Fix row group alignment in ORC writer (#15789)

Closes #15775

ORC writer encodes null mask bits in multiples of eight to avoid issues with other readers reading partial encoded bytes. When this does not align with row groups, the null mask encode boundaries are moved to align to multiples of eight. There was a bug in the alignment code that caused a pointless shift by 8 bits and, then, issues in encode. This PR fixes the unnecessary shift.

Authors:
  - Vukasin Milovanovic (https://github.com/vuule)

Approvers:
  - Nghia Truong (https://github.com/ttnghia)
  - Muhammad Haseeb (https://github.com/mhaseeb123)
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #15789

cpp/src/io/orc/writer_impl.cu

@@ -782,8 +782,16 @@ std::vector<std::vector<rowgroup_rows>> calculate_aligned_rowgroup_bounds(
         } else {
           // pushdown mask present; null mask bits w/ set pushdown mask bits will be encoded
           // Use the number of set bits in pushdown mask as size
-          auto bits_to_borrow =
-            8 - (d_pd_set_counts[rg_idx][parent_col_idx] - previously_borrowed) % 8;
+          auto bits_to_borrow = [&]() {
+            auto const parent_valid_count = d_pd_set_counts[rg_idx][parent_col_idx];
+            if (parent_valid_count < previously_borrowed) {
+              // Borrow to make an empty rowgroup
+              return previously_borrowed - parent_valid_count;
+            }
+            auto const misalignment = (parent_valid_count - previously_borrowed) % 8;
+            return (8 - misalignment) % 8;
+          }();
+
           if (bits_to_borrow == 0) {
             // Didn't borrow any bits for this rowgroup
             previously_borrowed = 0;

python/cudf/cudf/tests/test_orc.py

@@ -1954,3 +1954,16 @@ def test_writer_lz4():
 
     got = pd.read_orc(buffer)
     assert_eq(gdf, got)
+
+
+def test_row_group_alignment(datadir):
+    path = datadir / "TestOrcFile.MapManyNulls.parquet"
+
+    expected = cudf.read_parquet(path)
+
+    buffer = BytesIO()
+    expected.to_orc(buffer)
+
+    got = cudf.read_orc(buffer)
+
+    assert_eq(expected, got)