Building GHC 5.02 on Solaris x86

[Simon Marlow]

> Simon Marlow <simonmar at microsoft.com> writes:
> >
> > Having just looked at the code, it seems we could recover the
> > platform-independentness in the I/O library with just a 
> small amount of
> > wrapperage: most of the offending code is in PrelPosix.hsc, 
> with a few
> > #const's scattered around PrelHandle and PrelIO, and none of the
> > affected areas are particularly performance-critical.
> >
> > Also needing wrappers are Time, CPUTime, Directory & 
> System.  I'd say
> > stop there, though - this is purely to make bootstrapping 
> on a new OS
> > easier, so we only need to apply the no-tool restriction to 
> libraries
> > needed by the compiler.
> >
> 
> Good, let's sort this one out. ghc/compiler/ has its fingers 
> deep in the pie
> that is hslibs/, so presumably the restriction applies there also?

I would extend it as far as necessary in order to satisfy the
dependencies of a non-GHCi compiler.  That means only converting IOExts
from hslibs, I think.  net isn't needed, and neither is Readline.  I'll
have to be careful with Posix when I rewrite it.
 
> I'll happily convert some of these files (once I know what 
> wrapper story you have in mind); let me know.

That'd be great.

PrelPosix.fdFileSize is a tricky one, because we can't return an Integer
from C land and we can't return an off_t because the size of that is
OS-dependent.  Probably we'll have to return an Int64 even though that's
quite ugly.  The other functions in PrelPosix shouldn't present too much
of a problem - I imagine most of PrelPosix can be moved into C land
leaving a small Haskell wrapper.

For the #const's, we'll have to do something like this:

  something.c 
	HsInt get_BUFSIZ(void) { return BUFSIZ; }

  PrelHandle.hs: 
	bUFSIZ = unsafePerformIO get_BUFSIZ
	foreign import get_BUFSIZ :: IO Int

Directory needs some #const wrappers, and it can get isDirectory from
the I/O library.  System is already wrappered.  Time will be ugly, but
then it already is.  The CPUTime wrapper is probably going to have to
use Int64.

Cheers,
	Simon