[Mlir-commits] [mlir] [mlir][tensor] Loosen restrictions on folding dynamic reshapes (PR #137963)

[Artem Gindinson]

AGindinson wrote:

 > I'm trying to think of a generalized rule for when it becomes ambiguous. One possible way might be to check against a reversed target/source shape:
> ...
> The idea is that the current algorithm is "anti-greedy" and will try go minimize the size of the reassociations towards the beginning. So reversing the ordering will compare anti-greedy with greedy. The results will be the same if there is non-ambiguous reassociation.

Thanks, great catch first and foremost!

This may be a good approach, but I'm not sure how to conveniently let the ambiguity for collapsed dimensions of 1 pass through. Either tweaking the reverse iteration logic to push the 1's over to the next iteration (assuming this wouldn't lead to more peculiar edge cases, it might clutter the readability further) or writing the comparison by hand... I'll mull this over and submit a new version next week.

https://github.com/llvm/llvm-project/pull/137963