2025-02-20 Ordinary Meeting Notes

Feb 20, 2025 | RV-LFX Performance Event Sampling TG

Attendees: Beeman Strong

Notes

• Attendees: Beeman, Daniel, Robert, Chun, Bruce, Greg, Hasan

• Slides, video

• Agenda:

  • Resume discussion on late instr sample filtering

  • New perf events for Instr Sampling

  • Sample Collection

• Resume discussion of late filtering

  • For latencies, could make full threshold vs mask & match optional, trace does it this way

    • Triggers make masks optional

    • We can work this out once we start work on the spec

  • If filtering for only long latency insts, could take a long time to get many

    • Can control selection rate by counter value

  • Option to filter by instr type at selection vs retire forks the spec

    • Many RV extensions have optionality

    • Worried that selection filtering may have timing challenges for some implementations

    • Worried that’s too big a difference, hard to work with implementations with big differences

    • Agreed on approach to require filtering by instr type at retire, with the option for an implementation to add the optimization of filtering by type at selection

  • Can we filter to only record speculative insts

    • Yes, by filtering on “cleared by”, select all clear sources

    • Some usages:

      • Sample on all insts, spec and non-spec.  Get an idea of the percent discarded, like TMA’s BAD_SPEC

      • Sample only on spec ops, those cleared by mispredicts.  Thanks to recording of PC of branch/jump that mispredicted, can identify which mispredicts are most costly

  • What’s the benefit of filtering at selection?

    • Narrows the sample population to only select instruction types that have a chance to be recorded

    • Speeds up the process of filling the buffer, one of the chief complaints about instr sampling vs event sampling

• Need to be clear that only 1 inst at a time is sampled

• Need to be clear what can be filtered at selection vs at retire

• Latencies that count until retire may need to do threshold compare some cycle(s) before retire, in order to support precise LCOFI

• Should we accelerate selection after a sample is discarded?

  • Speeds up filling buffer, but alters sample rate

  • Discussion for next time, out of time

Action items

• Feb 13, 2025 - Beeman Strong - Start thread on latency options