It’s not necessarily about the load, it’s about the algorithmic complexity. Going from lists (lines in a file, characters in a line) to trees introduces a potentially exponential increase in complexity due to the number of ways the same list of elements can be organized into a tree.
Also, you’re underestimating the amount of processing. It’s not about pure CPU computations but RAM access or even I/O. Even existing non-semantic diff implementations are unexpectedly inadequate in terms of performance. You clearly haven’t tried diffing multi-GB log files.
Log files themselves don’t, but I’m just comparing it with simpler files with simpler structure with simpler algorithms with better complexity.