Re: Declaring Lisp function types

[Andrea Corallo <acorallo@gnu.org>]

Stefan Monnier via "Emacs development discussions."
<emacs-devel@gnu.org> writes:

>> (declaim (ftype (function (integer integer) integer) sum))
>> ;;                         ^^inputs         ^^output [optional]
>> (defun sum (a b)
>>   (declare (integer a b))
>>   (+ a b))
>
> Non-starter for me: the separation into two steps makes it unclear what
> the declaration applies to (e.g. when re-running the above code, does
> the `declaim` apply to the old definition (the one active when the
> `declaim` is executed)® the the one that's about to be installed)?
>
>> ;;2
>> (defun sum (a b)
>>   (declare (integer a b))
>>   (+ a b))
>
> None starter because of how we defined `declare`, where we'd have to
> define every existing type as a valid declaration idenitifer.
>
>> ;;3 through 'defstar' (a CL library not in the standard)
>> (defun* sum ((a integer) (b integer))
>>   (+ a b))
>> ;;4 again through 'defstar'
>> (defun* (sum -> integer) ((a integer) (b integer))
>>   (+ a b))
>
> Acceptable, with some tweaks to better fit my favorite bikeshed color.
>
>> (defun sum (a b)
>>   (declare (ftype (function (integer integer) integer)))
>>   (+ a b))
>
> The `f` of `ftype` is redundant with the following `function`, so we
> could shorten that to:
>
>     (defun sum (a b)
>       (declare (ftype (integer integer) integer))
>       (+ a b))
>
>> (defun sum (a b)
>>   (declare (function (integer integer) integer))
>>   (+ a b))
>
> It's cute, I guess.  Whether to prefer `function`, `ftype`, or Adam's `type`,
> is largely a "bikeshed color" choice.  I do prefer the latter two
> because we already know that this is a function, whereas we don't know
> that this is a *type* (and they're shorter, to boot).
>
> Later you said:
>> Fact is, we already use the form (function (ATYPES) RTYPE) as type
>> specifier for functions.  So (ftype (function (ATYPES) RTYPE)) would be
>> the most correct form semantically, where `ftype` (or `type` or really
>> what we prefer) would be the declaration which takes the type specifier
>> as argument.
>
> Of course (declare (ftype (integer integer) integer))
> would still end up generating something like
>
>     (foo-declare-type 'sum '(function (integer integer) integer))

My fear it's this is a bit more convoluted and this extra step makes it
less understandable/justifiable.  I like the symmetry of having
'function' both in the input (the declaration) and the output (the final
type itself).  Maybe my background as physicist makes symmetry too
central for me? :)

> so I see no semantic issue with using `ftype` or `type` here, unless
> there are functions whose type could take another form than (function
> <args> <rettype>)?  Are you thinking of types like
> (or (function (int) int) (function (float) float))?

That's a good example why it would be good to be able to accept the type
specifier as a declaration with no tricks. On the specific case I'm not
sure we want to support this in the inner machinery (at least for now).

> More important I think is to document what such annotations mean and
> what they should look like (currently, this is not super important,
> because the annotations live together with the code that uses them, but
> if we move them outside of `comp.el`, the "contract" needs to be made
> more explicit).
>
> - How they interact with `&optional` and `&rest` (or even `&key` for
>   `c-defun`).

ATM we already support in type specifiers `&optional` and `&rest`:

(subr-type (native-compile '(lambda (x &optional y &rest z)))) =>
(function (t &optional t &rest t) null)

Not sure we want to handle &key as well as it looks to me not very
native to the elisp machinery.  OTOH cl-defun just expands to the native
elisp call convention.

> - What will/could happen if one of the arguments does not have the
>   specified type?

I think if ones does a declaration has to declare the type of all
arguments (rest should be optional).

> - What will/could happen if the result does not have the
>   specified type?

I think we want to complete it with the inferred return type if we have
it or t otherwise.

> - Do we have types to say "arg unused" or "no return value"?

We don't have "arg unused" because the function type (or signature) is
like the contract with the outside word, it should not matter how (and
if) and arg is used inside.

OTOH we have "no return value" and it's nil

(subr-type (native-compile '(lambda (x) (error x)))) =>
(function (t) nil)

> - Can we have higher-order function types, like
>
>      (function (proc (function (proc string) void)) void)
>
>   and if so, again, what does it mean in terms of what can happen if the
>   runtime values don't actually match the announced types (e.g. what
>   happens (and when) if we pass a function that has the "wrong type")?

I don't kwnow if we want to allow this to be future proof, ATM certanly
the compiler does not use it and I don't think it could help code
generation.  OTOH might be nice for documentation?

As a note: AFAIR SBCL doesn't go beyond something like:
(function (integer function) function)

That if arguments/ret values are functions it forgets the inner details
of their type specifier.

Anyway I certanly agree we should better document this once it's shaped,
and I'll try my best.

But generally speaking I've the feeling there might be other cases we
don't see ATM where accepting directly the type specifier as valid
declaration graciously/naturally solves potential issues we could hit
otherwise.

Thanks

  Andrea