[Bug string/27457] vzeroupper use in AVX2 multiarch string functions cause HTM aborts

["hjl.tools at gmail dot com" <sourceware-bugzilla@sourceware.org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=27457

--- Comment #19 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Richard Biener from comment #17)
> (In reply to H.J. Lu from comment #11)
> > (In reply to Florian Weimer from comment #9)
> > > (In reply to rguenther from comment #7)
> > > > On Mon, 1 Mar 2021, fweimer at redhat dot com wrote:
> > > > > The sources do not seem to show the selection logic.
> > > > > __memset_avx2_unaligned_rtm seems to simply have dropped the VZEROUPPER
> > > > > instruction, which can't be good for general application performance.
> > > > 
> > > > The intent was to use %ymm16+ only which does not cause any transition
> > > > penalty even w/o vzeroupper.
> > > 
> > > I still saw %ymm0 usage in the disassembly, if I recall correctly. And for
> > > AVX2, there isn't much choice. I didn't try to reverse-engineer the
> > > corresponding IFUNC selector.
> > 
> > At function exit, there is
> > 
> >   f3:	0f 01 d6             	xtest  
> >   f6:	74 04                	je     fc <__memset_avx2_unaligned_erms_rtm+0x5c>
> >   f8:	c5 fc 77             	vzeroall 
> >   fb:	c3                   	ret    
> >   fc:	c5 f8 77             	vzeroupper 
> >   ff:	c3                   	ret    
> 
> Btw, if the 'je' mispredicts to the vzeroupper case inside an transaction
> will the speculative execution of vzeroupper abort the transaction or
> does it only abort the transaction when retired?

My branch includes RTM test:

  for (i = 0; i < loop; i++)
    {
      if (_xbegin() == _XBEGIN_STARTED)
        {
          failed |= function ();
          _xend();
        }
      else
        {
          failed |= function ();
          ++naborts;
        }
    }

It passes with xtest + je.

-- 
You are receiving this mail because:
You are on the CC list for the bug.