I think this is the right way to do this for Linux. @Homebrew/linux, can one of you have a look?

See also #75449.

From the bottle json files:

❯ jq '.[].bottle.cellar' <abseil--20210324.0.*.bottle.json
"any"
"any"
"any"
"any"

Shouldn't the RPATH be rewritten by install_name_tool on macOS? My understanding of binary relocation is that the RPATH can be updated using install_name_tool on macOS and patchelf.rb on Linux. The harder problem is when there are strings in the .rodata section of an ELF (or the equivalent location in a Mach-O on macOS), and that's why I want to implement Homebrew/brew#10846 to fix that.

I don't think we do any RPATH re-writing on macOS, so we rely on the build system to specify this correctly. When it doesn't, we can give it hints, but it seems we've given a suboptimal one here.

Do you do RPATH rewriting on Linux?

Homebrew/brew#10846 would be good to have, but I think that wherever we can get relocation for free, we should take it.

I don't think we do any RPATH re-writing on macOS, so we rely on the build system to specify this correctly. When it doesn't, we can give it hints, but it seems we've given a suboptimal one here.

It looks like we sort of do this by rewriting the lib paths in https://github.com/Homebrew/brew/blob/a654730accaaadff4d28ed8678b88baed693d621/Library/Homebrew/extend/os/mac/keg_relocate.rb#L42, and then try to remove the RPATHs afterward if the lib paths no longer use them. In theory I think we should fully rewrite the RPATH on macOS when pouring bottles and I will have to investigate this as part of Homebrew/brew#10846, but that can't be changed now.

Do you do RPATH rewriting on Linux?

That is what patchelf.rb (and formerly the patchelf formula) is used for.

Homebrew/brew#10846 would be good to have, but I think that wherever we can get relocation for free, we should take it.

100% agreed on this!

My guess is that this change will have no affect on Linux either way because the RPATHs are ignored when determining a bottle cellar. I can test this locally if you want me to just to be sure.

Actually it looks like this is already relocatable on Linux:

https://github.com/Homebrew/linuxbrew-core/blob/008365c9444c63af980f2302be56f2f401536bdc/Formula/abseil.rb

I'm just more concerned about whether the specific rpath argument here might break anything. I suppose I could always just set it to opt_lib, as it is in the current version of the formula on Linux, but we don't want that on macOS as that breaks relocatability.

Knowing the correct thing to pass to CMAKE_INSTALL_RPATH might be useful for other formulae though.

Actually it looks like this is already relocatable on Linux:

https://github.com/Homebrew/linuxbrew-core/blob/008365c9444c63af980f2302be56f2f401536bdc/Formula/abseil.rb

I'm just more concerned about whether the specific rpath argument here might break anything. I suppose I could always just set it to opt_lib, as it is in the current version of the formula on Linux, but we don't want that on macOS as that breaks relocatability.

I would say we should take the safe route in that case and use opt_lib on Linux, since it will be ignored when determining the cellar, and rewritten as needed if the user has a non-standard install prefix.

Knowing the correct thing to pass to CMAKE_INSTALL_RPATH might be useful for other formulae though.

Yes, that could be useful if it improves relocatability on macOS. It's possible that as part of Homebrew/brew#10846 we will eventually not need to set the RPATH like this, but until then I think we should use this approach to eliminate "trivial" cases where the only reason a binary can't be relocated is because of the RPATH.

Done.

Needs the next brew release tag.

🤖 A scheduled task has triggered a merge.