[PATCH] ipa-inline: Improve growth accumulation for recursive calls

[Segher Boessenkool]

Hi!

On Wed, Aug 12, 2020 at 09:03:35PM +0200, Richard Biener wrote:
> On August 12, 2020 7:53:07 PM GMT+02:00, Jan Hubicka <hubicka@ucw.cz> wrote:
> >> From: Xiong Hu Luo <luoxhu@linux.ibm.com>
> >> 523.xalancbmk_r +1.32%
> >> 541.leela_r     +1.51%
> >> 548.exchange2_r +31.87%
> >> 507.cactuBSSN_r +0.80%
> >> 526.blender_r   +1.25%
> >> 538.imagick_r   +1.82%

> >> diff --git a/gcc/cgraph.h b/gcc/cgraph.h
> >> index 0211f08964f..11903ac1960 100644
> >> --- a/gcc/cgraph.h
> >> +++ b/gcc/cgraph.h
> >> @@ -3314,6 +3314,8 @@ cgraph_edge::recursive_p (void)
> >>    cgraph_node *c = callee->ultimate_alias_target ();
> >>    if (caller->inlined_to)
> >>      return caller->inlined_to->decl == c->decl;
> >> +  else if (caller->clone_of && c->clone_of)
> >> +    return caller->clone_of->decl == c->clone_of->decl;
> >>    else
> >>      return caller->decl == c->decl;
> >
> >If you clone the function so it is no longer self recursive, it does
> >not
> >make much sense to lie to optimizers that the function is still
> >recursive.

Like Richard says below (if I understand him right, sorry if not), the
function still *is* recursive in its group of clones.

> >The inlining would be harmful even if the programer did cloning by
> >hand.
> >I guess main problem is the extreme register pressure issue combining
> >loop depth of 10 in caller with loop depth of 10 in callee just because
> >the function is called once.
> >
> >The negative effect is most likely also due to wrong profile estimate
> >which drives IRA to optimize wrong spot.  But I wonder if we simply
> >don't want to teach inlining function called once to not construct
> >large
> >loop depths?  Something like do not inline if caller&callee loop depth
> >is over 3 or so?
> 
> I don't think that's good by itself (consider leaf functions and x86 xmm reg ABI across calls). Even with large loop depth abstraction penalty removal can make inlining worth it. For the testcase the recursiveness is what looks special (recursion from a deeper loop nest level). 

Yes, the loop stuff / register pressure issues might help for the
exchange result, but what about the other five above?


Segher