1 | //===--- SyntheticCountsUtils.cpp - synthetic counts propagation utils ---===// |
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 | // This file defines utilities for propagating synthetic counts. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/Analysis/SyntheticCountsUtils.h" |
14 | #include "llvm/ADT/DenseSet.h" |
15 | #include "llvm/ADT/SCCIterator.h" |
16 | #include "llvm/Analysis/CallGraph.h" |
17 | #include "llvm/IR/ModuleSummaryIndex.h" |
18 | |
19 | using namespace llvm; |
20 | |
21 | // Given an SCC, propagate entry counts along the edge of the SCC nodes. |
22 | template <typename CallGraphType> |
23 | void SyntheticCountsUtils<CallGraphType>::propagateFromSCC( |
24 | const SccTy &SCC, GetProfCountTy GetProfCount, AddCountTy AddCount) { |
25 | |
26 | DenseSet<NodeRef> SCCNodes; |
27 | SmallVector<std::pair<NodeRef, EdgeRef>, 8> SCCEdges, NonSCCEdges; |
28 | |
29 | for (auto &Node : SCC) |
30 | SCCNodes.insert(Node); |
31 | |
32 | // Partition the edges coming out of the SCC into those whose destination is |
33 | // in the SCC and the rest. |
34 | for (const auto &Node : SCCNodes) { |
35 | for (auto &E : children_edges<CallGraphType>(Node)) { |
36 | if (SCCNodes.count(CGT::edge_dest(E))) |
37 | SCCEdges.emplace_back(Node, E); |
38 | else |
39 | NonSCCEdges.emplace_back(Node, E); |
40 | } |
41 | } |
42 | |
43 | // For nodes in the same SCC, update the counts in two steps: |
44 | // 1. Compute the additional count for each node by propagating the counts |
45 | // along all incoming edges to the node that originate from within the same |
46 | // SCC and summing them up. |
47 | // 2. Add the additional counts to the nodes in the SCC. |
48 | // This ensures that the order of |
49 | // traversal of nodes within the SCC doesn't affect the final result. |
50 | |
51 | DenseMap<NodeRef, Scaled64> AdditionalCounts; |
52 | for (auto &E : SCCEdges) { |
53 | auto OptProfCount = GetProfCount(E.first, E.second); |
54 | if (!OptProfCount) |
55 | continue; |
56 | auto Callee = CGT::edge_dest(E.second); |
57 | AdditionalCounts[Callee] += *OptProfCount; |
58 | } |
59 | |
60 | // Update the counts for the nodes in the SCC. |
61 | for (auto &Entry : AdditionalCounts) |
62 | AddCount(Entry.first, Entry.second); |
63 | |
64 | // Now update the counts for nodes outside the SCC. |
65 | for (auto &E : NonSCCEdges) { |
66 | auto OptProfCount = GetProfCount(E.first, E.second); |
67 | if (!OptProfCount) |
68 | continue; |
69 | auto Callee = CGT::edge_dest(E.second); |
70 | AddCount(Callee, *OptProfCount); |
71 | } |
72 | } |
73 | |
74 | /// Propgate synthetic entry counts on a callgraph \p CG. |
75 | /// |
76 | /// This performs a reverse post-order traversal of the callgraph SCC. For each |
77 | /// SCC, it first propagates the entry counts to the nodes within the SCC |
78 | /// through call edges and updates them in one shot. Then the entry counts are |
79 | /// propagated to nodes outside the SCC. This requires \p GraphTraits |
80 | /// to have a specialization for \p CallGraphType. |
81 | |
82 | template <typename CallGraphType> |
83 | void SyntheticCountsUtils<CallGraphType>::propagate(const CallGraphType &CG, |
84 | GetProfCountTy GetProfCount, |
85 | AddCountTy AddCount) { |
86 | std::vector<SccTy> SCCs; |
87 | |
88 | // Collect all the SCCs. |
89 | for (auto I = scc_begin(CG); !I.isAtEnd(); ++I) |
90 | SCCs.push_back(*I); |
91 | |
92 | // The callgraph-scc needs to be visited in top-down order for propagation. |
93 | // The scc iterator returns the scc in bottom-up order, so reverse the SCCs |
94 | // and call propagateFromSCC. |
95 | for (auto &SCC : reverse(SCCs)) |
96 | propagateFromSCC(SCC, GetProfCount, AddCount); |
97 | } |
98 | |
99 | template class llvm::SyntheticCountsUtils<const CallGraph *>; |
100 | template class llvm::SyntheticCountsUtils<ModuleSummaryIndex *>; |
101 | |