Make cumulative_compile_time_ns into a global, atomic variable.

Instead of keeping per-task compilation time, this change keeps a
global counter of compilation time, protected with atomic mutations.

Fixes #41739

```julia
julia> include("/Users/nathandaly/Downloads/compilation-task-migration-17-example.jl")
start thread: 2
end thread: 2
  5.185706 seconds (3.53 M allocations: 2.570 GiB, 7.34% gc time, 15.57% compilation time)

julia> include("/Users/nathandaly/Downloads/compilation-task-migration-17-example.jl")
start thread: 3
WARNING: replacing module M.
end thread: 1
  4.110316 seconds (18.23 k allocations: 2.391 GiB, 5.67% gc time, 0.24% compilation time)
```

base/timing.jl

@@ -60,13 +60,6 @@ cumulative_compile_time_ns_before() = ccall(:jl_cumulative_compile_time_ns_befor
 cumulative_compile_time_ns_after() = ccall(:jl_cumulative_compile_time_ns_after, UInt64, ())
 # cumulative total time this thread has spent on compilation since process start.
 cumulative_compile_time_ns() = ccall(:jl_cumulative_compile_time_ns, UInt64, ())
-function process_cumulative_compile_time_ns()
-    out = fill(UInt(0), Threads.nthreads())
-    Threads.@threads for i in 1:Threads.nthreads()
-        out[i] = cumulative_compile_time_ns()
-    end
-    return sum(out)
-end
 
 # total time spend in garbage collection, in nanoseconds
 gc_time_ns() = ccall(:jl_gc_total_hrtime, UInt64, ())

src/aotcompile.cpp

@@ -287,7 +287,7 @@ void *jl_create_native(jl_array_t *methods, const jl_cgparams_t cgparams, int _p
     JL_LOCK(&codegen_lock);
     uint64_t compiler_start_time = 0;
     int tid = jl_threadid();
-    if (jl_measure_compile_time[tid])
+    if (jl_atomic_load(&jl_measure_compile_time))
         compiler_start_time = jl_hrtime();
 
     CompilationPolicy policy = (CompilationPolicy) _policy;
@@ -415,8 +415,8 @@ void *jl_create_native(jl_array_t *methods, const jl_cgparams_t cgparams, int _p
     }
 
     data->M = std::move(clone);
-    if (jl_measure_compile_time[tid])
-        jl_cumulative_compile_time[tid] += (jl_hrtime() - compiler_start_time);
+    if (jl_atomic_load(&jl_measure_compile_time))
+        jl_atomic_fetch_add(&jl_measure_compile_time, (jl_hrtime() - compiler_start_time));
     if (policy == CompilationPolicy::ImagingMode)
         imaging_mode = 0;
     JL_UNLOCK(&codegen_lock); // Might GC
@@ -917,7 +917,7 @@ void *jl_get_llvmf_defn(jl_method_instance_t *mi, size_t world, char getwrapper,
         JL_LOCK(&codegen_lock);
         uint64_t compiler_start_time = 0;
         int tid = jl_threadid();
-        if (jl_measure_compile_time[tid])
+        if (jl_atomic_load(&jl_measure_compile_time))
             compiler_start_time = jl_hrtime();
         std::tie(m, decls) = jl_emit_code(mi, src, jlrettype, output);
 
@@ -942,8 +942,8 @@ void *jl_get_llvmf_defn(jl_method_instance_t *mi, size_t world, char getwrapper,
             m.release(); // the return object `llvmf` will be the owning pointer
         }
         JL_GC_POP();
-        if (jl_measure_compile_time[tid])
-            jl_cumulative_compile_time[tid] += (jl_hrtime() - compiler_start_time);
+        if (jl_atomic_load(&jl_measure_compile_time))
+            jl_atomic_fetch_add(&jl_measure_compile_time, (jl_hrtime() - compiler_start_time));
         JL_UNLOCK(&codegen_lock); // Might GC
         if (F)
             return F;

src/gf.c

@@ -3164,15 +3164,14 @@ static uint64_t inference_start_time = 0;
 JL_DLLEXPORT void jl_typeinf_begin(void)
 {
     JL_LOCK(&typeinf_lock);
-    if (jl_measure_compile_time[jl_threadid()])
+    if (jl_atomic_load(&jl_measure_compile_time))
         inference_start_time = jl_hrtime();
 }
 
 JL_DLLEXPORT void jl_typeinf_end(void)
 {
-    int tid = jl_threadid();
-    if (typeinf_lock.count == 1 && jl_measure_compile_time[tid])
-        jl_cumulative_compile_time[tid] += (jl_hrtime() - inference_start_time);
+    if (typeinf_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
+        jl_atomic_fetch_add(&jl_measure_compile_time, (jl_hrtime() - inference_start_time));
     JL_UNLOCK(&typeinf_lock);
 }
 

src/jitlayers.cpp

@@ -78,22 +78,19 @@ void jl_jit_globals(std::map<void *, GlobalVariable*> &globals)
 extern "C" JL_DLLEXPORT
 uint64_t jl_cumulative_compile_time_ns_before()
 {
-    int tid = jl_threadid();
-    jl_measure_compile_time[tid] += 1;
-    return jl_cumulative_compile_time[tid];
+    jl_atomic_fetch_add(&jl_measure_compile_time, 1);
+    return jl_atomic_load(&jl_cumulative_compile_time);
 }
 extern "C" JL_DLLEXPORT
 uint64_t jl_cumulative_compile_time_ns_after()
 {
-    int tid = jl_threadid();
-    jl_measure_compile_time[tid] -= 1;
-    return jl_cumulative_compile_time[tid];
+    jl_atomic_fetch_add(&jl_measure_compile_time, -1);
+    return jl_atomic_load(&jl_cumulative_compile_time);
 }
 
 extern "C" JL_DLLEXPORT
 uint64_t jl_cumulative_compile_time_ns() {
-    int tid = jl_threadid();
-    return jl_cumulative_compile_time[tid];
+    return jl_atomic_load(&jl_cumulative_compile_time);
 }
 
 // this generates llvm code for the lambda info
@@ -239,8 +236,7 @@ int jl_compile_extern_c(void *llvmmod, void *p, void *sysimg, jl_value_t *declrt
 {
     JL_LOCK(&codegen_lock);
     uint64_t compiler_start_time = 0;
-    int tid = jl_threadid();
-    if (jl_measure_compile_time[tid])
+    if (jl_atomic_load(&jl_measure_compile_time))
         compiler_start_time = jl_hrtime();
     jl_codegen_params_t params;
     jl_codegen_params_t *pparams = (jl_codegen_params_t*)p;
@@ -264,8 +260,8 @@ int jl_compile_extern_c(void *llvmmod, void *p, void *sysimg, jl_value_t *declrt
         if (success && llvmmod == NULL)
             jl_add_to_ee(std::unique_ptr<Module>(into));
     }
-    if (codegen_lock.count == 1 && jl_measure_compile_time[tid])
-        jl_cumulative_compile_time[tid] += (jl_hrtime() - compiler_start_time);
+    if (codegen_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
+        jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
     JL_UNLOCK(&codegen_lock);
     return success;
 }
@@ -321,8 +317,7 @@ jl_code_instance_t *jl_generate_fptr(jl_method_instance_t *mi JL_PROPAGATES_ROOT
 {
     JL_LOCK(&codegen_lock); // also disables finalizers, to prevent any unexpected recursion
     uint64_t compiler_start_time = 0;
-    int tid = jl_threadid();
-    if (jl_measure_compile_time[tid])
+    if (jl_atomic_load(&jl_measure_compile_time))
         compiler_start_time = jl_hrtime();
     // if we don't have any decls already, try to generate it now
     jl_code_info_t *src = NULL;
@@ -360,8 +355,8 @@ jl_code_instance_t *jl_generate_fptr(jl_method_instance_t *mi JL_PROPAGATES_ROOT
     else {
         codeinst = NULL;
     }
-    if (codegen_lock.count == 1 && jl_measure_compile_time[tid])
-        jl_cumulative_compile_time[tid] += (jl_hrtime() - compiler_start_time);
+    if (codegen_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
+        jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
     JL_UNLOCK(&codegen_lock);
     JL_GC_POP();
     return codeinst;
@@ -375,8 +370,7 @@ void jl_generate_fptr_for_unspecialized(jl_code_instance_t *unspec)
     }
     JL_LOCK(&codegen_lock);
     uint64_t compiler_start_time = 0;
-    int tid = jl_threadid();
-    if (jl_measure_compile_time[tid])
+    if (jl_atomic_load(&jl_measure_compile_time))
         compiler_start_time = jl_hrtime();
     if (unspec->invoke == NULL) {
         jl_code_info_t *src = NULL;
@@ -404,8 +398,8 @@ void jl_generate_fptr_for_unspecialized(jl_code_instance_t *unspec)
         }
         JL_GC_POP();
     }
-    if (codegen_lock.count == 1 && jl_measure_compile_time[tid])
-        jl_cumulative_compile_time[tid] += (jl_hrtime() - compiler_start_time);
+    if (codegen_lock.count == 1 && jl_atomic_load(&jl_measure_compile_time))
+        jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
     JL_UNLOCK(&codegen_lock); // Might GC
 }
 
@@ -428,8 +422,7 @@ jl_value_t *jl_dump_method_asm(jl_method_instance_t *mi, size_t world,
             // so create an exception here so we can print pretty our lies
             JL_LOCK(&codegen_lock); // also disables finalizers, to prevent any unexpected recursion
             uint64_t compiler_start_time = 0;
-            int tid = jl_threadid();
-            if (jl_measure_compile_time[tid])
+            if (jl_atomic_load(&jl_measure_compile_time))
                 compiler_start_time = jl_hrtime();
             specfptr = (uintptr_t)codeinst->specptr.fptr;
             if (specfptr == 0) {
@@ -454,8 +447,8 @@ jl_value_t *jl_dump_method_asm(jl_method_instance_t *mi, size_t world,
                 }
                 JL_GC_POP();
             }
-            if (jl_measure_compile_time[tid])
-                jl_cumulative_compile_time[tid] += (jl_hrtime() - compiler_start_time);
+            if (jl_atomic_load(&jl_measure_compile_time))
+                jl_atomic_fetch_add(&jl_cumulative_compile_time, (jl_hrtime() - compiler_start_time));
             JL_UNLOCK(&codegen_lock);
         }
         if (specfptr != 0)

src/julia_internal.h

@@ -154,8 +154,9 @@ static inline uint64_t cycleclock(void)
 
 #include "timing.h"
 
-extern uint8_t *jl_measure_compile_time;
-extern uint64_t *jl_cumulative_compile_time;
+// Global *atomic* integers controlling *process-wide* measurement of compilation time.
+extern volatile uint8_t jl_measure_compile_time;
+extern volatile uint64_t jl_cumulative_compile_time;
 
 #ifdef _COMPILER_MICROSOFT_
 #  define jl_return_address() ((uintptr_t)_ReturnAddress())

src/task.c

@@ -561,7 +561,7 @@ static void JL_NORETURN throw_internal(jl_task_t *ct, jl_value_t *exception JL_M
     ptls->io_wait = 0;
     // @time needs its compile timer disabled on error,
     // and cannot use a try-finally as it would break scope for assignments
-    jl_measure_compile_time[ptls->tid] = 0;
+    jl_atomic_fetch_add(&jl_measure_compile_time, -1);
     JL_GC_PUSH1(&exception);
     jl_gc_unsafe_enter(ptls);
     if (exception) {

src/threading.c

@@ -287,8 +287,8 @@ void jl_pgcstack_getkey(jl_get_pgcstack_func **f, jl_pgcstack_key_t *k)
 #endif
 
 jl_ptls_t *jl_all_tls_states JL_GLOBALLY_ROOTED;
-uint8_t *jl_measure_compile_time = NULL;
-uint64_t *jl_cumulative_compile_time = NULL;
+volatile uint8_t jl_measure_compile_time = 0;
+volatile uint64_t jl_cumulative_compile_time = 0;
 
 // return calling thread's ID
 // Also update the suspended_threads list in signals-mach when changing the
@@ -467,8 +467,6 @@ void jl_init_threading(void)
     }
     if (jl_n_threads <= 0)
         jl_n_threads = 1;
-    jl_measure_compile_time = (uint8_t*)calloc(jl_n_threads, sizeof(*jl_measure_compile_time));
-    jl_cumulative_compile_time = (uint64_t*)calloc(jl_n_threads, sizeof(*jl_cumulative_compile_time));
 #ifndef __clang_analyzer__
     jl_all_tls_states = (jl_ptls_t*)calloc(jl_n_threads, sizeof(void*));
 #endif