@inducer: this test does exercise the problem, because without the fix the iname "i" gets included the host code and so generates a runtime error.

However, I'm not sure this kernel should even compile. How many times should the k loop execute?

I agree that it's a tad weird. By the (current... :) semantics, bounds for k and j are computed by projecting out i.

>>> s = isl.BasicSet("{[i,j,k]: 0 < k < i and 0 < j < 10 and 0 < i < 10}")
>>> s.project_out(isl.dim_type.set, 0, 1)
BasicSet("{ [j, k] : 0 < j <= 9 and 0 < k <= 8 }")

We could yell if we have to project out anything but "boxy" dimensions, but that would be a pretty drastic change to the semantics.

That's what I thought was going on. I was not sure if the fact that i ought to be nested inside k should have anything to do with it - but I can live with this semantics.