diff --git a/compiler/tests/check_tests/Dialect/BConcrete/bufferization-nonzero-offets.mlir b/compiler/tests/check_tests/Dialect/BConcrete/bufferization-nonzero-offets.mlir
new file mode 100644
index 000000000..49d482762
--- /dev/null
+++ b/compiler/tests/check_tests/Dialect/BConcrete/bufferization-nonzero-offets.mlir
@@ -0,0 +1,16 @@
+// RUN: concretecompiler --action=dump-llvm-dialect %s
+
+// Extracted from the source referenced in Issue 663. This should
+// trigger the folding of memrefs of itermediate results to memrefs
+// with non-zero offsets. Prior to the use of symbolic offsets in the
+// memref used in the memref.cast operation produced by the BConcrete
+// bufferizer, bufferization of the function below would fail.
+func.func @main(%arg0: tensor<32x!FHE.eint<8>>, %arg1: tensor<256xi64>) -> !FHE.eint<8>
+{
+  %c0 = arith.constant 0 : index
+  %719 = "FHELinalg.apply_lookup_table"(%arg0, %arg1) : (tensor<32x!FHE.eint<8>>, tensor<256xi64>) -> tensor<32x!FHE.eint<8>>
+  %720 = tensor.extract_slice %719[16] [16] [1] : tensor<32x!FHE.eint<8>> to tensor<16x!FHE.eint<8>>
+  %722 = tensor.extract %720[%c0] : tensor<16x!FHE.eint<8>>
+  %755 = "FHE.add_eint"(%722, %722) : (!FHE.eint<8>, !FHE.eint<8>) -> !FHE.eint<8>
+  return %755 : !FHE.eint<8>
+}