Tuesday 10th January 2012
This month's meetings
<< < 2012 / 1 > >>
- Andrew Stubbs
- Åsa Sandahl
- Michael Hope
- Ramana Radhakrishnan
- Ulrich Weigand
- Review action items from last meeting
http://gcc.gnu.org/ml/gcc-patches/2012-01/msg00398.html (A15 branch tuning backport?)
- vld1 on quad loads
- Connect planning - any topics we want to discuss there ? partial-partial PRE sounds like one, looking at the WIP patches is another topic.
Action Items from this Meeting
- Åsa results for Sunspider were on android ? Follow-up with Michael / Android team on a panda / snowball ?
- ACTION on michael to find the testrun for SMS and send it through to everyone else.
Action Items from Previous Meeting
- Ramana to blueprint movw / movt vs constant pool [NOTDONE]
- A15 backport should be done in time for the next release.
- Uli spoke with all three of them -
- RichardS will followup on sched-pressure upstream in any case. We need to support with any benchmarking activities on ARM.
- A9 scheduler changes is pending because changes got stuck with genautomata issues - autoincdec changes and further benchmarking needs to be done.
- Running SPEC, EEMBC and ubutest with the sched-pressure. Action on Michael to document the benchmarks and ubutest for a number of the cases.
- Uli to look into sched-pressure and specify flags that we need testing on - fsched-pressure and send them to Michael and Åsa.
- Mem rtx_costs has an interaction with the vld1 rework that we've been talking about and needs to be scheduled accordingly. Go ahead with current mem rtx_costs and find the extra write back latency.
- Handoff vld1 to Uli so that he can start on that . We can get the vld1 started after mem rtx_costs.
- Åsa to send out about Sunspider results.
- SMS - Memory dependencies patch was waiting on Revital's patch for register pressure estimation for SMS. Richard remembers a degradation which might be due to the pipeline description and vmla and bypasses that it should have. vmla . SMS + register pressure appears still to have some regressions. ACTION on michael to find the testrun and send it through to everyone else. Maybe look at Revital's patch a little bit more.
Redoing the benchmarks on the fwprop.c change would be interesting. Lets get some numbers but with EEMBC Denbench, libav micro-benchmarks, EEMBC v1 v2. What about fate ? http://libav.org/fate.html
- Future work - ivopts analysis and loop invariants analysis.
- Revital's work - SMS + Register pressure posted to gcc-patches - SMS specific parts have been approved by Ayal . We might as well do the backport. SMS + register pressure appears still to have some regressions. ACTION on michael to find the testrun and send it through to everyone else. Maybe look at Revital's patch a little bit more. Lower priority than the other patches outstanding.
- Backport ICE for MODE_CC and SMS and close out that issue that was raised in LP879725.
Revital pointed out that http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49789 . Revital's long term goal is to turn on SMS at O3 but there would still probably be some regressions so we've got to decide how to proceed further.
- Ira's patch - vectorizer regressions at O3. The major regression was with fsched-pressure and follow-through with the vectorizer. All the other issues were related to alignment and peeling for alignment issues. There are 2 patches
- Disable peeling for alignment of small loops. Benchmark results ? ACTION: michaelh. Are the heuristics chosen the best ? Can we choose better heuristics for this ? Should we be more aggressive and disable peeling unless the loop is a lot bigger.
- Can we teach the vectorizer to align stack variable half-way ?
WorkingGroups/ToolChain/Meetings/Archive/2012-01-10 (last modified 2013-08-30 11:50:00)