Between opt-level=1 and opt-level=2, LLVM deoptimizes the output of ptr::addr, if it ptr::addr is #[inline]

[saethlin]

godbolt demo: https://godbolt.org/z/51o3hYjvv

This code:

#[inline]
fn inline_addr<T>(ptr: *const T) -> usize {
    unsafe {
        core::mem::transmute(ptr)
    }
}

pub fn inline_is_nonoverlapping(src: *const u8, dst: *const u8) -> usize {
    let src_usize = inline_addr(src);
    let dst_usize = inline_addr(dst);
    if src_usize > dst_usize { let _ = src_usize - dst_usize; }
    src_usize
}

At opt-level=2 compiles to this:

example::inline_is_nonoverlapping:
        cmp     rdi, rsi
        jbe     .LBB0_3
        jb      .LBB0_2
.LBB0_3:
        mov     rax, rdi
        ret
.LBB0_2:
        push    rax
        lea     rdi, [rip + str.0]
        lea     rdx, [rip + .L__unnamed_1]
        mov     esi, 33
        call    qword ptr [rip + core::panicking::panic@GOTPCREL]
        ud2

Decreasing to opt-level=1 or removing the #[inline] on the transmute wrapper function causes the panic path to go away.

As far as I can tell, this doesn't happen on stable but it happens on beta and nightly. This code pattern is used in the standard library here:

rust/library/core/src/intrinsics.rs

Lines 2234 to 2238 in cb94675

	pub(crate) fn is_nonoverlapping<T>(src: *const T, dst: *const T, count: usize) -> bool {
	let src_usize = src.addr();
	let dst_usize = dst.addr();
	let size = mem::size_of::<T>().checked_mul(count).unwrap();
	let diff = if src_usize > dst_usize { src_usize - dst_usize } else { dst_usize - src_usize };

@rustbot label +A-LLVM +regression-from-stable-to-beta

[nikic]

Works fine with LLVM 16: https://llvm.godbolt.org/z/67WPE19xa Presumably due to llvm/llvm-project@926e731.

Assigning to myself to check back on this after the next LLVM update.

[nikic]

Eh, we might as well fix this on the rustc side: #103299

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-medium