Improve handling of types in constraint system.

[rudkx]

Several fixes in order to make handling of types more consistent.

In particular:

• Expression types used within the constraint system itself always use
  the type map.
• Prior to calling out to any TypeChecker functions, the types are
  written back into the expression tree. After the call, the types are
  read back into the constraint system type map and cleared from the
  expression tree.
• Some calls to directly set types on the expressions are reintroduced
  in places they make sense (e.g. when building up new expressions that
  are then passed to typeCheckExpressionShallow).

ConstraintSystem::setType() is still setting the type on the expression
nodes at the moment, because there is still some incorrect handling of
types that I need to track down (in particular some issues related to
closures do not appear to be handled correctly).

Also, the diagnostics have not been updated at all to use the types from
the constraint system where appropriate.

[rudkx]

@swift-ci Please test

[rudkx]

@swift-ci Please test

[swift-ci]

Build failed
Jenkins build - Swift Test Linux Platform
Git Commit - fefae4979f12253c55097891d50be0e81d0422ab
Test requested by - @rudkx

[swift-ci]

Build failed
Jenkins build - Swift Test OS X Platform
Git Commit - fefae4979f12253c55097891d50be0e81d0422ab
Test requested by - @rudkx

[rudkx]

@swift-ci Please test

[swift-ci]

Build failed
Jenkins build - Swift Test Linux Platform
Git Commit - 98fc1ab3211e23fee80b2d7ec4b08455fccc9b03
Test requested by - @rudkx

[swift-ci]

Build failed
Jenkins build - Swift Test OS X Platform
Git Commit - 98fc1ab3211e23fee80b2d7ec4b08455fccc9b03
Test requested by - @rudkx

[rudkx]

@swift-ci Please smoke test

[DougGregor]

These places where we're calling back into the TypeChecker... should they be folded into the ConstraintSystem so that we don't have to do the write-back-into-expressions step? Is this mainly callWitness?

[rudkx]

Yes it's possible a few of these should be moved into the constraint system. I did that with one or two functions in a previous commit. I looked at callWitness and it appeared to be used from other places outside of CSApply so I didn't look at moving it in.

There are a few other places where we are currently writing types back, but I did that somewhat defensively at the moment to get closer to something that works. I intend on going back and removing ones that I suspect aren't actually needed (e.g. typeCheckExpressionShallow generally only looks at the untyped top-level node so perhaps it's not needed).

It's difficult to do that as we're still writing the types in ConstraintSystem::setType(). I may be able to determine that today before actually merging any of this in.

[DougGregor]

For something like callWitness, we could still provide a TypeChecker version that constructs a constraint system, records the types of the subexpressions in the constraint system, and then hops over to the implementation that doesn't ever use Expr::getType().

Not that we have to do the above before merging this excellent PR; I'm thinking about the trajectory here.

[rudkx]

Okay I found a pile of new work I had missed originally where we pass Expr* into functions in Expr.cpp that read and write types, so I'm going to use this as a checkpoint and then address those issues. It looks like at least some of that functionality is only used from the constraint system and should just be moved there, and other pieces can probably be addressed in other ways.

I think I'm going to try to encapsulate as much of the cruft here (by cruft I mean reading/writing types back and forth) into some entry points in ConstraintSystem that do the calls into the type checker so that there is one place it's happening rather than N.

[rudkx]

@swift-ci Please smoke test and merge

[rudkx]

@swift-ci Please smoke test