GitLab

This means it should be possible to set GhcUnregisterised=YES in
build.mk and the rest of the settings should follow automatically
(GhcWithNativeCodeGen, SplitObjs, GhcWithSMP).

* -{short,long}-ghci-banner are now dynamic options, so you can put
  ":set -short-ghci-banner" in .ghci
* The -v2 banner information now always tells you what compiler booted GHC,
  and what stage the compiler is. Thus we no longer assume that stage > 1
  iff GHCI is defined.

Add -short-ghci-banner and -long-ghci-banner.  The default is long, which is
the current behavior.  The short banner prints a one-line introduction with
only the version, web site, and ":? for help" message.

Cabal defaults to the value we want on Linux, but uses
"C:\\Program Files\\Common Files" on Windows.

Bindists should now work again, when doing "make install" at least.
"make in-place" is probably still broken.

Key changes:
* Always build as if BIN_DIST is 1. BIN_DIST is thus removed.
* Libraries are configured with prefix set to $$topdir rather than $(prefix)

foreign import ccall unsafe hs_hpc_write :: CString -> IO ()
foreign import ccall unsafe hs_hpc_read  :: CString -> IO ()

These write a Hpc description of the state of the world to a file, 
or read a description into the current Hpc tickbox subsystem.

- mkDataConIds forced an algebraic worker/wrapper on newtype instances (not
  on vanilla newtypes).  I am not sure where this came from (from code I
  wrote or from the recent restructuring of MkIds), but its not the right
  thing to do.  In particular, it requires newtype instances to have con_info
  labels in the generated code, which makes no sense.
- Worker construction for newtypes is smart enough that it will construct the
  right signature for newtype instances (ie, one that mentions the family
  type constructor, not the instance tycon).

- Fix typo introduced when moving deriving-specific code from TcEnv to TcDeriv
- Thanks to Roman for the bug report

This patch implements the suggestion in Trac #1398.  It's obviously
stupid to suggest -fno-monomorphism-restriction if the user is already
using it.  

(Maybe another suggestion would be good, but this one clearly bogus.)

The renamer used to be responsible for making sure that all interfaces
with instance decls (other than orphans) were loaded.  But TH makes that
impossible, so the typechecker does it, via checkWiredInTyCon.

This patch simply removes redundant additions to the free-variable set
in the renamer, which were there, I believe, solely to ensure that the 
instances came in.  Removing them should change nothing, but it's a 
useful clean up.

Typechecking of explicit tuples is done "by hand" (rather than using
boxySplitTyConApp) so that rigidity is maintained.  In making that change
I'd forgotten to retain the call to checkWiredInTyCon, which is important
to make sure instances are brought into scope.

Test is tc227.

I skipped utils/hsc2hs/Main.hs since its ifs also involved
checking for old versions of nhc98 (I don't want to figure that out),
but removed everything else I found relating to building with pre-6.0