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

[ioana ghiban]

================
@@ -85,14 +84,14 @@ func.func private @concat_strided(%src : memref<1x1xf32>,
 
   // CHECK-NOT:  memref.copy
   // CHECK:      %[[C0:.*]] = arith.constant 0 : index
-  // CHECK:      %[[C0_0:.*]] = arith.constant 0 : index
   // CHECK:      %[[VAL:.*]] = memref.load %[[SRC]][%[[C0]], %[[C0]]] : memref<1x1xf32>
-  // CHECK:      memref.store %[[VAL]], %[[DST]][%[[C0]], %[[C0_0]]] : memref<1x108xf32>
+  // CHECK:      memref.store %[[VAL]], %[[DST]][%[[C0]], %[[C0]]] : memref<1x108xf32>
   memref.copy %src, %reinterpret_cast
     : memref<1x1xf32> to memref<1x1xf32, strided<[107, 2]>>
   return
 }
 
+// Dynamic strides are irrelevant because all view indices are zero.
----------------
ioghiban wrote:

The semantics of `memref.copy` constrain the source and destination operands to have the same shape. Therefore, if the destination (the view/strided memref) has all dimensions one, then the copied memref must also, given the IR is valid. This means we're only copying one element. Does this clear up the comment or did I misunderstand your question?

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