On Mon, May 15, 2023 at 5:03 PM Aldy Hernandez <al...@redhat.com> wrote:
>
>
>
> On 5/15/23 13:08, Richard Biener wrote:
> > On Mon, May 15, 2023 at 12:35 PM Aldy Hernandez <al...@redhat.com> wrote:
> >>
> >> <tldr>
> >> We can now have int_range<N, RESIZABLE=false> for automatically
> >> resizable ranges. int_range_max is now int_range<3, true>
> >> for a 69X reduction in size from current trunk, and 6.9X reduction from
> >> GCC12. This incurs a 5% performance penalty for VRP that is more than
> >> covered by our > 13% improvements recently.
> >> </tldr>
> >>
> >> int_range_max is the temporary range object we use in the ranger for
> >> integers. With the conversion to wide_int, this structure bloated up
> >> significantly because wide_ints are huge (80 bytes a piece) and are
> >> about 10 times as big as a plain tree. Since the temporary object
> >> requires 255 sub-ranges, that's 255 * 80 * 2, plus the control word.
> >> This means the structure grew from 4112 bytes to 40912 bytes.
> >>
> >> This patch adds the ability to resize ranges as needed, defaulting to
> >> no resizing, while int_range_max now defaults to 3 sub-ranges (instead
> >> of 255) and grows to 255 when the range being calculated does not fit.
> >>
> >> For example:
> >>
> >> int_range<1> foo; // 1 sub-range with no resizing.
> >> int_range<5> foo; // 5 sub-ranges with no resizing.
> >> int_range<5, true> foo; // 5 sub-ranges with resizing.
> >>
> >> I ran some tests and found that 3 sub-ranges cover 99% of cases, so
> >> I've set the int_range_max default to that:
> >>
> >> typedef int_range<3, /*RESIZABLE=*/true> int_range_max;
> >>
> >> We don't bother growing incrementally, since the default covers most
> >> cases and we have a 255 hard-limit. This hard limit could be reduced
> >> to 128, since my tests never saw a range needing more than 124, but we
> >> could do that as a follow-up if needed.
> >>
> >> With 3-subranges, int_range_max is now 592 bytes versus 40912 for
> >> trunk, and versus 4112 bytes for GCC12! The penalty is 5.04% for VRP
> >> and 3.02% for threading, with no noticeable change in overall
> >> compilation (0.27%). This is more than covered by our 13.26%
> >> improvements for the legacy removal + wide_int conversion.
> >
> > Thanks for doing this.
> >
> >> I think this approach is a good alternative, while providing us with
> >> flexibility going forward. For example, we could try defaulting to a
> >> 8 sub-ranges for a noticeable improvement in VRP. We could also use
> >> large sub-ranges for switch analysis to avoid resizing.
> >>
> >> Another approach I tried was always resizing. With this, we could
> >> drop the whole int_range<N> nonsense, and have irange just hold a
> >> resizable range. This simplified things, but incurred a 7% penalty on
> >> ipa_cp. This was hard to pinpoint, and I'm not entirely convinced
> >> this wasn't some artifact of valgrind. However, until we're sure,
> >> let's avoid massive changes, especially since IPA changes are coming
> >> up.
> >>
> >> For the curious, a particular hot spot for IPA in this area was:
> >>
> >> ipcp_vr_lattice::meet_with_1 (const value_range *other_vr)
> >> {
> >> ...
> >> ...
> >> value_range save (m_vr);
> >> m_vr.union_ (*other_vr);
> >> return m_vr != save;
> >> }
> >>
> >> The problem isn't the resizing (since we do that at most once) but the
> >> fact that for some functions with lots of callers we end up a huge
> >> range that gets copied and compared for every meet operation. Maybe
> >> the IPA algorithm could be adjusted somehow??.
> >
> > Well, the above just wants to know whether the union_ operation changed
> > the range. I suppose that would be an interesting (and easy to compute?)
> > secondary output of union_ and it seems it already computes that (but
> > maybe not correctly?). So I suggest to change the above to
>
> union_ returns a value specifically for that, which Andrew uses for
> cache optimization. For that matter, your suggestion was my first
> approach, but I quickly found out we were being overly pessimistic in
> some cases, and I was too lazy to figure out why.
>
> >
> > bool res;
> > if (flag_checking)
> > {
> > value_range save (m_vr);
> > res = m_vr.union_ (*other_vr);
> > gcc_assert (res == (m_vr != save));
> > }
> > else
> > res = m_vr.union (*other_vr);
> > return res;
>
> With your suggested sanity check I chased the problem to a minor
> inconsistency when unioning nonzero masks. The issue wasn't a bug, just
> a pessimization. I'm attaching a patch that corrects the oversight
> (well, not oversight, everything was more expensive with trees)... It
> yields a 6.89% improvement to the ipa-cp pass!!! Thanks.
>
> I'll push it if it passes tests.
Tests passed. Pushed patch.
I've also pushed the original patch in this email. We can address
anything else as a follow-up.
Thanks for everyone's feedback.
Aldy
