view mercurial/ancestor.py @ 42745:4d20b1fe8a72

rust-discovery: using from Python code

As previously done in other topics, the Rust version is used if it's been
built.

The version fully in Rust of the partialdiscovery class has the performance
advantage over the Python version (actually using the Rust MissingAncestor) if
the undecided set is big enough. Otherwise no sampling occurs, and the
discovery is reasonably fast anyway.

Note: it's hard to predict the size of the initial undecided set, it can
depend on the kind of topological changes between the local and remote graphs.
The point of the Rust version is to make the bad cases acceptable.

More specifically, the performance advantages are:

- faster sampling, especially takefullsample()
- much faster addmissings() in almost all cases (see commit message in
  grandparent of the present changeset)
- no conversion cost of the undecided set at the interface between Rust and
  Python

== Measurements with big undecided sets

For an extreme example, discovery between mozilla-try and mozilla-unified
(over one million undecided revisions, same case as in dbd0fcca6dfc), we
get roughly a x2.5/x3 better performance:

Growing sample size (5% starting with 200): time goes down from
210 to 72 seconds.
Constant sample size of 200: time down from 1853 to 659 seconds.

With a sample size computed from number of roots and heads of the
undecided set (`respectsize` is `False`), here are perfdiscovery results:

Before ! wall 9.358729 comb 9.360000 user 9.310000 sys 0.050000 (median of 50)
After  ! wall 3.793819 comb 3.790000 user 3.750000 sys 0.040000 (median of 50)

In that later case, the sample sizes are routinely in the hundreds of
thousands of revisions. While still faster, the Rust iteration in
addmissings has less of an advantage than with smaller sample sizes, but
one sees addcommons becoming faster, probably a consequence of not having
to copy big sets back and forth.

This example is not a goal in itself, but it showcases several different
areas in which the process can become slow, due to different factors, and
how this full Rust version can help.

== Measurements with small undecided sets

In cases the undecided set is small enough than no sampling occurs,
the Rust version has a disadvantage at init if `targetheads` is really big
(some time is lost in the translation to Rust data structures),
and that is compensated by the faster `addmissings()`.

On a private repository with over one million commits, we still get a minor
improvement, of 6.8%:

  Before ! wall 0.593585 comb 0.590000 user 0.550000 sys 0.040000 (median of 50)
  After  ! wall 0.553035 comb 0.550000 user 0.520000 sys 0.030000 (median of 50)

What's interesting in that case is the first addinfo() at 180ms for Rust and
233ms for Python+C, mostly due to add_missings and the children cache
computation being done in less than 0.2ms on the Rust side vs over 40ms on the
Python side.

The worst case we have on hand is with mozilla-try, prepared with
discovery-helper.sh for 10 heads and depth 10, time goes up 2.2% on the median.
In this case `targetheads` is really huge with 165842 server heads.

  Before ! wall 0.823884 comb 0.810000 user 0.790000 sys 0.020000 (median of 50)
  After  ! wall 0.842607 comb 0.840000 user 0.800000 sys 0.040000 (median of 50)

If that would be considered a problem, more adjustments can be made, which are
prematurate at this stage: cooking special variants of methods of the inner
MissingAncestors object, retrieving local heads directly from Rust to avoid
the cost of conversion. Effort would probably be better spent at this point
improving the surroundings if needed.

Here's another data point with a smaller repository, pypy, where performance
is almost identical

 Before ! wall 0.015121 comb 0.030000 user 0.020000 sys 0.010000 (median of 186)
 After  ! wall 0.015009 comb 0.010000 user 0.010000 sys 0.000000 (median of 184)

Differential Revision: https://phab.mercurial-scm.org/D6430

# ancestor.py - generic DAG ancestor algorithm for mercurial
#
# Copyright 2006 Matt Mackall <[e-mail]>
#
# 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 heapq

from .node import nullrev
from . import (
    dagop,
    policy,
    pycompat,
)

parsers = policy.importmod(r'parsers')

def commonancestorsheads(pfunc, *nodes):
    """Returns a set with the heads of all common ancestors of all nodes,
    heads(::nodes[0] and ::nodes[1] and ...) .

    pfunc must return a list of parent vertices for a given vertex.
    """
    if not isinstance(nodes, set):
        nodes = set(nodes)
    if nullrev in nodes:
        return set()
    if len(nodes) <= 1:
        return nodes

    allseen = (1 << len(nodes)) - 1
    seen = [0] * (max(nodes) + 1)
    for i, n in enumerate(nodes):
        seen[n] = 1 << i
    poison = 1 << (i + 1)

    gca = set()
    interesting = len(nodes)
    nv = len(seen) - 1
    while nv >= 0 and interesting:
        v = nv
        nv -= 1
        if not seen[v]:
            continue
        sv = seen[v]
        if sv < poison:
            interesting -= 1
            if sv == allseen:
                gca.add(v)
                sv |= poison
                if v in nodes:
                    # history is linear
                    return {v}
        if sv < poison:
            for p in pfunc(v):
                sp = seen[p]
                if p == nullrev:
                    continue
                if sp == 0:
                    seen[p] = sv
                    interesting += 1
                elif sp != sv:
                    seen[p] |= sv
        else:
            for p in pfunc(v):
                if p == nullrev:
                    continue
                sp = seen[p]
                if sp and sp < poison:
                    interesting -= 1
                seen[p] = sv
    return gca

def ancestors(pfunc, *orignodes):
    """
    Returns the common ancestors of a and b that are furthest from a
    root (as measured by longest path).

    pfunc must return a list of parent vertices for a given vertex.
    """
    def deepest(nodes):
        interesting = {}
        count = max(nodes) + 1
        depth = [0] * count
        seen = [0] * count
        mapping = []
        for (i, n) in enumerate(sorted(nodes)):
            depth[n] = 1
            b = 1 << i
            seen[n] = b
            interesting[b] = 1
            mapping.append((b, n))
        nv = count - 1
        while nv >= 0 and len(interesting) > 1:
            v = nv
            nv -= 1
            dv = depth[v]
            if dv == 0:
                continue
            sv = seen[v]
            for p in pfunc(v):
                if p == nullrev:
                    continue
                dp = depth[p]
                nsp = sp = seen[p]
                if dp <= dv:
                    depth[p] = dv + 1
                    if sp != sv:
                        interesting[sv] += 1
                        nsp = seen[p] = sv
                        if sp:
                            interesting[sp] -= 1
                            if interesting[sp] == 0:
                                del interesting[sp]
                elif dv == dp - 1:
                    nsp = sp | sv
                    if nsp == sp:
                        continue
                    seen[p] = nsp
                    interesting.setdefault(nsp, 0)
                    interesting[nsp] += 1
                    interesting[sp] -= 1
                    if interesting[sp] == 0:
                        del interesting[sp]
            interesting[sv] -= 1
            if interesting[sv] == 0:
                del interesting[sv]

        if len(interesting) != 1:
            return []

        k = 0
        for i in interesting:
            k |= i
        return set(n for (i, n) in mapping if k & i)

    gca = commonancestorsheads(pfunc, *orignodes)

    if len(gca) <= 1:
        return gca
    return deepest(gca)

class incrementalmissingancestors(object):
    '''persistent state used to calculate missing ancestors incrementally

    Although similar in spirit to lazyancestors below, this is a separate class
    because trying to support contains and missingancestors operations with the
    same internal data structures adds needless complexity.'''
    def __init__(self, pfunc, bases):
        self.bases = set(bases)
        if not self.bases:
            self.bases.add(nullrev)
        self.pfunc = pfunc

    def hasbases(self):
        '''whether the common set has any non-trivial bases'''
        return self.bases and self.bases != {nullrev}

    def addbases(self, newbases):
        '''grow the ancestor set by adding new bases'''
        self.bases.update(newbases)

    def basesheads(self):
        return dagop.headrevs(self.bases, self.pfunc)

    def removeancestorsfrom(self, revs):
        '''remove all ancestors of bases from the set revs (in place)'''
        bases = self.bases
        pfunc = self.pfunc
        revs.difference_update(bases)
        # nullrev is always an ancestor
        revs.discard(nullrev)
        if not revs:
            return
        # anything in revs > start is definitely not an ancestor of bases
        # revs <= start needs to be investigated
        start = max(bases)
        keepcount = sum(1 for r in revs if r > start)
        if len(revs) == keepcount:
            # no revs to consider
            return

        for curr in pycompat.xrange(start, min(revs) - 1, -1):
            if curr not in bases:
                continue
            revs.discard(curr)
            bases.update(pfunc(curr))
            if len(revs) == keepcount:
                # no more potential revs to discard
                break

    def missingancestors(self, revs):
        '''return all the ancestors of revs that are not ancestors of self.bases

        This may include elements from revs.

        Equivalent to the revset (::revs - ::self.bases). Revs are returned in
        revision number order, which is a topological order.'''
        revsvisit = set(revs)
        basesvisit = self.bases
        pfunc = self.pfunc
        bothvisit = revsvisit.intersection(basesvisit)
        revsvisit.difference_update(bothvisit)
        if not revsvisit:
            return []

        start = max(max(revsvisit), max(basesvisit))
        # At this point, we hold the invariants that:
        # - revsvisit is the set of nodes we know are an ancestor of at least
        #   one of the nodes in revs
        # - basesvisit is the same for bases
        # - bothvisit is the set of nodes we know are ancestors of at least one
        #   of the nodes in revs and one of the nodes in bases. bothvisit and
        #   revsvisit are mutually exclusive, but bothvisit is a subset of
        #   basesvisit.
        # Now we walk down in reverse topo order, adding parents of nodes
        # already visited to the sets while maintaining the invariants. When a
        # node is found in both revsvisit and basesvisit, it is removed from
        # revsvisit and added to bothvisit. When revsvisit becomes empty, there
        # are no more ancestors of revs that aren't also ancestors of bases, so
        # exit.

        missing = []
        for curr in pycompat.xrange(start, nullrev, -1):
            if not revsvisit:
                break

            if curr in bothvisit:
                bothvisit.remove(curr)
                # curr's parents might have made it into revsvisit through
                # another path
                for p in pfunc(curr):
                    revsvisit.discard(p)
                    basesvisit.add(p)
                    bothvisit.add(p)
                continue

            if curr in revsvisit:
                missing.append(curr)
                revsvisit.remove(curr)
                thisvisit = revsvisit
                othervisit = basesvisit
            elif curr in basesvisit:
                thisvisit = basesvisit
                othervisit = revsvisit
            else:
                # not an ancestor of revs or bases: ignore
                continue

            for p in pfunc(curr):
                if p == nullrev:
                    pass
                elif p in othervisit or p in bothvisit:
                    # p is implicitly in thisvisit. This means p is or should be
                    # in bothvisit
                    revsvisit.discard(p)
                    basesvisit.add(p)
                    bothvisit.add(p)
                else:
                    # visit later
                    thisvisit.add(p)

        missing.reverse()
        return missing

# Extracted from lazyancestors.__iter__ to avoid a reference cycle
def _lazyancestorsiter(parentrevs, initrevs, stoprev, inclusive):
    seen = {nullrev}
    heappush = heapq.heappush
    heappop = heapq.heappop
    heapreplace = heapq.heapreplace
    see = seen.add

    if inclusive:
        visit = [-r for r in initrevs]
        seen.update(initrevs)
        heapq.heapify(visit)
    else:
        visit = []
        heapq.heapify(visit)
        for r in initrevs:
            p1, p2 = parentrevs(r)
            if p1 not in seen:
                heappush(visit, -p1)
                see(p1)
            if p2 not in seen:
                heappush(visit, -p2)
                see(p2)

    while visit:
        current = -visit[0]
        if current < stoprev:
            break
        yield current
        # optimize out heapq operation if p1 is known to be the next highest
        # revision, which is quite common in linear history.
        p1, p2 = parentrevs(current)
        if p1 not in seen:
            if current - p1 == 1:
                visit[0] = -p1
            else:
                heapreplace(visit, -p1)
            see(p1)
        else:
            heappop(visit)
        if p2 not in seen:
            heappush(visit, -p2)
            see(p2)

class lazyancestors(object):
    def __init__(self, pfunc, revs, stoprev=0, inclusive=False):
        """Create a new object generating ancestors for the given revs. Does
        not generate revs lower than stoprev.

        This is computed lazily starting from revs. The object supports
        iteration and membership.

        cl should be a changelog and revs should be an iterable. inclusive is
        a boolean that indicates whether revs should be included. Revs lower
        than stoprev will not be generated.

        Result does not include the null revision."""
        self._parentrevs = pfunc
        self._initrevs = revs = [r for r in revs if r >= stoprev]
        self._stoprev = stoprev
        self._inclusive = inclusive

        self._containsseen = set()
        self._containsiter = _lazyancestorsiter(self._parentrevs,
                                                self._initrevs,
                                                self._stoprev,
                                                self._inclusive)

    def __nonzero__(self):
        """False if the set is empty, True otherwise."""
        try:
            next(iter(self))
            return True
        except StopIteration:
            return False

    __bool__ = __nonzero__

    def __iter__(self):
        """Generate the ancestors of _initrevs in reverse topological order.

        If inclusive is False, yield a sequence of revision numbers starting
        with the parents of each revision in revs, i.e., each revision is
        *not* considered an ancestor of itself. Results are emitted in reverse
        revision number order. That order is also topological: a child is
        always emitted before its parent.

        If inclusive is True, the source revisions are also yielded. The
        reverse revision number order is still enforced."""
        return _lazyancestorsiter(self._parentrevs, self._initrevs,
                                  self._stoprev, self._inclusive)

    def __contains__(self, target):
        """Test whether target is an ancestor of self._initrevs."""
        seen = self._containsseen
        if target in seen:
            return True
        iter = self._containsiter
        if iter is None:
            # Iterator exhausted
            return False
        # Only integer target is valid, but some callers expect 'None in self'
        # to be False. So we explicitly allow it.
        if target is None:
            return False

        see = seen.add
        try:
            while True:
                rev = next(iter)
                see(rev)
                if rev == target:
                    return True
                if rev < target:
                    return False
        except StopIteration:
            # Set to None to indicate fast-path can be used next time, and to
            # free up memory.
            self._containsiter = None
            return False

class rustlazyancestors(object):

    def __init__(self, index, revs, stoprev=0, inclusive=False):
        self._index = index
        self._stoprev = stoprev
        self._inclusive = inclusive
        # no need to prefilter out init revs that are smaller than stoprev,
        # it's done by rustlazyancestors constructor.
        # we need to convert to a list, because our ruslazyancestors
        # constructor (from C code) doesn't understand anything else yet
        self._initrevs = initrevs = list(revs)

        self._containsiter = parsers.rustlazyancestors(
            index, initrevs, stoprev, inclusive)

    def __nonzero__(self):
        """False if the set is empty, True otherwise.

        It's better to duplicate this essentially trivial method than
        to subclass lazyancestors
        """
        try:
            next(iter(self))
            return True
        except StopIteration:
            return False

    def __iter__(self):
        return parsers.rustlazyancestors(self._index,
                                         self._initrevs,
                                         self._stoprev,
                                         self._inclusive)

    def __contains__(self, target):
        return target in self._containsiter