perf: SIMD percent_encode against the character-set bitmap

[anonrig]

Summary

percent_encode was still a scalar find_if plus byte loop on the parse / setter / url_search_params hot paths.

This adds a 16-byte encode kernel, while keeping setter codegen on the original scalar code:

• SSSE3 pshufb looks up the existing 32-byte character_set bitmap (cs[b >> 3] & (1 << (b & 7))). gcc/clang x86-64 uses target("ssse3") so a baseline SSE2 build still gets the kernel, matching parser.cpp. Clean 32-byte pairs are appended in one go.
• NEON uses vqtbl2q_u8 over the same bitmap, also pairing clean 32-byte runs.
• RVV uses indexed loads (vluxei8).
• Dirty windows walk the match mask and emit %XX only for set bits. The suffix output is reserved at 3x remaining length to avoid realloc.

Short / setter paths stay on the original code in the ada.cpp unity TU:

• percent_encode_index is the original inline 8-byte unroll.
• The percent_encode template (used by set_hash / set_search) is the original find_if plus byte loop.
• Allocating percent_encode overloads use that same scalar loop when the remaining suffix is under 48 bytes.
• The SIMD kernel lives in unicode_percent_encode.cpp (its own TU) so it does not change setter inlining. Single-header builds still amalgamate it through ada.cpp.

No per-call nibble LUT (the #1124 regression).

Why this path

Issue #1120 identified percent-encode as the remaining scalar bottleneck. Host/path/query/hash scans are already SIMD; encode was still byte-at-a-time on long query/fragment/form strings.

Performance

Release, g++ 13.3, Xeon. Long query/fragment (~200 B) uses the SIMD suffix. Setter microbenchmarks match main.

CodSpeed

Earlier revisions inlined SIMD into the unity TU and regressed SetHash. Setter templates stay on the scalar loop; the kernel is a separate TU. AVX2 runtime dispatch was dropped after it moved unrelated setter benches.

Correctness

Oracle tests for every official character set, lengths 0–80 plus 96/128/256-byte windows, encode bytes at every offset, non-ASCII, dense punctuation, and append/replace templates.

The percent_encode<append> template is explicitly instantiated so Apple Clang Release still exports the symbol for tests.

No public signature or object-layout change.

Closes #1120.

  

[codspeed-hq]

Merging this PR will not alter performance

✅ 45 untouched benchmarks
🆕 2 new benchmarks
⏩ 4 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
🆕	LongFragment	N/A	7.3 µs	N/A
🆕	LongQuery	N/A	7.3 µs	N/A

---

_{Comparing cursor/simd-percent-encode-5263 (0848284) with main (18ca958)}

Footnotes

1. 4 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. ↩

[codecov]

Codecov Report

❌ Patch coverage is 86.07595% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 63.25%. Comparing base (18ca958) to head (0848284).

Files with missing lines	Patch %	Lines
src/unicode.cpp	50.00%	0 Missing and 6 partials ⚠️
src/unicode_percent_encode.cpp	92.53%	0 Missing and 5 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1230      +/-   ##
==========================================
+ Coverage   63.00%   63.25%   +0.25%     
==========================================
  Files          38       39       +1     
  Lines        7628     7699      +71     
  Branches     3496     3514      +18     
==========================================
+ Hits         4806     4870      +64     
  Misses        749      749              
- Partials     2073     2080       +7     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.