The simdutf project, https://github.com/simdutf/simdutf, in addition to having highly optimized unicode conversions has highly optimized/tested/fuzzed base64 encode/decode implementations, importantly including support for WHATWG-forgiving base64 -> https://github.com/simdutf/simdutf?tab=readme-ov-file#base64. simdutf is licensed under a compatible Apache-2.0 / MIT licenses, so we should consider an import. (our current implementation is entirely scalar)
<rdar://problem/129694621>
Yes, I'm working on this right now.
We already adopted simdutf for Uint8Array toBase64. I will apply fromBase64 part too (having some patch already).