MIR optimization produces worse code by losing provenance information

I tried this code:

unsafe extern "C" {
    safe fn foo();
}

pub fn example() {
    let x = 1;
    println!("{x}");
    if x != 1 {
        foo();
    }
}

With -O, the foo() call cannot be optimized away. It can be optimized away however, if I use -Zmir-opt-level=0.

https://godbolt.org/z/c5qE7Tvsf

My hypothesis is that:

• format_args takes x by reference
• This x is annotated as noalias when passing to relevant Argument constructors
• If inlining is done by LLVM, it makes use of this information and correctly deduce that x cannot change during std::io::stdio::_print
• If inlining is done by Rust at MIR level, this information is lost, and all LLVM sees is that a pointer is put into Argument and this Argument is passed to std::io::stdio::_print, and hence it has to assume conservatively that the value may change.

This was affecting Rust-for-Linux as we have a build_assert! which performs build-time assertions by requiring calls to be optimized away (Rust equivalent of C's BUILD_BUG_ON), and @joelagnel discovered that adding a print can cause such assertion to trip up.

@rustbot label: +A-rust-for-linux +I-slow +I-heavy