Optimize stack-encoded instructions after emission

[alexcrichton]

Currently we encode functions to raw wasm and the we're done. Due to our representation with a tree-like AST, however, there are possible optimizations in the encoded instructions we can't necessarily represent in the AST. For example:

call $foo
drop
i32.const 1
return

can be more efficiently encoded as:

call $foo
i32.const 1
return

Or similarly whatever AST we settle on for:

i32.const 1
i32.const 1
call $foo ;; no inputs or outputs
i32.xor

can be optimally represented as above, but the AST is likely to look more like:

(i32.xor
  (i32.const 1)
  (block (result i32)
    (call $foo)
    (i32.const 1)))

[fitzgen]

can be optimally represented as above, but the AST is likely to look more like:

(i32.xor
  (i32.const 1)
  (block (result i32)
    (call $foo)
    (i32.const 1)))

I think this will take us pretty far, especially assuming wasm-opt follows us (for production builds at least)

[alexcrichton]

Oh actually that was an incorrect AST, it'd actually have to be:

call $a
call $a
call $b
i32.xor

transforms to:

(local i32)
(i32.xor 
  (call $a)
  (block (result i32)
    (local.set 0 (call $a))
    (call $b)
    (local.get 0)))

(can't reorder call $a and call $b!)