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::minhas poor performance compared to a hand written function (which was included in the Baseline). This is rather curious and concerning.