Merge pull request #3016 from secondlife/rye/tracyfixes

Tracy Improvements

4 files changed, 87 insertions, 76 deletions

diff --git a/indra/llcommon/linden_common.h b/indra/llcommon/linden_common.h
index a918caa2e8..a41af153fe 100644
--- a/indra/llcommon/linden_common.h
+++ b/indra/llcommon/linden_common.h
@@ -28,12 +28,6 @@
 #define LL_LINDEN_COMMON_H
 
 #include "llprofiler.h"
-#if TRACY_ENABLE && !defined(LL_PROFILER_ENABLE_TRACY_OPENGL)  // hooks for memory profiling
-void *tracy_aligned_malloc(size_t size, size_t alignment);
-void  tracy_aligned_free(void *memblock);
-#define _aligned_malloc(X, Y) tracy_aligned_malloc((X), (Y))
-#define _aligned_free(X)      tracy_aligned_free((X))
-#endif
 
 // *NOTE:  Please keep includes here to a minimum!
 //
diff --git a/indra/llcommon/llcommon.cpp b/indra/llcommon/llcommon.cpp
index f1f3958fe0..110a1c32a1 100644
--- a/indra/llcommon/llcommon.cpp
+++ b/indra/llcommon/llcommon.cpp
@@ -33,23 +33,23 @@
 #include "lltracethreadrecorder.h"
 #include "llcleanup.h"
 
-thread_local bool gProfilerEnabled = false;
-
-#if (TRACY_ENABLE)
+#if LL_PROFILER_CONFIGURATION > 1 && TRACY_ENABLE
 // Override new/delete for tracy memory profiling
 
 void* ll_tracy_new(size_t size)
 {
-    void* ptr;
-    if (gProfilerEnabled)
-    {
-        //LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
-        ptr = (malloc)(size);
-    }
-    else
+    void* ptr = (malloc)(size);
+    if (!ptr)
     {
-        ptr = (malloc)(size);
+        throw std::bad_alloc();
     }
+    TracyAlloc(ptr, size);
+    return ptr;
+}
+
+void* ll_tracy_aligned_new(size_t size, size_t alignment)
+{
+    void* ptr = ll_aligned_malloc_fallback(size, alignment);
     if (!ptr)
     {
         throw std::bad_alloc();
@@ -58,6 +58,18 @@ void* ll_tracy_new(size_t size)
     return ptr;
 }
 
+void ll_tracy_delete(void* ptr)
+{
+    TracyFree(ptr);
+    (free)(ptr);
+}
+
+void ll_tracy_aligned_delete(void* ptr)
+{
+    TracyFree(ptr);
+    ll_aligned_free_fallback(ptr);
+}
+
 void* operator new(size_t size)
 {
     return ll_tracy_new(size);
@@ -68,18 +80,14 @@ void* operator new[](std::size_t count)
     return ll_tracy_new(count);
 }
 
-void ll_tracy_delete(void* ptr)
+void* operator new(size_t size, std::align_val_t align)
 {
-    TracyFree(ptr);
-    if (gProfilerEnabled)
-    {
-        //LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
-        (free)(ptr);
-    }
-    else
-    {
-        (free)(ptr);
-    }
+    return ll_tracy_aligned_new(size, (size_t)align);
+}
+
+void* operator new[](std::size_t count, std::align_val_t align)
+{
+    return ll_tracy_aligned_new(count, (size_t)align);
 }
 
 void operator delete(void *ptr) noexcept
@@ -92,27 +100,17 @@ void operator delete[](void* ptr) noexcept
     ll_tracy_delete(ptr);
 }
 
-// C-style malloc/free can't be so easily overridden, so we define tracy versions and use
-// a pre-processor #define in linden_common.h to redirect to them. The parens around the native
-// functions below prevents recursive substitution by the preprocessor.
-//
-// Unaligned mallocs are rare in LL code but hooking them causes problems in 3p lib code (looking at
-// you, Havok), so we'll only capture the aligned version.
-
-void *tracy_aligned_malloc(size_t size, size_t alignment)
+void operator delete(void *ptr, std::align_val_t align) noexcept
 {
-    auto ptr = ll_aligned_malloc_fallback(size, alignment);
-    if (ptr) TracyAlloc(ptr, size);
-    return ptr;
+    ll_tracy_aligned_delete(ptr);
 }
 
-void tracy_aligned_free(void *memblock)
+void operator delete[](void* ptr, std::align_val_t align) noexcept
 {
-    TracyFree(memblock);
-    ll_aligned_free_fallback(memblock);
+    ll_tracy_aligned_delete(ptr);
 }
 
-#endif
+#endif // TRACY_ENABLE && !LL_PROFILER_ENABLE_TRACY_OPENGL
 
 //static
 bool LLCommon::sAprInitialized = false;
diff --git a/indra/llcommon/llmemory.h b/indra/llcommon/llmemory.h
index 80cfe554c4..72aec57080 100644
--- a/indra/llcommon/llmemory.h
+++ b/indra/llcommon/llmemory.h
@@ -222,7 +222,7 @@ inline void* ll_aligned_realloc_16(void* ptr, size_t size, size_t old_size) // r
         ll_aligned_free_16(ptr);
     }
 #endif
-    LL_PROFILE_ALLOC(ptr, size);
+    LL_PROFILE_ALLOC(ret, size);
     return ret;
 }
 
@@ -231,8 +231,6 @@ inline void* ll_aligned_malloc_32(size_t size) // returned hunk MUST be freed wi
     LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
 #if defined(LL_WINDOWS)
     void* ret = _aligned_malloc(size, 32);
-#elif defined(LL_DARWIN)
-    void* ret = ll_aligned_malloc_fallback( size, 32 );
 #else
     void *ret;
     if (0 != posix_memalign(&ret, 32, size))
@@ -248,8 +246,31 @@ inline void ll_aligned_free_32(void *p)
     LL_PROFILE_FREE(p);
 #if defined(LL_WINDOWS)
     _aligned_free(p);
-#elif defined(LL_DARWIN)
-    ll_aligned_free_fallback( p );
+#else
+    free(p); // posix_memalign() is compatible with heap deallocator
+#endif
+}
+
+inline void* ll_aligned_malloc_64(size_t size) // returned hunk MUST be freed with ll_aligned_free_32().
+{
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
+#if defined(LL_WINDOWS)
+    void* ret = _aligned_malloc(size, 64);
+#else
+    void *ret;
+    if (0 != posix_memalign(&ret, 64, size))
+        return nullptr;
+#endif
+    LL_PROFILE_ALLOC(ret, size);
+    return ret;
+}
+
+inline void ll_aligned_free_64(void *p)
+{
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
+    LL_PROFILE_FREE(p);
+#if defined(LL_WINDOWS)
+    _aligned_free(p);
 #else
     free(p); // posix_memalign() is compatible with heap deallocator
 #endif
@@ -261,19 +282,23 @@ LL_FORCE_INLINE void* ll_aligned_malloc(size_t size)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
     void* ret;
-    if (LL_DEFAULT_HEAP_ALIGN % ALIGNMENT == 0)
+    if constexpr (LL_DEFAULT_HEAP_ALIGN % ALIGNMENT == 0)
     {
         ret = malloc(size);
         LL_PROFILE_ALLOC(ret, size);
     }
-    else if (ALIGNMENT == 16)
+    else if constexpr (ALIGNMENT == 16)
     {
         ret = ll_aligned_malloc_16(size);
     }
-    else if (ALIGNMENT == 32)
+    else if constexpr (ALIGNMENT == 32)
     {
         ret = ll_aligned_malloc_32(size);
     }
+    else if constexpr (ALIGNMENT == 64)
+    {
+        ret = ll_aligned_malloc_64(size);
+    }
     else
     {
         ret = ll_aligned_malloc_fallback(size, ALIGNMENT);
@@ -285,16 +310,20 @@ template<size_t ALIGNMENT>
 LL_FORCE_INLINE void ll_aligned_free(void* ptr)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
-    if (ALIGNMENT == LL_DEFAULT_HEAP_ALIGN)
+    if constexpr (ALIGNMENT == LL_DEFAULT_HEAP_ALIGN)
     {
         LL_PROFILE_FREE(ptr);
         free(ptr);
     }
-    else if (ALIGNMENT == 16)
+    else if constexpr (ALIGNMENT == 16)
     {
         ll_aligned_free_16(ptr);
     }
-    else if (ALIGNMENT == 32)
+    else if constexpr (ALIGNMENT == 32)
+    {
+        return ll_aligned_free_32(ptr);
+    }
+    else if constexpr (ALIGNMENT == 64)
     {
         return ll_aligned_free_32(ptr);
     }
diff --git a/indra/llcommon/llprofiler.h b/indra/llcommon/llprofiler.h
index f6a4d24747..6ff4341b69 100644
--- a/indra/llcommon/llprofiler.h
+++ b/indra/llcommon/llprofiler.h
@@ -74,23 +74,18 @@
 #define LL_PROFILER_CONFIGURATION           LL_PROFILER_CONFIG_FAST_TIMER
 #endif
 
-extern thread_local bool gProfilerEnabled;
-
 #if defined(LL_PROFILER_CONFIGURATION) && (LL_PROFILER_CONFIGURATION > LL_PROFILER_CONFIG_NONE)
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY || LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY_FAST_TIMER
         #include "tracy/Tracy.hpp"
 
-        // Enable OpenGL profiling
-        #define LL_PROFILER_ENABLE_TRACY_OPENGL 0
-
         // Enable RenderDoc labeling
-        #define LL_PROFILER_ENABLE_RENDER_DOC 0
+        //#define LL_PROFILER_ENABLE_RENDER_DOC 0
 
     #endif
 
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY
         #define LL_PROFILER_FRAME_END                   FrameMark
-        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );    gProfilerEnabled = true;
+        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );
         #define LL_RECORD_BLOCK_TIME(name)              ZoneScoped // Want descriptive names; was: ZoneNamedN( ___tracy_scoped_zone, #name, true );
         #define LL_PROFILE_ZONE_NAMED(name)             ZoneNamedN( ___tracy_scoped_zone, name, true );
         #define LL_PROFILE_ZONE_NAMED_COLOR(name,color) ZoneNamedNC( ___tracy_scopped_zone, name, color, true ) // RGB
@@ -133,7 +128,7 @@ extern thread_local bool gProfilerEnabled;
     #endif
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY_FAST_TIMER
         #define LL_PROFILER_FRAME_END                   FrameMark
-        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );    gProfilerEnabled = true;
+        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );
         #define LL_RECORD_BLOCK_TIME(name)              ZoneNamedN(___tracy_scoped_zone, #name, true);   const LLTrace::BlockTimer& LL_GLUE_TOKENS(block_time_recorder, __LINE__)(LLTrace::timeThisBlock(name)); (void)LL_GLUE_TOKENS(block_time_recorder, __LINE__);
         #define LL_PROFILE_ZONE_NAMED(name)             ZoneNamedN( ___tracy_scoped_zone, #name, true );
         #define LL_PROFILE_ZONE_NAMED_COLOR(name,color) ZoneNamedNC( ___tracy_scopped_zone, name, color, true ) // RGB
@@ -158,23 +153,20 @@ extern thread_local bool gProfilerEnabled;
 #endif // LL_PROFILER
 
 #if LL_PROFILER_ENABLE_TRACY_OPENGL
-#define LL_PROFILE_GPU_ZONE(name)        TracyGpuZone(name)
-#define LL_PROFILE_GPU_ZONEC(name,color) TracyGpuZoneC(name,color)
+#define LL_PROFILE_GPU_ZONE(name)         TracyGpuZone(name)
+#define LL_PROFILE_GPU_ZONEC(name,color)  TracyGpuZoneC(name,color)
 #define LL_PROFILER_GPU_COLLECT           TracyGpuCollect
 #define LL_PROFILER_GPU_CONTEXT           TracyGpuContext
-
-// disable memory tracking (incompatible with GPU tracing
-#define LL_PROFILE_ALLOC(ptr, size)             (void)(ptr); (void)(size);
-#define LL_PROFILE_FREE(ptr)                    (void)(ptr);
+#define LL_PROFILER_GPU_CONTEXT_NAMED     TracyGpuContextName
 #else
-#define LL_PROFILE_GPU_ZONE(name)        (void)name;
-#define LL_PROFILE_GPU_ZONEC(name,color) (void)name;(void)color;
+#define LL_PROFILE_GPU_ZONE(name)           (void)name;
+#define LL_PROFILE_GPU_ZONEC(name,color)    (void)name;(void)color;
 #define LL_PROFILER_GPU_COLLECT
 #define LL_PROFILER_GPU_CONTEXT
+#define LL_PROFILER_GPU_CONTEXT_NAMED(name) (void)name;
+#endif // LL_PROFILER_ENABLE_TRACY_OPENGL
 
-#define LL_LABEL_OBJECT_GL(type, name, length, label)
-
-#if !LL_DARWIN && LL_PROFILER_CONFIGURATION > 1
+#if LL_PROFILER_CONFIGURATION > 1
 #define LL_PROFILE_ALLOC(ptr, size)             TracyAlloc(ptr, size)
 #define LL_PROFILE_FREE(ptr)                    TracyFree(ptr)
 #else
@@ -182,9 +174,7 @@ extern thread_local bool gProfilerEnabled;
 #define LL_PROFILE_FREE(ptr)                    (void)(ptr);
 #endif
 
-#endif
-
-#if LL_PROFILER_ENABLE_RENDER_DOC
+#ifdef LL_PROFILER_ENABLE_RENDER_DOC
 #define LL_LABEL_OBJECT_GL(type, name, length, label) glObjectLabel(type, name, length, label)
 #else
 #define LL_LABEL_OBJECT_GL(type, name, length, label)