panic msg for slicing used to contain path to user's code, now contains path to core

[carols10cents]

I just noticed this regression from 1.60 to 1.61 because TRPL reasons.

Code

I tried this code:

fn main() {
    let s = "ü";
    println!("{}", &s[0..1]);
}

I expected to see a panic with this message, pointing to the line in my code that caused the panic, which is what I see if I run cargo +1.60 run:

thread 'main' panicked at 'byte index 1 is not a char boundary; it is inside 'ü' (bytes 0..2) of `ü`', 
src/main.rs:3:21

Instead, if I run cargo +1.61 run, I see a panic with a message pointing to a path in core:

thread 'main' panicked at 'byte index 1 is not a char boundary; it is inside 'ü' (bytes 0..2) of `ü`', 
library/core/src/str/mod.rs:127:5

For clearer viewing:

- [...] it is inside 'ü' (bytes 0..2) of `ü`', src/main.rs:3:21
+ [...] it is inside 'ü' (bytes 0..2) of `ü`', library/core/src/str/mod.rs:127:5

The core path is still what I see with cargo +1.63 run.

Version it worked on

It most recently worked on: Rust 1.60.0

Version with regression

It stopped working with 1.61.0

rustc --version --verbose:

rustc 1.61.0 (fe5b13d68 2022-05-18)
binary: rustc
commit-hash: fe5b13d681f25ee6474be29d748c65adcd91f69e
commit-date: 2022-05-18
host: aarch64-apple-darwin
release: 1.61.0
LLVM version: 14.0.0

Other info

What's fun is because of the infrastructure I have with the book, I can see that Rust 1.47 was the first version to point to the user's code; before then it pointed to core as well.

This does NOT seem to occur with all panics coming from core; for example, trying to unwrap a None points to main:

fn main() {
    None.unwrap()
}

thread 'main' panicked at 'called `Option::unwrap()` on a `None` value', src/main.rs:2:10

Indexing out of bounds of an array points to main:

use rand;

fn main() {
    let index: usize = rand::random(); // to get around the compile-time catch of out-of-bound constants
    let foo = [1, 2, 3][index];
}

thread 'main' panicked at 'index out of bounds: the len is 3 but the index is 13912094962097789605', 
src/main.rs:5:15

Interestingly, slicing out of bounds of an array points to core (that is, this is the same issue):

use rand;

fn main() {
    let index: usize = rand::random(); // to get around the compile-time catch of out-of-bound constants
    let foo = &[1, 2, 3][index..index + 1];
}

thread 'main' panicked at 'range end index 17812862265448483785 out of range for slice of length 3', 
library/core/src/slice/index.rs:73:5

so it doesn't appear to be a problem with panic messages in general, but something specific about slicing panics in particular.

[compiler-errors]

I didn't bisect this, but I'm about 90% sure that this is due to #94657 cc @fee1-dead

[fee1-dead]

This is a flaw in the const_eval_select intrinsic's implementation. We currently rewire it to use the implementation which calls the functions from the FnOnce trait. That call cannot be track_caller, and therefore panics will trace to the call instead. The previous design was to use lang item machinery to rewire it to call function items, which would fix this.

[Mark-Simulacrum]

This seems like a pretty unfortunate regression in ability to debug these - do we have line of sight to fixing that? If I follow the PR, it's for unstable indexing or so? (i.e., perhaps we should revert it if we can't improve the diagnostic quality here)?

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-high

[fee1-dead]

I'll try to fix this without reverting. If you don't hear from me within the next four hours, feel free to do a revert.

[fee1-dead]

Tried and it is a bit hard. Feel free to revert now and I will followup later.

[estebank]

Lets make sure we add a regression test, regardless of the solution (revert or fix).

[fee1-dead]

#100759 is in the queue and contains a regression test.