fix @invokelatest performance regression (#58582)

A bit of hacking to get back near to the same performance as before by
using the GlobalRef to optimize the getglobal lookup for now and
avoiding the extra Vararg function indirection which forced some extra
boxing and lookups.

    julia> @Btime foo(1.5)
      22.892 ns (1 allocation: 16 bytes)  # v1.11
     141.543 ns (3 allocations: 48 bytes) # master
      38.759 ns (2 allocations: 32 bytes) # PR
  
The remaining difference is split about equally between the need now to
box the world counter value for invoke_in_world and the extra cost of
scanning the partition table for `Base.sin` to find the current entry.

Fix #58334

Author: vtjnash

base/reflection.jl

@@ -1269,16 +1269,17 @@ macro invoke(ex)
     return esc(out)
 end
 
-apply_gr(gr::GlobalRef, @nospecialize args...) = getglobal(gr.mod, gr.name)(args...)
-apply_gr_kw(@nospecialize(kwargs::NamedTuple), gr::GlobalRef, @nospecialize args...) = Core.kwcall(kwargs, getglobal(gr.mod, gr.name), args...)
+getglobalref(gr::GlobalRef, world::UInt) = ccall(:jl_eval_globalref, Any, (Any, UInt), gr, world)
 
-function invokelatest_gr(gr::GlobalRef, @nospecialize args...; kwargs...)
+function invokelatest_gr(gr::GlobalRef, args...; kwargs...)
     @inline
     kwargs = merge(NamedTuple(), kwargs)
+    world = get_world_counter()
+    f = getglobalref(gr, world)
     if isempty(kwargs)
-        return invokelatest(apply_gr, gr, args...)
+        return invoke_in_world(world, f, args...)
     end
-    return invokelatest(apply_gr_kw, kwargs, gr, args...)
+    return invoke_in_world(world, Core.kwcall, kwargs, f, args...)
 end
 
 """

src/ast.c

@@ -1328,7 +1328,7 @@ JL_DLLEXPORT jl_value_t *jl_lower(jl_value_t *expr, jl_module_t *inmodule,
 {
     jl_value_t *core_lower = NULL;
     if (jl_core_module)
-        core_lower = jl_get_global_value(jl_core_module, jl_symbol("_lower"));
+        core_lower = jl_get_global_value(jl_core_module, jl_symbol("_lower"), jl_current_task->world_age);
     if (!core_lower || core_lower == jl_nothing) {
         return jl_fl_lower(expr, inmodule, filename, line, world, warn);
     }

src/builtins.c

@@ -1340,7 +1340,7 @@ JL_CALLABLE(jl_f_getglobal)
         jl_atomic_error("getglobal: module binding cannot be read non-atomically");
     else if (order >= jl_memory_order_seq_cst)
         jl_fence();
-    jl_value_t *v = jl_eval_global_var(mod, sym); // relaxed load
+    jl_value_t *v = jl_eval_global_var(mod, sym, jl_current_task->world_age); // relaxed load
     if (order >= jl_memory_order_acquire)
         jl_fence();
     return v;

src/gf.c

@@ -512,13 +512,13 @@ JL_DLLEXPORT jl_code_info_t *jl_gdbcodetyped1(jl_method_instance_t *mi, size_t w
     ct->world_age = jl_typeinf_world;
     jl_value_t **fargs;
     JL_GC_PUSHARGS(fargs, 4);
-    jl_module_t *CC = (jl_module_t*)jl_get_global_value(jl_core_module, jl_symbol("Compiler"));
+    jl_module_t *CC = (jl_module_t*)jl_get_global_value(jl_core_module, jl_symbol("Compiler"), ct->world_age);
     if (CC != NULL && jl_is_module(CC)) {
         JL_GC_PROMISE_ROOTED(CC);
-        fargs[0] = jl_get_global_value(CC, jl_symbol("NativeInterpreter"));;
+        fargs[0] = jl_get_global_value(CC, jl_symbol("NativeInterpreter"), ct->world_age);
         fargs[1] = jl_box_ulong(world);
         fargs[1] = jl_apply(fargs, 2);
-        fargs[0] = jl_get_global_value(CC, jl_symbol("typeinf_code"));
+        fargs[0] = jl_get_global_value(CC, jl_symbol("typeinf_code"), ct->world_age);
         fargs[2] = (jl_value_t*)mi;
         fargs[3] = jl_true;
         ci = (jl_code_info_t*)jl_apply(fargs, 4);

src/init.c

@@ -249,7 +249,7 @@ JL_DLLEXPORT void jl_atexit_hook(int exitcode) JL_NOTSAFEPOINT_ENTER
     if (jl_base_module) {
         size_t last_age = ct->world_age;
         ct->world_age = jl_get_world_counter();
-        jl_value_t *f = jl_get_global_value(jl_base_module, jl_symbol("_atexit"));
+        jl_value_t *f = jl_get_global_value(jl_base_module, jl_symbol("_atexit"), ct->world_age);
         if (f != NULL) {
             jl_value_t **fargs;
             JL_GC_PUSHARGS(fargs, 2);
@@ -357,7 +357,7 @@ JL_DLLEXPORT void jl_postoutput_hook(void)
         jl_task_t *ct = jl_get_current_task();
         size_t last_age = ct->world_age;
         ct->world_age = jl_get_world_counter();
-        jl_value_t *f = jl_get_global_value(jl_base_module, jl_symbol("_postoutput"));
+        jl_value_t *f = jl_get_global_value(jl_base_module, jl_symbol("_postoutput"), ct->world_age);
         if (f != NULL) {
             JL_TRY {
                 JL_GC_PUSH1(&f);

src/interpreter.c

@@ -162,18 +162,18 @@ static jl_value_t *do_invoke(jl_value_t **args, size_t nargs, interpreter_state
 }
 
 // get the global (throwing if null) in the current world
-jl_value_t *jl_eval_global_var(jl_module_t *m, jl_sym_t *e)
+jl_value_t *jl_eval_global_var(jl_module_t *m, jl_sym_t *e, size_t world)
 {
-    jl_value_t *v = jl_get_global_value(m, e);
+    jl_value_t *v = jl_get_global_value(m, e, world);
     if (v == NULL)
         jl_undefined_var_error(e, (jl_value_t*)m);
     return v;
 }
 
 // get the global (throwing if null) in the current world, optimized
-jl_value_t *jl_eval_globalref(jl_globalref_t *g)
+jl_value_t *jl_eval_globalref(jl_globalref_t *g, size_t world)
 {
-    jl_value_t *v = jl_get_globalref_value(g);
+    jl_value_t *v = jl_get_globalref_value(g, world);
     if (v == NULL)
         jl_undefined_var_error(g->name, (jl_value_t*)g->mod);
     return v;
@@ -218,10 +218,10 @@ static jl_value_t *eval_value(jl_value_t *e, interpreter_state *s)
         return jl_quotenode_value(e);
     }
     if (jl_is_globalref(e)) {
-        return jl_eval_globalref((jl_globalref_t*)e);
+        return jl_eval_globalref((jl_globalref_t*)e, jl_current_task->world_age);
     }
     if (jl_is_symbol(e)) {  // bare symbols appear in toplevel exprs not wrapped in `thunk`
-        return jl_eval_global_var(s->module, (jl_sym_t*)e);
+        return jl_eval_global_var(s->module, (jl_sym_t*)e, jl_current_task->world_age);
     }
     if (jl_is_pinode(e)) {
         jl_value_t *val = eval_value(jl_fieldref_noalloc(e, 0), s);

src/jl_uv.c

@@ -162,7 +162,8 @@ static void jl_uv_call_close_callback(jl_value_t *val)
     JL_GC_PUSHARGS(args, 2); // val is "rooted" in the finalizer list only right now
     args[0] = jl_eval_global_var(
             jl_base_relative_to(((jl_datatype_t*)jl_typeof(val))->name->module),
-            jl_symbol("_uv_hook_close")); // topmod(typeof(val))._uv_hook_close
+            jl_symbol("_uv_hook_close"),
+            jl_current_task->world_age); // topmod(typeof(val))._uv_hook_close
     args[1] = val;
     jl_apply(args, 2); // TODO: wrap in try-catch?
     JL_GC_POP();

src/jlapi.c

@@ -955,7 +955,7 @@ static NOINLINE int true_main(int argc, char *argv[])
     ct->world_age = jl_get_world_counter();
 
     jl_function_t *start_client = jl_base_module ?
-        (jl_function_t*)jl_get_global_value(jl_base_module, jl_symbol("_start")) : NULL;
+        (jl_function_t*)jl_get_global_value(jl_base_module, jl_symbol("_start"), ct->world_age) : NULL;
 
     if (start_client) {
         int ret = 1;

src/jltypes.c

@@ -3287,6 +3287,7 @@ void jl_init_types(void) JL_GC_DISABLED
                         jl_perm_symsvec(3, "mod", "name", "binding"),
                         jl_svec(3, jl_module_type, jl_symbol_type, jl_binding_type),
                         jl_emptysvec, 0, 0, 3);
+    jl_globalref_type->name->mayinlinealloc = 0; // not at all worthwhile, since the only constructor returns a boxed object
 
     jl_core_module = jl_new_module(jl_symbol("Core"), NULL);
 

src/julia.h

@@ -2094,9 +2094,7 @@ JL_DLLEXPORT jl_value_t *jl_get_existing_strong_gf(jl_binding_t *b JL_PROPAGATES
 JL_DLLEXPORT int jl_boundp(jl_module_t *m, jl_sym_t *var, int allow_import);
 JL_DLLEXPORT int jl_is_const(jl_module_t *m, jl_sym_t *var);
 JL_DLLEXPORT int jl_globalref_is_const(jl_globalref_t *gr);
-JL_DLLEXPORT jl_value_t *jl_get_globalref_value(jl_globalref_t *gr);
 JL_DLLEXPORT jl_value_t *jl_get_global(jl_module_t *m JL_PROPAGATES_ROOT, jl_sym_t *var);
-JL_DLLEXPORT jl_value_t *jl_get_global_value(jl_module_t *m, jl_sym_t *var);
 JL_DLLEXPORT void jl_set_global(jl_module_t *m JL_ROOTING_ARGUMENT, jl_sym_t *var, jl_value_t *val JL_ROOTED_ARGUMENT);
 JL_DLLEXPORT void jl_set_const(jl_module_t *m JL_ROOTING_ARGUMENT, jl_sym_t *var, jl_value_t *val JL_ROOTED_ARGUMENT);
 void jl_set_initial_const(jl_module_t *m JL_ROOTING_ARGUMENT, jl_sym_t *var, jl_value_t *val JL_ROOTED_ARGUMENT, int exported);