[Mlir-commits] [mlir] [MLIR] Make `OneShotModuleBufferize` use `OpInterface` (PR #107295)

[Tzung-Han Juang]

================
@@ -19,6 +19,7 @@
 // tensors annotated with attributes from the `bufferization` dialect. These
 // attributes hint the bufferization pass to assume buffers can be directly
 // used for these tensors without reshaping.
+module @payload attributes { transform.target_tag = "payload" } {
 func.func @conv(
     %input: !tinput {bufferization.writable = false,
                      bufferization.access = "read",
----------------
tzunghanjuang wrote:

Hi @matthias-springer , we are trying to use `debug-payload-root-tag=payload` to limit the `moduleOP bufferization` to the payload module. (Or it will trigger an error for no `ReturnLike`.)

It seems that the added payload module in this test blocks some passes after `transform-interpreter`.  The checks fail since the IRs like `llvm.intr.fma` are not produced. Maybe there are some integration issues? Is there any better way to exclude `transform.named_sequence` from bufferziation? Thank you very much.
 
https://github.com/llvm/llvm-project/blob/828783177f71d95522763fba92ef1e42dc6101c7/mlir/test/Examples/transform/ChH/full.mlir#L379-L388

https://github.com/llvm/llvm-project/pull/107295