GitLab

Create a new syntax for depending on any library of any package.
The syntax is

    build-depends: pkgname:{pkgname, sublibname} -any

where the second `pkgname` specifies a dependency on the main unnamed
library.

Closes #4206.

This gives us more type safety when dealing with libraries only

Before, a 3-tuple was used

The old --dependency could only do --dependency=pkg=cid,
but with public sublibraries this will become insufficient.

Now there is the option to also specify a component name using
--dependency=pkg:component=cid

So we don't try to symlink failed packages

Fixes #5550 (the dead symlink part, not the error reporting (#5641))

It conflicted with the flag with the same name from configFlags

Fixes #5550

[ci skip]

Monolithic inplace packages have a directory for each exe, so we allow
multiple paths and return all the exe paths of the package.

Fixes #5104

Before, new-install could exit successfully while actually _not_ symlinking
anything.
Moreover, without an overwrite option, upgrading cabal was impossible
without moving links around.

* Error out on overwrite
* Overwrite if --force-overwrite is passed

Using windows/msys paths can contain both kinds of slashes.
This causes issues when we compare paths but don't normalize
them, leading us to assume they refer to different locations.

As the case of `cpphs-1.20.8` has shown in #5379, in cases when a
monolithic package gets installed due to its executable components
(i.e. `exe:cpphs`) into the store we *have* to register if there's a
buildable public library (i.e. `lib:cpphs`) that was built and installed
into the same store folder as otherwise this will cause build failures
once a target actually depends on `lib:cpphs`

This ought to fix #5379

For some reason repl invocation was skipped for indefinite components.
However, GHCi seems to handle them just fine.

Fixes #5619.

Using the Unix epoch is causing problems on Windows because it's
before MS-DOS's epoch of 1980-01-01. Picking a different, memorable
one ought to solve that.

Closes #5596.

Fixes #5469

This adds support to `cabal init` for selecting the Cabal spec
version (i.e. the `cabal-version:` field) and also takes care
of translating between pre-SPDX and SPDX license ids.

Does not warn if the existing symlink already points to the correct location

Partial solution to #5491

Lazy I/O strikes again.

- Disable cabal-install's haddock on CI
- Temporarily disable cabal check (`-main-is` in tests)
- Remove this is not a lib note
- Tell to run git update-index --assume-unchanged

    ... when changing to -dev or -monolithic configuration.

    Run --no-assume-unchanged unconditionally, when
    changing to the production build.
- Improve auto-gen note in cabal-install.cabal
- Try to build haddock (with GHC-8.4.3 only)

Previously, the solver represented an unknown package with an empty PChoice
node, because an unknown package is like a package with no versions.  Now that
--reject-unconstrained-dependencies also needs to reject whole packages, it is
simpler to handle both types of failures similarly, and use a failure node for
both.  Another advantage of using a failure node now is that the node can store
more information about why the package was rejected.  The explicit node also
simplifies generating the solver log.

This commit treats a failure node under a goal choice as a backjump in
D.S.Modular.Explore.exploreLog, which fixes #5502.

The case gave the same result as a combination of two other simpler cases.

This prevents the `--store-dir` value from being cached
and thus fixes #5504

We only error-out when an explicit `--(haddock-)quickjump` is specified on a version of
Haddock that doesn't support it.

However, if `--(haddock-)quickjump` was implied by `--(haddock-)for-hackage`, we should just
warn.

Fixes #5494.