backend: lazy closure restore

[ggreif]

This gives us a les intrusive way of closure variable restoration. Instead of

# function prelude
  local.get $clos
  i32.load offs_a
  local.set $a
  ...
  local.get $a
  # use $a

one has

# empty prelude
  ...
  local.get $clos
  i32.load offs_a
  # use it instead of $a

Some functions only use closed-over variables to pass them on to other closures! So it makes total sense to avoid pulling them into locals. Such functions arise from (non-inlined) local-async functions.

Note: Right now this change is enabled globally, so I expect that the code size will increase due to closure variables used more than once. When we'll do this selectively, I expect a reduction in code size.

[ggreif]

Suggested change

	Used for immutable and mutable, non-captured data *)
	Used for immutable and mutable data *)

[ggreif]

Suggested change

	Used for immutable and mutable, non-captured data *)
	Used for immutable and mutable data *)

[crusso]

This might also reduce stack pressure - but won't it require repeated forwarding in the incremental gc? @luc-blaeser

[ggreif]

This might also reduce stack pressure - but won't it require repeated forwarding in the incremental gc? @luc-blaeser

Definitely, it's a tradeoff. I expect it to be a total win when we just wrap an inner function (we can detect this in freevars) and use all closure vars linearly. A good example is CPS-converted func() : async () { ... }, where the async block boils down to a function call.