`#[link_section]` does not work in libraries on release

[AsafFisher]

I tried this code while compiling for release:
bin.rs

use elfredo::extended_data;


fn main() {
    print!("{}", do_stuff_with_extended_data());
}

lib.rs

#[link_section = ".extended"]
#[used]
#[no_mangle]
pub static extended_data: u8 = 3;

fn do_stuff_with_extended_data() -> &str{
// Uses extended_data
}

I expected to see this happen (compiled with debug):

asaffisher@ubuntu:/mnt/hgfs/elfredo/target/debug$ readelf -S test_echo_patch_elf | grep extend
  [18] .extended         PROGBITS         000000000003cec8  0003cec8

Instead, this happened(compiled in release):

asaffisher@ubuntu:/mnt/hgfs/elfredo/target/release$ readelf -S test_echo_patch_elf | grep extend
asaffisher@ubuntu:/mnt/hgfs/elfredo/target/release$

Meta

rustc --version --verbose:

rustc 1.53.0-nightly (42816d61e 2021-04-24)
binary: rustc
commit-hash: 42816d61ead7e46d462df997958ccfd514f8c21c
commit-date: 2021-04-24
host: x86_64-unknown-linux-gnu
release: 1.53.0-nightly
LLVM version: 12.0.0

[nagisa]

#[used] can only ensure that the symbols get to the linker, but not that the linker won't remove them. Can you verify that this is not the linker removing your otherwise unused sections?

[AsafFisher]

The section is obviously used, see the println!
I added the #[used]just for the memes...
Also this behavior does not happen on debug.

[nagisa]

It is used at a source level, but there's nothing preventing optimizations from inlining the value of the (immutable) static into the printing routines. At which point the static can become otherwise unused. FWIW I'm not able to reproduce this, regardless:

# rustc lib.rs --crate-type lib -C opt-level=3 --crate-name=elfredo
# readelf -S libelfredo.rlib | grep .extended
  [ 5] .extended         PROGBITS         0000000000000000  0000077c
# rustc main.rs --crate-type bin --extern elfredo=libelfredo.rlib -Copt-level=3 --edition=2018
# readelf -S main | grep extended
  [17] .extended         PROGBITS         0000000000036cc8  00036cc8

The only recommendation I have is the same as before. Verify that it isn't the linker removing this section. Perhaps the easiest way would be to utilize a linker script with KEEP directive or an equivalent inside it.

[AsafFisher]

You are right its an optimization thing. But I still think it is an unwanted behavior.
Why would one ever want to both make sure a variable is allocated at a certain section and to have such an optimization... I think that when you use #[link_section = ".extended"] This behavior is unwanted.

Note that I changed the example

[bjorn3]

Since #91504 on nightly we now have #[used(linker)] behind the used_with_arg feature. #[used] behaves identical to #[used(compiler)]. Note that the linker actually needs to have support for this or it may be discarded anyway. At https://llvm.org/docs/LangRef.html#the-llvm-used-global-variable LLVM does however suggest that #[used(compiler)] should not be exposed to source languages, unlike what rust does.