bug#64646: Master: Native compiler doesn't always compile lambda forms.

[Andrea Corallo]

Alan Mackenzie <acm@muc.de> writes:

> Hello, Andrea.
>
> This bug doesn't seem to be moving, so ....
>
> On Wed, Jul 26, 2023 at 10:57:01 -0400, Andrea Corallo wrote:
>> Alan Mackenzie <acm@muc.de> writes:
>
>> > Hello, Andrea.
>
>> > On Mon, Jul 17, 2023 at 09:17:13 -0400, Andrea Corallo wrote:
>> >> Alan Mackenzie <acm@muc.de> writes:
>
>> >> > In the master branch:
>
>> >> > (i) emacs -Q
>> >> > (ii) C-x b foo.el <RET>
>> >> > (iii) Insert into foo.el:
>> >> >     ;; -*- lexical-binding:t -*-
>> >> > (iv) M-x emacs-lisp-mode
>> >> > (v) Insert into foo.el:
>> >> >     (defun foo () "foo doc string"
>> >> >       (lambda (bar) "lambda doc string" (car bar)))
>> >> > (vi) With point after the function, C-x C-e to evaluate it.
>
>> >> > (vii) M-: (native-compile 'foo)
>> >> > This returns #<subr foo>
>> >> > (viii) M-: (foo)
>> >> > This returns the lambda form as a byte-compiled function.  This is a 
>> >> > bug:
>> >> > it should return the lambda form as a native-compiled function.
>
>> >> > Note: this bug is also in the emacs-29 branch.
>
>> >> Hi Alan,
>
>> >> I can reproduce, (native-compile 'foo) compiles only foo, compiling the
>> >> whole compilation unit with eg `emacs-lisp-native-compile-and-load'
>> >> compiles as expected also the inner lambda.
>
>> > Why would compiling a .el file compile inner lambda forms, but
>> > native-compile doesn't?
>
>> >> I'm not 100% convinced this behaviour is a bug tho.
>
>> > I don't understand that.  Why might it be incorrect to compile that inner
>> > lambda natively?
>
>> Hi Alan,
>
>> I'm not saying it would be incorrect.  I'm suggesting that if is not
>> specified what's the expected behaviour of compiling by name the outer
>> lambda it might not be a bug.
>
>> When we compile a whole compilation unit we indeed have to compile all
>> functions, in this case what we promised is I think not defined.
>
> I still don't understand that.  The doc string for native-compile says:
>
>     Compile FUNCTION-OR-FILE into native code.
>
> ..  I can't see any reason not also to compile inner lambda functions
> natively.
>
> Anyhow, to fix this bug (if such it be) is easy:
>
> diff --git a/lisp/emacs-lisp/comp.el b/lisp/emacs-lisp/comp.el
> index 181e5ca96a1..2360fbaa494 100644
> --- a/lisp/emacs-lisp/comp.el
> +++ b/lisp/emacs-lisp/comp.el
> @@ -1359,7 +1359,12 @@ comp-add-func-to-ctxt
>                (comp-ctxt-top-level-forms comp-ctxt)
>                (list (make-byte-to-native-func-def :name function-name
>                                                    :c-name c-name)))
> -        (comp-add-func-to-ctxt func))))
> +        (comp-add-func-to-ctxt func))
> +      ;; Handle any lambda functions in BYTE-CODE.
> +      (maphash (lambda (key val)
> +                 (unless (eq key (aref byte-code 1))
> +                   (comp-intern-func-in-ctxt key val)))
> +               byte-to-native-lambdas-h)))
>
>  (cl-defmethod comp-spill-lap-function ((form list))
>    "Byte-compile FORM, spilling data from the byte compiler."
>
>
> What do you say?

LGTM as long as indeed it does not regress any test. Speaking of which
with the patch I guess we want a test to cover this.

> Incidentally, the code in the various comp-spill-lap-function methods
> together with comp-intern-func-in-ctxt appears to have some code
> duplication.  Would it be possible to have the symbol and list methods of
> comp-spill-lap-function simply call comp-intern-func-in-ctxt the way the
> string method does?  That would simplify those two methods quite a bit.

Mmmh maybe, I think one has to try to see if the result is satisfactory.

Thanks

  Andrea