Seems in compose 1.7 ui-graphics now depends on gr...
# composet
Tolriq
12/09/2023, 9:15 AMSeems in compose 1.7 ui-graphics now depends on graphic-path that includes huge .so files. Any plans to try to reduce the sizes if they are mandatory now? 340K for ARMV8 is huge given the small code. (No log link, MinSizeRel) ?
r
romainguy
12/09/2023, 5:37 PMThe .so shouldn't be so big indeed. I'll look into this, thanks for raising this problem
t
Tolriq
12/09/2023, 5:43 PMThanks from a French compatriot 🙂
r
romainguy
12/09/2023, 5:44 PMDid you try a release build btw? AGP should do pretty aggressive symbols removal
t
Tolriq
12/09/2023, 5:44 PMIt is on release builds, I carefully check aab size in CI.
r
romainguy
12/09/2023, 5:44 PMThanks
t
Tolriq
12/09/2023, 5:45 PMSeems there's a very huge .text in the .so I suppose it's built with a recent NDK?
Tolriq
12/09/2023, 5:45 PMThere was a lot of size regressions after 22.
Tolriq
12/09/2023, 5:46 PMhttps://github.com/android/ndk/issues/1402 Maybe your example will give them enough data to report to upper project.
r
romainguy
12/09/2023, 5:51 PMThis reminds me of how much I dislike the C++ stdlib
t
Tolriq
12/09/2023, 5:53 PMnot linking with log, tweaking -Os or forcing MinSizeRel that is not the default should already help a lot, but yes I hate the very little control we have 🙂
r
romainguy
12/09/2023, 5:54 PMBased on what I'm reading in the bug reports it doesn't seem -Os will help that much
t
Tolriq
12/09/2023, 5:55 PMNot for that bug no, but I'm not sure this is the main or the only issue here, just pointing that issue due to the big .text, but there's maybe room to gain on the rest too.
r
romainguy
12/09/2023, 5:55 PMWe're already pretty aggressive at stripping stuff: https://cs.android.com/androidx/platform/frameworks/support/+/androidx-main:graphics/graphics-path/build.gradle;l=50
romainguy
12/09/2023, 5:56 PMBut yeah liblog shouldn't be linked in release
t
Tolriq
12/09/2023, 5:56 PMYes I've seen I use nearly the same.
r
romainguy
12/09/2023, 5:56 PMI'll fix that
romainguy
12/09/2023, 5:57 PMAnd this library can probably not use libc++ at all
t
Tolriq
12/09/2023, 5:58 PMWell I did not dig that down I'll leave you the fun 😛
r
romainguy
12/11/2023, 9:15 PMalright, I shaved off 150 KiB
romainguy
12/11/2023, 9:15 PMdoing more would require ditching the stdlib entirely it seems
romainguy
12/11/2023, 10:28 PMAnd another 10 KiB
romainguy
12/12/2023, 1:23 AMHmm seems I’m down to 14 KiB total
t
Tolriq
12/12/2023, 7:30 AMWow, nice change indeed. Seeing the PR the last iteration is just about removing stdlib something I can't do. Do you have some details about what you did for the first 150k in case there's some trickery I can reuse?
r
romainguy
12/12/2023, 8:05 AMYes. The first 150 KiB were won by doing something I should have done in the first place for safety: marking all symbols but JNI_OnLoad as local so they can be removed by the stripping step
romainguy
12/12/2023, 8:06 AMI saved another 10 KiB by ditching call_once and std_vector. I was left with only one call to "new" as an stdlib non-header-only call and that's enough to bring exceptions and libunwind 😕
romainguy
12/12/2023, 8:07 AMActually just linking libstdc++ seems enough. I guess static initializers are to blame?
t
Tolriq
12/12/2023, 9:44 AMThanks a lot, I missed the already merged parts, will look them in details. And thanks for the work on that .so really appreciated.