Performance regression with ark 0.5 on deserializing PrimeField elements.

[dkales]

We recently upgraded to ark-* 0.5 and observed noticeable performance degradation on deserialization of large Vec<Primefield>.

This is due to the compiler no longer being able to optimize out the code in https://github.com/arkworks-rs/algebra/blob/48b5348e8998b927fe92f14bac348e0423e017b0/serialize/src/lib.rs#L65C1-L87C2, even though the check function is just Ok(()) for these.

Below is this new block in a flamegraph, this is completely missing when using ark-* 0.4.

The init + teardown of the rayon call here is adding about 10x total runtime in our benchmarks, since we do quite a bit of serialization there (although I am fully aware that doing that many deserializations is probably not a common scenario).
Is this just left up to the compiler to optimize out and for some reason it no longer manages to do so? We can workaround it by disabling these checks completely, but we only would like to disable them for such cases anyway, which seems impossible to do without specialization.

Would be glad to hear any other ideas or workarounds.