Tagging subscribers to this area: @dotnet/area-system-numerics
See info in area-owners.md if you want to be subscribed.

However, I would suggest adding a unit test validating that vectors populated from two different NaN patterns (if such a concept exists) still compare as equal.

Such a concept does exist. There are actually 2^n (where n is 23 for float and 52 for double) different variations of NaN. However, all NaN follow the principle of NaN != NaN (regardless of sign or payload) and we are explicitly checking for x == x to get the nan mask.

This combined with our other NaN tests (throughout the various APIs) should be sufficient coverage of correctness in total.

The suggestion was mainly to ensure that somebody didn't try to be overly clever and "optimize" this code:

Vector128<T> result = Vector128.Equals(this, other) | ~(Vector128.Equals(this, this) | Vector128.Equals(other, other));

To this:

Vector128<T> result = Vector128.Equals(this, other) | Vector128.Equals(this.AsTSizedInt(), other.AsTSizedInt()).AsT();

Your unit tests would still pass even though this code is incorrect.

Your unit tests would still pass even though this code is incorrect.

I'd hope no one goes touching this code and ignores the comments in place. I can add a test covering it, but will do so as part of a separate PR (building other NaNs is a bit of a pain and there is no guarantee we won't normalize them regardless, as that's allowed by the IEEE 754 spec).

I'd hope no one goes touching this code and ignores the comments in place.

I'd hope so to. Ultimately this was just a suggestion, and I think you have better judgment than I do on what people may / may not try here. Perhaps I'm a naturally paranoid person. :)

If you think this is overkill I'll defer to your decision.

I do think its good to add as an extra safety measure. I just don't want to re-spin CI and block the bug fix from going in on it ;)

Logged #68710 to track said test enhancement

We're seeing a number of regressions caused by this PR on the 7.0 RC2 vs .NET 6.0 perf report, but we assume they are all by design and expected.