Panics with `interned value in LRU so must be in key_map`

[Draggu]

on 0.23.0

salsa/src/interned.rs

Line 542 in a0e7a06

.expect("interned value in LRU so must be in key_map")

expect is failing.

[Draggu]

We managed to find out why this does not work.

In the current implementation, when a value is interned, it is hashed and then added to the lru list, then the hash is used in the key_map to store an Id.

When interning value which holds reference to a query return value like this

    #[salsa::interned]
    struct Interned {
        str_slice: &'db str
    }

    #[salsa::tracked(returns(ref))]
    fn my_query() -> String {
        ...
    }

this will read str_slice value while hashing and it will work correctly.
However when this value will be taken from LRU list in new revision, String it points to can be already dealocated if query returned different value.
Then we will try to hash str_slice again and if we are lucky and allocator did NOT reused memory freed by String it will still work, thus making reproduction nondeterministic.
But if it does, we will read a bunch of random bytes leading to wrong hash here and in consequence to panic while removing entry from key_map.

[ibraheemdev]

Ah, I see. Interned values should not actually be allowed to hold references, this is currently unsound. See #985.