GitLab

Clearer code and a wider interface to give more control.

so we do not accidentally inherit the default flags for the underlying
configure command when we in fact want the empty flags. I am beginning
to suspect that having any defaults at all was a mistake.

We do not want to pass a load of default command line flags. We only want
to pass flags where we're overriding the defaults.
Also it's because old Cabal versions do not understand new flags.

Make it usable for callers that want to check
properties before constructing an InstallPlan.

This is joint work with Kolmodin.
An InstallPlan records what we want to install and records the outcome
of installing packages. It has a fairly strict internal invariant which
should guarantee that each package can be configured sucessfully.

The install --only doesn't do any planning or build
reporting so make it completely separate so we do
not have to make up fake build plans or results.

Rather than using global 'constants' for the current os and arch.

Externally we currently use a DepGraph, though we'll move to a InstallPlan.
This is a step towards making it easier to swap in different package
dependency resolution algorithms. Because we're hiding ResolvedDependency
we have to disable the extended Hackage.Info stuff for the moment.

Mostly just changing showVersion, showPackageId to display instead

The baseDir was passed instead of the full path.

This patch takes advantage of the new OptionField structure in Distribution.Simple.Command
to provide support for all the configure command line options in the .cabal/config file.
This has a global effect on all the packages managed by cabal-install.

The ticket also mentions support for per-package sections in the config file. This patch
does not take care of that.

When installing a bunch of package we have to catch exceptions since
we carry on building other packages that did not depend on the
failing package. We were recording what phase the failure was in but
not the actual exception. We now record that too and print it along
with the more general explanation of what package failed and in
which phase. It's not perfect, eg when a package fails to compile we
end up printing that the exception was "ExitFailure 1" which is not
very useful.

We always know the base path for construction or extraction so don't
bother using Maybe FilePath. Also use GZip qualified.

The current directory is a global variable, we should not mutate it.
So instead, pass a base and relative path when generating tar entries.
Also change sanitizePath to be pure and use FilePath.Poisx.

Use lazy bytestring when packing tar entry headers rather than strict and
then making a single chunk lazy bytestring later. The lazy bytestring pack
will only generate a single chunk for a String that short (<4k).
Use openBinaryFile rather than openFile + hSetBinaryMode.
Add a haddock module header with copyright note.

we don't call setup sdist anymore but we use functions from Distribution.Simple.SrcDist

That is, allow quoted and unquoted paths.

It also adds the internal flag --only so we can avoid building a setup when build-type == Simple.

+1 for using the same version of the ReadP lib everywhere,
previously we would only have noticed this type error with nhc98.

It tells you what to do if you don't supply any args, rather than doing
nothing and it check for simple mistakes and gives helpful error messages.