Count invalidations

Author: pchintalapudi

src/gf.c

@@ -1642,6 +1642,7 @@ static void do_nothing_with_codeinst(jl_code_instance_t *ci) {}
 // recursively invalidate cached methods that had an edge to a replaced method
 static void invalidate_method_instance(void (*f)(jl_code_instance_t*), jl_method_instance_t *replaced, size_t max_world, int depth)
 {
+    jl_timing_counter_inc(JL_TIMING_COUNTER_Invalidations, 1);
     if (_jl_debug_method_invalidation) {
         jl_value_t *boxeddepth = NULL;
         JL_GC_PUSH1(&boxeddepth);

src/timing.c

@@ -48,6 +48,8 @@ const char *jl_timing_names[(int)JL_TIMING_LAST] =
 #undef X
     };
 
+JL_DLLEXPORT jl_timing_counter_t jl_timing_counters[JL_TIMING_COUNTER_LAST];
+
 #ifdef USE_ITTAPI
 JL_DLLEXPORT __itt_event jl_timing_ittapi_events[(int)JL_TIMING_EVENT_LAST];
 #endif
@@ -60,11 +62,22 @@ void jl_print_timings(void)
         root_time -= jl_timing_counts[i];
     }
     jl_timing_counts[0] = root_time;
+    fprintf(stderr, "\nJULIA TIMINGS\n");
     for (int i = 0; i < JL_TIMING_LAST; i++) {
         if (jl_timing_counts[i] != 0)
             fprintf(stderr, "%-25s : %5.2f %%   %" PRIu64 "\n", jl_timing_names[i],
                     100 * (((double)jl_timing_counts[i]) / total_time), jl_timing_counts[i]);
     }
+
+    fprintf(stderr, "\nJULIA COUNTERS\n");
+    uint64_t val = 0;
+#ifdef USE_TIMING_COUNTS
+#define X(name) val = jl_atomic_load_relaxed(&jl_timing_counters[(int)JL_TIMING_COUNTER_##name].basic_counter); \
+    if (val != 0) \
+        fprintf(stderr, "%-25s : %" PRIu64 "\n", #name, val);
+    JL_TIMING_COUNTERS
+#undef X
+#endif
 }
 
 void jl_init_timing(void)
@@ -78,6 +91,10 @@ void jl_init_timing(void)
 #ifdef USE_ITTAPI
 #define X(name) jl_timing_ittapi_events[i++] = __itt_event_create(#name, strlen(#name));
     JL_TIMING_EVENTS
+#undef X
+    i = 0;
+#define X(name) jl_timing_counters[i++].ittapi_counter = __itt_counter_create(#name, "julia.runtime");
+    JL_TIMING_COUNTERS
 #undef X
 #endif
 }

src/timing.h

@@ -78,6 +78,9 @@ extern uint32_t jl_timing_print_limit;
 #define jl_timing_block_exit_task(ct, ptls) ((jl_timing_block_t *)NULL)
 #define jl_pop_timing_block(blk)
 
+#define jl_timing_counter_inc(counter, value)
+#define jl_timing_counter_dec(counter, value)
+
 #define jl_profile_lock_init(lock, name)
 #define jl_profile_lock_start_wait(lock)
 #define jl_profile_lock_acquired(lock)
@@ -181,6 +184,10 @@ void jl_timing_puts(jl_timing_block_t *cur_block, const char *str);
         X(NATIVE_Create) \
 
 
+#define JL_TIMING_COUNTERS \
+        X(Invalidations) \
+
+
 enum jl_timing_owners {
 #define X(name) JL_TIMING_ ## name,
     JL_TIMING_OWNERS
@@ -195,6 +202,13 @@ enum jl_timing_events {
     JL_TIMING_EVENT_LAST
 };
 
+enum jl_timing_counter_types {
+#define X(name) JL_TIMING_COUNTER_ ## name,
+    JL_TIMING_COUNTERS
+#undef X
+    JL_TIMING_COUNTER_LAST
+};
+
 /**
  * Timing back-ends differ in terms of whether they support nested
  * and asynchronous events.
@@ -404,6 +418,44 @@ struct jl_timing_suspend_cpp_t {
     _jl_timing_suspend_ctor(&__timing_suspend, #subsystem, ct)
 #endif
 
+// Counting
+#ifdef USE_ITTAPI
+#define _ITTAPI_COUNTER_MEMBER __itt_counter ittapi_counter;
+#else
+#define _ITTAPI_COUNTER_MEMBER
+#endif
+
+#ifdef USE_TIMING_COUNTS
+#define _COUNTS_MEMBER _Atomic(uint64_t) basic_counter;
+#else
+#define _COUNTS_MEMBER
+#endif
+
+typedef struct {
+    _ITTAPI_COUNTER_MEMBER
+    _COUNTS_MEMBER
+} jl_timing_counter_t;
+
+JL_DLLEXPORT extern jl_timing_counter_t jl_timing_counters[JL_TIMING_COUNTER_LAST];
+
+static inline void jl_timing_counter_inc(int counter, uint64_t val) JL_NOTSAFEPOINT {
+#ifdef USE_ITTAPI
+    __itt_counter_inc_delta(jl_timing_counters[counter].ittapi_counter, val);
+#endif
+#ifdef USE_TIMING_COUNTS
+    jl_atomic_fetch_add_relaxed(&jl_timing_counters[counter].basic_counter, val);
+#endif
+}
+
+static inline void jl_timing_counter_dec(int counter, uint64_t val) JL_NOTSAFEPOINT {
+#ifdef USE_ITTAPI
+    __itt_counter_dec_delta(jl_timing_counters[counter].ittapi_counter, val);
+#endif
+#ifdef USE_TIMING_COUNTS
+    jl_atomic_fetch_add_relaxed(&jl_timing_counters[counter].basic_counter, -(int64_t)val);
+#endif
+}
+
 // Locking profiling
 static inline void jl_profile_lock_init(jl_mutex_t *lock, const char *name) {
 #ifdef USE_ITTAPI