join38.q (without map join) produces incorrect result when testing with multiple reducers

I have this query from a modified version of join38.q, which does NOT use map join:

FROM src a JOIN tmp b ON (a.key = b.col11)
SELECT a.value, b.col5, count(1) as count
where b.col11 = 111
group by a.value, b.col5;

If I set mapred.reduce.tasks to 1, the result is correct. But, if I set it to be a larger number (3 for instance), then result will be

val_111	105	1

which is wrong.

I think the issue is that, for this case, ConstantPropagationProcFactory will overwrite the partition cols for the reduce sink desc, with an empty list. Then, later on in ReduceSinkOperator#computeHashCode, since partitionEval is length 0, it will use an random number as hashcode, for each separate row. As result, rows with same key will be distributed to different reducers, and hence leads to incorrect result.