Automatic timer fallback

2 files changed, 105 insertions, 37 deletions

diff --git a/client/TracyProfiler.cpp b/client/TracyProfiler.cpp
index ff9ce9c7..73384eab 100644
--- a/client/TracyProfiler.cpp
+++ b/client/TracyProfiler.cpp
@@ -322,24 +322,50 @@ static void InitFailure( const char* msg )
     exit( 1 );
 }
 
-static int64_t SetupHwTimer()
+static bool checkHardwareSupportsInvariantTSC()
 {
-#if !defined TRACY_TIMER_QPC && !defined TRACY_TIMER_FALLBACK
+    const char* noCheck = GetEnvVar( "TRACY_NO_INVARIANT_CHECK" );
+    if( noCheck && noCheck[0] == '1' ) 
+    {
+        return true;
+    }
+
     uint32_t regs[4];
     CpuId( regs, 1 );
-    if( !( regs[3] & ( 1 << 4 ) ) ) InitFailure( "CPU doesn't support RDTSC instruction." );
+    if( !( regs[3] & ( 1 << 4 ) ) )
+    {
+#if !defined TRACY_TIMER_QPC && !defined TRACY_TIMER_FALLBACK
+        InitFailure( "CPU doesn't support RDTSC instruction." );
+#endif
+        return false;
+    }
     CpuId( regs, 0x80000007 );
-    if( !( regs[3] & ( 1 << 8 ) ) )
+    if( regs[3] & ( 1 << 8 ) )
+    {
+        return true;
+    }
+
+    return false;
+}
+
+#if defined TRACY_TIMER_FALLBACK && defined TRACY_HW_TIMER
+bool hardwareSupportsInvariantTSC()
+{
+    static bool cachedResult = checkHardwareSupportsInvariantTSC();
+    return cachedResult;
+}
+#endif
+
+static int64_t SetupHwTimer()
+{
+#if !defined TRACY_TIMER_QPC && !defined TRACY_TIMER_FALLBACK
+    if(!checkHardwareSupportsInvariantTSC())
     {
-        const char* noCheck = GetEnvVar( "TRACY_NO_INVARIANT_CHECK" );
-        if( !noCheck || noCheck[0] != '1' )
-        {
 #if defined _WIN32
-            InitFailure( "CPU doesn't support invariant TSC.\nDefine TRACY_NO_INVARIANT_CHECK=1 to ignore this error, *if you know what you are doing*.\nAlternatively you may rebuild the application with the TRACY_TIMER_QPC or TRACY_TIMER_FALLBACK define to use lower resolution timer." );
+        InitFailure( "CPU doesn't support invariant TSC.\nDefine TRACY_NO_INVARIANT_CHECK=1 to ignore this error, *if you know what you are doing*.\nAlternatively you may rebuild the application with the TRACY_TIMER_QPC or TRACY_TIMER_FALLBACK define to use lower resolution timer." );
 #else
-            InitFailure( "CPU doesn't support invariant TSC.\nDefine TRACY_NO_INVARIANT_CHECK=1 to ignore this error, *if you know what you are doing*.\nAlternatively you may rebuild the application with the TRACY_TIMER_FALLBACK define to use lower resolution timer." );
+        InitFailure( "CPU doesn't support invariant TSC.\nDefine TRACY_NO_INVARIANT_CHECK=1 to ignore this error, *if you know what you are doing*.\nAlternatively you may rebuild the application with the TRACY_TIMER_FALLBACK define to use lower resolution timer." );
 #endif
-        }
     }
 #endif
 
@@ -3451,23 +3477,32 @@ void Profiler::HandleDisconnect()
 
 void Profiler::CalibrateTimer()
 {
-#ifdef TRACY_HW_TIMER
-    std::atomic_signal_fence( std::memory_order_acq_rel );
-    const auto t0 = std::chrono::high_resolution_clock::now();
-    const auto r0 = GetTime();
-    std::atomic_signal_fence( std::memory_order_acq_rel );
-    std::this_thread::sleep_for( std::chrono::milliseconds( 200 ) );
-    std::atomic_signal_fence( std::memory_order_acq_rel );
-    const auto t1 = std::chrono::high_resolution_clock::now();
-    const auto r1 = GetTime();
-    std::atomic_signal_fence( std::memory_order_acq_rel );
-
-    const auto dt = std::chrono::duration_cast<std::chrono::nanoseconds>( t1 - t0 ).count();
-    const auto dr = r1 - r0;
-
-    m_timerMul = double( dt ) / double( dr );
-#else
     m_timerMul = 1.;
+
+#ifdef TRACY_HW_TIMER
+
+#  if !defined TRACY_TIMER_QPC && defined TRACY_TIMER_FALLBACK
+    const bool needCalibration = hardwareSupportsInvariantTSC();
+#  else
+    const bool needCalibration = true;
+#  endif
+    if (needCalibration)
+    {
+        std::atomic_signal_fence( std::memory_order_acq_rel );
+        const auto t0 = std::chrono::high_resolution_clock::now();
+        const auto r0 = GetTime();
+        std::atomic_signal_fence( std::memory_order_acq_rel );
+        std::this_thread::sleep_for( std::chrono::milliseconds( 200 ) );
+        std::atomic_signal_fence( std::memory_order_acq_rel );
+        const auto t1 = std::chrono::high_resolution_clock::now();
+        const auto r1 = GetTime();
+        std::atomic_signal_fence( std::memory_order_acq_rel );
+
+        const auto dt = std::chrono::duration_cast<std::chrono::nanoseconds>( t1 - t0 ).count();
+        const auto dr = r1 - r0;
+
+        m_timerMul = double( dt ) / double( dr );
+    }
 #endif
 }
 
diff --git a/client/TracyProfiler.hpp b/client/TracyProfiler.hpp
index 005feb6e..ca50dd24 100644
--- a/client/TracyProfiler.hpp
+++ b/client/TracyProfiler.hpp
@@ -26,11 +26,11 @@
 #  include <mach/mach_time.h>
 #endif
 
-#if !defined TRACY_TIMER_FALLBACK && ( defined _WIN32 || ( defined __i386 || defined _M_IX86 || defined __x86_64__ || defined _M_X64 ) || ( defined TARGET_OS_IOS && TARGET_OS_IOS == 1 ) )
+#if ( defined _WIN32 || ( defined __i386 || defined _M_IX86 || defined __x86_64__ || defined _M_X64 ) || ( defined TARGET_OS_IOS && TARGET_OS_IOS == 1 ) )
 #  define TRACY_HW_TIMER
 #endif
 
-#if !defined TRACY_HW_TIMER
+#if defined TRACY_TIMER_FALLBACK || !defined TRACY_HW_TIMER
 #  include <chrono>
 #endif
 
@@ -67,6 +67,23 @@ TRACY_API uint32_t GetThreadHandle();
 TRACY_API bool ProfilerAvailable();
 TRACY_API int64_t GetFrequencyQpc();
 
+#if defined TRACY_TIMER_FALLBACK && defined TRACY_HW_TIMER && ( defined __i386 || defined _M_IX86 || defined __x86_64__ || defined _M_X64 )
+TRACY_API bool hardwareSupportsInvariantTSC();  // check, if we need fallback scenario
+#else
+#  if defined TRACY_HW_TIMER
+tracy_force_inline bool hardwareSupportsInvariantTSC()
+{
+    return true;  // this is checked at startup
+}
+#  else
+tracy_force_inline bool hardwareSupportsInvariantTSC()
+{
+    return false;
+}
+#  endif
+#endif
+
+
 struct SourceLocationData
 {
     const char* name;
@@ -166,25 +183,39 @@ public:
     {
 #ifdef TRACY_HW_TIMER
 #  if defined TARGET_OS_IOS && TARGET_OS_IOS == 1
-        return mach_absolute_time();
+        if (hardwareSupportsInvariantTSC())
+        {
+            return mach_absolute_time();
+        }
 #  elif defined _WIN32
 #    ifdef TRACY_TIMER_QPC
         return GetTimeQpc();
 #    else
-        return int64_t( __rdtsc() );
+        if (hardwareSupportsInvariantTSC())
+        {
+            return int64_t( __rdtsc() );
+        }
 #    endif
 #  elif defined __i386 || defined _M_IX86
-        uint32_t eax, edx;
-        asm volatile ( "rdtsc" : "=a" (eax), "=d" (edx) );
-        return ( uint64_t( edx ) << 32 ) + uint64_t( eax );
+        if (hardwareSupportsInvariantTSC())
+        {
+            uint32_t eax, edx;
+            asm volatile ( "rdtsc" : "=a" (eax), "=d" (edx) );
+            return ( uint64_t( edx ) << 32 ) + uint64_t( eax );
+        }
 #  elif defined __x86_64__ || defined _M_X64
-        uint64_t rax, rdx;
-        asm volatile ( "rdtsc" : "=a" (rax), "=d" (rdx) );
-        return (int64_t)(( rdx << 32 ) + rax);
+        if (hardwareSupportsInvariantTSC())
+        {
+            uint64_t rax, rdx;
+            asm volatile ( "rdtsc" : "=a" (rax), "=d" (rdx) );
+            return (int64_t)(( rdx << 32 ) + rax);
+        }
 #  else
 #    error "TRACY_HW_TIMER detection logic needs fixing"
 #  endif
-#else
+#endif
+
+#if !defined TRACY_HW_TIMER || defined TRACY_TIMER_FALLBACK
 #  if defined __linux__ && defined CLOCK_MONOTONIC_RAW
         struct timespec ts;
         clock_gettime( CLOCK_MONOTONIC_RAW, &ts );
@@ -193,6 +224,8 @@ public:
         return std::chrono::duration_cast<std::chrono::nanoseconds>( std::chrono::high_resolution_clock::now().time_since_epoch() ).count();
 #  endif
 #endif
+
+        return 0;  // unreacheble branch
     }
 
     tracy_force_inline uint32_t GetNextZoneId()