This fixes #220: new-build now builds, installs and adds executables to
PATH automatically if they show up in 'build-tools'.  However, there is
still more that could be done: the new behavior only applies to a
specific list of 'build-tools' (alex, happy, etc) which Cabal recognizes
out of the box.  The plan is to introduce a new 'tool-depends' field to
allow dependencies on other executables as well.

Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>

Previously, the solver filtered out redundant backjumping messages twice, once
in 'Log.logToProgress' and again in 'Message.showMessages. However,
'showMessages' relied on the backjumping messages to determine where to insert
messages about missing packages. This led to missing "unknown package" messages
(part of issue #3617).

This commit removes the filtering in 'logToProgress', because it was redundant.

This change avoids holding the whole log in memory.

Added this so that we can verify the unknown package error.

This is relevant for unknown packages. Without the current
goal in the initial conflict set, the unknown package itself
does not end up in its own conflict set.

But we want it to, if only to make sure that the error slicing
machinery keeps the message about the unknown goal around.

It is also the correct thing to do. The initial conflict set
corresponds to a virtual "avoid" choice underneath the package
goal. That choice corresponds to avoiding assigning any proper
value to the goal at all, which might be possible if the goal
was not needed. The conflict set for this "avoid" goal is always
between the current package and its immediate goal reason.

I have added comments to the code explaining this as well.

kristenk authored Apr 11, 2016 and Edsko de Vries committed Apr 21, 2016

This commit replaces a call to 'lgBlame' with a call to 'lgConflictSet'.
'lgConflictSet' adds the members of the link group to the conflict set. The
other members are required, for example, when a linked package's version doesn't
match a constraint, and the solver must try other versions for the linked-to
package.

All tests now pass.

and fix the error message in showMessages; this keeps the construction of
conflict sets uniform.

This still isn't quite right though. If we remove B from db21 as well, we don't see a line "unknown package B" appearing in the log. I have no idea why.

Also, @kosmikus says:

> ok, so first problem: we still need to add C to the conflict set
> one way to do this now is to change the "initial" conflict set in the backjump calls to be the node + the goal reason again
> I was assuming that the initial set is only ever added to an existing conflict set that we already have
> but this is now no longer true

I have no idea what that means :)

In the original implemented we recorded only the two conflicting packages; in
the "fix" in #3221 we include the GRC for one of the two packages, but not the
other. This is also wrong, although slightly less so and @kosmikus tells me
that it's actually rather difficult to find an example that exposes the
remaining inaccuracy. I'll have to take him at his word on that :) In this
commit we use the GRC for both goals instead.

This commit is a pure refactoring, no semantic changes. Submitting as separate
PR at @kosmikus request.

Tihs moves `ConflictSet` and `Var` into their own modules.

This prevents individual tests from running too long and limits the size of the
solver log.

This commit removes the 'ExTest' constructor from 'ExampleDependency'.
Test dependencies are now represented using the same
'ExampleDependency' constructors as library dependencies.  The only
difference between dependencies of different components is that they
are placed under different keys in 'D.C.ComponentDeps.ComponentDeps'.

Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu>