Correctly account for -package-db ordering when picking packages.

Authored by ezyang on Mar 19 2017, 6:07 PM.

Description

Correctly account for -package-db ordering when picking packages.

Summary:
When I originally implemented ABI-based shadowing as per
ee4e1654c31b9c6f6ad9b19ece25f040bbbcbd72, I switched our strategy
from pasting together lists to creating a map of all units first,
and then selecting packages from this. However, what I did
not realize when doing this was that we actually depended
on the *ordering* of these lists later, when we selected
a preferred package to use.

The crux is if I have -package-db db1 -package-db db2 -package p-0.1,
and p-0.1 is provided by both db1 and db2, which one does the
-package flag select? Previously, this was undetermined; now
we always select the instance from the LATEST package database.
(If p-0.1 shows up multiple times in the same database, once again
the chosen package is undefined.)

The reason why cabal08 intermittently failed was that, in practice,
we were sorting on the UnitId, so when we bumped version numbers,
that often wibbled the UnitIds so that they compared oppositely.
I've extended the test so that we check that the relation is
antisymmetric.

Fixes Trac #13313

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

Test Plan: validate

Reviewers: bgamari, austin

Subscribers: rwbarton, thomie

Differential Revision: https://phabricator.haskell.org/D3369