Adding some type info to haXe

There are two good things about open source software; it’s open, and there’s source. So I had a hack with haXe and produced haxe-typed.exe, as a replacement for haxe.exe. This executable adds strong typing to strings, ints, floats and bools when they are local variables. It does not add types to function parameters and return types.

The results were promising:

AS3 haXe haXe typed
CarDemo 7.6 ms 38.6 ms 38.9 ms
SimpleLoop – for 115 ms 1218 ms 74 ms
SimpleLoop – while 115 ms 1020 ms 76 ms
PerfTest 1 – string manipulation, join 238 ms 383 ms 720 ms
PerfTest 2 – string sort 620 ms 1284 ms 195 ms
PerfTest 3 – string addition 125 ms 138 ms 130 ms
PerfTest 4 – bit string manipulation 125 ms 134 ms 125 ms
PerfTest 5 – floating point ops 2 ms 24 ms 2 ms
PerfTest 6 – substr 2 ms 7 ms 7 ms
PerfTest 7 – string indexOf 58 ms 75 ms 75 ms
PerfTest 8 – Math.round/random 6 ms 38 ms 24 ms
PerfTest 9 – integer addition 31 ms 343 ms 27 ms
PerfTest 10 – string function calls 34 ms 103 ms 104 ms
PerfTest 11 – MD5 68 ms 163 ms 171 ms
PerfTest 12 – integer function calls 1 ms 14 ms 9 ms

Here we see that where the processing is local, and in a tight loop, huge performance gains can be achieved. When functions calls are involved, there are no gains – sometimes it is even slower (presumably because of the conversion of specific type to generic type). Unfortunately, CarDemo makes extensive use of properties and object orientation nad therefore does not benefit at all from these optimisations.

HaXe even outperforms in a few tests – the integer loop and the sort function. My initial guess is haXe’s use of local variables for constants, rather than integer constants is what helps this loop, and I have no idea about the sort function. Perhaps there is a bug somewhere.

Use the above haxe-typed.exe at your own risk. There it was more hacking than science used to produce it. Also, there is a bunch of text-out that probably means nothing – this is basically my attempt at understanding what was going on. If you are after the source-code changes just let me know. The changes are probably not very nice to look at, or done very well – but at least I got it compiling.

The exe is not very widely tested (only on these 3 programmes !). When there is a type mismatch, the flash player throws an exception and the execution stops – often leaving you looking at a white screen. Use the “player/debug/SAFlashPlayer.exe”, provided with the flex SDK, to see these exceptions. More often than not, they are something like “Int and * can not be reconciled”. This means there is a bug in bytecode created by the modified haXe. The process of fixing the exe was to basically track these down one by one.

The code also relies on expressions having the correct type. This is generally the case with haXe, except for the case of Floats with constant ints, eg: “var f:Float = 2”. I had to change the source to “var f:Float = 2.0” to get these to compile.

I had no reference to the AS3 assembly syntax other than what I could see in the haXe code and what I could deduce from the “abcdump.exe” of both haXe and flex outputs. The only principle for typing local varibles I can see is to initialise them with a specifically typed value. And this must be done at the “top” of the function – I’m not sure what logic is used to determine where the “top” finishes, but the following pseudo code works:

  var i:Int = 0;
  while (something)

Whereas the following does not define i as an integer:

  While (something)
     var i:Int = 0;

So the task of adding type information amounts to initialising all the local variables (including those that are limited in scope by blocks) at the top of the function.

I did not add strong types to object variables. This should be quite possible, since haXe knows the obejct types. External information suggests that this would be a very good thing to do, and I may get around to it one day. Maybe.

I also did not add types to function arguements or return values. I’m sure doing so would close the gap between haXe and AS3 native code, however the code removes this type information reasonably early and it would take a bit of effort to get this working.

I found editing the ocaml code a bit of a slog. Firstly, “ml” style languages were new to me and it took me a while to grok it. It took me a while to workout how to even decompose the syntax into a series of statements (oh, it’s returning a function!). Secondly, the code is pretty much bereft of comments so it was hard to know where to start – I couldn’t really look a a bit of code in isolation whithout a good feeling for the whole code base. And Lastly, the ocaml penchant for terse variable names (eg, “let field_access ctx get f t e p =”).

This may be as far as I take these optimisations for now. The main goal was to see if the performance can be improved – and I think the answer is yes. I don’t think my changes are of high enough quality to re-submit them into the code base (in fact, I would almost certainly do a lot differently if I were to start again). It has been a great exercise and I’ve learnt a lot – including a new programming language.

This entry was posted in Blog. Bookmark the permalink.

Leave a Reply

Your email address will not be published. Required fields are marked *