Making best use of Trex needs a different mindset vs H98 datatypes with named fields. (Or indeed vs named-fields data structures in many languages) There's some examples around in the Trex/Hugs material, but I've seen nothing showing this:
> type WPoint r = (x :: Float, y :: Float | r) -- r extended "with Point" > type WRadius r = (radius :: Float | r) > type WColour r = (colour :: String | r) > type ColouredCircle = Rec( WColour (WRadius (WPoint EmptyRow)))
The idea is rather than nest 'Point' as a datatype inside 'Circle' 'plug and play' components (such as Point, Radius) to give a structural type. It's interesting and impressive that the Wxxx synonyms take an argument of kind row, return a result of kind row. `EmptyRow` in the last line is a Hugs-defined Row constant that 'plugs' the end of the composed components, just as `emptyRec` does for Trex Records.
I'd love to write that last type as
> type MkRec r = Rec( r EmptyRow)
> type ColouredCircle = MkRec (WColour . WRadius . WPoint)
But of course type synonyms -- even with parameters -- are not functions; can't be composed; can't be passed to higher-order type synonyms.
Components like WPoint can be arguments to type-parametric functions for record subtyping: such as a generic routine to 'move' any shape by modifying its x, y coordinates. 'Move' in scare quotes because that's returning a copy of the shape's data structure with modified coordinates, not poking values into an object/reference.