On a tangentially related point, portability of several of the existing standard hierarchical libraries is rather patchy too. For instance, the main reason that nhc98 does not currently ship with System.Time from the base package, or any of the network or unix packages, is that the checked-in implementations have a tool-dependency on hsc2hs. (The latter does not appear to have a stand-alone distribution independent of ghc, so it would be necessary to incorporate it into the nhc98 tree, and I don't currently have the time available to untangle hsc2hs's configuration and build system and re-tangle it for nhc98.)
I guess I'm with Sigbjorn on wanting to reduce the number of tool dependencies, even though I happen to like GreenCard and hsc2hs for what they can automate. Or perhaps, rather than reducing the dependencies, I want to reduce their impact, by making the tools much easier to get hold of, configure, and install.
Splitting off hsc2hs into a standalone distribution wouldn't be too difficult, and is definitely a worthwhile task if anyone would like to take it on. Looking at the sources, it has some #ifdef __NHC__ so it looks like it worked on nhc98 at some point. It's a single module Haskell program with an auxiliary file (template-hsc.h). It needs to know one or two things about the system: what the C compiler is called, or where GHC can be found. That's about it, as far as I can tell. Cheers, Simon