1 //=- SyntheticCountsPropagation.cpp - Propagate function counts --*- C++ -*-=//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements a transformation that synthesizes entry counts for
11 // functions and attaches !prof metadata to functions with the synthesized
12 // counts. The presence of !prof metadata with counter name set to
13 // 'synthesized_function_entry_count' indicate that the value of the counter is
14 // an estimation of the likely execution count of the function. This transform
15 // is applied only in non PGO mode as functions get 'real' profile-based
16 // function entry counts in the PGO mode.
17 //
18 // The transformation works by first assigning some initial values to the entry
19 // counts of all functions and then doing a top-down traversal of the
20 // callgraph-scc to propagate the counts. For each function the set of callsites
21 // and their relative block frequency is gathered. The relative block frequency
22 // multiplied by the entry count of the caller and added to the callee's entry
23 // count. For non-trivial SCCs, the new counts are computed from the previous
24 // counts and updated in one shot.
25 //
26 //===----------------------------------------------------------------------===//
27 
28 #include "llvm/Transforms/IPO/SyntheticCountsPropagation.h"
29 #include "llvm/ADT/DenseSet.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/Analysis/BlockFrequencyInfo.h"
32 #include "llvm/Analysis/CallGraph.h"
33 #include "llvm/Analysis/SyntheticCountsUtils.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
41 
42 using namespace llvm;
43 using Scaled64 = ScaledNumber<uint64_t>;
44 
45 #define DEBUG_TYPE "synthetic-counts-propagation"
46 
47 /// Initial synthetic count assigned to functions.
48 static cl::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.
54 static 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.
59 static 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.
64 static void
65 initializeCounts(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 
96 PreservedAnalyses SyntheticCountsPropagation::run(Module &M,
97                                                   ModuleAnalysisManager &MAM) {
98   FunctionAnalysisManager &FAM =
99       MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
100   DenseMap<Function *, uint64_t> Counts;
101   // Set initial entry counts.
102   initializeCounts(M, [&](Function *F, uint64_t Count) { Counts[F] = Count; });
103 
104   // Compute the relative block frequency for a callsite. Use scaled numbers
105   // and not integers since the relative block frequency could be less than 1.
106   auto GetCallSiteRelFreq = [&](CallSite CS) {
107     Function *Caller = CS.getCaller();
108     auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(*Caller);
109     BasicBlock *CSBB = CS.getInstruction()->getParent();
110     Scaled64 EntryFreq(BFI.getEntryFreq(), 0);
111     Scaled64 BBFreq(BFI.getBlockFreq(CSBB).getFrequency(), 0);
112     BBFreq /= EntryFreq;
113     return BBFreq;
114   };
115 
116   CallGraph CG(M);
117   // Propgate the entry counts on the callgraph.
118   propagateSyntheticCounts(
119       CG, GetCallSiteRelFreq, [&](Function *F) { return Counts[F]; },
120       [&](Function *F, uint64_t New) { Counts[F] += New; });
121 
122   // Set the counts as metadata.
123   for (auto Entry : Counts)
124     Entry.first->setEntryCount(Entry.second, true);
125 
126   return PreservedAnalyses::all();
127 }
128