From: Richard Sandiford <richard.sandiford@arm.com>
To: Oluwatamilore Adebayo <oluwatamilore.adebayo@arm.com>
Cc: <gcc-patches@gcc.gnu.org>, <richard.guenther@gmail.com>
Subject: Re: [PATCH] rtl: AArch64: New RTL for ABD
Date: Tue, 06 Jun 2023 18:03:37 +0100 [thread overview]
Message-ID: <mptjzwgplp2.fsf@arm.com> (raw)
In-Reply-To: <20230606151153.53653-1-oluwatamilore.adebayo@arm.com> (Oluwatamilore Adebayo's message of "Tue, 6 Jun 2023 16:11:53 +0100")
Oluwatamilore Adebayo <oluwatamilore.adebayo@arm.com> writes:
>> It would be good to mark all of these functions with __attribute__((noipa)),
>> since I think interprocedural optimisations might otherwise defeat the
>> runtime test in abd_run_1.c (in the sense that we might end up folding
>> things at compile time and not testing the vector versions of the functions).
>
> Done.
>
>> There are 14 tests, and it looks like 6 of them are expected to produce
>> ABD instructions while 8 aren't. It isn't really clear which tests are
>> which though.
>>
>> I think it'd help to split the file into two:
>>
>> - one containing only the tests that should produce ABD, so that the
>> scan-assembler counts sum up to the number of tests
>>
>> - one containing only the tests that cannot use ABD, with:
>>
>> { dg-final { scan-assembler-not {\tsabd\t} } }
>> { dg-final { scan-assembler-not {\tuabd\t} } }
>>
>> to enforce that
>
> After adjustments made to the vectoriser part, all tests now use an abd
> instruction.
Ah, that's a problem. Sorry, I didn't review 1/2 closely enough.
For:
> + /* Failed to find a widen operation so we check for a regular MINUS_EXPR. */
> + if (diff
> + && gimple_assign_rhs_code (diff) == MINUS_EXPR
> + && (TYPE_UNSIGNED (abs_type) || TYPE_OVERFLOW_UNDEFINED (abs_type)))
> + {
> + *half_type = NULL_TREE;
> + return true;
> + }
the condition should instead be:
if (diff
&& gimple_assign_rhs_code (diff) == MINUS_EXPR
&& TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (abs_oprnd)))
{
*half_type = NULL_TREE;
return true;
}
That is, we rely on overflow being undefined, so we need to check
TYPE_OVERFLOW_UNDEFINED on the type of the subtraction (rather than
abs_type, which is the type of ABS input, and at this point can be
different from TREE_TYPE (abs_oprnd)).
Then fn_unsigned_int and fn_unsigned_char_int_short correctly
avoid using SABD. The output for the other tests looks right.
It would be good to add a:
/* { dg-final { scan-assembler-not {\tabs\t} } } */
to be the positive tests, to make it more obvious that all separate
ABS instructions are elided.
Thanks,
Richard
next prev parent reply other threads:[~2023-06-06 17:03 UTC|newest]
Thread overview: 8+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-05-09 16:14 Oluwatamilore Adebayo
2023-05-16 12:30 ` Richard Sandiford
2023-06-06 15:11 ` Oluwatamilore Adebayo
2023-06-06 15:17 ` [PATCH 2/2] " Oluwatamilore Adebayo
2023-06-06 17:03 ` Richard Sandiford [this message]
2023-06-08 10:37 ` [PATCH] rtl: " Oluwatamilore Adebayo
2023-06-08 10:38 ` [PATCH 2/2] " Oluwatamilore Adebayo
2023-06-13 8:27 ` Oluwatamilore Adebayo
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=mptjzwgplp2.fsf@arm.com \
--to=richard.sandiford@arm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=oluwatamilore.adebayo@arm.com \
--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).