[dinwwwh]

Requested by dinwwwh · project thread

Summary

A cold start builds the tree in V8's interpreter and baseline code, where each call and each property load goes through a builtin. In the profile of the first request for a method (PATCH in the REST set, 100 routes), loads and calls are most of the time outside split. Two changes in expand's part loop:

• decode(part) became part.includes('%') ? decode(part) : part, so a part without % (nearly all of them) skips the call.
• literals = Math.min(literals, index) became if (index < literals) literals = index, which drops a global load, a property load and a call per waiting route and per param.

Stryker showed a gap the second change exposed: with the comparison mutated to true, literals moves to the last param, and a late-added route sharing a param part with the route before it skipped that param. A case in late-adds.test.ts now covers it (/s/{x}/t/{y} then /s/{x}/t/u, added after a request).

Measurements

Callgrind instructions, --single-threaded, hash seeds 1 and 2 (averaged), main vs this PR:

Stage	main	PR	Change
REST, 3 requests	3.892M	3.757M	-3.5%
REST, 3 requests on 3 methods	5.697M	5.487M	-3.7%
REST, 10 requests	6.615M	6.452M	-2.5%
100 real nouns, 10 requests	6.588M	6.426M	-2.5%
RPC (statics only), 10 requests	420k	420k	same
Router import			+6k (+0.4% of the router's share)

• Size (esbuild --minify, gzip -9): 5068 B to 5076 B minified, 2112 B to 2114 B gzip.
• Steady state: unchanged, since expand only runs while building the tree.

Checks

• pnpm run check green; router tests 73/73 with 100% coverage.
• All fuzzers at 3 seeds: 0 diffs.
• Stryker: 64 survivors (61 on main). The three new ones are equivalent: index <= literals at both checks (assigning an equal value), and part.includes(""), which only sends every part through decode, which checks for % itself. (An earlier count of 54 here came from a run that shared the CPU with benchmarks, where test timeouts counted as kills.)

🤖 Generated with Claude Code

https://claude.ai/code/session_01VVaw7Sutbw6xMTMyDFeNfw

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

1 warning

⚠️ 1 benchmark is dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

✅ 29 untouched benchmarks
⏩ 2 skipped benchmarks¹

---

_{Comparing claude/project-thread-m32ra9 (78440c8) with main (b2a5f7d)}

Footnotes

1. 2 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[pullfrog]

✅ No new issues found.

Reviewed changes

• expand skips decode for %-free parts — part.includes('%') ? decode(part) : part; decode already returns its input unchanged when it holds no %, so the value is identical and only the call frame is saved. Same shape as the existing split branch.
• literals = Math.min(...) → if (index < literals) literals = index (both the wait and param sites) — equivalent for the finite values literals/index ever hold; drops a global Math load, property load, and call per waiting route/param.
• New late-add case — /s/{x}/t/{y} then /s/{x}/t/u queued after a request, asserting /s/1/t/u still yields { x: '1' }.

I verified the two edits are behavior-preserving by reading decode (its %-guard makes the ternary a no-op) and the literals value domain. I also mutated the param-branch guard to always assign (if (index >= 0)) and ran late-adds: it fails exactly as the PR describes, with /s/1/t/u returning params: undefined because the over-long shared-prefix walk skips the {x} param — so the added case is non-vacuous and genuinely pins that regression. Reverted; router suite is 73/73 green and the tree is clean.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}