1//=- SyntheticCountsPropagation.cpp - Propagate function counts --*- C++ -*-=//
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 a transformation that synthesizes entry counts for
10// functions and attaches !prof metadata to functions with the synthesized
11// counts. The presence of !prof metadata with counter name set to
12// 'synthesized_function_entry_count' indicate that the value of the counter is
13// an estimation of the likely execution count of the function. This transform
14// is applied only in non PGO mode as functions get 'real' profile-based
15// function entry counts in the PGO mode.
16//
17// The transformation works by first assigning some initial values to the entry
18// counts of all functions and then doing a top-down traversal of the
19// callgraph-scc to propagate the counts. For each function the set of callsites
20// and their relative block frequency is gathered. The relative block frequency
21// multiplied by the entry count of the caller and added to the callee's entry
22// count. For non-trivial SCCs, the new counts are computed from the previous
23// counts and updated in one shot.
24//
25//===----------------------------------------------------------------------===//
26
27#include "llvm/Transforms/IPO/SyntheticCountsPropagation.h"
28#include "llvm/ADT/DenseSet.h"
29#include "llvm/ADT/STLExtras.h"
30#include "llvm/Analysis/BlockFrequencyInfo.h"
31#include "llvm/Analysis/CallGraph.h"
32#include "llvm/Analysis/ProfileSummaryInfo.h"
33#include "llvm/Analysis/SyntheticCountsUtils.h"
34#include "llvm/IR/Function.h"
35#include "llvm/IR/Instructions.h"
36#include "llvm/IR/Module.h"
37#include "llvm/Support/CommandLine.h"
38#include "llvm/Support/Debug.h"
39#include "llvm/Support/raw_ostream.h"
40
41using namespace llvm;
42using Scaled64 = ScaledNumber<uint64_t>;
43using ProfileCount = Function::ProfileCount;
44
45#define DEBUG_TYPE "synthetic-counts-propagation"
46
47/// Initial synthetic count assigned to functions.
48cl::opt<int>
49    InitialSyntheticCount("initial-synthetic-count", cl::Hidden, cl::init(10),
50                          cl::ZeroOrMore,
51                          cl::desc("Initial value of synthetic entry count."));
52
53/// Initial synthetic count assigned to inline functions.
54static cl::opt<int> InlineSyntheticCount(
55    "inline-synthetic-count", cl::Hidden, cl::init(15), cl::ZeroOrMore,
56    cl::desc("Initial synthetic entry count for inline functions."));
57
58/// Initial synthetic count assigned to cold functions.
59static cl::opt<int> ColdSyntheticCount(
60    "cold-synthetic-count", cl::Hidden, cl::init(5), cl::ZeroOrMore,
61    cl::desc("Initial synthetic entry count for cold functions."));
62
63// Assign initial synthetic entry counts to functions.
64static void
65initializeCounts(Module &M, function_ref<void(Function *, uint64_t)> SetCount) {
66  auto MayHaveIndirectCalls = [](Function &F) {
67    for (auto *U : F.users()) {
68      if (!isa<CallInst>(U) && !isa<InvokeInst>(U))
69        return true;
70    }
71    return false;
72  };
73
74  for (Function &F : M) {
75    uint64_t InitialCount = InitialSyntheticCount;
76    if (F.isDeclaration())
77      continue;
78    if (F.hasFnAttribute(Attribute::AlwaysInline) ||
79        F.hasFnAttribute(Attribute::InlineHint)) {
80      // Use a higher value for inline functions to account for the fact that
81      // these are usually beneficial to inline.
82      InitialCount = InlineSyntheticCount;
83    } else if (F.hasLocalLinkage() && !MayHaveIndirectCalls(F)) {
84      // Local functions without inline hints get counts only through
85      // propagation.
86      InitialCount = 0;
87    } else if (F.hasFnAttribute(Attribute::Cold) ||
88               F.hasFnAttribute(Attribute::NoInline)) {
89      // Use a lower value for noinline and cold functions.
90      InitialCount = ColdSyntheticCount;
91    }
92    SetCount(&F, InitialCount);
93  }
94}
95
96PreservedAnalyses SyntheticCountsPropagation::run(Module &M,
97                                                  ModuleAnalysisManager &MAM) {
98  FunctionAnalysisManager &FAM =
99      MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
100  DenseMap<Function *, Scaled64> Counts;
101  // Set initial entry counts.
102  initializeCounts(
103      M, [&](Function *F, uint64_t Count) { Counts[F] = Scaled64(Count, 0); });
104
105  // Edge includes information about the source. Hence ignore the first
106  // parameter.
107  auto GetCallSiteProfCount = [&](const CallGraphNode *,
108                                  const CallGraphNode::CallRecord &Edge) {
109    Optional<Scaled64> Res = None;
110    if (!Edge.first)
111      return Res;
112    CallBase &CB = *cast<CallBase>(*Edge.first);
113    Function *Caller = CB.getCaller();
114    auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(*Caller);
115
116    // Now compute the callsite count from relative frequency and
117    // entry count:
118    BasicBlock *CSBB = CB.getParent();
119    Scaled64 EntryFreq(BFI.getEntryFreq(), 0);
120    Scaled64 BBCount(BFI.getBlockFreq(CSBB).getFrequency(), 0);
121    BBCount /= EntryFreq;
122    BBCount *= Counts[Caller];
123    return Optional<Scaled64>(BBCount);
124  };
125
126  CallGraph CG(M);
127  // Propgate the entry counts on the callgraph.
128  SyntheticCountsUtils<const CallGraph *>::propagate(
129      &CG, GetCallSiteProfCount, [&](const CallGraphNode *N, Scaled64 New) {
130        auto F = N->getFunction();
131        if (!F || F->isDeclaration())
132          return;
133
134        Counts[F] += New;
135      });
136
137  // Set the counts as metadata.
138  for (auto Entry : Counts) {
139    Entry.first->setEntryCount(ProfileCount(
140        Entry.second.template toInt<uint64_t>(), Function::PCT_Synthetic));
141  }
142
143  return PreservedAnalyses::all();
144}
145