ARC optimizations

[dcodeIO]

In order to perform ARC optimizations with a PostAssemblyScript Binaryen pass, this PR adds better hints for the optimizer to determine (and then follow) a __retain. Reason is that previously the compiler would emit code like

(drop
 (call $__retain
  (local.get $0)
 )
)

when just retaining something that already exists, like a function argument that magically exists but isn't set explicitly, but without a local.set we cannot follow the influences of the retain through the local graph.

With the change the compiler emits

(local.set $0
 (call $__retain
  (local.get $0)
 )
)

so we can easily see that the local $0 is retained from here onward but wasn't before, and any local.get influenced by the local.set is safe to assume to inherit the retain.

This was especially problematic with flattening, leading to code like

(local.set $1
 (local.get $0)
)
(local.set $2
 (call $__retain
  (local.get $1)
 )
)
(drop
 (local.get $2)
)
;; continues using $0

which completely misses the side effect if we start at local.set $2 that is never used again.

Interestingly, this moves quite some code around in our test fixtures as well, even though its functionally the same.

[dcodeIO]

Alright, last commit uses a custom Binaryen build with the PostAssemblyScript stuff I came up with so far. Also copies Binaryen's pass order instead of inheriting -OX so we can adapt it. Haven't looked through all changes in detail, but pushing as I assume @MaxGraey is also interested in taking a look :)

[dcodeIO]

Interesting cases for further inspection

Size reduced

• std/typedarray.optimized.wat (-2077)
• std/array.optimized.wat (-1189)
• std/string.optimized.wat (-324)
• std/string-encoding.optimized.wat (-95)
• std/arraybuffer.optimized.wat (-35)

Size increased

• std/set.optimized.wat (+16)
• std/symbol.optimized.wat (+22)