Update tracy version to 0.1

Author: Causeless

external/sources/tracy/meson.build

@@ -1,4 +1,4 @@
-project('tracy', ['cpp'])
+project('tracy', ['cpp'], version: '0.10.0')
 
 if get_option('tracy_enable')
   add_project_arguments('-DTRACY_ENABLE', language : 'cpp')
@@ -64,6 +64,10 @@ if get_option('tracy_no_system_tracing')
   add_project_arguments('-DTRACY_NO_SYSTEM_TRACING', language : 'cpp')
 endif
 
+if get_option('tracy_patchable_nopsleds')
+  add_project_arguments('-DTRACY_PATCHABLE_NOPSLEDS', language : 'cpp')
+endif
+
 if get_option('tracy_delayed_init')
   add_project_arguments('-DTRACY_DELAYED_INIT', language : 'cpp')
 endif
@@ -161,18 +165,17 @@ if tracy_shared_libs
     include_directories : tracy_public_include_dirs,
     override_options    : override_options,
     install             : true)
-    install_headers(includes)
-    install_headers(common_includes, subdir : 'common')
-    install_headers(client_includes, subdir : 'client')
 else
   tracy = static_library('tracy', tracy_src, tracy_header_files,
     dependencies        : [ threads_dep ],
     include_directories : tracy_public_include_dirs,
     override_options    : override_options,
-    install             : false)
+    install             : true)
 endif
 
-
+install_headers(includes)
+install_headers(common_includes, subdir : 'common')
+install_headers(client_includes, subdir : 'client')
 
 tracy_dep_compile_args = []
 

external/sources/tracy/meson_options.txt

@@ -1,6 +1,6 @@
-option('tracy_enable', type : 'boolean', yield: true, value : true, description : 'Enable profiling')
-option('tracy_on_demand', type : 'boolean', value : false, yield: true, description : 'On-demand profiling')
-option('tracy_callstack', type : 'boolean', value : false, yield: true, description : 'Enfore callstack collection for tracy regions')
+option('tracy_enable', type : 'boolean', value : true, description : 'Enable profiling')
+option('tracy_on_demand', type : 'boolean', value : false, description : 'On-demand profiling')
+option('tracy_callstack', type : 'boolean', value : false, description : 'Enfore callstack collection for tracy regions')
 option('tracy_no_callstack', type : 'boolean', value : false, description : 'Disable all callstack related functionality')
 option('tracy_no_callstack_inlines', type : 'boolean', value : false, description : 'Disables the inline functions in callstacks')
 option('tracy_only_localhost', type : 'boolean', value : false, description : 'Only listen on the localhost interface')
@@ -14,6 +14,7 @@ option('tracy_no_verify', type : 'boolean', value : false, description : 'Disabl
 option('tracy_no_vsync_capture', type : 'boolean', value : false, description : 'Disable capture of hardware Vsync events')
 option('tracy_no_frame_image', type : 'boolean', value : false, description : 'Disable the frame image support and its thread')
 option('tracy_no_system_tracing', type : 'boolean', value : false, description : 'Disable systrace sampling')
+option('tracy_patchable_nopsleds', type : 'boolean', value : false, description : 'Enable nopsleds for efficient patching by system-level tools (e.g. rr)')
 option('tracy_timer_fallback', type : 'boolean', value : false, description : 'Use lower resolution timers')
 option('tracy_delayed_init', type : 'boolean', value : false, description : 'Enable delayed initialization of the library (init on first call)')
 option('tracy_manual_lifetime', type : 'boolean', value : false, description : 'Enable the manual lifetime management of the profile')

external/sources/tracy/public/TracyClient.cpp

@@ -22,6 +22,7 @@
 #include "common/tracy_lz4.cpp"
 #include "client/TracyProfiler.cpp"
 #include "client/TracyCallstack.cpp"
+#include "client/TracySysPower.cpp"
 #include "client/TracySysTime.cpp"
 #include "client/TracySysTrace.cpp"
 #include "common/TracySocket.cpp"

external/sources/tracy/public/client/TracyCallstack.cpp

@@ -686,7 +686,9 @@ void InitCallstackCritical()
 void InitCallstack()
 {
     cb_bts = backtrace_create_state( nullptr, 0, nullptr, nullptr );
+#ifndef TRACY_DEMANGLE
     ___tracy_init_demangle_buffer();
+#endif
 
 #ifdef __linux
     InitKernelSymbols();
@@ -761,7 +763,9 @@ debuginfod_client* GetDebuginfodClient()
 
 void EndCallstack()
 {
+#ifndef TRACY_DEMANGLE
     ___tracy_free_demangle_buffer();
+#endif
 #ifdef TRACY_DEBUGINFOD
     ClearDebugInfoVector( s_di_known );
     debuginfod_end( s_debuginfod );

external/sources/tracy/public/client/TracyLock.hpp

@@ -21,7 +21,7 @@ class LockableCtx
         , m_active( false )
 #endif
     {
-        assert( m_id != std::numeric_limits<uint32_t>::max() );
+        assert( m_id != (std::numeric_limits<uint32_t>::max)() );
 
         auto item = Profiler::QueueSerial();
         MemWrite( &item->hdr.type, QueueType::LockAnnounce );
@@ -154,7 +154,7 @@ class LockableCtx
 
     tracy_force_inline void CustomName( const char* name, size_t size )
     {
-        assert( size < std::numeric_limits<uint16_t>::max() );
+        assert( size < (std::numeric_limits<uint16_t>::max)() );
         auto ptr = (char*)tracy_malloc( size );
         memcpy( ptr, name, size );
         auto item = Profiler::QueueSerial();
@@ -235,7 +235,7 @@ class SharedLockableCtx
         , m_active( false )
 #endif
     {
-        assert( m_id != std::numeric_limits<uint32_t>::max() );
+        assert( m_id != (std::numeric_limits<uint32_t>::max)() );
 
         auto item = Profiler::QueueSerial();
         MemWrite( &item->hdr.type, QueueType::LockAnnounce );
@@ -450,7 +450,7 @@ class SharedLockableCtx
 
     tracy_force_inline void CustomName( const char* name, size_t size )
     {
-        assert( size < std::numeric_limits<uint16_t>::max() );
+        assert( size < (std::numeric_limits<uint16_t>::max)() );
         auto ptr = (char*)tracy_malloc( size );
         memcpy( ptr, name, size );
         auto item = Profiler::QueueSerial();

external/sources/tracy/public/client/TracyProfiler.cpp

@@ -83,7 +83,9 @@
 #endif
 
 #ifdef __APPLE__
-#  define TRACY_DELAYED_INIT
+#  ifndef TRACY_DELAYED_INIT
+#    define TRACY_DELAYED_INIT
+#  endif
 #else
 #  ifdef __GNUC__
 #    define init_order( val ) __attribute__ ((init_priority(val)))
@@ -1072,7 +1074,9 @@ static void CrashHandler( int signal, siginfo_t* info, void* /*ucontext*/ )
     }
     closedir( dp );
 
+#ifdef TRACY_HAS_CALLSTACK
     if( selfTid == s_symbolTid ) s_symbolThreadGone.store( true, std::memory_order_release );
+#endif
 
     TracyLfqPrepare( QueueType::Crash );
     TracyLfqCommit;
@@ -1353,6 +1357,7 @@ Profiler::Profiler()
     , m_queryImage( nullptr )
     , m_queryData( nullptr )
     , m_crashHandlerInstalled( false )
+    , m_programName( nullptr )
 {
     assert( !s_instance );
     s_instance = this;
@@ -1711,13 +1716,24 @@ void Profiler::Worker()
             if( m_sock ) break;
 #ifndef TRACY_ON_DEMAND
             ProcessSysTime();
+#  ifdef TRACY_HAS_SYSPOWER
+            m_sysPower.Tick();
+#  endif
 #endif
 
             if( m_broadcast )
             {
                 const auto t = std::chrono::high_resolution_clock::now().time_since_epoch().count();
                 if( t - lastBroadcast > 3000000000 )  // 3s
                 {
+                    m_programNameLock.lock();
+                    if( m_programName )
+                    {
+                        broadcastMsg = GetBroadcastMessage( m_programName, strlen( m_programName ), broadcastLen, dataPort );
+                        m_programName = nullptr;
+                    }
+                    m_programNameLock.unlock();
+
                     lastBroadcast = t;
                     const auto ts = std::chrono::duration_cast<std::chrono::seconds>( std::chrono::system_clock::now().time_since_epoch() ).count();
                     broadcastMsg.activeTime = int32_t( ts - m_epoch );
@@ -1828,6 +1844,9 @@ void Profiler::Worker()
         for(;;)
         {
             ProcessSysTime();
+#ifdef TRACY_HAS_SYSPOWER
+            m_sysPower.Tick();
+#endif
             const auto status = Dequeue( token );
             const auto serialStatus = DequeueSerial();
             if( status == DequeueStatus::ConnectionLost || serialStatus == DequeueStatus::ConnectionLost )
@@ -4149,6 +4168,7 @@ TRACY_API void ___tracy_emit_frame_image( const void* image, uint16_t w, uint16_
 TRACY_API void ___tracy_emit_plot( const char* name, double val ) { tracy::Profiler::PlotData( name, val ); }
 TRACY_API void ___tracy_emit_plot_float( const char* name, float val ) { tracy::Profiler::PlotData( name, val ); }
 TRACY_API void ___tracy_emit_plot_int( const char* name, int64_t val ) { tracy::Profiler::PlotData( name, val ); }
+TRACY_API void ___tracy_emit_plot_config( const char* name, int type, int step, int fill, uint32_t color ) { tracy::Profiler::ConfigurePlot( name, tracy::PlotFormatType(type), step, fill, color ); }
 TRACY_API void ___tracy_emit_message( const char* txt, size_t size, int callstack ) { tracy::Profiler::Message( txt, size, callstack ); }
 TRACY_API void ___tracy_emit_messageL( const char* txt, int callstack ) { tracy::Profiler::Message( txt, callstack ); }
 TRACY_API void ___tracy_emit_messageC( const char* txt, size_t size, uint32_t color, int callstack ) { tracy::Profiler::MessageColor( txt, size, color, callstack ); }
@@ -4167,7 +4187,7 @@ TRACY_API void ___tracy_emit_gpu_zone_begin( const struct ___tracy_gpu_zone_begi
 {
     TracyLfqPrepareC( tracy::QueueType::GpuZoneBegin );
     tracy::MemWrite( &item->gpuZoneBegin.cpuTime, tracy::Profiler::GetTime() );
-    tracy::MemWrite( &item->gpuNewContext.thread, tracy::GetThreadHandle() );
+    tracy::MemWrite( &item->gpuZoneBegin.thread, tracy::GetThreadHandle() );
     tracy::MemWrite( &item->gpuZoneBegin.srcloc, data.srcloc );
     tracy::MemWrite( &item->gpuZoneBegin.queryId, data.queryId );
     tracy::MemWrite( &item->gpuZoneBegin.context, data.context );
@@ -4190,7 +4210,7 @@ TRACY_API void ___tracy_emit_gpu_zone_begin_alloc( const struct ___tracy_gpu_zon
 {
     TracyLfqPrepareC( tracy::QueueType::GpuZoneBeginAllocSrcLoc  );
     tracy::MemWrite( &item->gpuZoneBegin.cpuTime, tracy::Profiler::GetTime() );
-    tracy::MemWrite( &item->gpuNewContext.thread, tracy::GetThreadHandle() );
+    tracy::MemWrite( &item->gpuZoneBegin.thread, tracy::GetThreadHandle() );
     tracy::MemWrite( &item->gpuZoneBegin.srcloc, data.srcloc );
     tracy::MemWrite( &item->gpuZoneBegin.queryId, data.queryId );
     tracy::MemWrite( &item->gpuZoneBegin.context, data.context );
@@ -4202,7 +4222,7 @@ TRACY_API void ___tracy_emit_gpu_zone_begin_alloc_callstack( const struct ___tra
     tracy::GetProfiler().SendCallstack( data.depth );
     TracyLfqPrepareC( tracy::QueueType::GpuZoneBeginAllocSrcLocCallstack  );
     tracy::MemWrite( &item->gpuZoneBegin.cpuTime, tracy::Profiler::GetTime() );
-    tracy::MemWrite( &item->gpuNewContext.thread, tracy::GetThreadHandle() );
+    tracy::MemWrite( &item->gpuZoneBegin.thread, tracy::GetThreadHandle() );
     tracy::MemWrite( &item->gpuZoneBegin.srcloc, data.srcloc );
     tracy::MemWrite( &item->gpuZoneBegin.queryId, data.queryId );
     tracy::MemWrite( &item->gpuZoneBegin.context, data.context );
@@ -4292,7 +4312,7 @@ TRACY_API void ___tracy_emit_gpu_zone_begin_alloc_serial( const struct ___tracy_
     auto item = tracy::Profiler::QueueSerial();
     tracy::MemWrite( &item->hdr.type, tracy::QueueType::GpuZoneBeginAllocSrcLocSerial );
     tracy::MemWrite( &item->gpuZoneBegin.cpuTime, tracy::Profiler::GetTime() );
-    tracy::MemWrite( &item->gpuNewContext.thread, tracy::GetThreadHandle() );
+    tracy::MemWrite( &item->gpuZoneBegin.thread, tracy::GetThreadHandle() );
     tracy::MemWrite( &item->gpuZoneBegin.srcloc, data.srcloc );
     tracy::MemWrite( &item->gpuZoneBegin.queryId, data.queryId );
     tracy::MemWrite( &item->gpuZoneBegin.context, data.context );
@@ -4304,7 +4324,7 @@ TRACY_API void ___tracy_emit_gpu_zone_begin_alloc_callstack_serial( const struct
     auto item = tracy::Profiler::QueueSerialCallstack( tracy::Callstack( data.depth ) );
     tracy::MemWrite( &item->hdr.type, tracy::QueueType::GpuZoneBeginAllocSrcLocCallstackSerial );
     tracy::MemWrite( &item->gpuZoneBegin.cpuTime, tracy::Profiler::GetTime() );
-    tracy::MemWrite( &item->gpuNewContext.thread, tracy::GetThreadHandle() );
+    tracy::MemWrite( &item->gpuZoneBegin.thread, tracy::GetThreadHandle() );
     tracy::MemWrite( &item->gpuZoneBegin.srcloc, data.srcloc );
     tracy::MemWrite( &item->gpuZoneBegin.queryId, data.queryId );
     tracy::MemWrite( &item->gpuZoneBegin.context, data.context );

external/sources/tracy/public/client/TracyProfiler.hpp

@@ -10,6 +10,7 @@
 #include "tracy_concurrentqueue.h"
 #include "tracy_SPSCQueue.h"
 #include "TracyCallstack.hpp"
+#include "TracySysPower.hpp"
 #include "TracySysTime.hpp"
 #include "TracyFastVector.hpp"
 #include "../common/TracyQueue.hpp"
@@ -208,7 +209,22 @@ class Profiler
         if( HardwareSupportsInvariantTSC() )
         {
             uint64_t rax, rdx;
+#ifdef TRACY_PATCHABLE_NOPSLEDS
+            // Some external tooling (such as rr) wants to patch our rdtsc and replace it by a
+            // branch to control the external input seen by a program. This kind of patching is
+            // not generally possible depending on the surrounding code and can lead to significant
+            // slowdowns if the compiler generated unlucky code and rr and tracy are used together.
+            // To avoid this, use the rr-safe `nopl 0(%rax, %rax, 1); rdtsc` instruction sequence,
+            // which rr promises will be patchable independent of the surrounding code.
+            asm volatile (
+                    // This is nopl 0(%rax, %rax, 1), but assemblers are inconsistent about whether
+                    // they emit that as a 4 or 5 byte sequence and we need to be guaranteed to use
+                    // the 5 byte one.
+                    ".byte 0x0f, 0x1f, 0x44, 0x00, 0x00\n\t"
+                    "rdtsc" : "=a" (rax), "=d" (rdx) );
+#else
             asm volatile ( "rdtsc" : "=a" (rax), "=d" (rdx) );
+#endif
             return (int64_t)(( rdx << 32 ) + rax);
         }
 #  else
@@ -288,7 +304,7 @@ class Profiler
     {
 #ifndef TRACY_NO_FRAME_IMAGE
         auto& profiler = GetProfiler();
-        assert( profiler.m_frameCount.load( std::memory_order_relaxed ) < std::numeric_limits<uint32_t>::max() );
+        assert( profiler.m_frameCount.load( std::memory_order_relaxed ) < (std::numeric_limits<uint32_t>::max)() );
 #  ifdef TRACY_ON_DEMAND
         if( !profiler.IsConnected() ) return;
 #  endif
@@ -305,6 +321,12 @@ class Profiler
         fi->flip = flip;
         profiler.m_fiQueue.commit_next();
         profiler.m_fiLock.unlock();
+#else
+        static_cast<void>(image); // unused
+        static_cast<void>(w); // unused
+        static_cast<void>(h); // unused
+        static_cast<void>(offset); // unused
+        static_cast<void>(flip); // unused
 #endif
     }
 
@@ -362,7 +384,7 @@ class Profiler
 
     static tracy_force_inline void Message( const char* txt, size_t size, int callstack )
     {
-        assert( size < std::numeric_limits<uint16_t>::max() );
+        assert( size < (std::numeric_limits<uint16_t>::max)() );
 #ifdef TRACY_ON_DEMAND
         if( !GetProfiler().IsConnected() ) return;
 #endif
@@ -399,7 +421,7 @@ class Profiler
 
     static tracy_force_inline void MessageColor( const char* txt, size_t size, uint32_t color, int callstack )
     {
-        assert( size < std::numeric_limits<uint16_t>::max() );
+        assert( size < (std::numeric_limits<uint16_t>::max)() );
 #ifdef TRACY_ON_DEMAND
         if( !GetProfiler().IsConnected() ) return;
 #endif
@@ -442,7 +464,7 @@ class Profiler
 
     static tracy_force_inline void MessageAppInfo( const char* txt, size_t size )
     {
-        assert( size < std::numeric_limits<uint16_t>::max() );
+        assert( size < (std::numeric_limits<uint16_t>::max)() );
         auto ptr = (char*)tracy_malloc( size );
         memcpy( ptr, txt, size );
         TracyLfqPrepare( QueueType::MessageAppInfo );
@@ -676,6 +698,13 @@ class Profiler
         return m_isConnected.load( std::memory_order_acquire );
     }
 
+    tracy_force_inline void SetProgramName( const char* name )
+    {
+        m_programNameLock.lock();
+        m_programName = name;
+        m_programNameLock.unlock();
+    }
+
 #ifdef TRACY_ON_DEMAND
     tracy_force_inline uint64_t ConnectionId() const
     {
@@ -730,7 +759,7 @@ class Profiler
     static tracy_force_inline uint64_t AllocSourceLocation( uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz )
     {
         const auto sz32 = uint32_t( 2 + 4 + 4 + functionSz + 1 + sourceSz + 1 + nameSz );
-        assert( sz32 <= std::numeric_limits<uint16_t>::max() );
+        assert( sz32 <= (std::numeric_limits<uint16_t>::max)() );
         const auto sz = uint16_t( sz32 );
         auto ptr = (char*)tracy_malloc( sz );
         memcpy( ptr, &sz, 2 );
@@ -941,6 +970,10 @@ class Profiler
     void ProcessSysTime() {}
 #endif
 
+#ifdef TRACY_HAS_SYSPOWER
+    SysPower m_sysPower;
+#endif
+
     ParameterCallback m_paramCallback;
     void* m_paramCallbackData;
     SourceContentsCallback m_sourceCallback;
@@ -959,6 +992,9 @@ class Profiler
     } m_prevSignal;
 #endif
     bool m_crashHandlerInstalled;
+
+    const char* m_programName;
+    TracyMutex m_programNameLock;
 };
 
 }

external/sources/tracy/public/client/TracyScoped.hpp

@@ -108,7 +108,7 @@ class ScopedZone
 
     tracy_force_inline void Text( const char* txt, size_t size )
     {
-        assert( size < std::numeric_limits<uint16_t>::max() );
+        assert( size < (std::numeric_limits<uint16_t>::max)() );
         if( !m_active ) return;
 #ifdef TRACY_ON_DEMAND
         if( GetProfiler().ConnectionId() != m_connectionId ) return;
@@ -123,7 +123,7 @@ class ScopedZone
 
     tracy_force_inline void Name( const char* txt, size_t size )
     {
-        assert( size < std::numeric_limits<uint16_t>::max() );
+        assert( size < (std::numeric_limits<uint16_t>::max)() );
         if( !m_active ) return;
 #ifdef TRACY_ON_DEMAND
         if( GetProfiler().ConnectionId() != m_connectionId ) return;