[Mlir-commits] [mlir] [MLIR][XeGPU] make offsets optional for create_nd_tdesc (PR #148335)

[Chao Chen]

================
@@ -163,9 +167,30 @@ def XeGPU_CreateNdDescOp: XeGPU_Op<"create_nd_tdesc", [Pure, ViewLikeOpInterface
     }
 
     ArrayRef<int64_t> getStaticOffsets(){
-      return getConstOffsets();
+      auto attr = getConstOffsetsAttr();
+
+      if (attr) 
+        return attr;
+
+      auto memrefType = llvm::dyn_cast<MemRefType>(getSourceType());
+      int rank = 0;
+      if (memrefType) {
+        //use source memref's rank, as source memref rank may be higher
+        rank = memrefType.getRank();
+      } else {
+        //nd_tdesc created from ui64, use nd_tdesc's rank
+        rank = getTensorDescShape().size();
----------------
chencha3 wrote:

It is not safe to infer the rank from the TensorDesc, since TensorDesc could have fewer rank than offset. You can simply use `int rank = getStaticSizes().size()` instead; 

https://github.com/llvm/llvm-project/pull/148335