profile locks

profile.bat

@@ -3,9 +3,9 @@
 :: `ping` is being used in place of `TIMEOUT`
 :: https://www.ibm.com/support/pages/timeout-command-run-batch-job-exits-immediately-and-returns-error-input-redirection-not-supported-exiting-process-immediately
 
-taskkill /im tracy.exe /f 2> nul
+taskkill /im tracy-profiler.exe /f 2> nul
 
-start %UserProfile%\Home\apps\tracy\capture.exe -o .tracy -f
+start tracy-capture.exe -o .tracy -f
 
 echo Launching app...
 build\clang-user-profiling\bin\powerplay.exe
@@ -21,4 +21,4 @@ if NOT %errorlevel% == 0 (
 ping -n 2 127.0.0.1 >NUL
 
 echo Launching tracy...
-start "" "%UserProfile%\Home\apps\tracy\Tracy.exe" ".tracy"
+start "" "tracy-profiler.exe" ".tracy"

src/app.c

@@ -129,6 +129,7 @@ void app_entry_point(void)
     G.exit_sf = sync_flag_alloc();
     G.exit_callbacks_mutex = sys_mutex_alloc();
     G.exit_callbacks_arena = arena_alloc(GIGABYTE(64));
+    G.arena = arena_alloc(GIGABYTE(64));
 
     u32 worker_count = 4;
     {
@@ -147,7 +148,6 @@ void app_entry_point(void)
 #endif
     }
 
-    G.arena = arena_alloc(GIGABYTE(64));
     G.write_path = initialize_write_directory(&G.arena, STR(WRITE_DIR));
 
     /* Startup logging */

src/common.h

@@ -373,6 +373,8 @@ GLOBAL const f64 *_f64_nan          = (f64 *)&_f64_nan_u64;
 typedef __uint128_t u128;
 
 INLINE b32 u128_eq(u128 a, u128 b) { return a == b; }
+INLINE b32 u128_lt(u128 a, u128 b) { return a < b; }
+INLINE b32 u128_gt(u128 a, u128 b) { return a > b; }
 INLINE u128 u128_xor_u8(u128 a, u8 b) { return a ^ b; }
 INLINE u128 u128_mul(u128 a, u128 b) { return a * b; }
 
@@ -389,6 +391,8 @@ INLINE u128 u128_mul(u128 a, u128 b) { return a * b; }
 typedef struct { u64 hi; u64 lo; } u128;
 
 INLINE b32 u128_eq(u128 a, u128 b) { return a.hi == b.hi && a.lo == b.lo; }
+INLINE b32 u128_lt(u128 a, u128 b) { return ((((a.hi > b.hi) - (a.hi < b.hi)) << 1) + ((a.lo > b.lo) - (a.lo < b.lo))) < 0; }
+INLINE b32 u128_gt(u128 a, u128 b) { return ((((a.hi > b.hi) - (a.hi < b.hi)) << 1) + ((a.lo > b.lo) - (a.lo < b.lo))) > 0; }
 INLINE u128 u128_xor_u8(u128 a, u8 b) { return U128(a.hi, a.lo ^ b); }
 
 /* https://www.codeproject.com/Tips/784635/UInt-Bit-Operations */
@@ -685,13 +689,28 @@ INLINE f64 clamp_f64(f64 v, f64 min, f64 max) { return v < min ? min : v > max ?
 #endif
 INLINE void __prof_zone_cleanup_func(TracyCZoneCtx *__tracy_ctx) { TracyCZoneEnd(*__tracy_ctx); }
 
-#define __profalloc(ptr, size)      TracyCAlloc(ptr, size)
+#define __profalloc(ptr, size)      TracyCAlloc((ptr), (size))
-#define __proffree(ptr)             TracyCFree(ptr)
+#define __proffree(ptr)             TracyCFree((ptr))
-#define __profmsg(txt, len, col)    TracyCMessageC(txt, len, col);
+#define __profmsg(txt, len, col)    TracyCMessageC((txt), (len), (col));
-#define __profframe(name)           TracyCFrameMarkNamed(name)
+#define __profframe(name)           TracyCFrameMarkNamed((name))
+#define __profthread(name)          TracyCSetThreadName((name))
+
+#define __proflock_ctx TracyCSharedLockCtx
+#define __proflock_alloc(ctx) TracyCSharedLockAnnounce((ctx))
+#define __proflock_release(ctx) TracyCSharedLockTerminate((ctx))
+#define __proflock_before_exclusive_lock(ctx) TracyCSharedLockBeforeExclusiveLock((ctx))
+#define __proflock_after_exclusive_lock(ctx) TracyCSharedLockAfterExclusiveLock((ctx))
+#define __proflock_after_exclusive_unlock(ctx) TracyCSharedLockAfterExclusiveUnlock((ctx))
+#define __proflock_after_try_exclusive_lock(ctx, acquired) TracyCSharedLockAfterTryExclusiveLock((ctx), (acquired))
+#define __proflock_before_shared_lock(ctx) TracyCSharedLockBeforeSharedLock((ctx))
+#define __proflock_after_shared_lock(ctx) TracyCSharedLockAfterSharedLock((ctx))
+#define __proflock_after_shared_unlock(ctx) TracyCSharedLockAfterSharedUnlock((ctx))
+#define __proflock_after_try_shared_lock(ctx, acquired) TracyCSharedLockAfterTrySharedLock((ctx), (acquired))
+#define __proflock_mark(ctx) TracyCSharedLockMark((ctx))
+#define __proflock_custom_name(ctx, name, len) TracyCSharedLockCustomName((ctx), (name), (len))
 
 #if PROFILING_CAPTURE_FRAME_IMAGE
-# define __profframeimage(image, width, height, offset, flipped)    TracyCFrameImage(image, width, height, offset, flipped);
+# define __profframeimage(image, width, height, offset, flipped)    TracyCFrameImage((image), (width), (height), (offset), (flipped));
 #else
 # define __profframeimage(image, width, height, offset, flipped)
 #endif  /* PROFILING_CAPTURE_FRAME_IMAGE */
@@ -706,7 +725,21 @@ INLINE void __prof_zone_cleanup_func(TracyCZoneCtx *__tracy_ctx) { TracyCZoneEnd
 #define __proffree(ptr)
 #define __profmsg(txt, len, col)
 #define __profframe(name)
+#define __profthread(name)
 #define __profframeimage(image, width, height, offset, flipped)
+#define __proflock_ctx
+#define __proflock_alloc(ctx)
+#define __proflock_release(ctx)
+#define __proflock_before_exclusive_lock(ctx)
+#define __proflock_after_exclusive_lock(ctx)
+#define __proflock_after_exclusive_unlock(ctx)
+#define __proflock_after_try_exclusive_lock(ctx, acquired)
+#define __proflock_before_shared_lock(ctx)
+#define __proflock_after_shared_lock(ctx)
+#define __proflock_after_shared_unlock(ctx)
+#define __proflock_after_try_shared_lock(ctx, acquired)
+#define __proflock_mark(ctx)
+#define __proflock_custom_name(ctx, name, len)
 
 #endif  /* PROFILING */
 

src/game.c

@@ -302,7 +302,8 @@ INTERNAL void spawn_test_entities(void)
         e->layer = GAME_LAYER_RELATIVE_WEAPON;
 
         entity_enable_prop(e, ENTITY_PROP_WEAPON);
-        e->trigger_delay = 1.0f / 10.0f;
+        //e->trigger_delay = 1.0f / 10.0f;
+        e->trigger_delay = 1.0f / 100.0f;
 
         player_ent->equipped = e->handle;
     }

src/intrinsics.h

@@ -115,7 +115,7 @@ INLINE void ix_pause(void)
     _mm_pause();
 }
 
-INLINE i64 ix_clock(void)
+INLINE u64 ix_clock(void)
 {
     return __rdtsc();
 }

src/math.h

@@ -224,7 +224,7 @@ INLINE f32 math_ln(f32 x)
     LOCAL_PERSIST const f32 ln2_hi = 6.9313812256e-01f;
     LOCAL_PERSIST const f32 ln2_lo = 9.0580006145e-06f;
 
-    i32 x_int = *(u32 *)&x;
+    i32 x_int = *(i32 *)&x;
 
     i32 k = 0;
     if (x_int < 0x00800000) {
@@ -238,7 +238,7 @@ INLINE f32 math_ln(f32 x)
         }
         k -= 25;
         x *= two_p25;
-        x_int = *(u32 *)&x;
+        x_int = *(i32 *)&x;
     } else if (x_int >= 0x7f800000) {
         return x + x;
     }
@@ -300,7 +300,7 @@ INLINE f32 math_exp(f32 x)
     LOCAL_PERSIST const f32 two_m100 = 7.8886090522e-31f;
 
     u32 x_uint = *(u32 *)&x;
-    i32 x_sign_bit = (x_uint >> 31) & 1;
+    u32 x_sign_bit = (x_uint >> 31) & 1;
     x_uint &= 0x7fffffff;
 
     /* Filter out non-finite argument */

src/sys.h

@@ -336,6 +336,9 @@ void sys_window_cursor_disable_clip(struct sys_window *sys_window);
 
 struct sys_mutex {
     u64 handle;
+#if PROFILING
+    __proflock_ctx profiling_ctx;
+#endif
 #if RTC
     u64 owner_tid;
     struct atomic_i64 count;

src/sys_win32.c

@@ -679,6 +679,7 @@ INTERNAL void win32_update_window_from_system(struct win32_window *window);
 
 INTERNAL void win32_window_process_event(struct win32_window *window, struct sys_event event)
 {
+    __prof;
     struct sys_lock lock = sys_mutex_lock_e(&window->event_callbacks_mutex);
     for (u64 i = 0; i < window->event_callbacks_count; ++i) {
         window->event_callbacks[i](event);
@@ -1356,10 +1357,12 @@ void sys_window_cursor_disable_clip(struct sys_window *sys_window)
 struct sys_mutex sys_mutex_alloc(void)
 {
     __prof;
+    struct sys_mutex mutex = ZI;
+
+    __proflock_alloc(mutex.profiling_ctx);
     SRWLOCK srwlock = SRWLOCK_INIT;
-    struct sys_mutex mutex = {
+    mutex.handle = *(u64 *)&srwlock;
-        .handle = *(u64 *)&srwlock
+
-    };
     return mutex;
 }
 
@@ -1367,6 +1370,7 @@ void sys_mutex_release(struct sys_mutex *mutex)
 {
     __prof;
     (UNUSED)mutex;
+    __proflock_release(mutex->profiling_ctx);
     /* Mutex should be unlocked */
     ASSERT(atomic_i64_eval(&mutex->count) == 0);
 }
@@ -1374,7 +1378,9 @@ void sys_mutex_release(struct sys_mutex *mutex)
 struct sys_lock sys_mutex_lock_e(struct sys_mutex *mutex)
 {
     __prof;
+    __proflock_before_exclusive_lock(mutex->profiling_ctx);
     AcquireSRWLockExclusive((SRWLOCK *)&mutex->handle);
+    __proflock_after_exclusive_lock(mutex->profiling_ctx);
 #if RTC
     mutex->owner_tid = (u64)GetCurrentThreadId();
     atomic_i64_inc_eval(&mutex->count);
@@ -1388,7 +1394,9 @@ struct sys_lock sys_mutex_lock_e(struct sys_mutex *mutex)
 struct sys_lock sys_mutex_lock_s(struct sys_mutex *mutex)
 {
     __prof;
+    __proflock_before_shared_lock(mutex->profiling_ctx);
     AcquireSRWLockShared((SRWLOCK *)&mutex->handle);
+    __proflock_after_shared_lock(mutex->profiling_ctx);
 #if RTC
     atomic_i64_inc_eval(&mutex->count);
 #endif
@@ -1406,8 +1414,10 @@ void sys_mutex_unlock(struct sys_lock *lock)
 #endif
     if (lock->exclusive) {
         ReleaseSRWLockExclusive((SRWLOCK *)&lock->mutex->handle);
+        __proflock_after_exclusive_unlock(lock->mutex->profiling_ctx);
     } else {
         ReleaseSRWLockShared((SRWLOCK *)&lock->mutex->handle);
+        __proflock_after_shared_unlock(lock->mutex->profiling_ctx);
     }
     MEMZERO_STRUCT(lock);
 }
@@ -1489,7 +1499,22 @@ void sys_condition_variable_wait(struct sys_condition_variable *cv, struct sys_l
     atomic_i64_dec_eval(&mutex->count);
 #endif
     struct win32_condition_variable *w32cv = (struct win32_condition_variable *)cv->handle;
+
+    /* TODO: Correct profiling of internal condition variable sleep / wait mutex state */
+    if (exclusive) {
+        __proflock_after_exclusive_unlock(mutex->profiling_ctx);
+    } else {
+        __proflock_after_shared_unlock(mutex->profiling_ctx);
+    }
     SleepConditionVariableSRW(&w32cv->condition_variable, (SRWLOCK *)&mutex->handle, INFINITE, exclusive ? 0 : CONDITION_VARIABLE_LOCKMODE_SHARED);
+    if (exclusive) {
+        __proflock_before_exclusive_lock(mutex->profiling_ctx);
+        __proflock_after_exclusive_lock(mutex->profiling_ctx);
+    } else {
+        __proflock_before_shared_lock(mutex->profiling_ctx);
+        __proflock_after_shared_lock(mutex->profiling_ctx);
+    }
+
 #if RTC
     atomic_i64_inc_eval(&mutex->count);
     if (exclusive) {
@@ -1513,7 +1538,22 @@ void sys_condition_variable_wait_time(struct sys_condition_variable *cv, struct
 #endif
     struct win32_condition_variable *w32cv = (struct win32_condition_variable *)cv->handle;
     u32 ms = (u32)math_round_to_int((f32)seconds * 1000.f);
+
+    /* TODO: Correct profiling of internal condition variable sleep / wait mutex state */
+    if (exclusive) {
+        __proflock_after_exclusive_unlock(mutex->profiling_ctx);
+    } else {
+        __proflock_after_shared_unlock(mutex->profiling_ctx);
+    }
     SleepConditionVariableSRW(&w32cv->condition_variable, (SRWLOCK *)&mutex->handle, ms, exclusive ? 0 : CONDITION_VARIABLE_LOCKMODE_SHARED);
+    if (exclusive) {
+        __proflock_before_exclusive_lock(mutex->profiling_ctx);
+        __proflock_after_exclusive_lock(mutex->profiling_ctx);
+    } else {
+        __proflock_before_shared_lock(mutex->profiling_ctx);
+        __proflock_after_shared_lock(mutex->profiling_ctx);
+    }
+
 #if RTC
     atomic_i64_inc_eval(&mutex->count);
     if (exclusive) {
@@ -1639,6 +1679,7 @@ INTERNAL void win32_thread_release_locked(struct sys_lock *lock, struct win32_th
 INTERNAL DWORD WINAPI win32_thread_proc(LPVOID vt)
 {
     struct win32_thread *t = (struct win32_thread *)vt;
+    __profthread(t->thread_name_cstr);
 
     /* Initialize COM */
     CoInitializeEx(NULL, COINIT_APARTMENTTHREADED);

src/third_party/tracy/client/TracyProfiler.cpp

@@ -79,6 +79,7 @@
 #include "TracyThread.hpp"
 #include "TracyArmCpuTable.hpp"
 #include "TracySysTrace.hpp"
+#include "TracyLock.hpp"
 #include "../tracy/TracyC.h"
 
 #if defined TRACY_MANUAL_LIFETIME && !defined(TRACY_DELAYED_INIT)
@@ -4792,171 +4793,128 @@ TRACY_API void ___tracy_emit_gpu_time_sync_serial( const struct ___tracy_gpu_tim
 
 struct __tracy_lockable_context_data
 {
-    uint32_t m_id;
+    tracy::LockableCtx ctx;
-#ifdef TRACY_ON_DEMAND
-    std::atomic<uint32_t> m_lockCount;
-    std::atomic<bool> m_active;
-#endif
 };
 
 TRACY_API struct __tracy_lockable_context_data* ___tracy_announce_lockable_ctx( const struct ___tracy_source_location_data* srcloc )
 {
-    struct __tracy_lockable_context_data *lockdata = (__tracy_lockable_context_data*)tracy::tracy_malloc( sizeof( __tracy_lockable_context_data ) );
+    static_assert(sizeof(struct ___tracy_source_location_data) == sizeof(tracy::SourceLocationData));  // C srcloc struct must fit in cpp srcloc struct
-    lockdata->m_id =tracy:: GetLockCounter().fetch_add( 1, std::memory_order_relaxed );
+    struct __tracy_lockable_context_data *lockdata = (__tracy_lockable_context_data *)tracy::tracy_malloc(sizeof(__tracy_lockable_context_data));
-#ifdef TRACY_ON_DEMAND
-    new(&lockdata->m_lockCount) std::atomic<uint32_t>( 0 );
-    new(&lockdata->m_active) std::atomic<bool>( false );
-#endif
-    assert( lockdata->m_id != (std::numeric_limits<uint32_t>::max)() );
 
-    auto item = tracy::Profiler::QueueSerial();
+    // Because Tracy deletes the "=" operator
-    tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockAnnounce );
+    tracy::LockableCtx ctx { (tracy::SourceLocationData *)srcloc };
-    tracy::MemWrite( &item->lockAnnounce.id, lockdata->m_id );
+    memcpy(&lockdata->ctx, &ctx, sizeof(ctx));
-    tracy::MemWrite( &item->lockAnnounce.time, tracy::Profiler::GetTime() );
-    tracy::MemWrite( &item->lockAnnounce.lckloc, (uint64_t)srcloc );
-    tracy::MemWrite( &item->lockAnnounce.type, tracy::LockType::Lockable );
-#ifdef TRACY_ON_DEMAND
-    tracy::GetProfiler().DeferItem( *item );
-#endif
-    tracy::Profiler::QueueSerialFinish();
 
     return lockdata;
 }
 
 TRACY_API void ___tracy_terminate_lockable_ctx( struct __tracy_lockable_context_data* lockdata )
 {
-    auto item = tracy::Profiler::QueueSerial();
+    lockdata->ctx.~LockableCtx();
-    tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockTerminate );
-    tracy::MemWrite( &item->lockTerminate.id, lockdata->m_id );
-    tracy::MemWrite( &item->lockTerminate.time, tracy::Profiler::GetTime() );
-#ifdef TRACY_ON_DEMAND
-    tracy::GetProfiler().DeferItem( *item );
-#endif
-    tracy::Profiler::QueueSerialFinish();
-
-#ifdef TRACY_ON_DEMAND
-    lockdata->m_lockCount.~atomic();
-    lockdata->m_active.~atomic();
-#endif
     tracy::tracy_free((void*)lockdata);
 }
 
 TRACY_API int32_t ___tracy_before_lock_lockable_ctx( struct __tracy_lockable_context_data* lockdata )
 {
-#ifdef TRACY_ON_DEMAND
+    return lockdata->ctx.BeforeLock();
-    bool queue = false;
-    const auto locks = lockdata->m_lockCount.fetch_add( 1, std::memory_order_relaxed );
-    const auto active = lockdata->m_active.load( std::memory_order_relaxed );
-    if( locks == 0 || active )
-    {
-        const bool connected = tracy::GetProfiler().IsConnected();
-        if( active != connected ) lockdata->m_active.store( connected, std::memory_order_relaxed );
-        if( connected ) queue = true;
-    }
-    if( !queue ) return static_cast<int32_t>(false);
-#endif
-
-    auto item = tracy::Profiler::QueueSerial();
-    tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockWait );
-    tracy::MemWrite( &item->lockWait.thread, tracy::GetThreadHandle() );
-    tracy::MemWrite( &item->lockWait.id, lockdata->m_id );
-    tracy::MemWrite( &item->lockWait.time, tracy::Profiler::GetTime() );
-    tracy::Profiler::QueueSerialFinish();
-    return static_cast<int32_t>(true);
 }
 
 TRACY_API void ___tracy_after_lock_lockable_ctx( struct __tracy_lockable_context_data* lockdata )
 {
-    auto item = tracy::Profiler::QueueSerial();
+    lockdata->ctx.AfterLock();
-    tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockObtain );
-    tracy::MemWrite( &item->lockObtain.thread, tracy::GetThreadHandle() );
-    tracy::MemWrite( &item->lockObtain.id, lockdata->m_id );
-    tracy::MemWrite( &item->lockObtain.time, tracy::Profiler::GetTime() );
-    tracy::Profiler::QueueSerialFinish();
 }
 
 TRACY_API void ___tracy_after_unlock_lockable_ctx( struct __tracy_lockable_context_data* lockdata )
 {
-#ifdef TRACY_ON_DEMAND
+    lockdata->ctx.AfterUnlock();
-    lockdata->m_lockCount.fetch_sub( 1, std::memory_order_relaxed );
-    if( !lockdata->m_active.load( std::memory_order_relaxed ) ) return;
-    if( !tracy::GetProfiler().IsConnected() )
-    {
-        lockdata->m_active.store( false, std::memory_order_relaxed );
-        return;
-    }
-#endif
-
-    auto item = tracy::Profiler::QueueSerial();
-    tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockRelease );
-    tracy::MemWrite( &item->lockRelease.id, lockdata->m_id );
-    tracy::MemWrite( &item->lockRelease.time, tracy::Profiler::GetTime() );
-    tracy::Profiler::QueueSerialFinish();
 }
 
 TRACY_API void ___tracy_after_try_lock_lockable_ctx( struct __tracy_lockable_context_data* lockdata, int32_t acquired )
 {
-#ifdef TRACY_ON_DEMAND
+    lockdata->ctx.AfterTryLock(acquired);
-    if( !acquired ) return;
-
-    bool queue = false;
-    const auto locks = lockdata->m_lockCount.fetch_add( 1, std::memory_order_relaxed );
-    const auto active = lockdata->m_active.load( std::memory_order_relaxed );
-    if( locks == 0 || active )
-    {
-        const bool connected = tracy::GetProfiler().IsConnected();
-        if( active != connected ) lockdata->m_active.store( connected, std::memory_order_relaxed );
-        if( connected ) queue = true;
-    }
-    if( !queue ) return;
-#endif
-
-    if( acquired )
-    {
-        auto item = tracy::Profiler::QueueSerial();
-        tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockObtain );
-        tracy::MemWrite( &item->lockObtain.thread, tracy::GetThreadHandle() );
-        tracy::MemWrite( &item->lockObtain.id, lockdata->m_id );
-        tracy::MemWrite( &item->lockObtain.time, tracy::Profiler::GetTime() );
-        tracy::Profiler::QueueSerialFinish();
-    }
 }
 
-TRACY_API void ___tracy_mark_lockable_ctx( struct __tracy_lockable_context_data* lockdata, const struct ___tracy_source_location_data* srcloc )
+TRACY_API void ___tracy_mark_lockable_ctx(struct __tracy_lockable_context_data *lockdata, const struct ___tracy_source_location_data *srcloc)
 {
-#ifdef TRACY_ON_DEMAND
+    lockdata->ctx.Mark((tracy::SourceLocationData *)srcloc);
-    const auto active = lockdata->m_active.load( std::memory_order_relaxed );
-    if( !active ) return;
-    const auto connected = tracy::GetProfiler().IsConnected();
-    if( !connected )
-    {
-        if( active ) lockdata->m_active.store( false, std::memory_order_relaxed );
-        return;
-    }
-#endif
-
-    auto item = tracy::Profiler::QueueSerial();
-    tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockMark );
-    tracy::MemWrite( &item->lockMark.thread, tracy::GetThreadHandle() );
-    tracy::MemWrite( &item->lockMark.id, lockdata->m_id );
-    tracy::MemWrite( &item->lockMark.srcloc, (uint64_t)srcloc );
-    tracy::Profiler::QueueSerialFinish();
 }
 
-TRACY_API void ___tracy_custom_name_lockable_ctx( struct __tracy_lockable_context_data* lockdata, const char* name, size_t nameSz )
+TRACY_API void ___tracy_custom_name_lockable_ctx(struct __tracy_lockable_context_data *lockdata, const char *name, size_t nameSz)
 {
-    assert( nameSz < (std::numeric_limits<uint16_t>::max)() );
+    lockdata->ctx.CustomName(name, nameSz);
-    auto ptr = (char*)tracy::tracy_malloc( nameSz );
+}
-    memcpy( ptr, name, nameSz );
+
-    auto item = tracy::Profiler::QueueSerial();
+struct __tracy_shared_lockable_context_data
-    tracy::MemWrite( &item->hdr.type, tracy::QueueType::LockName );
+{
-    tracy::MemWrite( &item->lockNameFat.id, lockdata->m_id );
+    tracy::SharedLockableCtx ctx;
-    tracy::MemWrite( &item->lockNameFat.name, (uint64_t)ptr );
+};
-    tracy::MemWrite( &item->lockNameFat.size, (uint16_t)nameSz );
+
-#ifdef TRACY_ON_DEMAND
+TRACY_API struct __tracy_shared_lockable_context_data* ___tracy_announce_shared_lockable_ctx( const struct ___tracy_source_location_data* srcloc )
-    tracy::GetProfiler().DeferItem( *item );
+{
-#endif
+    static_assert(sizeof(struct ___tracy_source_location_data) == sizeof(tracy::SourceLocationData));  // C srcloc struct must fit in cpp srcloc struct
-    tracy::Profiler::QueueSerialFinish();
+    struct __tracy_shared_lockable_context_data *lockdata = (__tracy_shared_lockable_context_data *)tracy::tracy_malloc(sizeof(__tracy_shared_lockable_context_data));
+
+    // Because Tracy deletes the "=" operator
+    tracy::SharedLockableCtx ctx { (tracy::SourceLocationData *)srcloc };
+    memcpy(&lockdata->ctx, &ctx, sizeof(ctx));
+
+    return lockdata;
+}
+
+TRACY_API void ___tracy_terminate_shared_lockable_ctx( struct __tracy_shared_lockable_context_data* lockdata )
+{
+    lockdata->ctx.~SharedLockableCtx();
+    tracy::tracy_free((void*)lockdata);
+}
+
+TRACY_API int32_t ___tracy_before_exclusive_lock_shared_lockable_ctx( struct __tracy_shared_lockable_context_data* lockdata )
+{
+    return lockdata->ctx.BeforeLock();
+}
+
+TRACY_API void ___tracy_after_exclusive_lock_shared_lockable_ctx( struct __tracy_shared_lockable_context_data* lockdata )
+{
+    lockdata->ctx.AfterLock();
+}
+
+TRACY_API void ___tracy_after_exclusive_unlock_shared_lockable_ctx( struct __tracy_shared_lockable_context_data* lockdata )
+{
+    lockdata->ctx.AfterUnlock();
+}
+
+TRACY_API void ___tracy_after_try_exclusive_lock_shared_lockable_ctx( struct __tracy_shared_lockable_context_data* lockdata, int32_t acquired )
+{
+    lockdata->ctx.AfterTryLock(acquired);
+}
+
+TRACY_API int32_t ___tracy_before_shared_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata)
+{
+    return lockdata->ctx.BeforeLockShared();
+}
+
+TRACY_API void ___tracy_after_shared_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata)
+{
+    lockdata->ctx.AfterLockShared();
+}
+
+TRACY_API void ___tracy_after_try_shared_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata, int32_t acquired)
+{
+    lockdata->ctx.AfterTryLockShared(acquired);
+}
+
+TRACY_API void ___tracy_after_shared_unlock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata)
+{
+    lockdata->ctx.AfterUnlockShared();
+}
+
+TRACY_API void ___tracy_mark_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata, const struct ___tracy_source_location_data *srcloc)
+{
+    lockdata->ctx.Mark((tracy::SourceLocationData *)srcloc);
+}
+
+TRACY_API void ___tracy_custom_name_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata, const char *name, size_t nameSz)
+{
+    lockdata->ctx.CustomName(name, nameSz);
 }
 
 TRACY_API int32_t ___tracy_connected( void )

src/third_party/tracy/common/TracyQueue.hpp

@@ -331,7 +331,7 @@ struct QueuePlotDataInt : public QueuePlotDataBase
     int64_t val;
 };
 
-struct QueuePlotDataFloat : public QueuePlotDataBase 
+struct QueuePlotDataFloat : public QueuePlotDataBase
 {
     float val;
 };
@@ -466,7 +466,7 @@ struct QueueGpuTimeSync
     int64_t cpuTime;
     uint8_t context;
 };
-    
+
 struct QueueGpuContextName
 {
     uint8_t context;

src/third_party/tracy/tracy/TracyC.h

@@ -39,6 +39,7 @@ TRACY_API void ___tracy_set_thread_name( const char* name );
 typedef const void* TracyCZoneCtx;
 
 typedef const void* TracyCLockCtx;
+typedef const void* TracyCSharedLockCtx;
 
 #define TracyCZone(c,x)
 #define TracyCZoneN(c,x,y)
@@ -111,6 +112,20 @@ typedef const void* TracyCLockCtx;
 #define TracyCLockMark(l)
 #define TracyCLockCustomName(l,x,y)
 
+#define TracyCSharedLockCtx(l)
+#define TracyCSharedLockAnnounce(l)
+#define TracyCSharedLockTerminate(l)
+#define TracyCSharedLockBeforeExclusiveLock(l)
+#define TracyCSharedLockAfterExclusiveLock(l)
+#define TracyCSharedLockAfterExclusiveUnl(l)
+#define TracyCSharedLockAfterTryExclusiveLock(l,x)
+#define TracyCSharedLockBeforeSharedLock(l)
+#define TracyCSharedLockAfterSharedLock(l)
+#define TracyCSharedLockAfterSharedUnl(l)
+#define TracyCSharedLockAfterTrySharedLock(l,x)
+#define TracyCSharedLockMark(l)
+#define TracyCSharedLockCustomName(l,x,y)
+
 #define TracyCIsConnected 0
 #define TracyCIsStarted 0
 
@@ -200,6 +215,7 @@ struct __tracy_lockable_context_data;
 typedef /*const*/ struct ___tracy_c_zone_context TracyCZoneCtx;
 
 typedef struct __tracy_lockable_context_data* TracyCLockCtx;
+typedef struct __tracy_lockable_context_data *TracyCSharedLockCtx;
 
 #ifdef TRACY_MANUAL_LIFETIME
 TRACY_API void ___tracy_startup_profiler(void);
@@ -365,6 +381,32 @@ TRACY_API void ___tracy_custom_name_lockable_ctx( struct __tracy_lockable_contex
 #define TracyCLockMark( lock ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__,  TracyFile, (uint32_t)TracyLine, 0 }; ___tracy_mark_lockable_ctx( lock, &TracyConcat(__tracy_source_location,TracyLine) );
 #define TracyCLockCustomName( lock, name, nameSz ) ___tracy_custom_name_lockable_ctx( lock, name, nameSz );
 
+TRACY_API struct __tracy_shared_lockable_context_data *___tracy_announce_shared_lockable_ctx(const struct ___tracy_source_location_data *srcloc);
+TRACY_API void ___tracy_terminate_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata);
+TRACY_API int32_t ___tracy_before_exclusive_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata);
+TRACY_API void ___tracy_after_exclusive_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata);
+TRACY_API void ___tracy_after_exclusive_unlock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata);
+TRACY_API void ___tracy_after_try_exclusive_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata, int32_t acquired);
+TRACY_API int32_t ___tracy_before_shared_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata);
+TRACY_API void ___tracy_after_shared_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata);
+TRACY_API void ___tracy_after_shared_unlock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata);
+TRACY_API void ___tracy_after_try_shared_lock_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata, int32_t acquired);
+TRACY_API void ___tracy_mark_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata, const struct ___tracy_source_location_data *srcloc);
+TRACY_API void ___tracy_custom_name_shared_lockable_ctx(struct __tracy_shared_lockable_context_data *lockdata, const char *name, size_t nameSz);
+
+#define TracyCSharedLockAnnounce( lock ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__,  TracyFile, (uint32_t)TracyLine, 0 }; lock = ___tracy_announce_shared_lockable_ctx( &TracyConcat(__tracy_source_location,TracyLine) );
+#define TracyCSharedLockTerminate( lock ) ___tracy_terminate_shared_lockable_ctx( lock );
+#define TracyCSharedLockBeforeExclusiveLock( lock ) ___tracy_before_exclusive_lock_shared_lockable_ctx( lock );
+#define TracyCSharedLockAfterExclusiveLock( lock ) ___tracy_after_exclusive_lock_shared_lockable_ctx( lock );
+#define TracyCSharedLockAfterExclusiveUnlock( lock ) ___tracy_after_exclusive_unlock_shared_lockable_ctx( lock );
+#define TracyCSharedLockAfterTryExclusiveLock( lock, acquired ) ___tracy_after_try_exclusive_lock_shared_lockable_ctx( lock, acquired );
+#define TracyCSharedLockBeforeSharedLock( lock ) ___tracy_before_shared_lock_shared_lockable_ctx( lock );
+#define TracyCSharedLockAfterSharedLock( lock ) ___tracy_after_shared_lock_shared_lockable_ctx( lock );
+#define TracyCSharedLockAfterSharedUnlock( lock ) ___tracy_after_shared_unlock_shared_lockable_ctx( lock );
+#define TracyCSharedLockAfterTrySharedLock( lock, acquired ) ___tracy_after_try_shared_lock_shared_lockable_ctx( lock, acquired );
+#define TracyCSharedLockMark( lock ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__,  TracyFile, (uint32_t)TracyLine, 0 }; ___tracy_mark_shared_lockable_ctx( lock, &TracyConcat(__tracy_source_location,TracyLine) );
+#define TracyCSharedLockCustomName( lock, name, nameSz ) ___tracy_custom_name_shared_lockable_ctx( lock, name, nameSz );
+
 #define TracyCIsConnected ___tracy_connected()
 
 #ifdef TRACY_FIBERS

src/user.c

@@ -462,9 +462,9 @@ INTERNAL SORT_COMPARE_FUNC_DEF(entity_draw_order_cmp, arg_a, arg_b, udata)
     }
     if (res == 0) {
         /* Sort by sprite */
-        u64 a_cmp = a->sprite.hash;
+        u128 a_cmp = a->sprite.hash;
-        u64 b_cmp = b->sprite.hash;
+        u128 b_cmp = b->sprite.hash;
-        res = (a_cmp < b_cmp) - (a_cmp > b_cmp);
+        res = u128_lt(a_cmp, b_cmp) - u128_gt(a_cmp, b_cmp);
     }
     if (res == 0) {
         /* Sort by activation */
@@ -733,7 +733,7 @@ INTERNAL void user_update(void)
         if (!entity_is_valid_and_active(ent)) continue;
 
         /* How much time between camera shakes */
-        const f32 frequency = 0.01;
+        const f32 frequency = 0.01f;
         f32 shake = ent->shake;
         if (shake > 0) {
             u64 basis = hash_fnv64(HASH_FNV64_BASIS, BUFFER_FROM_STRUCT(&ent->handle));
@@ -960,7 +960,7 @@ INTERNAL void user_update(void)
 
                 f32 opacity_b = clamp_f32(1.f - (v2_dot(vdc, vdb) / v2_len_sq(vdc)), 0, 1);
 
-                f32 thickness = 0.01;
+                f32 thickness = 0.01f;
                 u32 color_start = RGBA_32_F(1, 0.5, 0, opacity_a);
                 u32 color_end = RGBA_32_F(1, 0.8, 0.4, opacity_b);