[Mlir-commits] [mlir] [mlir][xegpu] Distribute three more convert_layout cases to lanes (PR #222693)

[Sang Ik Lee]

================

----------------
silee2 wrote:

It lowers to the 8-bit integer sub-group shuffle. The LLVM type converter maps `f8E8M0FNU` to `i8`, so `GPUToLLVMSPV`'s mangling picks the `char` overload:
>
> ```mlir
> %8 = builtin.unrealized_conversion_cast %7 : f8E8M0FNU to i8
> %9 = llvm.call spir_funccc @_Z21sub_group_shuffle_xorcj(%8, %c8_i32)
>        {convergent, no_unwind, will_return} : (i8, i32) -> i8
> %10 = builtin.unrealized_conversion_cast %9 : i8 to f8E8M0FNU

https://github.com/llvm/llvm-project/pull/222693