[m-rev.] for post-commit review: register typeclass_info args more directly

[Julien Fischer]

Hi Zoltan,

On Sat, 19 Jun 2021, Zoltan Somogyi wrote:

> Update rtti_varmaps more directly.
> 
> Replace three traversals over the arguments of a typeclass_info
> with one traversal.
> 
> Add an XXX about an old problem.


> diff --git a/compiler/polymorphism_type_class_info.m b/compiler/polymorphism_type_class_info.m
> index 949f6dd23..c3791292e 100644
> --- a/compiler/polymorphism_type_class_info.m
> +++ b/compiler/polymorphism_type_class_info.m

...

> @@ -899,44 +898,51 @@ record_constraint_type_info_locns(Constraint, ExtraHeadVar, !Info) :-
>      % Find all the type variables in the constraint, and remember what
>      % index they appear in the typeclass info.
> 
> +    poly_info_get_rtti_varmaps(!.Info, RttiVarMaps0),
>      % The first type_info will be just after the superclass infos.
> -    First = NumSuperClasses + 1,
> -    Last = NumSuperClasses + ClassArity,
> -    assoc_list.from_corresponding_lists(ClassTypes, First `..` Last,
> -        IndexedClassTypes),
> +    record_tci_slots_for_unseen_or_in_type_info_tvars(ExtraHeadVar,
> +        ClassTypes, NumSuperClasses + 1, RttiVarMaps0, RttiVarMaps),
> +    poly_info_set_rtti_varmaps(RttiVarMaps, !Info).
>
>      % Work out which type variables we haven't seen before, or which we
>      % assumed earlier would be produced in a type_info (this can happen for
>      % code which needs mode reordering and which calls existentially quantified
>      % predicates or deconstructs existentially quantified terms).
> -    poly_info_get_rtti_varmaps(!.Info, RttiVarMaps0),
> -    NewTVarAndIndex =
> -        ( pred(TVarAndIndex::out) is nondet :-
> -            list.member(Type - Index, IndexedClassTypes),
> -            type_vars(Type, TypeVars),
> -            list.member(TypeVar, TypeVars),
> -            ( if
> -                rtti_search_type_info_locn(RttiVarMaps0, TypeVar, TypeInfoLocn)
> -            then
> -                TypeInfoLocn = type_info(_)
> -            else
> -                true
> -            ),
> -            TVarAndIndex = TypeVar - Index
> -        ),
> -    solutions(NewTVarAndIndex, NewClassTypeVars),
> -
> -    % Make an entry in the TypeInfo locations map for each new type variable.
> -    % The type variable can be found at the previously calculated offset
> +    %
> +    % Then make an entry in the TypeInfo locations map for each new type
> +    % variable. The type variable can be found at the calculated offset
>      % with the new typeclass_info.
> -    MakeEntry =
> -        ( pred(IndexedTypeVar::in, R0::in, R::out) is det :-
> -            IndexedTypeVar = TheTypeVar - Index,
> -            Location = typeclass_info(ExtraHeadVar, Index),
> -            rtti_set_type_info_locn(TheTypeVar, Location, R0, R)
> +    %
> +:- pred record_tci_slots_for_unseen_or_in_type_info_tvars(prog_var::in,
> +    list(mer_type)::in, int::in, rtti_varmaps::in, rtti_varmaps::out) is det.
> +
> +record_tci_slots_for_unseen_or_in_type_info_tvars(_, [], _, !RttiVarMaps).
> +record_tci_slots_for_unseen_or_in_type_info_tvars(ExtraHeadVar,
> +        [ClassType | ClassTypes], CurIndex, !RttiVarMaps) :-
> +    type_vars(ClassType, TypeVars),
> +    list.filter(is_unseen_or_in_type_info_tvar(!.RttiVarMaps),
> +        TypeVars, UnSeenOrInTypeInfoTypeVars),
> +    Location = typeclass_info(ExtraHeadVar, CurIndex),
> +    InsertIntoRttiVarMap =
> +        ( pred(TVar::in, R0::in, R::out) is det :-
> +            rtti_set_type_info_locn(TVar, Location, R0, R)
>          ),
> -    list.foldl(MakeEntry, NewClassTypeVars, RttiVarMaps0, RttiVarMaps),
> -    poly_info_set_rtti_varmaps(RttiVarMaps, !Info).
> +    % XXX If ClassType contains more than one type variable, this records
> +    % Location as applying to ALL OF THEM. This code is inherited from
> +    % the time when the parameters of typeclasses *had* to be type variables,

I'm not sure what that means; the parameters of typeclasses still have to be
type variables.  Do you mean typeclass constraints?

The diff is fine.

Julien.