The typechecking of arrow forms (in GHC 7.6) is known to be bogus, as
described in Trac #5609, because it marches down tuple types that may
not yet be fully worked out, depending on when constraint solving
happens.  Moreover, coercions are generated and simply discarded.  The
fact that it works at all is a miracle.

This refactoring is based on a conversation with Ross, where we
rearranged the typing of the argument stack, so that the arrows
have the form
   a (env, (arg1, (arg2, ...(argn, ())))) res
rather than
   a (arg1, (arg2, ...(argn, env))) res
as it was before.

This is vastly simpler to typecheck; just look at the beautiful,
simple type checking of arrow forms now!

We need a new HsCmdCast to capture the coercions generated from
the argument stack.

This leaves us in a better position to tackle the open arrow tickets
 * Trac #5777 still fails.  (I was hoping this patch would cure it.)
 * Trac #5609 is too complicated for me to grok.  Ross?
 * Trac #344
 * Trac #5333

[ ] is m4 syntax, so we need to use 'test foo' rather than '[ foo ]'.

It's now only used once

pipeLoop now returns the final dflags, rather than having
runPipeline' extract it from the end state.

Signed-off-by: Edward Z. Yang <ezyang@mit.edu>

AnnProvenance now has Functor, Foldable, Traversable instances.

(# is now always a lexeme, even if followed by a symbol.

This fixes certain older GCCs which do not accept link options when assembling or compiling:

  ppc_85xx-gcc: --hash-size=31: linker input file unused because linking not done
  ppc_85xx-gcc: --reduce-memory-overheads: linker input file unused because linking not done

and diagnose this to stderr.

The loop exit condition was testing ${LLC} instead of $1, which was
incorrect. While I'm here, quote the path being tested since it may contain
spaces (e.g. on Windows), and don't search paths that don't exist, which
eliminates un-useful error messages from find.

We only use the unicode characters if the locale supports them.

Simon Marlow authored 12 years ago and ian@well-typed.com committed 12 years ago

I don't think the x86-64 version is quite right, but this ought to be
enough to pass cgrun071.

This code is terrible and needs a complete refactor.  There's a lot of
duplication, and we ought to be specifying the ABI in a much more
abstract way (like LLVM).

Using { } with sed on OS X requires using newlines rather than
semicolons. I've rewritten the sed so that no { } are necessary.

If a file is created by a %.hi rule, and the actual filename isn't
mentioned in the makefiles, then make will treat it as an 'intermediate
file' and delete it when it is finished.

We'd been lucky so far that .hi files weren't actually being built due
to our rules (but rather, as side-effects of the .o rules). However,
when using -dynamic-too to build, we had a rule
    $1/$2/build/%.$$(dyn_osuf): $1/$2/build/%.$$(v_hisuf)
which meant that building a .dyn_o could cause the rule for the
corresponding .hi to be used, and the .hi may then be deleted later on.
This was exacerbated by a bug in GNU make 3.81 which caused make to
enter an infinite loop if running in parallel mode:
    http://lists.gnu.org/archive/html/bug-make/2013-02/msg00020.html

Adding
    .SECONDARY:
would stop make from deleting the intermediate files. However, this
caused make to take a pathologically long time (it appeared to be
live-locked for 2 hours before I killed it) with our build system.

This patch instead creates lines like
    $(eval $(call hi-rule,libraries/base/dist-install/build/Unsafe/Coerce.dyn_hi libraries/base/dist-install/build/Unsafe/Coerce.hi  : %hi: %o  libraries/base/Unsafe/Coerce.hs))
in the .depend files, which results in a rule like
    libraries/base/dist-install/build/Unsafe/Coerce.dyn_hi libraries/base/dist-install/build/Unsafe/Coerce.hi  : %hi: %o  libraries/base/Unsafe/Coerce.hs ;
which, as the files are now all named in the makefiles, means they are
no longer intermediate files so do not get deleted.

Thanks to Daniel Pratt <colorblinddad@gmail.com> for pointing out the failure
and fix.

This means we don't define them multiple times

The hsc2hs, alex and happy options variables are now also
non-way-specific, as the files are shared between all ways.