patchcat-file: only use bitmaps when filtering

Commit 8002e8ee18 (builtin/cat-file: use bitmaps to efficiently filter by object type, 2025-04-02) introduced a performance regression when we are not filtering objects: it uses bitmaps even when they won't help, incurring extra costs. For example, running the new perf tests from this commit, which check the performance of listing objects by oid:

  $ export GIT_PERF_LARGE_REPO=/path/to/linux.git
  $ git -C "$GIT_PERF_LARGE_REPO" repack -adb
  $ GIT_SKIP_TESTS=p1006.1 ./run 8002e8ee18^ 8002e8ee18 p1006-cat-file.sh
  [...]
  Test                                  8002e8ee18^       8002e8ee18
  -------------------------------------------------------------------------------
  1006.2: list all objects (sorted)     1.48(1.44+0.04)   6.39(6.35+0.04) +331.8%
  1006.3: list all objects (unsorted)   3.01(2.97+0.04)   3.40(3.29+0.10) +13.0%
  1006.4: list blobs                    4.85(4.67+0.17)   1.68(1.58+0.10) -65.4%

An invocation that filters, like listing all blobs (1006.4), does benefit from using the bitmaps; it now doesn't have to check the type of each object from the pack data, so the tradeoff is worth it.

But for listing all objects in sorted idx order (1006.2), we otherwise would never open the bitmap nor the revindex file. Worse, our sorting step gets much worse. Normally we append into an array in pack .idx order, and the sort step is trivial. But with bitmaps, we get the objects in pack order, which is apparently random with respect to oid, and have to sort the whole thing. (Note that this freshly-packed state represents the best case for .idx sorting; if we had two packs, then we'd have their objects one after the other and qsort would have to interleave them).

The unsorted test in 1006.3 is interesting: there we are going in pack order, so we load the revindex for the pack anyway. And though we don't sort the result, we do use an oidset to check for duplicates. So we can see in the 8002e8ee18^ timings that those two things cost ~1.5s over the sorted case (mostly the oidset hash cost). We also incur the extra cost to open the bitmap file as of 8002e8ee18, which seems to be ~400ms. (This would probably be faster with a bitmap lookup table, but writing that out is not yet the default).

So we know that bitmaps help when there's filtering to be done, but otherwise make things worse. Let's only use them when there's a filter.

The perf script shows that we've fixed the regressions without hurting the bitmap case:

  Test                                  8002e8ee18^       8002e8ee18                HEAD
  --------------------------------------------------------------------------------------------------------
  1006.2: list all objects (sorted)     1.56(1.53+0.03)   6.44(6.37+0.06) +312.8%   1.62(1.54+0.06) +3.8%
  1006.3: list all objects (unsorted)   3.04(2.98+0.06)   3.45(3.38+0.07) +13.5%    3.04(2.99+0.04) +0.0%
  1006.4: list blobs                    5.14(4.98+0.15)   1.76(1.68+0.06) -65.8%    1.73(1.64+0.09) -66.3%

Note that there's another related case: we might have a filter that cannot be used with bitmaps. That check is handled already for us in for_each_bitmapped_object(), though we'd still load the bitmap and revindex files pointlessly in that case. I don't think it can happen in practice for cat-file, though, since it allows only blob:none, blob:limit, and object:type filters, all of which work with bitmaps.

It would be easy-ish to insert an extra check like:

  can_filter_bitmap(&opt->objects_filter);

into the conditional, but I didn't bother here. It would be redundant with the call in for_each_bitmapped_object(), and the can_filter helper function is static local in the bitmap code (so we'd have to make it public).

Signed-off-by: Jeff King <peff@peff.net>
---
There are some timing tests in 8002e8ee18 that claim the non-filter case
is not regressed, but it's not clear to me exactly which commands were
run. Given the size of the repo there and the fact that it's more I/O
bound, I'd guess it is using an output format that requires looking at
the packed objects.

You can see the mean user CPU does jump by almost 2 seconds in the timings given there. So it may be that the problem was there but drowned out by I/O noise. At any rate, the new t/perf tests isolate it better and reproduce consistently for me.

 builtin/cat-file.c       |  8 +++++---
 t/perf/p1006-cat-file.sh | 14 ++++++++++++++
 2 files changed, 19 insertions(+), 3 deletions(-)

On Tue, Jan 06, 2026 at 05:25:58AM -0500, Jeff King wrote:

I honestly can't remember anymore, either. I really should stick with the actual command run in the Hyperfine benchmark names.

Could be, yeah.

Yeah, this seems like a reasonable change to me. I would've preferred to avoid the assignment in the conditional, but other than that this looks good to me.

Thanks!

Patrick

On Wed, Jan 07, 2026 at 09:20:59AM +0100, Patrick Steinhardt wrote:

Yeah, I tried to rewrite this to avoid the assignment-in-conditional, but the logic gets even more convoluted because we need to get to the "else" clause from multiple places then.

I do think that the for_each_bitmapped_object() interface is making this a bit harder. Before it was added, the main bitmap entry point was always prepare_bitmap_walk(), which opened the bitmap file itself (and only after doing the cheap can_filter_bitmap() check).

But here that doesn't quite work, because we need the bitmap_index to persist after the for_each_bitmapped_object() call so that we can check bitmap_index_contains_pack() on it.

I was tempted to suggest that for_each_bitmapped_object() should return the bitmap_index itself, and then this code would become:

   if (filter.choice != LOFC_DISABLED)
	   bitmap = for_each_bitmapped_object(filter, cb, &payload);
   if (bitmap) {
	   /* we iterated those objects; check for other packs */
   } else {
	   /* we did nothing; look at all packs */
   }

   free_bitmap_index(bitmap);

which is not too bad. Mostly I wanted to make the fix as small as possible, but I was also a little hesitant to tweak the API when we have only one caller (and we don't know what a second caller might want). But we could always revisit it on top.

-Peff