evaluate benefits of TFA tracking constant values and lengths of arrays

[mraleph]

There are some indications that tracking this information might allow to tree-shake more code from AOT compiled applications.

One of the motivating examples would be to fold away defaultTargetPlatform getter (which returns a constant value based on the target compilation platform) and propagate it's value into the uses - which should enable more constant folding and subsequently tree-shaking of code that is only used on some specific platforms.

There are some challenges associated with it - currently TFA is control-flow insensitive. To successfully propagate defaultTargetPlatform we would need certain control flow sensitivity. These changes might result in a significant increase in compile times for AOT code - but potential gains are worth the try.

[mraleph]

/cc @alexmarkov

[mraleph]

NBody benchmark and vector math use the following pattern:

class Body {
  final Float64List _data = Float64List(7);
  double get x => _data[0];
  set x (double v) { _data[0] = v; }
}

I have done a small dirty experiment where I have propagated specifically the length of Body._data to be 7 in the LoadFieldInstr::Canonicalize rule. This improved NBody performance by around 20% - which means it will probably be similarly beneficial for the vector_math package as well (it might also improve the code size).