xref: /llvm-project/lld/ELF/CallGraphSort.cpp (revision bcc1e584483c8246b651290b0c2d696bd57006a9)
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(Ctx &);
92 
93   DenseMap<const InputSectionBase *, int> run();
94 
95 private:
96   Ctx &ctx;
97   std::vector<Cluster> clusters;
98   std::vector<const InputSectionBase *> sections;
99 };
100 
101 // Maximum amount the combined cluster density can be worse than the original
102 // cluster to consider merging.
103 constexpr int MAX_DENSITY_DEGRADATION = 8;
104 
105 // Maximum cluster size in bytes.
106 constexpr uint64_t MAX_CLUSTER_SIZE = 1024 * 1024;
107 } // end anonymous namespace
108 
109 using SectionPair =
110     std::pair<const InputSectionBase *, const InputSectionBase *>;
111 
112 // Take the edge list in ctx.arg.callGraphProfile, resolve symbol names to
113 // Symbols, and generate a graph between InputSections with the provided
114 // weights.
115 CallGraphSort::CallGraphSort(Ctx &ctx) : ctx(ctx) {
116   MapVector<SectionPair, uint64_t> &profile = ctx.arg.callGraphProfile;
117   DenseMap<const InputSectionBase *, int> secToCluster;
118 
119   auto getOrCreateNode = [&](const InputSectionBase *isec) -> int {
120     auto res = secToCluster.try_emplace(isec, clusters.size());
121     if (res.second) {
122       sections.push_back(isec);
123       clusters.emplace_back(clusters.size(), isec->getSize());
124     }
125     return res.first->second;
126   };
127 
128   // Create the graph.
129   for (std::pair<SectionPair, uint64_t> &c : profile) {
130     const auto *fromSB = cast<InputSectionBase>(c.first.first);
131     const auto *toSB = cast<InputSectionBase>(c.first.second);
132     uint64_t weight = c.second;
133 
134     // Ignore edges between input sections belonging to different output
135     // sections.  This is done because otherwise we would end up with clusters
136     // containing input sections that can't actually be placed adjacently in the
137     // output.  This messes with the cluster size and density calculations.  We
138     // would also end up moving input sections in other output sections without
139     // moving them closer to what calls them.
140     if (fromSB->getOutputSection() != toSB->getOutputSection())
141       continue;
142 
143     int from = getOrCreateNode(fromSB);
144     int to = getOrCreateNode(toSB);
145 
146     clusters[to].weight += weight;
147 
148     if (from == to)
149       continue;
150 
151     // Remember the best edge.
152     Cluster &toC = clusters[to];
153     if (toC.bestPred.from == -1 || toC.bestPred.weight < weight) {
154       toC.bestPred.from = from;
155       toC.bestPred.weight = weight;
156     }
157   }
158   for (Cluster &c : clusters)
159     c.initialWeight = c.weight;
160 }
161 
162 // It's bad to merge clusters which would degrade the density too much.
163 static bool isNewDensityBad(Cluster &a, Cluster &b) {
164   double newDensity = double(a.weight + b.weight) / double(a.size + b.size);
165   return newDensity < a.getDensity() / MAX_DENSITY_DEGRADATION;
166 }
167 
168 // Find the leader of V's belonged cluster (represented as an equivalence
169 // class). We apply union-find path-halving technique (simple to implement) in
170 // the meantime as it decreases depths and the time complexity.
171 static int getLeader(int *leaders, int v) {
172   while (leaders[v] != v) {
173     leaders[v] = leaders[leaders[v]];
174     v = leaders[v];
175   }
176   return v;
177 }
178 
179 static void mergeClusters(std::vector<Cluster> &cs, Cluster &into, int intoIdx,
180                           Cluster &from, int fromIdx) {
181   int tail1 = into.prev, tail2 = from.prev;
182   into.prev = tail2;
183   cs[tail2].next = intoIdx;
184   from.prev = tail1;
185   cs[tail1].next = fromIdx;
186   into.size += from.size;
187   into.weight += from.weight;
188   from.size = 0;
189   from.weight = 0;
190 }
191 
192 // Group InputSections into clusters using the Call-Chain Clustering heuristic
193 // then sort the clusters by density.
194 DenseMap<const InputSectionBase *, int> CallGraphSort::run() {
195   std::vector<int> sorted(clusters.size());
196   std::unique_ptr<int[]> leaders(new int[clusters.size()]);
197 
198   std::iota(leaders.get(), leaders.get() + clusters.size(), 0);
199   std::iota(sorted.begin(), sorted.end(), 0);
200   llvm::stable_sort(sorted, [&](int a, int b) {
201     return clusters[a].getDensity() > clusters[b].getDensity();
202   });
203 
204   for (int l : sorted) {
205     // The cluster index is the same as the index of its leader here because
206     // clusters[L] has not been merged into another cluster yet.
207     Cluster &c = clusters[l];
208 
209     // Don't consider merging if the edge is unlikely.
210     if (c.bestPred.from == -1 || c.bestPred.weight * 10 <= c.initialWeight)
211       continue;
212 
213     int predL = getLeader(leaders.get(), c.bestPred.from);
214     if (l == predL)
215       continue;
216 
217     Cluster *predC = &clusters[predL];
218     if (c.size + predC->size > MAX_CLUSTER_SIZE)
219       continue;
220 
221     if (isNewDensityBad(*predC, c))
222       continue;
223 
224     leaders[l] = predL;
225     mergeClusters(clusters, *predC, predL, c, l);
226   }
227 
228   // Sort remaining non-empty clusters by density.
229   sorted.clear();
230   for (int i = 0, e = (int)clusters.size(); i != e; ++i)
231     if (clusters[i].size > 0)
232       sorted.push_back(i);
233   llvm::stable_sort(sorted, [&](int a, int b) {
234     return clusters[a].getDensity() > clusters[b].getDensity();
235   });
236 
237   DenseMap<const InputSectionBase *, int> orderMap;
238   int curOrder = -clusters.size();
239   for (int leader : sorted) {
240     for (int i = leader;;) {
241       orderMap[sections[i]] = curOrder++;
242       i = clusters[i].next;
243       if (i == leader)
244         break;
245     }
246   }
247   if (!ctx.arg.printSymbolOrder.empty()) {
248     std::error_code ec;
249     raw_fd_ostream os(ctx.arg.printSymbolOrder, ec, sys::fs::OF_None);
250     if (ec) {
251       ErrAlways(ctx) << "cannot open " << ctx.arg.printSymbolOrder << ": "
252                      << ec.message();
253       return orderMap;
254     }
255 
256     // Print the symbols ordered by C3, in the order of increasing curOrder
257     // Instead of sorting all the orderMap, just repeat the loops above.
258     for (int leader : sorted)
259       for (int i = leader;;) {
260         // Search all the symbols in the file of the section
261         // and find out a Defined symbol with name that is within the section.
262         for (Symbol *sym : sections[i]->file->getSymbols())
263           if (!sym->isSection()) // Filter out section-type symbols here.
264             if (auto *d = dyn_cast<Defined>(sym))
265               if (sections[i] == d->section)
266                 os << sym->getName() << "\n";
267         i = clusters[i].next;
268         if (i == leader)
269           break;
270       }
271   }
272 
273   return orderMap;
274 }
275 
276 // Sort sections by the profile data using the Cache-Directed Sort algorithm.
277 // The placement is done by optimizing the locality by co-locating frequently
278 // executed code sections together.
279 static DenseMap<const InputSectionBase *, int>
280 computeCacheDirectedSortOrder(Ctx &ctx) {
281   SmallVector<uint64_t, 0> funcSizes;
282   SmallVector<uint64_t, 0> funcCounts;
283   SmallVector<codelayout::EdgeCount, 0> callCounts;
284   SmallVector<uint64_t, 0> callOffsets;
285   SmallVector<const InputSectionBase *, 0> sections;
286   DenseMap<const InputSectionBase *, size_t> secToTargetId;
287 
288   auto getOrCreateNode = [&](const InputSectionBase *inSec) -> size_t {
289     auto res = secToTargetId.try_emplace(inSec, sections.size());
290     if (res.second) {
291       // inSec does not appear before in the graph.
292       sections.push_back(inSec);
293       funcSizes.push_back(inSec->getSize());
294       funcCounts.push_back(0);
295     }
296     return res.first->second;
297   };
298 
299   // Create the graph.
300   for (std::pair<SectionPair, uint64_t> &c : ctx.arg.callGraphProfile) {
301     const InputSectionBase *fromSB = cast<InputSectionBase>(c.first.first);
302     const InputSectionBase *toSB = cast<InputSectionBase>(c.first.second);
303     // Ignore edges between input sections belonging to different sections.
304     if (fromSB->getOutputSection() != toSB->getOutputSection())
305       continue;
306 
307     uint64_t weight = c.second;
308     // Ignore edges with zero weight.
309     if (weight == 0)
310       continue;
311 
312     size_t from = getOrCreateNode(fromSB);
313     size_t to = getOrCreateNode(toSB);
314     // Ignore self-edges (recursive calls).
315     if (from == to)
316       continue;
317 
318     callCounts.push_back({from, to, weight});
319     // Assume that the jump is at the middle of the input section. The profile
320     // data does not contain jump offsets.
321     callOffsets.push_back((funcSizes[from] + 1) / 2);
322     funcCounts[to] += weight;
323   }
324 
325   // Run the layout algorithm.
326   std::vector<uint64_t> sortedSections = codelayout::computeCacheDirectedLayout(
327       funcSizes, funcCounts, callCounts, callOffsets);
328 
329   // Create the final order.
330   DenseMap<const InputSectionBase *, int> orderMap;
331   int curOrder = -sortedSections.size();
332   for (uint64_t secIdx : sortedSections)
333     orderMap[sections[secIdx]] = curOrder++;
334 
335   return orderMap;
336 }
337 
338 // Sort sections by the profile data provided by --callgraph-profile-file.
339 //
340 // This first builds a call graph based on the profile data then merges sections
341 // according either to the C³ or Cache-Directed-Sort ordering algorithm.
342 DenseMap<const InputSectionBase *, int>
343 elf::computeCallGraphProfileOrder(Ctx &ctx) {
344   if (ctx.arg.callGraphProfileSort == CGProfileSortKind::Cdsort)
345     return computeCacheDirectedSortOrder(ctx);
346   return CallGraphSort(ctx).run();
347 }
348