translating statics/consts segfaults LLVM if order is wrong

[nikomatsakis]

This test case (a permutation of const-autoderef.rs) causes rustc to segfault (somewhere in LLVM):

static B: u8 = (&A)[0];
static A: [u8; 1] = ['h' as u8];
static D: u8 = (&C)[0];
static C: &'static &'static &'static &'static [u8; 1] = & & & &A;

pub fn main() {
    assert_eq!(B, A[0]);
    assert_eq!(D, A[0]);
}

Uncovered while working on separating items from HIR, since the new visitors can sometimes visit items in different orders (unless you force them to do otherwise).

[nikomatsakis]

I believe the cause of this is #26630.

[nikomatsakis]

cc @eefriedman -- this is triggered by recursive statics (see also some notes on #29719, if you haven't already)

[arielb1]

duplicate of #28324 (or pseudo-duplicate - the issue was intended to capture the "actually linkable" case too).

[nikomatsakis]

@arielb1 ah, yes, I remember this issue. And there is a legit question also of whether we should support the kind of derefs etc that are evident here vs having people create intermediate statics. I personally am not terribly fond of that, since the value of a static is not necessarily fixed over the course of an execution.

[nagisa]

Reminder, that LLVM asserts turn into undefined behaviour (I believe?) when built in release mode. Probably would be good to compile a debug LLVM to see what’s the assert?

[arielb1]

@nagisa

The issue is known - we are attempting to dereference a static that is not defined.

[eefriedman]

This is fixed.

[nikomatsakis]

The original test case, indeed, does not seem to compile.

[nagisa]

Shouldn’t we add a regression test or something?