Many thanks to Darrin Chandler and dcoppa@ for testing, reporting about
broken stuff, missing dependencies here and in ports depending on ghc.
Notes and rants:
- Bootstrapping is done using precompiled binaries, since .hc
bootstrapping still doesn't work. I really hate this.
THIS MEANS THAT GHC IS NOW AND WILL STAY LEGACY-ONLY (i386 and amd64)
At least until someone fixes it. I tried for more than two year
(well, only in my spare time and during my vacations) and failed.
- libgmp is currently disabled, because I didn't yet hack the GHC build
system to use the system libgmp instead of the patched one included
in GHC.
- The haddock ncluded in the ghc distfile is replaced by the version
of haddock found in devel/haddock. Haddock itself is @commented
in the ghc PLIST. Unfortunately, this needs an ugly hack that
introduces an otherwise useless pseudo flavor `no_deps' in
devel/haddock.
- CLDouble has been removed from GHC some time ago, because it was
an alias for double (AFAIK there's now support for long double
in GHC). As this isn't a really big problem, it currently breaks
c2hs, which I'll mark broken temporarily before committing the
ghc update.
- The external codeset defaults to latin1 (suggested by Simon Marlow)
and can be overridden by setting the HS_ENCODING to any codeset
supported by libiconv.
- ghc.port.mk still needs some love, especially for letting a port add
additional parameters to certain invocations of ${MODGHC_SETUP_PROG}.
get process informations. Remove the mkbundle.cs patch and replace it
with one that uses configure.in so the libgc dependency ends up in the
pkg-config file. Move disabled features to configure.in from our Makefile
Zoltan Varga from the mono team helped me to debug the problem we
were having. So the correct problem description is that on OpenBSD,
ctx can be NULL if we are interrupting poll().
The upstream diff also fixes the issue where we are interrupting
unmanaged code.
it up before having my morning coffee.
So it turns out that sometimes the context on openbsd can be 0. This will
lead to problems. I suspect that this is because of our fantastic pthread
library. With the following diff we just skip doing some JIT stuff if the
context is 0 because the thread probably exited before.
Anyways this way everything *seems* to work fine and almost all regression
tests are passing now except for 2 out of 380.
Pepole understanding pthreads are welcome to look at it and come up with
something that is not a tripe XXX hack.
so work around by tricking it to write into WRKINST through a symlink.
keep this here, so that affected ports need minimal modifications, and we
can work out a better fix long term.
Agreed with wcmaier...
by jasper@'s fixes to the libffi port.
Mark mips64 (aka, sgi) as BROKEN, as the build fails. We'll look closer
at this after unlock.
OK jasper@
ports is unlocked for a while only for those who have been informed.
breaking cd /usr/ports && SUBDIR=some/path make something for
category makefiles. While there, also put spaces around += uniformously.
okay naddy@, jasper@
