speed up hg log --patch
Changing dodiff to read the manifest/changelog for node1 before calling
repo.update allows us to take advantage of the revlog revision cache.
Before this patch and my previous "speed up hg log --debug" patch, when
using hg log -p to display three revisions (A, B and C), dodiff and
repo.changes would end up reading the manifests in this order:
B A B A C B C B
With both patches, this order becomes:
A A B B B B C C
(This considers only dodiff and repo.changes. I'm not sure how other
parts of hg log enter the picture.)
The speed up will depend on the revisions being displayed. (All
"before" times already have my previous "speed up hg log --debug" patch
applied.)
hg repo (tip =
414e81ae971f). hg log -p
before after
real 0m50.981s 0m45.279s
user 0m47.930s 0m42.560s
sys 0m2.526s 0m2.523s
output size: 6917897 bytes
kernel repo (tip =
9d4e135960ed). hg log -p -l64
before after
real 2m14.995s 1m45.025s
user 2m9.509s 1m33.900s
sys 0m3.663s 0m2.942s
output size:
31497621 bytes
same kernel repo. hg log -p -l64 -r
c84c2069592f:0
before after
real 1m48.045s 1m0.076s
user 1m44.094s 0m58.492s
sys 0m2.603s 0m1.103s
output size: 197983 bytes
c84c2069592f was the tip of a 10 day old kernel repo that I had lying
around and was where I first tested this patch. For some weird
coincidence it's also a place where the patch makes a huge difference.
speed up hg log --debug
hg repo:
before:
real 0m6.998s
user 0m6.952s
sys 0m0.046s
after:
real 0m3.884s
user 0m3.839s
sys 0m0.042s
Replaced fixed window size for walkchangerevs with an increasing one. Window
sizes starts at 8 (for good interactiveness) and doubles with each window
until it is 512, which seems to be the maximum efficient value.
hgmerge: do not use file with markers in tools merging themselves
This time I forgot about a crucial part of the logic change.
The patch makes kdiff3 and tkdiff useful again.