ARM linker support!

[Ben Gamari]

Yesterday I finally had a large block of time to devote to the ARM
linker. While I spent a large fraction of the day tracking down an
unrelated bug (ticket #5824, David, your feedback would be greatly
appreciated here), the linker is looking quite good,

  $ inplace/lib/ghc-stage2 -B/opt/ghc/ghc/ghc/inplace/lib --interactive +RTS -N1 
  GHCi, version 7.4.0.20120126: http://www.haskell.org/ghc/  :? for help
  Loading package ghc-prim ... linking ... done.
  Loading package integer-gmp ... linking ... done.
  Loading package base ... linking ... done.
  Prelude> print "hello world!"
  "hello world!"
  Prelude> import System.IO
  Prelude System.IO> writeFile "asdf" "Hello World"
  Prelude System.IO> 
  Leaving GHCi.

Quite exciting! I've pushed the latest state of my tree to Github
(http://www.github.com/bgamari/ghc/arm-ghci-7.4). As you might infer
from the URL, I'm now working on top of the ghc-7.4 branch. I just got
things working a few minutes ago, so the code isn't too well tested, but
must be pretty close to correct to get as far as it does. The code needs
to be cleaned up (whitespace fixed, debug output removed) and there is
still an instruction decoding issue that I need to figure out, but I'd
say things are looking pretty good.

I've been testing against a very recent LLVM commit (git
7750ff1e3cbb87e68f406e6fa7c43a80a61a0ccb, svn commit 149182) which fails
fewer tests than the 3.0 release and seems to emit reasonable code.
While you're looking at patches, have a look at ticket #5814 which adds
support for LLVM's "fence" instruction, added in the 3.0 release. The
patch appears to work but I'm unsure of the exact syncronization
requirements imposed by the Cmm barrier primitive. Also, I currently
always emit "fence", where we should revert to using the
"llvm.memory.barrier" intrinsic used <3.0.

I'd appreciate any comments on any of the above.

Cheers,

- Ben