Original patch by brianlsmith@gmail.com

This has the undesitrable effect that stg_gc_gen_hp in the RTS was
caught as an HPC symbol and put in the data segment.

This is a minor fix up to the patch
 * retain arity for let-bound vars in simplifier

This time, the PowerPC NCG suffered from a typo.

(I've really grown to hate all those #if foo_TARGET_ARCH
 lines. I need to build GHC for at least four different
 platforms just to make sure that there are no typos...)

We were freeing the tasks in exitScheduler (stopTaskManager) before
exitStorage (stat_exit), but the latter needs to walk down the list
printing stats. Resulted in segfaults with commands like
    ghc -v0 -e main q.hs -H32m -H32m +RTS -Sstderr
(where q.hs is trivial), but very sensitive to exact commandline and
libc version or something.

We now check for freeing memory that isn't allocated, and overwrite memory
we are about to free with 0xaa.
It might be nice to also have a flag to control whether the memory
actually gets free'd.

This patch fixes several bugs in the handling of impliciation
constraints, thereby fixing several regression-suite failures.

On the way I managed to further simplify the code in TcSimplify;
the extra lines are comments.

Continue repairing breakage from patch
"x86_64: support PIC and therefore, Mac OS X in the NCG",
this time something affecting (linux && !i386)

GHCi evaluates a command to flush the stdout/stderr buffers, but this
command can only use names that are in scope. It used to need IO.stderr
from the haskell98 package, but this patch means it only needs things
from the base package.

We weren't getting the bitmap right for the continuation BCO in a case-of-case.

Make things compile on Linux again

On x86[-64], MachCodeGen uses the old XOR trick to zero a register.
This patch makes regUsage not list the register as an input in this
case.
Listing the register as an input for the instruction could make it 
appear like an unitialised value, which is bad because unitialised
values can cause the register allocator to panic (at least in the
presence of a loop).

Supporting x86_64-apple-darwin in the NCG basically boils down to supporting
position-independent code in the NCG.
PIC code works almost exactly the same as on x86_64-linux, while position-dependent
code is not supported at all.
This patch implements -fPIC for x86_64-linux, too, but that is untested.

* So far, family instance modules was only available for external modules.
* This fixes the "Over" test in the testsuite under indexed-types/

* This is only a slight generalisation of the parser, so that family
  declarations on the toplevel and in classes are uniform.
* I didn't allow that right away as it is a bit tricky to avoid reduce/reduce 
  conflicts.

I recentl changed the scoring system used by dependency analysis for
recursive bindings, that it used the *form* of the RHS of a binding,
rather than just its type. In doing so I inadvertently made recursive
dictionary bindings unravel less well, because I'd missed the case
of 	c = /\a. C (...) (...)

This patch fixes the problem.  A good example is the instance for
Monad (ST s) or Show (ST s a) in GHC.ST.  It's vital for these
dictionaries to be inlinable.