Why is --graph --max-count=n so much slower than --graph HEAD~n..?

I've noticed that --max-count doesn't seem to speed up `git log --graph' computation time. Here are some numbers using the linux kernel repository:

    | command                          | time* |
    |----------------------------------+-------|
    | git log --graph --max-count=5000 | 4.11s |
    | git log --graph --max-count=1000 | 4.05s |
    | git log --graph --max-count=500  | 3.95s |
    | git log --graph --max-count=50   | 3.95s |
    | git log --graph --max-count=10   | 3.97s |
    | git log --graph --max-count=1    | 3.96s |

However, specifying a manual revision range results in a much more dramatic speedup as `n' decreases:

    | command                          | time* |
    |----------------------------------+-------|
    | git log --graph HEAD~5000..      | 6.69s |
    | git log --graph HEAD~1000..      | 1.89s |
    | git log --graph HEAD~500..       | 0.03s |
    | git log --graph HEAD~50..        | 0.02s |
    | git log --graph HEAD~10..        | 0.02s |
    | git log --graph HEAD~1..         | 0.00s |

* All times are "steady state" measurements after warming up the disk
  cache by running the command a few times.

I see that the manual revision range doesn't seem to go back as far in history when I diff the output of the two commands. However, if I add `--max-count=50' to the `HEAD~50..' command I get the exact same commits but with fewer "dangling" lines at the bottom of the graph in the `HEAD~50..' approach.

Anyone care to provide any insight about what's going on here? Is this expected behavior? Is there any low-hanging fruit for optimizing --max-count --graph?

Mitchel Humpherys <mitch.special@gmail.com> writes:

AFAIK, --graph wants to compute the whole history and the max-count only affects the output phase after --graph does its computation.

Besides, "log --max-count=n" and "log HEAD~n.." compute completely different things, so the comparison is apples and oranges.

On Tue, May 20 2014 at 03:50:43 PM, Junio C Hamano <gitster@pobox.com> wrote:

Yes, apples and oranges in a black box :). I provided the HEAD~n.. measurements just to show that we can get (almost) the exact same output another way and it's much faster. It just "seems like" --max-count=n should speed things up as n decreases...

Surely. I am on a bus with terrible WiFi that does not let me use the usual terminal, but you would find a code in revision.c that sets revs->topo_order = 1 when it parses "--graph" option. If you disable it, that would stop "--graph" from wanting to compute the whole history before starting to emit stuff (and then stop at nth one with --max-count).

I do not know what other side effects such a change would have, though.

On Tue, May 20, 2014 at 5:13 PM, Mitchel Humpherys <mitch.special@gmail.com> wrote: