Skip to content
Ferroni

CSS Performance Optimization Log

This log tracks attempted CSS-path optimizations and whether they were kept. Goal is to avoid re-trying low-value ideas without new evidence.

2026-03-02 - RegSet context quickcheck (rejected)

  • Hypothesis: Add branch-aware/context-aware quickchecks in regset to reduce VM calls for CSS patterns (especially @charset-style alternations with look-behind).
  • Implementation sketch:
    • first-byte mask inference from opcode graph
    • branch-root extraction from push-chain
    • start-only + previous-literal context rules
    • pre-match_at reject in regset_search_body_position_lead
  • Outcome: no meaningful CSS hotspot gain (typically noise / low single-digit percent, unstable across reruns).
  • Cost: large complexity increase in hot-path code and maintenance burden.
  • Decision: removed from main.
  • Revisit only if: we have new profiling proof that failed match_at calls are dominated by context-guardable branches and we can show >=10% stable win.

2026-03-02 - RegSet exact_prefix2 prefilter (rejected)

  • Hypothesis: Add a second-byte prefilter for case-sensitive exact prefixes before calling match_at.
  • Implementation sketch:
    • store optional 2-byte exact prefix per regset entry
    • reject candidates early if str[s+1] does not match
  • Outcome: no stable gain in CSS tokenize hotspots; improvements/regressions were within run-to-run noise.
  • Cost: extra branch/state in regset hot path for little value.
  • Decision: removed from main.
  • Revisit only if: candidate fanout profile shows many false positives where second-byte filtering is dominant.

2026-03-02 - ASCII fastpaths in regexec (kept)

  • Hypothesis: Reduce per-character overhead in VM inner loop for ASCII-heavy CSS input.
  • Implementation:
    • regexec::enclen: ASCII-compatible + ASCII-byte fast return 1
    • regexec::prev_char_head: ASCII-compatible + ASCII-byte fast return s - 1
    • OpCode::WordStar: ASCII/single-byte fast traversal and AltLazy.ascii=true where safe
  • Outcome: small but repeatable improvement in CSS tokenizer benchmarks (roughly low single-digit percent).
  • Decision: kept.

Measurement notes

  • Use direct benchmark binary invocation to avoid mixing with compile-time warnings and unrelated runs:
    • target/release/deps/onig_bench-... --bench scanner/css_tm_15_patterns_lines_tokenize_no_id_rust
    • target/release/deps/onig_bench-... --bench scanner/css_tm_15_patterns_lines_tokenize_with_id_rust
  • Always compare against a clean baseline worktree (52ba6d9 in this round) and rerun at least twice to avoid thermal/noise misreads.