When generating Subversion diffs, RBTools does a lot of diff processing

to:

Ensure that added/deleted empty files are present
Ensure renamed files have the right information
Convert relative paths to absolute paths
Filter for any excluded files

Each step of this would iterate over the previous diff and then generate

a new one. If the diff had, say, 10,000 lines, we'd parse and then

re-build those 10,000 lines in almost every one of those stages, before

finally returning the result. This was slow.
This change converts each processing stage to a generator, allowing one

pass through the diff. Each stage will iterate through and yield lines

for the next stage, the result of which is iteratively joined into a

final byte string.
The stages themselves have been optimized a bit. We used to perform

repeated lookups on SVNClient for the same attributes (compiled

regexes and constants) on each loop. Those are now pulled out of the

loops.
A potential bug was also found and fixed that could have led to an

infinite loop during diff generation when processing empty files. If a

very specific (unlikely) set of conditions were met (deleted empty file,

revision values of 0 for the diff operation, and no Revision found for

the file in svn info), we'd repeate the loop, but with no advancement.

That'd cause the same condition to be hit over and over. We now advance

the iterator.

Unit tests pass.