1 //===- bolt/Passes/ReorderFunctions.cpp - Function reordering pass --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements ReorderFunctions class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "bolt/Passes/ReorderFunctions.h"
14 #include "bolt/Passes/HFSort.h"
15 #include "llvm/Support/CommandLine.h"
16 #include <fstream>
17 
18 #define DEBUG_TYPE "hfsort"
19 
20 using namespace llvm;
21 
22 namespace opts {
23 
24 extern cl::OptionCategory BoltOptCategory;
25 extern cl::opt<unsigned> Verbosity;
26 extern cl::opt<uint32_t> RandomSeed;
27 
28 extern size_t padFunction(const bolt::BinaryFunction &Function);
29 
30 cl::opt<bolt::ReorderFunctions::ReorderType>
31 ReorderFunctions("reorder-functions",
32   cl::desc("reorder and cluster functions (works only with relocations)"),
33   cl::init(bolt::ReorderFunctions::RT_NONE),
34   cl::values(clEnumValN(bolt::ReorderFunctions::RT_NONE,
35       "none",
36       "do not reorder functions"),
37     clEnumValN(bolt::ReorderFunctions::RT_EXEC_COUNT,
38       "exec-count",
39       "order by execution count"),
40     clEnumValN(bolt::ReorderFunctions::RT_HFSORT,
41       "hfsort",
42       "use hfsort algorithm"),
43     clEnumValN(bolt::ReorderFunctions::RT_HFSORT_PLUS,
44       "hfsort+",
45       "use hfsort+ algorithm"),
46     clEnumValN(bolt::ReorderFunctions::RT_PETTIS_HANSEN,
47       "pettis-hansen",
48       "use Pettis-Hansen algorithm"),
49     clEnumValN(bolt::ReorderFunctions::RT_RANDOM,
50       "random",
51       "reorder functions randomly"),
52     clEnumValN(bolt::ReorderFunctions::RT_USER,
53       "user",
54       "use function order specified by -function-order")),
55   cl::ZeroOrMore,
56   cl::cat(BoltOptCategory));
57 
58 static cl::opt<bool> ReorderFunctionsUseHotSize(
59     "reorder-functions-use-hot-size",
60     cl::desc("use a function's hot size when doing clustering"), cl::init(true),
61     cl::cat(BoltOptCategory));
62 
63 static cl::opt<std::string>
64 FunctionOrderFile("function-order",
65   cl::desc("file containing an ordered list of functions to use for function "
66            "reordering"),
67   cl::cat(BoltOptCategory));
68 
69 static cl::opt<std::string>
70 GenerateFunctionOrderFile("generate-function-order",
71   cl::desc("file to dump the ordered list of functions to use for function "
72            "reordering"),
73   cl::cat(BoltOptCategory));
74 
75 static cl::opt<std::string>
76 LinkSectionsFile("generate-link-sections",
77   cl::desc("generate a list of function sections in a format suitable for "
78            "inclusion in a linker script"),
79   cl::cat(BoltOptCategory));
80 
81 static cl::opt<bool>
82     UseEdgeCounts("use-edge-counts",
83                   cl::desc("use edge count data when doing clustering"),
84                   cl::init(true), cl::cat(BoltOptCategory));
85 
86 static cl::opt<bool>
87 CgFromPerfData("cg-from-perf-data",
88   cl::desc("use perf data directly when constructing the call graph"
89            " for stale functions"),
90   cl::init(true),
91   cl::ZeroOrMore,
92   cl::cat(BoltOptCategory));
93 
94 static cl::opt<bool> CgIgnoreRecursiveCalls(
95     "cg-ignore-recursive-calls",
96     cl::desc("ignore recursive calls when constructing the call graph"),
97     cl::init(true), cl::cat(BoltOptCategory));
98 
99 static cl::opt<bool>
100 CgUseSplitHotSize("cg-use-split-hot-size",
101   cl::desc("use hot/cold data on basic blocks to determine hot sizes for "
102            "call graph functions"),
103   cl::init(false),
104   cl::ZeroOrMore,
105   cl::cat(BoltOptCategory));
106 
107 } // namespace opts
108 
109 namespace llvm {
110 namespace bolt {
111 
112 using NodeId = CallGraph::NodeId;
113 using Arc = CallGraph::Arc;
114 using Node = CallGraph::Node;
115 
116 void ReorderFunctions::reorder(std::vector<Cluster> &&Clusters,
117                                std::map<uint64_t, BinaryFunction> &BFs) {
118   std::vector<uint64_t> FuncAddr(Cg.numNodes()); // Just for computing stats
119   uint64_t TotalSize = 0;
120   uint32_t Index = 0;
121 
122   // Set order of hot functions based on clusters.
123   for (const Cluster &Cluster : Clusters) {
124     for (const NodeId FuncId : Cluster.targets()) {
125       Cg.nodeIdToFunc(FuncId)->setIndex(Index++);
126       FuncAddr[FuncId] = TotalSize;
127       TotalSize += Cg.size(FuncId);
128     }
129   }
130 
131   if (opts::ReorderFunctions == RT_NONE)
132     return;
133 
134   if (opts::Verbosity == 0) {
135 #ifndef NDEBUG
136     if (!DebugFlag || !isCurrentDebugType("hfsort"))
137       return;
138 #else
139     return;
140 #endif
141   }
142 
143   bool PrintDetailed = opts::Verbosity > 1;
144 #ifndef NDEBUG
145   PrintDetailed |=
146     (DebugFlag && isCurrentDebugType("hfsort") && opts::Verbosity > 0);
147 #endif
148   TotalSize   = 0;
149   uint64_t CurPage     = 0;
150   uint64_t Hotfuncs    = 0;
151   double TotalDistance = 0;
152   double TotalCalls    = 0;
153   double TotalCalls64B = 0;
154   double TotalCalls4KB = 0;
155   double TotalCalls2MB = 0;
156   if (PrintDetailed)
157     outs() << "BOLT-INFO: Function reordering page layout\n"
158            << "BOLT-INFO: ============== page 0 ==============\n";
159   for (Cluster &Cluster : Clusters) {
160     if (PrintDetailed)
161       outs() << format(
162           "BOLT-INFO: -------- density = %.3lf (%u / %u) --------\n",
163           Cluster.density(), Cluster.samples(), Cluster.size());
164 
165     for (NodeId FuncId : Cluster.targets()) {
166       if (Cg.samples(FuncId) > 0) {
167         Hotfuncs++;
168 
169         if (PrintDetailed)
170           outs() << "BOLT-INFO: hot func " << *Cg.nodeIdToFunc(FuncId) << " ("
171                  << Cg.size(FuncId) << ")\n";
172 
173         uint64_t Dist = 0;
174         uint64_t Calls = 0;
175         for (NodeId Dst : Cg.successors(FuncId)) {
176           if (FuncId == Dst) // ignore recursive calls in stats
177             continue;
178           const Arc &Arc = *Cg.findArc(FuncId, Dst);
179           const auto D = std::abs(FuncAddr[Arc.dst()] -
180                                   (FuncAddr[FuncId] + Arc.avgCallOffset()));
181           const double W = Arc.weight();
182           if (D < 64 && PrintDetailed && opts::Verbosity > 2)
183             outs() << "BOLT-INFO: short (" << D << "B) call:\n"
184                    << "BOLT-INFO:   Src: " << *Cg.nodeIdToFunc(FuncId) << "\n"
185                    << "BOLT-INFO:   Dst: " << *Cg.nodeIdToFunc(Dst) << "\n"
186                    << "BOLT-INFO:   Weight = " << W << "\n"
187                    << "BOLT-INFO:   AvgOffset = " << Arc.avgCallOffset() << "\n";
188           Calls += W;
189           if (D < 64)        TotalCalls64B += W;
190           if (D < 4096)      TotalCalls4KB += W;
191           if (D < (2 << 20)) TotalCalls2MB += W;
192           Dist += Arc.weight() * D;
193           if (PrintDetailed)
194             outs() << format("BOLT-INFO: arc: %u [@%lu+%.1lf] -> %u [@%lu]: "
195                              "weight = %.0lf, callDist = %f\n",
196                              Arc.src(),
197                              FuncAddr[Arc.src()],
198                              Arc.avgCallOffset(),
199                              Arc.dst(),
200                              FuncAddr[Arc.dst()],
201                              Arc.weight(), D);
202         }
203         TotalCalls += Calls;
204         TotalDistance += Dist;
205         TotalSize += Cg.size(FuncId);
206 
207         if (PrintDetailed) {
208           outs() << format("BOLT-INFO: start = %6u : avgCallDist = %lu : ",
209                            TotalSize, Calls ? Dist / Calls : 0)
210                  << Cg.nodeIdToFunc(FuncId)->getPrintName() << '\n';
211           const uint64_t NewPage = TotalSize / HugePageSize;
212           if (NewPage != CurPage) {
213             CurPage = NewPage;
214             outs() << format(
215                 "BOLT-INFO: ============== page %u ==============\n", CurPage);
216           }
217         }
218       }
219     }
220   }
221   outs() << "BOLT-INFO: Function reordering stats\n"
222          << format("BOLT-INFO:  Number of hot functions: %u\n"
223                    "BOLT-INFO:  Number of clusters: %lu\n",
224                    Hotfuncs, Clusters.size())
225          << format("BOLT-INFO:  Final average call distance = %.1lf "
226                    "(%.0lf / %.0lf)\n",
227                    TotalCalls ? TotalDistance / TotalCalls : 0, TotalDistance,
228                    TotalCalls)
229          << format("BOLT-INFO:  Total Calls = %.0lf\n", TotalCalls);
230   if (TotalCalls)
231     outs() << format("BOLT-INFO:  Total Calls within 64B = %.0lf (%.2lf%%)\n",
232                      TotalCalls64B, 100 * TotalCalls64B / TotalCalls)
233            << format("BOLT-INFO:  Total Calls within 4KB = %.0lf (%.2lf%%)\n",
234                      TotalCalls4KB, 100 * TotalCalls4KB / TotalCalls)
235            << format("BOLT-INFO:  Total Calls within 2MB = %.0lf (%.2lf%%)\n",
236                      TotalCalls2MB, 100 * TotalCalls2MB / TotalCalls);
237 }
238 
239 namespace {
240 
241 std::vector<std::string> readFunctionOrderFile() {
242   std::vector<std::string> FunctionNames;
243   std::ifstream FuncsFile(opts::FunctionOrderFile, std::ios::in);
244   if (!FuncsFile) {
245     errs() << "Ordered functions file \"" << opts::FunctionOrderFile
246            << "\" can't be opened.\n";
247     exit(1);
248   }
249   std::string FuncName;
250   while (std::getline(FuncsFile, FuncName))
251     FunctionNames.push_back(FuncName);
252   return FunctionNames;
253 }
254 
255 }
256 
257 void ReorderFunctions::runOnFunctions(BinaryContext &BC) {
258   auto &BFs = BC.getBinaryFunctions();
259   if (opts::ReorderFunctions != RT_NONE &&
260       opts::ReorderFunctions != RT_EXEC_COUNT &&
261       opts::ReorderFunctions != RT_USER) {
262     Cg = buildCallGraph(BC,
263                         [](const BinaryFunction &BF) {
264                           if (!BF.hasProfile())
265                             return true;
266                           if (BF.getState() != BinaryFunction::State::CFG)
267                             return true;
268                           return false;
269                         },
270                         opts::CgFromPerfData,
271                         false, // IncludeColdCalls
272                         opts::ReorderFunctionsUseHotSize,
273                         opts::CgUseSplitHotSize,
274                         opts::UseEdgeCounts,
275                         opts::CgIgnoreRecursiveCalls);
276     Cg.normalizeArcWeights();
277   }
278 
279   std::vector<Cluster> Clusters;
280 
281   switch (opts::ReorderFunctions) {
282   case RT_NONE:
283     break;
284   case RT_EXEC_COUNT:
285     {
286       std::vector<BinaryFunction *> SortedFunctions(BFs.size());
287       uint32_t Index = 0;
288       std::transform(BFs.begin(),
289                      BFs.end(),
290                      SortedFunctions.begin(),
291                      [](std::pair<const uint64_t, BinaryFunction> &BFI) {
292                        return &BFI.second;
293                      });
294       std::stable_sort(SortedFunctions.begin(), SortedFunctions.end(),
295                        [&](const BinaryFunction *A, const BinaryFunction *B) {
296                          if (A->isIgnored())
297                            return false;
298                          const size_t PadA = opts::padFunction(*A);
299                          const size_t PadB = opts::padFunction(*B);
300                          if (!PadA || !PadB) {
301                            if (PadA)
302                              return true;
303                            if (PadB)
304                              return false;
305                          }
306                          return !A->hasProfile() &&
307                            (B->hasProfile() ||
308                             (A->getExecutionCount() > B->getExecutionCount()));
309                        });
310       for (BinaryFunction *BF : SortedFunctions)
311         if (BF->hasProfile())
312           BF->setIndex(Index++);
313     }
314     break;
315   case RT_HFSORT:
316     Clusters = clusterize(Cg);
317     break;
318   case RT_HFSORT_PLUS:
319     Clusters = hfsortPlus(Cg);
320     break;
321   case RT_PETTIS_HANSEN:
322     Clusters = pettisAndHansen(Cg);
323     break;
324   case RT_RANDOM:
325     std::srand(opts::RandomSeed);
326     Clusters = randomClusters(Cg);
327     break;
328   case RT_USER:
329     {
330       uint32_t Index = 0;
331       for (const std::string &Function : readFunctionOrderFile()) {
332         std::vector<uint64_t> FuncAddrs;
333 
334         BinaryData *BD = BC.getBinaryDataByName(Function);
335         if (!BD) {
336           uint32_t LocalID = 1;
337           while(1) {
338             // If we can't find the main symbol name, look for alternates.
339             const std::string FuncName =
340                 Function + "/" + std::to_string(LocalID);
341             BD = BC.getBinaryDataByName(FuncName);
342             if (BD)
343               FuncAddrs.push_back(BD->getAddress());
344             else
345               break;
346             LocalID++;
347           }
348         } else {
349           FuncAddrs.push_back(BD->getAddress());
350         }
351 
352         if (FuncAddrs.empty()) {
353           errs() << "BOLT-WARNING: Reorder functions: can't find function for "
354                  << Function << ".\n";
355           continue;
356         }
357 
358         for (const uint64_t FuncAddr : FuncAddrs) {
359           const BinaryData *FuncBD = BC.getBinaryDataAtAddress(FuncAddr);
360           assert(FuncBD);
361 
362           BinaryFunction *BF = BC.getFunctionForSymbol(FuncBD->getSymbol());
363           if (!BF) {
364             errs() << "BOLT-WARNING: Reorder functions: can't find function for "
365                    << Function << ".\n";
366             break;
367           }
368           if (!BF->hasValidIndex())
369             BF->setIndex(Index++);
370           else if (opts::Verbosity > 0)
371             errs() << "BOLT-WARNING: Duplicate reorder entry for " << Function
372                    << ".\n";
373         }
374       }
375     }
376     break;
377   }
378 
379   reorder(std::move(Clusters), BFs);
380 
381   std::unique_ptr<std::ofstream> FuncsFile;
382   if (!opts::GenerateFunctionOrderFile.empty()) {
383     FuncsFile = std::make_unique<std::ofstream>(opts::GenerateFunctionOrderFile,
384                                                 std::ios::out);
385     if (!FuncsFile) {
386       errs() << "BOLT-ERROR: ordered functions file "
387              << opts::GenerateFunctionOrderFile << " cannot be opened\n";
388       exit(1);
389     }
390   }
391 
392   std::unique_ptr<std::ofstream> LinkSectionsFile;
393   if (!opts::LinkSectionsFile.empty()) {
394     LinkSectionsFile =
395         std::make_unique<std::ofstream>(opts::LinkSectionsFile, std::ios::out);
396     if (!LinkSectionsFile) {
397       errs() << "BOLT-ERROR: link sections file " << opts::LinkSectionsFile
398              << " cannot be opened\n";
399       exit(1);
400     }
401   }
402 
403   if (FuncsFile || LinkSectionsFile) {
404     std::vector<BinaryFunction *> SortedFunctions(BFs.size());
405     std::transform(BFs.begin(), BFs.end(), SortedFunctions.begin(),
406                    [](std::pair<const uint64_t, BinaryFunction> &BFI) {
407                      return &BFI.second;
408                    });
409 
410     // Sort functions by index.
411     std::stable_sort(
412       SortedFunctions.begin(),
413       SortedFunctions.end(),
414       [](const BinaryFunction *A, const BinaryFunction *B) {
415         if (A->hasValidIndex() && B->hasValidIndex())
416           return A->getIndex() < B->getIndex();
417         if (A->hasValidIndex() && !B->hasValidIndex())
418           return true;
419         if (!A->hasValidIndex() && B->hasValidIndex())
420           return false;
421         return A->getAddress() < B->getAddress();
422       });
423 
424     for (const BinaryFunction *Func : SortedFunctions) {
425       if (!Func->hasValidIndex())
426         break;
427       if (Func->isPLTFunction())
428         continue;
429 
430       if (FuncsFile)
431         *FuncsFile << Func->getOneName().str() << '\n';
432 
433       if (LinkSectionsFile) {
434         const char *Indent = "";
435         std::vector<StringRef> AllNames = Func->getNames();
436         std::sort(AllNames.begin(), AllNames.end());
437         for (StringRef Name : AllNames) {
438           const size_t SlashPos = Name.find('/');
439           if (SlashPos != std::string::npos) {
440             // Avoid duplicates for local functions.
441             if (Name.find('/', SlashPos + 1) != std::string::npos)
442               continue;
443             Name = Name.substr(0, SlashPos);
444           }
445           *LinkSectionsFile << Indent << ".text." << Name.str() << '\n';
446           Indent = " ";
447         }
448       }
449     }
450 
451     if (FuncsFile) {
452       FuncsFile->close();
453       outs() << "BOLT-INFO: dumped function order to "
454              << opts::GenerateFunctionOrderFile << '\n';
455     }
456 
457     if (LinkSectionsFile) {
458       LinkSectionsFile->close();
459       outs() << "BOLT-INFO: dumped linker section order to "
460              << opts::LinkSectionsFile << '\n';
461     }
462   }
463 }
464 
465 } // namespace bolt
466 } // namespace llvm
467