J3/03-138r1

To: J3
From: John Reid
Subject: Edits for UK comment TC4 (Type-bound generics)
Date: 1 April 2003

1. Comment TC4

    Change type-bound generics to be sets of specific named type-bound
    procedures.

    The generic type-bound procedure facility is difficult to understand and
    unnecessarily difficult to implement. It's confusing for the user when
    you explain that generic resolution produces a "slot number" in the
    dispatch table instead of a name.

    This confusion arises at least partly because although normal generics
    are a collection of (named) normal procedures, type-bound generics are a
    collection of UNNAMED type-bound procedures created by compiler magic
    from the list of the actual procedures implementing them at some level.

    It's also difficult to understand when, during inheritance, one is
    extending the generic vs. overriding an already-existing specific,
    precisely because the specifics don't have names.

    Section 4.5.1
    Page 44
    Syntax change: Instead of GENERIC :: <generic-spec> => <procedure-list>
      have GENERIC :: <generic-spec> => <type-bound-proc-list>

    Concomitant change:
    (a) The GENERIC statement always adds new specifics to the list, it
    never overrides them.
    (b) To override a specific within a generic, you use the specific name
    on the PROCEDURE statement, just like nongeneric tbps.
    i.e. to override a specific, you specific *the specific name*,
    to extend the generic, you specify *the generic spec*.

    Advantages:
    (1) Simpler exposition.
    (2) Simpler implementation.
    (3) Programs will be simpler to understand and maintain, because of the
    differentiation between overriding and extending.

    Disadvantages:
    (1) The user has to name the specifics instead of the compiler doing
    all the work.
    (2) The specific names have to be public if the user wants them to be
    overridable.

2. Specifications

That generic type-bound procedures be specified explicitly as a set of
non-generic type-bound procedures, instead of as the set of implementation
procedures.

A generic type-bound procedure may have an accessibility, but does not have
any of the other binding attributes, viz PASS/NOPASS and NON_OVERRIDABLE.
Each individual specific is individually overridable (or not).  Each
individual specific either has the PASS attribute or the NOPASS attribute.


3. Syntax

The syntax of the generic proc-binding statement shall be:

   GENERIC [ [ , <access-spec> ] :: ] <generic-spec> => <binding-name-list>

   Constraint: Each <binding-name> in the <binding-name-list> shall be the name
               of a specific binding of the type.

4. Edits to 02-007r3 and 03-166

Edits marked as "ESSENTIAL:" should be done (or at least considered) even if
the technical change proposed is not done, as they correct mistakes in the
current version.  Edits marked "USEFUL" are wording improvements that also
should be considered regardless.

[52:20] Change "<binding-attr-list>" to "<access-spec>", and
        change "<binding-list>" to "<binding-name-list>".
{Syntax change.}

[53:1-2] Delete.
{ESSENTIAL: This constraint is more than adequately covered by 16.2.  A generic
binding is a local entity of class 2, as is a specific binding.  There is no
"escape" clause for generic bindings, so "an identifier of a local of one class
shall not be used to identify another local entity of the same class" is quite
sufficient for our purposes.  The presence of this unnecessary constraint is
confusing.}

[53:2+] Add constraints
  "C447a (R442) Each <binding-name> in <binding-name-list> shall be the name of
         a specific binding of the type.
   C447b (R442) If <generic-spec> is not <generic-name>, each of its specific
         bindings shall have a passed-object dummy argument (4.5.3.3)."

[52:20+] Move the constraint at [53:23-25] to here and change
"<derived-type-def>" to "derived type".
{ESSENTIAL: Misplaced.}

[53:14,16] Delete ",R445" twice.
{No longer applicable.}

[53:18] Change "R443" to "R446".
{ESSENTIAL: This was the wrong rule number.  See C451 for a similarly scoped
constraint which has the correct rule number, viz R436.}

[53:19-20] Delete.
{This constraint was misplaced anyway.  The reworded version is C447b.}

[53:21-22] Delete.
{It is no longer possible to override a specific with a GENERIC statement.}

[53:23-25] See above: this constraint to be moved to [52:20+].

[53:30] Before last "." insert
  "for its specific bindings."
{Say what goes into the generic interface.  Obvious, but I don't think we
come right out and say it anywhere else for generic tbps.  We do it for
ordinary generic interfaces at [257:1-2].}

[53:30+] Insert new paragraph
  "A type-bound procedure may be identified by a <<binding name>> in the scope
   of the type definition.  This name is the <binding-name> for a specific
   binding, and the <generic-name> for a generic binding whose <generic-spec>
   is <generic-name>.  A final binding, or a generic binding whose <generic-
   spec> is not <generic-name>, has no binding name."
{ESSENTIAL: It's nice to say what we all know we mean.  And some have been
confused about whether final bindings can be invoked through the type via a
binding name equal to their procedure name.
EDITOR: Please index "binding name" to here.}

[53:31] Add "specific" before "binding".

[54:1] Append sentence and add note:
  "Each additional <generic-binding> with the same <generic-spec> extends the
   generic interface.
   Note 4.41a
   Unlike the situation with generic procedure names, a generic type-bound
   procedure name is not permitted to be the same as a specific type-bound
   procedure name in the same type (16.2)."
{It's nice to say what the syntax means, as well as allowing it.
ESSENTIAL: Some edit should be done here even if this TC is not accepted,
but not this one.}

[54:21+3] Replace "procedure bindings" with "procedures".
{USEFUL: These are called either "type-bound procedures" or "bindings".}

[58:1-3] Replace with
  "If a generic binding specified in a type definition has the same
   <generic-spec> as an inherited binding, it extends the generic interface
   and shall satisfy the requirements specified in 16.2.3."
{It is no longer possible to override a specific with a GENERIC statement.}

[58:5-7] Replace with
  "it extends the type-bound generic interface for <dtio-generic-spec> and
   shall satisfy the requirements specified in 16.2.3."
{It is no longer possible to override a specific with a GENERIC statement.}

[122:13,32] Delete "type-bound".
{USEFUL: The adjective is redundant here, and not needed.
Alternatively, change "type-bound generic binding" to "generic type-bound
procedure".}

[142:4] Delete "type-bound".
{USEFUL: As above, the unnecessary adjective is redundant.}

[199:8] Change "derived-type procedure" to "generic".
{USEFUL: The correct term for the means of specification is "generic binding",
and the correct term for what is specified is "generic type-bound procedure";
never "derived-type procedure binding".}

[203:10] Change "type-bound generic binding" to "generic type-bound procedure".
{USEFUL: The right term for a generic tbp is generic tbp.}

[262:26] Change "the name of a procedure binding" to "a binding name".
{USEFUL: Simplify and improve wording.}

[262:28-263:10] Replace with
  "Resolving references to type-bound procedures is described in 12.4.5."
{ESSENTIAL: It is a serious disservice to the reader to bury the details of
type-bound procedure invocation in the middle of 12.4.0.  It deserves its
own subsection.}

[272:19] Insert "named" before "procedure".
{ESSENTIAL: This section only handles named procedures, not procedure
components or type-bound procedures.}

[274:30+] Insert new section
  "12.4.5 Resolving type-bound procedure references

   If the <binding-name> in a <procedure-designator> (R1219) is that of a
   specific type-bound procedure, the procedure referenced is the one
   identified by the binding with that name in the dynamic type of the
   <data-ref>.

   If the <binding-name> in a <procedure-designator> is that of a generic type
   bound procedure, the generic binding with that name in the declared type of
   the <data-ref> is used to select a specific binding:
     (1) If the reference is consistent with one of the specific bindings of
         that generic binding, that specific binding is selected.
     (2) Otherwise, the reference shall be consistent with an elemental
         reference to one of the specific bindings of that generic binding;
         that specific binding is selected.
   The reference is to the procedure identified by the binding with the same
   name as the selected specific binding, in the dynamic type of the
   <data-ref>."
{Describe the simpler resolution procedure for generic type-bound procedures.
ESSENTIAL: Resolution of type-bound procedures deserves its own section.}

[397:20+3] Delete "type-bound" before "generic bindings".
{USEFUL: Adjective doesn't fit this noun phrase.}

[398:28] Change "A ... <generic-name>" to
  "The binding name (4.5.1.5) of a type-bound procedure"
{ESSENTIAL: The text is poor at the moment because it implies that a
generic binding for which the <generic-spec> is <generic-name> does not
have a binding name.}

[425:35] Change ", or" to ",".
[425:36] Before "." insert ", or as part of the finalization process".
{ESSENTIAL: Existing text omits the final subroutine case.}

[470:1-20] Delete.
{There is no longer anything complicated about generic type-bound procedure
dispatch, so no need to explain it.  Generic tbps now use generic resolution
to select a specific binding, and that dispatches in the normal way.}
===END===