Run DCE after a LoopFlatten test to reduce spurious output [nfc]
[llvm-project.git] / bolt / lib / Passes / PettisAndHansen.cpp
blobf138c609b689add10b20e6078c45d5d640a2349e
1 //===- bolt/Passes/PettisAndHansen.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 implements Pettis and Hansen code-layout algorithm.
11 //===----------------------------------------------------------------------===//
13 #include "bolt/Passes/HFSort.h"
14 #include "llvm/Support/Debug.h"
15 #include "llvm/Support/Format.h"
16 #include "llvm/Support/raw_ostream.h"
17 #include <set>
18 #include <unordered_map>
20 #define DEBUG_TYPE "hfsort"
22 namespace llvm {
23 namespace bolt {
25 using NodeId = CallGraph::NodeId;
26 using Arc = CallGraph::Arc;
27 using Node = CallGraph::Node;
29 namespace {
30 class ClusterArc {
31 public:
32 ClusterArc(Cluster *Ca, Cluster *Cb, double W = 0)
33 : C1(std::min(Ca, Cb)), C2(std::max(Ca, Cb)), Weight(W) {}
35 friend bool operator==(const ClusterArc &Lhs, const ClusterArc &Rhs) {
36 return Lhs.C1 == Rhs.C1 && Lhs.C2 == Rhs.C2;
39 Cluster *const C1;
40 Cluster *const C2;
41 mutable double Weight;
44 class ClusterArcHash {
45 public:
46 int64_t operator()(const ClusterArc &Arc) const {
47 std::hash<int64_t> Hasher;
48 return hashCombine(Hasher(int64_t(Arc.C1)), int64_t(Arc.C2));
52 using ClusterArcSet = std::unordered_set<ClusterArc, ClusterArcHash>;
54 void orderFuncs(const CallGraph &Cg, Cluster *C1, Cluster *C2) {
55 NodeId C1head = C1->targets().front();
56 NodeId C1tail = C1->targets().back();
57 NodeId C2head = C2->targets().front();
58 NodeId C2tail = C2->targets().back();
60 double C1headC2head = 0;
61 double C1headC2tail = 0;
62 double C1tailC2head = 0;
63 double C1tailC2tail = 0;
65 for (const Arc &Arc : Cg.arcs()) {
66 if ((Arc.src() == C1head && Arc.dst() == C2head) ||
67 (Arc.dst() == C1head && Arc.src() == C2head))
68 C1headC2head += Arc.weight();
69 else if ((Arc.src() == C1head && Arc.dst() == C2tail) ||
70 (Arc.dst() == C1head && Arc.src() == C2tail))
71 C1headC2tail += Arc.weight();
72 else if ((Arc.src() == C1tail && Arc.dst() == C2head) ||
73 (Arc.dst() == C1tail && Arc.src() == C2head))
74 C1tailC2head += Arc.weight();
75 else if ((Arc.src() == C1tail && Arc.dst() == C2tail) ||
76 (Arc.dst() == C1tail && Arc.src() == C2tail))
77 C1tailC2tail += Arc.weight();
80 const double Max = std::max(std::max(C1headC2head, C1headC2tail),
81 std::max(C1tailC2head, C1tailC2tail));
83 if (C1headC2head == Max) {
84 // flip C1
85 C1->reverseTargets();
86 } else if (C1headC2tail == Max) {
87 // flip C1 C2
88 C1->reverseTargets();
89 C2->reverseTargets();
90 } else if (C1tailC2tail == Max) {
91 // flip C2
92 C2->reverseTargets();
95 } // namespace
97 std::vector<Cluster> pettisAndHansen(const CallGraph &Cg) {
98 // indexed by NodeId, keeps its current cluster
99 std::vector<Cluster *> FuncCluster(Cg.numNodes(), nullptr);
100 std::vector<Cluster> Clusters;
101 std::vector<NodeId> Funcs;
103 Clusters.reserve(Cg.numNodes());
105 for (NodeId F = 0; F < Cg.numNodes(); F++) {
106 if (Cg.samples(F) == 0)
107 continue;
108 Clusters.emplace_back(F, Cg.getNode(F));
109 FuncCluster[F] = &Clusters.back();
110 Funcs.push_back(F);
113 ClusterArcSet Carcs;
115 auto insertOrInc = [&](Cluster *C1, Cluster *C2, double Weight) {
116 auto Res = Carcs.emplace(C1, C2, Weight);
117 if (!Res.second)
118 Res.first->Weight += Weight;
121 // Create a std::vector of cluster arcs
123 for (const Arc &Arc : Cg.arcs()) {
124 if (Arc.weight() == 0)
125 continue;
127 Cluster *const S = FuncCluster[Arc.src()];
128 Cluster *const D = FuncCluster[Arc.dst()];
130 // ignore if s or d is nullptr
132 if (S == nullptr || D == nullptr)
133 continue;
135 // ignore self-edges
137 if (S == D)
138 continue;
140 insertOrInc(S, D, Arc.weight());
143 // Find an arc with max weight and merge its nodes
145 while (!Carcs.empty()) {
146 auto Maxpos =
147 std::max_element(Carcs.begin(), Carcs.end(),
148 [&](const ClusterArc &Carc1, const ClusterArc &Carc2) {
149 return Carc1.Weight < Carc2.Weight;
152 ClusterArc Max = *Maxpos;
153 Carcs.erase(Maxpos);
155 Cluster *const C1 = Max.C1;
156 Cluster *const C2 = Max.C2;
158 if (C1->size() + C2->size() > MaxClusterSize)
159 continue;
161 if (C1->frozen() || C2->frozen())
162 continue;
164 // order functions and merge cluster
166 orderFuncs(Cg, C1, C2);
168 LLVM_DEBUG(dbgs() << format("merging %s -> %s: %.1f\n",
169 C2->toString().c_str(), C1->toString().c_str(),
170 Max.Weight));
172 // update carcs: merge C1arcs to C2arcs
174 std::unordered_map<ClusterArc, Cluster *, ClusterArcHash> C2arcs;
175 for (const ClusterArc &Carc : Carcs) {
176 if (Carc.C1 == C2)
177 C2arcs.emplace(Carc, Carc.C2);
178 if (Carc.C2 == C2)
179 C2arcs.emplace(Carc, Carc.C1);
182 for (auto It : C2arcs) {
183 Cluster *const C = It.second;
184 ClusterArc const C2arc = It.first;
186 insertOrInc(C, C1, C2arc.Weight);
187 Carcs.erase(C2arc);
190 // update FuncCluster
192 for (NodeId F : C2->targets())
193 FuncCluster[F] = C1;
195 C1->merge(*C2, Max.Weight);
196 C2->clear();
199 // Return the set of Clusters that are left, which are the ones that
200 // didn't get merged.
202 std::set<Cluster *> LiveClusters;
203 std::vector<Cluster> OutClusters;
205 for (NodeId Fid : Funcs)
206 LiveClusters.insert(FuncCluster[Fid]);
207 for (Cluster *C : LiveClusters)
208 OutClusters.push_back(std::move(*C));
210 llvm::sort(OutClusters, compareClustersDensity);
212 return OutClusters;
215 } // namespace bolt
216 } // namespace llvm