1 //===- CallGraphSort.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 /// The file is responsible for sorting sections using LLVM call graph profile
10 /// data by placing frequently executed code sections together. The goal of the
11 /// placement is to improve the runtime performance of the final executable by
12 /// arranging code sections so that i-TLB misses and i-cache misses are reduced.
13 ///
14 /// The algorithm first builds a call graph based on the profile data and then
15 /// iteratively merges "chains" (ordered lists) of input sections which will be
16 /// laid out as a unit. There are two implementations for deciding how to
17 /// merge a pair of chains:
18 ///  - a simpler one, referred to as Call-Chain Clustering (C^3), that follows
19 ///    "Optimizing Function Placement for Large-Scale Data-Center Applications"
20 /// https://research.fb.com/wp-content/uploads/2017/01/cgo2017-hfsort-final1.pdf
21 /// - a more advanced one, referred to as Cache-Directed-Sort (CDSort), which
22 ///   typically produces layouts with higher locality, and hence, yields fewer
23 ///   instruction cache misses on large binaries.
24 //===----------------------------------------------------------------------===//
25 
26 #include "CallGraphSort.h"
27 #include "InputFiles.h"
28 #include "InputSection.h"
29 #include "Symbols.h"
30 #include "llvm/Support/FileSystem.h"
31 #include "llvm/Transforms/Utils/CodeLayout.h"
32 
33 #include <numeric>
34 
35 using namespace llvm;
36 using namespace lld;
37 using namespace lld::elf;
38 
39 namespace {
40 struct Edge {
41   int from;
42   uint64_t weight;
43 };
44 
45 struct Cluster {
46   Cluster(int sec, size_t s) : next(sec), prev(sec), size(s) {}
47 
48   double getDensity() const {
49     if (size == 0)
50       return 0;
51     return double(weight) / double(size);
52   }
53 
54   int next;
55   int prev;
56   uint64_t size;
57   uint64_t weight = 0;
58   uint64_t initialWeight = 0;
59   Edge bestPred = {-1, 0};
60 };
61 
62 /// Implementation of the Call-Chain Clustering (C^3). The goal of this
63 /// algorithm is to improve runtime performance of the executable by arranging
64 /// code sections such that page table and i-cache misses are minimized.
65 ///
66 /// Definitions:
67 /// * Cluster
68 ///   * An ordered list of input sections which are laid out as a unit. At the
69 ///     beginning of the algorithm each input section has its own cluster and
70 ///     the weight of the cluster is the sum of the weight of all incoming
71 ///     edges.
72 /// * Call-Chain Clustering (C³) Heuristic
73 ///   * Defines when and how clusters are combined. Pick the highest weighted
74 ///     input section then add it to its most likely predecessor if it wouldn't
75 ///     penalize it too much.
76 /// * Density
77 ///   * The weight of the cluster divided by the size of the cluster. This is a
78 ///     proxy for the amount of execution time spent per byte of the cluster.
79 ///
80 /// It does so given a call graph profile by the following:
81 /// * Build a weighted call graph from the call graph profile
82 /// * Sort input sections by weight
83 /// * For each input section starting with the highest weight
84 ///   * Find its most likely predecessor cluster
85 ///   * Check if the combined cluster would be too large, or would have too low
86 ///     a density.
87 ///   * If not, then combine the clusters.
88 /// * Sort non-empty clusters by density
89 class CallGraphSort {
90 public:
91   CallGraphSort();
92 
93   DenseMap<const InputSectionBase *, int> run();
94 
95 private:
96   std::vector<Cluster> clusters;
97   std::vector<const InputSectionBase *> sections;
98 };
99 
100 // Maximum amount the combined cluster density can be worse than the original
101 // cluster to consider merging.
102 constexpr int MAX_DENSITY_DEGRADATION = 8;
103 
104 // Maximum cluster size in bytes.
105 constexpr uint64_t MAX_CLUSTER_SIZE = 1024 * 1024;
106 } // end anonymous namespace
107 
108 using SectionPair =
109     std::pair<const InputSectionBase *, const InputSectionBase *>;
110 
111 // Take the edge list in Config->CallGraphProfile, resolve symbol names to
112 // Symbols, and generate a graph between InputSections with the provided
113 // weights.
114 CallGraphSort::CallGraphSort() {
115   MapVector<SectionPair, uint64_t> &profile = config->callGraphProfile;
116   DenseMap<const InputSectionBase *, int> secToCluster;
117 
118   auto getOrCreateNode = [&](const InputSectionBase *isec) -> int {
119     auto res = secToCluster.try_emplace(isec, clusters.size());
120     if (res.second) {
121       sections.push_back(isec);
122       clusters.emplace_back(clusters.size(), isec->getSize());
123     }
124     return res.first->second;
125   };
126 
127   // Create the graph.
128   for (std::pair<SectionPair, uint64_t> &c : profile) {
129     const auto *fromSB = cast<InputSectionBase>(c.first.first);
130     const auto *toSB = cast<InputSectionBase>(c.first.second);
131     uint64_t weight = c.second;
132 
133     // Ignore edges between input sections belonging to different output
134     // sections.  This is done because otherwise we would end up with clusters
135     // containing input sections that can't actually be placed adjacently in the
136     // output.  This messes with the cluster size and density calculations.  We
137     // would also end up moving input sections in other output sections without
138     // moving them closer to what calls them.
139     if (fromSB->getOutputSection() != toSB->getOutputSection())
140       continue;
141 
142     int from = getOrCreateNode(fromSB);
143     int to = getOrCreateNode(toSB);
144 
145     clusters[to].weight += weight;
146 
147     if (from == to)
148       continue;
149 
150     // Remember the best edge.
151     Cluster &toC = clusters[to];
152     if (toC.bestPred.from == -1 || toC.bestPred.weight < weight) {
153       toC.bestPred.from = from;
154       toC.bestPred.weight = weight;
155     }
156   }
157   for (Cluster &c : clusters)
158     c.initialWeight = c.weight;
159 }
160 
161 // It's bad to merge clusters which would degrade the density too much.
162 static bool isNewDensityBad(Cluster &a, Cluster &b) {
163   double newDensity = double(a.weight + b.weight) / double(a.size + b.size);
164   return newDensity < a.getDensity() / MAX_DENSITY_DEGRADATION;
165 }
166 
167 // Find the leader of V's belonged cluster (represented as an equivalence
168 // class). We apply union-find path-halving technique (simple to implement) in
169 // the meantime as it decreases depths and the time complexity.
170 static int getLeader(int *leaders, int v) {
171   while (leaders[v] != v) {
172     leaders[v] = leaders[leaders[v]];
173     v = leaders[v];
174   }
175   return v;
176 }
177 
178 static void mergeClusters(std::vector<Cluster> &cs, Cluster &into, int intoIdx,
179                           Cluster &from, int fromIdx) {
180   int tail1 = into.prev, tail2 = from.prev;
181   into.prev = tail2;
182   cs[tail2].next = intoIdx;
183   from.prev = tail1;
184   cs[tail1].next = fromIdx;
185   into.size += from.size;
186   into.weight += from.weight;
187   from.size = 0;
188   from.weight = 0;
189 }
190 
191 // Group InputSections into clusters using the Call-Chain Clustering heuristic
192 // then sort the clusters by density.
193 DenseMap<const InputSectionBase *, int> CallGraphSort::run() {
194   std::vector<int> sorted(clusters.size());
195   std::unique_ptr<int[]> leaders(new int[clusters.size()]);
196 
197   std::iota(leaders.get(), leaders.get() + clusters.size(), 0);
198   std::iota(sorted.begin(), sorted.end(), 0);
199   llvm::stable_sort(sorted, [&](int a, int b) {
200     return clusters[a].getDensity() > clusters[b].getDensity();
201   });
202 
203   for (int l : sorted) {
204     // The cluster index is the same as the index of its leader here because
205     // clusters[L] has not been merged into another cluster yet.
206     Cluster &c = clusters[l];
207 
208     // Don't consider merging if the edge is unlikely.
209     if (c.bestPred.from == -1 || c.bestPred.weight * 10 <= c.initialWeight)
210       continue;
211 
212     int predL = getLeader(leaders.get(), c.bestPred.from);
213     if (l == predL)
214       continue;
215 
216     Cluster *predC = &clusters[predL];
217     if (c.size + predC->size > MAX_CLUSTER_SIZE)
218       continue;
219 
220     if (isNewDensityBad(*predC, c))
221       continue;
222 
223     leaders[l] = predL;
224     mergeClusters(clusters, *predC, predL, c, l);
225   }
226 
227   // Sort remaining non-empty clusters by density.
228   sorted.clear();
229   for (int i = 0, e = (int)clusters.size(); i != e; ++i)
230     if (clusters[i].size > 0)
231       sorted.push_back(i);
232   llvm::stable_sort(sorted, [&](int a, int b) {
233     return clusters[a].getDensity() > clusters[b].getDensity();
234   });
235 
236   DenseMap<const InputSectionBase *, int> orderMap;
237   int curOrder = 1;
238   for (int leader : sorted) {
239     for (int i = leader;;) {
240       orderMap[sections[i]] = curOrder++;
241       i = clusters[i].next;
242       if (i == leader)
243         break;
244     }
245   }
246   if (!config->printSymbolOrder.empty()) {
247     std::error_code ec;
248     raw_fd_ostream os(config->printSymbolOrder, ec, sys::fs::OF_None);
249     if (ec) {
250       error("cannot open " + config->printSymbolOrder + ": " + ec.message());
251       return orderMap;
252     }
253 
254     // Print the symbols ordered by C3, in the order of increasing curOrder
255     // Instead of sorting all the orderMap, just repeat the loops above.
256     for (int leader : sorted)
257       for (int i = leader;;) {
258         // Search all the symbols in the file of the section
259         // and find out a Defined symbol with name that is within the section.
260         for (Symbol *sym : sections[i]->file->getSymbols())
261           if (!sym->isSection()) // Filter out section-type symbols here.
262             if (auto *d = dyn_cast<Defined>(sym))
263               if (sections[i] == d->section)
264                 os << sym->getName() << "\n";
265         i = clusters[i].next;
266         if (i == leader)
267           break;
268       }
269   }
270 
271   return orderMap;
272 }
273 
274 // Sort sections by the profile data using the Cache-Directed Sort algorithm.
275 // The placement is done by optimizing the locality by co-locating frequently
276 // executed code sections together.
277 DenseMap<const InputSectionBase *, int> elf::computeCacheDirectedSortOrder() {
278   SmallVector<uint64_t, 0> funcSizes;
279   SmallVector<uint64_t, 0> funcCounts;
280   SmallVector<codelayout::EdgeCount, 0> callCounts;
281   SmallVector<uint64_t, 0> callOffsets;
282   SmallVector<const InputSectionBase *, 0> sections;
283   DenseMap<const InputSectionBase *, size_t> secToTargetId;
284 
285   auto getOrCreateNode = [&](const InputSectionBase *inSec) -> size_t {
286     auto res = secToTargetId.try_emplace(inSec, sections.size());
287     if (res.second) {
288       // inSec does not appear before in the graph.
289       sections.push_back(inSec);
290       funcSizes.push_back(inSec->getSize());
291       funcCounts.push_back(0);
292     }
293     return res.first->second;
294   };
295 
296   // Create the graph.
297   for (std::pair<SectionPair, uint64_t> &c : config->callGraphProfile) {
298     const InputSectionBase *fromSB = cast<InputSectionBase>(c.first.first);
299     const InputSectionBase *toSB = cast<InputSectionBase>(c.first.second);
300     // Ignore edges between input sections belonging to different sections.
301     if (fromSB->getOutputSection() != toSB->getOutputSection())
302       continue;
303 
304     uint64_t weight = c.second;
305     // Ignore edges with zero weight.
306     if (weight == 0)
307       continue;
308 
309     size_t from = getOrCreateNode(fromSB);
310     size_t to = getOrCreateNode(toSB);
311     // Ignore self-edges (recursive calls).
312     if (from == to)
313       continue;
314 
315     callCounts.push_back({from, to, weight});
316     // Assume that the jump is at the middle of the input section. The profile
317     // data does not contain jump offsets.
318     callOffsets.push_back((funcSizes[from] + 1) / 2);
319     funcCounts[to] += weight;
320   }
321 
322   // Run the layout algorithm.
323   std::vector<uint64_t> sortedSections = codelayout::computeCacheDirectedLayout(
324       funcSizes, funcCounts, callCounts, callOffsets);
325 
326   // Create the final order.
327   DenseMap<const InputSectionBase *, int> orderMap;
328   int curOrder = 1;
329   for (uint64_t secIdx : sortedSections)
330     orderMap[sections[secIdx]] = curOrder++;
331 
332   return orderMap;
333 }
334 
335 // Sort sections by the profile data provided by --callgraph-profile-file.
336 //
337 // This first builds a call graph based on the profile data then merges sections
338 // according either to the C³ or Cache-Directed-Sort ordering algorithm.
339 DenseMap<const InputSectionBase *, int> elf::computeCallGraphProfileOrder() {
340   if (config->callGraphProfileSort == CGProfileSortKind::Cdsort)
341     return computeCacheDirectedSortOrder();
342   return CallGraphSort().run();
343 }
344