wdir() implementation is still incomplete and shouldn't be advertised to
users. This patch will be backed out when
- template values such as {rev} and {node} are settled
- major commands and revsets work without crashing
I came across a case where an internal command was using a revset that I didn't
immediately pass in and it was difficult to debug what was going wrong with the
revset. This prints out the revset and informs the user that the error is with
a rebset so it should be more obvious what and where the error is.
This will be useful to execute actions after the tree is parsed and
before the revset returns a match. Finding symbols in the parse tree
will later allow hashes of hidden revisions to work on the command
line without the --hidden flag.
If self is a smartset and other is a fullreposet, nothing should be necessary.
A small win for trivial query in mozilla-central repo:
revset #0: (0:100000)
0) wall 0.017211 comb 0.020000 user 0.020000 sys 0.000000 (best of 163)
1) wall 0.001324 comb 0.000000 user 0.000000 sys 0.000000 (best of 2160)
This returns the csets where matching subrepos have changed with respect to the
containing repo's first parent. The second parent shouldn't matter, because it
is either syncing up to the first parent (i.e. it hasn't changed from the
current branch's POV), or the merge changed it with respect to the first parent
(which already adds it to the set).
There's already a 'subrepo' fileset, but it is prefixed with 'set:', so there
should be no ambiguity (in code anyway). The only test I see for it is to
revert subrepos named by a glob pattern (in test-subrepo.t, line 58). Since it
doesn't return a tracked file, neither 'log "set:subrepo()"' nor
'files "set:subrepo()"' print anything. Therefore, it seems useful to have a
revset that will return something for log (and can be added to a revsetalias to
be chained with 'file' revsets.)
It might be nice to be able to filter for added, modified and removed
separately, but add/remove should be rare. It might also be nice to be able to
do a 'contains' check, in addition to this mutated check. Maybe it is possible
to get those with the existing 'adds', 'contains', 'modifies' and 'removes' by
teaching them to chase explicit paths into subrepos.
I'm not sure if this should be added to the 'modifies adds removes' line in
revset.optimize() (since it is doing an AMR check on .hgsubstate), or if it is
OK to put into 'safesymbols' (things like 'file' are on the list, and that takes
a regex, among other patterns).
This patvh speeds up the computation of the not public() changeset
and incidentally speed up the computation of divergents() changeset on our big
repo by 100x from 50% to 0.5% of the time spent in smartlog with evolve.
In this patch we optimize not public() to _notpublic() (new revset) and use
the work on phaseset (from the previous commit) to be able to compute
_notpublic() quickly.
We use a non-lazy approach making the assumption the number of notpublic
change will not be in the order of magnitude of the repo size. Adopting a
lazy approach gives a speedup of 5x (vs 100x) only due to the overhead of the
code for lazy generation.
Before this change, doing ordered iteration over an 'addset' object composed of
operands without fastasc or fastdesc method could result in duplicated entries.
This was the result of applying '_iterordered' on an unordered set.
We fix it by ensuring we iterate over the set in a sorted order. Using the fast
iterator when it exists on any operand. We kill the '_iterator' method in the
process because it did not make a lot of sense independently.
Thanks goes to Yuya Nishihara for reporting the issue and analysing the cause.
The main purpose of wdir() is to annotate working-directory files.
Currently many commands and revsets cannot handle workingctx and may raise
exception. For example, -r ":wdir()" results in TypeError. This problem will
be addressed by future patches.
We could add "wdir" symbol instead, but it would conflict with the existing
tag, bookmark or branch. So I decided not to.
List of commands that will potentially support workingctx revision:
command default remarks
-------- ------- -----------------------------------------------------
annotate p1 useful
archive p1 might be useful
cat p1 might be useful on Windows (no cat)
diff p1:wdir (default)
export p1 might be useful if wctx can have draft commit message
files wdir (default)
grep tip:0 might be useful
identify wdir (default)
locate wdir (default)
log tip:0 might be useful with -p or -G option
parents wdir (default)
status wdir (default)
This patch includes minimal test of "hg status" that should be able to handle
the workingctx revision.
Previously we would instantiate the revbranchcache with a repo object, use it
briefly, then require it be passed in every time we wanted to fetch any
information. This seems unnecessary since it's obviously specific to that repo
(since it was constructed with it).
This patch stores the repo on the revbranchcache object, and removes the repo
parameter from the various functions on that class. This has the other nice
benefit of removing the double-revbranchcache-read that existed before (it was
read once for the branch revset, and once for the repo.revbranchcache).
Although Python supports `X = Y if COND else Z`, this was only
introduced in Python 2.5. Since we have to support Python 2.4, it was
a very common thing to write instead `X = COND and Y or Z`, which is a
bit obscure at a glance. It requires some intricate knowledge of
Python to understand how to parse these one-liners.
We change instead all of these one-liners to 4-liners. This was
executed with the following perlism:
find -name "*.py" -exec perl -pi -e 's,(\s*)([\.\w]+) = \(?(\S+)\s+and\s+(\S*)\)?\s+or\s+(\S*)$,$1if $3:\n$1 $2 = $4\n$1else:\n$1 $2 = $5,' {} \;
I tweaked the following cases from the automatic Perl output:
prev = (parents and parents[0]) or nullid
port = (use_ssl and 443 or 80)
cwd = (pats and repo.getcwd()) or ''
rename = fctx and webutil.renamelink(fctx) or []
ctx = fctx and fctx or ctx
self.base = (mapfile and os.path.dirname(mapfile)) or ''
I also added some newlines wherever they seemd appropriate for readability
There are probably a few ersatz ternary operators still in the code
somewhere, lurking away from the power of a simple regex.
As per fullreposet.__and__, it can omit the range check of rev. Therefore,
"null" revision is accepted automagically.
It seems this can fix many query results involving null symbol. Originally,
the simplest "(null)" query did fail if there were hidden revisions. Tests
are randomly chosen.
fullreposet mimics the behavior of localrepo, where "null" revision is not
listed but contained.
fcccbf073394 says we should avoid function calls in __contains__, so
super(fullreposet, self).__contains__(rev) is not an option.
Actually the super call doubled the benchmark result of trivial query:
revisions:
0) 6aa81b0c4658 (tip when I wrote this patch)
1) rev == node.nullrev or super(fullreposet, self).__contains__(rev)
revset #0: tip:0
0) wall 0.008441 comb 0.010000 user 0.010000 sys 0.000000 (best of 282)
1) wall 0.016152 comb 0.010000 user 0.010000 sys 0.000000 (best of 146)
I'm not sure if "all()" should filter out "null", but "all()" is stated as
'the same as "0:tip"' (except that it doesn't reorder the subset, I think.)
This patch is intended to avoid exposing a fullreposet to graphmod.dagwalker(),
which would result in strange drawing in future version:
|
o changeset: 0:f8035bb17114
| user: test
| date: Thu Jan 01 00:00:00 1970 +0000
| summary: add a
caused by:
parents = sorted(set([p.rev() for p in ctx.parents()
if p.rev() in revs]))
We cannot add "and p.rev() != nullrev" here because revs may actually include
"null" revision.
Transitioning to Mercurial versions with revision branch cache could be slow as
long as all operations were readonly (revset queries) and the cache would be
populated but not written back.
Instead, fall back to using the consistently slow path when readonly and the
cache doesn't exist yet. That avoids the overhead of populating the cache
without writing it back.
If not readonly, it will still populate all missing entries initially. That
avoids repeated writing of the cache file with small updates, and it also makes
sure a fully populated cache available for the readonly operations.
Before this patch, revset predicate "tag()" and "named('tags')" differ
from each other, because the former doesn't include "tip" but the
latter does.
For equivalence, "named('tags')" shouldn't include the revision
corresponded to "tip". But just removing "tip" from the "tags"
namespace causes breaking backward compatibility, even though "tip"
itself is planned to be eliminated, as mentioned below.
http://selenic.com/pipermail/mercurial-devel/2015-February/066157.html
To mask specific names ("tip" in this case) for "named()" predicate,
this patch introduces "deprecated" into "namespaces", and makes
"named()" predicate examine whether each names are masked by the
namespace, to which they belong.
"named()" will really work correctly after 3.3.1 (see a3c326a7f57a for
detail), and fixing this on STABLE before 3.3.1 can prevent initial
users of "named()" from expecting "named('tags')" to include "tip".
It is reason why this patch is posted for STABLE, even though problem
itself isn't so serious.
This may have to be flagged as "(BC)", if applied on DEFAULT.
Before this patch, revset predicate "named()" uses each nodes gotten
from target namespaces directly.
This causes problems below:
- combination of other predicates doesn't work correctly, because
they assume that revisions are listed up in number
- "hg log" doesn't show any revisions for "named()" result, because:
- "changeset_printer" stores formatted output for each revisions
into dict with revision number (= ctx.rev()) as a key of them
- "changeset_printer.flush(rev)" writes stored output for
the specified revision, but
- "commands.log" invokes it with the node, gotten from "named()"
- "hg debugrevspec" shows nodes (= may be binary) directly
Difference between revset predicate "tag()" and "named('tags')" in
tests is fixed in subsequent patch.
Before this patch, "bookmark()", "named()" and "tag()" predicates
raise "Abort", when the specified pattern doesn't match against
existing ones.
This prevents "present()" predicate from continuing the query, because
it only catches "RepoLookupError".
This patch raises "RepoLookupError" instead of "Abort", to make
"present()" predicate continue the query, even if "bookmark()",
"named()" or "tag()" in the sub-query of it are aborted.
This patch doesn't contain raising "RepoLookupError" for "re:" pattern
in "tag()", because "tag()" treats it differently from others. Actions
of each predicates at failure of pattern matching can be summarized as
below:
predicate "literal:" "re:"
---------- ----------- ------------
bookmark abort abort
named abort abort
tag abort continue (*1)
branch abort continue (*2)
---------- ----------- ------------
"tag()" may have to abort in the (*1) case for similarity, but this
change may break backward compatibility of existing revset queries. It
seems to have to be changed on "default" branch (with "BC" ?).
On the other hand, (*2) seems to be reasonable, even though it breaks
similarity, because "branch()" in this case doesn't check exact
existence of branches, but does pick up revisions of which branch
matches against the pattern.
This patch also adds tests for "branch()" to clarify behavior around
"present()" of similar predicates, even though this patch doesn't
change "branch()".
This can simplify the conversion from numeric revision to string. Without it,
we have to handle -1 specially because repo['-1'] != repo[-1].
The -1 revision is not officially documented, but this change makes sense
assuming that "rev(%d)" exists for scripting or third-party tools.
When running "hg log 'set:added()'", we create two matchers: one used
for producing the revset and one used for finding files to match. In
185b6b930e8c (graphlog: evaluate FILE/-I/-X filesets on the working
dir, 2012-02-26), we started passing a revision argument along from
what's currently in cmdutil._makelogrevset() to
revset._matchfiles(). When the revision was an empty string, it
referred to the working copy. This was subtly done with "repo[rev or
None]". Then, in 5ff5c5c9e69f (revset: avoid recalculating filesets,
2014-10-22), that conversion from empty string to None was lost. Note
that repo[''] is equivalent to repo['.'], not repo[None].
The consequence of this, to the user, is that when running "hg log
'set:added()'", the file matcher matches files added in the working
copy, while the revset matcher matches revisions that touch files
added in the parent of the working copy. As a result, only revisions
that touch any files added in the parent of the working copy will be
considered, but they will only be included if they also touch files
added in the working copy.
Fix the bug by converting '' to None again, but make it a little more
explicit this time (plus, we now have tests for it).
This change is intended to avoid exposing the implementation detail to
callers. I'm going to extend fullreposet to support "null" revision, so
these mfunc calls will have to use fullreposet() instead of spanset().
Before this patch, collisions between alias argument names in the
declaration are ignored, and this silently causes unexpected alias
evaluation.
This patch checks for such collisions, and aborts (or shows a warning) when
collisions are detected.
This patch doesn't add a test to "test-revset.t", because a doctest is
enough to test the collisions detection itself.
Before this patch, alias declaration is parsed by string base
operations: matching against "^([^(]+)\(([^)]+)\)$" and splitting by
",".
This overlooks many syntax errors like below (see the previous patch
introducing "_parsealiasdecl" for detail):
- un-closed parenthesis causes being treated as "alias symbol"
- symbol/function name aren't examined whether they are valid or not
- invalid argument list causes unexpected argument names
To parse alias declaration strictly, this patch replaces parsing
implementation by "_parsealiasdecl".
This patch tests only one typical declaration error case, because
error detection itself is already tested in the doctest of
"_parsealiasdecl".
This also removes class property "args" and "error", because these are
certainly initialized in "revsetalias.__init__".
This patch introduces "_parsealiasdecl" to parse alias declarations
strictly. For example, "_parsealiasdecl" can detect problems below,
which current implementation can't.
- un-closed parenthesis causes being treated as "alias symbol"
because all of declarations not in "func(....)" style are
recognized as "alias symbol".
for example, "foo($1, $2" is treated as the alias symbol.
- alias symbol/function names aren't examined whether they are valid
as symbol or not
for example, "foo bar" can be treated as the alias symbol, but of
course such invalid symbol can't be referred in revset.
- just splitting argument list by "," causes overlooking syntax
problems in the declaration
for example, all of invalid declarations below are overlooked:
- foo("bar") => taking one argument named as '"bar"'
- foo("unclosed) => taking one argument named as '"unclosed'
- foo(bar::baz) => taking one argument named as 'bar::baz'
- foo(bar($1)) => taking one argument named as 'bar($1)'
To decrease complication of patch, current implementation for alias
declarations is replaced by "_parsealiasdecl" in the subsequent
patch. This patch just introduces it.
This patch defines "_parsealiasdecl" not as a method of "revsetalias"
class but as a one of "revset" module, because of ease of testing by
doctest.
This patch factors some helper functions for "tree" out, because:
- direct accessing like "if tree[0] == 'func' and len(tree) > 1"
decreases readability
- subsequent patch (and also existing code paths, in the future) can
use them for readability
This patch also factors "_tokenizealias" out, because it can be used
also for parsing alias definitions strictly.
Before this patch, any errors in the declaration of revset alias
aren't detected at all, and there is no information about error source
in the error message.
As a part of preparation for parsing alias declarations and
definitions more strictly, this patch stores full detail into
"revsetalias.error" for error source distinction.
This makes raising "Abort" and warning potential errors just use
"revsetalias.error" without any message composing.
This patch defines the composing function not in "ParseError" class but
in "revset" module, because:
- "_()" shouldn't be used in "ParseError", to avoid adding "from
i18n import _" i18n" to "error" module
- generalizing message composition of"ParseError" for all code paths
other than revset isn't the purpose of this patch
we should also take care of showing "unexpected leading
whitespace" for some code paths, to generalize widely.
Before this patch, "tokenize" doesn't recognize the symbol starting
with "$" as a valid one.
This prevents revset alias declarations and definitions from being
parsed with "tokenize", because "$" may be used as the initial letter
of alias arguments.
BTW, the alias argument name doesn't require leading "$" itself, in
fact. But we have to assume that users may use "$" as the initial
letter of argument names in their aliases, because examples in "hg
help revsets" uses such names for a long time.
To make "tokenize" extensible to parse alias declarations and
definitions, this patch introduces optional arguments "syminitletters"
and "symletters". Giving these sets can change the policy of "valid
symbol" in tokenization easily.
This patch keeps original examination of letter validity for
reviewability, even though there is redundant interchanging between
"chr"/"ord" at initialization of "_syminitletters" and "_symletters".
At most 256 times examination (per initialization) is cheaper enough
than revset evaluation itself.
This patch is a part of preparation for parsing alias declarations and
definitions more strictly.
Because spanset.isascending() ignored the ascending flag, the result of
"fullreposet() & x" was always sorted in ascending order.
The test case is carefully chosen to call fullreposet.__and__.
Branch name filtering in revsets was expensive. For every rev it created a
changectx and called .branch() which retrieved the branch name from the
changelog.
Instead, use the revbranchcache.
The revbranchcache is used read-only. The revset implementation with generators
and callbacks makes it hard to figure out when we are done using/updating the
cache and could write it back. It would also be 'tricky' to lock the repo for
writing from within a revset execution. Finally, the branchmap update will
usually make sure that the cache is updated before any revset can be run.
The revbranchcache is used without any locking but is short-lived and used in a
tight loop where we can assume that the changelog doesn't change ... or where
it not is relevant to us if it does.
perfrevset 'branch(mobile)' on mozilla-central.
Before:
! wall 10.989637 comb 10.970000 user 10.940000 sys 0.030000 (best of 3)
After, no cache:
! wall 7.368656 comb 7.370000 user 7.360000 sys 0.010000 (best of 3)
After, with cache:
! wall 0.528098 comb 0.530000 user 0.530000 sys 0.000000 (best of 18)
The performance improvement even without cache come from being based on
branchinfo on the changelog instead of using ctx.branch().
Some tests are added to verify that the revbranchcache works and keep an eye on
when the cache files actually are updated.
It was introduced at deb42ca4dd93, where spanset.__contains__() did not exist.
Nowadays, we have to pay huge penalty for len(subset).
The following example showed that OR operation could be O(n * m^2)
(n: len(repo), m: number of OR operators, m >= 2) probably because of
filteredset.__len__.
revset #0: 0|1|2|3|4|5|6|7|8|9
0) wall 8.092713 comb 8.090000 user 8.090000 sys 0.000000 (best of 3)
1) wall 0.445354 comb 0.450000 user 0.430000 sys 0.020000 (best of 22)
2) wall 0.000389 comb 0.000000 user 0.000000 sys 0.000000 (best of 7347)
(0: 3.2.4, 1: 3.1.2, 2: this patch)
With this patch, we can make it much easier to specify 'only(A,B)' ->
A%B. Similarly, 'only(A)' -> A%.
On Windows, '%' is a semi-reserved symbol in the following way: using non-bash
shells (e.g. cmd.exe but NOT PowerShell, ConEmu, and cmder), %var% is only
expanded when 'var' exists and is surrounded by '%'.
That only leaves batch scripts which could prove to be problematic. I posit
that this isn't a big issue because any developer of batch scripts already
knows that to use '%' one needs to escape it by using a double '%%'.
Alternatives to '%' could be '=' but that might be limiting our future if we
ever decide to use temporary assignments in a revset.
fullreposet.__contains__() will be rewritten in order to support "null"
revision, and "rev()" won't be possible to rely on it.
This backs out 23ac42e12ce5, but there is no performance regression now.
revisions:
0) bd19f94d30e9 "l not in fullreposet(repo)"
1) this patch "l not in repo.changelog"
revset #0: rev(210000)
0) wall 0.000056 comb 0.000000 user 0.000000 sys 0.000000 (best of 48036)
1) wall 0.000049 comb 0.000000 user 0.000000 sys 0.000000 (best of 54969)
Before this patch, referring alias arguments is parsed by string base
operation "str.replace".
This causes problems below (see the previous patch introducing
"_parsealiasdefn" for detail)
- the shorter name argument breaks referring the longer name
- argument names in the quoted string are broken
This patch replaces parsing alias definition by "_parsealiasdefn" to
parse strictly.
This patch introduces "_parsealiasdefn" to parse alias definitions
strictly. For example, it can avoid problems below, which current
implementation can't.
- the shorter name argument breaks referring the longer name one in
the definition, if the former is completely prefix of the latter
for example, the alias definition "foo($1, $10) = $1 or $10" is
parsed as "_aliasarg('$1') or _aliasarg('$1')0" and causes parse
error, because tail "0" of "_aliasarg('$1')0" is invalid.
- argument names in the quoted string are broken
for example, the definition "foo($1) = $1 or desc('$1')" is parsed
as "_aliasarg('$1') or desc('_aliasarg(\'$1\')')" and causes
unexpected description matching against not '$1' but '_aliasarg(\'$1\')'.
To decrease complication of patch, current implementation for alias
definitions is replaced by "_parsealiasdefn" in the subsequent
patch. This patch just introduces it.
This patch defines "_parsealiasdefn" not as a method of "revsetalias"
class but as a one of "revset" module, because of ease of testing by
doctest.
Before this patch, there is no way to concatenate strings at runtime.
For example, to search for the issue ID "1234" in descriptions against
all of "issue 1234", "issue:1234", issue1234" and "bug(1234)"
patterns, the revset below should be written fully from scratch for
each issue ID.
grep(r"\bissue[ :]?1234\b|\bbug\(1234\)")
This patch introduces new infix operator "##" to concatenate
strings/symbols at runtime. Operator symbol "##" comes from the same
one of C pre-processor. This concatenation allows parametrizing a part
of strings in revset queries.
In the case of example above, the definition of the revset alias using
operator "##" below can search issue ID "1234" in complicated patterns
by "issue(1234)" simply:
issue($1) = grep(r"\bissue[ :]?" ## $1 ## r"\b|\bbug\(" ## $1 ## r"\)")
"##" operator does:
- concatenate not only strings but also symbols into the string
Exact distinction between strings and symbols seems not to be
convenience, because it is tiresome for users (and
"revset.getstring" treats both similarly)
For example of revset alias "issue()", "issue(1234)" is easier
than "issue('1234')".
- have higher priority than any other prefix, infix and postfix
operators (like as "##" of C pre-processor)
This patch (re-)assigns the priority 20 to "##", and 21 to "(",
because priority 19 is already assigned to "-" as prefix "negate".
The fix for linkrev pointing to hidden revision was crashing when the file was
missing from head's manifest. We now properly handle this case.
(yes I feel silly)
Before this patch, a problematic revset alias aborts execution
immediately, even if it isn't referred in the specified revset.
If old "hg" may be used too (for example, bisecting Mercurial itself),
it is also difficult to write alias definitions using features newly
introduced by newer "hg" into configuration files, because such alias
definitions cause unexpected abortion at parsing revset aliases with
old "hg".
This patch delays showing parse error for the revset alias until it is
actually referred at runtime.
This patch detects referring problematic aliases in "_expandaliases"
by examination of "revsetalias.error", which is initialized with the
error message only when parsing fails.
For usability, this patch also warns about problematic aliases, even
if they aren't referred at runtime. This should help users to know
potential problems in their alias definitions earlier.
If the file revision with a filtered linkrev does not have any
(unfiltered) children, we cannot use it to bound the search for
another introduction. Instead, we have to look at the file revision
used by each head changeset. If one of them uses this file revision, we
know there is another occurrence and we have a starting point. See
inline comments for details.
Adding some kind of permanent reference of all the introductions of a
file revision instead of just the first one would be much better. But
this is more difficult. I hope to take that into account in the next
repository format.
This revset is used by 'hg log FILENAME'. This prevent bugs when used on
a repository with hidden revisions.
Instead of just discarding file revisions whose linkrevs point to filtered
revisions, we put them aside and post-process them trying to find a non-filtered
introduction. See inline documentation for details about how it works.
This only fixes some of the problems. Once again, more will be needed when we can
cannot rely on child revisions of a file to find linkrev-shadowned revisions.
A test is added for 'hg log' catching such cases.
The follow revset (used by `hg log --follow`) now uses the new 'introrev'
method to bootstrap its traversal. This catches issues from linkrev-shadowing of
the changesets introducing the version of a file in source changeset.
A new test has been added to display pathological cases.
Another test is affected because it was meant to test this behavior but actually
failed to do so for half of the output. The output are now similar.
Before this patch, _revancestors were giving false result when a revision was
duplicated in the input. Duplicated entry are rare but may happen when using the
`%lx` notation internally.
This series has no visible impact on the performance of the function according
to benchmark.
The dequeue was actually just used to be able to pop value one at a time.
Building the dequeue means we are reading all the input value at once at the
beginning of the evaluation. This defeat the lazyness of revset.
We replace the deque with iterator usage for the sake of simplicity and
lazyness.
This provide massive speedup to get the first result if the input set is big
max(::all())
before) wall 0.001917 comb 0.000000 user 0.000000 sys 0.000000 (best of 1115)
after) wall 0.000107 comb 0.000000 user 0.000000 sys 0.000000 (best of 22222)
We usually use 'node' for variable containing 20 bytes hash. There is nothing
nodish in this variable, so we rename it to "inputrev" as it old the next entry
of the iteration.
The 'for r in self:' call could trigger full consumption of the generator while
we only need a single value. We also fast path if a single value got already
computed. See inline comment for more details.
This provide massive speedup for lazy operation using boolean testing.
max(::tip)
1e5463ae4044) wall 0.055609 comb 0.060000 user 0.060000 sys 0.000000 (best of 100)
after change) wall 0.000109 comb 0.000000 user 0.000000 sys 0.000000 (best of 19146)
It makes perfect sense for tokens to parse as existing revset symbols
(revset functions), and doesn't break anything, since parsing symbols
as functions works correctly in the presence of parens. For example,
if "only" is a bookmark, this used to error out,
hg log -r "only(only, @)"
which shouldn't, as the inner "only" is unambiguously not a function.
So we just remove the symbolset function and replace its calling site
with the stringset function.
For the tests, we confirm that "date" and "only" are both parsed as
revision names both inside revset expressions (e.g. an expression
containing ::) and inside old-style revision expressions (e.g. those
containing the name of the revision alone).
We were manually creating a base with explicit subset testing. We should let
smartset magic happen and optimise that logic if needed.
benchmark show some massive speedup when "parents set" is huge and "subset" is
small.
revset: 42:68 and roots(42:tip)
0) wall 0.011322 comb 0.010000 user 0.010000 sys 0.000000 (best of 161)
1) wall 0.002282 comb 0.010000 user 0.010000 sys 0.000000 (best of 1082)
Minor speedup in simple case (were fullreposet helps)
revset: roots(0::tip)
0) wall 0.095688 comb 0.100000 user 0.100000 sys 0.000000 (best of 85)
1) wall 0.084448 comb 0.080000 user 0.080000 sys 0.000000 (best of 95)
revset: roots((0:tip)::)
0) wall 0.146752 comb 0.140000 user 0.140000 sys 0.000000 (best of 58)
1) wall 0.143538 comb 0.140000 user 0.140000 sys 0.000000 (best of 59)
And small overhead then the "parents set" is fairly complicated (transforming it
into a revset once and for all appears to be faster).
revset: roots((tip~100::) - (tip~100::tip))
0) wall 0.004652 comb 0.010000 user 0.010000 sys 0.000000 (best of 544)
1) wall 0.004878 comb 0.010000 user 0.010000 sys 0.000000 (best of 479)
revset: roots((0::) - (0::tip))
0) wall 0.146587 comb 0.150000 user 0.150000 sys 0.000000 (best of 53)
1) wall 0.157192 comb 0.160000 user 0.160000 sys 0.000000 (best of 53)
revset: first(roots((0::) - (0::tip)))
0) wall 0.152924 comb 0.150000 user 0.150000 sys 0.000000 (best of 57)
1) wall 0.153192 comb 0.160000 user 0.160000 sys 0.000000 (best of 55)
Calling 'baseset(repo.changelog)' builds a list for all revisions in
the repo. And we already have the lazy and efficient 'fullreposet'
class for this purpose.
This gives us the usual benefits of the fullreposet but it is less visible
because the matching process itself is very expensive:
revset) matching(100)
before) wall 6.413281 comb 6.420000 user 5.910000 sys 0.510000 (best of 3)
after) wall 6.173608 comb 6.170000 user 5.750000 sys 0.420000 (best of 3)
However for some complex list, this provide a massive speedup
revset) matching(parents(100))
before) wall 23.890740 comb 23.890000 user 23.450000 sys 0.440000 (best of 3)
after) wall 6.382280 comb 6.390000 user 5.930000 sys 0.460000 (best of 3)
Calling 'baseset(repo.changelog)' builds a list for all revisions in
the repo. And we already have the lazy and efficient 'fullreposet'
class for this purpose.
This gives us the usual benefits of the fullreposet:
revset) 100^1
before) wall 0.002694 comb 0.000000 user 0.000000 sys 0.000000 (best of 897)
after) wall 0.000997 comb 0.000000 user 0.000000 sys 0.000000 (best of 2324)
revset) parents(100)^1
before) wall 0.003832 comb 0.000000 user 0.000000 sys 0.000000 (best of 587)
after) wall 0.001034 comb 0.000000 user 0.000000 sys 0.000000 (best of 2309)
revset) (100^1)^1
before) wall 0.005616 comb 0.000000 user 0.000000 sys 0.000000 (best of 405)
after) wall 0.001030 comb 0.000000 user 0.000000 sys 0.000000 (best of 2258)
Calling 'baseset(repo.changelog)' builds a list for all revisions in the
repo. And we already have the lazy and efficient 'fullreposet' class
for this purpose.
This gives us the usual benefits of the fullreposet:
revset) children(tip~100)
before) wall 0.007469 comb 0.010000 user 0.010000 sys 0.000000 (best of 338)
after) wall 0.003356 comb 0.000000 user 0.000000 sys 0.000000 (best of 755)
Calling 'baseset(repo.changelog)' builds a list for all revisions in
the repo. And we already have the lazy and efficient 'fullreposet'
class for this purpose.
This gives us the usual benefits of the fullreposet:
revset) 100~5
before) wall 0.002712 comb 0.000000 user 0.000000 sys 0.000000 (best of 918)
after) wall 0.000996 comb 0.000000 user 0.000000 sys 0.000000 (best of 2493)
revset) parents(100)~5
before) wall 0.003812 comb 0.010000 user 0.010000 sys 0.000000 (best of 667)
after) wall 0.001038 comb 0.000000 user 0.000000 sys 0.000000 (best of 2361)
revset) (100~5)~5
before) wall 0.005614 comb 0.000000 user 0.000000 sys 0.000000 (best of 446)
after) wall 0.001035 comb 0.000000 user 0.000000 sys 0.000000 (best of 2424)
Calling 'baseset(repo.changelog)' builds a list for all revisions in
the repo. And we already have the lazy and efficient 'fullreposet'
class for this purpose.
This gives us the usual benefit ofs the fullreposet:
revset) 10:100
before) wall 0.002774 comb 0.000000 user 0.000000 sys 0.000000 (best of 797)
after) wall 0.001977 comb 0.000000 user 0.000000 sys 0.000000 (best of 1244)
revset) parents(10):parents(100)
before) wall 0.005054 comb 0.000000 user 0.000000 sys 0.000000 (best of 481)
after) wall 0.002060 comb 0.000000 user 0.000000 sys 0.000000 (best of 1056)
The matcher variable 'm' in checkstatus() is reset to None on each
call, so the caching of the matcher no longer happens as it was
intended. This seems to be a regression in 6b9fbae54476 (revset: added
lazyset implementation to checkstatus, 2014-01-03).
Fix by moving the cached matcher into the enclosing function so it's
actually cached across calls. This speeds up
hg log -r 'modifies(mercurial/context.py)' >/dev/null
from 7.5s to 4s.
Also see similar fix in 5ff5c5c9e69f (revset: avoid recalculating
filesets, 2014-10-22).
hg log -r 1 ... -r 100 was never returning due to a regression in the
way addset computes __nonzero__. It used 'bool(self._r1 or self._r2)'
which required executing self._r1.__nonzero__ twice (once for the or,
once for the bool). hg log with a lot of -r's happens to build a one
sided addset tree of N length, which ends up being 2^N performance.
This patch fixes it by converting to bool before or'ing.
This problem can be repro'd with something as simple as:
hg log `for x in $(seq 1 50) ; do echo "-r $x "; done`
Adding '1 + 2 + ... + 20' to the revsetbenchmark.txt didn't seem to repro the
problem, so I wasn't able to add a revset benchmark for this issue.
0cc5c10d5dc7 was not the final version of that patch. It was really slow
because `l not in repo.changelog` iterates revisions up to `l`. Instead,
rev() should utilize spanset.__contains__().
revset #0: rev(210000)
0) wall 0.000039 comb 0.000000 user 0.000000 sys 0.000000 (best of 67978)
1) wall 0.002721 comb 0.000000 user 0.000000 sys 0.000000 (best of 1055)
2) wall 0.000059 comb 0.000000 user 0.000000 sys 0.000000 (best of 45599)
(0: 3.2-rc, 1: 0cc5c10d5dc7, 2: this patch)
Note that the benchmark result described in 0cc5c10d5dc7 is wrong because
it is the one of the initial version.
The recent optimization of "and" operation relies on the assumption that
the rhs set does not contain invalid revisions. So rev() has to remove
invalid revisions.
This is still faster than using `.filter(lambda r: r == l)`.
revset #0: rev(25)
0) wall 0.026341 comb 0.020000 user 0.020000 sys 0.000000 (best of 113)
1) wall 0.000038 comb 0.000000 user 0.000000 sys 0.000000 (best of 66567)
2) wall 0.000062 comb 0.000000 user 0.000000 sys 0.000000 (best of 43699)
(0: 428fa22fb2d1^, 1: 3.2-rc, 2: this patch)
The phase retrieval is fast enough to not require caching the result of the
functions.
draft()
0) wall 0.017209 comb 0.020000 user 0.020000 sys 0.000000 (best of 149)
1) wall 0.011654 comb 0.010000 user 0.010000 sys 0.000000 (best of 186)
public()
0) wall 0.018687 comb 0.010000 user 0.010000 sys 0.000000 (best of 128)
1) wall 0.013290 comb 0.010000 user 0.010000 sys 0.000000 (best of 181)
secret()
0) wall 0.017464 comb 0.020000 user 0.020000 sys 0.000000 (best of 127)
1) wall 0.011499 comb 0.000000 user 0.000000 sys 0.000000 (best of 196)
draft() - ::bookmark()
0) wall 0.020099 comb 0.020000 user 0.020000 sys 0.000000 (best of 127)
1) wall 0.014399 comb 0.020000 user 0.020000 sys 0.000000 (best of 169)
Instead of checking all elements of the subset against a single rev, just check
if this rev is in the subset. The old way was inherited from when the subset was
a list.
Non surprise, this provide massive speedup.
id("b7dc31e4baa4")
before) wall 0.008205 comb 0.000000 user 0.000000 sys 0.000000 (best of 302)
after) wall 0.000069 comb 0.000000 user 0.000000 sys 0.000000 (best of 34518)
revset #1: public() and id("b7dc31e4baa4")
before) wall 0.019763 comb 0.020000 user 0.020000 sys 0.000000 (best of 124)
after) wall 0.000101 comb 0.000000 user 0.000000 sys 0.000000 (best of 20130)
The & version is more likely to be optimised.
only(.)
before) wall 0.003216 comb 0.000000 user 0.000000 sys 0.000000 (best of 768)
after) wall 0.001086 comb 0.000000 user 0.000000 sys 0.000000 (best of 2231)
only(default, stable)
before) wall 0.018469 comb 0.020000 user 0.020000 sys 0.000000 (best of 138)
after) wall 0.015888 comb 0.010000 user 0.010000 sys 0.000000 (best of 156)
Except when stated otherwise, the condition used in `smartset.filter` will be
cached. A new argument has been introduced to disable that behavior. We use it
for filters created from `and` and `sub` operations.
This gives massive performance boosts for revsets with expensive conditions.
revset: branch(stable) or branch(default)
before) wall 4.329070 comb 4.320000 user 4.310000 sys 0.010000 (best of 3)
after) wall 2.356451 comb 2.360000 user 2.330000 sys 0.030000 (best of 4)
revset: author(mpm) or author(lmoscovicz)
before) wall 4.434719 comb 4.440000 user 4.440000 sys 0.000000 (best of 3)
after) wall 2.321720 comb 2.320000 user 2.320000 sys 0.000000 (best of 4)
Lazy revset broke the ordering of the `or` revset. We now stop assuming that
two ascending revset are combine into an ascending one.
Behavior in 3.0:
3:4 or 2:5 == [2, 3, 4, 5]
Behavior in 2.9:
3:4 or 2:5 == [3, 4, 2, 5]
We are adding a test for it.
For unclear reason, the performance `or` revset with expensive filter are
getting even worse than they used to be. This is probably caused by extra
uncached containment check or iteration.
revset #9: author(lmoscovicz) or author(mpm)
before) wall 3.487583 comb 3.490000 user 3.490000 sys 0.000000 (best of 3)
after) wall 4.481486 comb 4.480000 user 4.470000 sys 0.010000 (best of 3)
revset #10: author(mpm) or author(lmoscovicz)
before) wall 3.164839 comb 3.170000 user 3.160000 sys 0.010000 (best of 3)
after) wall 4.574965 comb 4.570000 user 4.570000 sys 0.000000 (best of 3)
We use the & operator to combine with subset (since this is more likely to be
optimised than filter) and we enforce the sorting of the result. Without this
enforced sorting, we may result in a different iteration order than the set
_descendent was computed from.
This reverts a bad `test-glog.t` change from 7904906883bd.
Another side effect is that `test-mq.t` shows `qparent::` including `-1` if
`qparent is -1`. This sound like a positive change.
This has good and bad impacts on the benchmarks, here is a good ones:
revset: 0::
before) wall 0.045489 comb 0.040000 user 0.040000 sys 0.000000 (best of 100)
after) wall 0.034330 comb 0.030000 user 0.030000 sys 0.000000 (best of 100)
revset: roots((0::) - (0::tip))
before) wall 0.134090 comb 0.140000 user 0.140000 sys 0.000000 (best of 63)
after) wall 0.128346 comb 0.130000 user 0.130000 sys 0.000000 (best of 69)
revset: ::p1(p1(tip))::
before) wall 0.143892 comb 0.140000 user 0.140000 sys 0.000000 (best of 55)
after) wall 0.124502 comb 0.130000 user 0.130000 sys 0.000000 (best of 65)
revset: roots((0:tip)::)
before) wall 0.204966 comb 0.200000 user 0.200000 sys 0.000000 (best of 43)
after) wall 0.184455 comb 0.180000 user 0.180000 sys 0.000000 (best of 47)
Here is a bad one:
revset: (20000::) - (20000)
before) wall 0.009592 comb 0.010000 user 0.010000 sys 0.000000 (best of 222)
after) wall 0.029837 comb 0.030000 user 0.030000 sys 0.000000 (best of 100)
The previous implementation was consuming the whole revset when asked for any
sort. The addset class is now doing lazy sorting like all other smarset classes.
This has no significant impact in the benchmark as-is. But this is important
to later change.
This add method is enforcing non-laziness, disabling multiple optimisations.
Benchmarks do not spot any significant difference but real usecase may. This
will also be important for further improvements to addset later in this series.
This add method is enforcing non-laziness, disabling multiple optimisations.
Benchmarks do not spot any significant regression but real usecase may. This
even gives some speedup in some cases:
revset #15: min(0::)
before) wall 0.001247 comb 0.000000 user 0.000000 sys 0.000000 (best of 1814)
after) wall 0.000942 comb 0.000000 user 0.000000 sys 0.000000 (best of 2367)
This will also be important for further improvement to addset later in this series.
This add method is enforcing non-laziness, disabling multiple optimisations.
Benchmarks do not spot any significant differences but real usecase may. This
will also be important for further improvements to addset later in this series.
A baseset starts without an explicit order. But as soon as a sort is requested,
we simply register that the baseset has an order and use the ordered version of
the list to behave accordingly.
We will want to properly record the order at creation time in the future. This
would unlock more optimisation and avoid some sorting.
Baseset contains already-computed revisions. It is considered "cheap" to do
operations on an already-computed set. So we add attributes to hold version of
the list in ascending and descending order and use them for `fastasc` and
`fastdesc`. Having distinct lists is important to provide correct iteration in
all cases. Altering a python list will impact an iterator connected to it.
eg: not preserving order at iterator creation time
>>> l = [0, 1]
>>> i = iter(l)
>>> l.reverse()
>>> list(i)
[1, 0]
eg: corrupting in progress iteration
>>> l = [0, 1]
>>> i = iter(l)
>>> i.next()
0
>>> l.reverse()
>>> i.next()
0
The baseset is doing more and more smartset magic and using its list-like
property less and less. So we store the list of revisions in an explicit
attribute and stop inheriting.
This requires reimplementing some basic methods.
In multiple places in the code, we use `someset[0]` or `someset[-1]`. This
works only because the `someset` is usually a baseset. For the same reason we
introduce a `first` and `last` methods to be implemented for all smartset
classes.
A `baseset` has multiple cached results and will get even more in the future.
Making it an object "populated once" like the other smartsets makes it both safer
and simpler. The append method will be removed at some point.
A `baseset` has multiple cached results and will get even more in the future.
Making it an object "populated once" like the other smartsets makes it both safer
and simpler. The append method will be removed at some point.
The expected iteration order may be different than the fast iteration order (eg:
ancestors(42) is expected to be iterated upward but is fast/lazy to compute
downward.
So we explicitly track the iteration order and enforce it if the manual
iteration is requested.
Default expected iteration order of a generator set is ascending because I'm
not aware of any descending revset that need a generatorset. The first to find
such descending revset will have the pleasure to make this configurable.
The utility of this cache is debatable (no visible benchmark impact) and using
generatorset for such purpose makes the code complicated.
We drop it for now. Someone can reintroduce a smart version of it in the future
if it is detected to be relevant.
When all revisions are known, we shortcut most of the class logic to use list
iteration instead. The cost of the sort is expected to be non-significant. The
list creation and sorting could be done lazily in the future. We have to copy
the list to not break existing iterator created before we finished consuming the
generator.
We gain a parameter to inform that the generator is ascending or descending. If
the generator is ordered, it is also used for the `fastasc` or `fastdesc`
version.
The _ascgeneratorset and _descgeneratorset class will be removed soon.
Better revset performance are also achieved with less overlay. There is no good
reason for addset to not be a smartset. We can replace the `_orderedsetmixin`
inheritance since `abstractsmartset` has efficient min and max too.
We have two goals here. First, we would like to restore the former iteration
order we had in 2.9. Second, we want this logic to be reusable for `fastasc`
and `fastdesc` methods.
All smartsets try to be lazy. The purpose of this class is to apply a
filter on another set. So we rename the class (and all its occurences) to
`filteredset`.
Just a bit of extra code makes the lazyset aware of order. This renders
orderedlazyset useless.
At some point, the `subset` will become responsible for this ordering logic. But
we are not there yet because the various objects used as subsets are not good enough.
When the filtered subset has such methods, we can use them. It is implemented
as properties to be able to quickly return None if no corresponding fastasc exists
on the subset.
Instead of having the direction of iteration enforced through the ordering of
`start` and `end` attributes of spanset, we encode the iteration direction in
an explicit attribute and always store start < end. The logic for sort and
reverse has to be updated. The __iter__ is now based on the newly introduced
`fastasc` and `fastdesc` methods.
This will allow other code simplifications in the future.
These two methods are actually silly aliases for `sort()` and
`sort(reverse=True)`. So we get that aliasing at the abstractsmartset level. We
will slowly phase out all the custom implementations and eventually remove any
mentions of it from the code.
Changeset 1440ec8e33c0 switched the order of the operand of the "&" computation
to work around an issue from repo-wide spanset. The need for a workaround has been
alleviated by the introduction of `fullreposet`. So we restore it to normal.
The benchmark shows no significant changes as expected.
We also revert the bogus test change introduced by 1440ec8e33c0. The order is
actually important.
This class documents all methods required by a smartset. This makes it easier
for people to respect the API and ensure we fail loudly when something does
not. It will later also contain common default implementations for multiple
methods, making it easier to have smartset classes with minimal work.
We are about to add a base class for `baseset` with an abstract `__nonzero__`
method. So we need this method to be explicitly defined to avoid issues. The
built-in list object apparently does not have a `__nonzero__` and relies on
`__len__` for this purpose?
Using `x.__contains__(r)` instead of `r in x` does not matter for built-in type
(set) but have a positive impact for all other classes. This will let us drop
some usage of baseset.set() in future patches. This also probably improves some
performance.
Doing manual iteration is expensible. We rely on built in list iteration
whenever possible. The other case has to become a closure we cannot have a both
yield and return in the same function.
This takes advantage of the `fullreposet` smartness with a nice
speedup. It's a similar speedup to `p1()` when a merge is in progress
(the non merge case is already lightning fast anyway.)
This takes advantage of the `fullreposet` smartness and yields a nice
speedup.
revset #0: p1()
0) wall 0.003256 comb 0.010000 user 0.010000 sys 0.000000 (best of 527)
1) wall 0.000066 comb 0.000000 user 0.000000 sys 0.000000 (best of 23224)
This takes advantage of the `fullreposet` smartness and yields a nice
speedup.
revset #0: rev(25)
0) wall 0.005480 comb 0.000000 user 0.000000 sys 0.000000 (best of 305)
1) wall 0.000052 comb 0.000000 user 0.000000 sys 0.000000 (best of 21891)
This takes advantage of the `fullreposet` smartness.
revset #0: origin(tip)
0) wall 0.005353 comb 0.000000 user 0.000000 sys 0.000000 (best of 354)
1) wall 0.003080 comb 0.000000 user 0.000000 sys 0.000000 (best of 446)
This takes advantage of the `fullreposet` smartness.
revset #0: follow(COPYING)
0) wall 0.002446 comb 0.000000 user 0.000000 sys 0.000000 (best of 735)
1) wall 0.000331 comb 0.000000 user 0.000000 sys 0.000000 (best of 5672)
This takes advantage of the `fullreposet` smartness.
revset #0: file(COPYING)
0) wall 3.179066 comb 3.180000 user 3.140000 sys 0.040000 (best of 3)
1) wall 2.723699 comb 2.730000 user 2.690000 sys 0.040000 (best of 4)
This takes advantage of the `fullreposet` smartness.
revset #0: divergent()
0) wall 0.002047 comb 0.000000 user 0.000000 sys 0.000000 (best of 813)
1) wall 0.000052 comb 0.000000 user 0.000000 sys 0.000000 (best of 22757)
This takes advantage of the `fullreposet` smartness.
revset #0: bisect(range)
0) wall 0.014007 comb 0.010000 user 0.010000 sys 0.000000 (best of 115)
1) wall 0.005556 comb 0.010000 user 0.010000 sys 0.000000 (best of 235)
This takes advantage of the `fullreposet` smartness.
revset #0: tip~25
0) wall 0.004800 comb 0.010000 user 0.010000 sys 0.000000 (best of 259)
1) wall 0.002475 comb 0.000000 user 0.000000 sys 0.000000 (best of 717)
This should give us the same benefit as elsewhere. Result is simpler (and
"faster").
Outgoing is dominated by the discovery so no benchmark is provided.
Python lookups are slow, so do all lookup outside of the for loop.
This provide a small but still significant speedup:
revset #0: 0::
0) wall 0.063258 comb 0.060000 user 0.060000 sys 0.000000 (best of 100)
1) wall 0.057776 comb 0.050000 user 0.050000 sys 0.000000 (best of 100)
Attribute lookup is slow in python. So this version is going to be a bit
faster. This does not have a visible impact since the rest of the stack is much
slower but this shaves the yak a few extra nanometers.
Moreover the new version is more readable so it worth doing this change for code
quality purpose.
This optimisation was approved by a core python dev.
"And" operation with something that contains the whole repo should be super
cheap. Check method docstring for details.
This provide massive boost to simple revset that use `subset & xxx`
revset #0: p1(20000)
0) wall 0.002447 comb 0.010000 user 0.010000 sys 0.000000 (best of 767)
1) wall 0.000529 comb 0.000000 user 0.000000 sys 0.000000 (best of 3947)
revset #1: p2(10000)
0) wall 0.002464 comb 0.000000 user 0.000000 sys 0.000000 (best of 913)
1) wall 0.000530 comb 0.000000 user 0.000000 sys 0.000000 (best of 4226)
No other regression spotted.
More performance improvements are expected in the future as more
revset predicate are converted to use `subset & xxx`
The relaxed way `fullreposet` handles "&" operation may cause some trouble for
people comparing smartset from different filter levels. I'm not sure such people
exist and we can improve that aspect in later patches.
We rename the `spanset` class to `_spanset`. `spanset` is now a function that
builds either a `fullreposet` or a `_spanset` according to the argument passed.
At some point, we may force people to explicitly use the `fullreposet`
constructor, but the current approach makes it easier to ensure we use the new
class whenever possible and focus on the benefits of this class.
Every revset evaluation starts from `subset = spanset(repo)` and a lot of
revset predicates build a `spansetrepo` for their internal needs.
`spanset` is a generic class that can handle any situation. As a result a lot
of operation between spanset result in an `orderedlazyset`, a safe object but
suboptimal in may situation.
So we introduce a `fullreposet` class where some of the operation will be
overwritten to produce more interesting results.
