[PATCH v2 2/3] x86: support AVX10.1/512
Jiang, Haochen
haochen.jiang@intel.com
Tue Sep 5 08:13:53 GMT 2023
> >> TBD: What about (VEX-encoded) FMA, F16C, AVX-VNNI, AVX-IFMA?
> Should we
> >> make those prereqs as well, along with the various AVX512* ones?
> >> (If so, the latter two would want VEX encodings generally preferred
> >> over EVEX ones when AVX10.1 is explicitly enabled, unless Disp8
> >> scaling helps shortening the resulting encoding.)
> >
> > Forgot to mention in previous thread, AVX-IFMA should not be implied
> > for AVX10.1 because it will not be present in Granite Rapids.
>
> Despite AVX512-IFMA being covered by AVX10? Very odd.
It is odd but it is caused by AVX512-IFMA is introduced first, sadly. We need
to make sure it is correct on HW before simplicity.
But in AVX10.2, the implication is open since we finally have AVX-IFMA shown.
Thx,
Haochen
>
> Jan
More information about the Binutils
mailing list