[Bug target/120941] [16 Regression] 24-40% slowdown of 519.lbm_r on Zen2 and 470.lbm on Zen5 since r16-1644-gaba3b9d3a48a07

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120941

--- Comment #32 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Filip Kastl from comment #30)
> (In reply to H.J. Lu from comment #29)
> > Created attachment 61973 [details]
> > A new patch
> > 
> > Please try this.
> 
> Sadly, this patch doesn't help.  Actually, lbm gets compiled into the same
> binary with and without the patch applied (at least on the current master). 
> If I disassemble the binary, I still see a spill (vbroadcast to %ymm3 and
> then almost immediately vmovapd %ymm3 onto stack).
> 
> Not sure, why the patch doesn't trigger.  It works for the testcase.  I
> currently don't see what's different between the testcase and unreduced lbm.
> In both cases, rrvl attempts to hoist a vbroadcastsd out of a loop.  For the
> testcase, the patch successfully forbids this.  For lbm itself, it fails to
> forbid this.  Perhaps bb->loop_father->latch info is corrupted?  Not sure
> how probable that is.
> 
> I guess that someone with access to lbm must debug this.  I'll look into it.

Please send me .273r.expand and .311r.rrvl files.  I will take a look.