atob and btoa are slower than node

[ry]

https://twitter.com/jarredsumner/status/1499715417773789185

[aapoalas]

Would it be best to have the op_base64_decode and _encode do all the work on Rust side?

ie.
op_base64_decode would return a String so that JS side doesn't have to do all the Uint8Array to Array to string mapping and joining.

op_base64_encode would take a String, check it for invalid characters and return an error if necessary (can ops directly return-throw a DomException? I see a lot of DomExceptionSomething in Rust code so I presume it's okay), convert to u8 slice and pass to base64::encode_config?

[AaronO]

Deno (main) now outperforms Node following PRs I've landed since Friday.
We're over 4x faster on long strings, ~10% faster on short strings.

Node

❯ node -v; node ./cli/bench/deno_common.js
v17.6.0
b64_rt_long:         	n = 100, dt = 2.755s, r = 36/s, t = 27549999ns/op
b64_rt_short:        	n = 1000000, dt = 0.507s, r = 1972387/s, t = 507ns/op

Deno (main 788553b)

❯ ./target/release/deno run --unstable -A ./cli/bench/deno_common.js
b64_rt_long:         	n = 100, dt = 0.626s, r = 160/s, t = 6260000ns/op
b64_rt_short:        	n = 1000000, dt = 0.451s, r = 2217295/s, t = 451ns/op

Conclusions

I would consider this fixed for now. There are further improvements in upcoming deno PRs, I've also submitted 2 PRs upstream to rust-base64: marshallpierce/rust-base64#180 and marshallpierce/rust-base64#179