Hey,
I'd consider it more like 1) model support is top priority and that dictates what ops are prioritized, 2) ML framework produced input dialects (like TOSA or StableHLO) coverage is then secondary (both of these are still changing though so any coverage numbers is slice in time), 3) everything else. LinAlg is the primary "consolidation" point codegen wise in IREE, so things that can be lowered to LinAlg are supported more readily.
That means while upstream MLIR dialects are supported, it's mostly those that have LinAlg lowerings or are produced along path 1 or 2 above. Now tensor::splat, I couldn't see any codegen for it even upstream. If it had a lowering to linalg then it would just work (the only Tensor dialect op that seems to have lowering to LinAlg is Pad it would seem). Else patches are welcome given priority is mostly focused on ML inputs (e.g., welcome but not priority)
-- Jacques