Linking a Rust staticlib unexpectedly changes C math functions from libm to bundled ones from compiler-builtins

[durin42]

Code

Given a staticlib with the source:

#[unsafe(no_mangle)]
pub extern "C" fn demo() -> f32 {
    let x : f32 = 1.1;
    x.ceil()
}

compiled as crate-type = ["staticlib"] and the C program:

#include <math.h>
#include <stdio.h>

float demo();

int main(int argc, char** argv) {
    float x = -1.2;
    printf("%f\n", ceilf(x));
    printf("%f\n", demo());
    return 0;
}

When linking the program like this: $(CC) main.o as-crate/target/debug/libas_crate.a -lm -o out/tgt-from-crate the resulting binary ends up using ceilf from compiler-builtins rather than libm.

I expected to see this happen: The binary should still get ceilf from libm, eg

$ nm -uD out/tgt-from-crate | grep ceilf || true
                 U ceilf@GLIBC_2.2.5

Instead, this happened: Instead, the ceilf symbol was satisfied by the matching symbol in compiler-builtins, and so the C call was surprise-migrated to a different implementation.

Version it worked on

1.86
rustc --version --verbose:

rustc 1.86.0 (05f9846f8 2025-03-31)
binary: rustc
commit-hash: 05f9846f893b09a1be1fc8560e33fc3c815cfecb
commit-date: 2025-03-31
host: x86_64-unknown-linux-gnu
release: 1.86.0
LLVM version: 19.1.7

Version with regression

1.87
rustc --version --verbose:

rustc 1.87.0 (17067e9ac 2025-05-09)
binary: rustc
commit-hash: 17067e9ac6d7ecb70e50f92c1944e545188d2359
commit-date: 2025-05-09
host: x86_64-unknown-linux-gnu
release: 1.87.0
LLVM version: 20.1.1

Additional information

I suspect this is related to some work tracked in #137578. I also found out Chromium has already stumbled on this: https://issues.chromium.org/issues/419258012#comment5 and for them it caused observable behavior changes. Apologies for not catching this sooner - we hadn't upgraded compiler-builtins for a while and didn't notice until a couple of weeks ago. :(

[Noratrieb]

bisection script

#!/usr/bin/env bash

set -ex

rustc --crate-type=staticlib a.rs
cc main.c liba.a -lm -o main

not nm -uD main | rg ceilf

searched nightlies: from nightly-2024-05-01 to nightly-2025-05-20
regressed nightly: nightly-2025-03-06
searched commit range: f9e0239...30f168e
regressed commit: ac951d3

bisected with cargo-bisect-rustc v0.6.8

Host triple: x86_64-unknown-linux-gnu
Reproduce with:

cargo bisect-rustc github --start 2024-05-01 --end 2025-05-20 --script repro.sh 

rust-lang/compiler-builtins#763

[Noratrieb]

It looks like rust-lang/compiler-builtins#763 just doesn't work, even a normal rust main reproducer picks ceilf from compiler-builtins instead of libc.

fn main() {
    let x : f32 = std::hint::black_box(1.1);
    dbg!(x.ceil());
}

nora@nixos /t/scratchBGHE6xUbh> rustc a.rs
nora@nixos /t/scratchBGHE6xUbh> nm -uD a | rg ceil
nora@nixos /t/scratchBGHE6xUbh [0|1]> rustc +1.84.1 a.rs
nora@nixos /t/scratchBGHE6xUbh> nm -uD a | rg ceil
                 U ceilf@GLIBC_2.2.5

[Noratrieb]

I'm a bit confused, this case for pure-rust programs was reported in #139487 but closed. But rust-lang/compiler-builtins#763 intended to not override anything, so that still seems like a bug? This here is just an extension of it where it will infect FFI as well, instead of just Rust.
@tgross35 is it intended that Rust programs now always call the compiler-builtins version?

[Noratrieb]

#137578 (comment) @Amanieu

So I previously thought that statically linked weak symbols always override strong dynamic symbols, but this may not actually be the case

It is indeed not the case. https://gist.github.com/Noratrieb/a58000d43527225ed5af0032890e00a5
This prints eak, so the weak symbol, instead of uwu, which would be the strong dynamic symbol.