Simplify getfield and isdefined tfuncs by treating DataType as immutable

[martinholters]

And a little cleanup: Neither UnionAll nor SimpleVector should need special-casing. The former is immutable and gets handled as well as any other immutable, the latter has no fields.

Technically, DataType can be mutated, but actually doing that is looking for more trouble than just wrong constant propagation, and we've been doing this for some fields of it anyway. But maybe we can throw in the setfield builtin when trying to modify a DataType? Or is it even possible to declare DataType immutable?

[JeffBezanson]

The types field can be undefined and then get set later.

[martinholters]

Ah. That's one of those where is_dt_const_field returned true, though.
Could we by any chance just trigger setting it here?

[vtjnash]

It can also get cleared later

[martinholters]

So you're saying that

julia> foo() = isdefined(VersionNumber, :types)
foo (generic function with 1 method)

julia> @code_typed foo()
CodeInfo(
1 ─     return true
) => Bool

is bad because foo won't get recompiled when VersionNumber.types gets cleared and foo() will then wrongly still return true? That is bad, of course, but nothing this PR introduces. It's the same on 1.0, 1.5, and master. I'm not sure how big an undertaking fixing that would be, considering that some code (e.g. fieldcount) might rely on this (and getfield) being const-inferred.

[martinholters]

Okay, I managed to construct a test case for the "types may get set later" case:

g37293(::Type{T}) where T = isdefined(T, :types)
struct T37293; x::Val{g37293(T37293)}; end # call g37293 on incomplete type with not-yet-set types
@test g37293(T37293) == isdefined(T37293, :types)

This fails on master, but I have an update to this PR in the working that will fix it.

@vtjnash how can I provoke clearing the types field?

[martinholters]

While at it, what about the fields of TypeName? Currently we treat only name, module, and wrapper as constant. These look ok? Is names constant? That would be helpful for fieldcount now that (::DataType).types cannot be treated as constant anymore. What about hash?

[JeffBezanson]

I believe TypeName.names is either constant or undefined.

[martinholters]

I believe TypeName.names is either constant or undefined.

But it can become set later if undefined now? So if undefined infer getfield as SimpleVector (not Bottom), if set to n, infer as Const(n) would be ok?

Coming back to (::DataType).types once more. Even if it can get cleared, as long as it cannot be set to a different value than before, infering getfield on it as Const (if not undefined) would still be ok, as that does not automatically imply nothrow. Would that be valid? (I've tried figuring out how/when types gets populated, but I couldn't even figure out where it would be cleared in a way visible to Julia code.)

So to summarize:

field	status	isdefined	getfield	remarks
(::DataType).types	undef	Bool	SimpleVector
	set	Bool	Const	assuming can get cleared, but not be set differently
(::TypeName).names	undef	Bool	SimpleVector	assuming can be set
	set	Const(true)	Const	...but neither cleared nor set differently

Does that look ok? Then I'll implement it in this PR.

[vtjnash]

Replaced by better effects modeling and const fields of mutables