Wrong results/performance regression when using BIND and graph pattern groups

I've been using SPARQL queries with BINDs and/or VALUESs that precede UNION blocks. These used to work efficiently (at least in Jena 2 times) with the bound value being bound also in the subpattern, but no longer do. In addition, at certain times, they produce nonsensical results. For example, the below returns rdf:type statements in the dataset without any regard to ?cl:

SELECT * {
  {
    BIND("nonexistant" AS ?cl)
    {
      BIND(?cl AS ?cl2)
      ?c a ?cl2 .
    } UNION {
      BIND(?cl AS ?cl2)
      ?c a ?cl2 .
    }
  }
}
LIMIT 10

while if you change just one of the subpatterns in the union to directly refer to ?cl, it returns an empty result set:

SELECT * {
  {
    BIND("nonexistant" AS ?cl)
    {
      BIND(?cl AS ?cl2)
      ?c a ?cl .
    } UNION {
      BIND(?cl AS ?cl2)
      ?c a ?cl2 .
    }
  }
}
LIMIT 10