Re: Honnor ix86_accumulate_outgoing_args again

[Vladimir Makarov <vmakarov at redhat dot com>]

On 13-10-19 4:30 PM, Jan Hubicka wrote:
> > Jan,
> >
> > Does this seem reasonable to you?
> Oops, sorry, I missed your email. (I was travelling and I am finishing a paper
> now).
> > Thanks,
> > Igor
> >
> > > -----Original Message-----
> > > From: Zamyatin, Igor
> > > Sent: Tuesday, October 15, 2013 3:48 PM
> > > To: Jan Hubicka
> > > Subject: RE: Honnor ix86_accumulate_outgoing_args again
> > >
> > > Jan,
> > >
> > > Now we have following prologue in, say, phi0 routine in equake
> > >
> > > 0x804aa90 1  push   %ebp
> > > 0x804aa91 2  mov    %esp,%ebp
> > > 0x804aa93 3  sub    $0x18,%esp
> > > 0x804aa96 4  vmovsd 0x80ef7a8,%xmm0
> > > 0x804aa9e 5  vmovsd 0x8(%ebp),%xmm1
> > > 0x804aaa3 6  vcomisd %xmm1,%xmm0   <-- we see big stall somewhere here
> > > or 1-2 instructions above
> > >
> > > While earlier it was
> > >
> > > 0x804abd0 1 sub    $0x2c,%esp
> > > 0x804abd3 2 vmovsd 0x30(%esp),%xmm1
> > > 0x804abd9 3 vmovsd 0x80efcc8,%xmm0
> > > 0x804abe1 4 vcomisd %xmm1,%xmm0
> Thanks for analysis! It is a different benchmark than for bulldozer, but
> apparently same case.  Again we used to eliminate frame pointer here but IRS
> now doesn't Do you see the same regression with -fno-omit-frame-pointer
> -maccumulate-outgoing-args?
>
> I suppose this is a conflict in between the push instruction hanled by stack
> engine and initialization of EBP that isn't.  That would explain why bulldozer
> don't seem to care about this particular benchmark (its stack engine seems to
> have quite different design).
>
> This is a bit sad situation - accumulate-outgoing-args is expensive code size
> wise and it seems we don't really need esp with -mno-accumulate-outgoing-args.
> The non-accumulation code path was mistakely disabled for too long ;(
>
> Vladimir, how much effort do you think it will be to fix the frame pointer
> elimination here?
My guess is a week.  The problem I am busy and having some problems with two
small projects right now which I'd like to include into gcc-4.9.

But I think, this still can be fixed on stage2 as it is a PR.

> I can imagine it is a quite tricky case. If so I would suggest adding m_CORE_ALL
> to X86_TUNE_ACCUMULATE_OUTGOING_ARGS with a comment explaining the problem and
> mentioning the regression on equake on core and mgrid on Bulldizer and opening
> an enhancement request for this...
>
> I also wonder if direct ESP use and push/pop instructions are causing so
> noticeable issues, I wonder if we can't "shrink wrap" this into red-zone in the
> 64bit compilation.  It seems that even with -maccumulate-outgoing-args pushing
> the frame allocation as late as possible in the function would be a good idea
> so it is not close to the push/pop/call/ret.