GitLab

The packages file has an extra component (for "darcs" or "git") so we need
to take that into account when grepping.

We were looking for executable things, but not requiring that they were
files. So we were adding lots of directories (and thus their contents)
to the bindists, massively inflating their size.

We now need to make sure that they have been toException'd.
Also, the RTS doesn't know about the Deadlock exception any more.

If
    GhcWithNativeCodeGen = NO
then we don't build the NCG modules, and we define OMIT_NATIVE_CODEGEN

Otherwise it can't work out how to make in ghc/ any more.

We now set Validating=YES in mk/are-validating.mk rather than on the
commandline. This means that if you build a tree with validate then
just running make in it will use the validate flags.
"make distclean" removes mk/are-validating.mk, putting us back in
standard build mode.

We only call register if we actually have a library to be registered

Now that we don't include any header files in .hc apart from our own,
the cc-options from packages are at best superfluous, so don't pass
them.

We still pass them to .c compilations, although I've just made changes
to Cabal so that cc-options from a .cabal file are not copied into the
InstalledPackageInfo.  Most uses of cc-options in Cabal are clearly
intended to be local to the package, but they were being propagated
everywhere, almost certainly unintentionally.

The way is left open for Cabal to allow packages to specify cc-options
that get propagated in the future, if we find a use case for this.

Some old experimental change leaked in with the parallel GC patches,
it seems

Condsider deriving two overlapping Data declarations for the same type
	deriving instance Data (T A)
	deriving instance Data (T B)

We were getting duplicate bindings for the data-con and tycon auxiliary
bindings for T.  This patch fixes the problem by doing these two decls
the same way as we do con2tag etc.  

(Why might you want such instances; see Trac #2456.)

This fixes nofib

This was broken in the System.FilePath switchover, since filepath's
splitSearchPath doesn't do what we want (it ignores empty
components on Windows, and treats them as "." on Unix).  So we use our
own hand-rolled version, just like GHC.

I'm not 100% sure if this is the right fix, but it seems sensible and
stops break008 segfaulting for me on amd64/Linux.