sapling/hgext3rd/fastannotate/context.py

802 lines
30 KiB
Python
Raw Normal View History

# Copyright 2016-present Facebook. All Rights Reserved.
#
# context: context needed to annotate a file
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.
from __future__ import absolute_import
import collections
import contextlib
import hashlib
import os
from mercurial import (
error,
lock as lockmod,
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
mdiff,
node,
scmutil,
util,
)
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
from mercurial.i18n import _
from . import (
error as faerror,
revmap as revmapmod,
)
import linelog as linelogmod
# given path, get filelog, cached
@util.lrucachefunc
def _getflog(repo, path):
return repo.file(path)
# extracted from mercurial.context.basefilectx.annotate
def _parents(f, follow=True):
# Cut _descendantrev here to mitigate the penalty of lazy linkrev
# adjustment. Otherwise, p._adjustlinkrev() would walk changelog
# from the topmost introrev (= srcrev) down to p.linkrev() if it
# isn't an ancestor of the srcrev.
f._changeid
pl = f.parents()
# Don't return renamed parents if we aren't following.
if not follow:
pl = [p for p in pl if p.path() == f.path()]
# renamed filectx won't have a filelog yet, so set it
# from the cache to save time
for p in pl:
if not '_filelog' in p.__dict__:
p._filelog = _getflog(f._repo, p.path())
return pl
# extracted from mercurial.context.basefilectx.annotate. slightly modified
# so it takes a fctx instead of a pair of text and fctx.
def _decorate(fctx):
text = fctx.data()
linecount = text.count('\n')
if text and not text.endswith('\n'):
linecount += 1
return ([(fctx, i) for i in xrange(linecount)], text)
# extracted from mercurial.context.basefilectx.annotate. slightly modified
# so it takes an extra "blocks" parameter calculated elsewhere, instead of
# calculating diff here.
def _pair(parent, child, blocks):
for (a1, a2, b1, b2), t in blocks:
# Changed blocks ('!') or blocks made only of blank lines ('~')
# belong to the child.
if t == '=':
child[0][b1:b2] = parent[0][a1:a2]
return child
# like scmutil.revsingle, but with lru cache, so their states (like manifests)
# could be reused
_revsingle = util.lrucachefunc(scmutil.revsingle)
def resolvefctx(repo, rev, path, resolverev=False, adjustctx=None):
"""(repo, str, str) -> fctx
get the filectx object from repo, rev, path, in an efficient way.
if resolverev is True, "rev" is a revision specified by the revset
language, otherwise "rev" is a nodeid, or a revision number that can
be consumed by repo.__getitem__.
if adjustctx is not None, the returned fctx will point to a changeset
that introduces the change (last modified the file). if adjustctx
is 'linkrev', trust the linkrev and do not adjust it. this is noticeably
faster for big repos but is incorrect for some cases.
"""
fastannotate: support replacing fctx.annotate directly Summary: Previously, I have performance concern about the signature of `fctx.annotate`: it returns fctxs, not ideal for performance since my initial goal is to get rid of reading revlogs in the best case. And letting the low-level `fctx.annotate` have some side effects writing to the disk is not that pretty. Because of that, fastannotate had re-invent part of the formatter (also optimized for performance somehow), and still cannot support all the features the original annotate supports, namely, templates. Now it makes sense to just replace `fctx.annotate` with a sub-optimal implementation, just for the flexibility of the template support of the original annotate command. We can use a "fake" or "lazy" fctx object to minimal the performance impact when people only need to print changeset nodes and line numbers - in which case we don't read revlog. Actually, the hgweb support already did a similar thing - converting fastannotate output to annotate output. So we can reuse some code. The next planned steps are: - Make the original "annotate" command aware of fastannotate protocol, do the pre-download stuff (downloading cache files at `fctx.annotate` is possible but inefficient because of not being batched). - "fastannotate" command remains a separated command optimized for perf and provides extra features like "--deleted". Because of the plan, the "commands" option does not make much sense - "fastannotate" command will not replace "annotate" directly thus dropped from the config interface. A new "modes" config option is added to control how fastannotate works. Test Plan: Modified existing tests Reviewers: #sourcecontrol, simonfar Reviewed By: simonfar Subscribers: mjpieters Differential Revision: https://phabricator.intern.facebook.com/D4238998 Signature: t1:4238998:1480447679:c48e0e565663c086293265e104d9cf414d913aa7
2016-11-29 14:51:08 +03:00
if resolverev and not isinstance(rev, int) and rev is not None:
ctx = _revsingle(repo, rev)
else:
ctx = repo[rev]
fctx = ctx[path]
if adjustctx is not None:
if adjustctx == 'linkrev':
introrev = fctx.linkrev()
else:
introrev = fctx.introrev()
if introrev != ctx.rev():
fctx._changeid = introrev
fctx._changectx = repo[introrev]
return fctx
# like mercurial.store.encodedir, but use linelog suffixes: .m, .l, .lock
def encodedir(path):
return (path
.replace('.hg/', '.hg.hg/')
.replace('.l/', '.l.hg/')
.replace('.m/', '.m.hg/')
.replace('.lock/', '.lock.hg/'))
def hashdiffopts(diffopts):
diffoptstr = str(sorted(
(k, getattr(diffopts, k))
for k in mdiff.diffopts.defaults.iterkeys()
))
return hashlib.sha1(diffoptstr).hexdigest()[:6]
_defaultdiffopthash = hashdiffopts(mdiff.defaultopts)
class annotateopts(object):
"""like mercurial.mdiff.diffopts, but is for annotate
followrename: follow renames, like "hg annotate -f"
followmerge: follow p2 of a merge changeset, otherwise p2 is ignored
"""
defaults = {
'diffopts': None,
'followrename': True,
'followmerge': True,
}
def __init__(self, **opts):
for k, v in self.defaults.iteritems():
setattr(self, k, opts.get(k, v))
@util.propertycache
def shortstr(self):
"""represent opts in a short string, suitable for a directory name"""
result = ''
if not self.followrename:
result += 'r0'
if not self.followmerge:
result += 'm0'
if self.diffopts is not None:
assert isinstance(self.diffopts, mdiff.diffopts)
diffopthash = hashdiffopts(self.diffopts)
if diffopthash != _defaultdiffopthash:
result += 'i' + diffopthash
return result or 'default'
defaultopts = annotateopts()
class _annotatecontext(object):
"""do not use this class directly as it does not use lock to protect
writes. use "with annotatecontext(...)" instead.
"""
def __init__(self, repo, path, linelogpath, revmappath, opts):
self.repo = repo
self.ui = repo.ui
self.path = path
self.opts = opts
self.linelogpath = linelogpath
self.revmappath = revmappath
self._linelog = None
self._revmap = None
self._node2path = {} # {str: str}
@property
def linelog(self):
if self._linelog is None:
self._linelog = linelogmod.linelog(self.linelogpath)
return self._linelog
@property
def revmap(self):
if self._revmap is None:
self._revmap = revmapmod.revmap(self.revmappath)
return self._revmap
def close(self):
if self._revmap is not None:
self._revmap.flush()
self._revmap = None
if self._linelog is not None:
self._linelog.close()
self._linelog = None
__del__ = close
def rebuild(self):
"""delete linelog and revmap, useful for rebuilding"""
self.close()
self._node2path.clear()
_unlinkpaths([self.revmappath, self.linelogpath])
@property
def lastnode(self):
"""return last node in revmap, or None if revmap is empty"""
if self._revmap is None:
# fast path, read revmap without loading its full content
return revmapmod.getlastnode(self.revmappath)
else:
return self._revmap.rev2hsh(self._revmap.maxrev)
def isuptodate(self, master, strict=True):
"""return True if the revmap / linelog is up-to-date, or the file
does not exist in the master revision. False otherwise.
it tries to be fast and could return false negatives, because of the
use of linkrev instead of introrev.
useful for both server and client to decide whether to update
fastannotate cache or not.
if strict is True, even if fctx exists in the revmap, but is not the
last node, isuptodate will return False. it's good for performance - no
expensive check was done.
if strict is False, if fctx exists in the revmap, this function may
return True. this is useful for the client to skip downloading the
cache if the client's master is behind the server's.
"""
lastnode = self.lastnode
try:
f = self._resolvefctx(master, resolverev=True)
# choose linkrev instead of introrev as the check is meant to be
# *fast*.
linknode = self.repo.changelog.node(f.linkrev())
if not strict and lastnode and linknode != lastnode:
# check if f.node() is in the revmap. note: this loads the
# revmap and can be slow.
return self.revmap.hsh2rev(linknode) is not None
# avoid resolving old manifest, or slow adjustlinkrev to be fast,
# false negatives are acceptable in this case.
return linknode == lastnode
except LookupError:
# master does not have the file, or the revmap is ahead
return True
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
def annotate(self, rev, master=None, showpath=False, showlines=False):
"""incrementally update the cache so it includes revisions in the main
branch till 'master'. and run annotate on 'rev', which may or may not be
included in the main branch.
if master is None, do not update linelog.
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
the first value returned is the annotate result, it is [(node, linenum)]
by default. [(node, linenum, path)] if showpath is True.
if showlines is True, a second value will be returned, it is a list of
corresponding line contents.
"""
# the fast path test requires commit hash, convert rev number to hash,
# so it may hit the fast path. note: in the "fctx" mode, the "annotate"
# command could give us a revision number even if the user passes a
# commit hash.
if isinstance(rev, int):
rev = node.hex(self.repo.changelog.node(rev))
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
# fast path: if rev is in the main branch already
directly, revfctx = self.canannotatedirectly(rev)
if directly:
if self.ui.debugflag:
self.ui.debug('fastannotate: %s: using fast path '
'(resolved fctx: %s)\n'
% (self.path, util.safehasattr(revfctx, 'node')))
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
return self.annotatedirectly(revfctx, showpath, showlines)
# resolve master
masterfctx = None
if master:
try:
masterfctx = self._resolvefctx(master, resolverev=True,
adjustctx=True)
except LookupError: # master does not have the file
pass
else:
if masterfctx in self.revmap: # no need to update linelog
masterfctx = None
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
# ... - @ <- rev (can be an arbitrary changeset,
# / not necessarily a descendant
# master -> o of master)
# |
# a merge -> o 'o': new changesets in the main branch
# |\ '#': revisions in the main branch that
# o * exist in linelog / revmap
# | . '*': changesets in side branches, or
# last master -> # . descendants of master
# | .
# # * joint: '#', and is a parent of a '*'
# |/
# a joint -> # ^^^^ --- side branches
# |
# ^ --- main branch (in linelog)
# these DFSes are similar to the traditional annotate algorithm.
# we cannot really reuse the code for perf reason.
# 1st DFS calculates merges, joint points, and needed.
# "needed" is a simple reference counting dict to free items in
# "hist", reducing its memory usage otherwise could be huge.
initvisit = [revfctx]
if masterfctx:
if masterfctx.rev() is None:
raise error.Abort(_('cannot update linelog to wdir()'),
hint=_('set fastannotate.mainbranch'))
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
initvisit.append(masterfctx)
visit = initvisit[:]
pcache = {}
needed = {revfctx: 1}
hist = {} # {fctx: ([(llrev or fctx, linenum)], text)}
while visit:
f = visit.pop()
if f in pcache or f in hist:
continue
if f in self.revmap: # in the old main branch, it's a joint
llrev = self.revmap.hsh2rev(f.node())
self.linelog.annotate(llrev)
result = self.linelog.annotateresult
hist[f] = (result, f.data())
continue
pl = self._parentfunc(f)
pcache[f] = pl
for p in pl:
needed[p] = needed.get(p, 0) + 1
if p not in pcache:
visit.append(p)
# 2nd (simple) DFS calculates new changesets in the main branch
# ('o' nodes in # the above graph), so we know when to update linelog.
newmainbranch = set()
f = masterfctx
while f and f not in self.revmap:
newmainbranch.add(f)
pl = pcache[f]
if pl:
f = pl[0]
else:
f = None
break
# f, if present, is the position where the last build stopped at, and
# should be the "master" last time. check to see if we can continue
# building the linelog incrementally. (we cannot if diverged)
if masterfctx is not None:
self._checklastmasterhead(f)
if self.ui.debugflag:
if newmainbranch:
self.ui.debug('fastannotate: %s: %d new changesets in the main'
' branch\n' % (self.path, len(newmainbranch)))
elif not hist: # no joints, no updates
self.ui.debug('fastannotate: %s: linelog cannot help in '
'annotating this revision\n' % self.path)
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
# prepare annotateresult so we can update linelog incrementally
self.linelog.annotate(self.linelog.maxrev)
# 3rd DFS does the actual annotate
visit = initvisit[:]
progress = 0
while visit:
f = visit[-1]
if f in hist:
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
visit.pop()
continue
ready = True
pl = pcache[f]
for p in pl:
if p not in hist:
ready = False
visit.append(p)
if not ready:
continue
visit.pop()
blocks = None # mdiff blocks, used for appending linelog
ismainbranch = (f in newmainbranch)
# curr is the same as the traditional annotate algorithm,
# if we only care about linear history (do not follow merge),
# then curr is not actually used.
assert f not in hist
curr = _decorate(f)
for i, p in enumerate(pl):
bs = list(self._diffblocks(hist[p][1], curr[1]))
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
if i == 0 and ismainbranch:
blocks = bs
curr = _pair(hist[p], curr, bs)
if needed[p] == 1:
del hist[p]
del needed[p]
else:
needed[p] -= 1
hist[f] = curr
del pcache[f]
if ismainbranch: # need to write to linelog
if not self.ui.quiet:
progress += 1
self.ui.progress(_('building cache'), progress,
total=len(newmainbranch))
bannotated = None
if len(pl) == 2 and self.opts.followmerge: # merge
bannotated = curr[0]
if blocks is None: # no parents, add an empty one
blocks = list(self._diffblocks('', curr[1]))
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
self._appendrev(f, blocks, bannotated)
elif showpath: # not append linelog, but we need to record path
self._node2path[f.node()] = f.path()
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
if progress: # clean progress bar
self.ui.write()
result = [
((self.revmap.rev2hsh(fr) if isinstance(fr, int) else fr.node()), l)
for fr, l in hist[revfctx][0]] # [(node, linenumber)]
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
return self._refineannotateresult(result, revfctx, showpath, showlines)
def canannotatedirectly(self, rev):
"""(str) -> bool, fctx or node.
return (True, f) if we can annotate without updating the linelog, pass
f to annotatedirectly.
return (False, f) if we need extra calculation. f is the fctx resolved
from rev.
"""
result = True
f = None
fastannotate: support replacing fctx.annotate directly Summary: Previously, I have performance concern about the signature of `fctx.annotate`: it returns fctxs, not ideal for performance since my initial goal is to get rid of reading revlogs in the best case. And letting the low-level `fctx.annotate` have some side effects writing to the disk is not that pretty. Because of that, fastannotate had re-invent part of the formatter (also optimized for performance somehow), and still cannot support all the features the original annotate supports, namely, templates. Now it makes sense to just replace `fctx.annotate` with a sub-optimal implementation, just for the flexibility of the template support of the original annotate command. We can use a "fake" or "lazy" fctx object to minimal the performance impact when people only need to print changeset nodes and line numbers - in which case we don't read revlog. Actually, the hgweb support already did a similar thing - converting fastannotate output to annotate output. So we can reuse some code. The next planned steps are: - Make the original "annotate" command aware of fastannotate protocol, do the pre-download stuff (downloading cache files at `fctx.annotate` is possible but inefficient because of not being batched). - "fastannotate" command remains a separated command optimized for perf and provides extra features like "--deleted". Because of the plan, the "commands" option does not make much sense - "fastannotate" command will not replace "annotate" directly thus dropped from the config interface. A new "modes" config option is added to control how fastannotate works. Test Plan: Modified existing tests Reviewers: #sourcecontrol, simonfar Reviewed By: simonfar Subscribers: mjpieters Differential Revision: https://phabricator.intern.facebook.com/D4238998 Signature: t1:4238998:1480447679:c48e0e565663c086293265e104d9cf414d913aa7
2016-11-29 14:51:08 +03:00
if not isinstance(rev, int) and rev is not None:
hsh = {20: bytes, 40: node.bin}.get(len(rev), lambda x: None)(rev)
if hsh is not None and (hsh, self.path) in self.revmap:
f = hsh
if f is None:
adjustctx = 'linkrev' if self._perfhack else True
f = self._resolvefctx(rev, adjustctx=adjustctx, resolverev=True)
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
result = f in self.revmap
if not result and self._perfhack:
# redo the resolution without perfhack - as we are going to
# do write operations, we need a correct fctx.
f = self._resolvefctx(rev, adjustctx=True, resolverev=True)
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
return result, f
def annotatealllines(self, rev, showpath=False, showlines=False):
"""(rev : str) -> [(node : str, linenum : int, path : str)]
the result has the same format with annotate, but include all (including
deleted) lines up to rev. call this after calling annotate(rev, ...) for
better performance and accuracy.
"""
revfctx = self._resolvefctx(rev, resolverev=True, adjustctx=True)
# find a chain from rev to anything in the mainbranch
if revfctx not in self.revmap:
chain = [revfctx]
a = ''
while True:
f = chain[-1]
pl = self._parentfunc(f)
if not pl:
break
if pl[0] in self.revmap:
a = pl[0].data()
break
chain.append(pl[0])
# both self.linelog and self.revmap is backed by filesystem. now
# we want to modify them but do not want to write changes back to
# files. so we create in-memory objects and copy them. it's like
# a "fork".
linelog = linelogmod.linelog()
linelog.copyfrom(self.linelog)
linelog.annotate(linelog.maxrev)
revmap = revmapmod.revmap()
revmap.copyfrom(self.revmap)
for f in reversed(chain):
b = f.data()
blocks = list(self._diffblocks(a, b))
self._doappendrev(linelog, revmap, f, blocks)
a = b
else:
# fastpath: use existing linelog, revmap as we don't write to them
linelog = self.linelog
revmap = self.revmap
lines = linelog.getalllines()
hsh = revfctx.node()
llrev = revmap.hsh2rev(hsh)
result = [(revmap.rev2hsh(r), l) for r, l in lines if r <= llrev]
# cannot use _refineannotateresult since we need custom logic for
# resolving line contents
if showpath:
result = self._addpathtoresult(result, revmap)
if showlines:
linecontents = self._resolvelines(result, revmap, linelog)
result = (result, linecontents)
return result
def _resolvelines(self, annotateresult, revmap, linelog):
"""(annotateresult) -> [line]. designed for annotatealllines.
this is probably the most inefficient code in the whole fastannotate
directory. but we have made a decision that the linelog does not
store line contents. so getting them requires random accesses to
the revlog data, since they can be many, it can be very slow.
"""
# [llrev]
revs = [revmap.hsh2rev(l[0]) for l in annotateresult]
result = [None] * len(annotateresult)
# {(rev, linenum): [lineindex]}
key2idxs = collections.defaultdict(list)
for i in xrange(len(result)):
key2idxs[(revs[i], annotateresult[i][1])].append(i)
while key2idxs:
# find an unresolved line and its linelog rev to annotate
hsh = None
try:
for (rev, _linenum), idxs in key2idxs.iteritems():
if revmap.rev2flag(rev) & revmapmod.sidebranchflag:
continue
hsh = annotateresult[idxs[0]][0]
break
except StopIteration: # no more unresolved lines
return result
if hsh is None:
# the remaining key2idxs are not in main branch, resolving them
# using the hard way...
revlines = {}
for (rev, linenum), idxs in key2idxs.iteritems():
if rev not in revlines:
hsh = annotateresult[idxs[0]][0]
if self.ui.debugflag:
self.ui.debug('fastannotate: reading %s line #%d '
'to resolve lines %r\n'
% (node.short(hsh), linenum, idxs))
fctx = self._resolvefctx(hsh, revmap.rev2path(rev))
lines = mdiff.splitnewlines(fctx.data())
revlines[rev] = lines
for idx in idxs:
result[idx] = revlines[rev][linenum]
assert all(x is not None for x in result)
return result
# run the annotate and the lines should match to the file content
self.ui.debug('fastannotate: annotate %s to resolve lines\n'
% node.short(hsh))
linelog.annotate(rev)
fctx = self._resolvefctx(hsh, revmap.rev2path(rev))
annotated = linelog.annotateresult
lines = mdiff.splitnewlines(fctx.data())
if len(lines) != len(annotated):
raise faerror.CorruptedFileError('unexpected annotated lines')
# resolve lines from the annotate result
for i, line in enumerate(lines):
k = annotated[i]
if k in key2idxs:
for idx in key2idxs[k]:
result[idx] = line
del key2idxs[k]
return result
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
def annotatedirectly(self, f, showpath, showlines):
"""like annotate, but when we know that f is in linelog.
f can be either a 20-char str (node) or a fctx. this is for perf - in
the best case, the user provides a node and we don't need to read the
filelog or construct any filecontext.
"""
if isinstance(f, str):
hsh = f
else:
hsh = f.node()
llrev = self.revmap.hsh2rev(hsh)
if not llrev:
raise faerror.CorruptedFileError('%s is not in revmap'
% node.hex(hsh))
if (self.revmap.rev2flag(llrev) & revmapmod.sidebranchflag) != 0:
raise faerror.CorruptedFileError('%s is not in revmap mainbranch'
% node.hex(hsh))
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
self.linelog.annotate(llrev)
result = [(self.revmap.rev2hsh(r), l)
for r, l in self.linelog.annotateresult]
return self._refineannotateresult(result, f, showpath, showlines)
def _refineannotateresult(self, result, f, showpath, showlines):
"""add the missing path or line contents, they can be expensive.
f could be either node or fctx.
"""
if showpath:
result = self._addpathtoresult(result)
if showlines:
if isinstance(f, str): # f: node or fctx
llrev = self.revmap.hsh2rev(f)
fctx = self._resolvefctx(f, self.revmap.rev2path(llrev))
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
else:
fctx = f
lines = mdiff.splitnewlines(fctx.data())
if len(lines) != len(result): # linelog is probably corrupted
raise faerror.CorruptedFileError()
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
result = (result, lines)
return result
def _appendrev(self, fctx, blocks, bannotated=None):
self._doappendrev(self.linelog, self.revmap, fctx, blocks, bannotated)
def _diffblocks(self, a, b):
return mdiff.allblocks(a, b, self.opts.diffopts)
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
@staticmethod
def _doappendrev(linelog, revmap, fctx, blocks, bannotated=None):
"""append a revision to linelog and revmap"""
def getllrev(f):
"""(fctx) -> int"""
# f should not be a linelog revision
if isinstance(f, int):
raise error.ProgrammingError('f should not be an int')
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
# f is a fctx, allocate linelog rev on demand
hsh = f.node()
rev = revmap.hsh2rev(hsh)
if rev is None:
rev = revmap.append(hsh, sidebranch=True, path=f.path())
return rev
# append sidebranch revisions to revmap
siderevs = []
siderevmap = {} # node: int
if bannotated is not None:
for (a1, a2, b1, b2), op in blocks:
if op != '=':
# f could be either linelong rev, or fctx.
siderevs += [f for f, l in bannotated[b1:b2]
if not isinstance(f, int)]
siderevs = set(siderevs)
if fctx in siderevs: # mainnode must be appended seperately
siderevs.remove(fctx)
for f in siderevs:
siderevmap[f] = getllrev(f)
# the changeset in the main branch, could be a merge
llrev = revmap.append(fctx.node(), path=fctx.path())
siderevmap[fctx] = llrev
for (a1, a2, b1, b2), op in reversed(blocks):
if op == '=':
continue
if bannotated is None:
linelog.replacelines(llrev, a1, a2, b1, b2)
else:
blines = [((r if isinstance(r, int) else siderevmap[r]), l)
for r, l in bannotated[b1:b2]]
linelog.replacelines_vec(llrev, a1, a2, blines)
def _addpathtoresult(self, annotateresult, revmap=None):
"""(revmap, [(node, linenum)]) -> [(node, linenum, path)]"""
if revmap is None:
revmap = self.revmap
def _getpath(nodeid):
path = self._node2path.get(nodeid)
if path is None:
path = revmap.rev2path(revmap.hsh2rev(nodeid))
self._node2path[nodeid] = path
return path
return [(n, l, _getpath(n)) for n, l in annotateresult]
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
def _checklastmasterhead(self, fctx):
"""check if fctx is the master's head last time, raise if not"""
if fctx is None:
llrev = 0
else:
llrev = self.revmap.hsh2rev(fctx.node())
if not llrev:
raise faerror.CannotReuseError()
fastannotate: implement the annotate algorithm Summary: This diff implements the `annotate` algorithm. Unlike the vanilla one, the annotate method takes 2 revisions: the revision specified for annotating, and the head of the main branch. The algorithm will do a "hybrid" annotate: incrementally update the linelog (the cache) so it can answer queries of any revision in the main branch. And use the traditional algorithm to deal with revisions not in the main branch: like a side branch of a merge commit, or the revision the user specified not in the main branch. The main branch is supposed to be something like `master` or `@`, and their p1s. Building up linelog with merge handled reasonably for the main branch, and the non-linelog part that produces final result share a lot internal states and logic so they are deeply coupled. Splitting them will probably reduce performance, or have difficulty (no clean way) to share internal states. If the caller only wants to build linelog without annotate things, just pass `rev = master`. While some attempts are made to support "merge" changeset, the result can still be different from the vanilla one sometimes. In those cases, both results make sense. It's really hard, if not impossible, to make the new implementation 100% same with the vanilla one because of the linear history restriction of linelog so I guess currently it's good enough. The differences will be covered by a `.t` test later. Test Plan: Code Review. A `.t` file will be added. Reviewers: #sourcecontrol, stash Reviewed By: stash Subscribers: stash, mjpieters Differential Revision: https://phabricator.intern.facebook.com/D3836438 Signature: t1:3836438:1473778829:27978479a01920833fa146f427178292ea1f5306
2016-09-08 18:54:21 +03:00
if self.linelog.maxrev != llrev:
raise faerror.CannotReuseError()
@util.propertycache
def _parentfunc(self):
"""-> (fctx) -> [fctx]"""
followrename = self.opts.followrename
followmerge = self.opts.followmerge
def parents(f):
pl = _parents(f, follow=followrename)
if not followmerge:
pl = pl[:1]
return pl
return parents
@util.propertycache
def _perfhack(self):
return self.ui.configbool('fastannotate', 'perfhack')
def _resolvefctx(self, rev, path=None, **kwds):
return resolvefctx(self.repo, rev, (path or self.path), **kwds)
def _unlinkpaths(paths):
"""silent, best-effort unlink"""
for path in paths:
try:
util.unlink(path)
except OSError:
pass
class pathhelper(object):
"""helper for getting paths for lockfile, linelog and revmap"""
fastannotate: implement a simple wireprotocol transferring the annotate cache Summary: This diff adds a simple wireprotocol that the client can download fastannotate cache files from the server. Design decisions: - Transfer full files. The linelog is not append-only and rsync-like incremental updates is not trivial. Since the files look small (usually < 1MB), send full copies for now. If file size becomes an issue, we can improve it later. - The server can send arbitrary number of files with paths defined. This provides the flexibility to send extra files without changing the protocol. For example, if we want to speed up `--deleted`, we may want to download `linemap` files. - No compression. The revmap is basically not compress-able because it contains the hashes. The linelog could be compressed though, but they are small enough, and we may rely on compression on other layers, like SSH. - `lastnode` check - if the last node of revmap the client reports matches what the server has, the server will assume it's up-to-date and do nothing. Note that this diff is only about the protocol and methods on peer, localrepo objects. The client-side command is not aware of the feature yet. Test Plan: - Clone from a local repo A to B, via the ssh protocol. - Config both repos so B is the client, A is the server. - Run `hg debugshell --debug` from B, make sure the client can get contents correctly (check the file contents from both the client and the server), also make sure the non-existed files will be ignored: In [1]: repo.prefetchfastannotate(['README', 'Makefile', 'NOT-EXIStED']) running ssh localhost ... sending hello command sending between command remote: 383 remote: capabilities: .... getannotate remote: 1 fastannotate: requesting 2 files sending batch command fastannotate: server returned fastannotate: writing 4976 bytes to fastannotate/default/README.l fastannotate: writing 813 bytes to fastannotate/default/README.m fastannotate: writing 9024 bytes to fastannotate/default/Makefile.l fastannotate: writing 2453 bytes to fastannotate/default/Makefile.m - Make sure the second time fetching these files, they will be skipped because of the uptodate check: In [2]: repo.prefetchfastannotate(['README', 'Makefile', 'NOT-EXIStED']) - Manually create a diverge case where the client and the server disagrees about what "master" (or "mainbranch") is. Make sure that although the request will be send by the client, the server can return nothing because of the lastnode check. Reviewers: #mercurial, durham, stash Reviewed By: stash Subscribers: mjpieters Differential Revision: https://phabricator.intern.facebook.com/D4063249 Signature: t1:4063249:1477920081:9bfab3031aa3118060c433f6e27c5b32aa25040c
2016-10-30 00:45:20 +03:00
def __init__(self, repo, path, opts=defaultopts):
# different options use different directories
self._vfspath = os.path.join('fastannotate',
opts.shortstr, encodedir(path))
self._repo = repo
@property
def dirname(self):
return os.path.dirname(self._repo.vfs.join(self._vfspath))
@property
def linelogpath(self):
return self._repo.vfs.join(self._vfspath + '.l')
def lock(self):
return lockmod.lock(self._repo.vfs, self._vfspath + '.lock')
@contextlib.contextmanager
def _lockflock(self):
"""the same as 'lock' but use flock instead of lockmod.lock, to avoid
creating temporary symlinks."""
import fcntl
lockpath = self.linelogpath
util.makedirs(os.path.dirname(lockpath))
lockfd = os.open(lockpath, os.O_RDONLY | os.O_CREAT, 0o664)
fcntl.flock(lockfd, fcntl.LOCK_EX)
try:
yield
finally:
fcntl.flock(lockfd, fcntl.LOCK_UN)
os.close(lockfd)
@property
def revmappath(self):
return self._repo.vfs.join(self._vfspath + '.m')
@contextlib.contextmanager
def annotatecontext(repo, path, opts=defaultopts, rebuild=False):
"""context needed to perform (fast) annotate on a file
an annotatecontext of a single file consists of two structures: the
linelog and the revmap. this function takes care of locking. only 1
process is allowed to write that file's linelog and revmap at a time.
when something goes wrong, this function will assume the linelog and the
revmap are in a bad state, and remove them from disk.
use this function in the following way:
with annotatecontext(...) as actx:
actx. ....
"""
helper = pathhelper(repo, path, opts)
util.makedirs(helper.dirname)
revmappath = helper.revmappath
linelogpath = helper.linelogpath
actx = None
try:
with helper.lock():
actx = _annotatecontext(repo, path, linelogpath, revmappath, opts)
if rebuild:
actx.rebuild()
yield actx
except Exception:
if actx is not None:
actx.rebuild()
repo.ui.debug('fastannotate: %s: cache broken and deleted\n' % path)
raise
finally:
if actx is not None:
actx.close()
def fctxannotatecontext(fctx, follow=True, diffopts=None, rebuild=False):
"""like annotatecontext but get the context from a fctx. convenient when
used in fctx.annotate
"""
repo = fctx._repo
path = fctx._path
if repo.ui.configbool('fastannotate', 'forcefollow', True):
follow = True
aopts = annotateopts(diffopts=diffopts, followrename=follow)
return annotatecontext(repo, path, aopts, rebuild)