Slow SQL Server CLR aggregates

I have always discovered the lack of a built-in aggregate function that simply concatenates values ​​using a specific custom separator, frustrating.

The closest thing I know is the XML hack:

select s.string as [text()] 
from strings s 
for xml path('')

But, in my opinion, a very strange way to do this, and you still have to handle the case with a trailing file or leading garbage. So, I thought I would write a CLR aggregate:

select dbo.Fold(s.string, ', ') 
from strings s

So nice. Except that it doesn’t work very well when the number of rows reaches 5000. I really do not expect to run it against this data, but, to my surprise, aggregate performance deteriorates quite quickly as the data set grows. At first I did not understand why, I realized that the UDF problem in SQL Server is a problem.

I use UDF with a class StringBuilder, because of this I need to implement IBinarySerialize. But for reasons I don’t understand, SQL Server calls a Write / Read pair between each iteration Accumulate. Obviously, this creates a performance problem when the string gets large. Is there anyway to avoid this expensive serialization?

FYI, an XML hack is orders of magnitude faster.