Re: Support loser tree for k-way merge

Hi,

Thank you for your reply.

> Can we drive the decision for what to do based on optimizer
> stats, i.e. n_distinct and row counts? Not sure what the calculation would
> be specifically, but something else to consider.
> 
> We can still provide the GUC to  override the optimizer decisions,
> but at least the optimizer, given up-to-date stats, may get it right most
> of the time.

That makes sense to me.

TODO
====
1) Consider optimizer statistics when deciding whether to use the heap or the loser tree.
2) Do we need a USEMEM() call to the array of losers?
3) Now the array length of losers is MAXORDER * 2, and in fact MAXORDER is enough, need some refactor of the code. (Is it worth doing?)
4) Add more code comments and doc.

Help are welcome!

--
Regards,
ChangAo Chen