Tuesday 7th August 2012
This month's meetings
<< < 2012 / 8 > >>
- Michael Hope
- Ramana Radhakrishnan
- Ulrich Weigand
- Zhenqiang Chen
- Matthew Gretton-Dann
- Review action items from last meeting
Review Todo list here . https://wiki.linaro.org/WorkingGroups/ToolChain/GCCPerfConnectQ212Notes
- Virtual connect sessions
- Vectorizer vs non-vectorizer issues.
- PGO , LTO status.
- uxtb vs and .
Action Items from this Meeting
- michael to follow-up and ping Mans on getting a hold of his benchmark runs.
- Outputs page and benchmark results for cortex-strings published. Not doing anything further for glibc as we have no resource.
- Vector alignments finished upstream . Uli to take forward backports to FSF release branches (4.7 and 4.6). See minutes for more discussion on this topic.
- michaelh to investigate regression with movw/movt patch by hand. Matt to investigate if regressions exist with SPEC2k on FSF trunk with this patch.
- michaelh to look at swapping the 2 virtual connect sessions.
Action Items from Previous Meeting
- Contact Mans and ask for details on his libav performance measurements - Michael - DONE
- Discuss status and next steps for cortex-strings at next meeting - Michael - DONE
- Vector alignments to 8 - Now fixed upstream - Uli DONE
- Create a blueprint to turn fvariable-expansion-in-unroller on when unroll-loops is on by default - Uli - DONE
- Ramana to update the priority list with current status and move it under the toolchain wiki page - DONE
- movw /movt splitting : Inexplicable differences between michaelh's runs and zqchen's runs. michael to look into it further - DONE
- Backporting the vector alignments patch shows some funny performance behaviours which should be investigated. The question of backporting this to Linaro 4.7 arose and it was agreed that given this is an ABI break across time it is important for Linaro GCC 4.7 to be in sync with what other 4.7 tools are released and out there. The ABI issues is across structure layouts of large 128 bit vectors if such a thing exists. This was thought to be rare when this issue was discussed last. However it is important that Linaro GCC remains in sync with what's upstream and having this discussion further upstream is to be noted. Given this situation a backport to the release branches should be investigated upstream.
- Michael sees regressions with the movw/movt splitting patch in a few cases - overall appears to be a wash but needs further investigation.
- Virtual connect
- Vectorizer vs non-vectorizer regressions if any. The one known issue is with the vectorizer cost-model. Is there anything else to be investigated currently ? Benchmark results sent to the list.
- PGO and LTO on ARM - what's the situation with this for ARM. Benchmark results sent to the list and discussion ensued around what might help ?
WorkingGroups/ToolChain/Meetings/Archive/2012-08-07 (last modified 2013-08-30 11:48:01)