Associated-const ranges in patterns are miscompiled in beta

[bytwise]

This code demonstrates the problem (playpen):

#![allow(dead_code)]

trait Range {
    const FIRST: u8;
    const LAST: u8;
}

struct OneDigit;
impl Range for OneDigit {
    const FIRST: u8 = 0;
    const LAST: u8 = 9;
}

struct TwoDigits;
impl Range for TwoDigits {
    const FIRST: u8 = 10;
    const LAST: u8 = 99;
}

struct ThreeDigits;
impl Range for ThreeDigits {
    const FIRST: u8 = 100;
    const LAST: u8 = 255;
}

fn digits(x: u8) -> u32 {
    match x {
        OneDigit::FIRST...OneDigit::LAST => 1,
        TwoDigits::FIRST...TwoDigits::LAST => 2,
        ThreeDigits::FIRST...ThreeDigits::LAST => 3,
        _ => unreachable!(),
    }
}

fn main() {
    println!("{}", digits(100));
}

With stable-1.32.0 it prints 3. With beta-1.33.0 it wrongly prints 1.

Beta also produces these wrong warnings:

warning: unreachable pattern
  --> src/main.rs:29:9
   |
29 |         TwoDigits::FIRST...TwoDigits::LAST => 2,
   |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   |
   = note: #[warn(unreachable_patterns)] on by default

warning: unreachable pattern
  --> src/main.rs:30:9
   |
30 |         ThreeDigits::FIRST...ThreeDigits::LAST => 3,
   |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

It seems that Rustc decides, that the first arm covers the whole range of a
u8 and optimizes out the other arms before they even reach MIR.

The code is very similar to this testcase.
But the testcase does not test that the other arms can be reached.

I was able to bisect the regression to #55937.

(This issue seems similar to #57894, but the reduced testcase there does not
use any associated consts)

[jonas-schievink]

Seems like a duplicate of #57894 (that issue applies to range matches in general, whether you're using an assoc. const or a literal doesn't matter) maybe not

Not the same as #57894

[pnkfelix]

triage: P-high, A-NLL, NLL-complete, assigning to self.

[pnkfelix]

cc @davidtwco

[pnkfelix]

I have confirmed that the bug appears to be specifically injected by commit 4be7214

(I've skimmed over that commit several times and still can't see the problem, but I'm going to try again now. Maybe the actual issue is that we are miscompiling PatternKind::AscribeUserType downstream, but that pattern is not sufficiently exercised elsewhere in the compiler today...)

[pnkfelix]

Well this is making my spidey-sense tingle:

rust/src/librustc_mir/build/matches/test.rs

Line 99 in 1aa2506

pub fn add_cases_to_switch<'pat>(&mut self,

rust/src/librustc_mir/build/matches/test.rs

Lines 134 to 139 in 1aa2506

	PatternKind::AscribeUserType { .. } |
	PatternKind::Leaf { .. } |
	PatternKind::Deref { .. } => {
	// don't know how to add these patterns to a switch
	false
	}

but even when that returns false, it seems like that would only imply that we fall through to presumably more general purpose code here:

rust/src/librustc_mir/build/matches/mod.rs

Line 1041 in 1aa2506

// perform the test, branching to one of N blocks. For each of

(still looking. Going to add a bit of instrumentation to these parts of the compiler.)

[davidtwco]

@pnkfelix I don't mean to step on any toes, but I looked into this a little as an earlier PR of mine introduced it. I've got a fix up at #58161 but if that's not the approach we want to take then I can leave this to you.