From: Jan Hubicka <hubicka@ucw.cz>
To: Richard Biener <rguenther@suse.de>
Cc: gcc-patches@gcc.gnu.org, d@dcepelik.cz
Subject: Re: Make nonoverlapping_component_refs work with duplicated main variants
Date: Mon, 08 Jul 2019 10:48:00 -0000 [thread overview]
Message-ID: <20190708094327.72wcfah7ezh5pxh5@kam.mff.cuni.cz> (raw)
In-Reply-To: <alpine.LSU.2.20.1907081057430.2976@zhemvz.fhfr.qr>
> > +/* FIELD1 and FIELD2 are two component refs whose bases are either
> > + both at the same address or completely disjoint.
> > + Return 1 if FIELD1 and FIELD2 are non-overlapping
> > + Return 0 if FIELD1 and FIELD2 are having same addresses or are
> > + completely disjoint.
>
> completely disjoint? I guess
>
> Return 0 if accesses to FIELD1 and FIELD2 are possibly overlapping.
>
> is better matching actual behavior. Likewise mentioning 'accesses'
> in the first because of the bitfield treatment (the fields may
> be non-overlapping but actual accesses might be).
I was trying to describe difference between 0 and -1.
We return 0 when we fully structurally matched the path and we know it
is same. -1 means that we arrived to somehting we can not handle (union,
mismatched offsets) and it would make sense to try disambiguating
further.
Currently it means that in addition to
nonoverlapping_component_refs_since_match_p we also do
nonoverlapping_component_refs_p which has some chance to recover from
the mismatched REF pair, match the types later on path and still
disambiguate. It seem to happen very rarely though.
>
> > + /* Different fields of the same record type cannot overlap.
> > + ??? Bitfields can overlap at RTL level so punt on them. */
> > + if (DECL_BIT_FIELD (field1) && DECL_BIT_FIELD (field2))
> > + return -1;
>
> This is similar as the DECL_BIT_FIELD_REPRESENTATIVE check so why
> return -1 instead of 0?
Well, my plan is to put this test before ref_and_offset which still have
chace to suceed if fields are far away. But i am happy to return 0 here
and mess with that later.
> > + else
> > + {
> > + if (operand_equal_p (DECL_FIELD_BIT_OFFSET (field1),
> > + DECL_FIELD_BIT_OFFSET (field2), 0))
> > + return 0;
>
> I think this is overly pessimistic - the offset of a field
> is DECL_FIELD_OFFSET + DECL_FIELD_BIT_OFFSET (the latter is
> only up to DECL_OFFSET_ALIGN, the rest of the constant
> offset spills into DECL_FIELD_OFFSET). Which also means ...
>
> > +
> > + /* Different fields of the same record type cannot overlap.
> > + ??? Bitfields can overlap at RTL level so punt on them. */
> > + if (DECL_BIT_FIELD (field1) && DECL_BIT_FIELD (field2))
> > + return -1;
> > +
> > + poly_uint64 offset1;
> > + poly_uint64 offset2;
> > + poly_uint64 size1;
> > + poly_uint64 size2;
> > + if (!poly_int_tree_p (DECL_FIELD_BIT_OFFSET (field1), &offset1)
> > + || !poly_int_tree_p (DECL_FIELD_BIT_OFFSET (field2), &offset2)
> > + || !poly_int_tree_p (DECL_SIZE (field1), &size1)
> > + || !poly_int_tree_p (DECL_SIZE (field2), &size2)
> > + || ranges_maybe_overlap_p (offset1, size1, offset2, size2))
>
> this is technically wrong in case we had DECL_FIELD_OFFSETs 4 and 8
> and DECL_FIELD_BIT_OFFSETs 32 and 0.
>
> So you have to compute the combined offsets first.
OK, I guess I can take look at the get_base_ref_and_offset there. Thanks
for pointing this out.
>
> > + return -1;
>
> I think it may make sense to return -1 if any of the !poly_int_tree_p
> tests fire, but otherwise? I'm not actually sure what -1 vs. 0
> means here - is 0 a must exactly overlap and -1 is a may overlap
> somehow?
Well, we have two fields that overlap partly from two different types
in >nonoverlapping_component_refs_since_match_p so it can not
continue walking (since the main invariant is broken)
we may still suceed with the nonoverlaping_component_refs
Thanks, I will update the patch.
Honza
next prev parent reply other threads:[~2019-07-08 9:43 UTC|newest]
Thread overview: 15+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-07-08 7:39 Jan Hubicka
2019-07-08 9:10 ` Richard Biener
2019-07-08 10:48 ` Jan Hubicka [this message]
2019-07-09 12:02 ` Jan Hubicka
2019-07-09 12:21 ` Richard Biener
2019-07-09 12:41 ` Jan Hubicka
2019-07-09 12:52 ` Richard Biener
2019-07-09 13:10 ` Jan Hubicka
2019-07-09 13:30 ` Richard Biener
2019-07-09 13:37 ` Jan Hubicka
2019-07-09 13:41 ` Richard Biener
2019-07-09 21:03 ` Bernhard Reutner-Fischer
2019-07-11 8:29 ` Rainer Orth
2019-07-16 9:30 ` Jan Hubicka
2019-07-16 11:58 ` Rainer Orth
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=20190708094327.72wcfah7ezh5pxh5@kam.mff.cuni.cz \
--to=hubicka@ucw.cz \
--cc=d@dcepelik.cz \
--cc=gcc-patches@gcc.gnu.org \
--cc=rguenther@suse.de \
/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).