LLVM misoptimization causes segfault

[mahkoh]

Consider the following code:

#[no_mangle]
pub unsafe extern fn memset(mut s: *mut u8, c: c_int, n: usize) -> *mut u8 {
    let end = s.add(n);
    let c = c as u8;
    while s != end {
        *s = c;
        s = s.add(1);
    }
    end.sub(n)
}

LLVM misoptimizes this function into an infinite recursion because it recognizes that the inner loop can be replaced by a call to memset. Normally this optimization is disabled inside functions called memset but since the body of the function is moved to a function with rust abi, this check doesn't work.

[rphmeier]

Confirmed here on stable, beta, and nightly using libc 0.2.4. I rewrote your function to use offset instead of "add", along with a relatively minimal test case.

extern crate libc;
#[no_mangle]
pub unsafe extern fn memset(mut s: *mut u8, c: libc::c_int, n: usize) -> *mut u8 {
    let end = s.offset(n as isize);
    let c = c as u8;
    while s != end {
        *s = c;
        s = s.offset(1);
    }
    s
}

fn main() {
    let mut arr = [0u8; 32];
    unsafe { memset(&mut arr[0], 1, arr.len()); }
    assert_eq!(arr, [1u8; 32]);
}

[huonw]

Dupe of #10116 and #18785, but thanks for filing! You may be interested in the no_builtins attribute (as a work-around until #10116 is fixed).