The float-in pass wasn't doing the right thing when you have

	let x{rule mentions y} = rhs in body

It allowed a binding mentioning y to float into the body, which is
obviously wrong.  I think this bug has been there a long time; I don't
really know why it has not come up before.

It showed up when compiling Text.Regex.Base.Context with WAY=p in
package regex-base.

Build an implication constraint if there is a non-trivial
refinement, even if there are no other 'given' constraints.

Test = gadt/set.hs

Fixes trac #929.
Merge to 6.6 branch.

I broke it during my recent interface-file overhaul

This is necessary to revert CAFs. Previously to this patch the user would get a msg "You may need to reload your modules". This patch takes care of that

A module loaded under debugging mode but owning no breakpoints was erroneously identified by ghci as a non-under debugging module, producing a confusing error msg when the user tried to set a breakpoint

The fix inserts an empty list of sites in the module-sites dictionary used by the debugger

Amusingly, this little error in the GHC manual came from the original
SPJ proposal for pattern guards from 1997 and even slipped into the
Haskell Workshop 2000 paper by SPJ and Martin Erwig.  It's almost 10
years old.

its root. Also gets rid of the getDeclMainBinder function which isn't needed 
anylonger.

This fixes #1047

This is so that exceptions raised by Main.main do exactly the same
thing as they would in a compiled program, including writing the
message to stderr and shutting down with the correct exit code.

- This patch cleans up the HsSyn representation of type family declarations.
- The new representation is not only less delicate, it also simplified teh code
  a bit.
- I took the opportunity of stream lining the terminology and function names
  at the same time.
- I also updated the description on the wiki at
  <http://hackage.haskell.org/trac/ghc/wiki/TypeFunctionsSyntax>

- The test for being able to derive the requested classes needs to be made
  with the representation tycon (not the family tycon).
- Standalone deriving for indexed types requires the instance types in the
  derive clause to match a data/newtype instance exactly (modulo alpha).

	(MERGE to STABLE branch, pls)

This patch makes us a bit more relaxed about ambiguous 
class method types.  See tc223 for an example.

Reported by Yitzchak Gale

There was a plain bug in the cast-optimiation code -- a call to
splitCoercionKind_maybe instead of coercionKind!  Result was that
we missed useful opportunities to move casts around.  Trac #995
is an example, but I bet there are more.

Now that coercion variables mention types, a type-lambda binder can
have free variables.  This patch adjusts the free-variable finder
to take account of this, by treating Ids and TyVars more uniformly.

In addition, I fixed a bug in the specialiser that was missing a 
free type variable in a binder.  And a bug in tyVarsOfInst that
was missing the type variables in the kinds of the quantified tyvars.

Fixes Trac #1053