all tasks profiler

src/julia_internal.h

@@ -206,13 +206,17 @@ JL_DLLEXPORT void jl_unlock_profile_wr(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_LEA
 int jl_lock_stackwalk(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_ENTER;
 void jl_unlock_stackwalk(int lockret) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_LEAVE;
 
-arraylist_t jl_get_all_tasks_arraylist(void) JL_NOTSAFEPOINT;
+arraylist_t *jl_get_all_tasks_arraylist(void) JL_NOTSAFEPOINT;
 int jl_rec_backtrace(jl_task_t *t) JL_NOTSAFEPOINT;
 extern volatile struct _jl_bt_element_t *profile_bt_data_prof;
 extern volatile size_t profile_bt_size_max;
 extern volatile size_t profile_bt_size_cur;
 extern volatile int profile_running;
 extern volatile int profile_all_tasks;
+STATIC_INLINE int all_tasks_profile_running(void) JL_NOTSAFEPOINT
+{
+    return profile_running && profile_all_tasks;
+}
 
 // number of cycles since power-on
 static inline uint64_t cycleclock(void) JL_NOTSAFEPOINT

src/signals-unix.c

@@ -724,11 +724,11 @@ void jl_profile_task_unix(void)
     // Get a random task for which we know that we won't
     // drop a sample in the profiler
     // TODO(Diogo, Nick): should we put an upper bound on the number of iterations?
-    arraylist_t tasks = jl_get_all_tasks_arraylist();
+    arraylist_t *tasks = jl_get_all_tasks_arraylist();
     jl_task_t *t = NULL;
     uint64_t seed = jl_rand();
     while (1) {
-        t = tasks.items[cong(tasks.len, UINT64_MAX, &seed)];
+        t = tasks->items[cong(tasks->len, UINT64_MAX, &seed)];
         assert(t == NULL || jl_is_task(t));
         if (t == NULL) {
             continue;
@@ -739,7 +739,8 @@ void jl_profile_task_unix(void)
         }
         break;
     }
-    arraylist_free(&tasks);
+    arraylist_free(tasks);
+    free(tasks);
     assert(t != NULL);
 
     int tid = jl_rec_backtrace(t);

src/signals-win.c

@@ -398,7 +398,7 @@ static DWORD WINAPI profile_bt( LPVOID lparam )
     while (1) {
         DWORD timeout_ms = nsecprof / (GIGA / 1000);
         Sleep(timeout_ms > 0 ? timeout_ms : 1);
-        if (running) {
+        if (profile_running) {
             if (jl_profile_is_buffer_full()) {
                 jl_profile_stop_timer(); // does not change the thread state
                 SuspendThread(GetCurrentThread());
@@ -420,16 +420,16 @@ static DWORD WINAPI profile_bt( LPVOID lparam )
 
                 jl_ptls_t ptls = jl_atomic_load_relaxed(&jl_all_tls_states)[0]; // given only profiling hMainThread
 
-                // store threadid but add 1 as 0 is preserved to indicate end of block
+                // META_OFFSET_THREADID store threadid but add 1 as 0 is preserved to indicate end of block
                 profile_bt_data_prof[profile_bt_size_cur++].uintptr = ptls->tid + 1;
 
-                // store task id (never null)
+                // META_OFFSET_TASKID store task id (never null)
                 profile_bt_data_prof[profile_bt_size_cur++].jlvalue = (jl_value_t*)jl_atomic_load_relaxed(&ptls->current_task);
 
-                // store cpu cycle clock
+                // META_OFFSET_CPUCYCLECLOCK store cpu cycle clock
                 profile_bt_data_prof[profile_bt_size_cur++].uintptr = cycleclock();
 
-                // store whether thread is sleeping but add 1 as 0 is preserved to indicate end of block
+                // META_OFFSET_SLEEPSTATE store whether thread is sleeping but add 1 as 0 is preserved to indicate end of block
                 profile_bt_data_prof[profile_bt_size_cur++].uintptr = jl_atomic_load_relaxed(&ptls->sleep_check_state) + 1;
 
                 // Mark the end of this block with two 0's
@@ -477,21 +477,21 @@ JL_DLLEXPORT int jl_profile_start_timer(uint8_t all_tasks)
             return -2;
         }
     }
-    if (running == 0) {
+    if (profile_running == 0) {
         // Failure to change the timer resolution is not fatal. However, it is important to
         // ensure that the timeBeginPeriod/timeEndPeriod is paired.
         if (TIMERR_NOERROR != timeBeginPeriod(timecaps.wPeriodMin))
             timecaps.wPeriodMin = 0;
     }
     profile_all_tasks = all_tasks;
-    running = 1; // set `running` finally
+    profile_running = 1; // set `profile_running` finally
     return 0;
 }
 JL_DLLEXPORT void jl_profile_stop_timer(void)
 {
-    if (running && timecaps.wPeriodMin)
+    if (profile_running && timecaps.wPeriodMin)
         timeEndPeriod(timecaps.wPeriodMin);
-    running = 0;
+    profile_running = 0;
     profile_all_tasks = 0;
 }
 

src/stackwalk.c

@@ -867,20 +867,14 @@ _os_ptr_munge(uintptr_t ptr)
 #define _OS_PTR_UNMUNGE(_ptr) _os_ptr_munge((uintptr_t)(_ptr))
 #endif
 
-
-STATIC_INLINE int all_tasks_profile_running(void)
-{
-    return profile_running && profile_all_tasks;
-}
-
 int jl_rec_backtrace(jl_task_t *t) JL_NOTSAFEPOINT
 {
     jl_task_t *ct = NULL;
     jl_ptls_t ptls = NULL;
     int16_t tid = INT16_MAX;
     if (!all_tasks_profile_running()) {
-        jl_task_t *ct = jl_current_task;
-        jl_ptls_t ptls = ct->ptls;
+        ct = jl_current_task;
+        ptls = ct->ptls;
         ptls->bt_size = 0;
         tid = ptls->tid;
     }

src/task.c

@@ -1116,12 +1116,11 @@ JL_DLLEXPORT jl_task_t *jl_get_current_task(void)
 }
 
 extern int gc_first_tid;
-
 // Select a task at random to profile. Racy: `live_tasks` can change at any time.
-arraylist_t jl_get_all_tasks_arraylist(void) JL_NOTSAFEPOINT
+arraylist_t *jl_get_all_tasks_arraylist(void) JL_NOTSAFEPOINT
 {
-    arraylist_t tasks;
-    arraylist_new(&tasks, 0);
+    arraylist_t *tasks = (arraylist_t*)malloc_s(sizeof(arraylist_t));
+    arraylist_new(tasks, 0);
     size_t nthreads = jl_atomic_load_acquire(&jl_n_threads);
     jl_ptls_t *allstates = jl_atomic_load_relaxed(&jl_all_tls_states);
     for (size_t i = 0; i < nthreads; i++) {
@@ -1135,21 +1134,20 @@ arraylist_t jl_get_all_tasks_arraylist(void) JL_NOTSAFEPOINT
         }
         jl_task_t *t = ptls2->root_task;
         if (t->stkbuf != NULL) {
-            arraylist_push(&tasks, t);
+            arraylist_push(tasks, t);
         }
         small_arraylist_t *live_tasks = &ptls2->gc_tls.heap.live_tasks;
         size_t n = mtarraylist_length(live_tasks);
         for (size_t i = 0; i < n; i++) {
             jl_task_t *t = (jl_task_t*)mtarraylist_get(live_tasks, i);
             if (t->stkbuf != NULL) {
-                arraylist_push(&tasks, t);
+                arraylist_push(tasks, t);
             }
         }
     }
     return tasks;
 }
 
-
 #ifdef JL_HAVE_ASYNCIFY
 JL_DLLEXPORT jl_ucontext_t *task_ctx_ptr(jl_task_t *t)
 {