Things to fix once specialization is in stable

[sgrif]

This list is probably non-exhaustive, but we should try and add to it as we remember issues or new ones come up

• We are currently brute forcing the impl of FromSqlRow. This should entirely be covered by FromSql implies FromSqlRow and our tuple impls. The addition of the DB parameter made this invalid under coherence, but we can hopefully go back to our old system w/ specialization
• We should be able to have FromSqlRow imply Queryable
• We should be able to remove the selectable_column_workaround! macro, and be able to write impl<Left, Right, T, ST> SelectableColumn<JoinSource<Left, Right>, ST> for T where T: SelectableColumn<Left> and impl<Left, Right, T, ST> SelectableColumn<JoinSource<Left, Right>, ST> for T where T: SelectableColumn<Right>. This would have required lattice
• Replace SupportsReturningClause and SupportsDefaultKeyword with a blanket impl and a refinement for SQLite. This will likely mean that we can simplify several where clauses as well. (As far as I'm aware, SQLite is the only case where these aren't supported. Other backends should be able to specialize if this is not the case). The where clauses about Insertable in the migrations mod can go away after this. What we really want here is negative reasoning. Most of these cases we don't want to specialize for SQLite, we want to implement a trait for everything except SQLite.
• Specialize the QueryFragment<Sqlite> impl for Now -- SQLite does not have a now function. While the string 'now' is technically valid on all platforms, I'd prefer to write something a little bit more "correct"

[sgrif]

We can probably use default impl/partial impl to make it easier to add new AST passes. We have a lot of AST nodes that are composed of exactly 1 or 2 elements, we can add a BinaryQueryFragment trait and do something like:

default impl<DB, T> QueryFragment<DB> for T where
    DB: Backend,
    T: BinaryQueryFragment<DB>,
{
    fn is_safe_to_cache_prepared(&self) -> bool {
        self.left().is_safe_to_cache_prepared() &&
            self.right().is_safe_to_cache_prepared()
    }
}

Right now I'm having to manually add these for a ton of nodes, and it's quite painful

[sgrif]

We should see if we can remove the need to ever specify QueryId as a constraint by adding a blanket impl for all T, maybe with the constraint of T: QueryFragment<Debug> or something. If I could, I would make QueryId a supertrait of QueryFragment but that would mean that I have to specify the query id type when it is boxed, which we don't want to do.

But right now if a type forgets to implement it, the error message is really hard to track down. It's similar to if QueryFragment wasn't implemented, but that's much less likely.

[sgrif]

I'd like to have

diesel/diesel/src/migrations/connection.rs

Line 23 in d120a23

for<'a> &'a NewMigration<'a>: Insertable<__diesel_schema_migrations, T::Backend>,

just be for<'a> &'a str: ToSql<VarChar, T::Backend>, and use an ad-hoc insert in that file. However, that's going to fail because ColumnInsertValue requires that DB: SupportsDefaultKeyword or DB == Sqlite. With specialization we can make the SQLite impl a blanket impl, and then specialize it for DB: SupportsDefaultKeyword.