* [PATCH 06/11] haifa-sched: Allow for NOTE_INSN_DELETED at start of epilogue
@ 2023-10-17 20:48 Alex Coplan
2023-10-19 14:54 ` Jeff Law
0 siblings, 1 reply; 3+ messages in thread
From: Alex Coplan @ 2023-10-17 20:48 UTC (permalink / raw)
To: gcc-patches
Cc: Richard Sandiford, Jim Wilson, Michael Meissner, Jeff Law,
Vladimir Makarov
[-- Attachment #1: Type: text/plain, Size: 603 bytes --]
haifa-sched.cc:remove_notes asserts that it lands on a real (non-note)
insn after advancing past NOTE_INSN_EPILOGUE_BEG, but with the upcoming
post-RA aarch64 load pair pass enabled, we can land on
NOTE_INSN_DELETED.
This patch adjusts remove_notes to remove these if they occur at the
start of the epilogue instead of asserting.
Bootstrapped/regtested as a series on aarch64-linux-gnu, OK for trunk?
gcc/ChangeLog:
* haifa-sched.cc (remove_notes): Allow for NOTE_INSN_DELETED at
the start of the epilgoue, remove these.
---
gcc/haifa-sched.cc | 11 +++++++++++
1 file changed, 11 insertions(+)
[-- Attachment #2: 0006-haifa-sched-Allow-for-NOTE_INSN_DELETED-at-start-of-.patch --]
[-- Type: text/x-patch, Size: 691 bytes --]
diff --git a/gcc/haifa-sched.cc b/gcc/haifa-sched.cc
index 8e8add709b3..9f45528fbe9 100644
--- a/gcc/haifa-sched.cc
+++ b/gcc/haifa-sched.cc
@@ -4249,6 +4249,17 @@ remove_notes (rtx_insn *head, rtx_insn *tail)
&& NOTE_KIND (next) == NOTE_INSN_BASIC_BLOCK
&& next != next_tail)
next = NEXT_INSN (next);
+
+ /* Skip over any NOTE_INSN_DELETED at the start of the epilogue.
+ */
+ while (NOTE_P (next)
+ && NOTE_KIND (next) == NOTE_INSN_DELETED)
+ {
+ auto tmp = NEXT_INSN (next);
+ delete_insn (next);
+ next = tmp;
+ }
+
gcc_assert (INSN_P (next));
add_reg_note (next, REG_SAVE_NOTE,
GEN_INT (NOTE_INSN_EPILOGUE_BEG));
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH 06/11] haifa-sched: Allow for NOTE_INSN_DELETED at start of epilogue
2023-10-17 20:48 [PATCH 06/11] haifa-sched: Allow for NOTE_INSN_DELETED at start of epilogue Alex Coplan
@ 2023-10-19 14:54 ` Jeff Law
2023-10-19 17:52 ` Alex Coplan
0 siblings, 1 reply; 3+ messages in thread
From: Jeff Law @ 2023-10-19 14:54 UTC (permalink / raw)
To: Alex Coplan, gcc-patches
Cc: Richard Sandiford, Jim Wilson, Michael Meissner, Vladimir Makarov
On 10/17/23 14:48, Alex Coplan wrote:
> haifa-sched.cc:remove_notes asserts that it lands on a real (non-note)
> insn after advancing past NOTE_INSN_EPILOGUE_BEG, but with the upcoming
> post-RA aarch64 load pair pass enabled, we can land on
> NOTE_INSN_DELETED.
>
> This patch adjusts remove_notes to remove these if they occur at the
> start of the epilogue instead of asserting.
>
> Bootstrapped/regtested as a series on aarch64-linux-gnu, OK for trunk?
>
> gcc/ChangeLog:
>
> * haifa-sched.cc (remove_notes): Allow for NOTE_INSN_DELETED at
> the start of the epilgoue, remove these.
One could argue that the pass should have actually deleted the insn
rather than just turned it into a NOTE_INSN_DELETED. Is there some
reason that's not done? A NOTE_INSN_DELETED carries no useful information.
> + /* Skip over any NOTE_INSN_DELETED at the start of the epilogue.
> + */
Don't bring the close comment down to a new line. If it fits, but it on
the last line of the actual comment. Otherwise bring down part of
comment so that we don't have the close comment on a line by itself.
Jeff
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH 06/11] haifa-sched: Allow for NOTE_INSN_DELETED at start of epilogue
2023-10-19 14:54 ` Jeff Law
@ 2023-10-19 17:52 ` Alex Coplan
0 siblings, 0 replies; 3+ messages in thread
From: Alex Coplan @ 2023-10-19 17:52 UTC (permalink / raw)
To: Jeff Law
Cc: gcc-patches, Richard Sandiford, Jim Wilson, Michael Meissner,
Vladimir Makarov
Hi Jeff,
On 19/10/2023 08:54, Jeff Law wrote:
>
>
> On 10/17/23 14:48, Alex Coplan wrote:
> > haifa-sched.cc:remove_notes asserts that it lands on a real (non-note)
> > insn after advancing past NOTE_INSN_EPILOGUE_BEG, but with the upcoming
> > post-RA aarch64 load pair pass enabled, we can land on
> > NOTE_INSN_DELETED.
> >
> > This patch adjusts remove_notes to remove these if they occur at the
> > start of the epilogue instead of asserting.
> >
> > Bootstrapped/regtested as a series on aarch64-linux-gnu, OK for trunk?
> >
> > gcc/ChangeLog:
> >
> > * haifa-sched.cc (remove_notes): Allow for NOTE_INSN_DELETED at
> > the start of the epilgoue, remove these.
> One could argue that the pass should have actually deleted the insn rather
> than just turned it into a NOTE_INSN_DELETED. Is there some reason that's
> not done? A NOTE_INSN_DELETED carries no useful information.
Yeah, I think we can teach rtl-ssa to do that. I'm testing a patch to
do so now. I think the pass is the first consumer of the deletion
functionality (insn_change::DELETE) in rtl-ssa, and I can't see a reason for it
to keep the NOTE_INSN_DELETED around.
So consider the scheduler patch withdrawn for now, thanks!
Alex
>
>
> > + /* Skip over any NOTE_INSN_DELETED at the start of the epilogue.
> > + */
>
> Don't bring the close comment down to a new line. If it fits, but it on the
> last line of the actual comment. Otherwise bring down part of comment so
> that we don't have the close comment on a line by itself.
>
> Jeff
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2023-10-19 17:52 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-10-17 20:48 [PATCH 06/11] haifa-sched: Allow for NOTE_INSN_DELETED at start of epilogue Alex Coplan
2023-10-19 14:54 ` Jeff Law
2023-10-19 17:52 ` Alex Coplan
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).