Max191

> is there a reason to do this with a separate pass instead of just doing this at the same time as `LinalgFoldUnitExtentDims`? Is it a matter of this pass...

> My sense is that this being done way too early. Agree with Quinn, maybe fold this into the `FoldUnitDims` pass. I decided to just change the nesting of FoldUnitExtentDims...

There is still a lot of code to add to this to actually implement the propagation piece. I think my plan is to make a new PR for each op...

> Love what it's doing, but terrified by the PackOp and static data layout coupling. We really are going to need to stop using PackOp at the global opt level...

Hmm, I thought there were already some existing patterns that would handle cases like this. Maybe I wrote some a while ago and never upstreamed it :p I can take...

> ``` > %subview_4 = memref.subview %7[%15, 0, 0, 0] [4, 512, 8, 8] [1, 1, 1, 1] : memref to memref > %collapse_shape_5 = memref.collapse_shape %subview_4 [[0], [1], [2,...

> Anyway, let's focus on the common case since that is the main blocker right now! Is it clear how to move forward, @Max191? Do you have code for this...

> I recently added the inference to canonicalization patterns ([llvm/llvm-project#80848](https://github.com/llvm/llvm-project/pull/80848)). What is the bug? > > ``` > %pack = tensor.pack %arg0 padding_value(%cst : f32) > outer_dims_perm = [1, 2,...

This is fixed by https://github.com/llvm/llvm-project/pull/82539

@MaheshRavishankar what is the reason for this check? https://github.com/openxla/iree/blob/b42e627e3761292676bbc6a405b7ef1829a8d53b/compiler/src/iree/compiler/Dialect/Flow/Transforms/FormDispatchRegions.cpp#L703-L705 This is the real thing blocking the fusion here, and it seems like it is very restrictive to me. Can we...