[Mlir-commits] [mlir] [memref] Support non-scalar copies in `reinterpret_cast` elision (PR #203873)

[ioana ghiban]

================
@@ -149,14 +146,165 @@ func.func private @concat_rank3(%src : memref<1x1x1xf32>,
 
   // CHECK-NOT:  memref.copy
   // CHECK:      %[[C0:.*]] = arith.constant 0 : index
-  // CHECK:      %[[C0_0:.*]] = arith.constant 0 : index
   // CHECK:      %[[VAL:.*]] = memref.load %[[SRC]][%[[C0]], %[[C0]], %[[C0]]] : memref<1x1x1xf32>
-  // CHECK:      memref.store %[[VAL]], %[[DST]][%[[C0]], %[[C0]], %[[C0_0]]] : memref<1x1x108xf32>
+  // CHECK:      memref.store %[[VAL]], %[[DST]][%[[C0]], %[[C0]], %[[C0]]] : memref<1x1x108xf32>
   memref.copy %src, %reinterpret_cast
     : memref<1x1x1xf32> to memref<1x1x1xf32>
   return
 }
 
+// CHECK-LABEL: func.func private @concat_0d(
+// CHECK-SAME:   %[[SRC:.*]]: memref<1x1x1xf32>
+// CHECK-SAME:   %[[DST:.*]]: memref<1x33x42xf32>
+func.func private @concat_0d(
----------------
ioghiban wrote:

Will rename `concat` to `copy_to_strided`.
There are no dimensions copied, therefore the `0d`. We're effectively copying a scalar:
```mlir
memref<1x1x1xf32> to memref<1x1x1xf32>
```
Perhaps using `scalar` in the naming is more suggestive.

https://github.com/llvm/llvm-project/pull/203873