MLIR optimization failure with DenseHashTable

Hey,

It’s not an MLIR issue rather than a TFG one (the former is used to build the latter, but the latter is a different group of people etc, mostly mentioning this as this also determines who triage and fix).

If memory serves (and it’s been a few years) on failure of this pass, it just becomes a NOP. So it doesn’t stop or affect any other optimization passes.

The error message appears to say that you have an invalid input which just happens to fail here first. Now this could also be as the type system is more precise here and possibly the runtime may have led this slide. In particular, the while loop is supposed to operate as a fixed point wrt type, while that input argument is changing from a single to “struct” tensor. Possibly just manually casting or setting the type should suffice.

– Jacques