[TIR] Fix crash on transform_layout

[quic-sanirudh]

When ScheduleTransformLayout is called, the block corresponding to the BlockRV named block is potentially changed. The symbol_table_ however is not updated for changes to BlockRV and hence when the same blockRV block is passed to ScheduleSetAxisSeparator, the sref corresponding to that is invalid and that results in crash as unknown type_index_

This PR is a temporary fix and a permanent fix would probably need to identify that the BlockRV is not pointing to a valid sref and throw a proper error or update the symbol_table_ appropriately to avoid this error completely.

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @Hzfengsy, @junrushao, @shingjan _{See #10317 for details}

_{Generated by tvm-bot}

[quic-sanirudh]

This PR provides a temporary fix to the error seen in #13957

[quic-sanirudh]

cc @vinx13 @Lunderberg

[quic-sanirudh]

@tvm-bot rerun

[quic-sanirudh]

This PR is a temporary fix and a permanent fix would probably need to identify that the BlockRV is not pointing to a valid sref and throw a proper error

Just realized that if we enable debug_mask="all" when creating the Schedule, we do get a proper error message related to the sref, so maybe this fix could be considered a permanent fix?

[quic-sanirudh]

@tvm-bot rerun

[vinx13]

I think as a permanent fix, the schedule primitive should always guarantee correct remapping of block reference. This is achieved by setting correct block_sref_reuse https://github.com/apache/tvm/blob/main/src/tir/schedule/primitive/layout_transformation.cc#L1165

[quic-sanirudh]

I think as a permanent fix, the schedule primitive should always guarantee correct remapping of block reference. This is achieved by setting correct block_sref_reuse https://github.com/apache/tvm/blob/main/src/tir/schedule/primitive/layout_transformation.cc#L1165

Thanks @vinx13 for the suggestion. I'll try to update this PR with the permanent fix.

[quic-sanirudh]

Closing in favor of #14023