GlobalStructInference: Optimize gets of immutable global struct data

[kripken]

A struct.get of a global, where the global creates a struct, can be
optimized by other passes, but not in all cases: Precompute can only
handle Literals (as it uses the interpreter), and GUFA works on a type-
based manner. The general case is important too, which looks like this:

;; Global struct, immutable.
(global $vtable (ref $vtable)
  (struct.new $vtable
    (global.get $imported)
  )
)

;; Code reads the global, then a field. This can be a read of $imported.
(struct.get $vtable 0
  (global.get $vtable)
)

This PR adds it to GlobalStructInference. That pass has a bunch of
useful infrastructure for it (parsing of struct.news, unnesting of values,
etc.), making it simple there.

Fixes #8002

[tlively]

It would be neat (and not too hard, I think) to expand this to support arbitrarily deep paths through struct.new, array.new, and global.get expressions in global initializers.

[kripken]

Yes, that is what the TODO about unnesting is about. (I think we need to unnest, like the other code here? Unless there is a simpler way I am missing.)

[tlively]

It would be fairly straightforward to have a utility that is given the root of some sequence of accessors (kind of like a gep in LLVM, except not all rolled into one instruction) and then follows the access path as long as it is known to remain constant, returning (a pointer to) the result.

[kripken]

Would that be useful anywhere else, do you think?

[tlively]

Hmm, I'd have to think about that. I saw that meanwhile you uploaded a solution that unnests the intermediate values into globals. The only benefit a path-walking utility would have over that would be fewer new globals to be optimized back out later. A downside would be potentially quadratic behavior, though.