Note that this change leaves cross-compilation to aarch64 as "broken" as it is now -- patchelf doesn't know about the target architecture being patched, so it will just use the 4K value from the host instead.

The better solution would be something like NixOS/patchelf#80 .

After a lot of rebuilding: I can confirm this leads to a minimal system which boots on both 4K pages and 64K pages configs, and the userland works well enough to build itself on either of the two configurations.

@delroth do you have an idea how to fix this in patchelf itself? The existing patch doesn't seem finished.

I think there are basically 2 approaches for a long term fix in patchelf.

1. What's done in patchelf#80, which is adding some minimal alignment per architecture in patchelf. I don't think there is much work to do on that patch, it probably just needs to be rebased, but it seems fine.
2. Changing patchelf to just never decrease the alignment when relocating segments -- basically, if a segment has alignment 0x10000, don't pick a new alignment, just reuse the 0x10000. I think this would be a better solution, but not knowing about patchelf I don't know how involved this is.

(1) will still require changes to patchelf for each new architecture with page size != 4K. (2) is theoretically better IMO by just following whatever the toolchain generated.

Both of these options would fix the cross-compilation problem and the reproducibility problem.

Could patchelf also check for aarch64 and set the proper with-page-size value?

I've made an attempt at a patchelf fix in NixOS/patchelf#216

I guess this PR isn't relevant anymore with the patchelf PR merged?

Yeah, though I'm starting to lose hope of ever seeing a new patchelf release...