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.

Lazy I/O strikes again.

This restricts the solver to a closed set of packages.

Previously, even with freeze files, there was no good way to instruct
cabal-install to avoid pulling in extra, unconstrained
packages. `--reject-unconstrained-dependencies=all` forces the solver
to stay within the set of packages that are either local packages,
explicit goals, or which are explicitly constrained in configuration
files or via the `--constraint` flag

Closes #2568.

This was failing on Mac because of clang preprocessor warnings.

In 7d39aa01, the severity of the glob
missing directory warning was downgraded, but the test wasn't updated.
Not sure why it was passing before.

/cc @quasicomputational

(cherry picked from commit 74b91213)

kristenk authored 6 years ago and Mikhail Glushenkov committed 6 years ago

new-build had assumed that each component could only directly depend on one
version of each package.  However, a component can depend on two different
versions of a package by listing it as a build-depends and build-tool-depends
dependency.  The incorrect assumption caused new-build to mix up the two
instances of the dependency when converting the solver's package-based install
plan (SolverInstallPlan) to the component-based install plan
(ElaboratedInstallPlan).  This commit fixes the bug by keeping build-depends and
build-tool-depends dependencies separate between reading them from the
SolverPackage and passing them to
D.Backpack.ConfiguredComponent.toConfiguredComponent.

This commit changes the signature of
D.Backpack.ConfiguredComponent.toConfiguredComponent to allow distinguishing
between the two types of dependencies.

quasicomputational authored 6 years ago and Mikhail Glushenkov committed 6 years ago

Now the behaviour doesn't switch on just --project-file's
presence.

quasicomputational authored 6 years ago and Mikhail Glushenkov committed 6 years ago

This doesn't get anywhere near the improvements suggested in #4831,
but it's a very respectable improvement over the status-quo for not
much effort.

VCS tests are *very* slow; it took a substantial amount of work (this is a squash of 35 commits by two authors with debugging help from @Mistuke as well) to get the test suite and build process fast enough that they can run along with the rest of it in under an hour (the Appveyor limit).

Co-authored-by: quasicomputational <quasicomputational@gmail.com>

quasicomputational authored 6 years ago and Alexis Williams committed 6 years ago

This has been a problem since #5372 began expanding globs in `cabal
check`. Now the logic of running a glob is separated from the parsing,
giving the caller the opportunity to handle parsing failures flexibly.

quasicomputational authored 6 years ago and Alexis Williams committed 6 years ago

These globs make `haddock`, `sdist` and `install` die, so they
definitely ought to be warned about!

Because of the dying behaviour, I made the checks dist-inexcusable;
starting from a clean slate I would probably have only made them
suspicious, but this isn't terrible.

quasicomputational authored 6 years ago and Alexis Williams committed 6 years ago

This also significantly improves the error when trying to refer to
missing directories, hopefully making it clear that it's coming from
Cabal. #5318 and snowleopard/hadrian#634 are two bugs which manifested
as Cabal trying to glob in a non-existent directory and both took some
debugging because of the obscurity of the error.

Previously, we were checking the package with a hard-coded root
directory of ".". This was not a problem before, but with #5372 we
have started to expand globs while checking packages, which breaks if
the CWD is not the directory containing the `.cabal` file and causes
snowleopard/hadrian#634.

Luckily, this is an easy fix: the correct directory is easy to
determine. Writing a test and making sure it's tickling the failing
case took longer than writing the fix!

"." is hard-coded as the root directory passed to `checkPackageFiles`
in a few other places, but those are (a) non-trivial to test, and (b)
already in places that have other assumptions about their CWD, so I
have simply documented the CWD requirement for those.

quasicomputational authored 6 years ago and Alexis Williams committed 6 years ago

The code does the right thing, but it's slightly fragile since it
depends on directory switching, so make sure it stays right.

quasicomputational authored 6 years ago and Alexis Williams committed 6 years ago

This one wasn't actually testing what it said it was testing. Oops.

quasicomputational authored 6 years ago and Alexis Williams committed 6 years ago

They're not all passing, but I think that that's a bug in the code
rather than the tests.

On all platforms, warn about their presence. On Windows, we should use
slashes (as opposed to backslashes) where possible, to avoid causing
things like #5386.

Closes #5386.

This has the effect of allowing a glob `*.html` to match the file
`foo.en.html`. For compatibility, this is only allowed with
`cabal-version: 3.0` or later; for earlier spec versions, a warning
will be generated by `cabal check` if there are files affected by this
change in behaviour.

Fixes #5057. Fixes #784. Closes #5061.