move timing_stack field from jl_task_t onto the stack also

Author: vtjnash

src/init.c

@@ -745,9 +745,6 @@ void _julia_init(JL_IMAGE_SEARCH rel)
 
     jl_init_tasks();
     jl_init_root_task(stack_lo, stack_hi);
-#ifdef ENABLE_TIMINGS
-    jl_root_task->timing_stack = jl_root_timing;
-#endif
     jl_init_common_symbols();
     jl_init_flisp();
     jl_init_serializer();

src/julia.h

@@ -1765,8 +1765,6 @@ typedef struct _jl_task_t {
 
     // saved gc stack top for context switches
     jl_gcframe_t *gcstack;
-
-    jl_timing_block_t *timing_stack;
 } jl_task_t;
 
 #define JL_TASK_STATE_RUNNABLE 0

src/julia_threads.h

@@ -207,6 +207,7 @@ struct _jl_tls_states_t {
     struct _jl_task_t *previous_task;
 #endif
     struct _jl_task_t *root_task;
+    struct _jl_timing_block_t *timing_stack;
     void *stackbase;
     size_t stacksize;
     jl_ucontext_t base_ctx; // base context of stack

src/rtutils.c

@@ -221,7 +221,7 @@ JL_DLLEXPORT void jl_enter_handler(jl_handler_t *eh)
     eh->world_age = ptls->world_age;
     current_task->eh = eh;
 #ifdef ENABLE_TIMINGS
-    eh->timing_stack = current_task->timing_stack;
+    eh->timing_stack = ptls->timing_stack;
 #endif
 }
 

src/task.c

@@ -519,9 +519,10 @@ JL_DLLEXPORT void jl_switch(void)
     ptls->finalizers_inhibited = 0;
 
 #ifdef ENABLE_TIMINGS
-    jl_timing_block_t *blk = ct->timing_stack;
+    jl_timing_block_t *blk = ptls->timing_stack;
     if (blk)
         jl_timing_block_stop(blk);
+    ptls->timing_stack = NULL;
 #endif
 
     ctx_switch(ptls);
@@ -546,7 +547,8 @@ JL_DLLEXPORT void jl_switch(void)
     ptls->finalizers_inhibited = finalizers_inhibited;
 
 #ifdef ENABLE_TIMINGS
-    assert(blk == ct->timing_stack);
+    assert(ptls->timing_stack == NULL);
+    ptls->timing_stack = blk;
     if (blk)
         jl_timing_block_start(blk);
 #else
@@ -600,7 +602,7 @@ static void JL_NORETURN throw_internal(jl_value_t *exception JL_MAYBE_UNROOTED)
     jl_handler_t *eh = ptls->current_task->eh;
     if (eh != NULL) {
 #ifdef ENABLE_TIMINGS
-        jl_timing_block_t *cur_block = ptls->current_task->timing_stack;
+        jl_timing_block_t *cur_block = ptls->timing_stack;
         while (cur_block && eh->timing_stack != cur_block) {
             cur_block = jl_pop_timing_block(cur_block);
         }
@@ -701,9 +703,6 @@ JL_DLLEXPORT jl_task_t *jl_new_task(jl_function_t *start, jl_value_t *completion
     t->started = 0;
     t->prio = -1;
     t->tid = -1;
-#ifdef ENABLE_TIMINGS
-    t->timing_stack = jl_root_timing;
-#endif
 
 #if defined(JL_DEBUG_BUILD)
     if (!t->copy_stack)

src/timing.c

@@ -17,7 +17,7 @@ extern "C" {
 #error Timings are not supported on your compiler
 #endif
 
-jl_timing_block_t *jl_root_timing;
+static uint64_t t0;
 uint64_t jl_timing_data[(int)JL_TIMING_LAST] = {0};
 const char *jl_timing_names[(int)JL_TIMING_LAST] =
     {
@@ -28,32 +28,32 @@ const char *jl_timing_names[(int)JL_TIMING_LAST] =
 
 void jl_print_timings(void)
 {
-    uint64_t total_time = 0;
+    uint64_t total_time = cycleclock() - t0;
+    uint64_t root_time = total_time;
     for (int i = 0; i < JL_TIMING_LAST; i++) {
-        total_time += jl_timing_data[i];
+        root_time -= jl_timing_data[i];
     }
+    jl_timing_data[0] = root_time;
     for (int i = 0; i < JL_TIMING_LAST; i++) {
         if (jl_timing_data[i] != 0)
-            fprintf(stderr,"%-25s : %5.2f %%   %" PRIu64 "\n", jl_timing_names[i],
+            fprintf(stderr, "%-25s : %5.2f %%   %" PRIu64 "\n", jl_timing_names[i],
                     100 * (((double)jl_timing_data[i]) / total_time), jl_timing_data[i]);
     }
 }
 
 void jl_init_timing(void)
 {
-    jl_root_timing = (jl_timing_block_t*)malloc_s(sizeof(jl_timing_block_t));
-    _jl_timing_block_init(jl_root_timing, JL_TIMING_ROOT);
-    jl_root_timing->prev = NULL;
+    t0 = cycleclock();
 }
 
 void jl_destroy_timing(void)
 {
-    jl_timing_block_t *stack = jl_current_task ? jl_current_task->timing_stack : jl_root_timing;
+    jl_ptls_t ptls = jl_get_ptls_states();
+    jl_timing_block_t *stack = ptls->timing_stack;
     while (stack) {
         _jl_timing_block_destroy(stack);
         stack = stack->prev;
     }
-    free(jl_root_timing);
 }
 
 jl_timing_block_t *jl_pop_timing_block(jl_timing_block_t *cur_block)

src/timing.h

@@ -23,7 +23,6 @@ extern "C" {
 #endif
 void jl_print_timings(void);
 jl_timing_block_t *jl_pop_timing_block(jl_timing_block_t *cur_block);
-extern jl_timing_block_t *jl_root_timing;
 void jl_timing_block_start(jl_timing_block_t *cur_block);
 void jl_timing_block_stop(jl_timing_block_t *cur_block);
 #ifdef __cplusplus
@@ -117,7 +116,8 @@ STATIC_INLINE uint64_t _jl_timing_block_init(jl_timing_block_t *block, int owner
 
 STATIC_INLINE void _jl_timing_block_ctor(jl_timing_block_t *block, int owner) {
     uint64_t t = _jl_timing_block_init(block, owner);
-    jl_timing_block_t **prevp = jl_current_task ? &jl_current_task->timing_stack : &jl_root_timing;
+    jl_ptls_t ptls = jl_get_ptls_states();
+    jl_timing_block_t **prevp = &ptls->timing_stack;
     block->prev = *prevp;
     if (block->prev)
         _jl_timing_block_stop(block->prev, t);
@@ -126,9 +126,10 @@ STATIC_INLINE void _jl_timing_block_ctor(jl_timing_block_t *block, int owner) {
 
 STATIC_INLINE void _jl_timing_block_destroy(jl_timing_block_t *block) {
     uint64_t t = cycleclock();
+    jl_ptls_t ptls = jl_get_ptls_states();
     _jl_timing_block_stop(block, t);
     jl_timing_data[block->owner] += block->total;
-    jl_timing_block_t **pcur = jl_current_task ? &jl_current_task->timing_stack : &jl_root_timing;
+    jl_timing_block_t **pcur = &ptls->timing_stack;
     assert(*pcur == block);
     *pcur = block->prev;
     if (block->prev)