Optimizer believes vector's length field may alias with its own buffer

[meithecatte]

Consider the following code:

pub fn treat_vec(v: &mut Vec<usize>) {
    if v.len() == 1 && v[0] == 42 {
        assert!(v.pop() == Some(42))
    }
}

(Playground)

I'd expect the assertion to get optimized-out. However, the following assembly is generated:

	pushq	%rax
	cmpq	$1, 16(%rdi)
	jne	.LBB0_4
# %bb.1:
	movq	(%rdi), %rax
	cmpq	$42, (%rax)
	jne	.LBB0_4
# %bb.2:
	movq	$0, 16(%rdi)
	cmpq	$42, (%rax)
	jne	.LBB0_3

.LBB0_4:
	popq	%rax
	retq

.LBB0_3:
	leaq	.L__unnamed_1(%rip), %rdi
	leaq	.L__unnamed_2(%rip), %rdx
	movl	$37, %esi
	callq	*core::panicking::panic@GOTPCREL(%rip)
	ud2

The only instruction between the two cmpq $42, (%rax) comparisons is a write to 16(%rdi). Thus, LLVM seems to believe that these two locations may alias.

[nagisa]

Duplicate of #54878. Thanks!

[nagisa]

And/or #71354

[jrmuizel]

Yep, this is fixed in Nightly