Re: PING^1 [PATCH v3] sched: Change no_real_insns_p to no_real_nondebug_insns_p [PR108273]

[Richard Biener <richard.guenther@gmail.com>]

On Thu, Nov 23, 2023 at 4:02 AM Kewen.Lin <linkw@linux.ibm.com> wrote:
>
> on 2023/11/22 18:25, Richard Biener wrote:
> > On Wed, Nov 22, 2023 at 10:31 AM Kewen.Lin <linkw@linux.ibm.com> wrote:
> >>
> >> on 2023/11/17 20:55, Alexander Monakov wrote:
> >>>
> >>> On Fri, 17 Nov 2023, Kewen.Lin wrote:
> >>>>> I don't think you can run cleanup_cfg after sched_init. I would suggest
> >>>>> to put it early in schedule_insns.
> >>>>
> >>>> Thanks for the suggestion, I placed it at the beginning of haifa_sched_init
> >>>> instead, since schedule_insns invokes haifa_sched_init, although the
> >>>> calls rgn_setup_common_sched_info and rgn_setup_sched_infos are executed
> >>>> ahead but they are all "setup" functions, shouldn't affect or be affected
> >>>> by this placement.
> >>>
> >>> I was worried because sched_init invokes df_analyze, and I'm not sure if
> >>> cfg_cleanup can invalidate it.
> >>
> >> Thanks for further explaining!  By scanning cleanup_cfg, it seems that it
> >> considers df, like compact_blocks checks df, try_optimize_cfg invokes
> >> df_analyze etc., but I agree that moving cleanup_cfg before sched_init
> >> makes more sense.
> >>
> >>>
> >>>>> I suspect this may be caused by invoking cleanup_cfg too late.
> >>>>
> >>>> By looking into some failures, I found that although cleanup_cfg is executed
> >>>> there would be still some empty blocks left, by analyzing a few failures there
> >>>> are at least such cases:
> >>>>   1. empty function body
> >>>>   2. block holding a label for return.
> >>>>   3. block without any successor.
> >>>>   4. block which becomes empty after scheduling some other block.
> >>>>   5. block which looks mergeable with its always successor but left.
> >>>>   ...
> >>>>
> >>>> For 1,2, there is one single successor EXIT block, I think they don't affect
> >>>> state transition, for 3, it's the same.  For 4, it depends on if we can have
> >>>> the assumption this kind of empty block doesn't have the chance to have debug
> >>>> insn (like associated debug insn should be moved along), I'm not sure.  For 5,
> >>>> a reduced test case is:
> >>>
> >>> Oh, I should have thought of cases like these, really sorry about the slip
> >>> of attention, and thanks for showing a testcase for item 5. As Richard as
> >>> saying in his response, cfg_cleanup cannot be a fix here. The thing to check
> >>> would be changing no_real_insns_p to always return false, and see if the
> >>> situation looks recoverable (if it breaks bootstrap, regtest statistics of
> >>> a non-bootstrapped compiler are still informative).
> >>
> >> As you suggested, I forced no_real_insns_p to return false all the time, some
> >> issues got exposed, almost all of them are asserting NOTE_P insn shouldn't be
> >> encountered in those places, so the adjustments for most of them are just to
> >> consider NOTE_P or this kind of special block and so on.  One draft patch is
> >> attached, it can be bootstrapped and regress-tested on ppc64{,le} and x86.
> >> btw, it's without the previous cfg_cleanup adjustment (hope it can get more
> >> empty blocks and expose more issues).  The draft isn't qualified for code
> >> review but I hope it can provide some information on what kinds of changes
> >> are needed for the proposal.  If this is the direction which we all agree on,
> >> I'll further refine it and post a formal patch.  One thing I want to note is
> >> that this patch disable one assertion below:
> >>
> >> diff --git a/gcc/sched-rgn.cc b/gcc/sched-rgn.cc
> >> index e5964f54ead..abd334864fb 100644
> >> --- a/gcc/sched-rgn.cc
> >> +++ b/gcc/sched-rgn.cc
> >> @@ -3219,7 +3219,7 @@ schedule_region (int rgn)
> >>      }
> >>
> >>    /* Sanity check: verify that all region insns were scheduled.  */
> >> -  gcc_assert (sched_rgn_n_insns == rgn_n_insns);
> >> +  // gcc_assert (sched_rgn_n_insns == rgn_n_insns);
> >>
> >>    sched_finish_ready_list ();
> >>
> >> Some cases can cause this assertion to fail, it's due to the mismatch on
> >> to-be-scheduled and scheduled insn counts.  The reason why it happens is that
> >> one block previously has only one INSN_P but while scheduling some other blocks
> >> it gets moved as well then we ends up with an empty block so that the only
> >> NOTE_P insn was counted then, but since this block isn't empty initially and
> >> NOTE_P gets skipped in a normal block, the count to-be-scheduled can't count
> >> it in.  It can be fixed with special-casing this kind of block for counting
> >> like initially recording which block is empty and if a block isn't recorded
> >> before then fix up the count for it accordingly.  I'm not sure if someone may
> >> have an argument that all the complication make this proposal beaten by
> >> previous special-casing debug insn approach, looking forward to more comments.
> >
> > Just a comment that the NOTE_P thing is odd - do we only ever have those for
> > otherwise empty BBs?  How are they skipped otherwise (and why does that not
> > work for otherwise empty BBs)?
>
> Yes, previously (bypassing empty BBs) there is no chance to encounter NOTE_P
> when scheduling insns, as for notes in normal BBs, when setting up the head
> and tail, some are skipped (like get_ebb_head_tail), and there are also some
> special handlings remove_notes and unlink_bb_notes to guarantee they are
> gone.  By disabling empty BB bypassing, all empty BBs will be actually
> uniformed as (head == tail && NOTE_P (head)), we have to deal with NOTE_P.

I see.  I expected most of them to be naturally part of another EBB.  So it's
rather a limitation of the head/tail representation.

I wonder if there's a more minimal fix though.  But iff head or tail
of an EBB then I
guess either head or tail has to point to a stmt in said block which necessarily
then means either a debug or note.

Richard.

>
> BR,
> Kewen