[Mlir-commits] [mlir] [mlir][xegpu] Lower lane_data repack convert_layout to bitcast_shuffle (PR #210837)

[Artem Kroviakov]

================
@@ -1491,6 +1491,72 @@ def XeGPU_TruncfOp
   let hasVerifier = 1;
 }
 
+def XeGPU_BitcastShuffleOp
+    : XeGPU_Op<"bitcast_shuffle", [Pure]> {
+  let summary = "Bit-preserving type conversion and re-distribution of a "
+                "subgroup's fragments across its lanes.";
+
+  let description = [{
+    `xegpu.bitcast_shuffle` is a lane-level operation. The `source` operand is
+    the fragment (a scalar or 1D vector) held by one lane, and the `result` is
+    the fragment that lane holds afterwards. It reinterprets the fragments as a
+    different element type, possibly with a different number of elements per
+    lane, and at the same time re-distributes the data across the lanes of the
+    subgroup.
+
+    Both `source` and `result` must be a scalar or a 1D vector of numerical
+    type, and they must have a different type. When they have the same number of
+    elements the operation degenerates into a plain bitcast. Otherwise the total
----------------
akroviakov wrote:

> When they have the same number of elements the operation degenerates into a plain bitcast

Why can the op not similarly degenerate into a shuffle for the same type?

https://github.com/llvm/llvm-project/pull/210837