create the combined branch condition and constant-fold the branches/selects | ||||
1753 | // in the hot path. | ||||
1754 | fixupBranchesAndSelects(Scope, PreEntryBlock, MergedBr, | ||||
1755 | ProfileCount ? ProfileCount.getValue() : 0); | ||||
1756 | } | ||||
1757 | |||||
1758 | // A helper for transformScopes. Clone the blocks in the scope (excluding the | ||||
1759 | // PreEntryBlock) to split into a hot path and a cold path and update the PHIs | ||||
1760 | // at the exit block. | ||||
1761 | void CHR::cloneScopeBlocks(CHRScope *Scope, | ||||
1762 | BasicBlock *PreEntryBlock, | ||||
1763 | BasicBlock *ExitBlock, | ||||
1764 | Region *LastRegion, | ||||
1765 | ValueToValueMapTy &VMap) { | ||||
1766 | // Clone all the blocks. The original blocks will be the hot-path | ||||
1767 | // CHR-optimized code and the cloned blocks will be the original unoptimized | ||||
1768 | // code. This is so that the block pointers from the | ||||
1769 | // CHRScope/Region/RegionInfo can stay valid in pointing to the hot-path code | ||||
1770 | // which CHR should apply to. | ||||
1771 | SmallVector<BasicBlock*, 8> NewBlocks; | ||||
1772 | for (RegInfo &RI : Scope->RegInfos) | ||||
1773 | for (BasicBlock *BB : RI.R->blocks()) { // This includes the blocks in the | ||||
1774 | // sub-Scopes. | ||||
1775 | assert(BB != PreEntryBlock && "Don't copy the preetntry block")((void)0); | ||||
1776 | BasicBlock *NewBB = CloneBasicBlock(BB, VMap, ".nonchr", &F); | ||||
1777 | NewBlocks.push_back(NewBB); | ||||
1778 | VMap[BB] = NewBB; | ||||
1779 | } | ||||
1780 | |||||
1781 | // Place the cloned blocks right after the original blocks (right before the | ||||
1782 | // exit block of.) | ||||
1783 | if (ExitBlock) | ||||
1784 | F.getBasicBlockList().splice(ExitBlock->getIterator(), | ||||
1785 | F.getBasicBlockList(), | ||||
1786 | NewBlocks[0]->getIterator(), F.end()); | ||||
1787 | |||||
1788 | // Update the cloned blocks/instructions to refer to themselves. | ||||
1789 | for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) | ||||
1790 | for (Instruction &I : *NewBlocks[i]) | ||||
1791 | RemapInstruction(&I, VMap, | ||||
1792 | RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); | ||||
1793 | |||||
1794 | // Add the cloned blocks to the PHIs of the exit blocks. ExitBlock is null for | ||||
1795 | // the top-level region but we don't need to add PHIs. The trivial PHIs | ||||
1796 | // inserted above will be updated here. | ||||
1797 | if (ExitBlock) | ||||
1798 | for (PHINode &PN : ExitBlock->phis()) | ||||
1799 | for (unsigned I = 0, NumOps = PN.getNumIncomingValues(); I < NumOps; | ||||
1800 | ++I) { | ||||
1801 | BasicBlock *Pred = PN.getIncomingBlock(I); | ||||
1802 | if (LastRegion->contains(Pred)) { | ||||
1803 | Value *V = PN.getIncomingValue(I); | ||||
1804 | auto It = VMap.find(V); | ||||
1805 | if (It != VMap.end()) V = It->second; | ||||
1806 | assert(VMap.find(Pred) != VMap.end() && "Pred must have been cloned")((void)0); | ||||
1807 | PN.addIncoming(V, cast<BasicBlock>(VMap[Pred])); | ||||
1808 | } | ||||
1809 | } | ||||
1810 | } | ||||
1811 | |||||
1812 | // A helper for transformScope. Replace the old (placeholder) branch with the | ||||
1813 | // new (merged) conditional branch. | ||||
1814 | BranchInst *CHR::createMergedBranch(BasicBlock *PreEntryBlock, | ||||
1815 | BasicBlock *EntryBlock, | ||||
1816 | BasicBlock *NewEntryBlock, | ||||
1817 | ValueToValueMapTy &VMap) { | ||||
1818 | BranchInst *OldBR = cast<BranchInst>(PreEntryBlock->getTerminator()); | ||||
1819 | assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == NewEntryBlock &&((void)0) | ||||
1820 | "SplitBlock did not work correctly!")((void)0); | ||||
1821 | assert(NewEntryBlock->getSinglePredecessor() == EntryBlock &&((void)0) | ||||
1822 | "NewEntryBlock's only pred must be EntryBlock")((void)0); | ||||
1823 | assert(VMap.find(NewEntryBlock) != VMap.end() &&((void)0) | ||||
1824 | "NewEntryBlock must have been copied")((void)0); | ||||
1825 | OldBR->dropAllReferences(); | ||||
1826 | OldBR->eraseFromParent(); | ||||
1827 | // The true predicate is a placeholder. It will be replaced later in | ||||
1828 | // fixupBranchesAndSelects(). | ||||
1829 | BranchInst *NewBR = BranchInst::Create(NewEntryBlock, | ||||
1830 | cast<BasicBlock>(VMap[NewEntryBlock]), | ||||
1831 | ConstantInt::getTrue(F.getContext())); | ||||
1832 | PreEntryBlock->getInstList().push_back(NewBR); | ||||
1833 | assert(NewEntryBlock->getSinglePredecessor() == EntryBlock &&((void)0) | ||||
1834 | "NewEntryBlock's only pred must be EntryBlock")((void)0); | ||||
1835 | return NewBR; | ||||
1836 | } | ||||
1837 | |||||
1838 | // A helper for transformScopes. Create the combined branch condition and | ||||
1839 | // constant-fold the branches/selects in the hot path. | ||||
1840 | void CHR::fixupBranchesAndSelects(CHRScope *Scope, | ||||
1841 | BasicBlock *PreEntryBlock, | ||||
1842 | BranchInst *MergedBR, | ||||
1843 | uint64_t ProfileCount) { | ||||
1844 | Value *MergedCondition = ConstantInt::getTrue(F.getContext()); | ||||
1845 | BranchProbability CHRBranchBias(1, 1); | ||||
1846 | uint64_t NumCHRedBranches = 0; | ||||
1847 | IRBuilder<> IRB(PreEntryBlock->getTerminator()); | ||||
1848 | for (RegInfo &RI : Scope->CHRRegions) { | ||||
1849 | Region *R = RI.R; | ||||
1850 | if (RI.HasBranch) { | ||||
1851 | fixupBranch(R, Scope, IRB, MergedCondition, CHRBranchBias); | ||||
1852 | ++NumCHRedBranches; | ||||
1853 | } | ||||
1854 | for (SelectInst *SI : RI.Selects) { | ||||
1855 | fixupSelect(SI, Scope, IRB, MergedCondition, CHRBranchBias); | ||||
1856 | ++NumCHRedBranches; | ||||
1857 | } | ||||
1858 | } | ||||
1859 | Stats.NumBranchesDelta += NumCHRedBranches - 1; | ||||
1860 | Stats.WeightedNumBranchesDelta += (NumCHRedBranches - 1) * ProfileCount; | ||||
1861 | ORE.emit([&]() { | ||||
1862 | return OptimizationRemark(DEBUG_TYPE"chr", | ||||
1863 | "CHR", | ||||
1864 | // Refer to the hot (original) path | ||||
1865 | MergedBR->getSuccessor(0)->getTerminator()) | ||||
1866 | << "Merged " << ore::NV("NumCHRedBranches", NumCHRedBranches) | ||||
1867 | << " branches or selects"; | ||||
1868 | }); | ||||
1869 | MergedBR->setCondition(MergedCondition); | ||||
1870 | uint32_t Weights[] = { | ||||
1871 | static_cast<uint32_t>(CHRBranchBias.scale(1000)), | ||||
1872 | static_cast<uint32_t>(CHRBranchBias.getCompl().scale(1000)), | ||||
1873 | }; | ||||
1874 | MDBuilder MDB(F.getContext()); | ||||
1875 | MergedBR->setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights)); | ||||
1876 | CHR_DEBUG(dbgs() << "CHR branch bias " << Weights[0] << ":" << Weights[1]do { } while (false) | ||||
1877 | << "\n")do { } while (false); | ||||
1878 | } | ||||
1879 | |||||
1880 | // A helper for fixupBranchesAndSelects. Add to the combined branch condition | ||||
1881 | // and constant-fold a branch in the hot path. | ||||
1882 | void CHR::fixupBranch(Region *R, CHRScope *Scope, | ||||
1883 | IRBuilder<> &IRB, | ||||
1884 | Value *&MergedCondition, | ||||
1885 | BranchProbability &CHRBranchBias) { | ||||
1886 | bool IsTrueBiased = Scope->TrueBiasedRegions.count(R); | ||||
1887 | assert((IsTrueBiased || Scope->FalseBiasedRegions.count(R)) &&((void)0) | ||||
1888 | "Must be truthy or falsy")((void)0); | ||||
1889 | auto *BI = cast<BranchInst>(R->getEntry()->getTerminator()); | ||||
1890 | assert(BranchBiasMap.find(R) != BranchBiasMap.end() &&((void)0) | ||||
1891 | "Must be in the bias map")((void)0); | ||||
1892 | BranchProbability Bias = BranchBiasMap[R]; | ||||
1893 | assert(Bias >= getCHRBiasThreshold() && "Must be highly biased")((void)0); | ||||
1894 | // Take the min. | ||||
1895 | if (CHRBranchBias > Bias) | ||||
1896 | CHRBranchBias = Bias; | ||||
1897 | BasicBlock *IfThen = BI->getSuccessor(1); | ||||
1898 | BasicBlock *IfElse = BI->getSuccessor(0); | ||||
1899 | BasicBlock *RegionExitBlock = R->getExit(); | ||||
1900 | assert(RegionExitBlock && "Null ExitBlock")((void)0); | ||||
1901 | assert((IfThen == RegionExitBlock || IfElse == RegionExitBlock) &&((void)0) | ||||
1902 | IfThen != IfElse && "Invariant from findScopes")((void)0); | ||||
1903 | if (IfThen == RegionExitBlock) { | ||||
1904 | // Swap them so that IfThen means going into it and IfElse means skipping | ||||
1905 | // it. | ||||
1906 | std::swap(IfThen, IfElse); | ||||
1907 | } | ||||
1908 | CHR_DEBUG(dbgs() << "IfThen " << IfThen->getName()do { } while (false) | ||||
1909 | << " IfElse " << IfElse->getName() << "\n")do { } while (false); | ||||
1910 | Value *Cond = BI->getCondition(); | ||||
1911 | BasicBlock *HotTarget = IsTrueBiased ? IfThen : IfElse; | ||||
1912 | bool ConditionTrue = HotTarget == BI->getSuccessor(0); | ||||
1913 | addToMergedCondition(ConditionTrue, Cond, BI, Scope, IRB, | ||||
1914 | MergedCondition); | ||||
1915 | // Constant-fold the branch at ClonedEntryBlock. | ||||
1916 | assert(ConditionTrue == (HotTarget == BI->getSuccessor(0)) &&((void)0) | ||||
1917 | "The successor shouldn't change")((void)0); | ||||
1918 | Value *NewCondition = ConditionTrue ? | ||||
1919 | ConstantInt::getTrue(F.getContext()) : | ||||
1920 | ConstantInt::getFalse(F.getContext()); | ||||
1921 | BI->setCondition(NewCondition); | ||||
1922 | } | ||||
1923 | |||||
1924 | // A helper for fixupBranchesAndSelects. Add to the combined branch condition | ||||
1925 | // and constant-fold a select in the hot path. | ||||
1926 | void CHR::fixupSelect(SelectInst *SI, CHRScope *Scope, | ||||
1927 | IRBuilder<> &IRB, | ||||
1928 | Value *&MergedCondition, | ||||
1929 | BranchProbability &CHRBranchBias) { | ||||
1930 | bool IsTrueBiased = Scope->TrueBiasedSelects.count(SI); | ||||
1931 | assert((IsTrueBiased ||((void)0) | ||||
1932 | Scope->FalseBiasedSelects.count(SI)) && "Must be biased")((void)0); | ||||
1933 | assert(SelectBiasMap.find(SI) != SelectBiasMap.end() &&((void)0) | ||||
1934 | "Must be in the bias map")((void)0); | ||||
1935 | BranchProbability Bias = SelectBiasMap[SI]; | ||||
1936 | assert(Bias >= getCHRBiasThreshold() && "Must be highly biased")((void)0); | ||||
1937 | // Take the min. | ||||
1938 | if (CHRBranchBias > Bias) | ||||
1939 | CHRBranchBias = Bias; | ||||
1940 | Value *Cond = SI->getCondition(); | ||||
1941 | addToMergedCondition(IsTrueBiased, Cond, SI, Scope, IRB, | ||||
1942 | MergedCondition); | ||||
1943 | Value *NewCondition = IsTrueBiased ? | ||||
1944 | ConstantInt::getTrue(F.getContext()) : | ||||
1945 | ConstantInt::getFalse(F.getContext()); | ||||
1946 | SI->setCondition(NewCondition); | ||||
1947 | } | ||||
1948 | |||||
1949 | // A helper for fixupBranch/fixupSelect. Add a branch condition to the merged | ||||
1950 | // condition. | ||||
1951 | void CHR::addToMergedCondition(bool IsTrueBiased, Value *Cond, | ||||
1952 | Instruction *BranchOrSelect, | ||||
1953 | CHRScope *Scope, | ||||
1954 | IRBuilder<> &IRB, | ||||
1955 | Value *&MergedCondition) { | ||||
1956 | if (IsTrueBiased) { | ||||
1957 | MergedCondition = IRB.CreateAnd(MergedCondition, Cond); | ||||
1958 | } else { | ||||
1959 | // If Cond is an icmp and all users of V except for BranchOrSelect is a | ||||
1960 | // branch, negate the icmp predicate and swap the branch targets and avoid | ||||
1961 | // inserting an Xor to negate Cond. | ||||
1962 | bool Done = false; | ||||
1963 | if (auto *ICmp = dyn_cast<ICmpInst>(Cond)) | ||||
1964 | if (negateICmpIfUsedByBranchOrSelectOnly(ICmp, BranchOrSelect, Scope)) { | ||||
1965 | MergedCondition = IRB.CreateAnd(MergedCondition, Cond); | ||||
1966 | Done = true; | ||||
1967 | } | ||||
1968 | if (!Done) { | ||||
1969 | Value *Negate = IRB.CreateXor( | ||||
1970 | ConstantInt::getTrue(F.getContext()), Cond); | ||||
1971 | MergedCondition = IRB.CreateAnd(MergedCondition, Negate); | ||||
1972 | } | ||||
1973 | } | ||||
1974 | } | ||||
1975 | |||||
1976 | void CHR::transformScopes(SmallVectorImpl<CHRScope *> &CHRScopes) { | ||||
1977 | unsigned I = 0; | ||||
1978 | DenseSet<PHINode *> TrivialPHIs; | ||||
1979 | for (CHRScope *Scope : CHRScopes) { | ||||
1980 | transformScopes(Scope, TrivialPHIs); | ||||
1981 | CHR_DEBUG(do { } while (false) | ||||
1982 | std::ostringstream oss;do { } while (false) | ||||
1983 | oss << " after transformScopes " << I++;do { } while (false) | ||||
1984 | dumpIR(F, oss.str().c_str(), nullptr))do { } while (false); | ||||
1985 | (void)I; | ||||
1986 | } | ||||
1987 | } | ||||
1988 | |||||
1989 | static void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) | ||||
1990 | dumpScopes(SmallVectorImpl<CHRScope *> &Scopes, const char *Label) { | ||||
1991 | dbgs() << Label << " " << Scopes.size() << "\n"; | ||||
1992 | for (CHRScope *Scope : Scopes) { | ||||
1993 | dbgs() << *Scope << "\n"; | ||||
1994 | } | ||||
1995 | } | ||||
1996 | |||||
1997 | bool CHR::run() { | ||||
1998 | if (!shouldApply(F, PSI)) | ||||
1999 | return false; | ||||
2000 | |||||
2001 | CHR_DEBUG(dumpIR(F, "before", nullptr))do { } while (false); | ||||
2002 | |||||
2003 | bool Changed = false; | ||||
2004 | { | ||||
2005 | CHR_DEBUG(do { } while (false) | ||||
2006 | dbgs() << "RegionInfo:\n";do { } while (false) | ||||
2007 | RI.print(dbgs()))do { } while (false); | ||||
2008 | |||||
2009 | // Recursively traverse the region tree and find regions that have biased | ||||
2010 | // branches and/or selects and create scopes. | ||||
2011 | SmallVector<CHRScope *, 8> AllScopes; | ||||
2012 | findScopes(AllScopes); | ||||
2013 | CHR_DEBUG(dumpScopes(AllScopes, "All scopes"))do { } while (false); | ||||
2014 | |||||
2015 | // Split the scopes if 1) the conditiona values of the biased | ||||
2016 | // branches/selects of the inner/lower scope can't be hoisted up to the | ||||
2017 | // outermost/uppermost scope entry, or 2) the condition values of the biased | ||||
2018 | // branches/selects in a scope (including subscopes) don't share at least | ||||
2019 | // one common value. | ||||
2020 | SmallVector<CHRScope *, 8> SplitScopes; | ||||
2021 | splitScopes(AllScopes, SplitScopes); | ||||
2022 | CHR_DEBUG(dumpScopes(SplitScopes, "Split scopes"))do { } while (false); | ||||
2023 | |||||
2024 | // After splitting, set the biased regions and selects of a scope (a tree | ||||
2025 | // root) that include those of the subscopes. | ||||
2026 | classifyBiasedScopes(SplitScopes); | ||||
2027 | CHR_DEBUG(dbgs() << "Set per-scope bias " << SplitScopes.size() << "\n")do { } while (false); | ||||
2028 | |||||
2029 | // Filter out the scopes that has only one biased region or select (CHR | ||||
2030 | // isn't useful in such a case). | ||||
2031 | SmallVector<CHRScope *, 8> FilteredScopes; | ||||
2032 | filterScopes(SplitScopes, FilteredScopes); | ||||
2033 | CHR_DEBUG(dumpScopes(FilteredScopes, "Filtered scopes"))do { } while (false); | ||||
2034 | |||||
2035 | // Set the regions to be CHR'ed and their hoist stops for each scope. | ||||
2036 | SmallVector<CHRScope *, 8> SetScopes; | ||||
2037 | setCHRRegions(FilteredScopes, SetScopes); | ||||
2038 | CHR_DEBUG(dumpScopes(SetScopes, "Set CHR regions"))do { } while (false); | ||||
2039 | |||||
2040 | // Sort CHRScopes by the depth so that outer CHRScopes comes before inner | ||||
2041 | // ones. We need to apply CHR from outer to inner so that we apply CHR only | ||||
2042 | // to the hot path, rather than both hot and cold paths. | ||||
2043 | SmallVector<CHRScope *, 8> SortedScopes; | ||||
2044 | sortScopes(SetScopes, SortedScopes); | ||||
2045 | CHR_DEBUG(dumpScopes(SortedScopes, "Sorted scopes"))do { } while (false); | ||||
2046 | |||||
2047 | CHR_DEBUG(do { } while (false) | ||||
2048 | dbgs() << "RegionInfo:\n";do { } while (false) | ||||
2049 | RI.print(dbgs()))do { } while (false); | ||||
2050 | |||||
2051 | // Apply the CHR transformation. | ||||
2052 | if (!SortedScopes.empty()) { | ||||
2053 | transformScopes(SortedScopes); | ||||
2054 | Changed = true; | ||||
2055 | } | ||||
2056 | } | ||||
2057 | |||||
2058 | if (Changed) { | ||||
2059 | CHR_DEBUG(dumpIR(F, "after", &Stats))do { } while (false); | ||||
2060 | ORE.emit([&]() { | ||||
2061 | return OptimizationRemark(DEBUG_TYPE"chr", "Stats", &F) | ||||
2062 | << ore::NV("Function", &F) << " " | ||||
2063 | << "Reduced the number of branches in hot paths by " | ||||
2064 | << ore::NV("NumBranchesDelta", Stats.NumBranchesDelta) | ||||
2065 | << " (static) and " | ||||
2066 | << ore::NV("WeightedNumBranchesDelta", Stats.WeightedNumBranchesDelta) | ||||
2067 | << " (weighted by PGO count)"; | ||||
2068 | }); | ||||
2069 | } | ||||
2070 | |||||
2071 | return Changed; | ||||
2072 | } | ||||
2073 | |||||
2074 | bool ControlHeightReductionLegacyPass::runOnFunction(Function &F) { | ||||
2075 | BlockFrequencyInfo &BFI = | ||||
2076 | getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); | ||||
2077 | DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||
2078 | ProfileSummaryInfo &PSI = | ||||
2079 | getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | ||||
2080 | RegionInfo &RI = getAnalysis<RegionInfoPass>().getRegionInfo(); | ||||
2081 | std::unique_ptr<OptimizationRemarkEmitter> OwnedORE = | ||||
2082 | std::make_unique<OptimizationRemarkEmitter>(&F); | ||||
2083 | return CHR(F, BFI, DT, PSI, RI, *OwnedORE.get()).run(); | ||||
2084 | } | ||||
2085 | |||||
2086 | namespace llvm { | ||||
2087 | |||||
2088 | ControlHeightReductionPass::ControlHeightReductionPass() { | ||||
2089 | parseCHRFilterFiles(); | ||||
2090 | } | ||||
2091 | |||||
2092 | PreservedAnalyses ControlHeightReductionPass::run( | ||||
2093 | Function &F, | ||||
2094 | FunctionAnalysisManager &FAM) { | ||||
2095 | auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(F); | ||||
2096 | auto &DT = FAM.getResult<DominatorTreeAnalysis>(F); | ||||
2097 | auto &MAMProxy = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(F); | ||||
2098 | auto &PSI = *MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); | ||||
2099 | auto &RI = FAM.getResult<RegionInfoAnalysis>(F); | ||||
2100 | auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F); | ||||
2101 | bool Changed = CHR(F, BFI, DT, PSI, RI, ORE).run(); | ||||
2102 | if (!Changed) | ||||
2103 | return PreservedAnalyses::all(); | ||||
2104 | return PreservedAnalyses::none(); | ||||
2105 | } | ||||
2106 | |||||
2107 | } // namespace llvm |
1 | //===- llvm/ADT/SmallVector.h - 'Normally small' vectors --------*- C++ -*-===// |
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 defines the SmallVector class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_ADT_SMALLVECTOR_H |
14 | #define LLVM_ADT_SMALLVECTOR_H |
15 | |
16 | #include "llvm/ADT/iterator_range.h" |
17 | #include "llvm/Support/Compiler.h" |
18 | #include "llvm/Support/ErrorHandling.h" |
19 | #include "llvm/Support/MemAlloc.h" |
20 | #include "llvm/Support/type_traits.h" |
21 | #include <algorithm> |
22 | #include <cassert> |
23 | #include <cstddef> |
24 | #include <cstdlib> |
25 | #include <cstring> |
26 | #include <functional> |
27 | #include <initializer_list> |
28 | #include <iterator> |
29 | #include <limits> |
30 | #include <memory> |
31 | #include <new> |
32 | #include <type_traits> |
33 | #include <utility> |
34 | |
35 | namespace llvm { |
36 | |
37 | /// This is all the stuff common to all SmallVectors. |
38 | /// |
39 | /// The template parameter specifies the type which should be used to hold the |
40 | /// Size and Capacity of the SmallVector, so it can be adjusted. |
41 | /// Using 32 bit size is desirable to shrink the size of the SmallVector. |
42 | /// Using 64 bit size is desirable for cases like SmallVector<char>, where a |
43 | /// 32 bit size would limit the vector to ~4GB. SmallVectors are used for |
44 | /// buffering bitcode output - which can exceed 4GB. |
45 | template <class Size_T> class SmallVectorBase { |
46 | protected: |
47 | void *BeginX; |
48 | Size_T Size = 0, Capacity; |
49 | |
50 | /// The maximum value of the Size_T used. |
51 | static constexpr size_t SizeTypeMax() { |
52 | return std::numeric_limits<Size_T>::max(); |
53 | } |
54 | |
55 | SmallVectorBase() = delete; |
56 | SmallVectorBase(void *FirstEl, size_t TotalCapacity) |
57 | : BeginX(FirstEl), Capacity(TotalCapacity) {} |
58 | |
59 | /// This is a helper for \a grow() that's out of line to reduce code |
60 | /// duplication. This function will report a fatal error if it can't grow at |
61 | /// least to \p MinSize. |
62 | void *mallocForGrow(size_t MinSize, size_t TSize, size_t &NewCapacity); |
63 | |
64 | /// This is an implementation of the grow() method which only works |
65 | /// on POD-like data types and is out of line to reduce code duplication. |
66 | /// This function will report a fatal error if it cannot increase capacity. |
67 | void grow_pod(void *FirstEl, size_t MinSize, size_t TSize); |
68 | |
69 | public: |
70 | size_t size() const { return Size; } |
71 | size_t capacity() const { return Capacity; } |
72 | |
73 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { return !Size; } |
74 | |
75 | /// Set the array size to \p N, which the current array must have enough |
76 | /// capacity for. |
77 | /// |
78 | /// This does not construct or destroy any elements in the vector. |
79 | /// |
80 | /// Clients can use this in conjunction with capacity() to write past the end |
81 | /// of the buffer when they know that more elements are available, and only |
82 | /// update the size later. This avoids the cost of value initializing elements |
83 | /// which will only be overwritten. |
84 | void set_size(size_t N) { |
85 | assert(N <= capacity())((void)0); |
86 | Size = N; |
87 | } |
88 | }; |
89 | |
90 | template <class T> |
91 | using SmallVectorSizeType = |
92 | typename std::conditional<sizeof(T) < 4 && sizeof(void *) >= 8, uint64_t, |
93 | uint32_t>::type; |
94 | |
95 | /// Figure out the offset of the first element. |
96 | template <class T, typename = void> struct SmallVectorAlignmentAndSize { |
97 | alignas(SmallVectorBase<SmallVectorSizeType<T>>) char Base[sizeof( |
98 | SmallVectorBase<SmallVectorSizeType<T>>)]; |
99 | alignas(T) char FirstEl[sizeof(T)]; |
100 | }; |
101 | |
102 | /// This is the part of SmallVectorTemplateBase which does not depend on whether |
103 | /// the type T is a POD. The extra dummy template argument is used by ArrayRef |
104 | /// to avoid unnecessarily requiring T to be complete. |
105 | template <typename T, typename = void> |
106 | class SmallVectorTemplateCommon |
107 | : public SmallVectorBase<SmallVectorSizeType<T>> { |
108 | using Base = SmallVectorBase<SmallVectorSizeType<T>>; |
109 | |
110 | /// Find the address of the first element. For this pointer math to be valid |
111 | /// with small-size of 0 for T with lots of alignment, it's important that |
112 | /// SmallVectorStorage is properly-aligned even for small-size of 0. |
113 | void *getFirstEl() const { |
114 | return const_cast<void *>(reinterpret_cast<const void *>( |
115 | reinterpret_cast<const char *>(this) + |
116 | offsetof(SmallVectorAlignmentAndSize<T>, FirstEl)__builtin_offsetof(SmallVectorAlignmentAndSize<T>, FirstEl ))); |
117 | } |
118 | // Space after 'FirstEl' is clobbered, do not add any instance vars after it. |
119 | |
120 | protected: |
121 | SmallVectorTemplateCommon(size_t Size) : Base(getFirstEl(), Size) {} |
122 | |
123 | void grow_pod(size_t MinSize, size_t TSize) { |
124 | Base::grow_pod(getFirstEl(), MinSize, TSize); |
125 | } |
126 | |
127 | /// Return true if this is a smallvector which has not had dynamic |
128 | /// memory allocated for it. |
129 | bool isSmall() const { return this->BeginX == getFirstEl(); } |
130 | |
131 | /// Put this vector in a state of being small. |
132 | void resetToSmall() { |
133 | this->BeginX = getFirstEl(); |
134 | this->Size = this->Capacity = 0; // FIXME: Setting Capacity to 0 is suspect. |
135 | } |
136 | |
137 | /// Return true if V is an internal reference to the given range. |
138 | bool isReferenceToRange(const void *V, const void *First, const void *Last) const { |
139 | // Use std::less to avoid UB. |
140 | std::less<> LessThan; |
141 | return !LessThan(V, First) && LessThan(V, Last); |
142 | } |
143 | |
144 | /// Return true if V is an internal reference to this vector. |
145 | bool isReferenceToStorage(const void *V) const { |
146 | return isReferenceToRange(V, this->begin(), this->end()); |
147 | } |
148 | |
149 | /// Return true if First and Last form a valid (possibly empty) range in this |
150 | /// vector's storage. |
151 | bool isRangeInStorage(const void *First, const void *Last) const { |
152 | // Use std::less to avoid UB. |
153 | std::less<> LessThan; |
154 | return !LessThan(First, this->begin()) && !LessThan(Last, First) && |
155 | !LessThan(this->end(), Last); |
156 | } |
157 | |
158 | /// Return true unless Elt will be invalidated by resizing the vector to |
159 | /// NewSize. |
160 | bool isSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
161 | // Past the end. |
162 | if (LLVM_LIKELY(!isReferenceToStorage(Elt))__builtin_expect((bool)(!isReferenceToStorage(Elt)), true)) |
163 | return true; |
164 | |
165 | // Return false if Elt will be destroyed by shrinking. |
166 | if (NewSize <= this->size()) |
167 | return Elt < this->begin() + NewSize; |
168 | |
169 | // Return false if we need to grow. |
170 | return NewSize <= this->capacity(); |
171 | } |
172 | |
173 | /// Check whether Elt will be invalidated by resizing the vector to NewSize. |
174 | void assertSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
175 | assert(isSafeToReferenceAfterResize(Elt, NewSize) &&((void)0) |
176 | "Attempting to reference an element of the vector in an operation "((void)0) |
177 | "that invalidates it")((void)0); |
178 | } |
179 | |
180 | /// Check whether Elt will be invalidated by increasing the size of the |
181 | /// vector by N. |
182 | void assertSafeToAdd(const void *Elt, size_t N = 1) { |
183 | this->assertSafeToReferenceAfterResize(Elt, this->size() + N); |
184 | } |
185 | |
186 | /// Check whether any part of the range will be invalidated by clearing. |
187 | void assertSafeToReferenceAfterClear(const T *From, const T *To) { |
188 | if (From == To) |
189 | return; |
190 | this->assertSafeToReferenceAfterResize(From, 0); |
191 | this->assertSafeToReferenceAfterResize(To - 1, 0); |
192 | } |
193 | template < |
194 | class ItTy, |
195 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
196 | bool> = false> |
197 | void assertSafeToReferenceAfterClear(ItTy, ItTy) {} |
198 | |
199 | /// Check whether any part of the range will be invalidated by growing. |
200 | void assertSafeToAddRange(const T *From, const T *To) { |
201 | if (From == To) |
202 | return; |
203 | this->assertSafeToAdd(From, To - From); |
204 | this->assertSafeToAdd(To - 1, To - From); |
205 | } |
206 | template < |
207 | class ItTy, |
208 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
209 | bool> = false> |
210 | void assertSafeToAddRange(ItTy, ItTy) {} |
211 | |
212 | /// Reserve enough space to add one element, and return the updated element |
213 | /// pointer in case it was a reference to the storage. |
214 | template <class U> |
215 | static const T *reserveForParamAndGetAddressImpl(U *This, const T &Elt, |
216 | size_t N) { |
217 | size_t NewSize = This->size() + N; |
218 | if (LLVM_LIKELY(NewSize <= This->capacity())__builtin_expect((bool)(NewSize <= This->capacity()), true )) |
219 | return &Elt; |
220 | |
221 | bool ReferencesStorage = false; |
222 | int64_t Index = -1; |
223 | if (!U::TakesParamByValue) { |
224 | if (LLVM_UNLIKELY(This->isReferenceToStorage(&Elt))__builtin_expect((bool)(This->isReferenceToStorage(&Elt )), false)) { |
225 | ReferencesStorage = true; |
226 | Index = &Elt - This->begin(); |
227 | } |
228 | } |
229 | This->grow(NewSize); |
230 | return ReferencesStorage ? This->begin() + Index : &Elt; |
231 | } |
232 | |
233 | public: |
234 | using size_type = size_t; |
235 | using difference_type = ptrdiff_t; |
236 | using value_type = T; |
237 | using iterator = T *; |
238 | using const_iterator = const T *; |
239 | |
240 | using const_reverse_iterator = std::reverse_iterator<const_iterator>; |
241 | using reverse_iterator = std::reverse_iterator<iterator>; |
242 | |
243 | using reference = T &; |
244 | using const_reference = const T &; |
245 | using pointer = T *; |
246 | using const_pointer = const T *; |
247 | |
248 | using Base::capacity; |
249 | using Base::empty; |
250 | using Base::size; |
251 | |
252 | // forward iterator creation methods. |
253 | iterator begin() { return (iterator)this->BeginX; } |
254 | const_iterator begin() const { return (const_iterator)this->BeginX; } |
255 | iterator end() { return begin() + size(); } |
256 | const_iterator end() const { return begin() + size(); } |
257 | |
258 | // reverse iterator creation methods. |
259 | reverse_iterator rbegin() { return reverse_iterator(end()); } |
260 | const_reverse_iterator rbegin() const{ return const_reverse_iterator(end()); } |
261 | reverse_iterator rend() { return reverse_iterator(begin()); } |
262 | const_reverse_iterator rend() const { return const_reverse_iterator(begin());} |
263 | |
264 | size_type size_in_bytes() const { return size() * sizeof(T); } |
265 | size_type max_size() const { |
266 | return std::min(this->SizeTypeMax(), size_type(-1) / sizeof(T)); |
267 | } |
268 | |
269 | size_t capacity_in_bytes() const { return capacity() * sizeof(T); } |
270 | |
271 | /// Return a pointer to the vector's buffer, even if empty(). |
272 | pointer data() { return pointer(begin()); } |
273 | /// Return a pointer to the vector's buffer, even if empty(). |
274 | const_pointer data() const { return const_pointer(begin()); } |
275 | |
276 | reference operator[](size_type idx) { |
277 | assert(idx < size())((void)0); |
278 | return begin()[idx]; |
279 | } |
280 | const_reference operator[](size_type idx) const { |
281 | assert(idx < size())((void)0); |
282 | return begin()[idx]; |
283 | } |
284 | |
285 | reference front() { |
286 | assert(!empty())((void)0); |
287 | return begin()[0]; |
288 | } |
289 | const_reference front() const { |
290 | assert(!empty())((void)0); |
291 | return begin()[0]; |
292 | } |
293 | |
294 | reference back() { |
295 | assert(!empty())((void)0); |
296 | return end()[-1]; |
297 | } |
298 | const_reference back() const { |
299 | assert(!empty())((void)0); |
300 | return end()[-1]; |
301 | } |
302 | }; |
303 | |
304 | /// SmallVectorTemplateBase<TriviallyCopyable = false> - This is where we put |
305 | /// method implementations that are designed to work with non-trivial T's. |
306 | /// |
307 | /// We approximate is_trivially_copyable with trivial move/copy construction and |
308 | /// trivial destruction. While the standard doesn't specify that you're allowed |
309 | /// copy these types with memcpy, there is no way for the type to observe this. |
310 | /// This catches the important case of std::pair<POD, POD>, which is not |
311 | /// trivially assignable. |
312 | template <typename T, bool = (is_trivially_copy_constructible<T>::value) && |
313 | (is_trivially_move_constructible<T>::value) && |
314 | std::is_trivially_destructible<T>::value> |
315 | class SmallVectorTemplateBase : public SmallVectorTemplateCommon<T> { |
316 | friend class SmallVectorTemplateCommon<T>; |
317 | |
318 | protected: |
319 | static constexpr bool TakesParamByValue = false; |
320 | using ValueParamT = const T &; |
321 | |
322 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
323 | |
324 | static void destroy_range(T *S, T *E) { |
325 | while (S != E) { |
326 | --E; |
327 | E->~T(); |
328 | } |
329 | } |
330 | |
331 | /// Move the range [I, E) into the uninitialized memory starting with "Dest", |
332 | /// constructing elements as needed. |
333 | template<typename It1, typename It2> |
334 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
335 | std::uninitialized_copy(std::make_move_iterator(I), |
336 | std::make_move_iterator(E), Dest); |
337 | } |
338 | |
339 | /// Copy the range [I, E) onto the uninitialized memory starting with "Dest", |
340 | /// constructing elements as needed. |
341 | template<typename It1, typename It2> |
342 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
343 | std::uninitialized_copy(I, E, Dest); |
344 | } |
345 | |
346 | /// Grow the allocated memory (without initializing new elements), doubling |
347 | /// the size of the allocated memory. Guarantees space for at least one more |
348 | /// element, or MinSize more elements if specified. |
349 | void grow(size_t MinSize = 0); |
350 | |
351 | /// Create a new allocation big enough for \p MinSize and pass back its size |
352 | /// in \p NewCapacity. This is the first section of \a grow(). |
353 | T *mallocForGrow(size_t MinSize, size_t &NewCapacity) { |
354 | return static_cast<T *>( |
355 | SmallVectorBase<SmallVectorSizeType<T>>::mallocForGrow( |
356 | MinSize, sizeof(T), NewCapacity)); |
357 | } |
358 | |
359 | /// Move existing elements over to the new allocation \p NewElts, the middle |
360 | /// section of \a grow(). |
361 | void moveElementsForGrow(T *NewElts); |
362 | |
363 | /// Transfer ownership of the allocation, finishing up \a grow(). |
364 | void takeAllocationForGrow(T *NewElts, size_t NewCapacity); |
365 | |
366 | /// Reserve enough space to add one element, and return the updated element |
367 | /// pointer in case it was a reference to the storage. |
368 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
369 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
370 | } |
371 | |
372 | /// Reserve enough space to add one element, and return the updated element |
373 | /// pointer in case it was a reference to the storage. |
374 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
375 | return const_cast<T *>( |
376 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
377 | } |
378 | |
379 | static T &&forward_value_param(T &&V) { return std::move(V); } |
380 | static const T &forward_value_param(const T &V) { return V; } |
381 | |
382 | void growAndAssign(size_t NumElts, const T &Elt) { |
383 | // Grow manually in case Elt is an internal reference. |
384 | size_t NewCapacity; |
385 | T *NewElts = mallocForGrow(NumElts, NewCapacity); |
386 | std::uninitialized_fill_n(NewElts, NumElts, Elt); |
387 | this->destroy_range(this->begin(), this->end()); |
388 | takeAllocationForGrow(NewElts, NewCapacity); |
389 | this->set_size(NumElts); |
390 | } |
391 | |
392 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
393 | // Grow manually in case one of Args is an internal reference. |
394 | size_t NewCapacity; |
395 | T *NewElts = mallocForGrow(0, NewCapacity); |
396 | ::new ((void *)(NewElts + this->size())) T(std::forward<ArgTypes>(Args)...); |
397 | moveElementsForGrow(NewElts); |
398 | takeAllocationForGrow(NewElts, NewCapacity); |
399 | this->set_size(this->size() + 1); |
400 | return this->back(); |
401 | } |
402 | |
403 | public: |
404 | void push_back(const T &Elt) { |
405 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
406 | ::new ((void *)this->end()) T(*EltPtr); |
407 | this->set_size(this->size() + 1); |
408 | } |
409 | |
410 | void push_back(T &&Elt) { |
411 | T *EltPtr = reserveForParamAndGetAddress(Elt); |
412 | ::new ((void *)this->end()) T(::std::move(*EltPtr)); |
413 | this->set_size(this->size() + 1); |
414 | } |
415 | |
416 | void pop_back() { |
417 | this->set_size(this->size() - 1); |
418 | this->end()->~T(); |
419 | } |
420 | }; |
421 | |
422 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
423 | template <typename T, bool TriviallyCopyable> |
424 | void SmallVectorTemplateBase<T, TriviallyCopyable>::grow(size_t MinSize) { |
425 | size_t NewCapacity; |
426 | T *NewElts = mallocForGrow(MinSize, NewCapacity); |
427 | moveElementsForGrow(NewElts); |
428 | takeAllocationForGrow(NewElts, NewCapacity); |
429 | } |
430 | |
431 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
432 | template <typename T, bool TriviallyCopyable> |
433 | void SmallVectorTemplateBase<T, TriviallyCopyable>::moveElementsForGrow( |
434 | T *NewElts) { |
435 | // Move the elements over. |
436 | this->uninitialized_move(this->begin(), this->end(), NewElts); |
437 | |
438 | // Destroy the original elements. |
439 | destroy_range(this->begin(), this->end()); |
440 | } |
441 | |
442 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
443 | template <typename T, bool TriviallyCopyable> |
444 | void SmallVectorTemplateBase<T, TriviallyCopyable>::takeAllocationForGrow( |
445 | T *NewElts, size_t NewCapacity) { |
446 | // If this wasn't grown from the inline copy, deallocate the old space. |
447 | if (!this->isSmall()) |
448 | free(this->begin()); |
449 | |
450 | this->BeginX = NewElts; |
451 | this->Capacity = NewCapacity; |
452 | } |
453 | |
454 | /// SmallVectorTemplateBase<TriviallyCopyable = true> - This is where we put |
455 | /// method implementations that are designed to work with trivially copyable |
456 | /// T's. This allows using memcpy in place of copy/move construction and |
457 | /// skipping destruction. |
458 | template <typename T> |
459 | class SmallVectorTemplateBase<T, true> : public SmallVectorTemplateCommon<T> { |
460 | friend class SmallVectorTemplateCommon<T>; |
461 | |
462 | protected: |
463 | /// True if it's cheap enough to take parameters by value. Doing so avoids |
464 | /// overhead related to mitigations for reference invalidation. |
465 | static constexpr bool TakesParamByValue = sizeof(T) <= 2 * sizeof(void *); |
466 | |
467 | /// Either const T& or T, depending on whether it's cheap enough to take |
468 | /// parameters by value. |
469 | using ValueParamT = |
470 | typename std::conditional<TakesParamByValue, T, const T &>::type; |
471 | |
472 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
473 | |
474 | // No need to do a destroy loop for POD's. |
475 | static void destroy_range(T *, T *) {} |
476 | |
477 | /// Move the range [I, E) onto the uninitialized memory |
478 | /// starting with "Dest", constructing elements into it as needed. |
479 | template<typename It1, typename It2> |
480 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
481 | // Just do a copy. |
482 | uninitialized_copy(I, E, Dest); |
483 | } |
484 | |
485 | /// Copy the range [I, E) onto the uninitialized memory |
486 | /// starting with "Dest", constructing elements into it as needed. |
487 | template<typename It1, typename It2> |
488 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
489 | // Arbitrary iterator types; just use the basic implementation. |
490 | std::uninitialized_copy(I, E, Dest); |
491 | } |
492 | |
493 | /// Copy the range [I, E) onto the uninitialized memory |
494 | /// starting with "Dest", constructing elements into it as needed. |
495 | template <typename T1, typename T2> |
496 | static void uninitialized_copy( |
497 | T1 *I, T1 *E, T2 *Dest, |
498 | std::enable_if_t<std::is_same<typename std::remove_const<T1>::type, |
499 | T2>::value> * = nullptr) { |
500 | // Use memcpy for PODs iterated by pointers (which includes SmallVector |
501 | // iterators): std::uninitialized_copy optimizes to memmove, but we can |
502 | // use memcpy here. Note that I and E are iterators and thus might be |
503 | // invalid for memcpy if they are equal. |
504 | if (I != E) |
505 | memcpy(reinterpret_cast<void *>(Dest), I, (E - I) * sizeof(T)); |
506 | } |
507 | |
508 | /// Double the size of the allocated memory, guaranteeing space for at |
509 | /// least one more element or MinSize if specified. |
510 | void grow(size_t MinSize = 0) { this->grow_pod(MinSize, sizeof(T)); } |
511 | |
512 | /// Reserve enough space to add one element, and return the updated element |
513 | /// pointer in case it was a reference to the storage. |
514 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
515 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
516 | } |
517 | |
518 | /// Reserve enough space to add one element, and return the updated element |
519 | /// pointer in case it was a reference to the storage. |
520 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
521 | return const_cast<T *>( |
522 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
523 | } |
524 | |
525 | /// Copy \p V or return a reference, depending on \a ValueParamT. |
526 | static ValueParamT forward_value_param(ValueParamT V) { return V; } |
527 | |
528 | void growAndAssign(size_t NumElts, T Elt) { |
529 | // Elt has been copied in case it's an internal reference, side-stepping |
530 | // reference invalidation problems without losing the realloc optimization. |
531 | this->set_size(0); |
532 | this->grow(NumElts); |
533 | std::uninitialized_fill_n(this->begin(), NumElts, Elt); |
534 | this->set_size(NumElts); |
535 | } |
536 | |
537 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
538 | // Use push_back with a copy in case Args has an internal reference, |
539 | // side-stepping reference invalidation problems without losing the realloc |
540 | // optimization. |
541 | push_back(T(std::forward<ArgTypes>(Args)...)); |
542 | return this->back(); |
543 | } |
544 | |
545 | public: |
546 | void push_back(ValueParamT Elt) { |
547 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
548 | memcpy(reinterpret_cast<void *>(this->end()), EltPtr, sizeof(T)); |
549 | this->set_size(this->size() + 1); |
550 | } |
551 | |
552 | void pop_back() { this->set_size(this->size() - 1); } |
553 | }; |
554 | |
555 | /// This class consists of common code factored out of the SmallVector class to |
556 | /// reduce code duplication based on the SmallVector 'N' template parameter. |
557 | template <typename T> |
558 | class SmallVectorImpl : public SmallVectorTemplateBase<T> { |
559 | using SuperClass = SmallVectorTemplateBase<T>; |
560 | |
561 | public: |
562 | using iterator = typename SuperClass::iterator; |
563 | using const_iterator = typename SuperClass::const_iterator; |
564 | using reference = typename SuperClass::reference; |
565 | using size_type = typename SuperClass::size_type; |
566 | |
567 | protected: |
568 | using SmallVectorTemplateBase<T>::TakesParamByValue; |
569 | using ValueParamT = typename SuperClass::ValueParamT; |
570 | |
571 | // Default ctor - Initialize to empty. |
572 | explicit SmallVectorImpl(unsigned N) |
573 | : SmallVectorTemplateBase<T>(N) {} |
574 | |
575 | public: |
576 | SmallVectorImpl(const SmallVectorImpl &) = delete; |
577 | |
578 | ~SmallVectorImpl() { |
579 | // Subclass has already destructed this vector's elements. |
580 | // If this wasn't grown from the inline copy, deallocate the old space. |
581 | if (!this->isSmall()) |
582 | free(this->begin()); |
583 | } |
584 | |
585 | void clear() { |
586 | this->destroy_range(this->begin(), this->end()); |
587 | this->Size = 0; |
588 | } |
589 | |
590 | private: |
591 | template <bool ForOverwrite> void resizeImpl(size_type N) { |
592 | if (N < this->size()) { |
593 | this->pop_back_n(this->size() - N); |
594 | } else if (N > this->size()) { |
595 | this->reserve(N); |
596 | for (auto I = this->end(), E = this->begin() + N; I != E; ++I) |
597 | if (ForOverwrite) |
598 | new (&*I) T; |
599 | else |
600 | new (&*I) T(); |
601 | this->set_size(N); |
602 | } |
603 | } |
604 | |
605 | public: |
606 | void resize(size_type N) { resizeImpl<false>(N); } |
607 | |
608 | /// Like resize, but \ref T is POD, the new values won't be initialized. |
609 | void resize_for_overwrite(size_type N) { resizeImpl<true>(N); } |
610 | |
611 | void resize(size_type N, ValueParamT NV) { |
612 | if (N == this->size()) |
613 | return; |
614 | |
615 | if (N < this->size()) { |
616 | this->pop_back_n(this->size() - N); |
617 | return; |
618 | } |
619 | |
620 | // N > this->size(). Defer to append. |
621 | this->append(N - this->size(), NV); |
622 | } |
623 | |
624 | void reserve(size_type N) { |
625 | if (this->capacity() < N) |
626 | this->grow(N); |
627 | } |
628 | |
629 | void pop_back_n(size_type NumItems) { |
630 | assert(this->size() >= NumItems)((void)0); |
631 | this->destroy_range(this->end() - NumItems, this->end()); |
632 | this->set_size(this->size() - NumItems); |
633 | } |
634 | |
635 | LLVM_NODISCARD[[clang::warn_unused_result]] T pop_back_val() { |
636 | T Result = ::std::move(this->back()); |
637 | this->pop_back(); |
638 | return Result; |
639 | } |
640 | |
641 | void swap(SmallVectorImpl &RHS); |
642 | |
643 | /// Add the specified range to the end of the SmallVector. |
644 | template <typename in_iter, |
645 | typename = std::enable_if_t<std::is_convertible< |
646 | typename std::iterator_traits<in_iter>::iterator_category, |
647 | std::input_iterator_tag>::value>> |
648 | void append(in_iter in_start, in_iter in_end) { |
649 | this->assertSafeToAddRange(in_start, in_end); |
650 | size_type NumInputs = std::distance(in_start, in_end); |
651 | this->reserve(this->size() + NumInputs); |
652 | this->uninitialized_copy(in_start, in_end, this->end()); |
653 | this->set_size(this->size() + NumInputs); |
654 | } |
655 | |
656 | /// Append \p NumInputs copies of \p Elt to the end. |
657 | void append(size_type NumInputs, ValueParamT Elt) { |
658 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumInputs); |
659 | std::uninitialized_fill_n(this->end(), NumInputs, *EltPtr); |
660 | this->set_size(this->size() + NumInputs); |
661 | } |
662 | |
663 | void append(std::initializer_list<T> IL) { |
664 | append(IL.begin(), IL.end()); |
665 | } |
666 | |
667 | void append(const SmallVectorImpl &RHS) { append(RHS.begin(), RHS.end()); } |
668 | |
669 | void assign(size_type NumElts, ValueParamT Elt) { |
670 | // Note that Elt could be an internal reference. |
671 | if (NumElts > this->capacity()) { |
672 | this->growAndAssign(NumElts, Elt); |
673 | return; |
674 | } |
675 | |
676 | // Assign over existing elements. |
677 | std::fill_n(this->begin(), std::min(NumElts, this->size()), Elt); |
678 | if (NumElts > this->size()) |
679 | std::uninitialized_fill_n(this->end(), NumElts - this->size(), Elt); |
680 | else if (NumElts < this->size()) |
681 | this->destroy_range(this->begin() + NumElts, this->end()); |
682 | this->set_size(NumElts); |
683 | } |
684 | |
685 | // FIXME: Consider assigning over existing elements, rather than clearing & |
686 | // re-initializing them - for all assign(...) variants. |
687 | |
688 | template <typename in_iter, |
689 | typename = std::enable_if_t<std::is_convertible< |
690 | typename std::iterator_traits<in_iter>::iterator_category, |
691 | std::input_iterator_tag>::value>> |
692 | void assign(in_iter in_start, in_iter in_end) { |
693 | this->assertSafeToReferenceAfterClear(in_start, in_end); |
694 | clear(); |
695 | append(in_start, in_end); |
696 | } |
697 | |
698 | void assign(std::initializer_list<T> IL) { |
699 | clear(); |
700 | append(IL); |
701 | } |
702 | |
703 | void assign(const SmallVectorImpl &RHS) { assign(RHS.begin(), RHS.end()); } |
704 | |
705 | iterator erase(const_iterator CI) { |
706 | // Just cast away constness because this is a non-const member function. |
707 | iterator I = const_cast<iterator>(CI); |
708 | |
709 | assert(this->isReferenceToStorage(CI) && "Iterator to erase is out of bounds.")((void)0); |
710 | |
711 | iterator N = I; |
712 | // Shift all elts down one. |
713 | std::move(I+1, this->end(), I); |
714 | // Drop the last elt. |
715 | this->pop_back(); |
716 | return(N); |
717 | } |
718 | |
719 | iterator erase(const_iterator CS, const_iterator CE) { |
720 | // Just cast away constness because this is a non-const member function. |
721 | iterator S = const_cast<iterator>(CS); |
722 | iterator E = const_cast<iterator>(CE); |
723 | |
724 | assert(this->isRangeInStorage(S, E) && "Range to erase is out of bounds.")((void)0); |
725 | |
726 | iterator N = S; |
727 | // Shift all elts down. |
728 | iterator I = std::move(E, this->end(), S); |
729 | // Drop the last elts. |
730 | this->destroy_range(I, this->end()); |
731 | this->set_size(I - this->begin()); |
732 | return(N); |
733 | } |
734 | |
735 | private: |
736 | template <class ArgType> iterator insert_one_impl(iterator I, ArgType &&Elt) { |
737 | // Callers ensure that ArgType is derived from T. |
738 | static_assert( |
739 | std::is_same<std::remove_const_t<std::remove_reference_t<ArgType>>, |
740 | T>::value, |
741 | "ArgType must be derived from T!"); |
742 | |
743 | if (I == this->end()) { // Important special case for empty vector. |
744 | this->push_back(::std::forward<ArgType>(Elt)); |
745 | return this->end()-1; |
746 | } |
747 | |
748 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
749 | |
750 | // Grow if necessary. |
751 | size_t Index = I - this->begin(); |
752 | std::remove_reference_t<ArgType> *EltPtr = |
753 | this->reserveForParamAndGetAddress(Elt); |
754 | I = this->begin() + Index; |
755 | |
756 | ::new ((void*) this->end()) T(::std::move(this->back())); |
757 | // Push everything else over. |
758 | std::move_backward(I, this->end()-1, this->end()); |
759 | this->set_size(this->size() + 1); |
760 | |
761 | // If we just moved the element we're inserting, be sure to update |
762 | // the reference (never happens if TakesParamByValue). |
763 | static_assert(!TakesParamByValue || std::is_same<ArgType, T>::value, |
764 | "ArgType must be 'T' when taking by value!"); |
765 | if (!TakesParamByValue && this->isReferenceToRange(EltPtr, I, this->end())) |
766 | ++EltPtr; |
767 | |
768 | *I = ::std::forward<ArgType>(*EltPtr); |
769 | return I; |
770 | } |
771 | |
772 | public: |
773 | iterator insert(iterator I, T &&Elt) { |
774 | return insert_one_impl(I, this->forward_value_param(std::move(Elt))); |
775 | } |
776 | |
777 | iterator insert(iterator I, const T &Elt) { |
778 | return insert_one_impl(I, this->forward_value_param(Elt)); |
779 | } |
780 | |
781 | iterator insert(iterator I, size_type NumToInsert, ValueParamT Elt) { |
782 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
783 | size_t InsertElt = I - this->begin(); |
784 | |
785 | if (I == this->end()) { // Important special case for empty vector. |
786 | append(NumToInsert, Elt); |
787 | return this->begin()+InsertElt; |
788 | } |
789 | |
790 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
791 | |
792 | // Ensure there is enough space, and get the (maybe updated) address of |
793 | // Elt. |
794 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumToInsert); |
795 | |
796 | // Uninvalidate the iterator. |
797 | I = this->begin()+InsertElt; |
798 | |
799 | // If there are more elements between the insertion point and the end of the |
800 | // range than there are being inserted, we can use a simple approach to |
801 | // insertion. Since we already reserved space, we know that this won't |
802 | // reallocate the vector. |
803 | if (size_t(this->end()-I) >= NumToInsert) { |
804 | T *OldEnd = this->end(); |
805 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
806 | std::move_iterator<iterator>(this->end())); |
807 | |
808 | // Copy the existing elements that get replaced. |
809 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
810 | |
811 | // If we just moved the element we're inserting, be sure to update |
812 | // the reference (never happens if TakesParamByValue). |
813 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
814 | EltPtr += NumToInsert; |
815 | |
816 | std::fill_n(I, NumToInsert, *EltPtr); |
817 | return I; |
818 | } |
819 | |
820 | // Otherwise, we're inserting more elements than exist already, and we're |
821 | // not inserting at the end. |
822 | |
823 | // Move over the elements that we're about to overwrite. |
824 | T *OldEnd = this->end(); |
825 | this->set_size(this->size() + NumToInsert); |
826 | size_t NumOverwritten = OldEnd-I; |
827 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
828 | |
829 | // If we just moved the element we're inserting, be sure to update |
830 | // the reference (never happens if TakesParamByValue). |
831 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
832 | EltPtr += NumToInsert; |
833 | |
834 | // Replace the overwritten part. |
835 | std::fill_n(I, NumOverwritten, *EltPtr); |
836 | |
837 | // Insert the non-overwritten middle part. |
838 | std::uninitialized_fill_n(OldEnd, NumToInsert - NumOverwritten, *EltPtr); |
839 | return I; |
840 | } |
841 | |
842 | template <typename ItTy, |
843 | typename = std::enable_if_t<std::is_convertible< |
844 | typename std::iterator_traits<ItTy>::iterator_category, |
845 | std::input_iterator_tag>::value>> |
846 | iterator insert(iterator I, ItTy From, ItTy To) { |
847 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
848 | size_t InsertElt = I - this->begin(); |
849 | |
850 | if (I == this->end()) { // Important special case for empty vector. |
851 | append(From, To); |
852 | return this->begin()+InsertElt; |
853 | } |
854 | |
855 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
856 | |
857 | // Check that the reserve that follows doesn't invalidate the iterators. |
858 | this->assertSafeToAddRange(From, To); |
859 | |
860 | size_t NumToInsert = std::distance(From, To); |
861 | |
862 | // Ensure there is enough space. |
863 | reserve(this->size() + NumToInsert); |
864 | |
865 | // Uninvalidate the iterator. |
866 | I = this->begin()+InsertElt; |
867 | |
868 | // If there are more elements between the insertion point and the end of the |
869 | // range than there are being inserted, we can use a simple approach to |
870 | // insertion. Since we already reserved space, we know that this won't |
871 | // reallocate the vector. |
872 | if (size_t(this->end()-I) >= NumToInsert) { |
873 | T *OldEnd = this->end(); |
874 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
875 | std::move_iterator<iterator>(this->end())); |
876 | |
877 | // Copy the existing elements that get replaced. |
878 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
879 | |
880 | std::copy(From, To, I); |
881 | return I; |
882 | } |
883 | |
884 | // Otherwise, we're inserting more elements than exist already, and we're |
885 | // not inserting at the end. |
886 | |
887 | // Move over the elements that we're about to overwrite. |
888 | T *OldEnd = this->end(); |
889 | this->set_size(this->size() + NumToInsert); |
890 | size_t NumOverwritten = OldEnd-I; |
891 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
892 | |
893 | // Replace the overwritten part. |
894 | for (T *J = I; NumOverwritten > 0; --NumOverwritten) { |
895 | *J = *From; |
896 | ++J; ++From; |
897 | } |
898 | |
899 | // Insert the non-overwritten middle part. |
900 | this->uninitialized_copy(From, To, OldEnd); |
901 | return I; |
902 | } |
903 | |
904 | void insert(iterator I, std::initializer_list<T> IL) { |
905 | insert(I, IL.begin(), IL.end()); |
906 | } |
907 | |
908 | template <typename... ArgTypes> reference emplace_back(ArgTypes &&... Args) { |
909 | if (LLVM_UNLIKELY(this->size() >= this->capacity())__builtin_expect((bool)(this->size() >= this->capacity ()), false)) |
910 | return this->growAndEmplaceBack(std::forward<ArgTypes>(Args)...); |
911 | |
912 | ::new ((void *)this->end()) T(std::forward<ArgTypes>(Args)...); |
913 | this->set_size(this->size() + 1); |
914 | return this->back(); |
915 | } |
916 | |
917 | SmallVectorImpl &operator=(const SmallVectorImpl &RHS); |
918 | |
919 | SmallVectorImpl &operator=(SmallVectorImpl &&RHS); |
920 | |
921 | bool operator==(const SmallVectorImpl &RHS) const { |
922 | if (this->size() != RHS.size()) return false; |
923 | return std::equal(this->begin(), this->end(), RHS.begin()); |
924 | } |
925 | bool operator!=(const SmallVectorImpl &RHS) const { |
926 | return !(*this == RHS); |
927 | } |
928 | |
929 | bool operator<(const SmallVectorImpl &RHS) const { |
930 | return std::lexicographical_compare(this->begin(), this->end(), |
931 | RHS.begin(), RHS.end()); |
932 | } |
933 | }; |
934 | |
935 | template <typename T> |
936 | void SmallVectorImpl<T>::swap(SmallVectorImpl<T> &RHS) { |
937 | if (this == &RHS) return; |
938 | |
939 | // We can only avoid copying elements if neither vector is small. |
940 | if (!this->isSmall() && !RHS.isSmall()) { |
941 | std::swap(this->BeginX, RHS.BeginX); |
942 | std::swap(this->Size, RHS.Size); |
943 | std::swap(this->Capacity, RHS.Capacity); |
944 | return; |
945 | } |
946 | this->reserve(RHS.size()); |
947 | RHS.reserve(this->size()); |
948 | |
949 | // Swap the shared elements. |
950 | size_t NumShared = this->size(); |
951 | if (NumShared > RHS.size()) NumShared = RHS.size(); |
952 | for (size_type i = 0; i != NumShared; ++i) |
953 | std::swap((*this)[i], RHS[i]); |
954 | |
955 | // Copy over the extra elts. |
956 | if (this->size() > RHS.size()) { |
957 | size_t EltDiff = this->size() - RHS.size(); |
958 | this->uninitialized_copy(this->begin()+NumShared, this->end(), RHS.end()); |
959 | RHS.set_size(RHS.size() + EltDiff); |
960 | this->destroy_range(this->begin()+NumShared, this->end()); |
961 | this->set_size(NumShared); |
962 | } else if (RHS.size() > this->size()) { |
963 | size_t EltDiff = RHS.size() - this->size(); |
964 | this->uninitialized_copy(RHS.begin()+NumShared, RHS.end(), this->end()); |
965 | this->set_size(this->size() + EltDiff); |
966 | this->destroy_range(RHS.begin()+NumShared, RHS.end()); |
967 | RHS.set_size(NumShared); |
968 | } |
969 | } |
970 | |
971 | template <typename T> |
972 | SmallVectorImpl<T> &SmallVectorImpl<T>:: |
973 | operator=(const SmallVectorImpl<T> &RHS) { |
974 | // Avoid self-assignment. |
975 | if (this == &RHS) return *this; |
976 | |
977 | // If we already have sufficient space, assign the common elements, then |
978 | // destroy any excess. |
979 | size_t RHSSize = RHS.size(); |
980 | size_t CurSize = this->size(); |
981 | if (CurSize >= RHSSize) { |
982 | // Assign common elements. |
983 | iterator NewEnd; |
984 | if (RHSSize) |
985 | NewEnd = std::copy(RHS.begin(), RHS.begin()+RHSSize, this->begin()); |
986 | else |
987 | NewEnd = this->begin(); |
988 | |
989 | // Destroy excess elements. |
990 | this->destroy_range(NewEnd, this->end()); |
991 | |
992 | // Trim. |
993 | this->set_size(RHSSize); |
994 | return *this; |
995 | } |
996 | |
997 | // If we have to grow to have enough elements, destroy the current elements. |
998 | // This allows us to avoid copying them during the grow. |
999 | // FIXME: don't do this if they're efficiently moveable. |
1000 | if (this->capacity() < RHSSize) { |
1001 | // Destroy current elements. |
1002 | this->clear(); |
1003 | CurSize = 0; |
1004 | this->grow(RHSSize); |
1005 | } else if (CurSize) { |
1006 | // Otherwise, use assignment for the already-constructed elements. |
1007 | std::copy(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
1008 | } |
1009 | |
1010 | // Copy construct the new elements in place. |
1011 | this->uninitialized_copy(RHS.begin()+CurSize, RHS.end(), |
1012 | this->begin()+CurSize); |
1013 | |
1014 | // Set end. |
1015 | this->set_size(RHSSize); |
1016 | return *this; |
1017 | } |
1018 | |
1019 | template <typename T> |
1020 | SmallVectorImpl<T> &SmallVectorImpl<T>::operator=(SmallVectorImpl<T> &&RHS) { |
1021 | // Avoid self-assignment. |
1022 | if (this == &RHS) return *this; |
1023 | |
1024 | // If the RHS isn't small, clear this vector and then steal its buffer. |
1025 | if (!RHS.isSmall()) { |
1026 | this->destroy_range(this->begin(), this->end()); |
1027 | if (!this->isSmall()) free(this->begin()); |
1028 | this->BeginX = RHS.BeginX; |
1029 | this->Size = RHS.Size; |
1030 | this->Capacity = RHS.Capacity; |
1031 | RHS.resetToSmall(); |
1032 | return *this; |
1033 | } |
1034 | |
1035 | // If we already have sufficient space, assign the common elements, then |
1036 | // destroy any excess. |
1037 | size_t RHSSize = RHS.size(); |
1038 | size_t CurSize = this->size(); |
1039 | if (CurSize >= RHSSize) { |
1040 | // Assign common elements. |
1041 | iterator NewEnd = this->begin(); |
1042 | if (RHSSize) |
1043 | NewEnd = std::move(RHS.begin(), RHS.end(), NewEnd); |
1044 | |
1045 | // Destroy excess elements and trim the bounds. |
1046 | this->destroy_range(NewEnd, this->end()); |
1047 | this->set_size(RHSSize); |
1048 | |
1049 | // Clear the RHS. |
1050 | RHS.clear(); |
1051 | |
1052 | return *this; |
1053 | } |
1054 | |
1055 | // If we have to grow to have enough elements, destroy the current elements. |
1056 | // This allows us to avoid copying them during the grow. |
1057 | // FIXME: this may not actually make any sense if we can efficiently move |
1058 | // elements. |
1059 | if (this->capacity() < RHSSize) { |
1060 | // Destroy current elements. |
1061 | this->clear(); |
1062 | CurSize = 0; |
1063 | this->grow(RHSSize); |
1064 | } else if (CurSize) { |
1065 | // Otherwise, use assignment for the already-constructed elements. |
1066 | std::move(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
1067 | } |
1068 | |
1069 | // Move-construct the new elements in place. |
1070 | this->uninitialized_move(RHS.begin()+CurSize, RHS.end(), |
1071 | this->begin()+CurSize); |
1072 | |
1073 | // Set end. |
1074 | this->set_size(RHSSize); |
1075 | |
1076 | RHS.clear(); |
1077 | return *this; |
1078 | } |
1079 | |
1080 | /// Storage for the SmallVector elements. This is specialized for the N=0 case |
1081 | /// to avoid allocating unnecessary storage. |
1082 | template <typename T, unsigned N> |
1083 | struct SmallVectorStorage { |
1084 | alignas(T) char InlineElts[N * sizeof(T)]; |
1085 | }; |
1086 | |
1087 | /// We need the storage to be properly aligned even for small-size of 0 so that |
1088 | /// the pointer math in \a SmallVectorTemplateCommon::getFirstEl() is |
1089 | /// well-defined. |
1090 | template <typename T> struct alignas(T) SmallVectorStorage<T, 0> {}; |
1091 | |
1092 | /// Forward declaration of SmallVector so that |
1093 | /// calculateSmallVectorDefaultInlinedElements can reference |
1094 | /// `sizeof(SmallVector<T, 0>)`. |
1095 | template <typename T, unsigned N> class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector; |
1096 | |
1097 | /// Helper class for calculating the default number of inline elements for |
1098 | /// `SmallVector<T>`. |
1099 | /// |
1100 | /// This should be migrated to a constexpr function when our minimum |
1101 | /// compiler support is enough for multi-statement constexpr functions. |
1102 | template <typename T> struct CalculateSmallVectorDefaultInlinedElements { |
1103 | // Parameter controlling the default number of inlined elements |
1104 | // for `SmallVector<T>`. |
1105 | // |
1106 | // The default number of inlined elements ensures that |
1107 | // 1. There is at least one inlined element. |
1108 | // 2. `sizeof(SmallVector<T>) <= kPreferredSmallVectorSizeof` unless |
1109 | // it contradicts 1. |
1110 | static constexpr size_t kPreferredSmallVectorSizeof = 64; |
1111 | |
1112 | // static_assert that sizeof(T) is not "too big". |
1113 | // |
1114 | // Because our policy guarantees at least one inlined element, it is possible |
1115 | // for an arbitrarily large inlined element to allocate an arbitrarily large |
1116 | // amount of inline storage. We generally consider it an antipattern for a |
1117 | // SmallVector to allocate an excessive amount of inline storage, so we want |
1118 | // to call attention to these cases and make sure that users are making an |
1119 | // intentional decision if they request a lot of inline storage. |
1120 | // |
1121 | // We want this assertion to trigger in pathological cases, but otherwise |
1122 | // not be too easy to hit. To accomplish that, the cutoff is actually somewhat |
1123 | // larger than kPreferredSmallVectorSizeof (otherwise, |
1124 | // `SmallVector<SmallVector<T>>` would be one easy way to trip it, and that |
1125 | // pattern seems useful in practice). |
1126 | // |
1127 | // One wrinkle is that this assertion is in theory non-portable, since |
1128 | // sizeof(T) is in general platform-dependent. However, we don't expect this |
1129 | // to be much of an issue, because most LLVM development happens on 64-bit |
1130 | // hosts, and therefore sizeof(T) is expected to *decrease* when compiled for |
1131 | // 32-bit hosts, dodging the issue. The reverse situation, where development |
1132 | // happens on a 32-bit host and then fails due to sizeof(T) *increasing* on a |
1133 | // 64-bit host, is expected to be very rare. |
1134 | static_assert( |
1135 | sizeof(T) <= 256, |
1136 | "You are trying to use a default number of inlined elements for " |
1137 | "`SmallVector<T>` but `sizeof(T)` is really big! Please use an " |
1138 | "explicit number of inlined elements with `SmallVector<T, N>` to make " |
1139 | "sure you really want that much inline storage."); |
1140 | |
1141 | // Discount the size of the header itself when calculating the maximum inline |
1142 | // bytes. |
1143 | static constexpr size_t PreferredInlineBytes = |
1144 | kPreferredSmallVectorSizeof - sizeof(SmallVector<T, 0>); |
1145 | static constexpr size_t NumElementsThatFit = PreferredInlineBytes / sizeof(T); |
1146 | static constexpr size_t value = |
1147 | NumElementsThatFit == 0 ? 1 : NumElementsThatFit; |
1148 | }; |
1149 | |
1150 | /// This is a 'vector' (really, a variable-sized array), optimized |
1151 | /// for the case when the array is small. It contains some number of elements |
1152 | /// in-place, which allows it to avoid heap allocation when the actual number of |
1153 | /// elements is below that threshold. This allows normal "small" cases to be |
1154 | /// fast without losing generality for large inputs. |
1155 | /// |
1156 | /// \note |
1157 | /// In the absence of a well-motivated choice for the number of inlined |
1158 | /// elements \p N, it is recommended to use \c SmallVector<T> (that is, |
1159 | /// omitting the \p N). This will choose a default number of inlined elements |
1160 | /// reasonable for allocation on the stack (for example, trying to keep \c |
1161 | /// sizeof(SmallVector<T>) around 64 bytes). |
1162 | /// |
1163 | /// \warning This does not attempt to be exception safe. |
1164 | /// |
1165 | /// \see https://llvm.org/docs/ProgrammersManual.html#llvm-adt-smallvector-h |
1166 | template <typename T, |
1167 | unsigned N = CalculateSmallVectorDefaultInlinedElements<T>::value> |
1168 | class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector : public SmallVectorImpl<T>, |
1169 | SmallVectorStorage<T, N> { |
1170 | public: |
1171 | SmallVector() : SmallVectorImpl<T>(N) {} |
1172 | |
1173 | ~SmallVector() { |
1174 | // Destroy the constructed elements in the vector. |
1175 | this->destroy_range(this->begin(), this->end()); |
1176 | } |
1177 | |
1178 | explicit SmallVector(size_t Size, const T &Value = T()) |
1179 | : SmallVectorImpl<T>(N) { |
1180 | this->assign(Size, Value); |
1181 | } |
1182 | |
1183 | template <typename ItTy, |
1184 | typename = std::enable_if_t<std::is_convertible< |
1185 | typename std::iterator_traits<ItTy>::iterator_category, |
1186 | std::input_iterator_tag>::value>> |
1187 | SmallVector(ItTy S, ItTy E) : SmallVectorImpl<T>(N) { |
1188 | this->append(S, E); |
1189 | } |
1190 | |
1191 | template <typename RangeTy> |
1192 | explicit SmallVector(const iterator_range<RangeTy> &R) |
1193 | : SmallVectorImpl<T>(N) { |
1194 | this->append(R.begin(), R.end()); |
1195 | } |
1196 | |
1197 | SmallVector(std::initializer_list<T> IL) : SmallVectorImpl<T>(N) { |
1198 | this->assign(IL); |
1199 | } |
1200 | |
1201 | SmallVector(const SmallVector &RHS) : SmallVectorImpl<T>(N) { |
1202 | if (!RHS.empty()) |
1203 | SmallVectorImpl<T>::operator=(RHS); |
1204 | } |
1205 | |
1206 | SmallVector &operator=(const SmallVector &RHS) { |
1207 | SmallVectorImpl<T>::operator=(RHS); |
1208 | return *this; |
1209 | } |
1210 | |
1211 | SmallVector(SmallVector &&RHS) : SmallVectorImpl<T>(N) { |
1212 | if (!RHS.empty()) |
1213 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1214 | } |
1215 | |
1216 | SmallVector(SmallVectorImpl<T> &&RHS) : SmallVectorImpl<T>(N) { |
1217 | if (!RHS.empty()) |
1218 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1219 | } |
1220 | |
1221 | SmallVector &operator=(SmallVector &&RHS) { |
1222 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1223 | return *this; |
1224 | } |
1225 | |
1226 | SmallVector &operator=(SmallVectorImpl<T> &&RHS) { |
1227 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
1228 | return *this; |
1229 | } |
1230 | |
1231 | SmallVector &operator=(std::initializer_list<T> IL) { |
1232 | this->assign(IL); |
1233 | return *this; |
1234 | } |
1235 | }; |
1236 | |
1237 | template <typename T, unsigned N> |
1238 | inline size_t capacity_in_bytes(const SmallVector<T, N> &X) { |
1239 | return X.capacity_in_bytes(); |
1240 | } |
1241 | |
1242 | /// Given a range of type R, iterate the entire range and return a |
1243 | /// SmallVector with elements of the vector. This is useful, for example, |
1244 | /// when you want to iterate a range and then sort the results. |
1245 | template <unsigned Size, typename R> |
1246 | SmallVector<typename std::remove_const<typename std::remove_reference< |
1247 | decltype(*std::begin(std::declval<R &>()))>::type>::type, |
1248 | Size> |
1249 | to_vector(R &&Range) { |
1250 | return {std::begin(Range), std::end(Range)}; |
1251 | } |
1252 | |
1253 | } // end namespace llvm |
1254 | |
1255 | namespace std { |
1256 | |
1257 | /// Implement std::swap in terms of SmallVector swap. |
1258 | template<typename T> |
1259 | inline void |
1260 | swap(llvm::SmallVectorImpl<T> &LHS, llvm::SmallVectorImpl<T> &RHS) { |
1261 | LHS.swap(RHS); |
1262 | } |
1263 | |
1264 | /// Implement std::swap in terms of SmallVector swap. |
1265 | template<typename T, unsigned N> |
1266 | inline void |
1267 | swap(llvm::SmallVector<T, N> &LHS, llvm::SmallVector<T, N> &RHS) { |
1268 | LHS.swap(RHS); |
1269 | } |
1270 | |
1271 | } // end namespace std |
1272 | |
1273 | #endif // LLVM_ADT_SMALLVECTOR_H |