Synched with versions from autoconf 2.59

A complete rewrite of ar-related tests:
* Modularized the tests
* Fixed (i.e. introduced :-) caching
* Made the test for support of "-input" more robust

Buglet in desugaring TH syntax

Document -fcontext-stack as static

The "rebindable-syntax" stuff wasn't dealing with the new location
information correctly.  This commit fixes the problem, and thereby
makes mdofail004 work right.  Maybe others too.

Revert to previous kind-check of for-alls, to avoid error message worsening

Document :kind

Wibble to :k

import wibble

Improve documentation about ambiguous types

Remove the entirely-redundant location from the argument of 
constructor HsPredTy,
    so that we have
	HsPredTy HsType
    rather than
	HsPredTy LHsType

Remove redundant case-analysis for single-constructor case of gunfold

Add :k(ind) command to ghci to find the kind of a type.
This works very, very similarly to :t(ype)

In the derived code for gunfold, use a wild-card for the 
final case, to avoid a redundant test, and to eliminate the
annoying warning about un-matched cases.

While I'm at it, rename HsUtils.wildPat to nlWildPat, for
consistency.

Don't convert Exact RdrNames into Orig ones when comparing RdrNames.
This implicit conversion (via nukeExact) was making the carefully-distinct
RdrNames conjured up by Template Haskell into the same RdrName, and
that in turn was reporting duplicate bindings.  

I'm not 100% certain that the implicit conversion isn't needed
for some other purpose, but it seems unclean anyway, so I'm removing
it, and we'll see if anything else breaks.

Import trimming

Fix parsing bug in new deriving stuff, which was killing the HEAD

Another stage-2 wibble to last commit

Wibble to last commit

Extend the "newtype deriving" feature a little bit more
(at the request of Wolfgang Jeltsch)

Here's the example:
    class C a b
    instance C [a] Char
    newtype T = T Char deriving( C [a] )

Perfectly sensible, and no reason it should not work.
Fixing this required me to generalise the abstract syntax of
a 'deriving' item, hence the non-local effects.

Some updates I've had sitting in my tree for a while:

  - when there are two logs only, print a summary table at the beginning
    of the output with columns for allocs, runtime, memory reads/writes, instrs.

  - handle multiple runs of each program, and average the results. (changes
    to nofib build system to follow).

cachegrind: update command line syntax for recent versions of valgrind

Document -fno-state-hack

Add a flag -fno-state-hack, which switches off the "state hack".

It's claims that every function over realWorldStatePrimTy is a
one-shot function.  This is pretty true in practice, and makes a big
difference.  For example, consider
	a `thenST` \ r -> ...E...
The early full laziness pass, if it doesn't know that r is one-shot
will pull out E (let's say it doesn't mention r) to give
	let lvl = E in a `thenST` \ r -> ...lvl...
When `thenST` gets inlined, we end up with
	let lvl = E in \s -> case a s of (r, s') -> ...lvl...
and we don't re-inline E.

Get rid of DiscardedInfo, and fix a Ghci bug at the same time.

The new story is this:

- We always read the whole interface file, as it exists on disk,
  not dropping pragmas or anything.

- We compare that from-the-disk copy with the new version before
  writing the new interface file.

- We drop the pragmas
  a) Before loading the interface payload into the declaration pools
  b) In the no-need-to-recompile case, before typechecking the
	interface decls.  Omitting this was the previous bug.

* Improve error message for overlapping instances

* Improve handling of type-variable-only constraints like (Foo a).
  Previously we never looked them up in the instance envt, *except* if
  -fallow-undecideable-instances was on, because that allows
	instance (...) => Foo a
  But -fallow-undecideable-instances might be on in the module with the
  instance decl, but off in the importing module.  Also it's really a per-class
  thing.  So now we just record in the instance env whether there are any such
  strange instances, a kind of short-cut for the lookup.

* Arrange that we are a bit more eager about resolving overloading in
  the case of existential pattern matching [George Russel suggestion]
  Here's the example (see comments in InstEnv)
	-- The key_tys can contain skolem constants, and we can guarantee that those
	-- are never going to be instantiated to anything, so we should not involve
	-- them in the unification test.  Example:
	--	class Foo a where { op :: a -> Int }
	--	instance Foo a => Foo [a] 	-- NB overlap
	--	instance Foo [Int]		-- NB overlap
	-- 	data T = forall a. Foo a => MkT a
	--	f :: T -> Int
	--	f (MkT x) = op [x,x]
	-- The op [x,x] means we need (Foo [a]).  Without the filterVarSet we'd
	-- complain, saying that the choice of instance depended on the instantiation
	-- of 'a'; but of course it isn't *going* to be instantiated.

Fix erroneous hi-boot

A preliminary step towards being able to identify existential
type variables separately.  That in turn helps when resolving
overloading; I think we want to resolve overloading without
worrying about what these type variables might instantiate to.

wibble

Comments mainly, plus use newTyConRhs

Remove unused result from dropInline

Comments

Import wibbles

Add support for foreign imports inside quasi-quotes.
Gave TH a few more uniques to play with and fixed a typo.

ClosureInfo.might_be_a_function: this predicate wasn't taking into
account abstract types, which might also hide functions underneath.

Fixes broken compiler when compiled without -O.

Removed derivation of fromConstr.
Added derivation of gunfold.
This avoids some hassle with bottoms and strict datatypes.

Add an extensible-records FAQ

Added the new-born haddock.js to the build process and the documentation.

Add a FAQ about finalizers and stdout/stderr.

-i on its own didn't delete the list of search paths as advertised.