This means we delay ambiguity resolution properly if we hit a depth
error, which means that we can come back to it later and hopefully be
able to resolve it by type on a second attempt.
This is the most likely cause of type checking being too slow, at least
in cases where there's not too much type level computation, so rather
than explore for ages, give up after going under too many ambiguous
names during elaboration.
The depth limit is configurable, with %ambiguity_depth <level>, but it's
very unlikely you'll ever want to do that rather than disambiguating
explicitly.
Do two passes through delayed elaborators (it's impossible to predict
dependency order). Possibly this should keep going as long as it's
making progress? I'll consider that later.
This is a tweak we used in Idris 1: In the end, if there is ambiguity,
and some names generate unification constraints but exactly one doesn't,
take the one that doesn't.
If we have lots of delayed elaborators, and more than one fails, we need
to print the name ambiguity error first since that is the most likely
cause of any other errors.
'case' goes first, because it might help resolve ambiguities. Ideally
we'd keep retrying delayed elaborators until they work, but that's going
to get slow...
This is needed for consistency with the self hosted version (where the
names aren't even available at that point - so it turns out we learn
something from erasability :))
We can work out what it should be since each case is supposed to have
the same type, so we can look ahead to the alternatives. This way, we
don't have to rely on the scrutinee alone to calculate the type of the
case block. (Idris 1 does this too but I've only just encountered the
need for it in Idris 2 now!)
Sorry for the less than informative title :). Lots going on here. It
started as an attempt to fix unification to deal with laziness coercions
in trickier places, but unearthed a couple of tricky and interconnected
issues that are hard to unpick into a single patch. So, this fixes a few
things:
- default hints should only be resolved on the current elaboration (e.g.
nested function definitions, not the outer definition which might not
yet be complete)
- delayed elaborators should be allowed to have nested delayed
elaborators, which means disambiguation is a little bit better
- we should delay elaborating arguments where the type isn't known yet,
because later arguments may resolve the type, and we can use this to
help with disambiguation/laziness coercions
- other bits and pieces arising
Lots of time is taken in finding the definitions to compile, mostly
because the definition itself needs to be decoded from the binary format
in the ttc, but we can at least make the rest of the process fast.
Inlining takes a while - we could improve this later by inlining as we
go, per file, rather than doing it just before compilation.
When we're checking if a definition is a hole and needs updating, when
loading ttc, the names might not fully resolved yet, so don't decode the
definition.
This is done during elaboration, but not for things which are solved by
unification or search, and they might be 0 if, say, they're a record
projection or other type level function.
Mostly if they arise from search, where the target of the search has a
let in the type. So, we need to consider this when checking
metavariables in the linearity check.
The old way did a lot of needless traversal, though you wouldn't
typically notice until you start having 40+ modules to build. Now make a
note of what we've looked at already.
Record name and constructor are in the outer namespace, fields in the
nested namepsace. Private names in the nested namespace is visible from
the outer namespace. Fixes#183
Also remove the 'nested' flag on INamespace, since it was only there for
records, and not really the right way to do things. The nested namespace
name is now given in the IRecord constructor (a front end might not want
to use a new namespace after all).
