[m-rev.] for review: start on replacing --intermod-directories and related options

[Julien Fischer]

On Thu, 14 Nov 2024 at 09:34, Zoltan Somogyi <zoltan.somogyi at runbox.com>
wrote:

>
> On 2024-11-13 23:41 +11:00 AEDT, "Zoltan Somogyi" <
> zoltan.somogyi at runbox.com> wrote:
> > > Regardless, it does
> >> not
> >> seem to be a useful behaviour *now*.
> >
> > Then I will next look into undoing it.
>
> I have now done so, with the result being the attached diff.
> It was much more interesting (in the Chinese curse sense)
> than I expected.
>
> For review by anyone. I would particularly like to know
> whether there is a better way to force the creation of .opt files
> before the compilation of their would-be consumers than the one
> adopted by this diff. Until now, I assumed that this was solved long ago,
> but apparently, this is not the case with mmake.
>

 > Don't treat .m files as substitutes for .opt files.
>
> compiler/write_deps_file.m:
>     When trying to check for the existence of .opt files, whether alone
>     or together with .trans_opt files, do not consider a .m file as
>     an acceptable substitute. Delete the option that can ask for this
>     behavior.

...

> diff --git a/tests/term/existential_error1.m
b/tests/term/existential_error1.m
> index f336133eb..4e54b4f1b 100644
> --- a/tests/term/existential_error1.m
> +++ b/tests/term/existential_error1.m
> @@ -4,10 +4,39 @@
>  %
>  % Regression test for term_norm.m
>  % Symptom: "Software Error: Unmatched lists in functor_norm_filter_args."
> -% This was caused by the list of counted arguments in the weight table
differing
> -% from the list of arguments the termination analyser provided when it
called
> -% functor norm.  The code that constructed the weight table was ignoring
> -% type_infos when constructing the list of counted arguments.
> +% This was caused by the list of counted arguments in the weight table
> +% differing from the list of arguments the termination analyser provided
> +% when it called functor norm. The code that constructed the weight table
> +% was ignoring type_infos when constructing the list of counted
arguments.
> +%
>
+%---------------------------------------------------------------------------%
> +%
> +% This test case happens to also expose another issue about intermodule
> +% optimization.
> +%
> +% Whether the deconstruct_univ predicate below terminates or not is
determined
> +% entirely by whether private_builtin.typed_unify terminates or not.
However,
> +% private_builtin.opt and private_builtin.trans_opt in the library
directory
> +% disagree about this: the .opt file gives typed_unify's termination
status
> +% as can_loop, while .trans_opt gives it as cannot_loop.

That initially struck me as strange, but it is correct. Proving termination
of typed_unify/2 requires knowing that type_of/1 terminates and we can only
know that from type_desc.opt.
(I think the comment in compiler/termination.m above
set_compiler_gen_terminates/6 is wrong: it implies that the termination
status
of *all* predicates in builtin and private_builtin is set by that code, but
doesn't seem to be what it actually does.)

The diff looks fine.

Julien.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mercurylang.org/archives/reviews/attachments/20241115/94ada0f4/attachment.html>