perf: speed up VLQ mappings decode

[Boshen]

Problem

Profiling parse of a ~115 KB sourcemap (samply + inline-expanded symbolication on the criterion bench) attributed ~40% of parse time to the VLQ mappings decode loop: per-byte delimiter compares before the table lookup, per-value array bounds checks, shift bookkeeping for every value, and a branchy sign decode. (The other big buckets — serde_json escaped-string unescaping and allocator traffic — are inherent to the JSON layer; the previously suspected token-vec shrink measured only ~1.2%.)

Changes

Rework parse_vlq_segment_into internals; its signature and all validation in decode_mapping are untouched:

• Fold , / ; into the decode table as negative sentinels → one table load per byte instead of two compares plus a load. Bytes outside the base64 alphabet map to 63, reproducing the historical lenient behavior exactly (they already decoded as payload 31 + continuation, i.e. like /).
• Single-byte fast path per value (small deltas dominate real mappings: ~80% of values are one byte), skipping continuation/shift bookkeeping and the overflow guard.
• Branchless sign-magnitude decode — delta signs are poorly predictable in real mappings, so the old data-dependent branch mispredicted often.

Behavior is locked, not just tested

• New unit tests written against the old implementation first: error precedence (VlqLeftover beats field validation, column check beats size check), exact >5-field counts, invalid-char ≡ / equivalence.
• Differential harness over all 142 .map/.json files under tests/ (fixtures + tc39 resources): decoded tokens, re-encoded bytes, and error variants are byte-identical old vs new.
• tc39 spec suite: identical pass/fail sets (86 passed, same 11 pre-existing failures).
• Full test suite + clippy --all-targets --all-features clean.

Benchmarks (criterion vs saved baseline, M-series)

bench	before	after	Δ
parse/real_xlarge	126.8 µs	109.7 µs	−13.6%
parse/real_large	3.82 µs	3.23 µs	−15.3%
parse/real_medium	592 ns	547 ns	−7.2%
smoke (inline parse)	548 ns	509 ns	−12.3%
serialize/* , lookup_table/* , concat/*			parity

Companion PR (independent, encoder side): fixes a latent unreserved first-token write in serialize_mappings.

[codspeed-hq]

Merging this PR will improve performance by 1.6%

⚠️ 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

⚡ 6 improved benchmarks
✅ 10 untouched benchmarks
⏩ 7 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	parse[real_small]	11.9 µs	11.6 µs	+2.42%
⚡	from_json_string_inline	14.4 µs	14.1 µs	+2.24%
⚡	parse[real_xlarge]	1.3 ms	1.2 ms	+1.43%
⚡	parse[real_medium]	14.7 µs	14.5 µs	+1.38%
⚡	parse[real_large]	47.2 µs	46.6 µs	+1.14%
⚡	build_single	5.8 µs	5.8 µs	+1.01%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf-vlq-decode (a193fad) with main (1423083)}

Footnotes

1. 7 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. ↩