GitLab

This allows us to work-around distribution-specific breakage easily.

I've ruined two trees already by failing to pass --flavour to hadrian.
Let's factor this out so it can be reused during troubleshooting.

Otherwise we are sensitive to libc's buffering strategy.
Similar to the issue fixed in 543dfaab.

Darwin sh doesn't support \e.

Needs to `write` bytes, not str.

This should fix #17861.

* FailablePattern can no longer be created since ab51bee4
  Therefore, Opt_WarnMissingMonadFailInstances has no effect anymore.
* XWrap is no longer used, it was moved to an extension field

Move the gmp configuration to its own binding.

This reverts commit 8924224e
and fixes #17787.

* Remove outdated Note [HsForAllTy tyvar binders] and [Context quantification].
  Since the wildcard refactor 1e041b73, HsForAllTy no longer has an flag
  controlling explicity. The field `hsq_implicit` is gone too.
  The current situation is covered by Note [HsType binders] which is already
  linked from LHsQTyVars.
* Small refactor in CoreLint, extracting common code to a function
* Remove "not so sure about WpFun" in TcEvidence, per Richard's comment
  !852 (comment 223226)
* Use mkIfThenElse in Foreign/Call, as it does exactly what we need.

Before this patch, GHC relied on Ord SrcSpan to identify source elements, by
using SrcSpan as Map keys:

	blackList :: Map SrcSpan ()      -- compiler/GHC/HsToCore/Coverage.hs
	instanceMap :: Map SrcSpan Name  -- compiler/GHC/HsToCore/Docs.hs

Firstly, this design is not valid in presence of UnhelpfulSpan, as it
distinguishes between  UnhelpfulSpan "X"  and  UnhelpfulSpan "Y", but those
strings are messages for the user, unfit to serve as identifiers for source
elements.

Secondly, this design made it hard to extend SrcSpan with additional data.
Recall that the definition of SrcSpan is:

	data SrcSpan =
	    RealSrcSpan !RealSrcSpan
	  | UnhelpfulSpan !FastString

Say we want to extend the RealSrcSpan constructor with additional information:

	data SrcSpan =
	    RealSrcSpan !RealSrcSpan !AdditionalInformation
	  | UnhelpfulSpan !FastString

	getAdditionalInformation :: SrcSpan -> AdditionalInformation
	getAdditionalInformation (RealSrcSpan _ a) = a

Now, in order for  Map SrcSpan  to keep working correctly, we must *ignore* additional
information when comparing SrcSpan values:

	instance Ord SrcSpan where
	  compare (RealSrcSpan r1 _) (RealSrcSpan r2 _) = compare r1 r2
	  ...

However, this would violate an important law:

	a == b  therefore  f a == f b

Ignoring  AdditionalInformation  in comparisons would mean that with
f=getAdditionalInformation, the law above does not hold.

A more robust design is to avoid  Ord SrcSpan  altogether, which is what this patch implements.
The mappings are changed to use RealSrcSpan instead:

	blackList :: Set RealSrcSpan         -- compiler/GHC/HsToCore/Coverage.hs
	instanceMap :: Map RealSrcSpan Name  -- compiler/GHC/HsToCore/Docs.hs

All SrcSpan comparisons are now done with explicit comparison strategies:

	SrcLoc.leftmost_smallest
	SrcLoc.leftmost_largest
	SrcLoc.rightmost_smallest

These strategies are not subject to the law mentioned above and can easily
discard both the string stored in  UnhelpfulSpan  and  AdditionalInformation.

Updates haddock submodule.

* Override rename, unlink, and remove
* Factor out wchar conversion

submodule updates: nofib, haddock

During parsing, GHC collects lexical information about AST nodes and
stores it in a map. It is needed to faithfully restore original source
code, e.g. compare these expressions:

	a =  b
	a  = b

The position of the equality sign is not recorded in the AST, so it must
be stored elsewhere.

This system is described in Note [Api annotations].

Before this patch, the mapping was represented by:

	Map (SrcSpan, AnnKeywordId) SrcSpan

After this patch, the mapping is represented by:

	Map (RealSrcSpan, AnnKeywordId) RealSrcSpan

The motivation behind this change is to avoid using the Ord SrcSpan
instance (required by Map here), as it interferes with #17632 (see the
discussion there).

SrcSpan is isomorphic to  Either String RealSrcSpan,  but we shouldn't
use those strings as Map keys. Those strings are intended as hints to
the user, e.g. "<interactive>" or "<compiler-generated code>", so they
are not a valid way to identify nodes in the source code.

The invariant which allowed the pervious method of splitting the type of
the body to find the type of the elements didn't work in the new
overloaded quotation world as the type can be something like
`WriterT () m a` rather than `Q a` like before.

Fixes #17839

We really need to make sure that these are shared because otherwise GHC
will allocate thousands of identical `TyConApp` nodes.

See #17292

-------------------------
Metric Decrease:
    haddock.Cabal
    T14683
-------------------------

The Overloaded Quotations proposal generalized the type of `lift`
to work over any `Quote`, but not the type of `liftData`, leading
to #17857. Thankfully, generalizing `liftData` is extremely
straightforward.

Fixes #17857.

Somehow I forgot to totally remove `gmpContext` in
d7029cc0. This patch fixes it and adds
some additional comments.

Fixes #17845.

It sometimes happened that occAnal would remove bindings
as dead code by relying on bindings to be in dependency
order. The fix was contributed by SPJ.

As noted in [proposal 0143][proposal] this is supposed to happen in
8.12.

Also fix an incorrect claim in the users guide that -Wstar-is-type is
enabled by default.

[proposal]: https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0143-remove-star-kind.rst

Previously our attempt to cache the cabal store would
`cp cabal-cache ~/.cabal`. However, if the latter already
 existed this meant that we would end up with
~/.cabal/cabal-cache. Not only would this not help caching
but it would exponentially grow the size of ~/.cabal. Not
good!