remlse opened PR #6255 from remlse:separate-lb-to-bindex-gen
to bytecodealliance:main
:
As discussed on Zulip.
This should be a refactor but may have some performance implications.
remlse requested elliottt for a review on PR #6255.
remlse requested wasmtime-compiler-reviewers for a review on PR #6255.
elliottt submitted PR review:
I caught up on the discussion on zulip, and think this is a nice refactoring towards using chaos mode to permute the block order!
jameysharp submitted PR review:
Benchmarks look fine, as I hoped.
- For a small program (bz2), compile time is somewhere in the 0-2% faster range.
- For a medium-sized program (pulldown-cmark), compile time is "1.00x faster" by instructions retired, and not significantly different by CPU cycles.
- For a large program (spidermonkey), compile time is again "1.00x faster" by instructions retired. By CPU cycles my measurements turned out noisier than I'd like, but report 1-5% faster.
So this change is a marginal compile-time performance improvement at all sizes of our benchmark programs. All I was really concerned about is that it's not worse, so this result is great!
jameysharp merged PR #6255.
Last updated: Nov 22 2024 at 16:03 UTC