Re: Via LLVM on arm (Was: ANNOUNCE: GHC 7.4.1 Release Candidate 1)
Hi,
On 01/ 2/12 12:46 PM, Joachim Breitner wrote:
Dear Karel,
I hope you don’t mind dragging this discussion onto d-haskell, the right
forum for this, as we are maintaining GHC in a team.
that's cool, let's see if I'm able to post to the list or it's private.
Am Montag, den 02.01.2012, 10:30 +0100 schrieb Karel Gardas:
let me ask, are you GHC debian developer/packager? If so, may I ask you
to switch to use registerised build on recent ARM machines and depend on
LLVM 3.0? i.e. when you use LLVM 3.0, then you are able to build GHC on
ARM in registerised way. GHC is using LLVM then for generating machine
code as there is no ARM NCG yet. I've tried a little bit to blog about
it on ghcarm.wordpress.com.
thanks for the heads up. We are very interested in improving the Haskell
situation on exotic architectures.
That's great to hear since I did not hold this hope seeing debian
provides all the builds unregisterised. So very well and thanks a lot
for your effort!
BTW: if you'd like to be even more motivated to switch to registerised
build on ARM, then have a look at some benchmark data here:
http://ghcarm.wordpress.com/2011/08/07/nofib-benchmarking/
I was just reading
http://ghcarm.wordpress.com/2011/12/24/ghc-7-4-1-release-candidate-1-is-out/
and it seems that ghc uses LLVM by default when compiling for arm – is
that right? So all we have to do is to set
GhcUnregisterised=NO
instead of
GhcUnregisterised=YES
and add llvm-3.0 to the Build-Depends on the architectures "armel" and
"armhf" – is that right?
Yes, that's right but you even don't need to do: GhcUnregisterised=NO --
as this is done by default on ARM now. (i.e. since 7.4.1). Also please
note that I don't have armhf here yet, so this is completely untested. I
do have only armel here (ubuntu 11.04 + 11.10) on i.MX53 and Pandaboard.
Do you expect any regressions in comparison with the unregistered,
non-LLVM-build?
If you are talking about regression from the GHC testsuite point of
view, then unfortunately the testsuite provides different sets of tests
for unregisterised and registerised builds, so this is not directly
comparable. Hmm, I think we might just gain if you, debian GHC team test
ARM registerised build on your set of haskell-based packages and report
directly to GHC trac any regression you see between unregisterised and
registerised builds. That would be absolutely fantastic if you do this
for us.
Also for now, GHCi is not working and the build fails on GHCi annotation
for vector package. I'd like to thank you a lot for submitting all the
proper bug reports to the GHC trac. That's indeed, great.
As per your suggestion on ARM-based buildbot for GHC, I'm going to
devote i.MX53 board to GHC buildbot in the future. I just need to get to
it...
Thanks!
Karel
Reply to: