[Sonicadvance1]

This is a relatively new Linux feature from 2017. Supported on AMD CPUs since Zen 4 (2022) and Intel CPUs since Ivy Bridge (2012). So it's been a long time coming for AMD to catch up. I had been slowly working towards implementing this but finally pushed it over the edge. This will slow down our CPUID emulation slightly but having it supported is worth it.

Some test results from my AmpereOneA.
Before:

Test, Total Cycles, Iterations, Cycles Average, Iter Time Average, iterations/Second
cpuid, 555942960, 53355000, 10.42, 10.42, 95972075.98

Before with const-prop disabled:

Test, Total Cycles, Iterations, Cycles Average, Iter Time Average, iterations/Second
cpuid, 701291860, 23713500, 29.57, 29.57, 33814024.31

After:

Test, Total Cycles, Iterations, Cycles Average, Iter Time Average, iterations/Second
cpuid, 606550700, 23713500, 25.58, 25.58, 39095660.10

TR7995WX for fun:

Test, Total Cycles, Iterations, Cycles Average, Iter Time Average, iterations/Second
cpuid, 1713884925, 23713500, 72.27, 28.95, 34538389.08

As seen most of the performance hit comes from the const-prop removal, but we actually manage to claw back a bit of the performance. The implementation is still faster than the threadripper which is kind of fun. Tested this with Proton with its CPUID emulation to ensure its working properly there but the unittest does what WINE does plus more.

[Sonicadvance1]

I'll finish this post-release.

[Sonicadvance1]

Alright, remaining bug actually wasn't anything spooky. I just forgot to say that the CPUID OpcodeDispatcher actually sets RIP.