Stop generating an intermediate deref when taking references to statics

[oli-obk]

r? @RalfJung @ecstatic-morse

So... the motivation for this change was to make things less fragile. There's still work to be done if we really want to go all the way to having more Rvalues for different ways to interact with statics and fully stop relying on metadata of locals for soundness.

fixes #54366

[ecstatic-morse]

This is bad. We no longer catch this pre-monomorphization even though it has&mut. You'll need to handle this case somehow.

[oli-obk]

I tried, but I see no way to distinguish this from the &mut STDERR_BUFFER_SPACE that we'd see for just let x = STDERR_BUFFER_SPACE without going all the way to having custom Rvalues for that

[ecstatic-morse]

There's always the HIR const-checker I guess 😆. I haven't got around to removing it yet.

As written, this PR allows stuff like let x = &mut STATIC_MUT in a const-context on stable, so you'll need to handle it somehow.

[oli-obk]

Unfortunately

struct Foo {
    field: *mut i32
}

static mut FOO: Foo = Foo {
    field: &mut [42] as *mut [i32] as *mut i32,
};

is stable, so I can't just ban mutable references.

[oli-obk]

But yea, I gottaa figure something out how to not expose it even more...

[RalfJung]

fixes #54366

That's a very different issue that the last PR with a similar title (#76982)? I am somewhat confused. I assume this is supposed to resolve #74840?

Also could you describe at a high level what the approach to solving #74840 is, compared to #76982? That would be easier than trying to reverse engineer that from the PR.^^

[oli-obk]

I'm trying to get rid of our reliance on the local_info for the soundness of accessing static items. The fact that this now fixes #54366 is a total unintended side effect that I only noticed because the canary test for it now fails.

This is not related to #74840 anymore, as that issue is now solely about making uninhabited extern types forbidden via a future incompat lint.

This PR takes &FOO expressions and directly converts them to constants of pointer type instead of having &*&FOO where the trailing &FOO is a pointer type. Though that causes us to stop being able to distinguish the user writing let x = FOO; and the user writing let x = *&FOO;, causing the former to now fail to compile in case FOO is a mutable static. So as it stands, this PR cannot work

[oli-obk]

I have explored the design space around Rvalue::ReadStatic and similar locally and have come to the conclusion that they are not feasible since they mostly just amount to more complexity without any reduction.

We're in a fairly good sweet spot with these &*&STATIC desugarings.

That said, we can still do the active PR once mutable references and dereferencing becomes legal in statics

[RalfJung]

Okay, I opened #77096 for the problems the current static handling is causing.