Avoid using PDEP and PEXT on AMD Zen

[Techjar]

For some unknown reason PDEP and PEXT are ridiculously slow on AMD Zen architecture, which is making DoubleToSingle an extremely costly operation on these CPUs, enough to cause severe slowdown in games. Unsure whether the usage in VertexLoaderX64 had any noticeable impact, but it's also been changed as a precautionary measure.

This should fix https://bugs.dolphin-emu.org/issues/11964

[delroth]

https://uops.info/html-instr/PEXT_R64_R64_R64.html confirms the fact that PEXT is slow on Zen / Zen 2. PDEP is similarly slow, we should make sure we don't emit these if we have a better alternative.

[andreas-abel]

https://uops.info/html-instr/PEXT_R64_R64_R64.html confirms the fact that PEXT is slow on Zen / Zen 2.

And this is actually just the best case. In the worst case, the latency is more than 289 cycles: https://twitter.com/uops_info/status/1202950247900684290

[JMC47]

Hopefully this is the cause of users on AMD machines suddenly saying Dolphin is really slow in partiuclar cases...

[MayImilae]

PDEP and PEXT latency issues are fixed on Zen 3. That said, since the family was bumped (Zen 3 is now family 25), and this code is exclusively family = 23, Zen 3 is already using the fast path for PDEP and PEXT. Basically, if anyone else spots that Zen 3 has fixed its PDEP and PEXT issues, don't worry, no changes are required.

It might be wise to update the comment though, and mention it is only Zen 1-2.

[shuffle2]

I've been going over the CPUDetect code recently and realized that some dolphin jit code checks bBMI2 while other code checks bFastBMI2. Is that intentional (for some paths on family 17h to still be using BMI2) or would it make sense to remove bFastBMI2 and have bBMI2 forced to false on family 17h?

edit: oh nvm, realized the issue only impacts a subset of BMI2.