Pavel> The reset of memory context is useless, because the reset of Pavel> perValueCxt is there already on the end of tfuncFetchRows Pavel> function
It's overkill, yes, but it's also harmless because resetting a context that's not been touched since the last reset has very little overhead.
The overhead is +/- zero. But the code will work exactly same without this line. So it is little bit confusing to use there.
Pavel> I don't understand to using tuple memory context
We still need a per-result-tuple memory context otherwise we're leaking whatever memory got allocated in each GetValue call into the per-input-value context. (We can use our projection econtext's per-tuple memory for this because we haven't done any evaluation of output items for the current cycle yet at the point this code is reached.)
Again, look at functionscan for how this is supposed to work.
Pavel> When we are running under perValueCxt, then there changing Pavel> memory context is useless
It's not useless at all; it's needed to avoid excess memory usage when a single XMLTABLE() call returns many rows. -- Andrew (irc:RhodiumToad)