perf: SIMD percent-decode and form-urlencoded encode/decode

[anonrig]

Percent-encode/decode and application/x-www-form-urlencoded were still scalar after the absolute-URL fast path grew nibble-table SIMD (#1216). Issue #1120 called this out. Prior attempts missed:

• perf: nibble-table SIMD scan for percent_encode_index #1218 / perf: SIMD percent_encode against the character-set bitmap #1230: SIMD percent_encode_index or 16-byte encode classify regresses short setters on CodSpeed
• optimize percent_encode, percent_encode_index, and percent_decode #1124: moved the index scan out of line and built LUTs at runtime

This keeps the inlined 8-byte percent_encode_index and converts the remaining hot tails.

What changed

• Decode: 16-byte SSE2/NEON scans for % (and + in form-urlencoded). Dense %XX runs decode five triplets per 16-byte load (SSSE3 pshufb + maddubs, or NEON tbl).
• Encode tail: consteval nibble tables from each official 32-byte percent-encode set (static_assert that every table matches bit_at). SIMD classify is noinline and only runs when the remaining suffix is at least 48 bytes, so SetHash / UserInfo / the official percent_encode examples stay on scalar bit_at.
• Search params: to_string() appends encoded keys/values in one pass, mapping space to + instead of percent-encode then replace.

Local results

ctest --output-on-failure --test-dir build: 345/345 passed (g++ 13.3, Debug).

Release, g++ 13.3, same inputs vs main at 18ca958 (wall-clock, noisy):

Path	main	this PR
percent_decode of 80×%3A	~216 ns	~98 ns
form_urlencoded_decode (prefix + dense %XX + +)	~199 ns	~117 ns
percent_encode tiny (hello world)	~32 ns	~29 ns
percent_encode ~200-byte fragment	~340 ns	~222 ns
url_search_params::to_string (dense keys/values)	~3.5 µs	~1.6 µs

No public signature or object-layout change. New form_urlencoded_encode / form_urlencoded_encode_append are internal (ada::unicode).

Closes #1120.

  

[codecov]

Codecov Report

❌ Patch coverage is 88.76404% with 20 lines in your changes missing coverage. Please review.
✅ Project coverage is 63.57%. Comparing base (18ca958) to head (9fcf35d).

Files with missing lines	Patch %	Lines
src/unicode.cpp	90.80%	1 Missing and 15 partials ⚠️
include/ada/url_search_params-inl.h	0.00%	0 Missing and 4 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1231      +/-   ##
==========================================
+ Coverage   63.00%   63.57%   +0.56%     
==========================================
  Files          38       38              
  Lines        7628     7741     +113     
  Branches     3496     3516      +20     
==========================================
+ Hits         4806     4921     +115     
- Misses        749      750       +1     
+ Partials     2073     2070       -3     

☔ 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.

[codspeed-hq]

Merging this PR will degrade performance by 5.48%

❌ 12 regressed benchmarks
✅ 33 untouched benchmarks
⏩ 4 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	SetHash	1.7 µs	2.1 µs	-15.58%
❌	Decode	3.6 µs	3.9 µs	-9.38%
❌	Fragment	3.3 µs	3.5 µs	-4.91%
❌	SetHostname	4.5 µs	4.7 µs	-4.33%
❌	SetProtocol	2.6 µs	2.7 µs	-4.32%
❌	SetHref	11 µs	11.5 µs	-4.16%
❌	C0Control	3.3 µs	3.5 µs	-4.11%
❌	Query	3.3 µs	3.5 µs	-4.07%
❌	SpecialQuery	3.3 µs	3.5 µs	-4.07%
❌	SetPort	2.9 µs	3 µs	-3.83%
❌	SetPassword	1.8 µs	1.9 µs	-3.23%
❌	SetPasswordInsertAndClear	1.9 µs	2 µs	-3.03%

Tip

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

---

_{Comparing cursor/simd-percent-encode-8260 (9fcf35d) 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. ↩