1 //===-- TimeProfiler.cpp - Hierarchical Time Profiler ---------------------===//
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 hierarchical time profiler.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Support/TimeProfiler.h"
14 #include "llvm/ADT/StringMap.h"
15 #include "llvm/Support/CommandLine.h"
16 #include "llvm/Support/FileSystem.h"
17 #include "llvm/Support/JSON.h"
18 #include <cassert>
19 #include <chrono>
20 #include <string>
21 #include <vector>
22 
23 using namespace std::chrono;
24 
25 namespace llvm {
26 
27 static cl::opt<unsigned> TimeTraceGranularity(
28     "time-trace-granularity",
29     cl::desc(
30         "Minimum time granularity (in microseconds) traced by time profiler"),
31     cl::init(500));
32 
33 TimeTraceProfiler *TimeTraceProfilerInstance = nullptr;
34 
35 typedef duration<steady_clock::rep, steady_clock::period> DurationType;
36 typedef std::pair<size_t, DurationType> CountAndDurationType;
37 typedef std::pair<std::string, CountAndDurationType>
38     NameAndCountAndDurationType;
39 
40 struct Entry {
41   time_point<steady_clock> Start;
42   DurationType Duration;
43   std::string Name;
44   std::string Detail;
45 
46   Entry(time_point<steady_clock> &&S, DurationType &&D, std::string &&N,
47         std::string &&Dt)
48       : Start(std::move(S)), Duration(std::move(D)), Name(std::move(N)),
49         Detail(std::move(Dt)){};
50 };
51 
52 struct TimeTraceProfiler {
53   TimeTraceProfiler() {
54     StartTime = steady_clock::now();
55   }
56 
57   void begin(std::string Name, llvm::function_ref<std::string()> Detail) {
58     Stack.emplace_back(steady_clock::now(), DurationType{}, std::move(Name),
59                        Detail());
60   }
61 
62   void end() {
63     assert(!Stack.empty() && "Must call begin() first");
64     auto &E = Stack.back();
65     E.Duration = steady_clock::now() - E.Start;
66 
67     // Only include sections longer than TimeTraceGranularity msec.
68     if (duration_cast<microseconds>(E.Duration).count() > TimeTraceGranularity)
69       Entries.emplace_back(E);
70 
71     // Track total time taken by each "name", but only the topmost levels of
72     // them; e.g. if there's a template instantiation that instantiates other
73     // templates from within, we only want to add the topmost one. "topmost"
74     // happens to be the ones that don't have any currently open entries above
75     // itself.
76     if (std::find_if(++Stack.rbegin(), Stack.rend(), [&](const Entry &Val) {
77           return Val.Name == E.Name;
78         }) == Stack.rend()) {
79       auto &CountAndTotal = CountAndTotalPerName[E.Name];
80       CountAndTotal.first++;
81       CountAndTotal.second += E.Duration;
82     }
83 
84     Stack.pop_back();
85   }
86 
87   void Write(raw_pwrite_stream &OS) {
88     assert(Stack.empty() &&
89            "All profiler sections should be ended when calling Write");
90 
91     json::Array Events;
92     const size_t ExpectedEntryCount =
93         Entries.size() + CountAndTotalPerName.size() + 1;
94     Events.reserve(ExpectedEntryCount);
95 
96     // Emit all events for the main flame graph.
97     for (const auto &E : Entries) {
98       auto StartUs = duration_cast<microseconds>(E.Start - StartTime).count();
99       auto DurUs = duration_cast<microseconds>(E.Duration).count();
100 
101       Events.emplace_back(json::Object{
102           {"pid", 1},
103           {"tid", 0},
104           {"ph", "X"},
105           {"ts", StartUs},
106           {"dur", DurUs},
107           {"name", E.Name},
108           {"args", json::Object{{"detail", E.Detail}}},
109       });
110     }
111 
112     // Emit totals by section name as additional "thread" events, sorted from
113     // longest one.
114     int Tid = 1;
115     std::vector<NameAndCountAndDurationType> SortedTotals;
116     SortedTotals.reserve(CountAndTotalPerName.size());
117     for (const auto &E : CountAndTotalPerName)
118       SortedTotals.emplace_back(E.getKey(), E.getValue());
119 
120     llvm::sort(SortedTotals.begin(), SortedTotals.end(),
121                [](const NameAndCountAndDurationType &A,
122                   const NameAndCountAndDurationType &B) {
123                  return A.second.second > B.second.second;
124                });
125     for (const auto &E : SortedTotals) {
126       auto DurUs = duration_cast<microseconds>(E.second.second).count();
127       auto Count = CountAndTotalPerName[E.first].first;
128 
129       Events.emplace_back(json::Object{
130           {"pid", 1},
131           {"tid", Tid},
132           {"ph", "X"},
133           {"ts", 0},
134           {"dur", DurUs},
135           {"name", "Total " + E.first},
136           {"args", json::Object{{"count", static_cast<int64_t>(Count)},
137                                 {"avg ms",
138                                  static_cast<int64_t>(DurUs / Count / 1000)}}},
139       });
140 
141       ++Tid;
142     }
143 
144     // Emit metadata event with process name.
145     Events.emplace_back(json::Object{
146         {"cat", ""},
147         {"pid", 1},
148         {"tid", 0},
149         {"ts", 0},
150         {"ph", "M"},
151         {"name", "process_name"},
152         {"args", json::Object{{"name", "clang"}}},
153     });
154 
155     assert(Events.size() == ExpectedEntryCount && "Size prediction failed!");
156 
157     OS << formatv("{0:2}", json::Value(json::Object(
158                                {{"traceEvents", std::move(Events)}})));
159   }
160 
161   SmallVector<Entry, 16> Stack;
162   SmallVector<Entry, 128> Entries;
163   StringMap<CountAndDurationType> CountAndTotalPerName;
164   time_point<steady_clock> StartTime;
165 };
166 
167 void timeTraceProfilerInitialize() {
168   assert(TimeTraceProfilerInstance == nullptr &&
169          "Profiler should not be initialized");
170   TimeTraceProfilerInstance = new TimeTraceProfiler();
171 }
172 
173 void timeTraceProfilerCleanup() {
174   delete TimeTraceProfilerInstance;
175   TimeTraceProfilerInstance = nullptr;
176 }
177 
178 void timeTraceProfilerWrite(raw_pwrite_stream &OS) {
179   assert(TimeTraceProfilerInstance != nullptr &&
180          "Profiler object can't be null");
181   TimeTraceProfilerInstance->Write(OS);
182 }
183 
184 void timeTraceProfilerBegin(StringRef Name, StringRef Detail) {
185   if (TimeTraceProfilerInstance != nullptr)
186     TimeTraceProfilerInstance->begin(Name, [&]() { return Detail; });
187 }
188 
189 void timeTraceProfilerBegin(StringRef Name,
190                             llvm::function_ref<std::string()> Detail) {
191   if (TimeTraceProfilerInstance != nullptr)
192     TimeTraceProfilerInstance->begin(Name, Detail);
193 }
194 
195 void timeTraceProfilerEnd() {
196   if (TimeTraceProfilerInstance != nullptr)
197     TimeTraceProfilerInstance->end();
198 }
199 
200 } // namespace llvm
201