From: Richard Henderson <rth@redhat.com>
To: uweigand@de.ibm.com
Cc: gcc-patches@gcc.gnu.org, rguenther@suse.de
Subject: [PATCH 0/2] Convert s390 to atomic optabs, v2
Date: Mon, 30 Jul 2012 22:33:00 -0000 [thread overview]
Message-ID: <1343687574-3244-1-git-send-email-rth@redhat.com> (raw)
In-Reply-To: <5016C81E.5020709@redhat.com>
The atomic_load/storedi_1 patterns are fixed to use LM, STM.
I've had a go at generating better code in the HQImode CAS
loop for aligned memory, but I don't know that I'd call it
the most efficient thing ever. Some of this is due to
deficiencies in other parts of the compiler (including the
s390 backend):
(1) MEM_ALIGN can't pass down full align+ofs data that we had
during cfgexpand. This means the opportunities for using
the "aligned" path are less than they ought to be.
(2) In get_pointer_alignment (used by get_builtin_sync_mem),
we don't consider an ADDR_EXPR to return the full alignment
that the type is due. I'm sure this is to work around some
other sort of usage via the <string.h> builtins, but it's
less-than-handy in this case.
I wonder if in get_builtin_sync_mem we ought to be using
get_object_alignment (build_fold_indirect_ref (addr)) instead?
Consider
struct S { int x; unsigned short y; } g_s;
unsigned short o, n;
void good() {
__builtin_compare_exchange (&g_s.y, &o, n, 0, 0, 0);
}
void bad(S *p_s) {
__builtin_compare_exchange (&p_s->y, &o, n, 0, 0, 0);
}
where GOOD produces the aligned MEM that we need, and BAD doesn't.
(3) Support for IC, and ICM via the insv pattern is lacking.
I've added a tiny bit of support here, in the form of using
the existing strict_low_part patterns, but most definitely we
could do better.
(4) The *sethighpartsi and *sethighpartdi_64 patterns ought to be
more different. As is, we can't insert into bits 48-56 of a
DImode quantity, because we don't generate ICM for DImode,
only ICMH.
(5) Missing support for RISBGZ in the form of an extv/z expander.
The existing *extv/z splitters probably ought to be conditionalized
on !Z10.
(6) The strict_low_part patterns should allow registers for at
least Z10. The SImode strict_low_part can use LR everywhere.
(7) RISBGZ could be used for a 3-address constant lshrsi3 before
srlk is available.
For the GOOD function above, and this patch set, for -O3 -march=z10:
larl %r3,s+4
lhrl %r0,o
lhi %r2,1
l %r1,0(%r3)
nilh %r1,0
.L2:
lr %r5,%r1
larl %r12,n
lr %r4,%r1
risbg %r4,%r0,32,47,16
icm %r5,3,0(%r12)
cs %r4,%r5,0(%r3)
je .L3
lr %r5,%r4
nilh %r5,0
cr %r5,%r1
lr %r1,%r5
jne .L2
lhi %r2,0
.L3:
srl %r4,16
sthrl %r4,o
Odd things:
* O is forced into a register before reaching the expander, so we
get the RISBG for that. N is left in a memory and so we commit
to using ICM for that. Further, because of how strict_low_part
is implemented we're committed to leaving that in memory.
* We don't optimize the loop and hoist the LARL of N outside the loop.
* Given that we're having to zap the mask in %r1 for the second
compare anyway, I wonder if RISBG is really beneficial over OR.
Is RISBG (or ICM for that matter) any faster (or even smaller)?
r~
Richard Henderson (2):
s390: Reorg s390_expand_insv
s390: Convert from sync to atomic optabs
gcc/config/s390/s390-protos.h | 3 +-
gcc/config/s390/s390.c | 270 ++++++++++++++++++----------
gcc/config/s390/s390.md | 401 +++++++++++++++++++++++++++++------------
3 files changed, 465 insertions(+), 209 deletions(-)
--
1.7.7.6
next prev parent reply other threads:[~2012-07-30 22:33 UTC|newest]
Thread overview: 29+ messages / expand[flat|nested] mbox.gz Atom feed top
2012-07-29 21:32 [CFT] s390: Convert from sync to atomic optabs Richard Henderson
2012-07-30 14:19 ` Ulrich Weigand
2012-07-30 15:12 ` Richard Henderson
2012-07-30 15:51 ` Ulrich Weigand
2012-07-30 18:53 ` Richard Henderson
2012-07-30 22:33 ` Richard Henderson [this message]
2012-07-30 22:33 ` [PATCH 1/2] s390: Reorg s390_expand_insv Richard Henderson
2012-07-30 22:36 ` [PATCH 2/2] s390: Convert from sync to atomic optabs Richard Henderson
2012-08-06 18:34 ` Ulrich Weigand
2012-08-06 18:51 ` Richard Henderson
2012-08-06 19:45 ` Richard Henderson
2012-08-06 22:40 ` s390: Avoid CAS boolean output inefficiency Richard Henderson
2012-08-07 17:02 ` Ulrich Weigand
2012-08-07 22:13 ` Richard Henderson
2012-08-08 18:05 ` Ulrich Weigand
2012-08-09 16:55 ` Eric Botcazou
2012-07-31 9:11 ` [PATCH 0/2] Convert s390 to atomic optabs, v2 Richard Guenther
2012-07-31 15:27 ` Andrew MacLeod
2012-07-31 16:07 ` Richard Henderson
2012-08-01 8:41 ` Richard Guenther
2012-08-01 15:59 ` Richard Henderson
2012-08-01 17:14 ` Richard Guenther
2012-08-01 19:42 ` Richard Henderson
2012-07-31 18:36 ` Ulrich Weigand
2012-07-31 19:54 ` Richard Henderson
2012-08-01 23:23 ` Richard Henderson
2012-08-03 12:20 ` Ulrich Weigand
2012-08-03 14:21 ` Ulrich Weigand
2012-08-06 16:44 ` Ulrich Weigand
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=1343687574-3244-1-git-send-email-rth@redhat.com \
--to=rth@redhat.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=rguenther@suse.de \
--cc=uweigand@de.ibm.com \
/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).