Calling to the graph-depends script is very costly, as it calls back to
'make' a lot of time.
It turns out that we already have the list of recursive dependencies, so
we can just print that.
As for the recursive reverse dependencies, we use the same memoisation
technique to cut-down on the expansion cost, which would otherwise be on
the order of 𝑶(𝑛²) (with 𝑛 enabled packages).
>From a defconfig, modified to use glibc, C++, wchar, locales, ssp, and
allyespackageconfig (tweaked to avoid multi providers, etc...), the
timings for X-show-recursive-rdepends are:
before after speedup #rdeps
libnss 0m22.932s 0m5.775s 3.97x 3
qt5base 0m41.176s 0m5.781s 7.12x 67
libjpeg 0m56.185s 0m5.749s 9.71x 228
libxml2 0m54.964s 0m5.795s 9.48x 271
freetype 0m46.754s 0m5.819s 8.07x 287
libpng 0m53.577s 0m5.760s 9.30x 303
sqlite 1m15.222s 0m5.807s 12.95x 801
libopenssl 1m25.471s 0m5.844s 14.63x 931
readline 1m13.805s 0m5.775s 12.78x 958
libzlib 1m11.807s 0m5.820s 12.34x 1039
toolchain 1m23.712s 0m6.080s 13.77x 2107
skeleton 1m27.839s 0m6.293s 13.96x 2111 (+1)
host-skeleton 1m27.405s 0m6.350s 13.76x 2172 (+2)
- speedup: ratio before/after
- #rdeps: number of recursive reverse dependencies, with the extra
dependencies returned with this patch, see below for the
reason.
So, for a low-level package with a lot of reverse dependencies, like
libzlibz, libopenssl or readline are, the timings are already very much
in favour of the change. This is less impressive with packages that
have few dependencies (libnss), but still much faster.
Also, remember that the config tested has as much packages enabled as
possible, so is in itself a degenerate case. With simpler and more
realistic configurations, the gains would probably be a bit lower than
reported above, but various tests still report good improvements
overall (note: coming up with a 'realistic' configuration is pretty
hard, as everyone and their dog have their notion of what is realistic
in their context, so nothing displayed here; timings are left as an
exercise for the interested parties to report aggravation in their
cases should they notice some regression).
Note that, more recursive reverse dependencies may be displayed now,
since we do not apply the exceptions applied in graph-depends. For
example, host-skeleton gains two new recursive reverse dependencies:
skeleton and toolchain, which are both exceptions in graph-depends.
As for direct (not reverse) dependencies: the gain is not as fantastic
as for reverse ones, but it is still noticeable, especially thanks to
a21212fb7cf (package/pkg-generic: speed up RECURSIVE_FINAL_DEPENDENCIES);
just a few examples for %-show-recursive-depends:
before after speedup #deps
libzlib 0m46.864s 0m5.902s 7.94x 17
qt5base 0m57.590s 0m5.848s 9.85x 190
sqlite 0m46.601s 0m5.816s 8.01x 24
Basically, displaying recursive dependencies, direct or reverse, is
almost a constant now: it only slightly varies by about 10% depending
on the complexity of the dependency chain, with the parsing of the
Makefiles still accounting for the large majority of the time.
(PS. Thanks to Joseph for suggesting a list of interesting packages
to test, and thanks to Trent for his example of memoisation!)
Signed-off-by: "Yann E. MORIN" <yann.morin.1998@free.fr>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Thomas De Schampheleire <patrickdepinguin@gmail.com>
Cc: Joseph Kogut <joseph.kogut@gmail.com>
Cc: Trent Piepho <tpiepho@impinj.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>