[Mlir-commits] [mlir] [MLIR][XeGPU] Add XeGPU scattered ops (PR #86594)

[Mehdi Amini]

================
@@ -32,6 +32,17 @@ void XeGPUDialect::initialize() {
 //===----------------------------------------------------------------------===//
 // XeGPU_TensorDescAttr
 //===----------------------------------------------------------------------===//
+TensorDescAttr TensorDescAttr::get(mlir::MLIRContext *context,
+                                   xegpu::MemoryScope memory_scope,
+                                   int array_length, bool boundary_check,
+                                   bool scattered) {
+  auto scopeAttr = MemoryScopeAttr::get(context, memory_scope);
+  auto lengthAttr =
+      IntegerAttr::get(IntegerType::get(context, 64), array_length);
+  auto boundaryAttr = BoolAttr::get(context, boundary_check);
+  auto scatteredAttr = BoolAttr::get(context, scattered);
+  return Base::get(context, scopeAttr, lengthAttr, boundaryAttr, scatteredAttr);
+}
----------------
joker-eph wrote:

This kind of builder encourages bad practices IMO: by hiding the attribute creation it incurs compile-time cost.
If instead these are moved to the user, they are more likely to reuse existing attributes across multiple builder call.

(It's something that I see in multiple places though, so feel free to keep it if you want: this is just food for thoughts)

https://github.com/llvm/llvm-project/pull/86594