This patch teaches SpecConstr about casts; see Note [SpecConstr for casts]

Note [Float coercions]
~~~~~~~~~~~~~~~~~~~~~~
When we find the binding
	x = e `cast` co
we'd like to transform it to
	x' = e
	x = x `cast` co		-- A trivial binding
There's a chance that e will be a constructor application or function, or something
like that, so moving the coerion to the usage site may well cancel the coersions
and lead to further optimisation.  Example:

     data family T a :: *
     data instance T Int = T Int

     foo :: Int -> Int -> Int
     foo m n = ...
        where
          x = T m
          go 0 = 0
          go n = case x of { T m -> go (n-m) }
		-- This case should optimise

The substitution used to carry "fragile" OccInfo to call sites via the
DoneId constructor of SimplEnv.SimplSR.  This was always a tricky thing 
to do, and for some time I've been removing the need for it.

Now at last I think we can nuke it altogether.  Hooray.

I did a full nonfib run, and got zero perf changes.

This is another attempt to fix the interaction between recursion and
RULES.  I just had it wrong before!  Now the significance of the
flag on IAmALoopBreaker is given in BasicTypes

  | IAmALoopBreaker	-- Used by the occurrence analyser to mark loop-breakers
			-- in a group of recursive definitions
	!RulesOnly	-- True <=> This loop breaker mentions the other binders
			--	    in its recursive group only in its RULES, not
			--	    in its rhs
			--  See OccurAnal Note [RulesOnly]

Note [Recursive unboxing]
~~~~~~~~~~~~~~~~~~~~~~~~~
Be careful not to try to unbox this!
	data T = MkT !T Int
But it's the *argument* type that matters. This is fine:
	data S = MkS S !Int
because Int is non-recursive.


Before this patch, we were only doing the unboxing if the *parent*
data type was non-recursive (eg that meant S was not unboxed), but
that is over-conservative. 

This showed up with indexed data types (thanks to Roman for finding it)
because indexed data types are conservatively regarded as always recursive.

Note [Scrutinee with cast]
~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this:
    f = \ t -> case (v `cast` co) of
	         V a b -> a : f t

Exactly the same optimistaion (unrolling one call to f) will work here, 
despite the cast.  See mk_alt_env in the Case branch of libCase.

This patch does the job.  For a change, it was really easy.

Remove ILX from the GHC altogether (although I left the source file IlxGen in case anyone wants to see it)

We want the universal and existential tyvars of a data constructor to
have distinct OccNames.  It's confusing if they don't (in error messages,
for example), and with the current way of generating IfaceSyn, it actally
generates bogus interface files.  (Which bit Roman.)

When IfaceSyn is full of Names, this won't matter so much, but it still
seems cleaner.

This patch adds a 'tidy' step to the generation of DataCon type 
variables in TcTyClDecls.tcResultType 

This is part 2 of the patch that improved the interaction of RULES and
recursion.  It's vital that all Ids that may be referred to from later in
the module are marked 'IAmALoopBreaker' because otherwise we may do
postInlineUnconditionally, and lose the binding altogether. 

So I've added a boolean rules-only flag to IAmALoopBreaker.  Now we can
do inlining for rules-only loop-breakers. 

Note [Case of cast]
~~~~~~~~~~~~~~~~~~~
Consider 	case (v `cast` co) of x { I# ->
		... (case (v `cast` co) of {...}) ...
We'd like to eliminate the inner case.  We can get this neatly by 
arranging that inside the outer case we add the unfolding
	v |-> x `cast` (sym co)
to v.  Then we should inline v at the inner case, cancel the casts, 
and away we go
	
This patch does the job, fixing a performance hole reported by Roman.

See Trac #683

This patch improves the interaction of recursion and RULES; at least I
hope it does.   The problem was that a RULE was being treated uniformly like
an "extra RHS". This worked badly when you have a non-recursive definition
that is made recursive only by RULE.

This patch maeks the occurrence analyser know whether a binder is referred to
only from RULES (the RulesOnly constructor in OccInfo).  Then we can ignore
such edges when deciding on the order of bindings in a letrec, and when
setting the LoopBreaker flag.

The remaining potential problem is this:
	rec{ f = ...g...
	   ; g = ...f...
	     RULE g True = ...
	   }

The RULE for g may not be visible in f's rhs.  This is fixable, but not
today.

I had forgotten to bring scoped type variables into scope at an expression
type signature, such as
	e :: forall s. <type>
where 's' should scope over the expression e.

Like everything to do with scoped type variables, fixing this took an 
unreasonable amount of work.  I'm sure there must be a better way to 
achitect this!

I updated the user manual too.

A test is tc213.

It would be good to push this into 6.6.1

Fixes Trac #919

Before the coKindFun could be applied to too many arguments;
now it expects exactly the right number of arguments.  That
makes it easier to write the coKindFuns, and localises the work.