simdutf-9.4.0

[Joe Locash]

New minor release.

[Joe Locash]

Version 9.4.0

What's Changed

• base64: find '=' while decoding when garbage is accepted
• base64: faster scalar encoding and decoding (supersedes #1037)
• haswell, westmere: cheaper surrogate range test in count_utf16
• base64: let the C++23 span overloads take any byte-like type
• validate_utf16: use the whole register on 256-bit targets
• Fix misleading comments about safety margins and store widths
• lasx: fix comments on the 2-byte UTF-8 to UTF-16 fast path
• icelake: keep converted output in convert_utf8_to_latin1_with_errors
• lasx: fix simd8::any() returning all() semantics
• base64: do not write past the output in the arm64/lsx/lasx tail
• Remove redundant defaulted default constructors
• README: document ABI compatibility constraints
• README: list Microsoft products among simdutf users
• README: list more projects that use simdutf
• cmake: allow building the benchmarks without the tests
• loongarch: enable LASX runtime dispatch with LLVM, keep LSX helpers out of the LASX target
• arm64: check all encodings in one pass in detect_encodings
• Name the Latin1 encoding in to_string
• icelake: stop checking failed encodings in detect_encodings
• base64: make base64_to_binary_safe with accept_garbage independent of buffer size
• base64: count trailing whitespace as read when the SIMD loop consumes all input (TC39)
• generic: one-pass detect_encodings for haswell, westmere, lsx and lasx

Full Changelog: ​v9.3.3...v9.4.0

Fixed at ad5f077f78.