* Stackage builds packages against fixed versions for lots of other
  packages, so some packages find weird or no build plans if the version
  constraints are not right.
* There are extra C library dependencies for the necessary libraries yet
* Some packages don't have a library component

Towards #72

Teo Camarasu authored 2 years ago and Matthew Pickering committed 2 years ago

This is to work around an issue where hydra was trying to update the flake.lock to a newer version of nixpkgs.

Finley McIlwaine authored 2 years ago and Matthew Pickering committed 2 years ago

The patches primarily deal with mtl-2.3.1 issues, just making consumers
of the package buildable. These were necessary to get some components of
HLS building on GHC 9.6.

During this process it was also discovered that the `vector-space`
package was affected by GHC issue
ghc#22912 so I've added it to
`extra_packages` in CI.

This commit also includes a small change to the `patch-tool` that fixes
the unpack script in the case that a package which shares a prefix with
the package being upacked is already in the packages dir (e.g. `Chart`
and `Chart-diagrams`).

Teo Camarasu authored 2 years ago and Matthew Pickering committed 2 years ago

*  Replaces the Gitlab CI based caching with the cache.zw3rk.com nix binary cache
*  Replaces the niv based input pinning with flakes and flake-compat (to keep the interface the same)
*  Adds the build env to the flake so it can be cached by cache.zw3rk.com
*  Update gitlab job definitions to minimise calls to nix

Rodrigo Mesquita authored 2 years ago and Matthew Pickering committed 2 years ago

With:
* c2hs (build_tool_package)
* servant-conduit
* servant-machines
* freer-simple
* linear-generics

This adds patches for freer-simple and servant-*, but adds all 5 to the
CI extra packages

Closes ghc/head.hackage#69

The hope is that by limiting concurrent jobs/ CPU usage for the first
call to nix, we will cache the environment without exhausting system
resources.

This version properly builds against `recv-0.1.0`, thereby fixing #74.

Code partially adapted from https://github.com/haskell-servant/servant/pull/1627

Code adapted from https://github.com/kazu-yamamoto/logger/pull/205

Code adapted from https://github.com/jacobstanley/unix-compat/pull/60

Code adapted from https://github.com/haskell-streaming/streaming/pull/113

`base-4.18.0.0` introduces `Eq` and `Ord` instances for `Generically1`, which
clash with the ones defined in `generic-data`. Guard the instances in
`generic-data` with CPP to fix this.

On nightly pipelines there are more than 20 jobs so we need to increase
this limit to make sure we get all the jobs we care about.

Fixes GHC #22663

We were doing this very convoluted thing with staged pipelines and
scheduled jobs but it's much more direct to just push to the other
branch directly when there is a job which updates the master branch.

This pipeline is for updating a special branch called "upstream-testing" which is used
when we want to trigger jobs to run on head.hackage. This branch is not protected so
the permissions about who can trigger jobs on upstream-testing are much more liberal than
the master branch.

It turns out that before_script doesn't get combined. Think it's about
time to start generating this CI config as well.. like we do for the
main GHC repo.

It seems that despite the documentation mentioning nothing about it, you
can always save a cache even if a job fails. Which is good for us as we
are caching the nix stuff, which often works when the whole job fails.

See: https://gitlab.com/gitlab-org/gitlab/-/issues/18969

The other option here (which introduces some more complexity) is to
separate the building of the tools into a separate job which then each
runner will download and unpack.