08:39 <ardb> arnd: the updated binutils with 64k segment alignment for armhf landed in trixie so things should gradually start working again

08:44 <arnd> ardb: nice! have you also looked into changing dpkg-deb into creating sparse tarballs so the rebuilt files won't cause a size regression?

08:48 <arnd> I suspect we just need this oneliner, but I haven't tried yet. This shold also help on arm64 https://www.irccloud.com/pastebin/CtmeY7o2/

09:17 <ardb> arnd: which source package does that apply against?

09:18 <arnd> dpkg

09:19 <arnd> I'm trying it out right now, built the same package again after installing it. Apparently the binaries in it happen to not change with sparseness though, need to try another one

09:26 <ardb> my build fails the selftest

09:36 <ardb> never mind

10:14 <arnd> ardb: I played around it some more, and found that it's (of course) more complicated:

10:15 <arnd> tar is able to store sparse files efficiently, but even with --sparse --sparse-format=1.0 --hole-detection=raw, it seems to only do this for files that are already sparse

10:16 <arnd> the random library package I looked at (libaom3) calls 'strip' on the .so file, which makes it non-sparse, even when it was sparse before stripping

10:17 <ardb> ugh

10:18 <ardb> so better to do it at installation time if page size > fs block size?

10:18 <arnd> I think hole-detection=raw is necessary for reproducible builds, otherwise the amount of sparseness depends on the block size of the file system on the build host. With =raw it uses the default 512 byte block size of the tar format

10:19 <arnd> I don't think there is any problem in making the file more sparse to save disk space

10:20 <arnd> e.g. with the extreme case of 64K pages and 512 byte FS blocks, there is still no reason to store the empty 512 byte blocks on disk, they just won't improve the memory usage in the page cache

10:34 <ardb> sure

10:35 <ardb> but if strip and objcopy etc don't preserve sparseness, we should probably fix that first

11:06 <arnd> I see that binutils-2.41 contains https://sourceware.org/git/?p=binutils-gdb.git;a=commitdiff;h=281309f3;hp=0d5adb56c85da38a0f95e872fda05cc6446010c3

11:06 <arnd> so it may just work. I'm still on Debian bookworm here, which has binutils-2.40

11:10 <arnd> I tried with 2.42 strip from my local collection of toolchains, and that does not create a sparse file. Interestingly a 2.28 strip does create a sparse file