[Bug rtl-optimization/125876] [13/14/15/16/17 Regression] x86: register-source vmovddup spilled to the stack instead of using the register form

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125876

--- Comment #11 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to vekumar from comment #9)
> (In reply to Sarvesh Chandra from comment #8)
> > Sure Richard, would an RFC to gcc-patches be apt for this issue?
> 
> Can u check if latency wise vunpcklpd is better than vmovddup for register
> variants, based on that you can post patch against this PR.

According to Agner Fog's table, vmovddup is same as vunpcklpd for thougput and
latency on Intel platform, but looks like Zen4/zen5 has better throuput for
vunpcklpd.