Using sourceforge gives release binaries which don't require us to
regenerate all of the autotools scripts. This removes the need for
dependencies like cppunit and libgcrypt and autoreconfHook.
cc @geerds
Currently it errors out at build time with:
/nix/store/HASH-cudatoolkit-6.5.19/usr_include/host_config.h:82:2:
error: #error -- unsupported GNU version! gcc 4.9 and up are not supported!
Instead of downgrading gcc to 4.8, this patch upgrades cuda to 7.0, which
I think is the better choice. (Cuda 7 dropped support for some older graphics
cards, but gained support for newer ones.)
* Build most of the stuff on /tmp, not in /nix/store.
* Generate hoogle database for all the dependencies.
* Generate haddock index and contents files.
* Cleanup.
Fixes CVE-2014-8121, CVE-2015-1781 and two unnumbered problems (apparently).
All these commits should be contained in the 2.22 release,
but we don't want that yet due to unresolved locale incompatibilites.
`man 1 info` says:
The first non-option argument, if present, is the menu entry to
start from; it is searched for in all `dir' files along INFOPATH.
If it is not present, info merges all `dir' files and shows the
result. Any remaining arguments are treated as the names of menu
items relative to the initial node visited.
Which means that this does what previous programs/info did and #8519
(on-the-fly infodir generation for Emacs) wanted to do, but for both
programs.
* Upstream switched to cmake => remove autoreconfHook
* Makefile.am is gone => remove darwin specific patch (untested)
The new version has a pkg-config file. Yay!
Tested locally with `nox-review wip`; no regressions found.
This is needed for cuda 6.5 or else nsight will not find them. In cuda 7.0 the
samples are already in $out/samples. Older versions do not build due to broken
dependency (python-minimal).
