Support native library modifiers (RFC 2951)

[petrochenkov]

cc rust-lang/rust#81490

I haven't added a test because stabilization of the feature has just landed on nightly and is not available on stable yet, but I've tested it manually.

Closes #689.

[dot-asm]

I for one would argue that link_modifies would be more appropriate name for suggested method. Because cc.modifiers("x") would appear rather as compiler modifier.

[petrochenkov]

I'd say the full name is "native library modifiers" because they are not necessarily about linking in the narrow sense, since they can also control bundling of native libraries into rlibs/staticlibs or potentially deduplication before linking.
Not sure how to better name it in this specific interface.
EDIT: But link_modifies should be fine too.

[dot-asm]

In the context "native" is effectively synonymous with "ffi" and cc-rs is all about "ffi." In other words "native" feels redundant in the context.

Just a reminder, I'm just a cc-rs user weathering #663.

[petrochenkov]

Renamed to link_lib_modifiers to match the modified cargo directive.

[petrochenkov]

Ping @m-ou-se.
Can this be merged in some form and released before the Rust 1.61 release?
It helps to address the first compatibility note from https://github.com/rust-lang/rust/pull/96539/files.

[m-ou-se]

Hi!

Is this cargo feature documented anywhere? The docs only shows the =[KIND=]NAME syntax, but say nothing about any modifiers.

[petrochenkov]

This is a rustc syntax (https://doc.rust-lang.org/nightly/rustc/command-line-arguments.html#-l-link-the-generated-crate-to-a-native-library), cargo just passes the STRING in cargo:rustc-link-lib=STRING directly to rustc.

But I agree that the cargo docs also need to be updated, I'll send a PR.
UPD: PR sent - rust-lang/cargo#10674.

[m-ou-se]

This is a rustc syntax (https://doc.rust-lang.org/nightly/rustc/command-line-arguments.html#-l-link-the-generated-crate-to-a-native-library), cargo just passes the STRING in cargo:rustc-link-lib=STRING directly to rustc.

Ah great. Can you add a link to that in the doc comment for link_lib_modifiers?

Then one more question:

The flag() method takes a single flag and adds it to a list of flags, similar for include(), but this new link_lib_modifiers() method takes a string with multiple modifiers and overrides any previously set value. Do you think it should be link_lib_modifier() that adds a single modifier, instead?

[petrochenkov]

Can you add a link to that in the doc comment for link_lib_modifiers?

Done.

The flag() method takes a single flag and adds it to a list of flags, similar for include(), but this new link_lib_modifiers() method takes a string with multiple modifiers and overrides any previously set value. Do you think it should be link_lib_modifier() that adds a single modifier, instead?

Both resetting and appending method could potentially make sense in both cases.
I added only a resetting method for the modifiers just because it's simpler, more general, and sufficient for now.

(I've changed the parameter from &str to Option<&str> though, so you can reset the modifiers back to none if necessary.)

[helgoboss]

Thanks @petrochenkov. I'm also very interested in your addition because the Rust 1.61 change rust-lang/rust#93901 breaks my application at runtime (although it compiles and links successfully). Using .link_lib_modifiers(Some("+whole-archive")) for one specific cc build invocation fixes that. For the record, here's the commit which fixes it: helgoboss/helgobox@0055d6c.

[petrochenkov]

@helgoboss
FWIW, +whole-archive can be passed even without this API, by using build.cargo_metadata(false) first, and then emitting the cargo:rust-ling-* directives manually:

println!("cargo:rustc-link-lib=static:+whole-archive={}", lib_name);
println!("cargo:rustc-link-search=native={}", env::var("OUT_DIR"));

Ping @m-ou-se again.

[helgoboss]

@petrochenkov Thanks, good to know. Your solution is much more convenient though.

[m-ou-se]

I added only a resetting method for the modifiers just because it's simpler, more general, and sufficient for now.

(I've changed the parameter from &str to Option<&str> though, so you can reset the modifiers back to none if necessary.)

The other methods like flags(), include(), file(), define(), are all append-only and provide no way of resetting. I think we should keep this consistent with that.

Right now it doesn't matter much with only "+whole-archive", but it'd be nice if we can keep things unchanged when more useful modifiers are added in the furture.

[petrochenkov]

Updated in accordance with #671 (comment).

[petrochenkov]

ping @m-ou-se

[thomcc]

This looks good to me, thanks.

Can this be merged in some form and released before the Rust 1.61 release?

Obviously we missed the boat here, but I'll try to cut a release after I finish going through the pending PRs and merging the ones that are viable.