The canonical way of doing 'roots(X)' is 'X - children(X)'. This is what the
implementation used to be. However, computing children is expensive because it
is unbounded. Any changesets in the repository may be a children of '0' so you
have to look at all changesets in the repository to compute children(0).
Moreover the current revsets implementation for children is not lazy, leading to
bad performance when fetching the first result.
There is a more restricted algorithm to compute roots:
roots(X) = [r for r in X if not parents(r) & X]
This achieve the same result while only looking for parent/children relation in
the X set itself, making the algorithm 'O(len(X))' membership operation.
Another advantages is that it turns the check into a simple filter, preserving
all laziness property of the underlying revsets.
The speed is very significant and some laziness is restored.
-) revset without 'roots(...)' to compare to base line
0) before this change
1) after this change
revset #0: roots((tip~100::) - (tip~100::tip))
plain min last
-) 0.001082 0.000993 0.000790
0) 0.001366 0.001385 0.001339
1) 0.001257 92% 0.001028 74% 0.000821 61%
revset #1: roots((0::) - (0::tip))
plain min last
-) 0.134551 0.144682 0.068453
0) 0.161822 0.171786 0.157683
1) 0.137583 85% 0.146204 85% 0.070012 44%
revset #2: roots(tip~100:)
plain min first last
-) 0.000219 0.000225 0.000231 0.000229
0) 0.000513 0.000529 0.000507 0.000539
1) 0.000463 90% 0.000269 50% 0.000267 52% 0.000463 85%
revset #3: roots(:42)
plain min first last
-) 0.000119 0.000146 0.000146 0.000146
0) 0.000231 0.000254 0.000253 0.000260
1) 0.000216 93% 0.000186 73% 0.000184 72% 0.000244 93%
revset #4: roots(not public())
plain min first
-) 0.000478 0.000502 0.000504
0) 0.000611 0.000639 0.000634
1) 0.000604 0.000560 87% 0.000558
revset #5: roots((0:tip)::)
plain min max first last
-) 0.057795 0.004905 0.058260 0.004908 0.038812
0) 0.132845 0.118931 0.130306 0.114280 0.127742
1) 0.111659 84% 0.005023 4% 0.111658 85% 0.005022 4% 0.092490 72%
revset #6: roots(0::tip)
plain min max first last
-) 0.032971 0.033947 0.033460 0.032350 0.033125
0) 0.083671 0.081953 0.084074 0.080364 0.086069
1) 0.074720 89% 0.035547 43% 0.077025 91% 0.033729 41% 0.083197
revset #7: 42:68 and roots(42:tip)
plain min max first last
-) 0.006827 0.000251 0.006830 0.000254 0.006771
0) 0.000337 0.000353 0.000366 0.000350 0.000366
1) 0.000318 94% 0.000297 84% 0.000353 0.000293 83% 0.000351
revset #8: roots(0:tip)
plain min max first last
-) 0.002119 0.000145 0.000147 0.000147 0.000147
0) 0.047441 0.040660 0.045662 0.040284 0.043435
1) 0.038057 80% 0.000187 0% 0.034919 76% 0.000186 0% 0.035097 80%
revset #0: roots(:42 + tip~42:)
plain min max first last sort
-) 0.000321 0.000317 0.000319 0.000308 0.000369 0.000343
0) 0.000772 0.000751 0.000811 0.000750 0.000802 0.000783
1) 0.000632 81% 0.000369 49% 0.000617 76% 0.000358 47% 0.000601 74% 0.000642 81%
If the computation of a set for each phase (done in C) is available,
we use it directly instead of applying a simple filter. This give a
massive speed-up in the vast majority of cases.
On my mercurial repo with about 15000 out of 40000 draft changesets:
revset: draft()
plain min first last
0) 0.011201 0.019950 0.009844 0.000074
1) 0.000284 2% 0.000312 1% 0.000314 3% 0.000315 x4.3
Bad performance for "last" come from the handling of the 15000 elements set
(memory allocation, filtering hidden changesets (99% of it) etc. compared to
applying the filter only on a handfuld of revisions (the first draft changesets
being close of tip).
This is not seen as an issue since:
* Timing is still pretty good and in line with all the other one,
* Current user of Vanilla Mercurial will not have 1/3 of their repo draft,
This bad effect disappears when phase's set is smaller. (about 200 secrets):
revset: secret()
plain min first last
0) 0.011181 0.022228 0.010851 0.000452
1) 0.000058 0% 0.000084 0% 0.000087 0% 0.000087 19%
Using 'repo[X]' is much slower because it creates a 'changectx' object and goes
though multiple layers of code to do so. It is also error prone if there is
tags, bookmarks, branch or other names that could map to a node hash and take
precedence (user are wicked).
This provides a significant performance boost on repository with a lot of
heads. Benchmark result for a repo with 1181 heads.
revset: head()
plain min last reverse
0) 0.014853 0.014371 0.014350 0.015161
1) 0.001402 9% 0.000975 6% 0.000874 6% 0.001415 9%
revset: head() - public()
plain min last reverse
0) 0.015121 0.014420 0.014560 0.015028
1) 0.001674 11% 0.001109 7% 0.000980 6% 0.001693 11%
revset: draft() and head()
plain min last reverse
0) 0.015976 0.014490 0.014214 0.015892
1) 0.002335 14% 0.001018 7% 0.000887 6% 0.002340 14%
The speed up is visible even when other more costly revset are in use
revset: head() and author("mpm")
plain min last reverse
0) 0.105419 0.090046 0.017169 0.108180
1) 0.090721 86% 0.077602 86% 0.003556 20% 0.093324 86%
This file should gather all revsets ever thought interesting by
anyone. That way one can check the impact of a change when touching
something revset-ish. See inline comments for details.
This file have been refilled with all the entry I could automatically
find from changeset descriptions. I assume we missed some not using
'revsetbenchmarks.py' output.