Extended Description

When feed IR to a tool like mlir-opt and going through transformations, the verifier failure are confusing:

input_to-tf-shape-inference.mlir:115:15: error: 'tf.Shape' op requires dynamic shape result for unranked operand
%27 = "tf.Shape"(%26) {T = i32, _output_shapes = ["tfshape$dim { size: 1 }"], device = "", out_type = i32} : (tensor<?xi32>) -> tensor<1xi32>

The IR displayed is actually extracted from the source file, and the actual operation that is failing the verifier may be entirely different.

It seems that for developers (or when IR is the input), the verifier would be better printing the current instruction.