Merge pull request #4300 from neobrain/feature_profiler_tracy

Profiler: Add Tracy backend

.gitmodules

@@ -47,3 +47,6 @@
 [submodule "External/jemalloc_glibc"]
 	path = External/jemalloc_glibc
 	url = https://github.com/FEX-Emu/jemalloc.git
+[submodule "External/tracy"]
+	path = External/tracy
+	url = https://github.com/wolfpld/tracy

CMakeLists.txt

@@ -30,7 +30,7 @@ option(ENABLE_VIXL_SIMULATOR "Enable use of VIXL simulator for emulation (only u
 option(ENABLE_VIXL_DISASSEMBLER "Enables debug disassembler output with VIXL" FALSE)
 option(USE_LEGACY_BINFMTMISC "Uses legacy method of setting up binfmt_misc" FALSE)
 option(ENABLE_FEXCORE_PROFILER "Enables use of the FEXCore timeline profiling capabilities" FALSE)
-set (FEXCORE_PROFILER_BACKEND "gpuvis" CACHE STRING "Set which backend you want to use for the FEXCore profiler")
+set (FEXCORE_PROFILER_BACKEND "gpuvis" CACHE STRING "Set which backend to use for the FEXCore profiler (gpuvis, tracy)")
 option(ENABLE_GLIBC_ALLOCATOR_HOOK_FAULT "Enables glibc memory allocation hooking with fault for CI testing")
 option(USE_PDB_DEBUGINFO "Builds debug info in PDB format" FALSE)
 
@@ -61,6 +61,22 @@ if (ENABLE_FEXCORE_PROFILER)
 
   if (FEXCORE_PROFILER_BACKEND STREQUAL "GPUVIS")
     add_definitions(-DFEXCORE_PROFILER_BACKEND=1)
+  elseif (FEXCORE_PROFILER_BACKEND STREQUAL "TRACY")
+    add_definitions(-DFEXCORE_PROFILER_BACKEND=2)
+    add_definitions(-DTRACY_ENABLE=1)
+    # Required so that Tracy will only start in the selected guest application
+    add_definitions(-DTRACY_MANUAL_LIFETIME=1)
+    add_definitions(-DTRACY_DELAYED_INIT=1)
+    # This interferes with FEX's signal handling
+    add_definitions(-DTRACY_NO_CRASH_HANDLER=1)
+    # Tracy can gather call stack samples in regular intervals, but this
+    # isn't useful for us since it would usually sample opaque JIT code
+    add_definitions(-DTRACY_NO_SAMPLING=1)
+    # This pulls in libbacktrace which allocators in global constructors (before FEX can set up its allocator hooks)
+    add_definitions(-DTRACY_NO_CALLSTACK=1)
+    if (MINGW_BUILD)
+      message(FATAL_ERROR "Tracy profiler not supported")
+    endif()
   else()
     message(FATAL_ERROR "Unknown FEXCore profiler backend ${FEXCORE_PROFILER_BACKEND}")
   endif()
@@ -270,6 +286,10 @@ if (BUILD_TESTS OR ENABLE_VIXL_DISASSEMBLER OR ENABLE_VIXL_SIMULATOR)
   include_directories(SYSTEM External/vixl/src/)
 endif()
 
+if (ENABLE_FEXCORE_PROFILER AND FEXCORE_PROFILER_BACKEND STREQUAL "TRACY")
+  add_subdirectory(External/tracy)
+endif()
+
 if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
   # This means we were attempted to get compiled with GCC
   message(FATAL_ERROR "FEX doesn't support getting compiled with GCC!")

FEXCore/Source/CMakeLists.txt

@@ -338,6 +338,10 @@ add_library(FEXCore_Base STATIC ${FEXCORE_BASE_SRCS})
 target_link_libraries(FEXCore_Base ${LIBS})
 AddDefaultOptionsToTarget(FEXCore_Base)
 
+if (ENABLE_FEXCORE_PROFILER AND FEXCORE_PROFILER_BACKEND STREQUAL "TRACY")
+  target_link_libraries(FEXCore_Base TracyClient)
+endif()
+
 function(AddObject Name Type)
   add_library(${Name} ${Type} ${SRCS})
 

FEXCore/Source/Interface/Core/Core.cpp

@@ -473,6 +473,7 @@ void ContextImpl::DestroyThread(FEXCore::Core::InternalThreadState* Thread) {
 void ContextImpl::UnlockAfterFork(FEXCore::Core::InternalThreadState* LiveThread, bool Child) {
   Allocator::UnlockAfterFork(LiveThread, Child);
 
+  Profiler::PostForkAction(Child);
   if (Child) {
     CodeInvalidationMutex.StealAndDropActiveLocks();
     if (Config.StrictInProcessSplitLocks) {

FEXCore/Source/Utils/Profiler.cpp

@@ -1,24 +1,22 @@
 // SPDX-License-Identifier: MIT
-#include <array>
 #include <cstdint>
 #include <fcntl.h>
-#include <limits.h>
 #ifndef _WIN32
 #include <linux/magic.h>
 #include <sys/stat.h>
 #include <sys/vfs.h>
-#include <time.h>
 #endif
 
 #include <FEXCore/Utils/LogManager.h>
 #include <FEXCore/Utils/Profiler.h>
 #include <FEXCore/fextl/fmt.h>
 #include <FEXCore/fextl/string.h>
 
-#define BACKEND_OFF 0
-#define BACKEND_GPUVIS 1
-
 #ifdef ENABLE_FEXCORE_PROFILER
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_GPUVIS
+#include <array>
+#include <limits.h>
+#include <time.h>
 #ifndef _WIN32
 static inline uint64_t GetTime() {
   // We want the time in the least amount of overhead possible
@@ -49,7 +47,6 @@ static inline uint64_t GetTime() {
 
 #endif
 
-#if FEXCORE_PROFILER_BACKEND == BACKEND_GPUVIS
 namespace FEXCore::Profiler {
 ProfilerBlock::ProfilerBlock(std::string_view const Format)
   : DurationBegin {GetTime()}
@@ -114,35 +111,125 @@ void TraceObject(std::string_view const Format) {
   }
 }
 } // namespace GPUVis
+#elif FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+#include "tracy/Tracy.hpp"
+namespace Tracy {
+static int EnableAfterFork = 0;
+static bool Enable = false;
+
+void Init(std::string_view ProgramName, std::string_view ProgramPath) {
+  const char* ProfileTargetName = getenv("FEX_PROFILE_TARGET_NAME"); // Match by application name
+  const char* ProfileTargetPath = getenv("FEX_PROFILE_TARGET_PATH"); // Match by path suffix
+  const char* WaitForFork = getenv("FEX_PROFILE_WAIT_FOR_FORK");     // Don't enable profiling until the process forks N times
+  bool Matched = (ProfileTargetName && ProgramName == ProfileTargetName) || (ProfileTargetPath && ProgramPath.ends_with(ProfileTargetPath));
+  if (Matched && WaitForFork) {
+    EnableAfterFork = std::atoi(WaitForFork);
+  }
+  Enable = Matched && !EnableAfterFork;
+  if (Enable) {
+    tracy::StartupProfiler();
+    LogMan::Msg::IFmt("Tracy profiling started");
+  } else if (EnableAfterFork) {
+    LogMan::Msg::IFmt("Tracy profiling will start after fork");
+  }
+}
+
+void PostForkAction(bool IsChild) {
+  if (Enable) {
+    // Tracy does not support multiprocess profiling
+    LogMan::Msg::EFmt("Warning: Profiling a process with forks is not supported. Set the environment variable "
+                      "FEX_PROFILE_WAIT_FOR_FORK=<n> to start profiling after the n-th fork.");
+  }
+
+  if (IsChild) {
+    Enable = false;
+    return;
+  }
+
+  if (EnableAfterFork > 1) {
+    --EnableAfterFork;
+    LogMan::Msg::IFmt("Tracy profiling will start after {} forks", EnableAfterFork);
+  } else if (EnableAfterFork == 1) {
+    Enable = true;
+    EnableAfterFork = 0;
+    tracy::StartupProfiler();
+    LogMan::Msg::IFmt("Tracy profiling started");
+  }
+}
+
+void Shutdown() {
+  if (Tracy::Enable) {
+    LogMan::Msg::IFmt("Stopping Tracy profiling");
+    tracy::ShutdownProfiler();
+  }
+}
+
+void TraceObject(std::string_view const Format, uint64_t Duration) {}
+
+void TraceObject(std::string_view const Format) {
+  if (Tracy::Enable) {
+    TracyMessage(Format.data(), Format.size());
+  }
+}
+} // namespace Tracy
 #else
 #error Unknown profiler backend
 #endif
 #endif
 
 namespace FEXCore::Profiler {
+
 #ifdef ENABLE_FEXCORE_PROFILER
-void Init() {
-#if FEXCORE_PROFILER_BACKEND == BACKEND_GPUVIS
+void Init(std::string_view ProgramName, std::string_view ProgramPath) {
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_GPUVIS
   GPUVis::Init();
+#elif FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+  Tracy::Init(ProgramName, ProgramPath);
+#endif
+}
+
+void PostForkAction(bool IsChild) {
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+  Tracy::PostForkAction(IsChild);
+#endif
+}
+
+bool IsActive() {
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_GPUVIS
+  // Always active
+  return true;
+#elif FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+  // Active if previously enabled
+  if (Tracy::Enable) {
+    LogMan::Msg::EFmt("PROFILE ENABLED");
+  }
+  return Tracy::Enable;
 #endif
 }
 
 void Shutdown() {
-#if FEXCORE_PROFILER_BACKEND == BACKEND_GPUVIS
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_GPUVIS
   GPUVis::Shutdown();
+#elif FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+  Tracy::Shutdown();
 #endif
 }
 
 void TraceObject(std::string_view const Format, uint64_t Duration) {
-#if FEXCORE_PROFILER_BACKEND == BACKEND_GPUVIS
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_GPUVIS
   GPUVis::TraceObject(Format, Duration);
+#elif FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+  Tracy::TraceObject(Format, Duration);
 #endif
 }
 
 void TraceObject(std::string_view const Format) {
-#if FEXCORE_PROFILER_BACKEND == BACKEND_GPUVIS
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_GPUVIS
   GPUVis::TraceObject(Format);
+#elif FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+  Tracy::TraceObject(Format);
 #endif
 }
+
 #endif
 } // namespace FEXCore::Profiler

FEXCore/include/FEXCore/Utils/Profiler.h

@@ -10,6 +10,14 @@
 
 #include <FEXCore/Utils/CompilerDefs.h>
 
+#define FEXCORE_PROFILER_BACKEND_OFF 0
+#define FEXCORE_PROFILER_BACKEND_GPUVIS 1
+#define FEXCORE_PROFILER_BACKEND_TRACY 2
+
+#if defined(ENABLE_FEXCORE_PROFILER) && FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+#include "tracy/Tracy.hpp"
+#endif
+
 namespace FEXCore::Profiler {
 // FEXCore live-stats
 constexpr uint8_t STATS_VERSION = 2;
@@ -69,11 +77,23 @@ static inline uint64_t GetCycleCounter() {
 }
 #endif
 
-FEX_DEFAULT_VISIBILITY void Init();
+FEX_DEFAULT_VISIBILITY void Init(std::string_view ProgramName, std::string_view ProgramPath);
+FEX_DEFAULT_VISIBILITY void PostForkAction(bool IsChild);
+FEX_DEFAULT_VISIBILITY bool IsActive();
 FEX_DEFAULT_VISIBILITY void Shutdown();
 FEX_DEFAULT_VISIBILITY void TraceObject(std::string_view const Format);
 FEX_DEFAULT_VISIBILITY void TraceObject(std::string_view const Format, uint64_t Duration);
 
+#define UniqueScopeName2(name, line) name##line
+#define UniqueScopeName(name, line) UniqueScopeName2(name, line)
+
+// Declare an instantaneous profiler event.
+#define FEXCORE_PROFILE_INSTANT(name) FEXCore::Profiler::TraceObject(name)
+
+#if FEXCORE_PROFILER_BACKEND == FEXCORE_PROFILER_BACKEND_TRACY
+// Declare a scoped profile block variable with a fixed name.
+#define FEXCORE_PROFILE_SCOPED(name) ZoneNamedN(___tracy_scoped_zone, name, ::FEXCore::Profiler::IsActive())
+#else
 // A class that follows scoping rules to generate a profile duration block
 class ProfilerBlock final {
 public:
@@ -86,14 +106,9 @@ class ProfilerBlock final {
   std::string_view const Format;
 };
 
-#define UniqueScopeName2(name, line) name##line
-#define UniqueScopeName(name, line) UniqueScopeName2(name, line)
-
-// Declare an instantaneous profiler event.
-#define FEXCORE_PROFILE_INSTANT(name) FEXCore::Profiler::TraceObject(name)
-
 // Declare a scoped profile block variable with a fixed name.
 #define FEXCORE_PROFILE_SCOPED(name) FEXCore::Profiler::ProfilerBlock UniqueScopeName(ScopedBlock_, __LINE__)(name)
+#endif
 
 template<typename T, size_t FlatOffset = 0>
 class AccumulationBlock final {
@@ -127,7 +142,9 @@ class AccumulationBlock final {
 
 #else
 [[maybe_unused]]
-static void Init() {}
+static void Init(std::string_view ProgramName, std::string_view ProgramPath) {}
+[[maybe_unused]]
+static void PostForkAction(bool IsChild) {}
 [[maybe_unused]]
 static void Shutdown() {}
 [[maybe_unused]]

Source/Tools/FEXLoader/FEXLoader.cpp

@@ -389,7 +389,6 @@ int main(int argc, char** argv, char** const envp) {
     std::this_thread::sleep_for(std::chrono::seconds(StartupSleep()));
   }
 
-  FEXCore::Profiler::Init();
   FEXCore::Telemetry::Initialize();
 
   if (!LDPath().empty() && Program.ProgramPath.starts_with(LDPath())) {
@@ -493,6 +492,8 @@ int main(int argc, char** argv, char** const envp) {
     free(data);
   }
 
+  FEXCore::Profiler::Init(Program.ProgramName, Program.ProgramPath);
+
   // System allocator is now system allocator or FEX
   FEXCore::Context::InitializeStaticTables(Loader.Is64BitMode() ? FEXCore::Context::MODE_64BIT : FEXCore::Context::MODE_32BIT);
 

Source/Windows/ARM64EC/Module.cpp

@@ -539,7 +539,7 @@ NTSTATUS ProcessInit() {
   // Not applicable to Windows
   FEXCore::Config::EraseSet(FEXCore::Config::ConfigOption::CONFIG_TSOAUTOMIGRATION, "0");
 
-  FEXCore::Profiler::Init();
+  FEXCore::Profiler::Init("", "");
 
   FEXCore::Context::InitializeStaticTables(FEXCore::Context::MODE_64BIT);
 

Source/Windows/WOW64/Module.cpp

@@ -461,7 +461,7 @@ void BTCpuProcessInit() {
   // Not applicable to Windows
   FEXCore::Config::EraseSet(FEXCore::Config::ConfigOption::CONFIG_TSOAUTOMIGRATION, "0");
 
-  FEXCore::Profiler::Init();
+  FEXCore::Profiler::Init("", "");
 
   FEXCore::Context::InitializeStaticTables(FEXCore::Context::MODE_32BIT);