On 08/10/2017 12:10 AM, Randy MacLeod wrote:
Yep, total time for 'without gtk-doc' is *more* than cut in half!
The drop in configure time is certainly expected but the compilation
stage should be dominated by the compiler rather than make or ninja.
The 'without gtk-doc' compile difference mostly confirms that
but the 9% drop is odd. A couple of sources I've found assert that
for large projects and parallel builds a full build time is essentially
the same with ninja:
That work was done on a MacBook (!!) so it would be interesting to
see what the results are on a 24+ core Linux system. I might
give that a try tonight if there's nothing good on NetFlix.
To be honest, achieving faster build times is secondary in importance to
me. The important part is that autotools is one of the most awful pieces
of software ever written, and the less we have to deal with it in Yocto,
the happier we all will be. Upstreans generally share that POV.
By the way, systemd-234 has meson support and you (Alex) have sent
a patch update to 242 to the oe-core list but without switching to meson
as is reasonable. Anyway, 232 takes 51 seconds on my 16+16 core machine
so it would be a useful benchmark as well. Want to take a stab at that?
Oe-core master does not yet have any support for meson. We've done some
private work to bring it in, but it's not ready for submission yet. I
also want to build a few different recipes with it to make sure it
'basically works' for several different things.
After meson support is in master, we can start converting recipes from
autotools to meson. That requires careful manual work, but the good news
is that it can be done piece-meal; there's no need to switch everything
My plan is to try gstreamer next, and after that, implement meson
support for gettext. It's a notorious bottleneck in many builds, due to
inexplicably slow autoconf (that even RP couldn't get down to the root
Openembedded-core mailing list