rust - re-export libceed-sys features through libceed to allow building non-static without using libceed-sys directly

[jedbrown]

Fix #1694

I will note that there is nothing wrong with depending on libceed-sys directly to set the feature at that level, so I don't think this is a necessary change, just a convenience.

Thanks-to: @eliasboegel

[eliasboegel]

Thanks for bringing this up again and creating the PR!

Out of curiosity: Are you sure that this isn't a necessary change in order to link libceed dynamically?
Because of how features are intended to be additive only and are unified under that assumption,

[dependencies]
libceed-sys = { version="0.12", default-features=false}

will result in libceed-sys being built as a shared library, but

[dependencies]
libceed-sys = { version="0.12", default-features=false}
libceed = "0.12

will result in libceed-sys being built as a static library even though the static feature is explicitly disabled with default-features=false. This is because libceed brings in the default static feature, overridingdefault-features=false in the libceed-sys specification.

The possible fixes are either:

1. Re-export of the libceed-sys features like in this PR to allow:

[dependencies]
libceed = { version="0.12", default-features=false}

2. Opting for the "inverse feature" in libceed-sys, i.e. having an optional (additive), non-default shared feature instead of a default static feature, such that libceed doesn't enable any feature. This should allow doing:

[dependencies]
libceed-sys = { version="0.12", shared=true}
libceed = "0.12"

Happy to hear your thoughts, and if you prefer 2), then I'm happy to implement that instead of the feature re-export.

[jedbrown]

I should have written more carefully. I was thinking of having libceed/Cargo.toml use

[dependencies]
libceed-sys = { version = "0.12", path = "../libceed-sys", default-features = false}

but skipping the re-exporting of features. I think that supports all use cases, but requires transitive naming of libceed-sys to configure it. However, one is supposed to avoid having mutually-exclusive features whenever possible. What if we removed the static feature from libceed-sys and just leave the non-default system feature? (We can re-export it as a convenience.)

[eliasboegel]

Ah, thanks for clarifying. I'm personally not a fan of depending on transitive naming and instead in favor of removing the static feature from libceed-sys. I still think it would be important to have a shared build available due to #1694, but that could be a non-default feature called shared or dynamic or similar (re-exported or not).

[jedbrown]

Do you mean something other than the existing system feature? To clarify, is there reason to support building a private shared libceed.so/libceed.dylib (in most cases implying that the user must set LD_LIBRARY_PATH or otherwise arrange for the library to be found), versus private (meaning non-system) builds always being static (libceed.a)?

[eliasboegel]

Yes, in my mind they serve different purposes:
static (or a proposed shared): Building libCEED through Cargo
system: Use externally an built libCEED.

I have to admit that my motivation for a Cargo-built shared libCEED is quite niche. I would like to retain the ability to do this purely because of the weak symbol problem on MacOS when linking statically (#1694), which I have not been able to solve, so I link libCEED dynamically as a bandaid solution.

I can absolutely see how thats not a common case though and can see an argument for removing private shared builds to streamline the features. In that case I'll need to separately build/install libCEED while on my MacOS machine instead of having Cargo build it. That would be a little inconvenient but acceptable since that'll be necessary for any device support anyways.

[jedbrown]

Okay, I pushed a commit to swap the feature names, but it'll use pkg-config --static when trying to find the system library, unless one sets both system and shared. (If we keep it, I'll document that behavior, but I'm not sure I like it.) What do you think?

[eliasboegel]

This works for me and I think this is a relatively good solution. What's your concern with it?

[jedbrown]

Just that activating system only will tend to overlink or possibly fail, since system libs are usually shared libs. So it's a bit more user complexity to remember to activate both system and shared.

[eliasboegel]

I see. If you feel that this is a significant factor then I find removing shared/static altogether acceptable. It will remove the ability to use a Cargo-built libCEED on MacOS while #1694 persists, but the workaround is dynamically linking an externally built libCEED. For me personally that wouldn't be a problem. Can you see any use-case for statically linking an externally built libCEED? If not, then I think we could move to only system.

[jedbrown]

Revisiting this while trying to understand the root issue of weak symbols not working with static linking on MacOS. I don't have a solution there, but I found that ar crD doesn't work because Apple ar does not support D (deterministic), but one can use the environment variable ZERO_AR_DATE=1.
rust-lang/rust#47086 (comment)

[eliasboegel]

I also faced the ar D issue, and worked around it by using a third party bintools distribution rather than that shipped with Apple Clang. Made no difference on the weak symbol issue though.