Send callstacks before sending events they belong to.

TracyD3D12.hpp

@@ -311,6 +311,10 @@ namespace tracy
 			m_queryId = ctx->NextQueryId();
 			cmdList->EndQuery(ctx->m_queryHeap.Get(), D3D12_QUERY_TYPE_TIMESTAMP, m_queryId);
 
+#if defined(TRACY_HAS_CALLSTACK) && defined(TRACY_CALLSTACK)
+			GetProfiler().SendCallstack(TRACY_CALLSTACK);
+#endif
+
 			auto* item = Profiler::QueueSerial();
 #if defined(TRACY_HAS_CALLSTACK) && defined(TRACY_CALLSTACK)
 			MemWrite(&item->hdr.type, QueueType::GpuZoneBeginCallstackSerial);
@@ -324,10 +328,6 @@ namespace tracy
 			MemWrite(&item->gpuZoneBegin.context, ctx->GetId());
 
 			Profiler::QueueSerialFinish();
-
-#if defined(TRACY_HAS_CALLSTACK) && defined(TRACY_CALLSTACK)
-			GetProfiler().SendCallstack(TRACY_CALLSTACK);
-#endif
 		}
 
 		tracy_force_inline ~D3D12ZoneScope()

TracyLua.hpp

@@ -197,6 +197,13 @@ static inline int LuaZoneBeginS( lua_State* L )
     if( !GetLuaZoneState().active ) return 0;
 #endif
 
+#ifdef TRACY_CALLSTACK
+    const uint32_t depth = TRACY_CALLSTACK;
+#else
+    const auto depth = uint32_t( lua_tointeger( L, 1 ) );
+#endif
+    SendLuaCallstack( L, depth );
+
     TracyLfqPrepare( QueueType::ZoneBeginAllocSrcLocCallstack );
     lua_Debug dbg;
     lua_getstack( L, 1, &dbg );
@@ -206,13 +213,6 @@ static inline int LuaZoneBeginS( lua_State* L )
     MemWrite( &item->zoneBegin.srcloc, srcloc );
     TracyLfqCommit;
 
-#ifdef TRACY_CALLSTACK
-    const uint32_t depth = TRACY_CALLSTACK;
-#else
-    const auto depth = uint32_t( lua_tointeger( L, 1 ) );
-#endif
-    SendLuaCallstack( L, depth );
-
     return 0;
 }
 
@@ -225,6 +225,13 @@ static inline int LuaZoneBeginNS( lua_State* L )
     if( !GetLuaZoneState().active ) return 0;
 #endif
 
+#ifdef TRACY_CALLSTACK
+    const uint32_t depth = TRACY_CALLSTACK;
+#else
+    const auto depth = uint32_t( lua_tointeger( L, 2 ) );
+#endif
+    SendLuaCallstack( L, depth );
+
     TracyLfqPrepare( QueueType::ZoneBeginAllocSrcLocCallstack );
     lua_Debug dbg;
     lua_getstack( L, 1, &dbg );
@@ -236,13 +243,6 @@ static inline int LuaZoneBeginNS( lua_State* L )
     MemWrite( &item->zoneBegin.srcloc, srcloc );
     TracyLfqCommit;
 
-#ifdef TRACY_CALLSTACK
-    const uint32_t depth = TRACY_CALLSTACK;
-#else
-    const auto depth = uint32_t( lua_tointeger( L, 2 ) );
-#endif
-    SendLuaCallstack( L, depth );
-
     return 0;
 }
 #endif

TracyOpenCL.hpp

@@ -246,6 +246,8 @@ namespace tracy {
 
             m_beginQueryId = ctx->NextQueryId(EventInfo{ nullptr, EventPhase::Begin });
 
+            GetProfiler().SendCallstack(depth);
+
             auto item = Profiler::QueueSerial();
             MemWrite(&item->hdr.type, QueueType::GpuZoneBeginCallstackSerial);
             MemWrite(&item->gpuZoneBegin.cpuTime, Profiler::GetTime());
@@ -254,8 +256,6 @@ namespace tracy {
             MemWrite(&item->gpuZoneBegin.queryId, (uint16_t)m_beginQueryId);
             MemWrite(&item->gpuZoneBegin.context, ctx->GetId());
             Profiler::QueueSerialFinish();
-
-            GetProfiler().SendCallstack(depth);
         }
 
         tracy_force_inline void SetEvent(cl_event event)

TracyOpenGL.hpp

@@ -215,6 +215,8 @@ public:
         const auto queryId = GetGpuCtx().ptr->NextQueryId();
         glQueryCounter( GetGpuCtx().ptr->TranslateOpenGlQueryId( queryId ), GL_TIMESTAMP );
 
+        GetProfiler().SendCallstack( depth );
+
         const auto thread = GetThreadHandle();
         TracyLfqPrepare( QueueType::GpuZoneBeginCallstack );
         MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
@@ -223,8 +225,6 @@ public:
         MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
         MemWrite( &item->gpuZoneBegin.context, GetGpuCtx().ptr->GetId() );
         TracyLfqCommit;
-
-        GetProfiler().SendCallstack( depth );
     }
 
     tracy_force_inline ~GpuCtxScope()

TracyVulkan.hpp

@@ -342,6 +342,8 @@ public:
         const auto queryId = ctx->NextQueryId();
         vkCmdWriteTimestamp( cmdbuf, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, ctx->m_query, queryId );
 
+        GetProfiler().SendCallstack( depth );
+
         auto item = Profiler::QueueSerial();
         MemWrite( &item->hdr.type, QueueType::GpuZoneBeginCallstackSerial );
         MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
@@ -350,8 +352,6 @@ public:
         MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
         MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
         Profiler::QueueSerialFinish();
-
-        GetProfiler().SendCallstack( depth );
     }
 
     tracy_force_inline ~VkCtxScope()

client/TracyProfiler.cpp

@@ -594,12 +594,12 @@ LONG WINAPI CrashFilter( PEXCEPTION_POINTERS pExp )
     }
 
     {
+        GetProfiler().SendCallstack( 60, "KiUserExceptionDispatcher" );
+
         TracyLfqPrepare( QueueType::CrashReport );
         item->crashReport.time = Profiler::GetTime();
         item->crashReport.text = (uint64_t)s_crashText;
         TracyLfqCommit;
-
-        GetProfiler().SendCallstack( 60, "KiUserExceptionDispatcher" );
     }
 
     HANDLE h = CreateToolhelp32Snapshot( TH32CS_SNAPTHREAD, 0 );
@@ -826,12 +826,12 @@ static void CrashHandler( int signal, siginfo_t* info, void* /*ucontext*/ )
     }
 
     {
+        GetProfiler().SendCallstack( 60, "__kernel_rt_sigreturn" );
+
         TracyLfqPrepare( QueueType::CrashReport );
         item->crashReport.time = Profiler::GetTime();
         item->crashReport.text = (uint64_t)s_crashText;
         TracyLfqCommit;
-
-        GetProfiler().SendCallstack( 60, "__kernel_rt_sigreturn" );
     }
 
     DIR* dp = opendir( "/proc/self/task" );
@@ -2992,14 +2992,13 @@ TRACY_API TracyCZoneCtx ___tracy_emit_zone_begin_callstack( const struct ___trac
         TracyLfqCommitC;
     }
 #endif
+    tracy::GetProfiler().SendCallstack( depth );
     {
         TracyLfqPrepareC( tracy::QueueType::ZoneBeginCallstack );
         tracy::MemWrite( &item->zoneBegin.time, tracy::Profiler::GetTime() );
         tracy::MemWrite( &item->zoneBegin.srcloc, (uint64_t)srcloc );
         TracyLfqCommitC;
     }
-
-    tracy::GetProfiler().SendCallstack( depth );
     return ctx;
 }
 
@@ -3058,14 +3057,13 @@ TRACY_API TracyCZoneCtx ___tracy_emit_zone_begin_alloc_callstack( uint64_t srclo
         TracyLfqCommitC;
     }
 #endif
+    tracy::GetProfiler().SendCallstack( depth );
     {
         TracyLfqPrepareC( tracy::QueueType::ZoneBeginAllocSrcLocCallstack );
         tracy::MemWrite( &item->zoneBegin.time, tracy::Profiler::GetTime() );
         tracy::MemWrite( &item->zoneBegin.srcloc, srcloc );
         TracyLfqCommitC;
     }
-
-    tracy::GetProfiler().SendCallstack( depth );
     return ctx;
 }
 

client/TracyProfiler.hpp

@@ -283,6 +283,8 @@ public:
 #ifdef TRACY_ON_DEMAND
         if( !GetProfiler().IsConnected() ) return;
 #endif
+        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
+
         TracyLfqPrepare( callstack == 0 ? QueueType::Message : QueueType::MessageCallstack );
         auto ptr = (char*)tracy_malloc( size );
         memcpy( ptr, txt, size );
@@ -290,8 +292,6 @@ public:
         MemWrite( &item->messageFat.text, (uint64_t)ptr );
         MemWrite( &item->messageFat.size, (uint16_t)size );
         TracyLfqCommit;
-
-        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
     }
 
     static tracy_force_inline void Message( const char* txt, int callstack )
@@ -299,12 +299,12 @@ public:
 #ifdef TRACY_ON_DEMAND
         if( !GetProfiler().IsConnected() ) return;
 #endif
+        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
+
         TracyLfqPrepare( callstack == 0 ? QueueType::MessageLiteral : QueueType::MessageLiteralCallstack );
         MemWrite( &item->messageLiteral.time, GetTime() );
         MemWrite( &item->messageLiteral.text, (uint64_t)txt );
         TracyLfqCommit;
-
-        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
     }
 
     static tracy_force_inline void MessageColor( const char* txt, size_t size, uint32_t color, int callstack )
@@ -313,6 +313,8 @@ public:
 #ifdef TRACY_ON_DEMAND
         if( !GetProfiler().IsConnected() ) return;
 #endif
+        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
+
         TracyLfqPrepare( callstack == 0 ? QueueType::MessageColor : QueueType::MessageColorCallstack );
         auto ptr = (char*)tracy_malloc( size );
         memcpy( ptr, txt, size );
@@ -323,8 +325,6 @@ public:
         MemWrite( &item->messageColorFat.b, uint8_t( ( color >> 16 ) & 0xFF ) );
         MemWrite( &item->messageColorFat.size, (uint16_t)size );
         TracyLfqCommit;
-
-        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
     }
 
     static tracy_force_inline void MessageColor( const char* txt, uint32_t color, int callstack )
@@ -332,6 +332,8 @@ public:
 #ifdef TRACY_ON_DEMAND
         if( !GetProfiler().IsConnected() ) return;
 #endif
+        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
+
         TracyLfqPrepare( callstack == 0 ? QueueType::MessageLiteralColor : QueueType::MessageLiteralColorCallstack );
         MemWrite( &item->messageColorLiteral.time, GetTime() );
         MemWrite( &item->messageColorLiteral.text, (uint64_t)txt );
@@ -339,8 +341,6 @@ public:
         MemWrite( &item->messageColorLiteral.g, uint8_t( ( color >> 8  ) & 0xFF ) );
         MemWrite( &item->messageColorLiteral.b, uint8_t( ( color >> 16 ) & 0xFF ) );
         TracyLfqCommit;
-
-        if( callstack != 0 ) tracy::GetProfiler().SendCallstack( callstack );
     }
 
     static tracy_force_inline void MessageAppInfo( const char* txt, size_t size )
@@ -401,8 +401,8 @@ public:
         auto callstack = Callstack( depth );
 
         profiler.m_serialLock.lock();
-        SendMemAlloc( QueueType::MemAllocCallstack, thread, ptr, size );
         SendCallstackMemory( callstack );
+        SendMemAlloc( QueueType::MemAllocCallstack, thread, ptr, size );
         profiler.m_serialLock.unlock();
 #else
         MemAlloc( ptr, size, secure );
@@ -423,8 +423,8 @@ public:
         auto callstack = Callstack( depth );
 
         profiler.m_serialLock.lock();
-        SendMemFree( QueueType::MemFreeCallstack, thread, ptr );
         SendCallstackMemory( callstack );
+        SendMemFree( QueueType::MemFreeCallstack, thread, ptr );
         profiler.m_serialLock.unlock();
 #else
         MemFree( ptr, secure );
@@ -473,9 +473,9 @@ public:
         auto callstack = Callstack( depth );
 
         profiler.m_serialLock.lock();
+        SendCallstackMemory( callstack );
         SendMemName( name );
         SendMemAlloc( QueueType::MemAllocCallstackNamed, thread, ptr, size );
-        SendCallstackMemory( callstack );
         profiler.m_serialLock.unlock();
 #else
         MemAlloc( ptr, size, secure );
@@ -496,9 +496,9 @@ public:
         auto callstack = Callstack( depth );
 
         profiler.m_serialLock.lock();
+        SendCallstackMemory( callstack );
         SendMemName( name );
         SendMemFree( QueueType::MemFreeCallstackNamed, thread, ptr );
-        SendCallstackMemory( callstack );
         profiler.m_serialLock.unlock();
 #else
         MemFree( ptr, secure );

client/TracyScoped.hpp

@@ -49,12 +49,12 @@ public:
 #ifdef TRACY_ON_DEMAND
         m_connectionId = GetProfiler().ConnectionId();
 #endif
+        GetProfiler().SendCallstack( depth );
+
         TracyLfqPrepare( QueueType::ZoneBeginCallstack );
         MemWrite( &item->zoneBegin.time, Profiler::GetTime() );
         MemWrite( &item->zoneBegin.srcloc, (uint64_t)srcloc );
         TracyLfqCommit;
-
-        GetProfiler().SendCallstack( depth );
     }
 
     tracy_force_inline ScopedZone( uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, bool is_active = true )
@@ -86,13 +86,13 @@ public:
 #ifdef TRACY_ON_DEMAND
         m_connectionId = GetProfiler().ConnectionId();
 #endif
+        GetProfiler().SendCallstack( depth );
+
         TracyLfqPrepare( QueueType::ZoneBeginAllocSrcLocCallstack );
         const auto srcloc = Profiler::AllocSourceLocation( line, source, sourceSz, function, functionSz, name, nameSz );
         MemWrite( &item->zoneBegin.time, Profiler::GetTime() );
         MemWrite( &item->zoneBegin.srcloc, srcloc );
         TracyLfqCommit;
-
-        GetProfiler().SendCallstack( depth );
     }
 
     tracy_force_inline ~ScopedZone()

common/TracyProtocol.hpp

@@ -9,7 +9,7 @@ namespace tracy
 
 constexpr unsigned Lz4CompressBound( unsigned isize ) { return isize + ( isize / 255 ) + 16; }
 
-enum : uint32_t { ProtocolVersion = 41 };
+enum : uint32_t { ProtocolVersion = 42 };
 enum : uint16_t { BroadcastVersion = 2 };
 
 using lz4sz_t = uint32_t;