Identified in the BF ticket, there's an 11% regression as linkbench2 load phase overall throughput. I spent some time trying to bisect this, but unfortunately there isn't one single change that caused a performance issue - there's a number of smaller drops spread over several months.

Instead of trying to investigate/revert these one at a time, we should profile __wt_row_modify to see where it's slow and why.

For further details see the linked BF.