The hack (optimisation?) to normalise integer literals when below some
threshold is fine on the RHS, but on the LHS causes problems since we
need them in normal form for pattern matching. Fixes#112
This improves coverage checking, because it can now see that things like
(Z = S x) and (x = S x) are empty. Previously, it only checked that all
possible constructors had a disjoint index. Now, it looks for matches
and checks for disjointness in the matches, which catches a lot more
things especially with equality.
This is added to functions which are guaranteed to be productive. The
check is currently very conservative - just added when every clause is
constructor headed (or headed by an AllGuarded function), and there are
no other function applications.
This allows a function to be inlined for totality checking purposes
only. So, for example, (>>=) might be a function, but if it evaluates to
something constructor guarded in some context, then it might still be
productive.
It needs to take into account that solving other names might cause
unification errors to succeed, so only give up if there's conflicting
concrete constructors
If we never evaluate under Delay at all, we won't inline interface
methods, which means productive things defined in an interface can never
be today. So, make sure to set the tcinline flag before quoting the
Delayed closure.
If we're delaying because of a unification failure, there's no point in
trying again. This can massively speed up (or maybe I should say "unslow
down") error reporting if there's ambiguity because the elaborator won't
be exploring some paths which can never succeed.
Run once ignoring errors to make progress on interfaces/determining
arguments, then again in full. Second step isn't needed since it was
just covering up an earlier bug.
This means that some errors under lots of delays are reported quicker,
though I still haven't completely got to the bottom of that one.
