I made as few changes as I could to get the tests to pass, but this
was a bit involved due to some churn in the blame code since someone
last gave fastannotate any TLC.
There's still follow-up work here to rip out support for old versions
of hg and to integrate the protocol with modern standards.
Some performance numbers (all on my 2016 MacBook Pro with a 2.6Ghz i7):
time: real 1.050 secs (user 0.990+0.000 sys 0.060+0.000)
time: real 5.900 secs (user 5.720+0.000 sys 0.110+0.000)
time: real 0.120 secs (user 0.100+0.000 sys 0.020+0.000)
time: real 3.330 secs (user 3.220+0.000 sys 0.070+0.000)
time: real 30.610 secs (user 30.190+0.000 sys 0.230+0.000)
time: real 0.180 secs (user 0.160+0.000 sys 0.020+0.000)
time: real 7.640 secs (user 7.210+0.000 sys 0.380+0.000)
time: real 98.650 secs (user 97.000+0.000 sys 0.950+0.000)
time: real 1.580 secs (user 1.340+0.000 sys 0.240+0.000)
time: real 17.110 secs (user 16.490+0.000 sys 0.500+0.000)
time: real 399.750 secs (user 394.520+0.000 sys 2.610+0.000)
time: real 1.780 secs (user 1.530+0.000 sys 0.240+0.000)
So building the cache is expensive (but might be faster with xdiff
enabled), but the blame results are *way* faster.