From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-bugzilla@gcc.gnu.org>
Received: by sourceware.org (Postfix, from userid 48)
	id 268783856DD0; Wed, 17 May 2023 14:04:17 +0000 (GMT)
DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 268783856DD0
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org;
	s=default; t=1684332257;
	bh=8KwrwHs7p53ywBrjH9cRmd2KQeU61QfVu/ZZJfq2kno=;
	h=From:To:Subject:Date:In-Reply-To:References:From;
	b=EzT+zrGAXKn9n1s7kJzwOdmyM+YrNwB07f5EOnQeh1oMv7fdQ/WYEnU2KTAAUNgTx
	 MpCxhvY+BZUAjA6JbEoJui2nZ13/ko8v/Wxl3mM1LNGKwf4IYU5IUocO+G7SqRN8NJ
	 y21ku1tYmpsiH4LAwU5HRPXU6MDuo1bUKhjrDsC0=
From: "hubicka at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>
To: gcc-bugs@gcc.gnu.org
Subject: [Bug target/109811] libjxl 0.7 is a lot slower in GCC 13.1 vs Clang
 16
Date: Wed, 17 May 2023 14:04:16 +0000
X-Bugzilla-Reason: CC
X-Bugzilla-Type: changed
X-Bugzilla-Watch-Reason: None
X-Bugzilla-Product: gcc
X-Bugzilla-Component: target
X-Bugzilla-Version: 13.1.1
X-Bugzilla-Keywords: missed-optimization
X-Bugzilla-Severity: normal
X-Bugzilla-Who: hubicka at gcc dot gnu.org
X-Bugzilla-Status: NEW
X-Bugzilla-Resolution: 
X-Bugzilla-Priority: P3
X-Bugzilla-Assigned-To: unassigned at gcc dot gnu.org
X-Bugzilla-Target-Milestone: ---
X-Bugzilla-Flags: 
X-Bugzilla-Changed-Fields: attachments.created
Message-ID: <bug-109811-4-cpMhgQFbhY@http.gcc.gnu.org/bugzilla/>
In-Reply-To: <bug-109811-4@http.gcc.gnu.org/bugzilla/>
References: <bug-109811-4@http.gcc.gnu.org/bugzilla/>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
X-Bugzilla-URL: http://gcc.gnu.org/bugzilla/
Auto-Submitted: auto-generated
MIME-Version: 1.0
List-Id: <gcc-bugs.sourceware.org>

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D109811

--- Comment #8 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Created attachment 55101
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=3D55101&action=3Dedit
hottest loop

jpegxl build machinery adds -fno-vectorize and -fno-slp-vectorize to clang
flags.  Adding -fno-tree-vectorize -fno-tree-slp-vectorize makes GCC genera=
ted
code more similar.  With this most difference is caused by
FindBestPatchDictionary or FindTextLikePatches if that function is not inli=
ned.

  15.22%  cjxl     libjxl.so.0.7.0       [.] jxl::(anonymous
namespace)::FindTextLikePatches=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20
  10.19%  cjxl     libjxl.so.0.7.0       [.] jxl::FindBestPatchDictionary=
=20=20=20=20=20=20=20
   5.27%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::QuantizeBlockAC=
=20=20=20=20=20=20=20
   5.06%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::EstimateEntropy=
=20=20=20=20=20=20=20
   4.82%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::EstimateEntropy=
=20=20=20=20=20=20=20
   4.35%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::QuantizeBlockAC=
=20=20=20=20=20=20=20
   4.21%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::(anonymous
namespace)::TransformFromPixels=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20
   3.87%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::(anonymous
namespace)::TransformFromPixels=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=20=
=20=20=20=20=20=20=20=20=20=20
   3.78%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::FindBestMultiplie=
r=20=20=20=20
   3.27%  cjxl     libjxl.so.0.7.0       [.] jxl::N_AVX2::FindBestMultiplie=
r=20=20=20=20

I think it is mostly register allocation not handling well the internal loop
quoted above.  I am adding preprocessed sources.=