From: "Kewen.Lin" <linkw@linux.ibm.com>
To: Richard Biener <richard.guenther@gmail.com>
Cc: Alexander Monakov <amonakov@ispras.ru>,
Jeff Law <jeffreyalaw@gmail.com>,
Maxim Kuvyrkov <maxim.kuvyrkov@linaro.org>,
GCC Patches <gcc-patches@gcc.gnu.org>,
Richard Sandiford <richard.sandiford@arm.com>,
Jeff Law <jlaw@ventanamicro.com>,
Vladimir Makarov <vmakarov@redhat.com>,
zhroma@ispras.ru, Andrey Belevantsev <abel@ispras.ru>,
Segher Boessenkool <segher@kernel.crashing.org>,
Peter Bergner <bergner@linux.ibm.com>,
Michael Meissner <meissner@linux.ibm.com>,
Alexandre Oliva <oliva@adacore.com>
Subject: Re: PING^1 [PATCH v3] sched: Change no_real_insns_p to no_real_nondebug_insns_p [PR108273]
Date: Thu, 23 Nov 2023 10:36:39 +0800 [thread overview]
Message-ID: <814d768f-b858-f377-0155-fdefed0aa078@linux.ibm.com> (raw)
In-Reply-To: <CAFiYyc1mosr6Kh0MV_w=DuvRaPj_609NVZm-1X+ewRoPoh77oA@mail.gmail.com>
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.
BR,
Kewen
next prev parent reply other threads:[~2023-11-23 3:02 UTC|newest]
Thread overview: 28+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-10-25 2:45 Kewen.Lin
2023-11-08 2:49 ` PING^1 " Kewen.Lin
2023-11-08 9:45 ` Richard Sandiford
2023-11-09 9:10 ` Maxim Kuvyrkov
2023-11-09 17:40 ` Alexander Monakov
2023-11-10 1:57 ` Kewen.Lin
2023-11-10 3:49 ` Jeff Law
2023-11-10 11:25 ` Alexander Monakov
2023-11-10 13:32 ` Richard Biener
2023-11-10 14:18 ` Alexander Monakov
2023-11-10 14:20 ` Richard Biener
2023-11-10 14:41 ` Alexander Monakov
2023-11-15 9:12 ` Kewen.Lin
2023-11-15 9:43 ` Alexander Monakov
2023-11-17 9:03 ` Kewen.Lin
2023-11-17 10:13 ` Richard Biener
2023-11-17 12:55 ` Alexander Monakov
2023-11-22 9:30 ` Kewen.Lin
2023-11-22 10:25 ` Richard Biener
2023-11-23 2:36 ` Kewen.Lin [this message]
2023-11-23 8:20 ` Richard Biener
2023-11-23 9:09 ` Kewen.Lin
2023-12-12 7:02 ` [PATCH draft v2] sched: Don't skip empty block in scheduling [PR108273] Kewen.Lin
2023-11-10 3:49 ` PING^1 [PATCH v3] sched: Change no_real_insns_p to no_real_nondebug_insns_p [PR108273] Jeff Law
2023-11-15 9:01 ` [PATCH] sched: Remove debug counter sched_block Kewen.Lin
2023-12-12 6:17 ` PING^1 " Kewen.Lin
2023-12-20 20:43 ` Jeff Law
2023-12-21 5:46 ` Kewen.Lin
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=814d768f-b858-f377-0155-fdefed0aa078@linux.ibm.com \
--to=linkw@linux.ibm.com \
--cc=abel@ispras.ru \
--cc=amonakov@ispras.ru \
--cc=bergner@linux.ibm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=jeffreyalaw@gmail.com \
--cc=jlaw@ventanamicro.com \
--cc=maxim.kuvyrkov@linaro.org \
--cc=meissner@linux.ibm.com \
--cc=oliva@adacore.com \
--cc=richard.guenther@gmail.com \
--cc=richard.sandiford@arm.com \
--cc=segher@kernel.crashing.org \
--cc=vmakarov@redhat.com \
--cc=zhroma@ispras.ru \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).