[Common] Update CombineSourceLayoutTransform pass to fold broadcast

[Abhishek-Varma]

This commit makes following updates :-

1. Folds named/generic broadcast op into MapLoadOp.
2. Inserts identity MapLoadOp only for "complex" relayout => chains length ≥ 2, has reshape (expand/collapse), and has non-extract_slice op (transpose, pad, copy, broadcast).
3. Avoids map_load for simple chains (e.g. single extract_slice) to prevent large memref.alloca and stack size issues.
4. Adds pass flag/option test-combine-non-complex-chains for lit-testing each op in isolation.

[Abhishek-Varma]

Ping for review.

[hanhanW]

Ping for review.

I'll take a look later today. I need to catch up context from the issue.

[hanhanW]

High-overview question:

This is done to prevent creating MapLoadOp for simple primitives
like tensor.extract_slice since it unneccesarily ends up creating an
empty tensor for MapLoad's destination -> which in turn leads to
creation of big memref.alloca ops later in the pipeline causing stack
size limit issue.

Do we really need the folding for extract_slice op? The implementation is looking for chained relayout op, should we explicitly match the case to reflect what you're saying here?

I approved the previous PR that allows extract_slice op because it is a mirror of the destionatino folding, but I wonder if we need it for source folding.

[Max191]

A couple of quick comments

[Max191]

Nice! It looks good now (% some nits) aside from the testing. I think we just need a pass option to test non-complex chains and then testing will be much easier.

[Max191]

Also, please update the description to be consistent with the new changes. It is still referring to the old implementation.

[Max191]

Nice work, it LGTM now. Just make sure to address the comments about the tests before landing.

[Max191]

Also, let's wait for @hanhanW to give one more review before landing.

[hanhanW]

Just one question for op collection and one nit for implementation.

[hanhanW]

Thanks!