[Mlir-commits] [mlir] [MLIR][XeGPU] Distribute load_gather/store_scatter op from Wg To Sg (PR #154420)

[Adam Siemieniuk]

================
@@ -763,6 +763,88 @@ struct WgToSgArithConstantOp : public OpConversionPattern<arith::ConstantOp> {
   }
 };
 
+// This pattern transforms the LoadGatherOp with explicit offsets to load
+// subgroup data, similar to WgToSgLoadNdOpWithOffset.
+struct WgToSgLoadGatherOpWithOffset
+    : public OpConversionPattern<xegpu::LoadGatherOp> {
+  using OpConversionPattern<xegpu::LoadGatherOp>::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(xegpu::LoadGatherOp op, OneToNOpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+
+    if (!op.getOffsets())
+      return failure();
+
+    Location loc = op.getLoc();
+    VectorType resultType = op.getResult().getType();
----------------
adam-smnk wrote:

I don't think it's guaranteed to be a vector anymore after #154653 

https://github.com/llvm/llvm-project/pull/154420