Hi
1. It IS available on Windows.
Not without Mingw/Cygwin - which in my mind makes it not Windows native. I also know that the release is made without testing on Windows, and that certain related tools like hmake rely on shell scripts. If there is someone using nhc seriously on Windows, it would be good if they documented exactly what they needed to do, and ideally supply a binary for the rest of the world.
2. Since it is interpreted, you should compare it with GHCi, not with compiled programs. The comparison is NOT BAD.
Really? Do you have a benchmark for that? (sadly nobench is down at the moment) Remember that GHCi does not interpret all code, but actually has massive amounts of it compiled (all the base libraries etc) Also for most users this is an irrelevant distinction. The language shootout does not benchmark Java against GHCi because both use bytecode - they benchmark the speed the user sees at the end.
3. Premises correct, conclusion speculative. Probable, but when you *accuse* somebody, then *prove* it.
To take a recent commit message: * Comment out incorrect kind inference. In fact, there is no kind inference at all - just an assignment of kinds to type variables, which turns out to be wrong of course. There are plenty of bugs in nhc. You can't implement the Haskell 98 spec without kind inference, yet nhc (and Yhc) both lack it entirely. There were about 3 changes required to XMonad to make one single module compile with nhc. Not all of these bugs are show stoppers, and its perfectly possible to use it without running into a bug, but there are more out there, and its more likely you'll run across one.
(who has nothing to do with nhc, but who respects enormously this effort)
I also respect this effort as well! We need more compilers, and not just toy compilers, but production compilers! Thanks Neil