--- Begin Message ---
Subject: |
30.0.50; `type-of` is not precise enough |
Date: |
Mon, 11 Mar 2024 19:19:40 -0400 |
Package: Emacs
Version: 30.0.50
`type-of` is supposed to return "the" type of its argument. Given that
ELisp has a notion of subtyping, "the" type is expected to mean "the
most precise type". This is used in `cl-generic` to decide which method
to apply, so it's important that it returns precise information.
Currently, there `type-of` fails to return precise enough information in
a few cases:
- When the argument is nil, it returns `symbol`. The problem here is
that `symbol` is not a subtype of `list`, whereas nil is a list.
- When the argument is a special form, a C primitive, or
a native-compiled function it returns `subr`. Currently our
type hierarchy says that `subr` is a subtype of `compiled-function`
(and hence of `function`), but a special form is *not* a function
(it fails the `functionp` test and can't be `funcall`ed).
Currently `cl-generic` works around the first point above by using
(if FOO (type-of FOO) 'null) instead of calling `type-of` directly.
Suggestion:
I suggest we change `type-of` to return `null` for `nil`,
`special-form` for subrs that are special forms, `subr-primitive`
for C primitives, and `subr-native-elisp` for native-compiled subrs.
There are a few other cases where we could improve the precision, tho
they are less important because they don't cause problems w.r.t
subtyping like the above does.
Further improvements could include:
- Return `boolean` for `t`. This would be nice otherwise (with the
above suggestion) `cl-generic` can dispatch on "nil is a boolean"
but not on "t is a boolean".
- Return `keyword` for symbols that are keywords.
- Return `fixnum` or `bignum` rather than just `integer`.
Probably not worth the trouble.
- We could go crazy and return `keyword-with-pos` for `symbols-with-pos`
that are keywords.
Of these further improvements, only the first (return `boolean` for `t`)
seems worth the trouble.
Still, any change as suggested here would be an incompatible change, so
there's risk it'll break some code out there (`type-of` is not used very
often, but it *is* used). Another option is to introduce a new function
which does the same as `type-of` but with changes like the ones above.
(we could even decide to give it a `cl-generic-` prefix to discourage
its use elsewhere so we can be more free to change its return value in
the future).
Stefan
--- End Message ---
--- Begin Message ---
Subject: |
Re: bug#69739: 30.0.50; `type-of` is not precise enough |
Date: |
Mon, 18 Mar 2024 09:33:58 -0400 |
User-agent: |
Gnus/5.13 (Gnus v5.13) |
>> +@example
>> +(object-type 1)
>> + @result{} fixnum
>> +@group
>> +(object-type 'nil)
>> + @result{} null
>> +(object-type (record 'foo))
>> + @result{} foo
>
> "object-type"?
Oops! thanks.
>> DEFUN ("type-of", Ftype_of, Stype_of, 1, 1, 0,
>> doc: /* Return a symbol representing the type of OBJECT.
>> The symbol returned names the object's basic type;
>> -for example, (type-of 1) returns `integer'. */)
>> +for example, (type-of 1) returns `integer'.
>> +Contrary to `cl-type-of' the returned type is not always the most
> ^^
> I think we want a comma there.
>> +DEFUN ("cl-type-of", Fcl_type_of, Scl_type_of, 1, 1, 0,
>> + doc: /* Return a symbol representing the type of OBJECT.
>> +The symbol returned names the most specific possible type of the object.
> ^^^^^^^^^^^^^^^^^^^
> I think "The returned symbol" is better here, as it prevents a
> possible confusion (whether "returned" alludes to "symbol" or to
> "names").
Agreed.
>> +for example, (object-type nil) returns `null'.
> ^^^^^^^^^^^
> "object-type"?
As you can see I had used `object-type` instead of `cl-type-of` in some
prior version of the code :-)
>> (defsubst subr-primitive-p (object)
>> - "Return t if OBJECT is a built-in primitive function."
>> + "Return t if OBJECT is a built-in primitive written in C."
>> (declare (side-effect-free error-free))
>> (and (subrp object)
>> (not (subr-native-elisp-p object))))
>>
>> +(defsubst primitive-function-p (object)
>> + "Return t if OBJECT is a built-in primitive function."
>> + (declare (side-effect-free error-free))
>> + (and (subrp object)
>> + (not (or (subr-native-elisp-p object)
>> + (eq (cdr (subr-arity object)) 'unevalled)))))
>
> Should these doc strings mention the special case of special form,
> which each one of them treats differently?
OK.
Pushed, thanks,
Stefan
--- End Message ---