https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127177

--- Comment #7 from Matthias Kretz (Vir) <mkretz at gcc dot gnu.org> ---
>AVX512 VFPCLASS put the result into the kmask, and there's inter-unit movement
>between kmask and gpr, it's not necessary faster.

kmov has 1 cycle latency, no? And typically we want to branch on scalar isinf,
in which case a ktest instruction without move to gpr should be emitted.
But I agree, we shouldn't change this without benchmarks.

Reply via email to