Some packages use LTO (link-time optimization) with the GNU toolchain. In the past, this was problematic because the generated debugging information was not quite usable. Today, it causes problems with some of the hardening quality control tooling because the annobin GCC plug-in may not be active when the final LTO link with new code generation happens (bug 1548821). What's our current take on using LTO for Fedora package builds?

On 02/25/2018 12:45 PM, Zbigniew Jędrzejewski-Szmek wrote: >>What's our current take on using LTO for Fedora package builds? >systemd would like to use it. Why? What are the benefits?

BTW size. Did you try to compile systemd with optimisation for size (-Os)? There are many Linux packages which should be optimised for size because they provide binaries which are speed critical like boot loaders and boot tools (grub, lilo, gubby) some diagnostics and system tools (dmidecode, strace, parted, shadow-utils, util-linux, lvm, mdadm ..) In many cases optimisations for size provides faster code at it has lower probability generate CPU cache misses.

But size is easy: $ du -s /var/tmp/inst-O* 74428 /var/tmp/inst-O3 56136 /var/tmp/inst-O3-lto 61992 /var/tmp/inst-Os 43572 /var/tmp/inst-Os-lto

(Note, it's all unstripped.)

On 26 February 2018 at 07:11, Zbigniew Jędrzejewski-Szmek <zbyszek(a)in.waw.pl > wrote: > On Mon, Feb 26, 2018 at 05:24:28AM +0100, Kevin Kofler wrote: > > Zbigniew Jędrzejewski-Szmek wrote: > > > But size is easy: > > > $ du -s /var/tmp/inst-O* > > > 74428 /var/tmp/inst-O3 > > > 56136 /var/tmp/inst-O3-lto > > > 61992 /var/tmp/inst-Os > > > 43572 /var/tmp/inst-Os-lto > > > > The current default is -O2, not -O3, though. > > Heh, I'm a bit surprised: > > 74428 /var/tmp/inst-O3 > 56136 /var/tmp/inst-O3-lto > 61992 /var/tmp/inst-Os > 43572 /var/tmp/inst-Os-lto > 71268 /var/tmp/inst-O2 > 50276 /var/tmp/inst-O2-lto > 47928 /var/tmp/inst-O0 > 32688 /var/tmp/inst-O0-lto > > So... I'd say that lto *does* give a measure benefit. > (Note, it's all unstripped.) Looking only on total file size is not proper method checking code size. Especially such method may be totally misleading if binaris are not stripped. Can you repeat above and show output of the command "size /var/tmp/inst-*"?

On Sun, Feb 25, 2018 at 01:19:24PM +0100, Florian Weimer wrote: > On 02/25/2018 12:45 PM, Zbigniew Jędrzejewski-Szmek wrote: > > > What's our current take on using LTO for Fedora package builds? > > systemd would like to use it. > > Why? What are the benefits? There's a small change in size. When I tested recently it was ~7%, but it was higher in the past. Fedora builds usually enable almost all optional code, but if some parts are disabled (which happens also in RHEL), I think the gains are bigger.

On 02/25/2018 12:45 PM, Zbigniew Jędrzejewski-Szmek wrote: > > What's our current take on using LTO for Fedora package builds? > systemd would like to use it. Why? What are the benefits?