If IO operations don't return an updated world token, the inliner might
think they're not used at all. So if this turns out to be an overhead,
we'll have to work out another way of eliding it.
Or, more accurately, by not doing a thing that's a waste of time. After
evaluating a local, if the result isn't applied to a stack, there's no
more reduction to do, so stop there.
Still go to the explicit named representation afterwards, since that's
an easier API for a code generator, and by then the names are guaranteed
unique.
Apologies to anyone who's working on a back end independently!
You can still use CExp for now, but I'm shortly going to try using
NamedCExp in inlining, to save the fairly large cost of maintaining the
de Bruijn indices when going under binders.
Names are kept unique in the translation, so you can assume you don't
need to worry about them.
Use a variant of MkVar which has the name in the type and therefore
means it can be erased to a newtype, so just an Integer. Then use
'insertNames' rather than repeatedly thinning.
A bit of reorganisation to ensure that the context membership proofs are
erased. It's only a small overhead in general to keep the proofs, but
there are places where it's really noticeable, particularly when
compiling large programs.
It's still too slow, but better than it was. Instead of all the
weakening, invent a new name when evaluating under a lambda and
substitute later.
Possibly we can improve this later by substituting in batches, like the
main evaluator does.
e.g. in a C file. This means we don't accidentally treat things as
empty, since previously we just defined these as empty types, but that
broke coverage checking. Fixes#240
The names the locals were being applied to weren't being updated
properly, so applications of local functions inside case blocks were
sometimes given the wrong arguments. This is one of the few places where
it's hard to keep track of names in the type system! So naturally that'd
be where things go wrong I suppose...
We never inspect it, so it carries no information - it just needs to be
there as a token to make sure that IO operations run at the right time.
So, IORes can be a newtype now and therefore optimised away.
Data types with one constructor, that has one unerased argument, are
translated to that argument for runtime. This doesn't have a huge effect
on its own, but doing this will expose other optimisations later (e.g.
increasing effect of inlining).
Specialisation happens on the run time case tree, so we need to know
separately which names occur at compile time, and which at run time.
Hence, we now have 'refersToRuntime' in addition to 'refersTo' since
they will be different. (In fact there's more differences, because
there's a lot of metavariable names which are needed at compile time but
erased at run time)
Make desugaring/elaboration of interfaces, interface implementations,
records, and parameter blocks take into account the pragma
`%unbound_implicits off`.
Main changes:
(a) Execute the pragma also during desugaring
(b) Check whether `isUnboundImplicits` is on at each desugaring step
Alternatives I didn't take:
(1) Changing `findBindableNames` to effectfully check the flag.
Rationale:
Apart from turning a pure function into an effectful one, this
would mean repeatedly calling `findBindableNames`, only to do
nothing once the flag is read.
(2) Adding another function that takes multiple places (list of terms)
that might contain bindable names, and before dispatching
`findBindableNames` on each term, checking the flag.
Rationale: I didn't want to add another abstraction. (weak
rationale)
@edwinb @gallais : if you prefer (2), I can do that.
My first attempt at refolding natural constants led to all names
being printed with their namespace prefixes. This fixes that,
most notably by introducing the massive `mapPTermM` function in
`Idris.Syntax`.
