From: Marc Glisse <marc.glisse@inria.fr>
To: Richard Biener <richard.guenther@gmail.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: update address taken: don't drop clobbers
Date: Sun, 07 Sep 2014 15:28:00 -0000 [thread overview]
Message-ID: <alpine.DEB.2.11.1409071654310.5483@laptop-mg.saclay.inria.fr> (raw)
In-Reply-To: <alpine.DEB.2.11.1407271853200.22972@stedding.saclay.inria.fr>
On Sun, 27 Jul 2014, Marc Glisse wrote:
> On Thu, 10 Jul 2014, Richard Biener wrote:
>
>>> --- gcc/tree-into-ssa.c (revision 212109)
>>> +++ gcc/tree-into-ssa.c (working copy)
>>> @@ -1831,26 +1831,38 @@ maybe_register_def (def_operand_p def_p,
>>> {
>>> tree def = DEF_FROM_PTR (def_p);
>>> tree sym = DECL_P (def) ? def : SSA_NAME_VAR (def);
>>>
>>> /* If DEF is a naked symbol that needs renaming, create a new
>>> name for it. */
>>> if (marked_for_renaming (sym))
>>> {
>>> if (DECL_P (def))
>>> {
>>> - tree tracked_var;
>>> -
>>> - def = make_ssa_name (def, stmt);
>>> + if (gimple_clobber_p (stmt) && is_gimple_reg (sym))
>>
>> sym should always be a gimple reg here (it's marked for renaming).
>>
>>> + {
>>> + /* Replace clobber stmts with a default def. Create a new
>>> + variable so we don't later think we must coalesce, which
>>> would
>>> + fail with some ada abnormal PHIs. Still, we try to keep
>>> a
>>> + similar name so error messages make sense. */
>>> + unlink_stmt_vdef (stmt);
>>
>> I think that's redundant with gsi_replace (note that using gsi_replace
>> looks dangerous here as it calls update_stmt during SSA rewrite...
>> that might open a can of worms).
>>
>>> + gsi_replace (&gsi, gimple_build_nop (), true);
>>> + tree id = DECL_NAME (sym);
>>> + const char* name = id ? IDENTIFIER_POINTER (id) : 0;
>>> + tree newvar = create_tmp_var (TREE_TYPE (sym), name);
>>> + def = get_or_create_ssa_default_def (cfun, newvar);
>>
>> So - can't you simply do
>>
>> gimple_assign_set_rhs_from_tree (&gsi,
>> get_or_create_dda_default_def (cfun, sym));
>>
>> ? Thus replace x = CLOBBER; with x_3 = x_2(D);
>>
>>> + }
>>> + else
>>
>> and of course still rewrite the DEF then. IMHO the copy-propagation
>> you do is premature optimization.
>
> Using your version, I end up with spurious warnings, in particular for
> va_list. pass_fold_builtins stops va_start/va_end taking the address of the
> list, so we get:
>
> list_6 = list_2(D);
>
> in place of the clobber at the end of the function. And there is no DCE-like
> pass afterwards, so we warn for the use of list_2(D).
> (passes.def contains a comment about running dce before uninit)
>
> I don't know if update_address_taken could avoid generating this assignment
> where the lhs has 0 use, but this shows the optimization is not completely
> premature.
>
> (uninit could also check for this case, but that feels like a bad hack)
I would like some guidance on this. I just tried this trivial patch:
NEXT_PASS (pass_split_crit_edges);
+ NEXT_PASS (pass_dce);
NEXT_PASS (pass_late_warn_uninitialized);
and it does not cause any regression, it even XPASS
gfortran.dg/reassoc_6.f for some reason. The FIXME note just above in
passes.def mentions 2 testcases that are already xfailed anyway.
Would that extra pass be acceptable?
Otherwise, what do you think should be responsible for cleaning up the
dead assignments?
--
Marc Glisse
next prev parent reply other threads:[~2014-09-07 15:28 UTC|newest]
Thread overview: 48+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-06-28 22:33 Marc Glisse
2014-06-29 23:38 ` SRA: " Marc Glisse
2014-07-07 8:56 ` Richard Biener
2014-07-07 9:32 ` Marc Glisse
2014-07-07 18:32 ` Richard Biener
2014-07-07 20:15 ` Marc Glisse
2014-07-07 16:59 ` Jeff Law
2014-07-10 14:55 ` Richard Biener
2014-07-10 15:01 ` Jakub Jelinek
2014-06-30 19:31 ` update address taken: " Jeff Law
2014-07-06 14:24 ` Marc Glisse
2014-07-06 14:54 ` pinskia
2014-07-06 15:01 ` Marc Glisse
2014-07-07 10:21 ` Richard Biener
2014-07-07 17:20 ` Jeff Law
2014-07-08 13:31 ` Marc Glisse
2014-07-10 15:10 ` Richard Biener
2014-07-10 15:49 ` Michael Matz
2014-07-10 18:23 ` Jeff Law
2014-07-11 8:10 ` Richard Biener
2014-07-11 8:14 ` Richard Biener
2014-07-11 12:06 ` Michael Matz
2014-07-11 17:16 ` Jeff Law
2014-07-12 6:15 ` Marc Glisse
2014-07-24 13:06 ` Richard Biener
2014-07-27 11:53 ` Marc Glisse
2014-07-27 18:01 ` Marc Glisse
2014-09-07 15:28 ` Marc Glisse [this message]
2014-10-15 14:36 ` Marc Glisse
2014-10-15 16:12 ` Jeff Law
2014-10-16 11:20 ` Richard Biener
2014-10-16 14:11 ` Marc Glisse
2014-10-16 14:34 ` Richard Biener
2014-10-16 17:29 ` Jeff Law
2014-10-16 17:58 ` Richard Biener
2014-10-16 18:37 ` Jeff Law
2014-10-17 20:46 ` Marc Glisse
2014-10-24 20:22 ` Jeff Law
2014-10-25 8:06 ` Marc Glisse
2014-10-31 21:06 ` Jeff Law
2014-10-25 17:14 ` Marc Glisse
2014-10-31 11:12 ` Richard Biener
2014-11-02 10:34 ` Marc Glisse
2014-11-03 9:06 ` Richard Biener
2014-10-31 21:16 ` Jeff Law
2014-10-16 17:23 ` Jeff Law
2014-10-18 22:23 ` Marc Glisse
2014-10-24 20:19 ` Jeff Law
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=alpine.DEB.2.11.1409071654310.5483@laptop-mg.saclay.inria.fr \
--to=marc.glisse@inria.fr \
--cc=gcc-patches@gcc.gnu.org \
--cc=richard.guenther@gmail.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).