anderj017 / levenshtein

levenshtein

Making rust as fast as it can go

Improving the performance of the Rust code in Christian's article

• Baseline: 2.077593106
• 6ff99a5: &str.chars().count() -> &str.len(): 2.022744823
• 1dede2f: reuse same vector: 1.880497681
• 6adaf13: build cache without .collect(): 1.7518643919999999

And breaking some of Christian's rules:

• 92d0b82: remove vector bounds checks: 1.475065986
• 23a7604: pre-allocate vector 100 chars: 1.455840814
• 7f47ea9: add notes of some interesting findings
• 73bdea8: use Wrapping to avoid int overflow checks (no real difference)
• 03f9200: eliminate branches (no real difference)

Interesting findings:

• Vector bounds checks are surprisingly expensive
• Rust's opt-in performance improvements look rather ugly :(
• std::cmp::min has poor performance compared to a hand written function (which was included in the Baseline). This is rather curious and concerning.