`alloc-variant-zeroed` causing issues with LTO

[clubby789]

I believe this causing https://crbug.com/441524277. In summary, the rust calloc function _RNvCsfo6nMSlQo1l_7___rustc19___rust_alloc_zeroed definition does not have the alloc-family attribute, but user modules of it have the alloc-family attribute on the calloc declaration. however, if ThinLTO decides to import the function, we lose the alloc-family attribute and the verifier complains 'alloc-variant-zeroed' must name a function belonging to the same 'alloc-family'.

Originally posted by @aeubanks in #144086 (comment)

[aeubanks]

followup to #144086 (comment)

$ cat src/main.rs

mod foobar {
    use std::alloc::{GlobalAlloc, Layout};

    struct Allocator;

    unsafe impl GlobalAlloc for Allocator {
        unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
            panic!()
        }

        unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
            panic!()
        }

        unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
            panic!()
        }

        unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
            panic!()
        }
    }

    #[global_allocator]
    static GLOBAL: Allocator = Allocator;
}


fn main() {
    println!("Hello, world!");
}

$ RUSTFLAGS='--emit=llvm-ir' cargo build --release

no sign of the alloc-family attribute in the .ll file, which does contain definitions for the alloc functions

[apiraino]

Assigning priority (discussion on Zulip).

@rustbot label -I-prioritize +P-high

[aeubanks]

any update on this? we have #144086 reverted in our ToT rustc build but that's not sustainable

[wesleywiser]

Visited during compiler team triage meeting, we recommend considering a revert & re-apply of the injecting PR after this issue has been resolved.