(This is not yet enabled; it will be turned on in a followup patch.)
The path encoding performed by fncache is complex and (perhaps
surprisingly) slow enough to negatively affect the overall performance
of Mercurial.
For a short path (< 120 bytes), the Python code can be reduced to a fairly
tractable state machine that either determines that nothing needs to be
done in a single pass, or performs the encoding in a second pass.
For longer paths, we avoid the more complicated hashed encoding scheme
for now, and fall back to Python.
Raw performance: I measured in a repo containing 150,000 files in its tip
manifest, with a median path name length of 57 bytes, and 95th percentile
of 96 bytes.
In this repo, the Python code takes 3.1 seconds to encode all path
names, while the hybrid C-and-Python code (called from Python) takes
0.21 seconds, for a speedup of about 14.
Across several other large repositories, I've measured the speedup from
the C code at between 26x and 40x.
For path names above 120 bytes where we must fall back to Python for
hashed encoding, the speedup is about 1.7x. Thus absolute performance
will depend strongly on the characteristics of a particular repository.
Here, we exclude hidden changesets from a rebase operation. If we
don't, a rewritten version of the hidden changesets will be created
by rebase. Those rewritten versions won't be hidden and will likely
conflict with other rewriting or revive pruned changeset. Moreover,
rewriting hidden revisions will surprise the user.
This change would not be necessary if changelog filtering were
already in core. But it's fairly cheap and helps to increase the
test-suite for such filtering.
Once changelog level filtering is added, hidden changes will be
automatically excluded or included according to the global --hidden
flags. Plain ignoring them is good enough for now.
In collapse mode, that content of state is not suitable to compute obsolescence
markers. We explicitly pass the resulting revision instead and use it as the
successors for all elements of the rebased set.
When obsolescence feature is enabled we now create markers from the rebased
set to the resulting set instead of stripping. The "state" mapping built by
rebase holds all necessary data.
Changesets "deleted" by the rebase are marked "succeeded" by the changeset they
would be rebased one. That the best guess of "successors" we have. Getting a
successors as meaningful as possible is important for automatic resolution of
obsolescence troubles. In other word, emptied changeset will looks collapsed
with their former parents. (see "empty changeset" section of the test if you are
still confused)
At the end of the rebase, rebased changesets are currently stripped. This
behavior will be eventually dropped in favor of obsolescence marker creation.
The main rebase function is already big and branchy enough. This changeset move
the clean-up logic in a dedicated function before we make it more complex.
For a netbeans clone on Windows 7 x64:
Before:
$ hg perffncacheencode
! wall 3.516000 comb 3.525623 user 3.525623 sys 0.000000 (best of 3)
After:
$ hg perffncacheencode
! wall 3.443000 comb 3.447622 user 3.447622 sys 0.000000 (best of 3)
Not yet used (will be enabled in a later patch).
This patch is a stripped down version of patches originally created by
Bryan O'Sullivan <bryano@fb.com>
For a netbeans clone on Windows 7 x64:
Before:
$ hg perffncacheload
! wall 0.124000 comb 0.124801 user 0.124801 sys 0.000000 (best of 76)
After:
$ hg perffncacheload
! wall 0.096000 comb 0.093601 user 0.078001 sys 0.015600 (best of 97)
hgweb has an incorrect padding calculation, causing the text to move further
away from the graph the more branches there are (issue3626). This patch fixes
all existing templates (gitweb, monoblue, paper and spartan).
Tests updated by Patrick Mezard <patrick@mezard.eu>
For a netbeans clone on Windows 7 x64:
Before:
$ hg perffncachewrite
! wall 0.210000 comb 0.218401 user 0.202801 sys 0.015600 (best of 47)
After:
$ hg perffncachewrite
! wall 0.104000 comb 0.109201 user 0.078000 sys 0.031200 (best of 95)
I don't think we will ever have anything in the store that resides inside a
directory that ends in .i or .d under store/ that we wouldn't want to have
direncoded. The files not under data/ surely don't need direncoding, but it
doesn't harm to let these few run through it. It hurts more to check whether the
thousands of other files start with 'data/'. They do anyway.
See also 67e6074ba430 (fixed with 0c522fe42894), which moved the direncoding
from filelog into store
by refactoring
for i, n in enumerate(res):
if n:
<main code block>
to
for i, n in enumerate(res):
if not n:
continue
<main code block>
(no functional change)
46cdcb89086f fixed display of non-ASCII names in file-selecting prompt, but
display in chunk selection remained broken. The reason is that using '%r' in
string formatting results in calling `repr` on file names, thus mangling
non-ASCII ones.
Merely creating and using a generator has a measurable impact,
particularly since the common case for stream_out is generators that
yield just once. Avoiding generators improves stream_out performance
by about 7%.
Auditing at this stage is both pointless (paths are already trusted by
the local repo) and expensive. Skipping the audits improves stream_out
performance by about 15%.
