opportunistically evaluate constants while populating `mir::Body::required_consts`

[oli-obk]

We can evaluate all constants that do not depend on generic parameters in

rust/src/librustc_mir/transform/required_consts.rs

Line 19 in 0b95879

if let ConstKind::Unevaluated(_, _, _) = const_kind {

We make that a MutVisitor that evaluates all constants that it can and replaces them in-place and only adds those that it couldn't evaluate to the required_const list.

This way we never have to evaluate another constant in the entire MIR optimization pipeline, because we know none can be evaluated (modulo those that we already evaluate for merging required_consts during inlining.

cc @rust-lang/wg-mir-opt I think I like having the above invariant. It will make working with constants in MIR much simpler and make any kind of error handling that would have to be implemented unified into a single location.

[RalfJung]

and make any kind of error handling that would have to be implemented unified into a single location.

Ideally that would actually dispatch through the const_eval error handling to have it all centralized there...

Same goes for the const error reporting that codegen is doing, FWIW.

Which makes me wonder, can we move this impl block to librustc_mir/const_eval/error.rs or so? This is solely CTFE error logic, so logically part of the "CTFE <-> rustc glue" that we otherwise have in librustc_mir::const_eval.

[oli-obk]

hmm right, I think nothing from librustc uses this anymore. We can't move the impl block, but we can make the functions free functions.

[RalfJung]

Yeah something like that.

So we also use this issue to track cleaning up the error reporting story more generally? If we also include cleaning up this then I think we can close #67191, AFAIK that's all that's left.

[RalfJung]

rust-lang/miri#1382 is also related, this is Miri lacking the code that codegen has to properly report post-monomorphization errors.