"arithmetic_overflow" and "unconditional_panic" lints do not fire on `const fn` that is only const-used in bin crate

[RalfJung]

I tried the following code:

const fn overflow() -> u32 {
    0 - 1
}

pub const X: u32 = overflow();

fn main() {}

On stable, in release mode, the emits a lint (err-by-default) about the arithmetic overflow. On nightly, it does not.

This probably is related to #78407. However, the part I do not understand is that I can only reproduce the problem with a binary crate, not with a library crate. Cc @oli-obk

[oli-obk]

That is odd indeed. I think it is a bug. Apparently the library crate will build the optimized_mir for overflow even though overflow is not public.

Is it a problem that we do not emit these lints on const fn that are only used in constants?

[RalfJung]

Is it a problem that we do not emit these lints on const fn that are only used in constants?

Yeah I think so. We should properly emit lints for all fn, no matter if they are const or not and no matter if they are used or not and no matter if this is a bin crate or a lib crate.

Emiting lints as part of an optimizations makes this hard. To me that just shows it's a mistake to emit lints as part of an optimization.^^

[oli-obk]

Note that there are a bunch of lints that we do not emit in dead code. I feel like this is a similar situation. Either the lint is triggered on dead code in the const fn, or the interpretation of the const fn as part of the interpretation of a const item will report the same error anyway.

[RalfJung]

Note that there are a bunch of lints that we do not emit in dead code.

If this is a dead code argument, then I assume we certainly should still emit the lint in a lib crate if the fn is pub?

or the interpretation of the const fn as part of the interpretation of a const item will report the same error anyway

No error will be reported in release mode, since overflow checks are disabled. So the fact that there is an overflow is entirely lost when no lint is emitted.

[oli-obk]

I assume we certainly should still emit the lint in a lib crate if the fn is pub?

Yes, and we would, because libraries generate optimized_mir for all their pub functions

No error will be reported in release mode, since overflow checks are disabled. So the fact that there is an overflow is entirely lost when no lint is emitted.

ugh, right.

So... if we want these lints unconditionally, we can run const prop just in analysis mode in the place where we also run borrowck and similar guaranteed MIR passes. in the optimization pipeline we'd then run const prop in mutating mode and possibly even not report any diagnostics from there.

[apiraino]

Assigning P-medium as discussed as part of the Prioritization Working Group procedure and removing I-prioritize.