Rust is inconsistently re-exporting rlib symbols from a binary

[erickt]

I noticed some strange behavior while trying to optimize file size. Consider this cargo setup:

Cargo.toml:

[package]
name = "foo"
version = "0.1.0"
authors = ["Erick Tryzelaar <etryzelaar@google.com>"]
edition = "2018"

src/lib.rs:

#[no_mangle]
pub fn big1() {}

#[no_mangle]
pub fn big2() {}

#[no_mangle]
pub fn big3() {}

examples/big1.rs:

fn main() {
    foo::big1();
}

examples/big2.rs:

fn main() {
    foo::big2();
}

examples/big3.rs:

fn main() {
    foo::big3();
}

On my mac using both rustc 1.51.0 (2fd73fabe 2021-03-23) and rustc 1.53.0-nightly (42816d61e 2021-04-24), I found that the rust binary for the two examples will be different:

% rustup run stable cargo build --release --all-targets && ls -l target/release/examples/{big1,big2,big3}
-rwxr-xr-x  1 etryzelaar  primarygroup  432536 May  4 17:02 target/release/examples/big1
-rwxr-xr-x  1 etryzelaar  primarygroup  432576 May  4 17:02 target/release/examples/big2
-rwxr-xr-x  1 etryzelaar  primarygroup  432608 May  4 17:02 target/release/examples/big3

Those handful of bytes of difference appears to be because for some reason, big2 binary is re-exporting the _big2 function, but the big1 binary does not re-export _big1:

% diff <(nm -g target/release/examples/big1) <(nm -g target/release/examples/big2)
192a193
> 00000001000037e0 T _big2
200c201
< 0000000100003740 T _main
---
> 0000000100003720 T _main

% diff <(nm -g target/release/examples/big1) <(nm -g target/release/examples/big3)
192a193,194
> 00000001000037e0 T _big2
> 00000001000037e0 T _big3
200c202
< 0000000100003740 T _main
---
> 0000000100003770 T _main

Ideally we'd optimize away all these symbols to shave off some bytes, but if that's difficult, we probably should at least make things consistent.

[tmandry]

I would guess this is an artifact caused by different functions being assigned to different codegen units, which affects LLVM's ability to inline them.

I can't explain why the function big2 appears in the big3 binary, but not big1..

[erickt]

I tried setting codgen-units=1, and while that seems to have made all the _main symbols identical, it's still inconsistently adding the _big* symbols:

Cargo.toml:

[package]
name = "foo"
version = "0.1.0"
authors = ["Erick Tryzelaar <etryzelaar@google.com>"]
edition = "2018"

[profile.release]
codegen-units = 1

% rustup run stable cargo clean && rustup run stable cargo build --release --all-targets -v && ls -l target/release/examples/{big1,big2,big3}
-rwxr-xr-x  1 etryzelaar  primarygroup  432480 May  4 19:15 target/release/examples/big1
-rwxr-xr-x  1 etryzelaar  primarygroup  432520 May  4 19:15 target/release/examples/big2
-rwxr-xr-x  1 etryzelaar  primarygroup  432544 May  4 19:15 target/release/examples/big3

% diff <(nm -g target/release/examples/big1) <(nm -g target/release/examples/big2)
192a193
> 00000001000037e0 T _big2
% diff <(nm -g target/release/examples/big1) <(nm -g target/release/examples/big3)
192a193,194
> 00000001000037e0 T _big2
> 00000001000037e0 T _big3