mir::Constant::span is a footgun for mir optimizations

[oli-obk]

The field is part of the PartialEq impl, and thus two constants that are equal except in their span will compare inequal. This causes problems in #77551 (comment) and #77549 (comment) and possibly in other places.

The "trivial" way to resolve this is to manually implement PartialEq and make it skip the Span field.
Downside: now we may cause weird diagnostics after such an optimization has been applied, because if one of the constants is eliminated, all diagnostics are reported on the other diagnostic's site. We'd need a multispan and properly combine it. That will be hard to ensure to happen if we just silently change PartialEq

cc @rust-lang/wg-mir-opt

[tmiasko]

The issue also applies to implicit const () at the end of block. Replacing it with an empty tuple aggregate in a MIR builder triggers a few extra optimizations in mir-opt test suite.

[camelid]

So are we going to use the "trivial" fix – adding a custom impl of PartialEq that ignores span – or some kind of more complex one – I think you mentioned a "multispan"?

[oli-obk]

Ideally we'd remove the PartialEq impl and leave a comment that it can't have that impl, but I'm not sure what the fallout of that is.

[camelid]

Perhaps we could add a method on Constant that checks equality while ignoring span?

[simonvandel]

Only adding a new method that ignores span equality does not remove the footgun though. Other code would still use the PartialEq which still compares unintuitively because of the span.

[camelid]

I meant that we should remove PartialEq as Oli suggested and then add a new method that explicitly ignores span.

[camelid]

@spastorino Is it okay if I work on this a bit?

[spastorino]

If someone wants to work on this, feel free to reuse my code. I don't remember how useful it is anymore since it has more than 1 year :). Anyway, feel free to contact me on a Zulip topic.