[Mlir-commits] [mlir] [mlir][transform] Add support for transform.param pad multiples in `PadOp` (PR #90755)

Oleksandr Alex Zinenko llvmlistbot at llvm.org
Thu May 2 07:27:56 PDT 2024


================
@@ -1021,8 +1023,7 @@ def PadOp : Op<Transform_Dialect, "structured.pad",
                       TransformHandleTypeInterface:$pad,
                       TransformHandleTypeInterface:$copy);
 
-  let assemblyFormat =
-    "$target attr-dict `:` functional-type(operands, results)";
+  let hasCustomAssemblyFormat = 1;
----------------
ftynse wrote:

> I actually wouldn't be opposed to using the dynamic list directive for this in favor of simpler code. even if I do personally think the resulting IR is uglier, but obviously will put personal preferences aside. But I think there should be consistency. So I would also suggest changing tile_using_for and vectorize if possible. Would like to hear what others think.

I strongly support consistency across transform ops. Going out of our way to add syntactic consistency with only tangentially-related interface is a waste of cycles. Take a look at https://github.com/llvm/llvm-project/blob/7dfb6f571c8090803a32b5611e519b77ae74de30/mlir/include/mlir/Dialect/Linalg/TransformOps/LinalgTransformOps.td#L2004-L2008, if _that_ can be reused here, let's do that. Otherwise, let's stick with custom syntax and eventually upgrade all syntax everywhere.

https://github.com/llvm/llvm-project/pull/90755


More information about the Mlir-commits mailing list