[m-dev.] proposed syntax for typed insts

Zoltan Somogyi zoltan.somogyi at runbox.com
Wed Sep 9 04:54:26 AEST 2015

On Tue, 8 Sep 2015 15:01:36 +1000 (AEST), Julien Fischer <jfischer at opturion.com> wrote:
> At this stage I would say no, but it depends a bit on the
> details of how typed insts are going to fit into the language.
> (For example, in their presence will we eventually restrict what
> untyped insts are allowed?)

The obvious first step is simply to record the type with the inst's definition,
and insist on *that* type matching the inst, not just *some* accessible type,
as we do now.

For this, we need to decide *what* type information we allow to be specified
for an inst.

I expect that 99% of defined insts involve function symbols from only one
type constructor. The list_skel and uniq_tree234 insts are both like that.
However, I can see that in some cases, one would want to define not just
insts like this:

:- inst at_most_one_key(K, V) for tree234(TK, TV) ==
        two(K, V, uniq_tree234(K, V), uniq_tree234(K, V))

but also insts like this:

:- inst at_most_one_yes_key(EK, V) for tree234(maybe(TEK), TV) ==
        two(yes(EK), V, uniq_tree234(K, V), uniq_tree234(K, V))

(The more realistic application I am thinking of is insts
that represent patterns involving more than one node type
in abstract syntax trees, but I didn't want to invent an AST
just for this post :-()

A type constructor applied to a list of unique type variables
can be represented by just its name and arity (as my original
proposal showed), but this shorthand doesn't work if the type
constructor's arguments are not variables (or aren't *distinct*

One could define at_most_one_yes_key in terms of at_most_one_key,
by providing the appropriate inst (bound(yes(EK)) for K, so this is
a tradeoff between simplicity of implementation and ease of use.

The second step would be modifying the mode checker to require
that defined inst that *do* specify what type or type template,
as above) they are supposed to apply to *are* only applied to
that type or type template.

At that point, we could (third step) announce the new capability,
and (fourth step, after waiting while everyone installs a new compiler)
we could start using it ourselves.

Fifth step, we could eventually generate warnings for inst definitions
that do not specify a type template. After the next release with that
warning, we could (sixth step) generate errors for such insts,
effectively deleting them from the language.

I expect that the second step would be by far the hardest to code,
especially if we decide to allow nontrivial type templates.
The longest would be the waiting periods (at least a few months,
probably a year) on both sides of step five.

Anyone see any flaws in the above? Would anyone want to keep
untyped insts around forever?

BTW, I added show_ops.m to extras, after documenting it a bit.


More information about the developers mailing list