Dear macdonc,
This seems like a really good idea that might be able to fill a gap between brook+ and il. Having finally gotten a particular IL program going I was thinking that there must be some simplifications possible, allowing e.g. names with array notation to be used instead of r100,r101,r110,r111 for a matrix etc.; I should try your compiler!
Another tool in this vein I'd love to see from AMD (or anybody!) would be a "simple" IL to gpuisa compiler or "translator" that doesn't attempt any optimization or anything, for example mapping vector il registers directly to vector gpu registers. I've noticed some il code ends up as isa code that uses many more actual registers than I think it should and has suboptimal burst-writing (I think this may be down to trying to use the "t" unit as much as possible), and would like to know whether a "simpler" version would actually perform better.
Good luck for your project,
Steven.