Peter Smith
2018-06-29 17:51:04 UTC
[TCWG-1424] Investigation into profile and size optimisations
- Completed the pass and got it working on both the old and new pass managers
- Spent quite a bit of time trying to understand how llvm classifies
hot and cold functions, a lot of the details are not documented well
or at all.
- Found another problem with --pgo and lnt when the exec-multisample
option is used. I've applied a local fix to stop the extra runs from
triggering profile collection
- Upstream lnt doesn't seem to collect code size information for a
benchmarks-only run, wrote a script to measure and inject it back.
- Got some benchmark figures on the pass. Difficult to interpret as I
suspect that too many of the benchmarks are too small to give useful
results. There seemed to be some huge regressions when a critical
function stopped getting inlined, but overall performance was
comparable. I need to find some better benchmarks and learn how to
make sense of results.
[TCWG-1368] Buildbot failure investigation
Decided to take a look as an exercise in using the packet.net machines.
Some libfuzzer tests are hanging or taking an extremely long time to
run on aarch64. Reproduced on one of TCWGs packet.net machines. At
early stage of investigation right now as there doesn't seem to be any
obvious answers or easily available diagnostics.
- Completed the pass and got it working on both the old and new pass managers
- Spent quite a bit of time trying to understand how llvm classifies
hot and cold functions, a lot of the details are not documented well
or at all.
- Found another problem with --pgo and lnt when the exec-multisample
option is used. I've applied a local fix to stop the extra runs from
triggering profile collection
- Upstream lnt doesn't seem to collect code size information for a
benchmarks-only run, wrote a script to measure and inject it back.
- Got some benchmark figures on the pass. Difficult to interpret as I
suspect that too many of the benchmarks are too small to give useful
results. There seemed to be some huge regressions when a critical
function stopped getting inlined, but overall performance was
comparable. I need to find some better benchmarks and learn how to
make sense of results.
[TCWG-1368] Buildbot failure investigation
Decided to take a look as an exercise in using the packet.net machines.
Some libfuzzer tests are hanging or taking an extremely long time to
run on aarch64. Reproduced on one of TCWGs packet.net machines. At
early stage of investigation right now as there doesn't seem to be any
obvious answers or easily available diagnostics.