commit 32057462c50865767d4bb4c34a69dc2d3fc34c4e
author Stella Laurenzo <stellaraccident@gmail.com> Mon Jul 25 13:28:03 2022 -0700
committer GitHub <noreply@github.com> Mon Jul 25 13:28:03 2022 -0700
tree e39f7a26e7b338f30c08069aa7b555ab3e9e5851
parent 9efad13a8ca1fede30f3ad292275defd2a87e655

Use the new memref->util->vm path. (#9883)

* Use the new memref->util->vm path.

This was a little more involved than anticipated:

* A number of type conversion issues and A->B->C issues in the patterns were fixed (only show up in type conversion settings).
* The memref->util buffer path introduces initializers for rodata constants, whereas the old one short-circuited these directly to vm rodata. This highlighted missing VM support for optimizing such cases, and it was an issue because VMVX does not support initializers. We may want to fix that last part, but how I have it at least gets the conversion to the same IR as before once it runs through the pipeline.
* The above alternate path highlighted a bad VM folding pattern which was not accounting for initializers and just folding to null (even though it had a value). Deleted (the new pass applies this optimization when it is correct to do so).
* Still more that can be done, but this at least gets us to deleting the original and switching.
* Linearize VMVX buffers to 1D, disallowing 0D. Aside from being one less thing to reason about, this interops with the invariants of the memref flattener and now lowers properly with the memref->util->vm flow.
* Enable last remaining test that was blocked on 0D.