The `ExtractFromShapeOfExtentTensor` canonicalization pattern was
unconditionally rewriting:
tensor.extract(shape.shape_of(%arg), %idx) -> tensor.dim(%arg, %idx)
even when `%arg` is a memref. This produced an invalid `tensor.dim`
(whose source operand must be a tensor), which then caused an assertion
failure in `DimOp::getSource()` when subsequent canonicalization
patterns tried to match the op:
Assertion `isa<To>(Val) && "cast<Ty>() argument of incompatible type\!"'
failed. [To = TypedValue<TensorType>, From = Value]
Fix: add an `IsTensorType` constraint to
`ExtractFromShapeOfExtentTensor` in `ShapeCanonicalization.td` so the
pattern only fires when `%arg` is a tensor type. The memref case is
intentionally left unfolded (the correct lowering to `memref.dim` would
require adding a MemRef dependency to the Shape dialect, which is not
desirable).
Tests cover both the positive case (tensor arg folds to tensor.dim) and
the negative case (memref arg is left unmodified).
Fixes#185248
Assisted-by: Claude Code
0572ad60f3