[Bug target/82735] _mm256_zeroupper does not invalidate previously computed registers

["crazylht at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82735

--- Comment #16 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #15)
> (In reply to Hongtao.liu from comment #14)
> > (In reply to Hongtao.liu from comment #12)
> > > (In reply to Jakub Jelinek from comment #10)
> > > > Last touched in PR99563.
> > > > I guess for the explicit user vzeroupper we need to add the clobbers/sets
> > > > earlier than in the vzeroupper pass, but ideally in a way that doesn't force
> > > > save/restore of registers that aren't really needed in the function.
> > > 
> > > Yes, if we want to add the clobbers/sets earlier(than CSE1), vzeroupper pass
> > > should be able to remove those unnecessary clobbers/sets.
> > 
> > Correct typo, add the clobbers/sets earlier(than RA)
> 
> I'm trying to add a post_reload splitter to add CLOBBERS of xmm to
> vzeroupper so that LRA knows vzeroupper will kill those xmm registers, then
> in pass_vzeroupper, transform those CLOBBERS to SET (xmm, xmm), it will
> benifit post_reload CSE which allow lower 128bits to cross vzeroupper, then
> in post_reload split2, drop those SETs, it's safe since there's no CSE
> between split2 and split3, problem is there's no update for data flow info
> between split2 and pro_and_epilog which mean even i manually drop those
> SETS, xmm6-xmm15 are still marked as used which causes redudant save and
> restore under 64-bit MSabi.
> 
> I'm thinking of adding a target_hook for updating df info just in the
> begenning of pass_pro_and_epilogue, the default behavior of the target_hook
> is doing nothing , and in i386 backend, df_analyse is called only under
> TARGET_AVX && cfun->machine->has_explicit_vzeroupper.

Oh, regs_ever_live isn't recomputed, that's why even i manually drop those SETs
is post_reload splitter, but xmm6 to xmm15 is still marked as live.

/* After reload, some ports add certain bits to regs_ever_live so
     this cannot be reset.  */