I think Rob has more context about it.

I'm not convinced that this is the fix... The result type and operand type should match based on my understanding about DSP interface. (I could be wrong though)

In D145160#4176453, @devajith-huawei wrote:

It will be great if someone can commit this for me (if it is good to go).

Hello. The original commit and the revert of this seem to have caused quite a lot of performance changes. Do you know by any chance what the difference between the code was? And by chance do you know if it was the SplittingGeps pass, LICM or CSE that gave the increase?

Hey @dmgreen, this is Hanhan working on IREE team (which is a MLIR based compiler project). We've noticed that this commit causes regressions in our tracking models. If there was a way to do control this flag through the C++ API, that would help us configure the options during setup of target machine easier. Could you help expose the option or give us some advices on how we can control the flag through some C++ API? Thank you!

https://reviews.llvm.org/D144604 is a better fix.

LGTM

In D144067#4134121, @mravishankar wrote:

This seems strange to me. There is nothing that prevents a subsequent transformation to collapse the cast into the slice. Indeed it is probably better to do so.

This is a similar fix to the issue we found in iree_linalg_ext.unpack version: https://github.com/iree-org/iree/issues/11607

rebase

rebase

add a test

In D143978#4125190, @chelini wrote:

Do we have a test case to exercise this code path?

fix clone method

update comments

rebase

address comments

In D143043#4105155, @nicolasvasilache wrote:

Thanks for extending!
Could you point me to your usage of this abstraction (offline is fine)?

fix tests

rebase and fix tests

rebase

nice work, thanks!

Thanks for fixing this!

LG, thanks!

address comments