I am therefore deeply reluctant to provide both GHC-for-mingw32 and GHC-for-cygwin. One build on Win32 is enough! We ended up with a mingw32 basis because it meant we could make GHC completely self-contained -- no dependence on cygwin1.dll etc.
Some comments from a binary-only GHC user who tends to depend a lot on cygwin while using windows:
This was *huge* step forward: GHC installs and runs with no problem on Windows now.
This *is* a huge step forward, and together with GHCi and HOpenGL, has even tempted me to use GHC a bit more:-) Having to wrestle with two different GHC installations on one platform would seem a step backward. So, I'd not only agree with one build only (GHC should be compilable under cygwin, though, for those who absolutely need to go the other way, or who want to track the CVS version), I'd like one build to work both in native and in cygwin mode. As a Haskell user, I'm interested in: - a standalone GHC, producing standalone executables and dlls, with good FFI interfaces to the non-Haskell world - portability across platforms, with as few code changes or restrictions as possible My approach to keeping windows/unix differences small is mostly based on cygwin, so I need to be able to use GHC and it's executables under cygwin, as I would use it under unix, in combination with other (windows/ cygwin) software. That doesn't mean that GHC-generated executables or libraries need to be cygwin-dependent, and cygwin is, by design, able to use windows executables (mixing of libraries is probably another story).
1. GHC does not understand cygwin paths in the file names passed to it on the command line.
Making GHC understand cygwin paths makes software more system dependent, not more portable. And what about the executables produced by GHC? Most of the cygwin path problems could, in theory, be solved without changing GHC, but with a lot of accumulated UNIX makefiles, that can be unpractical. As far as I understand, GHC can cope with both relative unix-style and relative and absolute windows-styles paths, so the remaining problem I tend to encounter are "absolute" unix-style paths which are really relative to the cygwin root directory. (I also seem to recall someone mentioning problems related to GHC passing "normalised" paths to other tools, but if GHC uses it's own toolchain, that seems unlikely?) My suggestion would be a --prefix <path> option, or GHC_PATH_PREFIX variable for GHC-produced executables (including GHC itself), telling them that any absolute, unix-style paths are to be interpreted relative to <path> (e.g., in cygwin makefiles, default installation, HC=ghc --prefix c:/cygwin). That wouldn't be platform-specific and might also come in handy for other purposes.
2. GHC on Win32 does not come with a Posix library. If we used a Cygwin basis, Posix would be easy because cygwin does all the hard work.
That looks like a real bugger to me as it impacts on portability of Haskell programs. Going from incomplete posix support to even less posix support was a step backwards.
3. I/O on Win32 is *blocking*. A blocking input operation freezes all the other Haskell threads.
No experience with that one, but in general, establishing consistent I/O behaviour across platforms would be a very useful asset.
Are there any other problems?
perhaps: 4. File I/O on windows differs from I/O in unix (locking of files instead of implicit maintainance of hidden handles, I think??). cygwin tries to smooth things over, but fails for more complex cases (open a file for reading, remove it, open it for writing, copy from read handle to write handle).We just traced a problem in building nhc on cygwin down to that one.. How does mingw fare in that respect? Better? Or even worse? 5. I assume that GHC and it's executables interface rather well with the windows world. What about interfacing to software ported from unix that depends on cygwin, though?
1. GHC already fudges filenames to take account of the Win32/Unix conventions. We could add more fudges, to change /cygwin/c/foo to c:\foo, for example. Perhaps controlled by a -cygwin flag to tell GHC-for-win32 whether to use cygwin fudges or not. Heuristic, yes; but might solve the problem for 99% of customers.
No heuristic, please, just some more flexibility for makefile authors (/cygwin/c/foo tends to be /cygdrive/c/foo these days, and sometimes is //c/foo, but c:/foo tends to work as well - do you want to track cygwin's mount table in GHC?-).
2. Mingw32 provides quite a lot of Posix, so if someone was prepared to put in a bit of work we could get a good part of the Posix library available on Win32, still via mingw. Any volunteers?
I step back to make space for the volunteers, but mingw being minimalistic, what are the chances that it's posix support is a suitable alternative to cygwin's, which isn't complete, either?
3. Non-blocking I/O is a soluble problem: Win32 provides suitable primitives. But they are different to the Unix primitives, so there is work to do in the runtime system to make it work. This is harder for a volunteer to do because it's in the runtime system, but not impossible.
As far as I know, cygwin (the *larger* cousin of mingw) is still waiting for someone to contribute solutions to I/O problems on top of windows.. Thanks for raising this issue, Claus PS.
The GHC core team is now down to Simon M and me. Sigbjorn heroically helps out on Win32 stuff, but it isn't his job. So we have strictly limited effort available.
[I have to admit that I'm a bit concerned about the overall number of people involved in Haskell implementations. Fortunately, the licenses are very open, and fortunately, we do have those heroic volunteers, but the Haskell implementation core team isn't going to be much (if any) bigger than the GHC core team soon. Given the interests of some (well, at least one;) companies, and the volunteers working for those, perhaps they could acknowledge the work done by those volunteers as being essential for their business? And given the wide-spread use in academia as a research and teaching platform, does anyone know of a way of securing some kind of international platform funding, to keep the current implementations alive and to develop them further (as an academic complement to Galois' (?) and MS' contributions in terms of man-power)?]