"Manuel M. T. Chakravarty" wrote:
>
> George Russell <[EMAIL PROTECTED]> wrote,
> This sounds interesting. So you want the branch that is not
> taken to be syntactically correct, but it need not type
> check. How about other semantic constraints (visibility of
> names etc)? If you want to use this mechanism for selecting
> whether or not to use non-standard features, you probably
> also have to allow that the branch not taken is
> syntactically illegal (eg, use of GHC's unboxed values in
> one branch and plain Haskell in the other). How is this
> exactly in Java?
I suspect that in Java both branches would have to be syntactically
correct, though I can't be bothered to look it up. Hmm, I hadn't
thought of that. But it shouldn't be hard to think up some way
round which is prettier than the CPP version; it could hardly be
uglier!
If we are going to have a different name for "if" when it is
used in this way (and I think we should anyway) we may as well
call it #if and be like C. This will mean that existing programs
should go through without much modification. So I suggest
(1) The compiler keeps track of which values at top level are
"constant expressions". A constant expression is an
Integer, Bool or String, or something obtained thereof
by the obvious operations (arithmetic, comparison,
perhaps if itself, ++). When a value bound to a constant
expression is exported, its constant expression status
is exported too.
(2) the CPP conditional syntax (#if, #else, and so on)
but NOTHING ELSE from CPP. The argument to #if must
be a constant expression.
(3) Where appropriate the compiler may allow values to be
bound to constant expressions on the command line.
Yes I know it's pretty ghastly, but it's still 10000 times better
than using CPP. And I don't think it should be too ghastly to
implement. The main bother would that the change would have to
be agreed precisely between the Hugs and GHC folk.
