1 //===- xray-account.h - XRay Function Call Accounting ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements basic function call accounting from an XRay trace.
12 //===----------------------------------------------------------------------===//
17 #include <system_error>
20 #include "xray-account.h"
21 #include "xray-registry.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/FormatVariadic.h"
24 #include "llvm/XRay/InstrumentationMap.h"
25 #include "llvm/XRay/Trace.h"
28 using namespace llvm::xray;
30 static cl::SubCommand Account("account", "Function call accounting");
31 static cl::opt<std::string> AccountInput(cl::Positional,
32 cl::desc("<xray log file>"),
33 cl::Required, cl::sub(Account));
35 AccountKeepGoing("keep-going", cl::desc("Keep going on errors encountered"),
36 cl::sub(Account), cl::init(false));
37 static cl::alias AccountKeepGoing2("k", cl::aliasopt(AccountKeepGoing),
38 cl::desc("Alias for -keep_going"),
40 static cl::opt<bool> AccountDeduceSiblingCalls(
41 "deduce-sibling-calls",
42 cl::desc("Deduce sibling calls when unrolling function call stacks"),
43 cl::sub(Account), cl::init(false));
45 AccountDeduceSiblingCalls2("d", cl::aliasopt(AccountDeduceSiblingCalls),
46 cl::desc("Alias for -deduce_sibling_calls"),
48 static cl::opt<std::string>
49 AccountOutput("output", cl::value_desc("output file"), cl::init("-"),
50 cl::desc("output file; use '-' for stdout"),
52 static cl::alias AccountOutput2("o", cl::aliasopt(AccountOutput),
53 cl::desc("Alias for -output"),
55 enum class AccountOutputFormats { TEXT, CSV };
56 static cl::opt<AccountOutputFormats>
57 AccountOutputFormat("format", cl::desc("output format"),
58 cl::values(clEnumValN(AccountOutputFormats::TEXT,
59 "text", "report stats in text"),
60 clEnumValN(AccountOutputFormats::CSV, "csv",
61 "report stats in csv")),
63 static cl::alias AccountOutputFormat2("f", cl::desc("Alias of -format"),
64 cl::aliasopt(AccountOutputFormat),
67 enum class SortField {
79 static cl::opt<SortField> AccountSortOutput(
80 "sort", cl::desc("sort output by this field"), cl::value_desc("field"),
81 cl::sub(Account), cl::init(SortField::FUNCID),
82 cl::values(clEnumValN(SortField::FUNCID, "funcid", "function id"),
83 clEnumValN(SortField::COUNT, "count", "funciton call counts"),
84 clEnumValN(SortField::MIN, "min", "minimum function durations"),
85 clEnumValN(SortField::MED, "med", "median function durations"),
86 clEnumValN(SortField::PCT90, "90p", "90th percentile durations"),
87 clEnumValN(SortField::PCT99, "99p", "99th percentile durations"),
88 clEnumValN(SortField::MAX, "max", "maximum function durations"),
89 clEnumValN(SortField::SUM, "sum", "sum of call durations"),
90 clEnumValN(SortField::FUNC, "func", "function names")));
91 static cl::alias AccountSortOutput2("s", cl::aliasopt(AccountSortOutput),
92 cl::desc("Alias for -sort"),
95 enum class SortDirection {
99 static cl::opt<SortDirection> AccountSortOrder(
100 "sortorder", cl::desc("sort ordering"), cl::init(SortDirection::ASCENDING),
101 cl::values(clEnumValN(SortDirection::ASCENDING, "asc", "ascending"),
102 clEnumValN(SortDirection::DESCENDING, "dsc", "descending")),
104 static cl::alias AccountSortOrder2("r", cl::aliasopt(AccountSortOrder),
105 cl::desc("Alias for -sortorder"),
108 static cl::opt<int> AccountTop("top", cl::desc("only show the top N results"),
109 cl::value_desc("N"), cl::sub(Account),
111 static cl::alias AccountTop2("p", cl::desc("Alias for -top"),
112 cl::aliasopt(AccountTop), cl::sub(Account));
114 static cl::opt<std::string>
115 AccountInstrMap("instr_map",
116 cl::desc("binary with the instrumentation map, or "
117 "a separate instrumentation map"),
118 cl::value_desc("binary with xray_instr_map"),
119 cl::sub(Account), cl::init(""));
120 static cl::alias AccountInstrMap2("m", cl::aliasopt(AccountInstrMap),
121 cl::desc("Alias for -instr_map"),
126 template <class T, class U> void setMinMax(std::pair<T, T> &MM, U &&V) {
127 if (MM.first == 0 || MM.second == 0)
128 MM = std::make_pair(std::forward<U>(V), std::forward<U>(V));
130 MM = std::make_pair(std::min(MM.first, V), std::max(MM.second, V));
133 template <class T> T diff(T L, T R) { return std::max(L, R) - std::min(L, R); }
137 bool LatencyAccountant::accountRecord(const XRayRecord &Record) {
138 setMinMax(PerThreadMinMaxTSC[Record.TId], Record.TSC);
139 setMinMax(PerCPUMinMaxTSC[Record.CPU], Record.TSC);
141 if (CurrentMaxTSC == 0)
142 CurrentMaxTSC = Record.TSC;
144 if (Record.TSC < CurrentMaxTSC)
147 auto &ThreadStack = PerThreadFunctionStack[Record.TId];
148 switch (Record.Type) {
149 case RecordTypes::ENTER: {
151 ThreadStack.emplace_back(Record.FuncId, Record.TSC);
154 case RecordTypes::EXIT: {
156 if (ThreadStack.back().first == Record.FuncId) {
157 const auto &Top = ThreadStack.back();
158 recordLatency(Top.first, diff(Top.second, Record.TSC));
159 ThreadStack.pop_back();
163 if (!DeduceSiblingCalls)
166 // Look for the parent up the stack.
168 std::find_if(ThreadStack.rbegin(), ThreadStack.rend(),
169 [&](const std::pair<const int32_t, uint64_t> &E) {
170 return E.first == Record.FuncId;
172 if (Parent == ThreadStack.rend())
175 // Account time for this apparently sibling call exit up the stack.
176 // Considering the following case:
182 // We might only ever see the following entries:
190 // Now we don't see the exit to g() because some older version of the XRay
191 // runtime wasn't instrumenting tail exits. If we don't deduce tail calls,
192 // we may potentially never account time for g() -- and this code would have
193 // already bailed out, because `<- f()` doesn't match the current "top" of
194 // stack where we're waiting for the exit to `g()` instead. This is not
195 // ideal and brittle -- so instead we provide a potentially inaccurate
196 // accounting of g() instead, computing it from the exit of f().
198 // While it might be better that we account the time between `-> g()` and
199 // `-> h()` as the proper accounting of time for g() here, this introduces
200 // complexity to do correctly (need to backtrack, etc.).
202 // FIXME: Potentially implement the more complex deduction algorithm?
203 auto I = std::next(Parent).base();
204 for (auto &E : make_range(I, ThreadStack.end())) {
205 recordLatency(E.first, diff(E.second, Record.TSC));
207 ThreadStack.erase(I, ThreadStack.end());
217 // We consolidate the data into a struct which we can output in various forms.
226 std::string DebugInfo;
227 std::string Function;
230 ResultRow getStats(std::vector<uint64_t> &Timings) {
231 assert(!Timings.empty());
233 R.Sum = std::accumulate(Timings.begin(), Timings.end(), 0.0);
234 auto MinMax = std::minmax_element(Timings.begin(), Timings.end());
235 R.Min = *MinMax.first;
236 R.Max = *MinMax.second;
237 auto MedianOff = Timings.size() / 2;
238 std::nth_element(Timings.begin(), Timings.begin() + MedianOff, Timings.end());
239 R.Median = Timings[MedianOff];
240 auto Pct90Off = std::floor(Timings.size() * 0.9);
241 std::nth_element(Timings.begin(), Timings.begin() + Pct90Off, Timings.end());
242 R.Pct90 = Timings[Pct90Off];
243 auto Pct99Off = std::floor(Timings.size() * 0.99);
244 std::nth_element(Timings.begin(), Timings.begin() + Pct90Off, Timings.end());
245 R.Pct99 = Timings[Pct99Off];
246 R.Count = Timings.size();
253 void LatencyAccountant::exportStats(const XRayFileHeader &Header, F Fn) const {
254 using TupleType = std::tuple<int32_t, uint64_t, ResultRow>;
255 std::vector<TupleType> Results;
256 Results.reserve(FunctionLatencies.size());
257 for (auto FT : FunctionLatencies) {
258 const auto &FuncId = FT.first;
259 auto &Timings = FT.second;
260 Results.emplace_back(FuncId, Timings.size(), getStats(Timings));
261 auto &Row = std::get<2>(Results.back());
262 if (Header.CycleFrequency) {
263 double CycleFrequency = Header.CycleFrequency;
264 Row.Min /= CycleFrequency;
265 Row.Median /= CycleFrequency;
266 Row.Pct90 /= CycleFrequency;
267 Row.Pct99 /= CycleFrequency;
268 Row.Max /= CycleFrequency;
269 Row.Sum /= CycleFrequency;
272 Row.Function = FuncIdHelper.SymbolOrNumber(FuncId);
273 Row.DebugInfo = FuncIdHelper.FileLineAndColumn(FuncId);
276 // Sort the data according to user-provided flags.
277 switch (AccountSortOutput) {
278 case SortField::FUNCID:
279 std::sort(Results.begin(), Results.end(),
280 [](const TupleType &L, const TupleType &R) {
281 if (AccountSortOrder == SortDirection::ASCENDING)
282 return std::get<0>(L) < std::get<0>(R);
283 if (AccountSortOrder == SortDirection::DESCENDING)
284 return std::get<0>(L) > std::get<0>(R);
285 llvm_unreachable("Unknown sort direction");
288 case SortField::COUNT:
289 std::sort(Results.begin(), Results.end(),
290 [](const TupleType &L, const TupleType &R) {
291 if (AccountSortOrder == SortDirection::ASCENDING)
292 return std::get<1>(L) < std::get<1>(R);
293 if (AccountSortOrder == SortDirection::DESCENDING)
294 return std::get<1>(L) > std::get<1>(R);
295 llvm_unreachable("Unknown sort direction");
299 // Here we need to look into the ResultRow for the rest of the data that
300 // we want to sort by.
301 std::sort(Results.begin(), Results.end(),
302 [&](const TupleType &L, const TupleType &R) {
303 auto &LR = std::get<2>(L);
304 auto &RR = std::get<2>(R);
305 switch (AccountSortOutput) {
306 case SortField::COUNT:
307 if (AccountSortOrder == SortDirection::ASCENDING)
308 return LR.Count < RR.Count;
309 if (AccountSortOrder == SortDirection::DESCENDING)
310 return LR.Count > RR.Count;
311 llvm_unreachable("Unknown sort direction");
313 if (AccountSortOrder == SortDirection::ASCENDING)
314 return LR.Min < RR.Min;
315 if (AccountSortOrder == SortDirection::DESCENDING)
316 return LR.Min > RR.Min;
317 llvm_unreachable("Unknown sort direction");
319 if (AccountSortOrder == SortDirection::ASCENDING)
320 return LR.Median < RR.Median;
321 if (AccountSortOrder == SortDirection::DESCENDING)
322 return LR.Median > RR.Median;
323 llvm_unreachable("Unknown sort direction");
324 case SortField::PCT90:
325 if (AccountSortOrder == SortDirection::ASCENDING)
326 return LR.Pct90 < RR.Pct90;
327 if (AccountSortOrder == SortDirection::DESCENDING)
328 return LR.Pct90 > RR.Pct90;
329 llvm_unreachable("Unknown sort direction");
330 case SortField::PCT99:
331 if (AccountSortOrder == SortDirection::ASCENDING)
332 return LR.Pct99 < RR.Pct99;
333 if (AccountSortOrder == SortDirection::DESCENDING)
334 return LR.Pct99 > RR.Pct99;
335 llvm_unreachable("Unknown sort direction");
337 if (AccountSortOrder == SortDirection::ASCENDING)
338 return LR.Max < RR.Max;
339 if (AccountSortOrder == SortDirection::DESCENDING)
340 return LR.Max > RR.Max;
341 llvm_unreachable("Unknown sort direction");
343 if (AccountSortOrder == SortDirection::ASCENDING)
344 return LR.Sum < RR.Sum;
345 if (AccountSortOrder == SortDirection::DESCENDING)
346 return LR.Sum > RR.Sum;
347 llvm_unreachable("Unknown sort direction");
349 llvm_unreachable("Unsupported sort order");
356 Results.erase(Results.begin() + AccountTop.getValue(), Results.end());
358 for (const auto &R : Results)
359 Fn(std::get<0>(R), std::get<1>(R), std::get<2>(R));
362 void LatencyAccountant::exportStatsAsText(raw_ostream &OS,
363 const XRayFileHeader &Header) const {
364 OS << "Functions with latencies: " << FunctionLatencies.size() << "\n";
366 // We spend some effort to make the text output more readable, so we do the
367 // following formatting decisions for each of the fields:
369 // - funcid: 32-bit, but we can determine the largest number and be
371 // a minimum of 5 characters, up to 9 characters, right aligned.
372 // - count: 64-bit, but we can determine the largest number and be
374 // a minimum of 5 characters, up to 9 characters, right aligned.
375 // - min, median, 90pct, 99pct, max: double precision, but we want to keep
376 // the values in seconds, with microsecond precision (0.000'001), so we
377 // have at most 6 significant digits, with the whole number part to be
379 // least 1 character. For readability we'll right-align, with full 9
381 // - debug info, function name: we format this as a concatenation of the
382 // debug info and the function name.
384 static constexpr char StatsHeaderFormat[] =
385 "{0,+9} {1,+10} [{2,+9}, {3,+9}, {4,+9}, {5,+9}, {6,+9}] {7,+9}";
386 static constexpr char StatsFormat[] =
387 R"({0,+9} {1,+10} [{2,+9:f6}, {3,+9:f6}, {4,+9:f6}, {5,+9:f6}, {6,+9:f6}] {7,+9:f6})";
388 OS << llvm::formatv(StatsHeaderFormat, "funcid", "count", "min", "med", "90p",
390 << llvm::formatv(" {0,-12}\n", "function");
391 exportStats(Header, [&](int32_t FuncId, size_t Count, const ResultRow &Row) {
392 OS << llvm::formatv(StatsFormat, FuncId, Count, Row.Min, Row.Median,
393 Row.Pct90, Row.Pct99, Row.Max, Row.Sum)
394 << " " << Row.DebugInfo << ": " << Row.Function << "\n";
398 void LatencyAccountant::exportStatsAsCSV(raw_ostream &OS,
399 const XRayFileHeader &Header) const {
400 OS << "funcid,count,min,median,90%ile,99%ile,max,sum,debug,function\n";
401 exportStats(Header, [&](int32_t FuncId, size_t Count, const ResultRow &Row) {
402 OS << FuncId << ',' << Count << ',' << Row.Min << ',' << Row.Median << ','
403 << Row.Pct90 << ',' << Row.Pct99 << ',' << Row.Max << "," << Row.Sum
404 << ",\"" << Row.DebugInfo << "\",\"" << Row.Function << "\"\n";
408 using namespace llvm::xray;
410 static CommandRegistration Unused(&Account, []() -> Error {
411 InstrumentationMap Map;
412 if (!AccountInstrMap.empty()) {
413 auto InstrumentationMapOrError = loadInstrumentationMap(AccountInstrMap);
414 if (!InstrumentationMapOrError)
415 return joinErrors(make_error<StringError>(
416 Twine("Cannot open instrumentation map '") +
417 AccountInstrMap + "'",
418 std::make_error_code(std::errc::invalid_argument)),
419 InstrumentationMapOrError.takeError());
420 Map = std::move(*InstrumentationMapOrError);
424 raw_fd_ostream OS(AccountOutput, EC, sys::fs::OpenFlags::F_Text);
426 return make_error<StringError>(
427 Twine("Cannot open file '") + AccountOutput + "' for writing.", EC);
429 const auto &FunctionAddresses = Map.getFunctionAddresses();
430 symbolize::LLVMSymbolizer::Options Opts(
431 symbolize::FunctionNameKind::LinkageName, true, true, false, "");
432 symbolize::LLVMSymbolizer Symbolizer(Opts);
433 llvm::xray::FuncIdConversionHelper FuncIdHelper(AccountInstrMap, Symbolizer,
435 xray::LatencyAccountant FCA(FuncIdHelper, AccountDeduceSiblingCalls);
436 auto TraceOrErr = loadTraceFile(AccountInput);
439 make_error<StringError>(
440 Twine("Failed loading input file '") + AccountInput + "'",
441 std::make_error_code(std::errc::executable_format_error)),
442 TraceOrErr.takeError());
444 auto &T = *TraceOrErr;
445 for (const auto &Record : T) {
446 if (FCA.accountRecord(Record))
448 for (const auto &ThreadStack : FCA.getPerThreadFunctionStack()) {
449 errs() << "Thread ID: " << ThreadStack.first << "\n";
450 auto Level = ThreadStack.second.size();
451 for (const auto &Entry : llvm::reverse(ThreadStack.second))
452 errs() << "#" << Level-- << "\t"
453 << FuncIdHelper.SymbolOrNumber(Entry.first) << '\n';
455 if (!AccountKeepGoing)
456 return make_error<StringError>(
457 Twine("Failed accounting function calls in file '") + AccountInput +
459 std::make_error_code(std::errc::executable_format_error));
461 switch (AccountOutputFormat) {
462 case AccountOutputFormats::TEXT:
463 FCA.exportStatsAsText(OS, T.getFileHeader());
465 case AccountOutputFormats::CSV:
466 FCA.exportStatsAsCSV(OS, T.getFileHeader());
470 return Error::success();