Fix construction of conflict sets during link validation

[edsko]

NOTE: This commit contains four commits, not two; the first two are #3221.

In #2834 @grayjay points out another conflict set construction bug. Actually, this is almost the same bug that was present in the check for the single instance restriction: we were forgetting to call simplifyVar on some variables before adding them to the conflict set.

This PR first adds @grayjay 's test case (another excellent test, thank you very much! :) ) . The next commit just adds some comments to the test explaining what is going on. Commit number 3 just moves some code out of the Modular.Dependency module into two separate Var and ConflictSet modules; it does not make any actual changes to the semantics. The actual fix is in the final commit, which makes ConflictSet abstract and makes sure that whenever we add a variable to a ConflictSet, or check whether a variable is a member of a conflict set, that we call simplifyVar first. This fixes this particular test case and also makes me feel a lot less anxious about other such bugs still being present.

Pinging @kosmikus and @dcoutts .

[edsko]

Pushed commit to try to fix build error on 7.4.

[edsko]

New travis failure seems to be transient?

[23Skidoo]

Restarted the build.

[23Skidoo]

The remaining failure is due to -Werror:

Distribution/Client/Dependency/Modular/ConflictSet.hs:65:1: warning: [-Wredundant-constraints]
    • Redundant constraint: Ord qpn
    • In the type signature for:
           filter :: Ord qpn =>
                     (Var qpn -> Bool) -> ConflictSet qpn -> ConflictSet qpn

[edsko]

Oh FFS 😧 Final attempt :-)

[dcoutts]

I've merged these patches to the nix-local-build branch (along with all the other recent solver PRs).

[edsko]

This has been merged into and subsumed by #3357 .