xref: /freebsd/contrib/llvm-project/lld/MachO/SectionPriorities.cpp (revision 525fe93dc7487a1e63a90f6a2b956abc601963c1)
1 //===- SectionPriorities.cpp ----------------------------------------------===//
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 is based on the ELF port, see ELF/CallGraphSort.cpp for the details
10 /// about the algorithm.
11 ///
12 //===----------------------------------------------------------------------===//
13 
14 #include "SectionPriorities.h"
15 #include "Config.h"
16 #include "InputFiles.h"
17 #include "Symbols.h"
18 #include "Target.h"
19 
20 #include "lld/Common/Args.h"
21 #include "lld/Common/CommonLinkerContext.h"
22 #include "lld/Common/ErrorHandler.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/MapVector.h"
25 #include "llvm/Support/Path.h"
26 #include "llvm/Support/TimeProfiler.h"
27 #include "llvm/Support/raw_ostream.h"
28 
29 #include <numeric>
30 
31 using namespace llvm;
32 using namespace llvm::MachO;
33 using namespace llvm::sys;
34 using namespace lld;
35 using namespace lld::macho;
36 
37 PriorityBuilder macho::priorityBuilder;
38 
39 namespace {
40 
41 size_t highestAvailablePriority = std::numeric_limits<size_t>::max();
42 
43 struct Edge {
44   int from;
45   uint64_t weight;
46 };
47 
48 struct Cluster {
49   Cluster(int sec, size_t s) : next(sec), prev(sec), size(s) {}
50 
51   double getDensity() const {
52     if (size == 0)
53       return 0;
54     return double(weight) / double(size);
55   }
56 
57   int next;
58   int prev;
59   uint64_t size;
60   uint64_t weight = 0;
61   uint64_t initialWeight = 0;
62   Edge bestPred = {-1, 0};
63 };
64 
65 class CallGraphSort {
66 public:
67   CallGraphSort(const MapVector<SectionPair, uint64_t> &profile);
68 
69   DenseMap<const InputSection *, size_t> run();
70 
71 private:
72   std::vector<Cluster> clusters;
73   std::vector<const InputSection *> sections;
74 };
75 // Maximum amount the combined cluster density can be worse than the original
76 // cluster to consider merging.
77 constexpr int MAX_DENSITY_DEGRADATION = 8;
78 } // end anonymous namespace
79 
80 // Take the edge list in callGraphProfile, resolve symbol names to Symbols, and
81 // generate a graph between InputSections with the provided weights.
82 CallGraphSort::CallGraphSort(const MapVector<SectionPair, uint64_t> &profile) {
83   DenseMap<const InputSection *, int> secToCluster;
84 
85   auto getOrCreateCluster = [&](const InputSection *isec) -> int {
86     auto res = secToCluster.try_emplace(isec, clusters.size());
87     if (res.second) {
88       sections.push_back(isec);
89       clusters.emplace_back(clusters.size(), isec->getSize());
90     }
91     return res.first->second;
92   };
93 
94   // Create the graph
95   for (const std::pair<SectionPair, uint64_t> &c : profile) {
96     const auto fromSec = c.first.first->canonical();
97     const auto toSec = c.first.second->canonical();
98     uint64_t weight = c.second;
99     // Ignore edges between input sections belonging to different output
100     // sections.  This is done because otherwise we would end up with clusters
101     // containing input sections that can't actually be placed adjacently in the
102     // output.  This messes with the cluster size and density calculations.  We
103     // would also end up moving input sections in other output sections without
104     // moving them closer to what calls them.
105     if (fromSec->parent != toSec->parent)
106       continue;
107 
108     int from = getOrCreateCluster(fromSec);
109     int to = getOrCreateCluster(toSec);
110 
111     clusters[to].weight += weight;
112 
113     if (from == to)
114       continue;
115 
116     // Remember the best edge.
117     Cluster &toC = clusters[to];
118     if (toC.bestPred.from == -1 || toC.bestPred.weight < weight) {
119       toC.bestPred.from = from;
120       toC.bestPred.weight = weight;
121     }
122   }
123   for (Cluster &c : clusters)
124     c.initialWeight = c.weight;
125 }
126 
127 // It's bad to merge clusters which would degrade the density too much.
128 static bool isNewDensityBad(Cluster &a, Cluster &b) {
129   double newDensity = double(a.weight + b.weight) / double(a.size + b.size);
130   return newDensity < a.getDensity() / MAX_DENSITY_DEGRADATION;
131 }
132 
133 // Find the leader of V's belonged cluster (represented as an equivalence
134 // class). We apply union-find path-halving technique (simple to implement) in
135 // the meantime as it decreases depths and the time complexity.
136 static int getLeader(std::vector<int> &leaders, int v) {
137   while (leaders[v] != v) {
138     leaders[v] = leaders[leaders[v]];
139     v = leaders[v];
140   }
141   return v;
142 }
143 
144 static void mergeClusters(std::vector<Cluster> &cs, Cluster &into, int intoIdx,
145                           Cluster &from, int fromIdx) {
146   int tail1 = into.prev, tail2 = from.prev;
147   into.prev = tail2;
148   cs[tail2].next = intoIdx;
149   from.prev = tail1;
150   cs[tail1].next = fromIdx;
151   into.size += from.size;
152   into.weight += from.weight;
153   from.size = 0;
154   from.weight = 0;
155 }
156 
157 // Group InputSections into clusters using the Call-Chain Clustering heuristic
158 // then sort the clusters by density.
159 DenseMap<const InputSection *, size_t> CallGraphSort::run() {
160   const uint64_t maxClusterSize = target->getPageSize();
161 
162   // Cluster indices sorted by density.
163   std::vector<int> sorted(clusters.size());
164   // For union-find.
165   std::vector<int> leaders(clusters.size());
166 
167   std::iota(leaders.begin(), leaders.end(), 0);
168   std::iota(sorted.begin(), sorted.end(), 0);
169 
170   llvm::stable_sort(sorted, [&](int a, int b) {
171     return clusters[a].getDensity() > clusters[b].getDensity();
172   });
173 
174   for (int l : sorted) {
175     // The cluster index is the same as the index of its leader here because
176     // clusters[L] has not been merged into another cluster yet.
177     Cluster &c = clusters[l];
178 
179     // Don't consider merging if the edge is unlikely.
180     if (c.bestPred.from == -1 || c.bestPred.weight * 10 <= c.initialWeight)
181       continue;
182 
183     int predL = getLeader(leaders, c.bestPred.from);
184     // Already in the same cluster.
185     if (l == predL)
186       continue;
187 
188     Cluster *predC = &clusters[predL];
189     if (c.size + predC->size > maxClusterSize)
190       continue;
191 
192     if (isNewDensityBad(*predC, c))
193       continue;
194 
195     leaders[l] = predL;
196     mergeClusters(clusters, *predC, predL, c, l);
197   }
198   // Sort remaining non-empty clusters by density.
199   sorted.clear();
200   for (int i = 0, e = (int)clusters.size(); i != e; ++i)
201     if (clusters[i].size > 0)
202       sorted.push_back(i);
203   llvm::stable_sort(sorted, [&](int a, int b) {
204     return clusters[a].getDensity() > clusters[b].getDensity();
205   });
206 
207   DenseMap<const InputSection *, size_t> orderMap;
208 
209   // Sections will be sorted by decreasing order. Absent sections will have
210   // priority 0 and be placed at the end of sections.
211   // NB: This is opposite from COFF/ELF to be compatible with the existing
212   // order-file code.
213   int curOrder = highestAvailablePriority;
214   for (int leader : sorted) {
215     for (int i = leader;;) {
216       orderMap[sections[i]] = curOrder--;
217       i = clusters[i].next;
218       if (i == leader)
219         break;
220     }
221   }
222   if (!config->printSymbolOrder.empty()) {
223     std::error_code ec;
224     raw_fd_ostream os(config->printSymbolOrder, ec, sys::fs::OF_None);
225     if (ec) {
226       error("cannot open " + config->printSymbolOrder + ": " + ec.message());
227       return orderMap;
228     }
229     // Print the symbols ordered by C3, in the order of decreasing curOrder
230     // Instead of sorting all the orderMap, just repeat the loops above.
231     for (int leader : sorted)
232       for (int i = leader;;) {
233         const InputSection *isec = sections[i];
234         // Search all the symbols in the file of the section
235         // and find out a Defined symbol with name that is within the
236         // section.
237         for (Symbol *sym : isec->getFile()->symbols) {
238           if (auto *d = dyn_cast_or_null<Defined>(sym)) {
239             if (d->isec == isec)
240               os << sym->getName() << "\n";
241           }
242         }
243         i = clusters[i].next;
244         if (i == leader)
245           break;
246       }
247   }
248 
249   return orderMap;
250 }
251 
252 std::optional<size_t>
253 macho::PriorityBuilder::getSymbolPriority(const Defined *sym) {
254   if (sym->isAbsolute())
255     return std::nullopt;
256 
257   auto it = priorities.find(sym->getName());
258   if (it == priorities.end())
259     return std::nullopt;
260   const SymbolPriorityEntry &entry = it->second;
261   const InputFile *f = sym->isec->getFile();
262   if (!f)
263     return entry.anyObjectFile;
264   // We don't use toString(InputFile *) here because it returns the full path
265   // for object files, and we only want the basename.
266   StringRef filename;
267   if (f->archiveName.empty())
268     filename = path::filename(f->getName());
269   else
270     filename = saver().save(path::filename(f->archiveName) + "(" +
271                             path::filename(f->getName()) + ")");
272   return std::max(entry.objectFiles.lookup(filename), entry.anyObjectFile);
273 }
274 
275 void macho::PriorityBuilder::extractCallGraphProfile() {
276   TimeTraceScope timeScope("Extract call graph profile");
277   bool hasOrderFile = !priorities.empty();
278   for (const InputFile *file : inputFiles) {
279     auto *obj = dyn_cast_or_null<ObjFile>(file);
280     if (!obj)
281       continue;
282     for (const CallGraphEntry &entry : obj->callGraph) {
283       assert(entry.fromIndex < obj->symbols.size() &&
284              entry.toIndex < obj->symbols.size());
285       auto *fromSym = dyn_cast_or_null<Defined>(obj->symbols[entry.fromIndex]);
286       auto *toSym = dyn_cast_or_null<Defined>(obj->symbols[entry.toIndex]);
287       if (fromSym && toSym &&
288           (!hasOrderFile ||
289            (!getSymbolPriority(fromSym) && !getSymbolPriority(toSym))))
290         callGraphProfile[{fromSym->isec, toSym->isec}] += entry.count;
291     }
292   }
293 }
294 
295 void macho::PriorityBuilder::parseOrderFile(StringRef path) {
296   assert(callGraphProfile.empty() &&
297          "Order file must be parsed before call graph profile is processed");
298   std::optional<MemoryBufferRef> buffer = readFile(path);
299   if (!buffer) {
300     error("Could not read order file at " + path);
301     return;
302   }
303 
304   MemoryBufferRef mbref = *buffer;
305   for (StringRef line : args::getLines(mbref)) {
306     StringRef objectFile, symbol;
307     line = line.take_until([](char c) { return c == '#'; }); // ignore comments
308     line = line.ltrim();
309 
310     CPUType cpuType = StringSwitch<CPUType>(line)
311                           .StartsWith("i386:", CPU_TYPE_I386)
312                           .StartsWith("x86_64:", CPU_TYPE_X86_64)
313                           .StartsWith("arm:", CPU_TYPE_ARM)
314                           .StartsWith("arm64:", CPU_TYPE_ARM64)
315                           .StartsWith("ppc:", CPU_TYPE_POWERPC)
316                           .StartsWith("ppc64:", CPU_TYPE_POWERPC64)
317                           .Default(CPU_TYPE_ANY);
318 
319     if (cpuType != CPU_TYPE_ANY && cpuType != target->cpuType)
320       continue;
321 
322     // Drop the CPU type as well as the colon
323     if (cpuType != CPU_TYPE_ANY)
324       line = line.drop_until([](char c) { return c == ':'; }).drop_front();
325 
326     constexpr std::array<StringRef, 2> fileEnds = {".o:", ".o):"};
327     for (StringRef fileEnd : fileEnds) {
328       size_t pos = line.find(fileEnd);
329       if (pos != StringRef::npos) {
330         // Split the string around the colon
331         objectFile = line.take_front(pos + fileEnd.size() - 1);
332         line = line.drop_front(pos + fileEnd.size());
333         break;
334       }
335     }
336     symbol = line.trim();
337 
338     if (!symbol.empty()) {
339       SymbolPriorityEntry &entry = priorities[symbol];
340       if (!objectFile.empty())
341         entry.objectFiles.insert(
342             std::make_pair(objectFile, highestAvailablePriority));
343       else
344         entry.anyObjectFile =
345             std::max(entry.anyObjectFile, highestAvailablePriority);
346     }
347 
348     --highestAvailablePriority;
349   }
350 }
351 
352 DenseMap<const InputSection *, size_t>
353 macho::PriorityBuilder::buildInputSectionPriorities() {
354   DenseMap<const InputSection *, size_t> sectionPriorities;
355   if (config->callGraphProfileSort) {
356     // Sort sections by the profile data provided by __LLVM,__cg_profile
357     // sections.
358     //
359     // This first builds a call graph based on the profile data then merges
360     // sections according to the C³ heuristic. All clusters are then sorted by a
361     // density metric to further improve locality.
362     TimeTraceScope timeScope("Call graph profile sort");
363     sectionPriorities = CallGraphSort(callGraphProfile).run();
364   }
365 
366   if (priorities.empty())
367     return sectionPriorities;
368 
369   auto addSym = [&](const Defined *sym) {
370     std::optional<size_t> symbolPriority = getSymbolPriority(sym);
371     if (!symbolPriority)
372       return;
373     size_t &priority = sectionPriorities[sym->isec];
374     priority = std::max(priority, *symbolPriority);
375   };
376 
377   // TODO: Make sure this handles weak symbols correctly.
378   for (const InputFile *file : inputFiles) {
379     if (isa<ObjFile>(file))
380       for (Symbol *sym : file->symbols)
381         if (auto *d = dyn_cast_or_null<Defined>(sym))
382           addSym(d);
383   }
384 
385   return sectionPriorities;
386 }
387