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 /// Implementation of Call-Chain Clustering from: Optimizing Function Placement
10 /// for Large-Scale Data-Center Applications
11 /// https://research.fb.com/wp-content/uploads/2017/01/cgo2017-hfsort-final1.pdf
12 ///
13 /// The goal of this algorithm is to improve runtime performance of the final
14 /// executable by arranging code sections such that page table and i-cache
15 /// misses are minimized.
16 ///
17 /// Definitions:
18 /// * Cluster
19 /// * An ordered list of input sections which are laid out as a unit. At the
20 /// beginning of the algorithm each input section has its own cluster and
21 /// the weight of the cluster is the sum of the weight of all incoming
22 /// edges.
23 /// * Call-Chain Clustering (C³) Heuristic
24 /// * Defines when and how clusters are combined. Pick the highest weighted
25 /// input section then add it to its most likely predecessor if it wouldn't
26 /// penalize it too much.
27 /// * Density
28 /// * The weight of the cluster divided by the size of the cluster. This is a
29 /// proxy for the amount of execution time spent per byte of the cluster.
30 ///
31 /// It does so given a call graph profile by the following:
32 /// * Build a weighted call graph from the call graph profile
33 /// * Sort input sections by weight
34 /// * For each input section starting with the highest weight
35 /// * Find its most likely predecessor cluster
36 /// * Check if the combined cluster would be too large, or would have too low
37 /// a density.
38 /// * If not, then combine the clusters.
39 /// * Sort non-empty clusters by density
40 ///
41 //===----------------------------------------------------------------------===//
42
43 #include "CallGraphSort.h"
44 #include "OutputSections.h"
45 #include "SymbolTable.h"
46 #include "Symbols.h"
47
48 #include <numeric>
49
50 using namespace llvm;
51 using namespace lld;
52 using namespace lld::elf;
53
54 namespace {
55 struct Edge {
56 int from;
57 uint64_t weight;
58 };
59
60 struct Cluster {
Cluster__anon724cdc3f0111::Cluster61 Cluster(int sec, size_t s) : next(sec), prev(sec), size(s) {}
62
getDensity__anon724cdc3f0111::Cluster63 double getDensity() const {
64 if (size == 0)
65 return 0;
66 return double(weight) / double(size);
67 }
68
69 int next;
70 int prev;
71 uint64_t size;
72 uint64_t weight = 0;
73 uint64_t initialWeight = 0;
74 Edge bestPred = {-1, 0};
75 };
76
77 class CallGraphSort {
78 public:
79 CallGraphSort();
80
81 DenseMap<const InputSectionBase *, int> run();
82
83 private:
84 std::vector<Cluster> clusters;
85 std::vector<const InputSectionBase *> sections;
86 };
87
88 // Maximum amount the combined cluster density can be worse than the original
89 // cluster to consider merging.
90 constexpr int MAX_DENSITY_DEGRADATION = 8;
91
92 // Maximum cluster size in bytes.
93 constexpr uint64_t MAX_CLUSTER_SIZE = 1024 * 1024;
94 } // end anonymous namespace
95
96 using SectionPair =
97 std::pair<const InputSectionBase *, const InputSectionBase *>;
98
99 // Take the edge list in Config->CallGraphProfile, resolve symbol names to
100 // Symbols, and generate a graph between InputSections with the provided
101 // weights.
CallGraphSort()102 CallGraphSort::CallGraphSort() {
103 MapVector<SectionPair, uint64_t> &profile = config->callGraphProfile;
104 DenseMap<const InputSectionBase *, int> secToCluster;
105
106 auto getOrCreateNode = [&](const InputSectionBase *isec) -> int {
107 auto res = secToCluster.try_emplace(isec, clusters.size());
108 if (res.second) {
109 sections.push_back(isec);
110 clusters.emplace_back(clusters.size(), isec->getSize());
111 }
112 return res.first->second;
113 };
114
115 // Create the graph.
116 for (std::pair<SectionPair, uint64_t> &c : profile) {
117 const auto *fromSB = cast<InputSectionBase>(c.first.first->repl);
118 const auto *toSB = cast<InputSectionBase>(c.first.second->repl);
119 uint64_t weight = c.second;
120
121 // Ignore edges between input sections belonging to different output
122 // sections. This is done because otherwise we would end up with clusters
123 // containing input sections that can't actually be placed adjacently in the
124 // output. This messes with the cluster size and density calculations. We
125 // would also end up moving input sections in other output sections without
126 // moving them closer to what calls them.
127 if (fromSB->getOutputSection() != toSB->getOutputSection())
128 continue;
129
130 int from = getOrCreateNode(fromSB);
131 int to = getOrCreateNode(toSB);
132
133 clusters[to].weight += weight;
134
135 if (from == to)
136 continue;
137
138 // Remember the best edge.
139 Cluster &toC = clusters[to];
140 if (toC.bestPred.from == -1 || toC.bestPred.weight < weight) {
141 toC.bestPred.from = from;
142 toC.bestPred.weight = weight;
143 }
144 }
145 for (Cluster &c : clusters)
146 c.initialWeight = c.weight;
147 }
148
149 // It's bad to merge clusters which would degrade the density too much.
isNewDensityBad(Cluster & a,Cluster & b)150 static bool isNewDensityBad(Cluster &a, Cluster &b) {
151 double newDensity = double(a.weight + b.weight) / double(a.size + b.size);
152 return newDensity < a.getDensity() / MAX_DENSITY_DEGRADATION;
153 }
154
155 // Find the leader of V's belonged cluster (represented as an equivalence
156 // class). We apply union-find path-halving technique (simple to implement) in
157 // the meantime as it decreases depths and the time complexity.
getLeader(std::vector<int> & leaders,int v)158 static int getLeader(std::vector<int> &leaders, int v) {
159 while (leaders[v] != v) {
160 leaders[v] = leaders[leaders[v]];
161 v = leaders[v];
162 }
163 return v;
164 }
165
mergeClusters(std::vector<Cluster> & cs,Cluster & into,int intoIdx,Cluster & from,int fromIdx)166 static void mergeClusters(std::vector<Cluster> &cs, Cluster &into, int intoIdx,
167 Cluster &from, int fromIdx) {
168 int tail1 = into.prev, tail2 = from.prev;
169 into.prev = tail2;
170 cs[tail2].next = intoIdx;
171 from.prev = tail1;
172 cs[tail1].next = fromIdx;
173 into.size += from.size;
174 into.weight += from.weight;
175 from.size = 0;
176 from.weight = 0;
177 }
178
179 // Group InputSections into clusters using the Call-Chain Clustering heuristic
180 // then sort the clusters by density.
run()181 DenseMap<const InputSectionBase *, int> CallGraphSort::run() {
182 std::vector<int> sorted(clusters.size());
183 std::vector<int> leaders(clusters.size());
184
185 std::iota(leaders.begin(), leaders.end(), 0);
186 std::iota(sorted.begin(), sorted.end(), 0);
187 llvm::stable_sort(sorted, [&](int a, int b) {
188 return clusters[a].getDensity() > clusters[b].getDensity();
189 });
190
191 for (int l : sorted) {
192 // The cluster index is the same as the index of its leader here because
193 // clusters[L] has not been merged into another cluster yet.
194 Cluster &c = clusters[l];
195
196 // Don't consider merging if the edge is unlikely.
197 if (c.bestPred.from == -1 || c.bestPred.weight * 10 <= c.initialWeight)
198 continue;
199
200 int predL = getLeader(leaders, c.bestPred.from);
201 if (l == predL)
202 continue;
203
204 Cluster *predC = &clusters[predL];
205 if (c.size + predC->size > MAX_CLUSTER_SIZE)
206 continue;
207
208 if (isNewDensityBad(*predC, c))
209 continue;
210
211 leaders[l] = predL;
212 mergeClusters(clusters, *predC, predL, c, l);
213 }
214
215 // Sort remaining non-empty clusters by density.
216 sorted.clear();
217 for (int i = 0, e = (int)clusters.size(); i != e; ++i)
218 if (clusters[i].size > 0)
219 sorted.push_back(i);
220 llvm::stable_sort(sorted, [&](int a, int b) {
221 return clusters[a].getDensity() > clusters[b].getDensity();
222 });
223
224 DenseMap<const InputSectionBase *, int> orderMap;
225 int curOrder = 1;
226 for (int leader : sorted) {
227 for (int i = leader;;) {
228 orderMap[sections[i]] = curOrder++;
229 i = clusters[i].next;
230 if (i == leader)
231 break;
232 }
233 }
234 if (!config->printSymbolOrder.empty()) {
235 std::error_code ec;
236 raw_fd_ostream os(config->printSymbolOrder, ec, sys::fs::OF_None);
237 if (ec) {
238 error("cannot open " + config->printSymbolOrder + ": " + ec.message());
239 return orderMap;
240 }
241
242 // Print the symbols ordered by C3, in the order of increasing curOrder
243 // Instead of sorting all the orderMap, just repeat the loops above.
244 for (int leader : sorted)
245 for (int i = leader;;) {
246 // Search all the symbols in the file of the section
247 // and find out a Defined symbol with name that is within the section.
248 for (Symbol *sym : sections[i]->file->getSymbols())
249 if (!sym->isSection()) // Filter out section-type symbols here.
250 if (auto *d = dyn_cast<Defined>(sym))
251 if (sections[i] == d->section)
252 os << sym->getName() << "\n";
253 i = clusters[i].next;
254 if (i == leader)
255 break;
256 }
257 }
258
259 return orderMap;
260 }
261
262 // Sort sections by the profile data provided by -callgraph-profile-file
263 //
264 // This first builds a call graph based on the profile data then merges sections
265 // according to the C³ heuristic. All clusters are then sorted by a density
266 // metric to further improve locality.
computeCallGraphProfileOrder()267 DenseMap<const InputSectionBase *, int> elf::computeCallGraphProfileOrder() {
268 return CallGraphSort().run();
269 }
270