Catching NaN at the MLIR Pass Boundary
Summary
The article discusses how an MLIR pass can generate a not-a-number (NaN) during constant folding and why NaN should be rejected at the earliest IR boundary. It covers IEEE 754 NaN behavior, propagation concerns, and the use of operation verifiers and diagnostic tests to prevent NaN from escaping into runtime. It also outlines a workflow to locate the offending pass, reduce the input, and harden a dialect with intrinsic invariants, using a fictional quant.rescale_fuse example and references to MLIR tooling.