From: Richard Biener <rguenther@suse.de>
To: "Kewen.Lin" <linkw@linux.ibm.com>
Cc: Stefan Schulze Frielinghaus <stefansf@linux.ibm.com>,
"juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai>,
Robin Dapp <rdapp.gcc@gmail.com>,
"richard.sandiford" <richard.sandiford@arm.com>,
GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: [PATCH V4] VECT: Support loop len control on EXTRACT_LAST vectorization
Date: Tue, 15 Aug 2023 06:58:02 +0000 (UTC) [thread overview]
Message-ID: <nycvar.YFH.7.77.849.2308150655490.12935@jbgna.fhfr.qr> (raw)
In-Reply-To: <3b7fafe9-fcc9-6ce7-f86d-5d4bdad72e3f@linux.ibm.com>
On Tue, 15 Aug 2023, Kewen.Lin wrote:
> Hi Stefan,
>
> on 2023/8/15 02:51, Stefan Schulze Frielinghaus wrote:
> > Hi everyone,
> >
> > I have bootstrapped and regtested the patch below on s390. For the
> > 64-bit target I do not see any changes regarding the testsuite. For the
> > 31-bit target I see the following failures:
> >
> > FAIL: gcc.dg/vect/no-scevccp-outer-14.c (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/no-scevccp-outer-14.c (test for excess errors)
> > FAIL: gcc.dg/vect/pr50451.c (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr50451.c (test for excess errors)
> > FAIL: gcc.dg/vect/pr50451.c -flto -ffat-lto-objects (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr50451.c -flto -ffat-lto-objects (test for excess errors)
> > FAIL: gcc.dg/vect/pr53773.c (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr53773.c (test for excess errors)
> > FAIL: gcc.dg/vect/pr53773.c -flto -ffat-lto-objects (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr53773.c -flto -ffat-lto-objects (test for excess errors)
> > FAIL: gcc.dg/vect/pr71407.c (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr71407.c (test for excess errors)
> > FAIL: gcc.dg/vect/pr71407.c -flto -ffat-lto-objects (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr71407.c -flto -ffat-lto-objects (test for excess errors)
> > FAIL: gcc.dg/vect/pr71416-1.c (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr71416-1.c (test for excess errors)
> > FAIL: gcc.dg/vect/pr71416-1.c -flto -ffat-lto-objects (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr71416-1.c -flto -ffat-lto-objects (test for excess errors)
> > FAIL: gcc.dg/vect/pr94443.c (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr94443.c (test for excess errors)
> > FAIL: gcc.dg/vect/pr94443.c -flto -ffat-lto-objects (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr94443.c -flto -ffat-lto-objects (test for excess errors)
> > FAIL: gcc.dg/vect/pr97558.c (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr97558.c (test for excess errors)
> > FAIL: gcc.dg/vect/pr97558.c -flto -ffat-lto-objects (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/pr97558.c -flto -ffat-lto-objects (test for excess errors)
> > FAIL: gcc.dg/vect/vect-reduc-pattern-3.c -flto -ffat-lto-objects (internal compiler error: in require, at machmode.h:313)
> > FAIL: gcc.dg/vect/vect-reduc-pattern-3.c -flto -ffat-lto-objects (test for excess errors)
> > UNRESOLVED: gcc.dg/vect/no-scevccp-outer-14.c compilation failed to produce executable
> > UNRESOLVED: gcc.dg/vect/pr53773.c -flto -ffat-lto-objects scan-tree-dump-times optimized "\\* 10" 2
> > UNRESOLVED: gcc.dg/vect/pr53773.c scan-tree-dump-times optimized "\\* 10" 2
> > UNRESOLVED: gcc.dg/vect/pr71416-1.c -flto -ffat-lto-objects compilation failed to produce executable
> > UNRESOLVED: gcc.dg/vect/pr71416-1.c compilation failed to produce executable
> > UNRESOLVED: gcc.dg/vect/vect-reduc-pattern-3.c -flto -ffat-lto-objects compilation failed to produce executable
> >
> > I've randomely picked pr50451.c and ran gcc against it which results in:
> >
> > during GIMPLE pass: vect
> > dump file: pr50451.c.174t.vect
> > /gcc-verify-workdir/patched/src/gcc/testsuite/gcc.dg/vect/pr50451.c: In function ?foo?:
> > /gcc-verify-workdir/patched/src/gcc/testsuite/gcc.dg/vect/pr50451.c:5:1: internal compiler error: in require, at machmode.h:313
> > 0x1265d21 opt_mode<scalar_int_mode>::require() const
> > /gcc-verify-workdir/patched/src/gcc/machmode.h:313
> > 0x1d7e4e9 opt_mode<machine_mode>::require() const
> > /gcc-verify-workdir/patched/src/gcc/vec.h:955
> > 0x1d7e4e9 vect_verify_loop_lens
> > /gcc-verify-workdir/patched/src/gcc/tree-vect-loop.cc:1471
> > 0x1da29ab vect_analyze_loop_2
> > /gcc-verify-workdir/patched/src/gcc/tree-vect-loop.cc:2929
> > 0x1da40c7 vect_analyze_loop_1
> > /gcc-verify-workdir/patched/src/gcc/tree-vect-loop.cc:3330
> > 0x1da499d vect_analyze_loop(loop*, vec_info_shared*)
> > /gcc-verify-workdir/patched/src/gcc/tree-vect-loop.cc:3484
> > 0x1deed27 try_vectorize_loop_1
> > /gcc-verify-workdir/patched/src/gcc/tree-vectorizer.cc:1064
> > 0x1deed27 try_vectorize_loop
> > /gcc-verify-workdir/patched/src/gcc/tree-vectorizer.cc:1180
> > 0x1def5c1 execute
> > /gcc-verify-workdir/patched/src/gcc/tree-vectorizer.cc:1296
> > Please submit a full bug report, with preprocessed source (by using -freport-bug).
> > Please include the complete backtrace with any bug report.
> > See <https://gcc.gnu.org/bugs/> for instructions.
> >
>
> It looks like s390 supports variable index vec_extract at -m31 but
> no vector with length. It seems we need to further check the vector
> with length capability, with something like:
>
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 5ae9f69c7eb..ef754467baf 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -10327,10 +10327,11 @@ vectorizable_live_operation (vec_info *vinfo, stmt_vec_info stmt_info,
> vect_record_loop_mask (loop_vinfo,
> &LOOP_VINFO_MASKS (loop_vinfo),
> 1, vectype, NULL);
> - else if (can_vec_extract_var_idx_p (
> + else if (get_len_load_store_mode (TYPE_MODE (vectype), true)
> + .exists ()
> + && can_vec_extract_var_idx_p (
> TYPE_MODE (vectype), TYPE_MODE (TREE_TYPE (vectype))))
> - vect_record_loop_len (loop_vinfo,
> - &LOOP_VINFO_LENS (loop_vinfo),
> + vect_record_loop_len (loop_vinfo, &LOOP_VINFO_LENS (loop_vinfo),
> 1, vectype, 1);
> else
> {
>
> sigh, the formatting looks odd.
I think the error is in vect_verify_loop_lens which assumes that
when we record _any_ length related op the target has to support
both len_load and len_store. Now that we have many other _len
functions that's certainly not true.
Instead a vect_verify_loop_lens-local "fix" would be to not use
.require () but instead when !.exists () simply return false.
That would still effectively require both len-load and len-store
for any -len predicated loop, but at least avoid the ICE.
Richard.
next prev parent reply other threads:[~2023-08-15 6:58 UTC|newest]
Thread overview: 19+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-08-11 14:24 juzhe.zhong
2023-08-14 6:53 ` Richard Biener
2023-08-14 7:09 ` juzhe.zhong
2023-08-14 8:51 ` Kewen.Lin
2023-08-14 8:58 ` Robin Dapp
2023-08-14 9:15 ` Richard Biener
2023-08-14 9:19 ` Kewen.Lin
2023-08-14 9:26 ` juzhe.zhong
2023-08-14 12:07 ` Richard Biener
2023-08-14 12:08 ` juzhe.zhong
2023-08-14 12:45 ` Kewen.Lin
2023-08-14 13:28 ` 钟居哲
2023-08-14 13:44 ` Richard Biener
2023-08-14 18:51 ` Stefan Schulze Frielinghaus
2023-08-15 6:25 ` Kewen.Lin
2023-08-15 6:58 ` Richard Biener [this message]
2023-08-15 8:52 ` Richard Sandiford
2023-08-15 8:55 ` Richard Biener
2023-08-15 9:05 ` Richard Sandiford
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=nycvar.YFH.7.77.849.2308150655490.12935@jbgna.fhfr.qr \
--to=rguenther@suse.de \
--cc=gcc-patches@gcc.gnu.org \
--cc=juzhe.zhong@rivai.ai \
--cc=linkw@linux.ibm.com \
--cc=rdapp.gcc@gmail.com \
--cc=richard.sandiford@arm.com \
--cc=stefansf@linux.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).