Emacs 28.1 built with x32 ABI outperforms x86-64

Emacs 28.1 built with x32 ABI outperforms x86-64

[Sven Hartrumpf]

Dear Emacs users.

With the new release of Emacs, I tried to build it with the x32 ABI.
I succeeded for the first time (gave up with earlier OS versions) and
it even worked without any problems (Ubuntu 21.10, gcc 11.2).

A small benchmark ("make check" followed by reproducing all .elc files)
outperforms the equivalent x86-64 version by 8 %.
Memory usage is almost half of x86-64, when opening files of several 100 MB.

Limitations:
- x32 has a hard memory limit of 4 GB.
- To avoid building too many dependencies for x32, I used the following
  configure line: ./configure --with-sound=no --without-dbus --without-x
- I had to build 12 x32 versions:
  gnutls autoconf ncurses nettle libidn2 libtasn1 p11-kit libgmp libtool libffi guile-2.2 gc
  (Ask me if you need more details.)

Greetings
Sven

Re: Emacs 28.1 built with x32 ABI outperforms x86-64

[Emanuel Berg via Users list for the GNU Emacs text editor]

Sven Hartrumpf wrote:

> With the new release of Emacs, I tried to build it with the
> x32 ABI. I succeeded for the first time (gave up with
> earlier OS versions) and it even worked without any problems
> (Ubuntu 21.10, gcc 11.2).
>
> A small benchmark ("make check" followed by reproducing all
> .elc files) outperforms the equivalent x86-64 version by 8
> %. Memory usage is almost half of x86-64, when opening files
> of several 100 MB.
>
> Limitations:
> - x32 has a hard memory limit of 4 GB.
> - To avoid building too many dependencies for x32, I used the following
>   configure line: ./configure --with-sound=no --without-dbus --without-x
> - I had to build 12 x32 versions:
>   gnutls autoconf ncurses nettle libidn2 libtasn1 p11-kit
> libgmp libtool libffi guile-2.2 gc
>   (Ask me if you need more details.)

Yes, can you post the script (or function) that carried out
the experiment?

-- 
underground experts united
https://dataswamp.org/~incal

Re: Emacs 28.1 built with x32 ABI outperforms x86-64

[Sven Hartrumpf]

Hello.

Emanuel wrote, 2022-04-08 19:17:
> Sven wrote:
> Yes, can you post the script (or function) that carried out
> the experiment?

Sure, see below.

Some notes:

- gccx32 is a tiny script that calls gcc in x32 mode
  (just passing CFLAGS to configure does not work as expected), containing:
 gcc-11 -mx32 -I/usr/local/include/x86_64-linux-gnux32 -I/usr/x86_64-linux-gnux32/include "$@"

- dropcaches-root is s tool that avoids filesystem cache effects, simlar to a:
  echo 3 > /proc/sys/vm/drop_caches

- I call the script emacs.install as follows:
    emacs.install x86-64 |& tee emacs.64
    emacs.install x32 |& tee emacs.x32
  To test the second line, you will need all the x32 builds that I mentioned in my first mail.

--- emacs.install

#!/bin/bash
set -e -u -x

ABI=$1
tar xf emacs-28.1.tar.gz
cd emacs-28.1
if [ "$ABI" == "x32" ] ; then
  CC=gccx32 ./configure --prefix=/usr/local --bindir=/usr/local/binx32 --includedir=/usr/local/includex32 --libdir=/usr/local/libx32 --with-sound=no --without-dbus --without-x
else # normal 64 bit
  CC=gcc-11 ./configure --prefix=/usr/local --bindir=/usr/local/bin64 --libdir=/usr/local/lib64 --with-sound=no --without-dbus --without-x
fi
make -j4
strip src/emacs
dropcaches-root
date -Is
make check
find lisp -name '*.elc' | grep -v -e '/c-by.elc$' -e 'bindat.elc' | xargs rm # grep to avoid 2 files where compilation fails with max-lisp-eval-depth error
date -Is
make
date -Is

---