1 //==-SummaryBasedOptimizations.cpp - Optimizations based on ThinLTO summary-==//
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 implements optimizations that are based on the module summaries.
10 // These optimizations are performed during the thinlink phase of the
11 // compilation.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/LTO/SummaryBasedOptimizations.h"
16 #include "llvm/Analysis/SyntheticCountsUtils.h"
17 #include "llvm/IR/ModuleSummaryIndex.h"
18 #include "llvm/Support/CommandLine.h"
19
20 using namespace llvm;
21
22 static cl::opt<bool> ThinLTOSynthesizeEntryCounts(
23 "thinlto-synthesize-entry-counts", cl::init(false), cl::Hidden,
24 cl::desc("Synthesize entry counts based on the summary"));
25
26 extern cl::opt<int> InitialSyntheticCount;
27
initializeCounts(ModuleSummaryIndex & Index)28 static void initializeCounts(ModuleSummaryIndex &Index) {
29 auto Root = Index.calculateCallGraphRoot();
30 // Root is a fake node. All its successors are the actual roots of the
31 // callgraph.
32 // FIXME: This initializes the entry counts of only the root nodes. This makes
33 // sense when compiling a binary with ThinLTO, but for libraries any of the
34 // non-root nodes could be called from outside.
35 for (auto &C : Root.calls()) {
36 auto &V = C.first;
37 for (auto &GVS : V.getSummaryList()) {
38 auto S = GVS.get()->getBaseObject();
39 auto *F = cast<FunctionSummary>(S);
40 F->setEntryCount(InitialSyntheticCount);
41 }
42 }
43 }
44
computeSyntheticCounts(ModuleSummaryIndex & Index)45 void llvm::computeSyntheticCounts(ModuleSummaryIndex &Index) {
46 if (!ThinLTOSynthesizeEntryCounts)
47 return;
48
49 using Scaled64 = ScaledNumber<uint64_t>;
50 initializeCounts(Index);
51 auto GetCallSiteRelFreq = [](FunctionSummary::EdgeTy &Edge) {
52 return Scaled64(Edge.second.RelBlockFreq, -CalleeInfo::ScaleShift);
53 };
54 auto GetEntryCount = [](ValueInfo V) {
55 if (V.getSummaryList().size()) {
56 auto S = V.getSummaryList().front().get()->getBaseObject();
57 auto *F = cast<FunctionSummary>(S);
58 return F->entryCount();
59 } else {
60 return UINT64_C(0);
61 }
62 };
63 auto AddToEntryCount = [](ValueInfo V, Scaled64 New) {
64 if (!V.getSummaryList().size())
65 return;
66 for (auto &GVS : V.getSummaryList()) {
67 auto S = GVS.get()->getBaseObject();
68 auto *F = cast<FunctionSummary>(S);
69 F->setEntryCount(
70 SaturatingAdd(F->entryCount(), New.template toInt<uint64_t>()));
71 }
72 };
73
74 auto GetProfileCount = [&](ValueInfo V, FunctionSummary::EdgeTy &Edge) {
75 auto RelFreq = GetCallSiteRelFreq(Edge);
76 Scaled64 EC(GetEntryCount(V), 0);
77 return RelFreq * EC;
78 };
79 // After initializing the counts in initializeCounts above, the counts have to
80 // be propagated across the combined callgraph.
81 // SyntheticCountsUtils::propagate takes care of this propagation on any
82 // callgraph that specialized GraphTraits.
83 SyntheticCountsUtils<ModuleSummaryIndex *>::propagate(&Index, GetProfileCount,
84 AddToEntryCount);
85 Index.setHasSyntheticEntryCounts();
86 }
87