Custom section generation under `wasm32-unknown-unknown` is inconsistent and unintuitive

[koute]

Assume we have a crate named dependency with the following content:

pub fn trigger() {
    submodule::call();
}
pub mod submodule {
    pub fn call() {
        #[link_section = "some-custom-section"]
        static SNIPPET: [u8; 3] = [b'X', b'Y', b'Z'];

        extern "C" {
            fn require_XYZ();
        }

        unsafe {
            require_XYZ();
        }
    }
}

And we have another crate which uses the dependency:

extern crate dependency;

#[no_mangle]
pub fn main() {
    dependency::trigger();
}

If I compile this crate like this:

$ cargo build --target=wasm32-unknown-unknown --release

and dump it with wasm-objdump then the "some-custom-section" custom section will be missing. However, if I change the dependency crate to look like this (I've moved the call function from the submodule to the crate root and even made it private):

pub fn trigger() {
    call();
}
fn call() {
    #[link_section = "some-custom-section"]
    static SNIPPET: [u8; 3] = [b'X', b'Y', b'Z'];

    extern "C" {
        fn require_XYZ();
    }

    unsafe {
        require_XYZ();
    }
}

and build the main crate again then the custom section is generated. Calling dependency::submodule::call directly instead of dependency::trigger also results in the custom section being generated.

Based on my experiments the custom section generation currently works like this:

In which crate is the custom section defined?	Where is the function containing the section?	How is the function containing the custom section called?	Is it generated?
External crate	Submodule	Not called	No
External crate	Submodule	Indirectly	No
External crate	Submodule	Directly	Yes
External crate	In root	Not called	No
External crate	In root	Indirectly	Yes
External crate	In root	Directly	Yes
Main crate	Any	Any	Yes

I have an example crate here which reproduces the issue:

$ git clone https://github.com/koute/rust-custom-section-issue
$ cd rust-custom-section-issue

# This will not generate a custom section:
$ cargo build --target=wasm32-unknown-unknown --release --features broken
$ wasm-objdump -s target/wasm32-unknown-unknown/release/rust_custom_section_issue.wasm

# This will:
$ cargo build --target=wasm32-unknown-unknown --release --features working
$ wasm-objdump -s target/wasm32-unknown-unknown/release/rust_custom_section_issue.wasm

I'm using the most recent nightly: rustc 1.32.0-nightly (4a45578bc 2018-12-07)

Could we make this somewhat consistent?

Some background: as a first step towards wasm-bindgen compatibility I'm converting stdweb's js! macro to use custom sections, however to make it not break existing downstream users I need to have either a) every custom section in the whole crate graph be generated, or b) always generated if the custom section is defined inside of a potentially reachable (at runtime) function. Otherwise I end up generating an import for a snippet for which the corresponding custom section entry doesn't exist.

cc @alexcrichton

[alexcrichton]

This is currently expected behavior, albeit somewhat unfortunately. This has to do with object files and what causes the linker to pull in an object file. Long story short a custom section always goes into the current codegen unit, but if nothing else is referenced from the codegen unit then the custom section doesn't get included in the final output because the linker never looks at the codegen unit/object file.

We could try to do better perhaps manually in rustc itself by doing something like placing all custom sections in a separate codegen unit and forcing it to be included, but that doesn't mirror what custom sections do in other targets, for example.

[koute]

Long story short a custom section always goes into the current codegen unit, but if nothing else is referenced from the codegen unit then the custom section doesn't get included in the final output because the linker never looks at the codegen unit/object file.

If that was the case wouldn't the custom section always be emitted when I call the function which contains it? Don't the function and the statics it contains inside of it go into the same codegen unit? Or am I simply misunderstanding this? (I'm sure there are a lot more moving parts here so it must not be so simple to explain in one sentence.)

We could try to do better perhaps manually in rustc itself by doing something like placing all custom sections in a separate codegen unit and forcing it to be included, but that doesn't mirror what custom sections do in other targets, for example.

If this behavior is the same for other targets too then I'd argue that it's, well, broken for those other targets too. As it stands right now the semantics of when the custom section is emitted depends on unintuitive (And I'm guessing not really documented nor guaranteed?) implementation details of the compiler, which is a subtle but fairly serious omission for an already stabilized feature.

[alexcrichton]

It suffices to say that there's a lot of moving parts with codegen units. What's happening above is you're being thwarted by LLVM's ThinLTO passes which performs inlining and reduces dependencies between codegen units.

It's true that it's difficult to use, and that's why it's very low level! We can try to improve it over time, I'm mostly just trying to explain what's happening.

[leighmcculloch]

This problem seems to be much worse on nightly lately. If I have codegen-units set very high with a stable build the link sections still get included, but recently noticed nightly builds are omitting custom sections. Unfortunately I don't have a definitive version of nightly when this started happening.

[thomcc]

If the static also has #[used] it may be related to #93718.

If so, checking that it worked prior to the 2022-07-20 nightly, and is broken after the 2022-07-26 nightly (note that the the nightlies between these two had a very broken version of #[used], which got fixed by #99676) would be pretty helpful, absent an actual regression range.

The #[used] handling in that code probably doesn't need the check for wasm, and if it's causing issues I think we could remove it, so that #[used] is llvm.compiler.used on wasm like it is on ELF. The primary¹ reason it does apply to wasm was just because I was trying to imitate clang's logic for implementing __attribute__((used)), which basically seemed to use llvm.compiler.used on ELF, and llvm.used for everything else.

That said, the relationship to codegen units is unclear to me, and if your static doesn't have #[used] it seems unlikely to be related (that said, if you don't have #[used], adding it might fix your issue...)

Footnotes

1. That is, aside from the fact that in an ideal world it should be legal to use llvm.used in place of llvm.compiler.used. Sadly, we don't live in such an ideal world, as use of llvm.used can expose toolchain bugs, which we already know happens with some ELF linkers (gold). ↩

[WorldSEnder]

I'm also experiencing issues related to this. Specifically, the following code placed in a private, not-used module will emit the custom section:

#[unsafe(no_mangle)]
static _MARKER: () = {
    #[used]
    #[link_section = "some-custom-section"]
    static SNIPPET: [u8; 3] = [b'X', b'Y', b'Z'];
};

but omitting the _MARKER around it will not. It seems the no_mangle annotation for _MARKER somehow influence the used annotation on SNIPPET. The issue is that the above generates an additional (global (;1;) i32 (i32.const <addr>)) in the wasm that I do not want to be there. I have not yet found a way to avoid this additional global+export and still have the custom section present. Even though, in my mind, no memory is shared here and the SNIPPET does not get duplicated into the data section.