torch.aten.batch_norm is relying on incorrect annotations

[silvasean]

As discussed in pytorch/pytorch#73050 (comment), there are a few ops that don't correctly annotate that they mutate their operands. It seems like those are aten::batch_norm and aten::layer_norm.

When I revamped our ODS generator code, I tried correcting those exceptions w.r.t. the HasValueSemantics and ReadOnly traits, but it seems we were relying on the old, incorrect annotation (which I think was okay, since it only matters in the training case, which we haven't implemented yet)

torch-mlir/python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/registry.py

Line 258 in a5fe0cf

# TODO: Handle some exceptions of incorrectly annotated ops.

To work on this, you just have to uncomment the code linked above and regenerate the ODS for torch.aten.batch_norm and see what breaks in the tests. I dug into it a little bit, and it seems like we will need some special handling in ReduceOpVariants to convert torch.aten.batch_norm to value semantics when training == false.

[byronyi]

Should be fixed with input mutation work in pytorch/pytorch#88817. See pytorch/pytorch#88375 (comment) for an example.