After a query is translated to ABT, we go through the process of "shredding" it into groups within the memo. This process involves walking the ABT and making a copy of each sub-tree only to override the child node. Since the ABT's copy constructor is a full deep copy by default, there is a lot of potentially wasted work.

As an example, for an input ABT such as

Root
   |
Filter
   |
Filter
   |
Evaluation
   |
Scan

the code linked above will get called for each node as we walk the ABT, starting with

Evaluation
   |
Scan

all the way up to the full Root node, each time making a copy only to replace the child with a delegator node.