[Mlir-commits] [mlir] [MLIR][XeGPU] Add unroll patterns for XeGPU (1/N) (PR #137010)

[Adam Siemieniuk]

================
@@ -141,46 +142,24 @@ void CreateNdDescOp::build(OpBuilder &builder, OperationState &state,
 }
 
 void CreateNdDescOp::build(OpBuilder &builder, OperationState &state,
-                           Type tdesc, TypedValue<MemRefType> source,
+                           Type tdesc, Value source,
                            llvm::ArrayRef<OpFoldResult> offsets,
                            llvm::ArrayRef<OpFoldResult> shape,
                            llvm::ArrayRef<OpFoldResult> strides) {
   assert(shape.size() && offsets.size() && strides.size() &&
          shape.size() == strides.size() && shape.size() == offsets.size());
 
-  llvm::SmallVector<int64_t> staticOffsets;
-  llvm::SmallVector<int64_t> staticShape;
-  llvm::SmallVector<int64_t> staticStrides;
+  auto intTy = dyn_cast<IntegerType>(source.getType());
+  auto memrefTy = dyn_cast<MemRefType>(source.getType());
+  assert(intTy || memrefTy && "Source has to be either int or memref.");
----------------
adam-smnk wrote:

nit: maybe let's cache type itself, use two `isa<>` here and only `dyn_cast` further below

https://github.com/llvm/llvm-project/pull/137010