[MLton] Vector.fromList[] performance

[Matthew Fluet]

> I have a largeish program that uses vectors a lot: it's a geometry
> code that is written in arbitrary-dimensional fashion.  Some of the
> code is in special dimension, so I often find myself converting tuples
> (x,y) into vectors using Vector.fromList[x,y].
>
> Unfortunately, the performance of this utterly sucks: I don't know how
> to read the generated bytecode to see exactly what's happening, but I
> can't fathom how it could be so bad without paying for building the
> list, then converting it to a vector.

Well, building the list and then converting it to a vector is exactly what 
it does, since that is what 'Vector.fromList [x,y]' does.

> How hard would it be to special-case Vector.fromList on a literal
> list?

It's not trivial.  'Vector.fromList' is implemented in SML, so the 
compiler doesn't treat it specially.  Trying to identify exactly this code 
would be pretty tough.

On the other hand, a loop-unrolling optimization would likely help both 
this idom and other code.

> SML/NJ allows this using the #[] construct, which helps SML/NJ
> on this test program knock the pants off mlton.  But really the
> compiler should just handle it automatically so that I can write
> standard-compliant code.

Right.  You could probably use MLton.Vector.unfoldi or the even more 
powerful MLton.Vector.create (in SVN) to build the vector directly, but it 
suffers from the same portability problem.