Re: Reducing the chunk header sizes on all memory context types

On Tue, 13 Sept 2022 at 20:27, David Rowley <dgrowleyml(at)gmail(dot)com> wrote:
> I see that one of the drawbacks from not using MemoryContextContains()
> is that window functions such as lead(), lag(), first_value(),
> last_value() and nth_value() may now do the datumCopy() when it's not
> needed. For example, with a window function call such as
> lead(byref_col ), the expression evaluation code in
> WinGetFuncArgInPartition() will just return the address in the
> tuplestore tuple for "byref_col". The datumCopy() is needed for that.
> However, if the function call was lead(byref_col || 'something') then
> we'd have ended up with a new allocation in CurrentMemoryContext to
> concatenate the two values. We'll now do a datumCopy() where we
> previously wouldn't have. I don't really see any way around that
> without doing some highly invasive surgery to the expression
> evaluation code.

It feels like a terrible idea, but I wondered if we could look at the
WindowFunc->args and make a decision if we should do the datumCopy()
based on the type of the argument. Vars would need to be copied as
they will point into the tuple's memory, but an OpExpr likely would
not need to be copied.

Aside from that, I don't have any ideas on how to get rid of the
possible additional datumCopy() from non-Var arguments to these window
functions. Should we just suffer it? It's quite likely that most
arguments to these functions are plain Vars anyway.

David