Re: Untyped "constant"?

[bywater!arnor!watson!prener.watson.ibm.com!ncohen@uunet.uu.net (Norman H . Cohen)]

In article <1991Sep23.164446.13412@aplcen.apl.jhu.edu>,
gralia@ddsdx2.jhuapl.edu (Mars J. Gralia) writes:

|> a) Suppose I have a compiler/RTE in which integer'last is 32_767.
|> b) And suppose the system has a long_integer type in which
|>    long_integer'last = 2_147_483_647.
|> c) And suppose I have the program:
|>           procedure Named1 is
|>              X : constant := integer'last+5;
|>           begin
|>               for j in 1..X loop
|>                   null;
|>               end loop;
|>           end Named1;
|>
|> d) Finally, suppose the "loop" variables are really used. E.g., by
|>    disabling optimization when I invoke the compiler.
|>
|>
|> Then:  is this a legal program or not?

As Dave Collard points out, Integer'Last+5 is of type Integer, and the
initial-value expression for a named number must be of type
universal_integer, so this is not legal.  However, we can get around this
technicality by rewriting the initial-value expression as

   Integer'Pos(Integer'Last)+5

since the 'Pos attribute yields a universal_integer result.

However, RM 3.6.1(2), when deciphered, states that a discrete range in
which each bound is either a named number, an integer literal, or an
attribute with a universal_integer value has each bound implicitly
converted to type Integer.  Thus 1 .. X is equivalent to
Integer(1) .. Integer(X).  Since X has a value greater than Integer'Last,
the conversion in the upper bound raises Constraint_Error.

|> Some compilers, admittedly archaic or for very small computers,
|> will raise a Constraint_Error, thereby forcing me to say:
|>              X : constant long_integer := integer'last+5;
|>
|> I would prefer the program given above, because it would appear to be
|> more portable.

Well, it's portable in the sense that it is guaranteed to raise
Constraint_Error (except when the loop is optimized away, in which case
the implicit conversion that would have raised Constraint_Error can be
eliminated pursuant to RM 11.6(7)).

A more useful form of portability would come from defining your own
integer type with the required range:

     procedure Named1 is
         type Loop_Index_Type is range 1 .. Integer'Last+5;
     begin
         for J in Loop_Index_Type loop
             null;
         end loop;
     end Named1;

Implementations for which Integer'Last = System.Max_Int will reject this
at compile time.  Implementations for which System.Max_Int >=
Integer'Last+5 are guaranteed to execute without an exception.