[GHC] #9955: ghc-stage1 compiles with bootstrapping ghc package, not the built one

[GHC]

#9955: ghc-stage1 compiles with bootstrapping ghc package, not the built one
-------------------------------------+-------------------------------------
        Reporter:  ezyang            |                   Owner:
            Type:  bug               |                  Status:  new
        Priority:  high              |               Milestone:  7.10.1
       Component:  Build System      |                 Version:  7.11
      Resolution:                    |                Keywords:
Operating System:  Unknown/Multiple  |            Architecture:
 Type of failure:  Incorrect result  |  Unknown/Multiple
  at runtime                         |               Test Case:
      Blocked By:                    |                Blocking:
 Related Tickets:                    |  Differential Revisions:
-------------------------------------+-------------------------------------
Changes (by ezyang):

 * priority:  highest => high


Comment:

 Falsified previous hypothesis. In fact, building GHC HEAD 7.10 works just
 fine; it's just the other way around that is a problem. Here's what's
 going on:

 When there are multiple wired in packages, we seem to ignore the hide-all-
 packages directive, and instead simply take all instances of the wired in
 package and *select the newest version*. Thus, when we compile GHC HEAD
 with GHC 7.10, the ghc package from GHC HEAD has a newer version than the
 system ghc package from 7.10, and we use that when compiling.  However,
 when we build 7.10 with GHC HEAD, we always select the system GHC HEAD
 library, which works about as well as you'd expect.

 What is going on with the `--show-iface`? The key is that package
 dependency description is done by a lookup in the package database. The
 package key for 'ghc' is always ghc, no matter what the version is (since
 it's wired in), but I notably have NOT told ghc about the inplace package
 database in my show-iface command. So we lookup the package keyed 'ghc' in
 the default database, which is 7.10's, voila.

 This behavior sort of makes sense. In particular, it does make sense for
 GHC to try to output something besides a package key when referring to
 packages in debug output, because in general you don't necessarily care
 about the key in question, and the output is a lot more readable if you
 just get package name-version. The downside is it's easy to mess up wired
 in packages.

 It's possible #9652 is caused by the same underlying mechanism.

 My proposed fix for the wired in problem is to properly teach GHC to
 respect `-hide-all-packages -package wired-in-0.1.2`: we really SHOULD
 pick 0.1.2, even if there is a newer version. Of course, if we don't
 mention the wired in package, it probably makes sense to default back to
 the newest version.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9955#comment:2>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler