Tuesday 15th November 2011
This month's meetings
<< < 2011 / 11 > >>
- O3 as a performance theme - Follow-up on discussions after the Summit .
- How do we take this forward ?
- Correctness -
- Bootstrap GCC in O3 regularly ?
- New optimizations (sched-pressure, SMS).
- Other Connect Hacking session topics - Biggest regressions -
partial_partial PRE - http://gcc.gnu.org/PR38785
- Loop unrolling at O2 ? Worth a set of benchmark runs to see what the status with this is ?
886124 Slowdown in vpr on ARM due to using the constant pool over MOVW/MOVT.
- Need for a day-light saving correction ?
- Any other business.
Action Items from this Meeting
Action Items from Previous Meeting
- Unrolling at O2 vs. SMS at O2 - Richard pointed out that SMS is probably better at O2 than unrolling.
- O3 as a performance theme - What are our goals ?
- Write-up at the end of it stating that it is as correct as it is the worth thing.
- What are the kind of performance benefits that we get at the end of this cycle ?
- ICEs are easier since it's large package.
- Bootstraps at O3 ?
- Test rebuilds at O3 from Mathias.
- eglibc at O3 .
- Packages we care about - GCC, GDB, eglibc, qemu, binutils, Qt (C++)
- Benchmarks that have correctness checks.
Nice to have: OpenEmbedded at O3 as a testsuite for correctness ?
- Not bringing the turning on SMS by default under this umbrella.
- Understand and fix some of the top regressions at O3.
- Michael to run benchmarks with unrolling -
- Change times for this call - 9:30 a.m. UTC Tuesdays.
- Ramana to check on the A9 description and mail relevant bits- Revital to check A8 versions.
- Next highest priority ticket for someone coming free to look at would be the vpr regression.
- 64 bit arithmetic regressions with Neon seems to show up. Would be good to look at that next time.
- Ken and Michael to take care of correctness infrastructure issues.
- Ramana to look at one of the major regressions before next time.
WorkingGroups/ToolChain/Meetings/Archive/2011-11-15 (last modified 2013-08-30 11:47:58)