Never consider flag choices as independent from their package.

This happened independently in a number of places, which was bad; and was about
to get worse with the base 3/4 thing.

The only problematic thing is that when we call `cabal clean` or `cabal
haddock` (and possibly others), _without_ first having called `configure`, we
attempt to build the setup script without calling the solver at all. This means
that if you do, say,

    cabal configure
    cabal clean
    cabal clean

for a package with a custom setup script that really needs setup dependencies
(for instance, because there are two versions of Cabal in the global package DB
and the setup script needs the _older_ one), then first call to `clean` will
succeed, but the second call will fail because we will try to build the setup
script without the solver and that will fail.

By chosing setup dependencies after regular dependencies we get more
opportunities for linking setup dependencies against regular dependencies.

(and, therefore, also to the modular solver's output)

Duncan Coutts authored 10 years ago and Edsko de Vries committed 9 years ago

This patch adds it to the package description types and to the parser.
There is a new custom setup section which contains the setup script's
dependencies. Also add some sanity checks.

Although we don't use the new setup dependency component anywhere yet, I've
replaced all uses of CD.flatDeps with CD.nonSetupDeps. This means that when we
do introduce the setup dependencies, all code in Cabal will still use all
dependencies except the setup dependencies, just like now. In other words,
using the setup dependencies in some places would be a conscious decision; the
default is that we leave the behaviour unchanged.

The crucial change in this commit is the change to PackageFixedDeps to return a
ComponentDeps structure, rather than a flat list of dependencies, as long with
corresponding changes in ConfiguredPackage and ReadyPackage to accomodate this.

We don't actually take _advantage_ of these more fine-grained dependencies yet;
any use of

    depends

is now a use of

   CD.flatDeps . depends

but we will :)

Note that I have not updated the top-down solver, so in the output of the
top-down solver we cheat and pretend that all dependencies are library
dependencies.

In this commit we modify the _output_ of the modular solver (CP, the modular's
solver internal version of ConfiguredPackage) to have fine-grained dependency.
This doesn't yet modify the rest of cabal-install, so once we translate from CP
to ConfiguredPackage we still lose the distinctions between different kinds of
dependencies; this will be the topic of the next commit.

In the modular solver (and elsewhere) we use Data.Graph to represent the
dependency graph (and the reverse dependency graph). However, now that we have
more fine-grained dependencies, we really want an _edge-labeled_ graph, which
unfortunately it not available in the `containers` package. Therefore I've
written a very simple wrapper around Data.Graph that supports edge labels; we
don't need many fancy graph algorithms, and can still use Data.Graph on these
edged graphs when we want (by calling them on the underlying unlabeled graph),
so adding a dependency on `fgl` does not seem worth it.

The modular solver has its own representation for a package (PInfo). In this
commit we modify PInfo to keep track of the different kinds of dependencies.

This is a bit intricate because the solver also regards top-level goals as
dependencies, but of course those dependencies are not part of any 'component'
as such, unlike "real" dependencies. We model this by adding a type parameter
to FlaggedDeps and go which indicates whether or not we have component
information; crucially, underneath flag choices we _always_ have component
information available.

Consequently, the modular solver itself will not make use of the ComponentDeps
datatype (but only using the Component type, classifying components); we will
use ComponentDeps when we translate out of the results from the modular solver
into cabal-install's main datatypes.

We don't yet _return_ fine-grained dependencies from the solver; this will be
the subject of the next commit.

The ComponentDeps datatype will give us fine-grained information about the
dependencies of a package's components.  This commit just introduces the
datatype, we don't use it anywhere yet.

This commit does nothing but rearrange the Modular.Dependency module into a
number of separate sections, so that's a bit clearer to see what's what. No
actual code changes here whatsoever.

I don't know why we we constructed this graph manually here rather than calling
`graphFromEdges`; it doesn't really matter except that we will want to change
the structure of this graph somewhat once we have more fine-grained
dependencies, and then the manual construction becomes a bit more painful;
easier to use the standard construction.

This address @23Skidoo's comment
https://github.com/haskell/cabal/pull/2500#issuecomment-8703532

Since we didn't really have a unit test setup for the solver yet, this
introduces some basic tests for solver, as well as tests for independent goals
specifically.

In particular, in the definition of dependencyInconsistencies.

One slightly annoying thing is that in order to validate an install plan, we
need to know if the goals are to be considered independent. This means we need
to pass an additional Bool to a few functions; to limit the number of functions
where this is necessary, also recorded whether or not goals are independent as
part of the InstallPlan itself.

This is implemented as a separate pass so that it can be understood
independently of the rest of the solver.

POption annotates a package choice with a "linked to" field. This commit
just introduces the datatype and deals with the immediate fallout, it doesn't
actually use the field for anything.

Duncan Coutts authored 10 years ago and Edsko de Vries committed 9 years ago

It turns out not to be the right solution for general private
dependencies and is just complicated. However we keep qualified
goals, just much simpler. Now dependencies simply inherit the
qualification of their parent goal. This gets us closer to the
intended behaviour for the --independent-goals feature, and for
the simpler case of private dependencies for setup scripts.

When not using --independent-goals, the solver behaves exactly as
before (tested by comparing solver logs for a hard hackage goal).
When using --independent-goals, now every dep of each independent
goal is qualified, so the dependencies are solved completely
independently (which is actually too much still).

Handle more 'strip --version' output formats.

Fixes #2497.

-Wall police

GHC HEAD now warns about redundant constraints.

Clean up graph traversal code