From: Mathieu Malaterre <malat@debian.org>
To: Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>
Cc: Florian Weimer <fweimer@redhat.com>, libc-help@sourceware.org
Subject: Re: glibc-hwcaps for armv7 (neon-vfpv4)
Date: Fri, 1 Sep 2023 09:18:33 +0200 [thread overview]
Message-ID: <CA+7wUszJw29JQ1A+Z82BAd75+MAnv5xC-Sg7+ug++wSgKR5dMw@mail.gmail.com> (raw)
In-Reply-To: <90243b3d-740a-f391-089e-45d5ada7096f@linaro.org>
Hi !
On Thu, Aug 31, 2023 at 5:30 PM Adhemerval Zanella Netto
<adhemerval.zanella@linaro.org> wrote:
>
>
>
> On 31/08/23 09:08, Mathieu Malaterre wrote:
> > Hi Florian !
> >
> > On Thu, Aug 31, 2023 at 1:24 PM Florian Weimer <fweimer@redhat.com> wrote:
> >>
> >> * Mathieu Malaterre:
> >>
> >>> Dear glibc maintainer,
> >>>
> >>> I fail to understand the ld.so man page (Debian/sid version: man-pages
> >>> 6.03) for hwcaps support. Specifically I'd like to install a shared
> >>> lib on a Debian/armhf system (baseline is neon-less) which was build
> >>> with gcc option:
> >>>
> >>> -march=armv7-a -mfpu=neon-vfpv4
> >>>
> >>> What subfolder should I be using ?
> >>
> >> You can use “LD_LIBRARY_PATH=. /bin/true” to see which subdirectories
> >> are probed. I think either neon/vfp or vfp/neon should be among the
> >> subdirectories. This functionality has since been removed from upstream
> >> glibc because it sometimes results in hundereds of extra openat system
> >> calls.
> >
> > oh, thanks for the trick. Could you confirm my conclusion below
> > (Debian/i386 sid system):
> >
> > % file /usr/lib/i386-linux-gnu/i686/sse2/libx264.so.164
> > /usr/lib/i386-linux-gnu/i686/sse2/libx264.so.164: ELF 32-bit LSB
> > shared object, Intel 80386, version 1 (SYSV), dynamically linked,
> > BuildID[sha1]=e66974d10aef77af7ed504266cde974d103484d6, stripped
> >
> > However:
> >
> > % /sbin/ldconfig -v 2>/dev/null | grep -A3 hwcap
> > -> nothing
> >
> > same goes with:
> >
> > % LD_DEBUG=libs LD_LIBRARY_PATH=. /bin/true
> > 3327956: find library=libc.so.6 [0]; searching
> > 3327956: search path=. (LD_LIBRARY_PATH)
> > 3327956: trying file=./libc.so.6
> > 3327956: search cache=/etc/ld.so.cache
> > 3327956: trying file=/lib/i386-linux-gnu/libc.so.6
> > 3327956:
> > 3327956:
> > 3327956: calling init: /lib/ld-linux.so.2
> > 3327956:
> > 3327956:
> > 3327956: calling init: /lib/i386-linux-gnu/libc.so.6
> > 3327956:
> > 3327956:
> > 3327956: initialize program: /bin/true
> > 3327956:
> > 3327956:
> > 3327956: transferring control: /bin/true
> > 3327956:
> > 3327956:
> > 3327956: calling fini: [0]
> > 3327956:
> > 3327956:
> > 3327956: calling fini: /lib/i386-linux-gnu/libc.so.6 [0]
> > 3327956:
> > 3327956:
> > 3327956: calling fini: /lib/ld-linux.so.2 [0]
> > 3327956:
> >
> >
> > If I understand correctly the file
> > /usr/lib/i386-linux-gnu/i686/sse2/libx264.so.164 cannot be used on
> > Debian/sid/i386 system since libc used is:
> >
> > % /lib/i386-linux-gnu/libc.so.6
> > GNU C Library (Debian GLIBC 2.37-7) stable release version 2.37.
> > Copyright (C) 2023 Free Software Foundation, Inc.
> > This is free software; see the source for copying conditions.
> > There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
> > PARTICULAR PURPOSE.
> > Compiled by GNU CC version 12.3.0.
> > libc ABIs: UNIQUE IFUNC ABSOLUTE
> > Minimum supported kernel: 3.2.0
> > For bug reporting instructions, please see:
> > <http://www.debian.org/Bugs/>.
>
> From your system configuration, it seems so. On a ubuntu22 I see:
>
> $ find /usr/ -iname libx264.so*
> /usr/lib/i386-linux-gnu/libx264.so.163
> /usr/lib/i386-linux-gnu/i686/sse2/libx264.so.163
> /usr/lib/x86_64-linux-gnu/libx264.so.163
> $ /sbin/ldconfig -v 2>/dev/null | grep -A3 hwcap
> /lib/i386-linux-gnu/i686: (hwcap: 0x0002000000000000) (from /etc/ld.so.conf.d/i386-linux-gnu.conf:3)
> /lib/i386-linux-gnu/i686/sse2: (hwcap: 0x0002000000000001) (from /etc/ld.so.conf.d/i386-linux-gnu.conf:3)
> libx264.so.163 -> libx264.so.163
>
> $ cat /etc/ld.so.conf.d/i386-linux-gnu.conf
> # Multiarch support
> /usr/local/lib/i386-linux-gnu
> /lib/i386-linux-gnu
> /usr/lib/i386-linux-gnu
> /usr/local/lib/i686-linux-gnu
> /lib/i686-linux-gnu
> /usr/lib/i686-linux-gnu
>
> $ /lib/ld-linux.so.2 --help
> [...]
> Shared library search path:
> (libraries located via /etc/ld.so.cache)
> /lib/i386-linux-gnu (system search path)
> /usr/lib/i386-linux-gnu (system search path)
> /lib (system search path)
> /usr/lib (system search path)
>
> No subdirectories of glibc-hwcaps directories are searched.
^very neat trick. I was not aware of this command.
> Legacy HWCAP subdirectories under library search path directories:
> i686 (AT_PLATFORM; supported, searched)
> tls (supported, searched)
> sse2 (supported, searched)
>
> So if you have '/usr/lib/i386-linux-gnu' add on your ld.so.cache, the
> <path>/i686/see will be search through legacy hwcap support.
This has been confirmed by Florian since: you'd need glibc 2.37 to
reproduce the behavior I am seeing in Debian/sid.
> >
> >>> Conversely how should I read the following:
> >>>
> >>> % sudo ldconfig -p | grep hwcap
> >>> libfoo.so.160 (libc6,x86-64, hwcap: 0x0004000000000000) =>
> >>> /lib/x86_64-linux-gnu/haswell/libfoo.so.160
> >>> libfoo.so.160 (libc6,x86-64, hwcap: 0x0000000000000004) =>
> >>> /lib/x86_64-linux-gnu/avx512_1/libfoo.so.160
> >>>
> >>> What does "haswell" / "avx512_1" subfolder implies in terms of gcc
> >>> compile options ?
> >>
> >> It's complicated. Nowdays, on x86-64, you can use
> >> glibc-hwcaps/x86-64-v3 and glibc-hwcaps/x86-64-v4, which are designed to
> >> correspond to -march=x86-64-v3 and -march=x86-64-v4, and x86-64-v4 is a
> >> superset of x86-64-v3.
> >
> > Ok, thanks for the update.
> >
> > Could you just review what I see on my Debian/sid/armhf system:
> >
> > % cat /proc/cpuinfo | grep Features | uniq
> > Features : half thumb fastmult vfp edsp neon vfpv3 tls vfpv4
> > idiva idivt lpae evtstrm
> > % LD_DEBUG=libs LD_LIBRARY_PATH=. /bin/true
> > 2964957: find library=libc.so.6 [0]; searching
> > 2964957: search path=. (LD_LIBRARY_PATH)
> > 2964957: trying file=./libc.so.6
> > 2964957: search cache=/etc/ld.so.cache
> > 2964957: trying file=/lib/arm-linux-gnueabihf/libc.so.6
> > 2964957:
> > 2964957:
> > 2964957: calling init: /lib/ld-linux-armhf.so.3
> > 2964957:
> > 2964957:
> > 2964957: calling init: /lib/arm-linux-gnueabihf/libc.so.6
> > 2964957:
> > 2964957:
> > 2964957: initialize program: /bin/true
> > 2964957:
> > 2964957:
> > 2964957: transferring control: /bin/true
> > 2964957:
> > 2964957:
> > 2964957: calling fini: [0]
> > 2964957:
> > 2964957:
> > 2964957: calling fini: /lib/arm-linux-gnueabihf/libc.so.6 [0]
> > 2964957:
> > 2964957:
> > 2964957: calling fini: /lib/ld-linux-armhf.so.3 [0]
> > 2964957:
> >
> > Based on the above, it seems glibc is setup to never search hwcaps
> > optimized libraries on this system (even if neon/vfpv4 is supported by
> > system).
>
> This is a ubuntu22 armhf chroot on aarch64:
>
> $ /lib/ld-linux-armhf.so.3 --help
> [...]
> Shared library search path:
> (libraries located via /etc/ld.so.cache)
> /lib/arm-linux-gnueabihf (system search path)
> /usr/lib/arm-linux-gnueabihf (system search path)
> /lib (system search path)
> /usr/lib (system search path)
>
> No subdirectories of glibc-hwcaps directories are searched.
>
> Legacy HWCAP subdirectories under library search path directories:
> v8l (AT_PLATFORM; supported, searched)
> tls (supported, searched)
> neon (supported, searched)
> vfp (supported, searched)
>
> So both neon and vfp; I can not say exactly how debian is configured since both
> ubuntu and debian carry out of tree patches; but I would expect them to be
> similar regarding libc.
next prev parent reply other threads:[~2023-09-01 7:18 UTC|newest]
Thread overview: 6+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-08-31 10:06 Mathieu Malaterre
2023-08-31 11:24 ` Florian Weimer
2023-08-31 12:08 ` Mathieu Malaterre
2023-08-31 15:30 ` Adhemerval Zanella Netto
2023-09-01 7:18 ` Mathieu Malaterre [this message]
2023-08-31 17:20 ` Florian Weimer
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=CA+7wUszJw29JQ1A+Z82BAd75+MAnv5xC-Sg7+ug++wSgKR5dMw@mail.gmail.com \
--to=malat@debian.org \
--cc=adhemerval.zanella@linaro.org \
--cc=fweimer@redhat.com \
--cc=libc-help@sourceware.org \
/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).