>Please find attached a POC patch to do just that.
>The switch to the single-datum tuplesort is done when there is only one
>attribute, it is byval (to avoid having to deal with copy of the references
>everywhere) and we are not in bound mode (to also avoid having to move things
>around).
Hi, nice results!
I have a few suggestions and questions to your patch:
1. Why do you moved the declaration of variable *plannode?
I think this is unnecessary, extend the scope.
2. Why do you declare a new variable TupleDesc out_tuple_desc at ExecInitSort?
I think this is unnecessary too, maybe I didn't notice something.
3. I inverted the order of check at this line, I think "!node-bounded" is more cheap that TupleDescAttr(tupDesc, 0) ->attbyval
4. Once that you changed the order of execution, this test is unlikely that happens, so add unlikely helps the branch.
5. I think that you add a invariant inside the loop "if(node->is_single_val)"?
Would not be better two fors?
For you convenience, I attached a v2 version (with styles changes), please take a look and can you repeat yours tests?
regards,
Ranier Vilela