Typical practice seems to be having a large collection of ad hoc local transformations that look for specific combinations of operations. I've only really seen more general static analysis aimed at not materializing intermediate arrays in academic research.

Expression tree IR is maybe the wrong setting to do this sort of fusion for APL. What you want for a+b+c is a 3-argument scalar function applied to a, b, and c, but you need something other than actual APL expressions for that.