perf(decode): skip ,/; pre-scan for large mappings

[Boshen]

Summary

Replaces the unconditional O(n) ,/; pre-scan in decode_mapping with a size-gated estimate, factored into a new estimate_token_capacity helper:

• mapping.len() < 256 — keep the exact ,/; count. Cheap on short mappings, and getting the capacity exact lets the trailing Vec::into_boxed_slice in decode_from_string skip its shrink-realloc (otherwise the realloc dominates per-iteration cost on tiny benchmarks).
• mapping.len() ≥ 256 — use the O(1) heuristic mapping.len() / 4 + 1. Realistic sourcemaps average ~4-5 bytes per segment, so the estimate is close to typical. Vec::push handles under-estimates via geometric growth; into_boxed_slice shrinks any over-allocation back to exact size, so RSS is unchanged.

No unsafe.

CodSpeed

Net: +3.61% improvement.

Wins (on the decode hot path — the algorithmic target)

Benchmark	BASE	HEAD	Δ
parse[real_xlarge]	1.4 ms	1.3 ms	+10.59%
parse[real_large]	50.5 µs	46.5 µs	+8.56%
parse[real_medium]	14.9 µs	14.5 µs	+2.74%

Regressions (binary-layout artifacts)

Benchmark	BASE	HEAD	Δ
build_single	7.0 µs	7.1 µs	-1.23%
lookup_table[real_medium]	1.4 µs	1.5 µs	-1.97%

Both regressed functions live in src/sourcemap_builder.rs and src/sourcemap.rs respectively, not in src/decode.rs — they aren't algorithmically touched. Inserting any code into decode.rs shifts the .text section and moves other functions to slightly different addresses, which changes cache behavior in CodSpeed's deterministic simulator. Verified by trying multiple variants (inline if-else, threshold 128 vs 256, helper-fn extraction) — the regression set is the same regardless.

The absolute delta on the regressed benchmarks is ~80ns and ~30ns respectively. The xlarge win is ~150µs (3,750× larger). Net real-world impact on bundler workloads is overwhelmingly positive.

RSS

Measured at 500 parsed xlarge maps:

• main: 223,200 KB
• perf(decode): skip ,/; pre-scan for large mappings #343: 223,200 KB (no change — into_boxed_slice shrinks Vec to exact size on every parse)

[codspeed-hq]

Merging this PR will improve performance by 3.61%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 3 improved benchmarks
❌ 2 regressed benchmarks
✅ 11 untouched benchmarks
⏩ 5 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	build_single	7 µs	7.1 µs	-1.23%
❌	lookup_table[real_medium]	1.4 µs	1.5 µs	-1.97%
⚡	parse[real_large]	50.5 µs	46.5 µs	+8.56%
⚡	parse[real_medium]	14.9 µs	14.5 µs	+2.74%
⚡	parse[real_xlarge]	1.4 ms	1.3 ms	+10.59%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf/decode-skip-prescan (4dabc46) with main (f9a6387)}

Footnotes

1. 5 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. ↩