Further wibbles to moving dependency analysis back to renamer; fixes ghci failures

Wibble; fixes TH failures

Fix bogon in reporting of duplicate sigs; make mod68 work again

Do 'tidying' on Kinds before printing them.  This avoids printing
stuff like 'k_43b' in user error messages.

To do this, I ended up adding an OccName to Kind.KindVar.  Even
then the implementation is a bit of hack (see comments with 
Type.tidyKind).  Still, it's a highly localised hack, whereas the
"right thing" entails making KindVar into a flavour of Var, which
seems like an uncomfortably big change.

   I think this change can merge to the stable branch

Rename exprIsValue to exprIsHNF, which is more accurate

It turned out that doing all binding dependency analysis in the typechecker
meant that the renamer's unused-binding error messages got worse.  So now
I've put the first dep anal back into the renamer, while the second (which
is specific to type checking) remains in the type checker.

I've also made the pretty printer sort the decls back into source order
before printing them (except with -dppr-debug).

Fixes rn041.

-Wall cleaning and dead code removal

Fix a bug caused by the change in representation for NPats;
the pattern-overlap checker didn't know about the change,
and so treated +1 and -1 as equal...
	
	MERGE to STABLE

Fix pretty-printing bug for NPat; MERGE to STABLE

Wibble to loadHomeInterface for TH quoting; MERGE to STABLE

Comments

Correct attribution of existential types; MERGE to STABLE

Add notes about Typable restrictions; merge to stable

mapUFM was strict in the structure of the tree, perhaps accidentally.
In any case, this interacted badly with the use of mapVarEnv in
SimplEnv.mkCoreSubst, which assumed that the map was lazy.

mapUFM is now lazy in the structure of the tree, which matches the
behaviour of FiniteMap.mapFM.

This fixes some long compilation times, in particular HTMLMonad98.hs
from WASH used to take 2 mins to compile without optimisation, now
takes less than 20 seconds.

callSiteInline: consider inlining even the once-in-one-branch
occurrences.  See comment for details.

preInlineUnconditionally: replace exprIsValue call with home-grown
version to fix loss-of-sharing that arises because exprIsValue doesn't
necessarily mean "ok to inline inside a lambda".  Fixes last night's
10queens test failures.

Check for RTLD_DEFAULT in dlfcn.h and use it if it is available.
(On Mac OS X its value is -2, and using NULL instead does not work.)

MERGE TO STABLE

fix bogus #ifdef in defaultHscTarget

Patch from SimonPJ (slightly tweaked by me after checking performance
results):

Fix occasional O(n^2) behaviour in the simplifier.  There was a
possibility that by inlining a binding, we could re-simplify an
arbitrary sized expression.  This patch fixes it by moving the
inlining of arbitrary-sized expressiong to the binding site
(preInlineUnconditionally), so the decision to inline happens before
simplifying the RHS.  To do this, we have to collect more information
during the occurrence analysis phase.

We still make inlining decisions at the call site, but they are always
size-limited, so we can't get quadratic blowup.

Fix the location of _stub.o files when compiling with -odir.  The
comment from the source:

-- The _stub.c file is derived from the haskell source file (but stored
-- in hscStubCOutName in the dflags for some reason, probably historical).
-- Consequently, we derive the _stub.o filename from the haskell object
-- filename.
--
-- This isn't necessarily the same as the object filename we
-- would get if we just compiled the _stub.c file using the pipeline.
-- For example:
--
--    ghc src/A.hs -odir obj
--
-- results in obj/A.o, and src/A_stub.c.  If we compile src/A_stub.c with
-- -odir obj, we would get obj/src/A_stub.o, which is wrong; we want
-- obj/A_stub.o.

Update documentation for ghc-pkg list and ghc-pkg latest.

Tidyup, and make all the commands that take a package identifier
behave the same.  Previously, naming a package without a version
number only matched the same package identifier, if you wanted to
match multiple packages you had to say P-*.  However, the new list and
latest commands behave differently, which was inconsistent.

I believe the previous behaviour was for backwards compatibility with
packages that had no version number.  However, it's unlikely that this
is ever useful, so I've changed the semantics so that a package name
on its own matches all versions of that package.

- Add "ghc-pkg list <pkg>" to list packages matching <pkg>
- Add "ghc-pkg latest <pkg>" to show the latest version of <pkg>
- Add --simple-output option to produce a more easily slurpable output
  from ghc-pkg list.

Patch from Lennart Kolmodin <kolmodin at dtek.chalmers.se>

small tidyup for memInventory

Fix up GC debugging output

Fix tiny debugging output bug

- move instance Show FastString into FastString module
- instance Outputable FastString should use ftext instead of unpackFS

A little bit of strictness (doesn't actually help much)

Fix an SCC

Annotate foreign calls with volatile registers to reduce the amount of
saving that needs to be done.

save + restore caller-saves registers around foreign calls

Add callerSaveVolatileRegs for saving registers around C calls

Add a ToDo

Allow an empty list of volatile regs on a call

Rmoeve a few lines that are provided by lib/compat/compat.mk

For Windows (moved from ghc-pkg/Makefile):

HC_BOOT_LIBS += -lshell32

Add explicit dependency for Distribution.Compat.FilePath

win32/AsyncIO.c:onIOComplete(), awaitRequests():
  Avoid completion table overflow through the use of
  a semaphore. Worker threads (the producers) wait for a free
  event slot to appear, the RTS thread (consumer) signal slot
  availability once the completion event has been de-queued and
  processed.

  This bug only manifests itself in 'massively concurrent' code
  (>200 threads), where the threads simultaneously engage in IO
  through the async IO layer. conc023 and conc036 in the regression
  test suite are examples of such.

Merge to STABLE.

- Remove -fno-cpr, which doesn't exist
- remove redundancy in description of -fno-full-laziness

Document a few more options