Make finalizer a built-in (#45423)

* Make finalizer a built-in

Split out from #45272. This is prepratory work towards adding
optimization passes that recognize this builtin. This PR adds
`Core.finalizer` with essentially the same interface as
`Base.finalizer`, but without the error checking or raw-C-pointer
feature. In future commits, the Core.finalizer interface will
likely expand slightly, but Base.finalizer will remain unchanged
and is the supported interface for this functionality.

* Update base/docs/basedocs.jl

Co-authored-by: Shuhei Kadowaki <40514306+aviatesk@users.noreply.github.com>

* Apply suggestions from code review

Co-authored-by: Ian Atol <ian.atol@juliacomputing.com>

Co-authored-by: Shuhei Kadowaki <40514306+aviatesk@users.noreply.github.com>
Co-authored-by: Ian Atol <ian.atol@juliacomputing.com>

Author: 3 people

base/compiler/tfuncs.jl

@@ -559,6 +559,7 @@ add_tfunc(atomic_pointerswap, 3, 3, (a, v, order) -> (@nospecialize; pointer_elt
 add_tfunc(atomic_pointermodify, 4, 4, atomic_pointermodify_tfunc, 5)
 add_tfunc(atomic_pointerreplace, 5, 5, atomic_pointerreplace_tfunc, 5)
 add_tfunc(donotdelete, 0, INT_INF, (@nospecialize args...)->Nothing, 0)
+add_tfunc(Core.finalizer, 2, 2, (@nospecialize args...)->Nothing, 5)
 
 # more accurate typeof_tfunc for vararg tuples abstract only in length
 function typeof_concrete_vararg(t::DataType)

base/docs/basedocs.jl

@@ -3068,4 +3068,25 @@ end
 """
 Base.donotdelete
 
+"""
+    Core.finalizer(f, o)
+
+This builtin is an implementation detail of [`Base.finalizer`](@ref) and end-users
+should use the latter instead.
+
+# Differences from `Base.finalizer`
+
+The interface of `Core.finalizer` is essentially the same as `Base.finalizer`,
+but there are a number of small differences. They are documented here for
+completeness only and (unlike `Base.finalizer`) have no stability guarantees.
+
+The current differences are:
+    - `Core.finalizer` does not check for mutability of `o`. Attempting to register
+      a finalizer for an immutable object is undefined behavior.
+    - The value `f` must be a Julia object. `Core.finalizer` does not support a
+      raw C function pointer.
+    - `Core.finalizer` returns `nothing` rather than `o`.
+"""
+Core.finalizer
+
 end

base/gcutils.jl

@@ -4,17 +4,27 @@
 ==(w::WeakRef, v) = isequal(w.value, v)
 ==(w, v::WeakRef) = isequal(w, v.value)
 
+# Used by `Base.finalizer` to validate mutability of an object being finalized.
+function _check_mutable(@nospecialize(o)) @noinline
+    if !ismutable(o)
+        error("objects of type ", typeof(o), " cannot be finalized")
+    end
+end
+
 """
     finalizer(f, x)
 
 Register a function `f(x)` to be called when there are no program-accessible references to
-`x`, and return `x`. The type of `x` must be a `mutable struct`, otherwise the behavior of
-this function is unpredictable.
+`x`, and return `x`. The type of `x` must be a `mutable struct`, otherwise the function
+will throw.
 
 `f` must not cause a task switch, which excludes most I/O operations such as `println`.
 Using the `@async` macro (to defer context switching to outside of the finalizer) or
 `ccall` to directly invoke IO functions in C may be helpful for debugging purposes.
 
+Note that there is no guaranteed world age for the execution of `f`. It may be
+called in the world age in which the finalizer was registered or any later world age.
+
 # Examples
 ```julia
 finalizer(my_mutable_struct) do x
@@ -42,18 +52,13 @@ end
 ```
 """
 function finalizer(@nospecialize(f), @nospecialize(o))
-    if !ismutable(o)
-        error("objects of type ", typeof(o), " cannot be finalized")
-    end
-    ccall(:jl_gc_add_finalizer_th, Cvoid, (Ptr{Cvoid}, Any, Any),
-          Core.getptls(), o, f)
+    _check_mutable(o)
+    Core.finalizer(f, o)
     return o
 end
 
 function finalizer(f::Ptr{Cvoid}, o::T) where T @inline
-    if !ismutable(o)
-        error("objects of type ", typeof(o), " cannot be finalized")
-    end
+    _check_mutable(o)
     ccall(:jl_gc_add_ptr_finalizer, Cvoid, (Ptr{Cvoid}, Any, Ptr{Cvoid}),
           Core.getptls(), o, f)
     return o

src/builtin_proto.h

@@ -57,6 +57,7 @@ DECLARE_BUILTIN(_typevar);
 DECLARE_BUILTIN(donotdelete);
 DECLARE_BUILTIN(getglobal);
 DECLARE_BUILTIN(setglobal);
+DECLARE_BUILTIN(finalizer);
 
 JL_CALLABLE(jl_f_invoke_kwsorter);
 #ifdef DEFINE_BUILTIN_GLOBALS
@@ -73,6 +74,7 @@ JL_CALLABLE(jl_f_get_binding_type);
 JL_CALLABLE(jl_f_set_binding_type);
 JL_CALLABLE(jl_f_donotdelete);
 JL_CALLABLE(jl_f_setglobal);
+JL_CALLABLE(jl_f_finalizer);
 
 #ifdef __cplusplus
 }

src/builtins.c

@@ -1600,6 +1600,14 @@ JL_CALLABLE(jl_f_donotdelete)
     return jl_nothing;
 }
 
+JL_CALLABLE(jl_f_finalizer)
+{
+    JL_NARGS(finalizer, 2, 2);
+    jl_task_t *ct = jl_current_task;
+    jl_gc_add_finalizer_(ct->ptls, args[1], args[0]);
+    return jl_nothing;
+}
+
 static int equiv_field_types(jl_value_t *old, jl_value_t *ft)
 {
     size_t nf = jl_svec_len(ft);
@@ -1970,6 +1978,7 @@ void jl_init_primitives(void) JL_GC_DISABLED
     jl_builtin__typebody = add_builtin_func("_typebody!", jl_f__typebody);
     add_builtin_func("_equiv_typedef", jl_f__equiv_typedef);
     jl_builtin_donotdelete = add_builtin_func("donotdelete", jl_f_donotdelete);
+    add_builtin_func("finalizer", jl_f_finalizer);
 
     // builtin types
     add_builtin("Any", (jl_value_t*)jl_any_type);

src/codegen.cpp

@@ -1126,7 +1126,8 @@ static const auto &builtin_func_map() {
           { jl_f_arrayset_addr,           new JuliaFunction{XSTR(jl_f_arrayset), get_func_sig, get_func_attrs} },
           { jl_f_arraysize_addr,          new JuliaFunction{XSTR(jl_f_arraysize), get_func_sig, get_func_attrs} },
           { jl_f_apply_type_addr,         new JuliaFunction{XSTR(jl_f_apply_type), get_func_sig, get_func_attrs} },
-          { jl_f_donotdelete_addr,        new JuliaFunction{XSTR(jl_f_donotdelete), get_donotdelete_sig, get_donotdelete_func_attrs} }
+          { jl_f_donotdelete_addr,        new JuliaFunction{XSTR(jl_f_donotdelete), get_donotdelete_sig, get_donotdelete_func_attrs} },
+          { jl_f_finalizer_addr,          new JuliaFunction{XSTR(jl_f_finalizer), get_func_sig, get_func_attrs} }
         };
     return builtins;
 }

src/gc.c

@@ -488,7 +488,7 @@ void jl_gc_run_all_finalizers(jl_task_t *ct)
     run_finalizers(ct);
 }
 
-static void gc_add_finalizer_(jl_ptls_t ptls, void *v, void *f) JL_NOTSAFEPOINT
+void jl_gc_add_finalizer_(jl_ptls_t ptls, void *v, void *f) JL_NOTSAFEPOINT
 {
     assert(jl_atomic_load_relaxed(&ptls->gc_state) == 0);
     arraylist_t *a = &ptls->finalizers;
@@ -518,7 +518,7 @@ static void gc_add_finalizer_(jl_ptls_t ptls, void *v, void *f) JL_NOTSAFEPOINT
 
 JL_DLLEXPORT void jl_gc_add_ptr_finalizer(jl_ptls_t ptls, jl_value_t *v, void *f) JL_NOTSAFEPOINT
 {
-    gc_add_finalizer_(ptls, (void*)(((uintptr_t)v) | 1), f);
+    jl_gc_add_finalizer_(ptls, (void*)(((uintptr_t)v) | 1), f);
 }
 
 JL_DLLEXPORT void jl_gc_add_finalizer_th(jl_ptls_t ptls, jl_value_t *v, jl_function_t *f) JL_NOTSAFEPOINT
@@ -527,7 +527,7 @@ JL_DLLEXPORT void jl_gc_add_finalizer_th(jl_ptls_t ptls, jl_value_t *v, jl_funct
         jl_gc_add_ptr_finalizer(ptls, v, jl_unbox_voidpointer(f));
     }
     else {
-        gc_add_finalizer_(ptls, v, f);
+        jl_gc_add_finalizer_(ptls, v, f);
     }
 }
 

src/julia_internal.h

@@ -466,6 +466,7 @@ void jl_gc_track_malloced_array(jl_ptls_t ptls, jl_array_t *a) JL_NOTSAFEPOINT;
 void jl_gc_count_allocd(size_t sz) JL_NOTSAFEPOINT;
 void jl_gc_run_all_finalizers(jl_task_t *ct);
 void jl_release_task_stack(jl_ptls_t ptls, jl_task_t *task);
+void jl_gc_add_finalizer_(jl_ptls_t ptls, void *v, void *f) JL_NOTSAFEPOINT;
 
 // Set GC memory trigger in bytes for greedy memory collecting
 void jl_gc_set_max_memory(uint64_t max_mem);

src/staticdata.c

@@ -314,7 +314,7 @@ static const jl_fptr_args_t id_to_fptrs[] = {
     &jl_f_ifelse, &jl_f__structtype, &jl_f__abstracttype, &jl_f__primitivetype,
     &jl_f__typebody, &jl_f__setsuper, &jl_f__equiv_typedef, &jl_f_get_binding_type,
     &jl_f_set_binding_type, &jl_f_opaque_closure_call, &jl_f_donotdelete,
-    &jl_f_getglobal, &jl_f_setglobal,
+    &jl_f_getglobal, &jl_f_setglobal, &jl_f_finalizer,
     NULL };
 
 typedef struct {