Hydra generates a GHC closure for Darwin that for no apparent reason
contains an ancient, broken Haddock binary -- probably because of an
impurity in the build system. That bug makes those GHC binaries
unusable: <https://github.com/NixOS/nixpkgs/issues/2689>.
rts/StgCRun.c: In function 'StgRunIsImplementedInAssembler':
rts/StgCRun.c:208:1:
error: frame pointer required, but reserved
StgRunIsImplementedInAssembler(void)
^
In file included from includes/Stg.h:230:0:
0,
from rts/StgCRun.c:70:
includes/stg/Regs.h:323:20:
note: for 'Sp'
GLOBAL_REG_DECL(P_,Sp,REG_Sp)
^
includes/stg/Regs.h:174:54:
note: in definition of macro 'GLOBAL_REG_DECL'
#define GLOBAL_REG_DECL(type,name,reg) register type name REG(reg);
^
In this case, we also need to specify compilation flags to mark stacks as
non-executable, otherwise PaX will not allow ghc or binaries built by ghc
to run. This is what gentoo-hardened does as well.
It sucks, I know, but GHC just doesn't compile reliably when built with
some -j<n> option. :-( We have build errors because of apparent race
conditions all over the place on Hydra. This causes so much trouble for
users that it's not worth keeping this option enabled, IMHO.
1) The wrapper erroneously used the ghc-pkg flag "--package-db" instead of
"--global-package-db". The result was that packages installed locally in
~/.ghc and ~/.cabal were invisible to GHC. This has been fixed.
2) The wrapper now deals gracefully with an empty package set: if no package
is requested to be included in the wrapped environment, the wrapper just
installs a pristine GHC.
3) Correctly configure the "docdir" path returned by ghc-paths.
4) Added some comments that describe the idea behind our ghc-paths patches and
gives users same sample shell code that can be used to import our special
environment variables into the currently running shell, so that programs
outside of the wrapped environment can use them, too.
The ghcWithPackage expression now has an argument 'ignoreCollisions' that
allows users to disable the path collision check like so:
(pkgs.haskellPackages.ghcWithPackages (pkgs: with pkgs; [ haskellPlatform ])).override { ignoreCollisions = true; };
See d64917ad17
for a long and detailed discussion of why these path collisions may occur.
Haskell packages -- i.e. packages built by our Cabal builder -- invariably have
the attributes 'pname' and 'version'. We use the absence of these attributes to
recognize non-Haskell packages and filter them from the closed package set
generated by closePropagation. We do this so that the generated Haskell
environment won't contain paths like "/lib/libz.a", which are part of the
closure but have nothing to do with Haskell.
The previous scheme used the attribute 'ghc' to accomplish the same thing, but
unfortunately other packages to contain a 'ghc' attribute, too, like the
old-style ghc-wrapper. Including the ghc-wrapper in this environment is
pointless, obviously. The new approach filters the ghc-wrapper successfully.
* There now is full support for building Haskell packages as shared libraries
for GHC versions 7.4.2 or later. The Cabal builder recognizes the following
attributes:
- enableSharedLibraries configures Cabal to build of shared libraries in
addition to static ones. This option requires that all dependencies of
the package have been compiled for use in shared libraries, too.
- enableSharedExecutables configures Cabal to prefer shared libraries when
linking executables.
The default values for these attributes are arguments to the haskellPackages
expression.
* Haskell builds now run in a LANG="en_US.UTF-8" environment to avoid plenty
of build and test suite errors. Without this setting, GHC seems unable to
deal with the UTF-8 character encoding that's generally considered standard
in the Haskell world.
* The Cabal builder supports a new attribute 'testTarget' to specify the exact
set of tests to be run during the check phase.
* The ghc-wrapper attribute ghcVersion has been removed. Instead, we use the
ghc.version attribute, which exists in unwrapped GHC derivations, too.
The change was supposed to trigger a re-build to fix a broken GHC binary
on the Hydra build farm, but now it turns out that the cause for the
errors we're seeing isn't GHC: all kinds of (non-Haskell) packages are
broken.
