[RAI-46860] Add Julia fptr compilation time metric

base/timing.jl

@@ -75,6 +75,10 @@ function cumulative_compile_time_ns()
     return comp, recomp
 end
 
+# Time spent JIT-compiling Julia method instances, a subset of
+# `cumulative_compile_time_ns`
+fptr_compile_time_ns() = ccall(:jl_fptr_compile_time_ns, UInt64, ())
+
 function cumulative_compile_timing(b::Bool)
     if b
         ccall(:jl_cumulative_compile_timing_enable, Cvoid, ())

src/jitlayers.cpp

@@ -534,6 +534,7 @@ jl_code_instance_t *jl_generate_fptr_impl(jl_method_instance_t *mi JL_PROPAGATES
                 jl_atomic_fetch_add_relaxed(&jl_cumulative_recompile_time, t_comp);
             }
             jl_atomic_fetch_add_relaxed(&jl_cumulative_compile_time, t_comp);
+            jl_atomic_fetch_add_relaxed(&jl_fptr_compile_time, t_comp);
         }
         ct->reentrant_timing &= ~1ull;
     }

src/jlapi.c

@@ -509,6 +509,11 @@ JL_DLLEXPORT uint64_t jl_cumulative_recompile_time_ns(void)
     return jl_atomic_load_relaxed(&jl_cumulative_recompile_time);
 }
 
+JL_DLLEXPORT uint64_t jl_fptr_compile_time_ns(void)
+{
+    return jl_atomic_load_relaxed(&jl_fptr_compile_time);
+}
+
 /**
  * @brief Enable per-task timing.
  */

src/julia_internal.h

@@ -307,6 +307,7 @@ static inline uint64_t cycleclock(void) JL_NOTSAFEPOINT
 extern JL_DLLEXPORT _Atomic(uint8_t) jl_measure_compile_time_enabled;
 extern JL_DLLEXPORT _Atomic(uint64_t) jl_cumulative_compile_time;
 extern JL_DLLEXPORT _Atomic(uint64_t) jl_cumulative_recompile_time;
+extern JL_DLLEXPORT _Atomic(uint64_t) jl_fptr_compile_time;
 
 // Global *atomic* integer controlling *process-wide* task timing.
 extern JL_DLLEXPORT _Atomic(uint8_t) jl_task_metrics_enabled;

src/threading.c

@@ -51,6 +51,8 @@ JL_DLLEXPORT _Atomic(uint64_t) jl_cumulative_recompile_time = 0;
 
 JL_DLLEXPORT _Atomic(uint8_t) jl_task_metrics_enabled = 0;
 
+JL_DLLEXPORT _Atomic(uint64_t) jl_fptr_compile_time = 0;
+
 JL_DLLEXPORT void *jl_get_ptls_states(void)
 {
     // mostly deprecated: use current_task instead