[m-rev.] for review: in_all_grades/in_tailrec_grades_only

[Julien Fischer]

On Tue, 19 May 2026 at 11:46, Zoltan Somogyi <zoltan.somogyi at runbox.com> wrote:
>
>
> Looking at the tests, I remembered that there is one more difference between LLDS
> and MLDS grades wrt tail recursion, beside the one in my last email. It applies to situations
> where two predicates such as even and odd in the invalid require_tailrec_? tests, call each
> other, but only the calls in one direction are tail calls; the calls in the other direction are not.
>
> In such situations, the LLDS code generator can make the tail calls reuse the stack frame
> (because it controls the code generated for the call), while the MLDS code generator
> cannot do that (since in that case, it is the target language compiler that is in control).
> Basically, the LLDS code generator works on tail *calls*, while the MLDS one works on
> *TSSCs*, meaning SCCs of procedures that *all* call each other via *tail* calls.
>
> I am inclined to ignore this difference, and let the new "in_tailrec_grades_only" mean
> grades that support *either* kind of tail recursion. We *could* split in_tailrec_grades_only,
> but I don't see much point. This is because
>
> - in such cases, the MLDS cannot reuse any stack frames, but even the LLDS can reuse
>   only half of the stack frames, so running out of stack is still a problem,
>
> - pretty much everyone will want to use an MLDS grade in production at some point,
>   at which time they need to conform to its requirements,
>
> - the diagnostic we generate for the mutually recursive calls whose behavior differs between
>   LLDS and MLDS explains those requirements quite well, and
>
> - most cases of mutual recursion can be made to conform to those requirements,
>   if they do not naturally conform to begin with.
>
> Do you guys agree?

Yes.

Julien.