A common mistake when writing binary search is to allow possible
integer overflow by using the simple average:
mid = (min + max) / 2;
Instead, use the overflow-safe version:
mid = min + (max - min) / 2;
This translation is safe since the operation occurs inside a loop
conditioned on "min < max". The included changes were found using
the following git grep:
git grep '/ *2;' '*.c'
Making this cleanup will prevent future review friction when a new
binary search is contructed based on existing code.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Long ago in 628522ec14 (sha1-lookup: more memory efficient
search in sorted list of SHA-1, 2007-12-29) we added
sha1_entry_pos(), a binary search that uses the uniform
distribution of sha1s to scale the selection of mid-points.
As this was a performance experiment, we tied it to the
GIT_USE_LOOKUP environment variable and never enabled it by
default.
This code was successful in reducing the number of steps in
each search. But the overhead of the scaling ends up making
it slower when the cache is warm. Here are best-of-five
timings for running rev-list on linux.git, which will have
to look up every object:
$ time git rev-list --objects --all >/dev/null
real 0m35.357s
user 0m35.016s
sys 0m0.340s
$ time GIT_USE_LOOKUP=1 git rev-list --objects --all >/dev/null
real 0m37.364s
user 0m37.045s
sys 0m0.316s
The USE_LOOKUP version might have more benefit on a cold
cache, as the time to fault in each page would dominate. But
that would be for a single lookup. In practice, most
operations tend to look up many objects, and the whole pack
.idx will end up warm.
It's possible that the code could be better optimized to
compete with a naive binary search for the warm-cache case,
and we could have the best of both worlds. But over the
years nobody has done so, and this is largely dead code that
is rarely run outside of the test suite. Let's drop it in
the name of simplicity.
This lets us remove sha1_entry_pos() entirely, as the .idx
lookup code was the only caller. Note that sha1-lookup.c
still contains sha1_pos(), which differs from
sha1_entry_pos() in two ways:
- it has a different interface; it uses a function pointer
to access sha1 entries rather than a size/offset pair
describing the table's memory layout
- it only scales the initial selection of "mi", rather
than each iteration of the search
We can't get rid of this function, as it's called from
several places. It may be that we could replace it with a
simple binary search, but that's out of scope for this patch
(and would need benchmarking).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If the first 18 bytes of the SHA1's of all entries are the same then
sha1_pos() dies and reports that the lower and upper limits of the
binary search were the same that this wasn't supposed to happen. This
is wrong because the remaining two bytes could still differ.
Furthermore: It wouldn't be a problem if they actually were the same,
i.e. if all entries have the same SHA1. The code already handles
duplicates just fine. Simply remove the erroneous check.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Acked-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The sha1_entry_pos function tries to be smart about
selecting the middle of a range for its binary search by
looking at the value differences between the "lo" and "hi"
constraints. However, it is unable to cope with entries with
duplicate keys in the sorted list.
We may hit a point in the search where both our "lo" and
"hi" point to the same key. In this case, the range of
values between our endpoints is 0, and trying to scale the
difference between our key and the endpoints over that range
is undefined (i.e., divide by zero). The current code
catches this with an "assert(lov < hiv)".
Moreover, after seeing that the first 20 byte of the key are
the same, we will try to establish a value from the 21st
byte. Which is nonsensical.
Instead, we can detect the case that we are in a run of
duplicates, and simply do a final comparison against any one
of them (since they are all the same, it does not matter
which). If the keys match, we have found our entry (or one
of them, anyway). If not, then we know that we do not need
to look further, as we must be in a run of the duplicate
key.
Signed-off-by: Jeff King <peff@peff.net>
Acked-by: Nicolas Pitre <nico@fluxnic.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This function has been copied from the "patch_pos" function in
"patch-ids.c" but an additional parameter has been added.
The new parameter is a function pointer, that is used to access the
sha1 of an element in the table.
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If we pick 'mi' between 'lo' and 'hi' at 50%, which was what the
simple binary search did, we are halving the search space
whether the entry at 'mi' is lower or higher than the target.
The previous patch was about picking not the middle but closer
to 'hi', when we know the target is a lot closer to 'hi' than it
is to 'lo'. However, if it turns out that the entry at 'mi' is
higher than the target, we would end up reducing the search
space only by the difference between 'mi' and 'hi' (which by
definition is less than 50% --- that was the whole point of not
using the simple binary search), which made the search less
efficient. And the risk of overshooting becomes very high, if
we try to be too precise.
This tweaks the selection of 'mi' to be a bit closer to the
middle than we would otherwise pick to avoid the problem.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently, when looking for a packed object from the pack idx, a
simple binary search is used.
A conventional binary search loop looks like this:
unsigned lo, hi;
do {
unsigned mi = (lo + hi) / 2;
int cmp = "entry pointed at by mi" minus "target";
if (!cmp)
return mi; "mi is the wanted one"
if (cmp > 0)
hi = mi; "mi is larger than target"
else
lo = mi+1; "mi is smaller than target"
} while (lo < hi);
"did not find what we wanted"
The invariants are:
- When entering the loop, 'lo' points at a slot that is never
above the target (it could be at the target), 'hi' points at
a slot that is guaranteed to be above the target (it can
never be at the target).
- We find a point 'mi' between 'lo' and 'hi' ('mi' could be
the same as 'lo', but never can be as high as 'hi'), and
check if 'mi' hits the target. There are three cases:
- if it is a hit, we have found what we are looking for;
- if it is strictly higher than the target, we set it to
'hi', and repeat the search.
- if it is strictly lower than the target, we update 'lo'
to one slot after it, because we allow 'lo' to be at the
target and 'mi' is known to be below the target.
If the loop exits, there is no matching entry.
When choosing 'mi', we do not have to take the "middle" but
anywhere in between 'lo' and 'hi', as long as lo <= mi < hi is
satisfied. When we somehow know that the distance between the
target and 'lo' is much shorter than the target and 'hi', we
could pick 'mi' that is much closer to 'lo' than (hi+lo)/2,
which a conventional binary search would pick.
This patch takes advantage of the fact that the SHA-1 is a good
hash function, and as long as there are enough entries in the
table, we can expect uniform distribution. An entry that begins
with for example "deadbeef..." is much likely to appear much
later than in the midway of a reasonably populated table. In
fact, it can be expected to be near 87% (222/256) from the top
of the table.
This is a work-in-progress and has switches to allow easier
experiments and debugging. Exporting GIT_USE_LOOKUP environment
variable enables this code.
On my admittedly memory starved machine, with a partial KDE
repository (3.0G pack with 95M idx):
$ GIT_USE_LOOKUP=t git log -800 --stat HEAD >/dev/null
3.93user 0.16system 0:04.09elapsed 100%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+55588minor)pagefaults 0swaps
Without the patch, the numbers are:
$ git log -800 --stat HEAD >/dev/null
4.00user 0.15system 0:04.17elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+60258minor)pagefaults 0swaps
In the same repository:
$ GIT_USE_LOOKUP=t git log -2000 HEAD >/dev/null
0.12user 0.00system 0:00.12elapsed 97%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+4241minor)pagefaults 0swaps
Without the patch, the numbers are:
$ git log -2000 HEAD >/dev/null
0.05user 0.01system 0:00.07elapsed 100%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+8506minor)pagefaults 0swaps
There isn't much time difference, but the number of minor faults
seems to show that we are touching much smaller number of pages,
which is expected.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Derrick Stolee
Jeff King
René Scharfe
Junio C Hamano
Christian Couder