RISE RP005 QEMU weekly report 2024-09-11

Work completed since last report

Results, statistics and graphs for this week.

WP2
- Added optimizations to vle.v and vse.v helper functions that improves performance of memcpy with odd numbers of bytes.
- Started optimization of TranslationBlock (TB) access functions.
  - RVV calls to TB access functions are in do_vsetvl and do_vsetivli.
  - cpu_get_tb_cpu_state accounts for 3-24% of execution time with memcpy, helper_lookup_tb_ptr accounts for 6-46% of execution time with memcpy, with the larger values being for small data sizes and small vectors (all data for latest optimized QEMU).
  - our initial approach of inlining these calls didn't produce significant improvements, we continue to work on this.
- Created initial optimization for vsetvl instructions, given that profiling shows that its helper accounts for 2-12% of execution time in memcpy.
  - no performance data let.

Work planned for the coming two weeks

WP2
- Explore new approaches to optimizing TranslationBlock access.
- Measure effectiveness of vsetvl optimization and extend this work.
- Continue optimization of the vle.v and vse.v helper functions:
Other
- Following discussion with RISE, we will switch to weekly calls.
- Paul Walmsley to send a follow up email about proposed tasks/priorities.

Planned absences

Jeremy Bennett will be away at conferences 12-16 September
- includes GNU Tools Cauldron 14-16 September
Paolo Savini will be on vacation 20-24 and 27 September