This has shown up a problem with 'case' which is hard to fix - since it
works by generating a function with the appropriate type, it's hard to
ensure that let bindings computational behaviour is propagated while
maintaining appropriate dependencies between arguments and keeping the
let so that it only evaluates once. So, I've disabled the computational
behaviour of 'let' inside case blocks. I hope this isn't a big
inconvenience (there are workarounds if it's ever needed, anyway).
Need to add by full name, due to ordering of loading (the name it's
attached to may not be resolved yet!). This doesn't seem to cause any
performance problems but we can revisit if it does.
Don't use the type of a scrutinee to restrict possible patterns, because
it might have been refined by a Rig0 argument that has a missing case.
Instead, generate all the possible cases and check that the generated
ones are impossible (there's no obvious change in performance)
Small change needed to fix one - assume given implicits which are of the
form x@_ arise from types. It's a bit of a hack but I don't think
there's any need for anything more complicated.
Only valid if unifying the pattern at the end doesn't solve any
metavariables. Also when elaborating applications of fromInteger etc to
constants on the LHS we need to be in expression mode, then reduce the
result later.
This was a slight difference from Blodwen that wasn't accounted for -
there might be lets in the nested environment, so when building the
expanded application type, make sure we go under them
This wasn't necessary before, since we always inlined, but since we can
now postpone things longer and don't always inline until much later, we
need to know what names everything refers to earlier.
Since lookup up a binding can be expensive in a big environment, and we
only need to reduce it if it turns out to be a let, caching it can be a
noticeable win
We can't refine by a name in Rig0 because we can't assume it covers all
possibilities, so only refine by them at the end, at which point we
check that there's only one case left (otherwise we have to match on it,
which is not allowed for an erased thing)
Then on reloading, only decode the Binary when it's first looked up.
This is a huge win on loading because it's decoding the binary blob that
costs, rather than loading it!
Also fix a bug where the elaborator state wasn't updated on completing
the delayed elaborator, which could cause issues with implicitly bound
names in particular.
Need to run delayed elaborators before binding implicits, since there
might be some inside the delayed elaborator. Also reorganise TTC
implementations so they're all in one place.
Needs to be allowed to search for things indexed by pattern variables,
which are not really holes in the same sense, or it might not find
things in types.
(Aside: we're currently implicitly binding failed searches in types,
which is, I think, wrong... better to report the failure)
Don't build these until we've completed the whole clause (that is, don't
do it if we're inside a case block) because we might need to collect
more information to complete checking the case block. e.g. there may be
constraints inside the case block that are resolved outside.
It's just in the way, we don't use it for anything, and everything we
might get out of it we're already getting out of the type. Revisit later
if it turns out to be worth it.
If we store them on the type constructor, we're required to have the
construction available, which might not be the case when reloading ttcs,
because we load and process modules *before* things they import.
With the --yaffle flag, you get the old behaviour which is to invoke the
checker for the core theory (and all the tests are updated appropriately
for this).
This gives useful information for expression search, because we can add
lambdas while we're still building the environment, and start looking at
locals after that.
It turns out we need a new kind of map, mapping possibly ambiguous names
to values. This is what 'Context' does but we can only have one of them
because it resolves names to a position in an array and it'd be
confusing to have more than one index per name.
Mostly direct from Blodwen (some minor modifications to deal with new
way of going into a new scope in the elaborator as well as the usual
bits dealing with name lookup and Glued terms)
And process them on loading. We record that hints need saving out when
adding them, and clear that list unless we happen to be reexporting the
thing we've just read (import public).
Like Idris 1, these are implicitly added on encountering a repeated name
or a non-constructor application. Unlike Idris 1 (and Blodwen) they are
checking by unification rather than matching (which means in particular
that function argument names can't be bound in dot patterns) which is
slightly less expressive, but better overall because matching is
potentially more error prone.
Slight change of plan: instead of having special names, add Lazy, Inf,
Delay and Force and keywords and elaborate them specially.
Correspondingly, add DelayCase for case trees. Given that implicit
laziness is important, it seems better to do it this way than constantly
check whether the name we're working with is important.
This turns out to make implicit laziness much easier, because the
unifier can flag whether it had to go under a 'Delayed' to succeed, and
report that back to the elaborator which can then insert the necessary
coercion.
This shortcuts ambiguity checking by only keeping the things which have
the target type. Given a set of ambiguous applications, the rules are:
- keep any which return the expected target type (a concrete match), or
a polymorphic type (a possible match)
- if there are any concrete matches, drop any possible match which has
the '%allow_overloads' flag set (in practice this will be interface
methods, but other things can be flagged).
- if there ar eno matches, keep everything so that the error messages
show everything that went wrong.
Most notably, when elaborating deferring argument, if the hole
standing for the argument is still a hole, fill it in directly rather
than going via unification. This prevents some needless evaluation.
This slows things down a bit because to find the holes and give them the
right multiplicities, we need to normalise all the arguments which might
have been metavariables. Maybe we should skip this if we're not using
anything linear, for efficiency?
As patterns are handled by deciding which side of the as is considered
'used'. In case blocks, that should be the variable name, but in general
it should be the pattern, so IAs now has a flag to say which one.
Need to record which holes are still to be solved (not counting
elaborations where there are user defined holes) and check at the end
that they are now solved.
There was a check on evaluating lets which was in Blodwen but I hadn't
added to the normaliser yet! Also, normalisation needs to reduce as
patterns for unification, but not when reducing finished LHS and
argument terms. This is a bit of a hack (but then, so is the
implementation of as patterns in general...).
So, when we're checking a nested expression, we have the as pattern as a
let bound variable (so that it has the necessary computational force)
but when we compile we just pass it as an ordinary argument, then it
gets the desired behaviour in case trees.
It's not quite there yet, though, because the treatment of 'as' patterns
isn't quite right and the slightly hacky approach we're taking might not
be the best. Rethinking now...
This is not quite an occurs check (there's no guarantee anyone will ask
for normalisation after all) but it catches the error which will result
if there's a cyclic metavariable solution.
My plan is to implement the occurs check by going through all the
metavariables at the end of elaboration of an expression, so that we
don't do too much work checking half-finished solutions.
Changed nested names (and case blocks) to store the resolved name as the
outer name, rather than the unresolved name, otherwise we'll have issues
when loading from TTC
This is for tracking locally defined names, so that we can expand their
applications to the full globally lifted name. At the moment, they don't
do anything, but since it's a change throughout the elaborator it
probably deserves its own commit.
Since the NF might refer to hole names, and those hole names might be
possible to evaluate now, we'll need to recalculate the expected type's
normal form before rerunning the delayed elaborator
Works by running all possible elaborators and checking that exactly one
succeeds. Still to do: pruning the list of elaborators by target type,
dealing with 'UniqueDefault', checking that delaying on failure works as
it should.
When we encounter them, not that they're a binding as normal, but also
record the thing they expand to. Then bind as a PLet, and convert that
to a Let on the RHS so it has computational force. The case tree
compiler knows about as patterns, so they get compiled to use the
appropriate name on the rhs (rather than a let).
Search on syntax rather than values, since this makes it easier to deal
with local variables accurately, and since we need to go back to syntax
anyway. We ensure that target types are normalised before search.
We'll need this for interface resolution - build up applications of fst
and snd as we go deeper. This also adds options for controlling names of
pairs, proofs, primitives, etc.
