Test polymorphic_recursion.rs takes a long time to compile with a debug-assertions=true compiler

[Zalathar]

#120912 reported that tests/mir-opt/inline/polymorphic_recursion.rs became very slow after #120584. A workaround for that slowness was added by a hack in #120942.

However, the test is still very slow when the compiler is built with [rust] debug-assertions=true.

On my system (Apple M2 Pro):

• Before the original change, it was fast enough that I never noticed it being slow when running x test mir-opt.
• After the original change but before the hack, it appeared to hang indefinitely (though perhaps it was just running very very slowly).
• After the hack, it still takes about 45 seconds to complete.

// skip-filecheck
// Make sure that the MIR inliner does not loop indefinitely on polymorphic recursion.
// compile-flags: --crate-type lib

// Randomize `def_path_hash` by defining them under a module with different names
macro_rules! emit {
    ($($m:ident)*) => {$(
        pub mod $m {
            pub trait Tr { type Next: Tr; }

            pub fn hoge<const N: usize, T: Tr>() {
                inner::<N, T>();
            }

            #[inline(always)]
            fn inner<const N: usize, T: Tr>()
            {
                inner::<N, T::Next>();
                inner::<N, T::Next>();
            }
        }
    )*};
}

// Increase the chance of triggering the bug
emit!(m00 m01 m02 m03 m04 m05 m06 m07 m08 m09 m10 m11 m12 m13 m14 m15 m16 m17 m18 m19);

[Zalathar]

I'm not sure what should be done here.

Obviously I can work around this issue, by just avoiding that specific test.

But I don't like the idea of just leaving this test lying around in its current state, if it becomes catastrophically slow in an otherwise-reasonable (dev) build configuration.

[compiler-errors]

:(

I wonder why it still takes so long. Perhaps it's just because of extremely deep type substitution? Let me try a few more hacks, I guess.

[compiler-errors]

#121123 will help; let's see if i can get that into a state that's landable though.

[Zalathar]

As of #121123 this test now takes about 4s for me, which is a lot less than 45 seconds, but still several times longer than the rest of mir-opt put together.