As Joseph S. Myers bentioned in comment 1 of PR46633, if
TREE_STRING_LENGTH is considered to be measured in BITS_PER_UNIT
rather than TYPE_PRECISION (char_type_node), fix_string_type actually
has one BITS_PER_UNIT uses too few rather than three to many.

The only other place I came up with in c-common also requires even more
BITS_PER_UNIT:

@@ -8541,8 +8544,14 @@ fold_offsetof_1 (tree expr, tree stop_re
 tree
 fold_offsetof (tree expr, tree stop_ref)
 {
+  tree size = fold_offsetof_1 (expr, stop_ref);
+
+  /* Convert in case a char is more than one unit.  */
+  size
+    = size_binop (CEIL_DIV_EXPR, size,
+                 size_int (TYPE_PRECISION (char_type_node) / BITS_PER_UNIT));
   /* Convert back from the internal sizetype to size_t.  */
-  return convert (size_type_node, fold_offsetof_1 (expr, stop_ref));
+  return convert (size_type_node, size);
 }

 /* Warn for A ?: C expressions (with B omitted) where A is a boolean

(In reply to Jorn Wolfgang Rennecke from comment #1)
> As Joseph S. Myers bentioned in comment 1 of PR46633, 

(this is the last bug open blocking that one, btw)

Confirmed.

in fold_offsetof, we still have:
```
    case COMPONENT_REF:
...
      off = size_binop_loc (input_location, PLUS_EXPR, DECL_FIELD_OFFSET (t),
                            size_int (tree_to_uhwi (DECL_FIELD_BIT_OFFSET (t))
                                      / BITS_PER_UNIT));

```

and:
```
    case ARRAY_REF:
...
      off = size_binop (MULT_EXPR, TYPE_SIZE_UNIT (TREE_TYPE (expr)), t);
```