From: Richard Biener <richard.guenther@gmail.com>
To: Robin Dapp <rdapp.gcc@gmail.com>
Cc: Richard Sandiford <richard.sandiford@arm.com>,
gcc-patches <gcc-patches@gcc.gnu.org>,
Tamar Christina <Tamar.Christina@arm.com>
Subject: Re: [PATCH] vect: Use statement vectype for conditional mask.
Date: Fri, 17 Nov 2023 09:47:47 +0100 [thread overview]
Message-ID: <CAFiYyc39g3Vb=xfPtoscRKCo6g4RZ7DC1+OAYvSNz-VjGvcUdw@mail.gmail.com> (raw)
In-Reply-To: <0424a09b-6867-4be9-8207-9d92a6187191@gmail.com>
On Fri, Nov 17, 2023 at 9:45 AM Robin Dapp <rdapp.gcc@gmail.com> wrote:
>
> > But note you can explicitly specify a vector type as well, there's an
> > overload for it, so we can fix the "invariant" case with the following
> > (OK if you can test this on relevant targets)
> >
> > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> > index 3f59139cb01..936a3de9534 100644
> > --- a/gcc/tree-vect-loop.cc
> > +++ b/gcc/tree-vect-loop.cc
> > @@ -8450,12 +8450,14 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
> > value. */
> > bool cond_fn_p = code.is_internal_fn ()
> > && conditional_internal_fn_code (internal_fn (code)) != ERROR_MARK;
> > + int cond_index = -1;
> > if (cond_fn_p)
> > {
> > gcc_assert (code == IFN_COND_ADD || code == IFN_COND_SUB
> > || code == IFN_COND_MUL || code == IFN_COND_AND
> > || code == IFN_COND_IOR || code == IFN_COND_XOR);
> > gcc_assert (op.num_ops == 4 && (op.ops[1] == op.ops[3]));
> > + cond_index = 0;
> > }
> >
> > bool masked_loop_p = LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
> > @@ -8486,12 +8488,13 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
> > vect_get_vec_defs (loop_vinfo, stmt_info, slp_node, ncopies,
> > single_defuse_cycle && reduc_index == 0
> > ? NULL_TREE : op.ops[0], &vec_oprnds0,
> > + cond_index == 0 ? truth_type_for (vectype_in) : NULL_TREE,
> > single_defuse_cycle && reduc_index == 1
> > - ? NULL_TREE : op.ops[1], &vec_oprnds1,
> > + ? NULL_TREE : op.ops[1], &vec_oprnds1, NULL_TREE,
> > op.num_ops == 4
> > || (op.num_ops == 3
> > && !(single_defuse_cycle && reduc_index == 2))
> > - ? op.ops[2] : NULL_TREE, &vec_oprnds2);
> > + ? op.ops[2] : NULL_TREE, &vec_oprnds2, NULL_TREE);
>
> Ah, yes that's what I meant. I had something along those lines:
>
> + if (!cond_fn_p)
> + {
> + vect_get_vec_defs (loop_vinfo, stmt_info, slp_node, ncopies,
> + single_defuse_cycle && reduc_index == 0
> + ? NULL_TREE : op.ops[0], &vec_oprnds0,
> + single_defuse_cycle && reduc_index == 1
> + ? NULL_TREE : op.ops[1], &vec_oprnds1,
> + op.num_ops == 3
> + && !(single_defuse_cycle && reduc_index == 2)
> + ? op.ops[2] : NULL_TREE, &vec_oprnds2);
> + }
> + else
> + {
> + /* For a conditional operation, pass the truth type as vectype for the
> + mask. */
> + gcc_assert (single_defuse_cycle && reduc_index == 1);
> + vect_get_vec_defs (loop_vinfo, stmt_info, slp_node, ncopies,
> + op.ops[0], &vec_oprnds0,
> + truth_type_for (vectype_in),
> + NULL_TREE, &vec_oprnds1, NULL_TREE,
> + op.ops[2], &vec_oprnds2, NULL_TREE);
> + }
>
> Even though this has a bit of duplication now I prefer it slightly
> because of the. I'd hope, once fully tested (it's already running
> but I'm having connection problems so don't know about the results
> yet), this could go in independently of the pattern fix?
Yes, your version is also OK.
Thanks,
Richard.
>
> Regards
> Robin
>
next prev parent reply other threads:[~2023-11-17 8:48 UTC|newest]
Thread overview: 12+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-08 16:18 Robin Dapp
2023-11-10 9:33 ` Richard Biener
2023-11-16 22:30 ` Robin Dapp
2023-11-17 8:24 ` Richard Biener
2023-11-17 8:45 ` Robin Dapp
2023-11-17 8:47 ` Richard Biener [this message]
2023-11-17 13:04 ` Robin Dapp
2023-12-03 18:32 ` [PATCH] testsuite: Fix up gcc.target/aarch64/pr112406.c for modern C [PR112406] Jakub Jelinek
2023-12-03 18:32 ` Jakub Jelinek
2023-12-03 18:59 ` Richard Biener
2023-11-17 19:20 ` [PATCH] vect: Use statement vectype for conditional mask Robin Dapp
2023-11-20 8:03 ` Richard Biener
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='CAFiYyc39g3Vb=xfPtoscRKCo6g4RZ7DC1+OAYvSNz-VjGvcUdw@mail.gmail.com' \
--to=richard.guenther@gmail.com \
--cc=Tamar.Christina@arm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=rdapp.gcc@gmail.com \
--cc=richard.sandiford@arm.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).