[llvmlinux] LLVMLinux Meeting minutes 2013-03-12
behanw at converseincode.com
Mon Mar 18 02:51:43 UTC 2013
On 03/16/13 10:56, Renato Golin wrote:
> Hi Behan,
> First of all, I'd like to thank you guys from working on the Linux
> side of LLVM, this work has long been push back in the agenda of many
> companies as "that will never happen", and you guys are proving all of
> them (and myself) wrong. I really appreciate it.
> Second, some introduction: So far, I'm the only LLVM guy at Linaro and
> am obviously focused on stabilizing the LLVM correctness and
> performance tests, so that we can start work on it without the fear of
> breaking things that are working. This will also reassure your patches
> that they have been properly tested in the open.
> Not just for ARM, but some of the fixes I sent, and will send in the
> next weeks also fixes x86 and PPC test-suites, getting them green as a
> quick and dirty way of spotting regressions. Unfortunately, the other
> back-ends fixes were "collateral damage", so I can't myself focus on
> getting them green as well as ARM (though I'd like to).
Thanks for the intro. It's really great to see Linaro taking an interest
> On 16 March 2013 00:48, Behan Webster <behanw at converseincode.com
> <mailto:behanw at converseincode.com>> wrote:
> o Behan:
> + Helped with clang port of Linaro Android
> Thanks Behan, that was great! We were "under pressure" to have
> something booting in less than a week and your help was fundamental to
> get there. We managed to have a Nexus 7 booting with the LLVM Kernel +
> Dalvik + some libs to show our member companies and our own. I think
> we managed to convince more people that LLVM is serious business.
I followed up with Bero when he was back online after Hong Kong. He
filled me in on your success. :)
> # Linaro team is investing on building Android with
> LLVM and publish their comparison with GCC.
That should have been "investigating building Android with LLVM". Apologies.
> Unfortunately, our benchmarks were tainted with GCC (basically
> everything else). The results we got were extremely close to each
> other, maybe indicating that the part compiled with LLVM didn't have
> much prime time in terms of performance (or we're just lucky that all
> of it was at the same performance level, which is unlikely).
From all the benchmarks I've seen, on average gcc and clang are close
in most instances. There are corner cases where one beats out the other.
But clang still has optimizations left to do even though it's already on
par with gcc. So who knows what will happen in the near future?
> So, not much to report so far, but there is a cross-team plan to get
> Android built with LLVM in the long term. Sadly, not priority number
> one, but something we'll not forget.
Of course. But it's a start which is promising. And I'm sure your work
on optimizing the ARM backend will only help.
> o LLVM 3.3 update
> + Patches that must be sent upstream before LLVM 3.3
> Feel free to copy me on your patches to llvm-commits if they're
> related to ARM in any way, and I'll be sure to review it as soon as
> possible. Though, depending on which area it is, I won't be able to
> approve straight away, but I can ping the right person within ARM or
Any help like this is much appreciated!
behanw at converseincode.com
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the LLVMLinux