Typeof operator in C (Re: An Interesting View of "Strong" Vs. "Weak" Typing)

[Karl Heuer]

Note: |typeof(x)| already exists, with that name, in gcc; it is syntactically
a type-name.  To avoid confusion, I'll refer to the original concept (an
operator that returns a value that somehow encodes the type information) as
|typecode(x)|.

>If it returns some integer, what does that integer mean, and what would you
>do with it?

My private library used to contain a function |void *alloc2(size_t, size_t)|
which would return a heap pointer that could be used like a two-dimensional
array.  The idea was that (int **)alloc2(m, n*sizeof(int)) would allocate and
return a single block of size m*sizeof(int *)+SLOP+m*n*sizeof(int), with the
first portion initialized to point into the last portion.  Then normal
subscripting |a[i][j]| would work, and the whole mess could be returned to the
freelist with a single call to |free()|.

I stopped using this function when I realized that it could not be implemented
on certain architectures.  (I'm willing to use library functions that cannot
be implemented portably, provided that they can be implemented nonportably on
any individual machine.  But this one can't be implemented *at all*, since
there exist two distinct results that take the same input.)  So instead I
started writing the necessary expansion inline, and bemoaned the fact that
this could never become a Standard Library function.

Or could it?  Suppose ANSI had mandated that a conforming implementation must
provide a macro |TYPE **malloc2d(size_t nr, size_t nc, TYPE)|.  Then a vaxlike
architecture simply has this expand into |(TYPE **)alloc2(nr,nc*sizeof(TYPE))|
where |alloc2()| is the function I described above.  Would this mean that
word-based architectures would be unable to implement ANSI C?  No, all that
they'd have to do is invent a minimal form of the |typecode()| operator, and
use |(TYPE **)alloc2x(nr,nc*sizeof(TYPE),typecode(TYPE *))| as the expansion,
where |alloc2x()| is like the impossible |alloc2()| except that it looks at
the third argument to decide whether to use char pointers or word pointers
when initializing the dope vectors.

Note that it's not necessary for |typecode| to actually distinguish among the
infinitely many types that can be constructed in C; all that is required *for
this application* is that any two pointer types with "different internal
representations" be distinguished.  On any particular architecture there are a
small number of these.

Let's return to the question.

>If it returns some integer, what does that integer mean, and what would you
>do with it?

If we were to agree that |typecode| is useful enough to actually require it
(in my example it was a "magic builtin" that the user would never need to know
about), I would say "implementation dependent".  More precisely, it should
return a value of type |typecode_t|, and it should be possible to compare two
such values for equality.  No other operations on |typecode_t| are necessary
or desirable.

Karl W. Z. Heuer (karl at haddock.isc.com or ima!haddock!karl), The Walking Lint