[Mlir-commits] [mlir] [mlir][ArmSVE] Add convert_to/from_svbool ops (PR #68586)

[Andrzej Warzyński]

================
@@ -66,6 +68,77 @@ using ScalableMaskedDivFOpLowering =
     OneToOneConvertToLLVMPattern<ScalableMaskedDivFOp,
                                  ScalableMaskedDivFIntrOp>;
 
+namespace {
+
+/// Unrolls a conversion to/from equivalent vector types, to allow using a
+/// conversion intrinsic that only supports 1-D vector types.
----------------
banach-space wrote:

I second Cullen here. It's not obvious to me what the right "pass" would be and where would it live, but we should add a TODO in the comments and in the commit message. Something along the lines:

> Extract the lowering of `convert_to_svbool` into a series of `vector.extract/insert` operations into a dedicated lowering layer/pass pre "LegalizeForLLVMExport". The latter should be as simple as 1-1 mapping.

https://github.com/llvm/llvm-project/pull/68586