TEAM-ADA Archives

Team Ada: Ada Programming Language Advocacy


Options: Use Forum View

Use Monospaced Font
Show Text Part by Default
Show All Mail Headers

Message: [<< First] [< Prev] [Next >] [Last >>]
Topic: [<< First] [< Prev] [Next >] [Last >>]
Author: [<< First] [< Prev] [Next >] [Last >>]

Print Reply
Randy Brukardt <[log in to unmask]>
Reply To:
Randy Brukardt <[log in to unmask]>
Fri, 17 May 2002 21:01:39 -0500
text/plain (151 lines)
> > 2) Language Revision
> > --------------------
> > This issue naturally leads to the question: can the next language
> > revision for Ada provide the missing guarantee, assuring a match
> > in comparisons of subprogram access values that designate the same
> > subprogram?
> No. Over my dead body.

I suppose you want a detailed technical explanation? Sigh.

There are various reasons that the profile of a subprogram may be different
depending on context. If that is the case, taking 'Access may not just be
taking the address of the machine code for the subprogram. A wrapper of some
sort will need to be generated in order to make the profile of the access
type match up with the profile of the actual subprogram.

The usual technique is to create the wrapper (which is really just a
compiler-generated subprogram) at the point of the use of 'Access (or a
convenient nearby point). Thus, the 'Access value itself is the address of
the wrapper, not the address of the subprogram, and two 'Access usages that
are too far apart will end up with different values.

This is a real issue with compilers that support any kind of generic
sharing. The way that generic units are shared is that each subprogram in
the generic includes an additional, compiler-generated parameter to the data
and description of the particular instance that is being used. Of course,
the profile of the subprogram in an instance matches that of a similar
subprogram declared outside of the instance. So, while the logical profiles
match, the actual machine code is quite different. The compiler has to use a
wrapper to make the profiles match.

A quick example may make this clearer:

    type Proc_Access is access procedure;

    procedure Proc1;

       Cnst : in Integer;
    procedure GProc;

    procedure Proc2 is new GProc (10);
    procedure Proc3 is new GProc (20);

    V : Proc_Access;

    if <complex expr.> then
       V := Proc1'Access;
       V := Proc2'Access;
    end if;


Assume GProc is a shared generic (meaning that there is only one instance of
the code for the body). In order for Proc2 and Proc3 to determine the value
of Cnst, it has to be passed to calls to GProc. (Usually, of course, there
are many parameters and private package data; the compiler will probably
pass a record of this stuff rather than one parameter for each item.) The
generated body for GProc would look something like:

    procedure GProc (~Cnst : in Integer) is ...

(where ~Cnst is a compiler-generated name).

The machine address of Proc2 and Proc3 is this body. If you wrote a call to


the compiler would generate:

    GProc (10);

to pass the generic data.

Now, if you just used the machine address of the subprogram (GProc) when
taking 'Access, how would you know whether to add the extra parameter, and
what its value is? The solution is to create a wrapper subprogram to
'stand-in' for the real body.

    procedure ~Proc2 is
        GProc (10);
    end ~Proc2;

(This, by the way, is the reason why you can't take 'Access in a generic
body for a type declared outside of the generic. In that case, there would
be no way to generate a thunk that could access the generic parameter data,
because the only way to access that data is to pass it in as a parameter --
but that would change the profile and defeat the purpose.)

One could imagine avoiding the 'different value' problem by building only
one such thunk for the entire program. However, consider separate
compilation: the only 'Access for a given subprogram might be buried in a
separate subprogram somewhere. How is the compiler going to figure out that
one exists?

An alternative would be to generate appropriate ones for every subprogram in
every generic instantiation. But 'Access is rare, so most of these would be
unused. Plus, it would have to be done for every possible access profile.
Even if the compiler has good code elimination, the cost of doing that
elimination would have a substantial drag on linking performance. (For
Janus/Ada, it probably would lengthen the linking times by 4-16 times, as
code elimination is quadratic in number of subprograms in the program.)

Another alternative would to use a "fat pointer" for access to subprograms
that would be self-describing. That is, it would contain information about
needed extra parameters and their values. However, such a thing would not be
compatible with "C", so the problem would reappear with Convention (C)

Certainly, we could get rid of the rule by disallowing the use of 'Access on
subprograms that come from generic instances, but that is a far worse wart
on the language than the current rule. We want subprograms in generics to be
as 'normal' as possible.

Janus/Ada uses universal sharing (meaning all generics are shared), but the
problem occurs for any sort of sharing. Certainly, if full program analysis
is needed before any generic sharing can occur, then there will be no
generic sharing.

The 'dead body' part comes from my reaction to any proposal to eliminate
generic sharing (which your suggestion is, even if you don't know it). This
is essentially the same as eliminating Janus/Ada as an Ada compiler, since
the cost to change from universal generic sharing to template generics is
prohibitive. Moreover, it would have little value to the users of Janus/Ada
(they would get a whole new set of bugs to deal with, and their programs
would have very different size and performance characteristics).

The best solution to your problem is a coding standard requiring a single
'Access for each subprogram in your program. To be effective, this would
have to be checked with a tool. As a practical matter, you won't see any
wrappers with library-level subprograms that aren't in generic units, so the
pattern usually works. (Which may be the most dangerous thing in this case.)

But the *really* best solution is an OOP-based solution. This is extensible
and type-safe and doesn't require unsafe compares and doesn't even need
(visible) access types. Admittedly, there is a bit more work up front, but I
think access types should avoided unless there is really dynamic allocation
(which can't happen with subprograms in Ada). One of the really nice things
about Ada 95 is that you can do useful OOP without using any access types.

              Randy Brukardt.