In order to build the package databases that we will use when compiling
a Haskell package, we iterate over the relevant dependencies, and if
they contain a package db, we copy its contents over.
So far so good, except when one of those dependencies is GHC. This
doesn't happen ordinarily, but it will happen when we construct the
package database for compiling `Setup.hs`. This is compiled for the
build architecture, so we get the build deps, including both the native
and the cross GHC (if there is one).
In this case, we end up copying the packages from the GHC's package
database. This is at best unnecessary, since we will get those packages
from the GHC when we compile with it.
At worst, however, this is semantically questionable. We can end up
having multiple copies of e.g. Cabal with the same version, but
(potentially) different contents. At the moment, GHC will expose one of
these at semi-random depending on which one it looks at "first".
However, there is a MR open [in
GHC](https://gitlab.haskell.org/ghc/ghc/merge_requests/545) which as a
side effect will instead expose both, leading to ambiguous module
warnings (which is not unreasonable, since it *is* ambiguous).
So what can we do about it? The simplest solution is just to not copy
the package databases from GHC. GHC is special in this regard, so I
think it's okay to treat it specially.
This PR should have no effect on anything now, but will prevent any
breakage when/if the GHC patch lands.
Closes https://github.com/NixOS/nixpkgs/pull/57706.
`all-functions` corresponds to `-fprof-auto` which places an SCC on every binding. It is well known that SCCs hinder GHC from doing its optimization magic and really slows down profiled code to a point where the profiling reports are completely skewed towards things that were completely optimized away in production settings. Concretely this shows up with things like lenses which do not carry runtime overhead when properly simplified.
`exported-functions` corresponds to GHCs `-fprof-auto-exported` which doesn't put SCCs on `INLINE`d code and in turn doesn't influence simplification of this basic but important stuff.
This commit is not really correct. The `hasActiveLibrary` check is wrong.
We can have an active library even if we do not ask for a static lirbary or
dynamic one; we can still have just a set of objet files and archives.
If empty directory isn't deleted, referer depenedencies will
fail with:
cp: missing destination file operand after '/tmp/nix-build-cabal-helper-0.8.0.2.drv-0/setup-package.conf.d/'
This is currently only the case for cabal-install, as cabal2nix
doesn't handle well buildable=False flags due to long-standing bugs
in Cabal itself.
This was reverted in 6e07a3a19a with the
other commits in the PR, as the number of overrides added in that PR was deemed
excessive. But this commit adds no overrides, and is valuable on its
own, so I am adding it back.
(cherry picked from commit 359e0ce4bb)
If we use a --docdir that's not specific to the package, then different builds
will install their license files into the same location, which leads to file
collisions if those are ever joined into the same environment.
Fixes https://github.com/NixOS/nixpkgs/issues/35024.
enableShared in generic-builder.nix should default to what the GHC
compiler was compiled with. Add a passthru to all of the GHC compilers
to hold the value of enableShared. If enableShared is not set in the
GHC we just use false as the default value for enableSharedLibraries.
Note: I may have missed some compilers. Only GHC & GHCJS are covered
by this commit but this shouldn’t break evaluation of anything else.
This decreases complexity and ensures setup dependencies are properly
specified with `setup-depends` as they should be. Testing will say if
this is a reasonable change.
The reason why this does not work is not that we can't built static
objects, we can, but we can't use `-staticlib` on GHC on windows.
`-staticlib` rolls all dependencies into a combined archive. While this
would work on windows if we used gnu ar and MRI script, GHC can't rely
on GNU ar, and as such has a quick archive concatenation module for GNU
and BSD archives only.
This is because they are just for Setup.hs, so they are just used at build time
and completely isolated from the normal components' dependencies.
This was previous implemented in 8a8f0408cd, but
reverted in e69c7f5641 because it broken
setup-depends non-cross in haskell shell environments (custom Setup.hs in cross
shell environments has never worked). This version adds a special native
exception to avoid that breakage.
Just like with the other `--extra-*` flags, cc/ld-wrapper already handles
this, but we need to make Cabal aware so that the haskell builds have
the correct metadata.
