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
35using namespace llvm;
36using namespace lld;
37using namespace lld::elf;
38
39namespace {
40struct Edge {
41 int from;
42 uint64_t weight;
43};
44
45struct 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 = {.from: -1, .weight: 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
89class CallGraphSort {
90public:
91 CallGraphSort();
92
93 DenseMap<const InputSectionBase *, int> run();
94
95private:
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.
102constexpr int MAX_DENSITY_DEGRADATION = 8;
103
104// Maximum cluster size in bytes.
105constexpr uint64_t MAX_CLUSTER_SIZE = 1024 * 1024;
106} // end anonymous namespace
107
108using 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.
114CallGraphSort::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(Key: isec, Args: clusters.size());
120 if (res.second) {
121 sections.push_back(x: isec);
122 clusters.emplace_back(args: clusters.size(), args: 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>(Val: c.first.first);
130 const auto *toSB = cast<InputSectionBase>(Val: 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.
162static 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.
170static 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
178static 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.
193DenseMap<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(first: leaders.get(), last: leaders.get() + clusters.size(), value: 0);
198 std::iota(first: sorted.begin(), last: sorted.end(), value: 0);
199 llvm::stable_sort(Range&: sorted, C: [&](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: leaders.get(), v: 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(a&: *predC, b&: c))
221 continue;
222
223 leaders[l] = predL;
224 mergeClusters(cs&: clusters, into&: *predC, intoIdx: predL, from&: c, fromIdx: 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(x: i);
232 llvm::stable_sort(Range&: sorted, C: [&](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(msg: "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>(Val: 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.
277DenseMap<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(Key: inSec, Args: sections.size());
287 if (res.second) {
288 // inSec does not appear before in the graph.
289 sections.push_back(Elt: inSec);
290 funcSizes.push_back(Elt: inSec->getSize());
291 funcCounts.push_back(Elt: 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>(Val: c.first.first);
299 const InputSectionBase *toSB = cast<InputSectionBase>(Val: 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(Elt: {.src: from, .dst: to, .count: 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(Elt: (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: funcSizes, FuncCounts: funcCounts, CallCounts: callCounts, CallOffsets: 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.
339DenseMap<const InputSectionBase *, int> elf::computeCallGraphProfileOrder() {
340 if (config->callGraphProfileSort == CGProfileSortKind::Cdsort)
341 return computeCacheDirectedSortOrder();
342 return CallGraphSort().run();
343}
344