ft: 部分优化调整

include/eTpf.hrl

@@ -1,56 +1,54 @@
+%% 三目元算符
+-define(IIF(Cond, Then, That), case Cond of true -> Then; _ -> That end).
+
 %% Ignore Msg Hole Name
 -define(eTpfHole, '$eTpfHole').
 
 %% 消息截取默认配置项
 -define(defTcmMap, #{
-   tcmIsCut => false
-   , tcmDepth => 5
-   , tcmListSize => 32
-   , tcmMapSize => 32
-   , tcmTupleSize => 32
-   , tcmBinSize => 128
-   , tcmBitSize => 128 * 8
-   , tcmNestStruct => 6
+	tcmIsCut => false
+	, tcmDepth => 5
+	, tcmListSize => 32
+	, tcmMapSize => 32
+	, tcmTupleSize => 32
+	, tcmBinSize => 128
+	, tcmBitSize => 128 * 8
+	, tcmNestStruct => 6
 }).
 
 -define(eTpfTracerId, eTpfTracerId).
 
 %% 类型声明
--export_type([input/0, userInput/0, traceOpts/0, tracerOpts/0]).
-
--type pattern() :: module() | {app, atom()} | {callback, module(), atom()}.
--type scope() :: {scope, [pid() | port() | all | processes | ports |existing | existing_processes | existing_ports |new | new_processes | new_ports]}.
--type input() :: [pattern() | scope()].
+-export_type([pidPortSpec/0, flagList/0, pattern/0, traceOpts/0, tracerOpts/0]).
 
--type userInput() :: pattern() | input().
+-type pidPortSpec() :: PidPortSpec :: pid() | port()| all | processes | ports| existing | existing_processes | existing_ports| new | new_processes | new_ports.
+-type flagList() :: erlang:trace_flag().
+-type pattern() :: module() | {module(), atom(), atom()} | {app, atom()} | {callback, module(), atom()}.
 
 -type traceOpts() :: #{
-   mode => trace | profile
-   , tracerId => any()
-   , send => boolean()
-   , running => boolean()
-   , process_dump => boolean()      %% 必需的trace选项 为了生成火焰图，我们目前在跟踪时需要使用一个附加选项。此选项将导致将堆栈跟踪信息添加到调用事件。选项为`process_dump`, 并且必须将其设置为`true`。
+	tracerId => any()
+	, stackTc => boolean()      %% 必需的trace选项 为了生成火焰图，我们目前在跟踪时需要使用一个附加选项。此选项将导致将堆栈跟踪信息添加到调用事件。选项为`process_dump`, 并且必须将其设置为`true`。
 }.
 
 -type tracerOpts() :: #{
-   %% tracer for socket
-   port => pos_integer()
-
-   %% tracer for file
-   , fDir      => string()                    %% 文件目录
-   , fBaseName => string()                    %% 文件base file name
-   , fMaxSize => pos_integer()                %% 最大的文件字节数
-   , fMaxMsg => pos_integer()                 %% 最大消息数量
-
-   %% tracer for console
-
-   %% trace msg term cut cfg
-   , tcmIsCut => boolean()                     %% 是否cut trace msg term
-   , tcmDepth => pos_integer()                 %% term cut时保留的最大深度
-   , tcmListSize => pos_integer()              %% List term cut时保留的最大数量
-   , tcmMapSize => pos_integer()               %% Map term cut时保留的最大数量
-   , tcmTupleSize => pos_integer()             %% Tuple term cut时保留的最大数量
-   , tcmBinSize => pos_integer()               %% Bin term cut时保留的最大字节数
-   , tcmBitSize => pos_integer()               %% Bits term cut时保留的最大bit数
-   , tcmNestStruct => pos_integer()            %% 复合结构 term cut时保留的最大数量
+	%% tracer for socket
+	port => pos_integer()
+
+	%% tracer for file
+	, fDir => string()                    %% 文件目录
+	, fBaseName => string()                    %% 文件base file name
+	, fMaxSize => pos_integer()                %% 最大的文件字节数
+	, fMaxMsg => pos_integer()                 %% 最大消息数量
+
+	%% tracer for console
+
+	%% trace msg term cut cfg
+	, tcmIsCut => boolean()                     %% 是否cut trace msg term
+	, tcmDepth => pos_integer()                 %% term cut时保留的最大深度
+	, tcmListSize => pos_integer()              %% List term cut时保留的最大数量
+	, tcmMapSize => pos_integer()               %% Map term cut时保留的最大数量
+	, tcmTupleSize => pos_integer()             %% Tuple term cut时保留的最大数量
+	, tcmBinSize => pos_integer()               %% Bin term cut时保留的最大字节数
+	, tcmBitSize => pos_integer()               %% Bits term cut时保留的最大bit数
+	, tcmNestStruct => pos_integer()            %% 复合结构 term cut时保留的最大数量
 }.

src/eTpf.erl

@@ -3,8 +3,16 @@
 -include("eTpf.hrl").
 
 -export([
+	%% trace函数
+	trace/1
+	, trace/2
+	, trace/3
+	, trace/6
+	, stop/0
+	, stop/1
+
 	%% 记录函数
-	sts/0             %% tpTracerShell
+	, sts/0             %% tpTracerShell
 	, stl/0           %% tpTracerLog
 	, stf/0           %% tpTracerFile
 
@@ -30,17 +38,9 @@
 
 	, gcPfs/1
 	, gcPfm/2
-
-	%% trace函数
-	, trace/1
-	, trace/2
-	, trace/3
-	, trace/4
-	, stop/0
-	, stop/1
 ]).
 
--compile([export_all]).
+-compile([export_all, nowarn_export_all]).
 
 sts() ->
 	trace([{app, eSync}, {scope, [whereis(esSyncSrv)]}]).
@@ -106,36 +106,37 @@ gcPfs(InputFile) ->
 gcPfm(InputFiles, Cwd) ->
 	tpGcColl:pfm(InputFiles, Cwd).
 
--spec trace(userInput()) -> ok.
-trace(Input) ->
-	trace(Input, tpTracerShell).
+-spec trace(pidPortSpec()) -> ok.
+trace(PidPortSpec) ->
+	trace(PidPortSpec, tpTracerShell).
 
--spec trace(userInput(), module()) -> ok.
-trace(Input, TracerMod) ->
-	trace(Input, TracerMod, #{}, #{}).
+-spec trace(pidPortSpec(), module()) -> ok.
+trace(PidPortSpec, TracerMod) ->
+	trace(PidPortSpec, TracerMod, #{}).
 
--spec trace(userInput(), module(), tracerOpts()) -> ok.
-trace(Input, TracerMod, TracerOpts) ->
-	trace(Input, TracerMod, TracerOpts, #{}).
+-spec trace(pidPortSpec(), module(), tracerOpts()) -> ok.
+trace(PidPortSpec, TracerMod, TracerOpts) ->
+	trace(PidPortSpec, [all], [{'_', '_', '_'}], TracerMod, TracerOpts, #{}).
 
--spec trace(userInput(), module(), tracerOpts(), traceOpts()) -> ok.
-trace(Input, TracerMod, TracerOpts, TraceOpts) when is_list(Input) ->
-	InputList = case is_list(Input) of true -> Input; _ -> [Input] end,
-	doTrace(InputList, TracerMod, TracerOpts, TraceOpts).
+-spec trace(pidPortSpec(), flagList(), pattern(), module(), tracerOpts(), traceOpts()) -> ok.
+trace(PidPortSpec, FlagList, TraceMods, TracerMod, TracerOpts, TraceOpts) ->
+	InputList = case is_list(PidPortSpec) of true -> PidPortSpec; _ -> [PidPortSpec] end,
+	doTrace(InputList, FlagList, TraceMods, TracerMod, TracerOpts, TraceOpts).
 
 doTrace(PidPortSpec, FlagList, TraceMods, TracerMod, TracerOpts, TraceOpts) ->
 	_ = application:ensure_all_started(eTpf),
 
 	TracerId = maps:get(tracerId, TraceOpts, ?eTpfTracerId),
-	case supervisor:start_child(eTpf_sup, #{
-		id => TracerId,
-		start => {TracerMod, start_link, [TracerOpts]},
-		restart => temporary,
-		type => worker,
-		shutdown => infinity
-	}) of
+	TracerSpec = #{
+		id => TracerId
+		, start => {TracerMod, start_link, [TracerOpts]}
+		, restart => temporary
+		, type => worker
+		, shutdown => infinity
+	},
+
+	case supervisor:start_child(eTpf_sup, TracerSpec) of
 		{ok, TracerPid} ->
-			_Mode = maps:get(mode, TraceOpts, trace),
 			LastTraceMods = flattenMods(TraceMods, []),
 			traceInput(PidPortSpec, FlagList, LastTraceMods, TracerPid, TraceOpts),
 			ok;
@@ -144,48 +145,38 @@ doTrace(PidPortSpec, FlagList, TraceMods, TracerMod, TracerOpts, TraceOpts) ->
 	end.
 
 flattenMods([], Acc) ->
-	case lists:flatten(Acc) of
-		[] ->
-			['_'];
-		AllMods ->
-			AllMods
-	end;
+	AllMFA = lists:flatten(Acc),
+	?IIF(AllMFA == [], [{'_', '_', '_'}], AllMFA);
 flattenMods([{callback, Mod, Fun} | Tail], Acc) when is_atom(Mod), is_atom(Fun) ->
 	Input = flattenMods(Mod:Fun(), []),
 	flattenMods(Tail, [Input | Acc]);
 flattenMods([{app, App} | Tail], Acc) when is_atom(App) ->
 	_ = application:load(App),
 	{ok, Mods} = application:get_key(App, modules),
-	flattenMods(Tail, [Mods | Acc]);
-flattenMods([Input | Tail], Acc) ->
-	flattenMods(Tail, [Input | Acc]).
-
-traceInput(PidPortSpec, FlagList, TraceMods, TracerState, TraceOpts) ->
+	MFA = [{OneMod, '_', '_'} || OneMod <- Mods],
+	flattenMods(Tail, [MFA | Acc]);
+flattenMods([{_, _, _} = MFA | Tail], Acc) ->
+	flattenMods(Tail, [MFA | Acc]);
+flattenMods([OneMod | Tail], Acc) ->
+	flattenMods(Tail, [{OneMod, '_', '_'} | Acc]).
+
+traceInput(PidPortSpec, FlagList, TraceMFAs, TracerState, TraceOpts) ->
 	% 每个进程只能由一个tracer进行跟踪。因此，跟踪已跟踪进程会失败。
-	%% @todo It might be useful to count the number of sends
-	%% or receives a function does.
-	%ExtraFlags = [running || maps:get(running, TraceOpts, false)] ++ [send || maps:get(send, TraceOpts, false)],
 	[
 		begin
-		%erlang:trace(PidPortSpec, true, [call, procs, arity, return_to, set_on_spawn, {tracer, tpTracerNif, TracerState} | ExtraFlags])
 			erlang:trace(OnePidPortSpec, true, [{tracer, tpTracerNif, TracerState} | FlagList])
 		end || OnePidPortSpec <- PidPortSpec
 	],
 
-	MatchSpec =
-		case TraceOpts of
-			#{process_dump := true} ->
-				[{'_', [], [{message, {process_dump}}]}];
-			_ ->
-				true
-		end,
+	PtFlags = maps:get(FlagList, TraceOpts, [local]),
+	MatchSpec = ?IIF(maps:get(stackTc, TraceOpts, false), [{'_', [], [{message, {process_dump}}]}], true),
 
 	[
-	   begin
+		begin
 		%% The module must be loaded before we attempt to trace it.
-			_ = code:ensure_loaded(OneTraceMods),
-			io:format("IMY******************  ~p~n  ", [OneTraceMods]),
-			_ = erlang:trace_pattern({OneTraceMods, '_', '_'}, [{'_', [], [{message, {process_dump}}]}], [local])
-		end || OneTraceMods <- TraceMods
+			_ = code:ensure_loaded(M),
+			io:format("IMY******************  ~p~n  ", [OneTraceMFA]),
+			_ = erlang:trace_pattern(OneTraceMFA, MatchSpec, PtFlags)
+		end || {M, _F, _A} = OneTraceMFA <- TraceMFAs
 	],
 	ok.

src/tc/tpTc.erl

@@ -0,0 +1,216 @@
+-module(tpTc).
+
+-compile(inline).
+-compile({inline_size, 128}).
+
+-export([
+   tc/1
+   , tc/2
+   , tc/3
+   , ts/4
+   , tm/5
+   , cvrTimeUnit/3
+   , test/1
+]).
+
+%% Measure the execution time (in nanoseconds) for Fun().
+-spec tc(Fun :: function()) -> {Time :: integer(), Value :: term()}.
+tc(F) ->
+   T1 = erlang:monotonic_time(),
+   Val = F(),
+   T2 = erlang:monotonic_time(),
+   Time = cvrTimeUnit(T2 - T1, native, nanosecond),
+   {Time, Val}.
+
+%% Measure the execution time (in nanoseconds) for Fun(Args).
+-spec tc(Fun :: function(), Arguments :: [term()]) -> {Time :: integer(), Value :: term()}.
+tc(F, A) ->
+   T1 = erlang:monotonic_time(),
+   Val = apply(F, A),
+   T2 = erlang:monotonic_time(),
+   Time = cvrTimeUnit(T2 - T1, native, nanosecond),
+   {Time, Val}.
+
+%% Measure the execution time (in nanoseconds) for an MFA.
+-spec tc(Module :: module(), Function :: atom(), Arguments :: [term()]) -> {Time :: integer(), Value :: term()}.
+tc(M, F, A) ->
+   T1 = erlang:monotonic_time(),
+   Val = apply(M, F, A),
+   T2 = erlang:monotonic_time(),
+   Time = cvrTimeUnit(T2 - T1, native, nanosecond),
+   {Time, Val}.
+
+-spec cvrTimeUnit(Time :: integer(), FromUnit :: erlang:time_unit(), ToUnit :: erlang:time_unit()) -> ConvertedTime :: integer().
+cvrTimeUnit(Time, FromUnit, ToUnit) ->
+   try
+      FU =
+         case FromUnit of
+            native -> erts_internal:time_unit();
+            perf_counter -> erts_internal:perf_counter_unit();
+            nanosecond -> 1000 * 1000 * 1000;
+            microsecond -> 1000 * 1000;
+            millisecond -> 1000;
+            second -> 1
+         end,
+      TU =
+         case ToUnit of
+            native -> erts_internal:time_unit();
+            perf_counter -> erts_internal:perf_counter_unit();
+            nanosecond -> 1000 * 1000 * 1000;
+            microsecond -> 1000 * 1000;
+            millisecond -> 1000;
+            second -> 1
+         end,
+      case Time < 0 of
+         true -> (TU * Time - (FU - 1)) div FU;
+         _ -> TU * Time div FU
+      end
+   catch
+      _ : _ ->
+         erlang:error(badarg, [Time, FromUnit, ToUnit])
+   end.
+
+%% 单进程循环测试：LoopTimes是循环次数
+%% utTc:ts(LoopTimes, Module, Function, ArgsList).
+%% 多进程并发测试：SpawnProcessesCount是并发的进程数 LoopTimes是循环次数
+%% utTc:tm(ProcessesCount, LoopTimes, Module, Function, ArgsList).
+
+doTc(M, F, A) ->
+   T1 = erlang:monotonic_time(),
+   apply(M, F, A),
+   T2 = erlang:monotonic_time(),
+   cvrTimeUnit(T2 - T1, native, nanosecond).
+
+distribution(List, Aver) ->
+   distribution(List, Aver, 0, 0).
+distribution([H | T], Aver, Greater, Less) ->
+   case H > Aver of
+      true ->
+         distribution(T, Aver, Greater + 1, Less);
+      false ->
+         distribution(T, Aver, Greater, Less + 1)
+   end;
+distribution([], _Aver, Greater, Less) ->
+   {Greater, Less}.
+
+%% ===================================================================
+%% test: one process test N times
+%% ===================================================================
+ts(LoopTime, M, F, A) ->
+   {Max, Min, Sum, Aver, Greater, Less} = loopTs(LoopTime, M, F, A, LoopTime, 0, 0, 0, []),
+   io:format("=====================~n"),
+   io:format("execute Args:~p~n", [A]),
+   io:format("execute Fun :~p~n", [F]),
+   io:format("execute Mod :~p~n", [M]),
+   io:format("execute LoopTime:~p~n", [LoopTime]),
+   io:format("MaxTime: ~10s(ns) ~10s(s)~n", [integer_to_binary(Max), float_to_binary(Max / 1000000000, [{decimals, 6}, compact])]),
+   io:format("MinTime: ~10s(ns) ~10s(s)~n", [integer_to_binary(Min), float_to_binary(Min / 1000000000, [{decimals, 6}, compact])]),
+   io:format("SumTime: ~10s(ns) ~10s(s)~n", [integer_to_binary(Sum), float_to_binary(Sum / 1000000000, [{decimals, 6}, compact])]),
+   io:format("AvgTime: ~10s(ns) ~10s(s)~n", [float_to_binary(Aver, [{decimals, 6}, compact]), float_to_binary(Aver / 1000000000, [{decimals, 6}, compact])]),
+   io:format("Grar   : ~10s(cn) ~10s(~s)~n", [integer_to_binary(Greater), float_to_binary(Greater / LoopTime, [{decimals, 2}]), <<"%">>]),
+   io:format("Less   : ~10s(cn) ~10s(~s)~n", [integer_to_binary(Less), float_to_binary(Less / LoopTime, [{decimals, 2}]), <<"%">>]),
+   io:format("=====================~n").
+
+
+loopTs(0, _M, _F, _A, LoopTime, Max, Min, Sum, List) ->
+   Aver = Sum / LoopTime,
+   {Greater, Less} = distribution(List, Aver),
+   {Max, Min, Sum, Aver, Greater, Less};
+loopTs(Index, M, F, A, LoopTime, Max, Min, Sum, List) ->
+   Nanosecond = doTc(M, F, A),
+   NewSum = Sum + Nanosecond,
+   if
+      Max == 0 ->
+         NewMax = NewMin = Nanosecond;
+      Max < Nanosecond ->
+         NewMax = Nanosecond,
+         NewMin = Min;
+      Min > Nanosecond ->
+         NewMax = Max,
+         NewMin = Nanosecond;
+      true ->
+         NewMax = Max,
+         NewMin = Min
+   end,
+   loopTs(Index - 1, M, F, A, LoopTime, NewMax, NewMin, NewSum, [Nanosecond | List]).
+
+
+%% ===================================================================
+%% Concurrency test: N processes each test one time
+%% ===================================================================
+
+tm(ProcCnt, LoopTime, M, F, A) ->
+   loopSpawn(ProcCnt, M, F, A, self(), LoopTime),
+   {Max, Min, Sum, Aver, Greater, Less} = collector(ProcCnt, 0, 0, 0, ProcCnt, []),
+   io:format("=====================~n"),
+   io:format("execute Args:~p~n", [A]),
+   io:format("execute Fun :~p~n", [F]),
+   io:format("execute Mod :~p~n", [M]),
+   io:format("execute LoopTime:~p~n", [LoopTime]),
+   io:format("execute ProcCnts:~p~n", [ProcCnt]),
+   io:format("MaxTime: ~10s(ns) ~10s(s)~n", [integer_to_binary(Max), float_to_binary(Max / 1000000000, [{decimals, 6}, compact])]),
+   io:format("MinTime: ~10s(ns) ~10s(s)~n", [integer_to_binary(Min), float_to_binary(Min / 1000000000, [{decimals, 6}, compact])]),
+   io:format("SumTime: ~10s(ns) ~10s(s)~n", [integer_to_binary(Sum), float_to_binary(Sum / 1000000000, [{decimals, 6}, compact])]),
+   io:format("AvgTime: ~10s(ns) ~10s(s)~n", [float_to_binary(Aver, [{decimals, 6}, compact]), float_to_binary(Aver / 1000000000, [{decimals, 6}, compact])]),
+   io:format("Grar   : ~10s(cn) ~10s(~s)~n", [integer_to_binary(Greater), float_to_binary(Greater / LoopTime, [{decimals, 2}]), <<"%">>]),
+   io:format("Less   : ~10s(cn) ~10s(~s)~n", [integer_to_binary(Less), float_to_binary(Less / LoopTime, [{decimals, 2}]), <<"%">>]),
+   io:format("=====================~n").
+
+
+loopSpawn(0, _, _, _, _, _) ->
+   ok;
+loopSpawn(ProcCnt, M, F, A, CollectorPid, LoopTime) ->
+   spawn_link(fun() -> worker(LoopTime, M, F, A, CollectorPid) end),
+   loopSpawn(ProcCnt - 1, M, F, A, CollectorPid, LoopTime).
+
+collector(0, Max, Min, Sum, ProcCnt, List) ->
+   Aver = Sum / ProcCnt,
+   {Greater, Less} = distribution(List, Aver),
+   {Max, Min, Sum, Aver, Greater, Less};
+collector(Index, Max, Min, Sum, ProcCnt, List) ->
+   receive
+      {result, Nanosecond} ->
+         NewSum = Sum + Nanosecond,
+         if
+            Max == 0 ->
+               NewMax = NewMin = Nanosecond;
+            Max < Nanosecond ->
+               NewMax = Nanosecond,
+               NewMin = Min;
+            Min > Nanosecond ->
+               NewMax = Max,
+               NewMin = Nanosecond;
+            true ->
+               NewMax = Max,
+               NewMin = Min
+         end,
+         collector(Index - 1, NewMax, NewMin, NewSum, ProcCnt, [Nanosecond | List])
+   after 1800000 ->
+      io:format("execute time out~n"),
+      ok
+   end.
+
+worker(LoopTime, M, F, A, CollectorPid) ->
+   SumTime = loopTm(LoopTime, M, F, A, 0),
+   CollectorPid ! {result, SumTime}.
+
+loopTm(0, _, _, _, SumTime) ->
+   SumTime;
+loopTm(LoopTime, M, F, A, SumTime) ->
+   Microsecond = doTc(M, F, A),
+   loopTm(LoopTime - 1, M, F, A, SumTime + Microsecond).
+
+test(N) ->
+   M1 = erlang:monotonic_time(),
+   timer:sleep(N),
+   M2 = erlang:monotonic_time(),
+   Time = cvrTimeUnit(M2 - M1, native, nanosecond),
+   io:format("IMY******************111 ~p~n", [Time]),
+
+   S1 = erlang:system_time(nanosecond),
+   timer:sleep(N),
+   S2 = erlang:system_time(nanosecond),
+   io:format("IMY******************222 ~p~n", [S2 - S1]).
+
+
+