[EH] Test StackIR's local2stack on translator output

[aheejin]

This adds STACKIR-OPT filecheck lines to translate-to-new-eh.wast to see if StackIR's local2stack optimization successfully removes some of unnecessary local.set/local.gets.

While supporting the whole Binayren optimization pipeline for the new EH instructions is not the goal for the very near-term future, StackIR's local2stack optimization can help with a very common pattern generated by this translator, which is:

(try $l
  (do ... )
  (catch_all
    (call $destructor)
    (rethrow $l)
  )
)

is translated to

(block $outer
  (local.set $exn     ;; can be optimized away
    (block $catch_all (result exnref)
      (try_table (catch_all_ref $catch_all)
        ...
      )
      (br $outer)
    )
  )
  (call $destructor)
  (throw_ref
    (local.get $exn)  ;; can be optimized away
  )
)

Here we don't really need local.set $exn and local.get $exn, and these can be optimized away using StackIR's local2stack. After optimizing them away in Stack IR, the code can be like

block $outer
  block $catch_all (result exnref)
    try_table (catch_all_ref $catch_all)
      ...
    end
    br $outer
  end
  call $destructor
  throw_ref
end

This optimization alone reduces the code size increased caused by translating significantly. For Adobe Photoshop, the code size increase goes down from 4.2% to 2.8%, and for Binaryen, it goes down from 3.8% to 2.0%.

[kripken]

Nice, looks like this is an example of those local.set/get operations getting removed.