Monday 23rd May 2011
This month's meetings
<< < 2011 / 5 > >>
- Review action items from last meeting
- Review 11.11 blueprints
- Work load vs uncertainty
- Thumb-2 performance
- NEON performance
- Better intrinsics
- Toolchain in Yocto/Android/Ubuntu
- Yocto/Ubuntu as a testsuite
- Good backtracing
- Any that require specifications
- In distributions?
- Deeper validation?
- Top level cross debug?
- Public plan review
- Should we be able to identify Linaro GCC?
Action Items from this Meeting
Action Items from Previous Meeting
- DONE: Michael to shift the call an hour later starting next week
- DONE: Michael to point Ramana at the logs for VFP vs NEON
- DONE: Ramana to check into the bintuils vmov patch and recommend on the upgrade
- DONE: Andrew to check inside CSL re: shrink-wrap clobbers to see what Bernd has done
- DONE: Peter to try today's hwpack and rootfs on vexpress and OMAP3
- DONE: gcc.dg/pr48335-2.c: Michael to check and log: is tracked as PR48493
Dave, Ken, Andrew, Chung-Lin, Richard, Ira, Matthias, Ramana, Revital
- Heard anything from Kate? Matthias: not yet
VFP vs NEON results:
- Are a few differences. Overall in line with previous versions.
- Not sure which is 'better'. Some may be test related.
- Should do an ARM vs Thumb-2 build as well?
Binutils and vmov:
- Ramana recommends binuitls patch for Natty as a SRU
- No packages should be affected as all native builds with no 64 bit targets
- Already in Oneiric
Cortex-A8 branch fix:
- Is now now in Natty and Oneiric
- Bernd is on other things at the moment
- No upstream replies for a month. Bernd will have to ping soon.
QEMU 2011.04 and current hwpack:
- Do work together, but you get more warnings than previousaly due to a kernel change
- Periodic unfortunatly
Blueprint review (Thumb-2):
- memset/memclr: tune the current versions, play with the macros that are currently there
- Driven by the DENBench example
- Also add unaligned support
- Dave: going through SPEC2k6 at the moment, looking at these
- Dynamic length ones are hard
- How do we decide the way to implement these? What is the 'best' way?
- How do we decide if we should align first?
- When you don't know the alignment, NEON on A* handles unaligned much better
- Ramana would rather not unless you know NEON is on
10 work items (10 weeks?) planned, want 40 weeks:
- Ramana will have a look at the efforts
- Revital has no more SMS work planned
- Note also that -mtune=a9 often regresses compared to -mtune=a8 when running on an A9
- Split most of these into at least two - they're not a week, not a month
- Add -fvect-cost-model causing a regression
Michael to talk to PM group about better estimates.
- Is there work to be done?
- Richard doesn't know of any specific area where they need improvement
- Will fix as found
- Review upstream bugs first? Yes
- Michael would like to do examples and proof in the future but it's wishlist
- Ken to write up the specs
- Needs a blueprint
- Needs work items on the hard issues and investigations so that they're captured
Public plan review:
- Middle of next week
- Slides due in a few days
Should we be able to identify Linaro GCC?
- From Matthias
Was a bug reported 780551 where on the FSF branch an ABI changed happened
- Worked around by the reporter by checking the GCC version
- Any other instances of a predefined value in other branches?
- Some use the package string, but no macros
Andrew says 'bad bad bad wrong wrong wrong'
- Correct way to fix is to do an autoconf test instead?
- Continue with the discussion next week
WorkingGroups/ToolChain/Meetings/Archive/2011-05-23 (last modified 2013-08-30 11:47:26)