AIX/RS6000 patches

Fix: mkPiType has moved.

Alter to match new CoreLint.

A few things here:

  * fix some comment typos
  * alter CoreLint architecture to permit warnings as well as errors
  * add `endPassWithRules' to permit printing of rules as well as binds
  * move mkPiType to CoreUtils (for UsageSP reasons)

printf => showFFloat. Now waugh_neural fails with
   Fail: Prelude.(!!): negative index,
but that's better than before.  :-)

Synched with output required by H98 report

*nix-ified the expected output. M$-people are more used to error
messages, and one has to lose...  >:-)

"Non-exhaustive guards" => "Non-exhaustive patterns"

Print only the first 6 characters of the result's decimal
representation, this "solves" the problem that without optimization
0.46920002 was printed and 0.4692 with -O. Perhaps this could be
turned into a test for constant folding, but this was probably not the
test writer's intention.

Use "foreign label" instead of litlit for stdout.

`foreign label foo :: Addr' was simply mapped to `foo' on the C side,
but this is wrong, resp. only works for C *functions*, not *variables*
(praise the implicit conversions of C!). It now correctly maps to `(&foo)'.

ATTENTION: I'm not sure if this fix breaks some SW which depends on
the old (wrong) behaviour (H/Direct?).

Added type signature to resolve ambiguity.

Synched with reality.

Synched with reality. The expected output is a little bit fragile...

-syslib => -package

HC_OPTS => SRC_HC_OPTS again, sorry about that

Added a few lines how to pass global options, just for oblivious
people like me.  :-}

* Changed `SRC_HC_OPTS += ...' to `HC_OPTS += ...', otherwise `make
  SRC_HC_OPTS=foo' does not work.

* -syslib => -package

Only add -o? FOO.stdBAR when it exists. I'm not sure if this
foreach-hack is the most elegant way to go though...

Delete whole swathes of old FFI-related stuff, after all no
documentation is better than wrong documentation :)

revert part of previous commit that sneaked in by mistake.

update link to deleted mkdependHS stuff

add an id to the dependency section

Pretty-printing EncodedStrings, next try...

Update docs on mkdependHS: it isn't a separate utility, so document it
under "Using GHC".

Also update section on Happy.

Undo the changes and make userStyle printing work again

Fix a nasty bug: the 'c' component of StgUnion was declared as type
StgChar, which sounds reasonable, but meant that when assigning a char
to R1 gcc would correctly arrange to only overwrite the low byte
rather than zero-extending the value.  We assume elsewhere that when
R1 contains a char that the upper 24 bits are zero, so this was wrong.

This bug must have been around for a *long* time, strange that it only
just showed up. <shrug>

Commiting version of STG -> GOO that seems to compile PrelBase successfully.

Many other wibbles; esp. String handling.

* Synched comments with reality

* Ensure that a f.e.d. function is never inlined, because the address
  of the C stub (a litlit) is might not be in scope in other modules.
  (untested fix).

*** merge ***

Adding change that handles trivial use of primitives (compares,
integer arithmetic, etc) better when generating Java.

Update instFunDepsOfTheta to accomodate Simon's recent commit (it hadn't
needed to generate fundeps for IParams previously).  This undoes Simon's
comment about `forall a. ?x::a => Int' being ambiguous.  It isn't, and
now it isn't flagged as such either.  I.e, there's nothing wrong with:
    foo :: ?x::a => Int
    foo = fst (13, ?x)
(of course, less contrived examples can also be given!).

Simon was kind enough to re-write elegantly a function that I had written
awkwardly, but revealed that beautification doesn't always yield better
code ;-)  Fixed a silly typo and further beautified the code (so that the
cut-n-paster error that Simon introduced would have been easier to spot).

ieee-flpt.h is in ../../includes, not ../includes.

GHC gets upset if you have types like

	Eq a => a->a

where 'a' is *not* universally quantified.
By "upset" I mean that the typechecker generates
rather bogus code, that subsequently kills Lint.

Such types used to be rejected in the renamer,
but Jeff removed that in favour of an ambiguity
check in TcMonoType.  I remember agreeing to move
the renamer check to the type checker, and Jeff did this,
but the check in TcMonoType was only checking for
*ambiguity*, which isn't quite the same.

I've restored the missing check and commented it
better in TcMonoType.

Jeff: if this isn't right for you, let's dicuss.

Incidentally, I also generalise the ambiguity check
to detect
	forall a. ?x::a => Int
which is ambiguous.

I did a few formatting changes too.

Don't group variable bindings with the same name together.  They
should be flagged as duplicate definitions.

remove NonExhaustiveGuards exception

Guard failure will now throw a PatternMatchFail, the
NonExhaustiveGuards exception is to be removed.

Part of the -split-objs commit that I forgot yesterday.

In my commit of 24 May I got this boolean condition
back to front:

    tryWW non_rec fn_id rhs
      | not (isNeverInlinePrag inline_prag)
      =  -- Don't split things that will never be inlined

The 'not' is obviously wrong!  As a result virtually nothing
is being worker-wrapper'd

How this has survived for more than two weeks beats me.

Commiting version of the STG->GOO code generator that works with fib.