[Bug rtl-optimization/98782] [11/12 Regression] Bad interaction between IPA frequences and IRA resulting in spills due to changes in BB frequencies

[tnfchris at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98782

--- Comment #27 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
(In reply to hubicka from comment #26)
> > It's with LTO, I'll see if non-LTO has the same benefit.  In terms of code-size
> > it looks like it accounts for a 20% increase for binary size, but the hot
> > function shrinks approx 6x. But of course this increase covers the entire
> > binaries and there are other hot functions like GetVirtualPixelsFromNexus that
> > also shrink significantly when specialized.
>
> Thanks.  I will check this.  We mey need tweaking cost model to first
> specialize the functions that reduce more (by doing multiple passes with
> decreasing threshold)

Thanks! It looks like it doesn't need LTO, just plain -Ofast is enough to get
the gain.

> 
> It is really nice to see some real improvmeents from ipa-cp. For years
> the pass did almost nothing on the benchmarks I tested (spec, Firefox,
> clang) but it has changed.  I also see measurable speedups on clang
> binary with ipa-cp on.

Indeed! Some of the gains from these cp's are quite nice!