Re: ForTran-Ada + flamette + question

Re: ForTran-Ada + flamette + question

[richw]

The following includes a suggestion for solving your problem, along
with my own defense of the decision to omit procedure-pointers from Ada:

The first solution I can think of (and, I guess, the simplest?) would
be to define an enumeration type, where each enumeral (is that a
word?) corresponds to one of the functions you'd like to bind to a
key-sequence.  Then just write a procedure that takes such an enumeral
as an "in" parameter and then "case"s on that parameter to the
appropriate call.

Abstractly, this new function is similar to "eval" in Lisp.  Now, instead
of binding characters sequences to function-pointers (as in C), just bind
the character sequences to enumerals.

Other comments about why I think the ability to pass procedures around
was omitted from Ada: (note that these are just opinions -- I'm not
quoting one of the design documents for Ada or anything like that)

In a language like Lisp, you can actually create procedures AT RUN-TIME
since they're just lists (i.e. data).  However, in C and Ada this isn't
true.  In C, all procedures that you can point to at run-time are defined
at compile time.  I think this last fact is an important one to consider;
the above method of mapping enumerals to subprogram calls works for Ada.
It wouldn't work in Lisp because you can't list all of the subprograms
that will exist when you run the program BEFORE you run it (because you
can build Lisp procedures "on the fly"); thus, you can't define an
enumeration type since that type has a finite number of compile-time
enumerals.

Anyway, my point is that I think that one can do all that you can do in
C in Ada by using the above enumerals-to-subprogram mapping method; I'd
be interested if people can provide counter-examples (I realize that
you'd have to do more if the procedures you call take different numbers
of arguments, but then again, you'd have to be more fancy in C, too).

If so, then omitting the ability to pass procedures around in Ada saved
the language designers from adding more syntactic and semantic and
run-time constructs to the language to support this.  Considering
the fact that Ada is already big, I think this was a wise decision --
especially since, with a minimal amount of work, you CAN do what you
want (using the enumeral-to-subprogram method).

Think about it -- strict type-checking becomes much more messy...
It is possible to add this to your language (e.g. CLU), but I personally
don't think it's worth complicating your language.  Even though C lets
you do it, C BY NO MEANS does type-checking of calls to procedures which
were passed in as arguments (it doesn't even check that the number of
arguments is correct).  Yes, C is an "unsafe" language...

Comments?  Remember, these are casual remarks and opinions -- please
think before you flame  :-)

Rich

Re: ForTran-Ada + flamette + question

[Russ Holden]

> 
> The following includes a suggestion for solving your problem, along
> with my own defense of the decision to omit procedure-pointers from Ada:
> 
> The first solution I can think of (and, I guess, the simplest?) would
> be to define an enumeration type, where each enumeral (is that a
> word?) corresponds to one of the functions you'd like to bind to a
> key-sequence.  Then just write a procedure that takes such an enumeral
> as an "in" parameter and then "case"s on that parameter to the
> appropriate call.
>
> Anyway, my point is that I think that one can do all that you can do in
> C in Ada by using the above enumerals-to-subprogram mapping method; I'd
> be interested if people can provide counter-examples (I realize that
> you'd have to do more if the procedures you call take different numbers
> of arguments, but then again, you'd have to be more fancy in C, too).


Rich, you are missing out on the beauty of procedure passing.  Simply put
it gives you the ability to dynamically bind actions without the invoker
of the procedure having to know about the procedure being called.  A 
couple of examples are:

1) Table-driven code - Using C all you have to do is add a procedure
   to a table.  Using the ADA enumerated-to-procedure mapping code
   modifications are involved (see below).

2) It provides the ability for higher levels of software to abstract
   procedures from lower levels such that lower levels call but do
   not "use" these procedures (in Parnas terminology).  For example,
   (and this is not hypothetical) is a DBMS which wants to perform
   qualification of records against user-specified predicates before
   retreiving them (copying them off their page essentially).  Since
   the code performing page level manipulations is generally at a much
   lower level in processing than the code which understands about
   the values in records a procedure made available to the page level
   is an appropriate mechanism.  The use of enumerated types does no
   good and there are no real good solutions that I know of in ADA
   (if anyone has any please send me mail).

   I have used procedure pointers in lots of other situations where
   a high-level wants or needs to provide a service to a lower level.
   The only requirement is that a "standard" invokation mechanism be
   defined.
-- 
Russell Holden
Computer Corporation of America
Four Cambridge Center
Cambridge, MA 02142

Re: ForTran-Ada + flamette + questi

[richw]

>>  ***** ada-uts:net.lang.ada / cca.UUCP!holden /  8:46 am  May 30, 1986
>>  Rich, you are missing out on the beauty of procedure passing.
>>  (Russell Holden)

Pardon me, but...
I've actually used this feature in many languages: Lisp, CLU, C, and
Smalltalk.  As a matter of fact, I'm in the process of rewriting a
C program (which used pointers to functions) in Ada.

>>  Simply put it gives you the ability to dynamically bind actions
>>  without the invoker of the procedure having to know about the
>>  procedure being called.

So does the enumeration-to-procedure mapping method I proposed.

>>                         ...  The use of enumerated types does no
>>  good and there are no real good solutions that I know of in ADA

I'm sorry, but I simply fail to see how your examples provide
conclusive arguments that this method will not work in general.
GRANTED, the mapping method is less convenient for the programmer,
but my point was that it is nevertheless possible in Ada at a
reasonable expense (which I think is justifiable given that omitting
procedure pointers from Ada prevented Ada from becoming even more
complex).

To be more concrete, the following outlines a package spec which
one could use:

    package Mapper is

          type No_Argument_Procedure_Pointer is (
              proc_A_0, proc_B_0, ... );

          type One_Integer_Argument_Procedure_Pointer is (
              proc_A_1_int, proc_B_1_int, ... );
                   :

          procedure Call (
              p : in No_Argument_Procedure_Pointer );

          procedure Call (
              p : in One_Integer_Argument_Procedure_Pointer;
              i : in Integer );
                   :
    end Mapper

The implementation of each "Call" procedure would simply be a case
statement to the appropriate "real" procedure call.

If anyone could provide an example where a Procedure_Pointer could
not be used wherever one would like to use a "real" pointer to a
procedure in a different langauge, I'd be grateful for a reply.

Again, I apologize if I'm just being dense, but...


Rich Wagner

Re: ForTran-Ada + flamette + questi

[Russ Holden]

I know this has been beaten to death (and put to sleep) but...

> >>  Rich, you are missing out on the beauty of procedure passing.
> >>  Simply put it gives you the ability to dynamically bind actions
 
> So does the enumeration-to-procedure mapping method I proposed.
> 
> To be more concrete, the following outlines a package spec which
> one could use:
> 
>     package Mapper is
> 
>           type No_Argument_Procedure_Pointer is (
>               proc_A_0, proc_B_0, ... );
> 
>           type One_Integer_Argument_Procedure_Pointer is (
>               proc_A_1_int, proc_B_1_int, ... );
>                    :
> 
>           procedure Call (
>               p : in No_Argument_Procedure_Pointer );
> 
>           procedure Call (
>               p : in One_Integer_Argument_Procedure_Pointer;
>               i : in Integer );
>                    :
>     end Mapper
> 
> The implementation of each "Call" procedure would simply be a case
> statement to the appropriate "real" procedure call.

There were a bunch of examples better than the ones I provided given
during the lengthy discussion about procedure parameters over the
past few weeks which point out the insufficiency of this (not to
mention its inherent ugliness and limits).  Also shown were mechanism
for making the checking of procedure parameters quite possible (for
example by requiring that procedures used in that manner be strongly 
typed).

> 
> If anyone could provide an example where a Procedure_Pointer could
> not be used wherever one would like to use a "real" pointer to a
> procedure in a different langauge, I'd be grateful for a reply.
> 
> Again, I apologize if I'm just being dense, but...
> 

OK, apology accepted.  


-- 
Russell Holden
Computer Corporation of America
Four Cambridge Center
Cambridge, MA 02142

Re: Enumerated "procedural" parameters

[Alastair Milne]

    Using an enumeration-indexed case to choose the procedure to be called
    will work if all the procedures to be called are known and available at
    the time that case is compiled.  It will therefore work if all the
    enumeration-passing calls are in the same program unit as the case, or if
    they are exported by a package which will permit only "procedures" known
    to it (its own, or others from packages used in the specification) to be 
    passed in.

    It will not work in a library package exporting routines to be applied to 
    routines from units not known to the package.  Routines that come to mind 
    immediately are in the field of numeric analysis: curve fitting, best fits,
    adaptive quadrature, etc. .  In these cases what one needs is a routine 
    which, given some routine x to produce original function values, calls 
    that routine as it needs for input to its own algorithm.  Such packages
    would be essential parts of libraries for numeric analysis.  Their routines
    could be applied to any data-producing routine a user supplied,
    with algorithms 50 years old or produced yesterday.  But in Ada, the old 
    approach of simply passing that routine as another parameter is 
    not available.

    A good point has already been made on this board that in using Ada, one
    can find that the way one used to do something is no longer available,
    only to discover that it can be done in new and superior, more secure ways.
    I hope that will turn out to be so of this case.


    Alastair Milne

Re: Enumerated "procedural" parameters

[Dick Dunn]

I think it's unfortunate that anyone ever suggested the idea of enumerating
procedures which might be called and suppying an enumeration value instead
of a procedural parameter.  The question [I hope!] was never whether one
could construct an Ada program which would have the same effect as a
program in another language using procedural parameters...Turing
equivalence is a nice tool but it's not the point.

The fact that not all languages have procedural parameters gives some
weight to the Ada position.

The fact that C, Pascal, Modula, FORTRAN, Algol, and many other languages
do have procedural parameters gives weight to the position that the lack of
procedural parameters is a shortcoming of Ada.

The argument that procedural parameters can be simulated is really
irrelevant.  The issues have to do with convenience and safety, not simple
expressive power in a theoretical sense.  (Most languages are equivalent in
that theoretical sense.)  Hence the discussion would be much more useful if
it concentrated on Ada solutions which represent something like procedural
parameters with a notation comparably concise and convenient.  Adding an
enumerated type, constructing the (unenforced!) correspondence between
enumeration values and procedures, and replacing a single procedure
invocation with a case statement over an enumeration value (assuming that
the set of procedures of interest is actually a priori enumerable!) is not
at all the same in any way of interest to a practicing programmer.  A
better answer is necessary here.
-- 
Dick Dunn	{hao,ucbvax,allegra}!nbires!rcd		(303)444-5710 x3086
   ...Simpler is better.

Re: ForTran-Ada + flamette + questi

[friesen]

In article <4700042@ada-uts> richw@ada-uts writes:
>
>>>  Simply put it gives you the ability to dynamically bind actions
>>>  without the invoker of the procedure having to know about the
>>>  procedure being called.
>
>So does the enumeration-to-procedure mapping method I proposed.
>
>>>                         ...  The use of enumerated types does no
>>>  good and there are no real good solutions that I know of in ADA
>
>I'm sorry, but I simply fail to see how your examples provide
>conclusive arguments that this method will not work in general.
>GRANTED, the mapping method is less convenient for the programmer,
>but my point was that it is nevertheless possible in Ada at a
>reasonable expense (which I think is justifiable given that omitting
>procedure pointers from Ada prevented Ada from becoming even more
>complex).
>
        I fail to see how an enumeration can solve the general
problem(even given your example). Given that at the time you write the
mapper package you do *not* know what procedures will be called by it,
and that *most* of the procedures to be used do not even exist yet,
*how* can you create an enumeration type that covers the set? The
whole point of function pointers as arguments is that it allows the
writing of *generic* procedures which can be "customized" by the
individual user by providing a pointer to a user supplied routine,
this allows the general routine to be used on data structures or
views that were not anticipated when the original routine was
written. If you can explain how to access non-existant functions,
whose names you do not even know, using an enumeration mapper, then I
will I will admit that this is equivalent to function pointers.
--

                                Sarima (Stanley Friesen)

UUCP: {ttidca|ihnp4|sdcrdcf|quad1|nrcvax|bellcore|logico}!psivax!friesen
ARPA: ??

Re: ForTran-Ada ... (really function pointers)

[markb]

For 95% of the cases (e.g., Intgrators, numerical analysis, graphics etc.)
Generics provide a reasonable replacement for function pointers.
The other 5% of the time (e.g., EMACS) the emuneration mapping seems
to provide a moderatly reasonable replacement.  I have still not seen an
example that is not covered by one of these two alternitives.

Mark Biggar
{allegra,burdvax,cbosgd,hplabs,ihnp4,akgua,sdcsvax}!sdcrdcf!markb

Re: ForTran-Ada + flamette + questi

[richw]

Sarima (Stanley Friesen) writes:

>>          I fail to see how an enumeration can solve the general
>>  problem (even given your example).  Given that at the time you
>>  write the mapper package you do *not* know what procedures will
>>  be called by it, and that *most* of the procedures to be used do
>>  not even exist yet, *how* can you create an enumeration type
>>  that covers the set?

Your point is valid; I did not explicitly state that modification of
the mapper package is necessary when newly written procedures are to
be passed (I didn't realize this wasn't obvious).  So, in terms of
program maintainance, yes, the mapping method is less convenient; if
you remember, I admitted from the onset that it WAS less convenient.
In any case, thanks for pointing out this ambiguity.

(Note, however, that the pain of having to modify the mapper package
 has nothing to do with WHAT you can do using the method; this mapping
 method STILL provides the same functionality as procedure passing.)

This raises an interesting question: can the mapper package be written
so that when new procedures are to be added, only the body of the
mapper package needs to be recompiled?  This is important when consider-
ing program maintainance because a change to the mapper package's spec
will force recompilation of all users of the mapper (something one would
like to avoid).  The last part of this note outlines a revised spec and
body which avoids having to modify the spec when adding procedures.

>>  The whole point of function pointers as arguments is that it
>>  allows the writing of *generic* procedures which can be
>>  "customized" by the individual user by providing a pointer to a
>>  user supplied routine, this allows the general routine to be
>>  used on data structures or views that were not anticipated when
>>  the original routine was written.

Ah, I'm GLAD you mentioned this.  The use of Ada's generics for such
situations is MUCH more advantageous than passing pointers.  If you
think about it, you can write a much more general "sort" procedure
using generics as opposed to procedure passing because generics lets
you not only vary the procedure used in comparing elements while
sorting -- it also lets you parameterize the TYPE of the elements in
the array you're sorting (assuming you're sorting arrays); can't do
that with procedure passing...


The following is that revised mapper package sketch I promised earlier;
it illustrates how one would pass two types of subprograms: simple
procedures which take no arguments and functions on Strings which return
Integers.

The basic idea involves moving the enumeration type declaration into
the package body.  The 'POS of the enumeration values are passed
around instead of the values themselves.  To make up for the
inability of users to refer to enumeration values by name (since
they no longer appear in the spec), the 'IMAGEs of the enumeration
values are used to create procedure pointers instead.  The costs of
changing the spec in these ways are: (1) More overhead due to the
use of 'VALUE in the body ('POS is not costly -- basically just a
type conversion) and (2) A misspelled procedure name is caught at
runtime (via the UNKNOWN_SUBPROGRAM exception) rather than link
time.

------------------------------------------------------------------------

package Mapper is

    type Procedure_with_0_arguments is private;
    type Function_with_1_in_String_return_Integer is private;

    UNKNOWN_SUBPROGRAM : exception;

    function Create (procedure_name : in String)
        return Procedure_with_0_arguments;
        -- Raises : UNKNOWN_SUBPROGRAM
    function Create (function_name : in String)
        return Function_with_1_in_String_return_Integer;
        -- Raises : UNKNOWN_SUBPROGRAM

    procedure Call (
        p : in Procedure_with_0_arguments);
    function  Call (
        f : in Function_with_1_in_String_return_Integer;
        s : in String)
            return Integer;

    procedure Call (
        procedure_name : in String);
        --
        -- Raises : UNKNOWN_SUBPROGRAM
        --
        -- This is functionally equivalent to:
        --           Call (Create (procedure_name));
        --
    function Call (
        function_name : in String;
        s : in String)
            return Integer;
        --
        -- Raises : UNKNOWN_SUBPROGRAM
        --
        -- This is functionally equivalent to:
        --           Call (Create (function_name), s);

private

    type Procedure_with_0_arguments is new Integer;
    type Function_with_1_in_String_return_Integer is new Integer;

end;

----------------------------------------------------------------------

with proc1;
with proc2;
with func1;
with func2;
     :

package body Mapper is

    type Enums_for_Procedure_with_0_arguments is (
        proc1_enum, proc2_enum, ...);
    type Enums_for_Function_with_1_in_String_return_Integer is (
        func1_enum, func2_enum, ...);
    --
    -- The above will need to be added to when new subprograms are
    -- to be passed.

    function Create (procedure_name : in String)
        return Procedure_with_0_arguments
    is begin
        return Enums_for_Procedure_with_0_arguments'POS (
               Enums_for_Procedure_with_0_arguments'VALUE (
                   procedure_name & "_enum"));
    exception
        when CONSTRAINT_ERROR =>
             raise UNKNOWN_SUBPROGRAM;
    end;

    function Create (function_name : in String)
        return Function_with_1_in_String_return_Integer
    is begin
        return Enums_for_Function_with_1_in_String_return_Integer'POS (
               Enums_for_Function_with_1_in_String_return_Integer'VALUE (
                   function_name & "_enum"));
    exception
        when CONSTRAINT_ERROR =>
             raise UNKNOWN_SUBPROGRAM;
    end;

    -- The following two subprograms will need to be added to when
    -- new subprograms are to be passed.
    --
    procedure Call (p : in Procedure_with_0_arguments)
    is begin
        case Enums_for_Procedure_with_0_arguments'VAL (p) is
            when proc1_enum => proc1;
            when proc2_enum => proc2;
                     :
        end case;
    end;

    function  Call (
        f : in Function_with_1_in_String_return_Integer;
        s : in String)
            return Integer
    is begin
        case Enums_for_Function_with_1_in_String_return_Integer'VAL (f) is
            when func1_enum => return func1 (s);
            when func2_enum => return func2 (s);
                     :
        end case;
    end;

    procedure Call (
        procedure_name : in String)
    is begin
        Call (Create (procedure_name));
    end;

    function Call (
        function_name : in String;
        s : in String)
            return Integer
    is begin
        return Call (Create (function_name), s);
    end;

end;

------------------------------------------------------------------------

Note the "_enum" suffixes for the enumeration literals; I included them
because I'm not sure offhand if there'd be any problem with overloading
the enumeration literals and the subprogram names they represent.

If not, then removing the "_enum" suffix (and removing the call to "&"
in the Create functions) would be a good thing; in any case, you get the
idea...

-- Rich Wagner

Re: ForTran-Ada + flamette + questi

[vilot]

In article <4700062@ada-uts> richw@ada-uts writes:
>
>Sarima (Stanley Friesen) writes:
>
>>>          I fail to see how an enumeration can solve the general
>>>  problem (even given your example).  Given that at the time you
>>>  write the mapper package you do *not* know what procedures will
>>>  be called by it, and that *most* of the procedures to be used do
>>>  not even exist yet, *how* can you create an enumeration type
>>>  that covers the set?
>
>Your point is valid; I did not explicitly state that modification of
>the mapper package is necessary when newly written procedures are to
>be passed (I didn't realize this wasn't obvious).  So, in terms of
>program maintainance, yes, the mapping method is less convenient; if
>you remember, I admitted from the onset that it WAS less convenient.
>In any case, thanks for pointing out this ambiguity.
>
>(Note, however, that the pain of having to modify the mapper package
> has nothing to do with WHAT you can do using the method; this mapping
> method STILL provides the same functionality as procedure passing.)
>

I noticed that passing objects of type task was rejected very early in
this discussion.  Why?  (Please don't plead  run-time  "performance".)

It seems to provide  a reasonable approximation of  C's semantics  for
passing  function  pointers.   Ignoring  the   aesthetic arguments and
assuming someone  Really  Needs to do  this, are there any significant
reasons why it cannot be used to solve the problem?

(P.S. A simple bit of code passed the Verdix compiler, at least.)

Re: ForTran-Ada + flamette + questi

[stt]

Unfortunately, passing a task doesn't do the trick
because the task type (which must be specified) implies
the code for the task.  Of course tasks are great for
holding onto a bit of state, if that is your primary goal.

Re: ForTran-Ada + flamette + questi

[Christopher Reedy]

Since I didn't see any responses to this:
In article <4700073@ada-uts> stt@ada-uts writes:
>
>Unfortunately, passing a task doesn't do the trick
>because the task type (which must be specified) implies
>the code for the task.  Of course tasks are great for
>holding onto a bit of state, if that is your primary goal.

There was a nice talk given at the 1985 Ada Europe conference in Paris
on this issue.  The thrust of the talk was that one needs an additional
task to act as an intermediary between the invoking and the invoked
tasks.  The intermediary task makes no entry calls, simply receiving
calls from both of the other tasks.  Thus, the invoking task need not
know the type of the invoked task.  (If anyone would like to see a
simple example, I have one.  However, it takes about three pages of
listing.)

In most cases this seems like a lot of overhead to go through to make a
variable procedure call.  However, the classical asynchronous buffer is
an example of this sort of approach.  In fact, if one was to combine the
intermediary task with an asynchronous buffer or some similar mechanism,
this approach might make a great deal of sense.

Chris Reedy
Computer Corporation of America
--Standard disclaimer: All opinions are my own.--