Tuesday 23rd August 2011
This month's meetings
<< < 2011 / 8 > >>
- Turning on new optimization passes (or other optimization passes we work on e.g. SMS) by default or any of the other optimizations that we do ? How do we go about it ? What's with turning on SMS by default atleast on ARM ?
- Tickets triaging - can we go back and look at some of the "Speed" tickets again to create work items or things to look at and move more stuff into the backlog and do a bit of gardening ? Use some of this time we have for talking through some of the outstanding "speed" tickets for a while. Might be worth trying this on IRC ?
Action Items from this Meeting
- MichaelH to set up the EEMBC runs with SMS flags and Revital to send the list of flags.
- Revital to take summary and write up what next to turn on SMS by default.
- Ramana to collate a list of options to turn on and off and send to Michael.
- Ramana to send out the cunroll example and create a ticket for it.
Action Items from Previous Meeting
- SMS - It needs some work before it can be turned on. We need to find cases where things are degraded . Should we need a cost-model ?
- SMS is currently limited in the data dependence analysis and can't determine dependence distance between memory references unless it can be proven that they are not aliasing. The conservative option ends up limiting the scheduling.
- SMS and auto-inc-dec patches.
- Need more testcases - Turn on testing and benchmarking.
- Richard tried it with EEMBC and there were some regressions in EEMBC. Run it and compare with what we are seeing.
- Where should it go - O2 vs O3 - no-ivopts in some cases improves, compile time ? code size.
- Where do we measure - Android vs no Android.
- Goal should be to turn this on only in the ARM backend.
- Would be good to clean up the tickets and see what comes out of it. -
Might also be interesting to look at the upstream enhancement requests - http://tinyurl.com/3jmx4dx and reconcile some of these with the launchpad database. Discuss them during the call and see if it is still valid ? And try and put a priority on them. Maybe have a look at some of them ahead of time and work through them during the call if they need to come up here.
- Might be interesting to look at ivopts vs no-ivopts on EEMBC (-fivopts and -fno-ivopts)
- sched1 vs no-sched1 - (find more cases where sched-pressure will help and correlate these 2 results.)
- cunroll seems to unroll quite a lot - one case from libav and send out the example. Is it worth playing with
- some of the parameters like --param max-peeled-insns ? Sometimes this will help but other times it really wont.
- Widening multiplies committed upstream
Query from michaelh if http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43862 is helped or not .
- ams_cs : Probably not.
- Backports waiting on approval
- auto-inc-dec patches held up by rtx_costs work and some ARM related changes which have been discussed upstream.
- Basic report on cbuild with automatic benchmarks and catch any regressions earlier.
- Running EEMBC on snowball - bit more unstable than Panda board . Have runs going.
- Trying to investigate errors in some of the automotive tests - and some of the CRC checks.
- SMS work related to libav. There is a kernel that has regressions with libav and looking to see how this is to be resolved.
- SMS inc-dec patch sent upstream again.
- vect-pack-trunc fixes upstream.
- tbh range improvements - Needs backport to 4.6
- VFP moves .
WorkingGroups/ToolChain/Meetings/Archive/2011-08-23 (last modified 2013-08-30 11:47:45)