[Mlir-commits] [mlir] [MLIR][XeGPU] Add offsets to load/store/prefetch (PR #150576)

[Jianhui Li]

================
@@ -189,6 +189,7 @@ def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
   let genVerifyDecl = 1;
 }
 
+def XeGPU_TensorDescOrMemRef : AnyTypeOf<[XeGPU_TensorDesc,Non0RankedMemRefOf<[XeGPU_ScalarType]>, UI64]>;
----------------
Jianhui-Li wrote:

This change is the purpose of this PR.  It allows the xegpu.load to accept memref+offset, which is the target OP semantics that we are migrating to.  The current xegpu.load accepts tdesc operand, which encapsulates "memref+offset".  This PR "fold" "memref+offset" directly to xegpu.load replacing "tdesc". The semantics are combination of "creating scattered_tdesc + xegpu.load with scattered_tdesc". 
The block_tdesc for load_nd was kept since it contains meaningful info (tile shape/stride) after offset being taken out. 

https://github.com/llvm/llvm-project/pull/150576