For most real use cases, repositories with alternates use them as main object storage. Checking the alternate for objects before the main repository should result in measurable speedups. Because of this, we're changing the sorting algorithm to prioritize alternates *in cases where two backends have the same priority*. This means that the pack backend for the alternate will be checked before the pack backend for the main repository *but* both of them will be checked before any loose backends.
In the current implementation of ODB backends, each backend is tasked with refreshing itself after a failed lookup. This is standard Git behavior: we want to e.g. reload the packfiles on disk in case they have changed and that's the reason we can't find the object we're looking for. This behavior, however, becomes pathological in repositories where multiple alternates have been loaded. Given that each alternate counts as a separate backend, a miss in the main repository (which can potentially be very frequent in cases where object storage comes from the alternate) will result in refreshing all its packfiles before we move on to the alternate backend where the object will most likely be found. To fix this, the code in `odb.c` has been refactored as to perform the refresh of all the backends externally, once we've verified that the object is nowhere to be found. If the refresh is successful, we then perform the lookup sequentially through all the backends, skipping the ones that we know for sure weren't refreshed (because they have no refresh API). The on-disk pack backend has been adjusted accordingly: it no longer performs refreshes internally.
As refdb and odb backends can be allocated by client code, libgit2 can’t know whether an alternative memory allocator was used, and thus should not try to call `git__free` on those objects. Instead, odb and refdb backend implementations must always provide their own `free` functions to ensure memory gets freed correctly.
odb: make the writestream's size a git_off_t
We currently first look in the loose object dir and then in the packs for objects. When performing operations on recent history this has a higher likelihood of hitting, but when we deal with operations which look further back into the past, we start spending a large amount of time getting ENOTENT from `access`. Reversing the priorities means that long-running operations can get to their objects faster, as we can look at the index data we have in memory (or rather mapped) to figure out whether we have an object, which is faster than going out to the filesystem. The packed backend already implements an optimistic read algorithm by first looking at the packs we know about and only going out to disk to referesh if the object is not found which means that in the case where we do have the object (which will be in the majority for anything that traverses the graph) we can avoid going to to disk entirely to determine whether an object exists. Operations which look at recent history may take a slight impact, but these would be operations which look a lot less at object and thus take less time regardless.
Restricting files to size_t is a silly limitation. The loose backend writes to a file directly, so there is no issue in using 63 bits for the size. We still assume that the header is going to fit in 64 bytes, which does mean quite a bit smaller files due to the run-length encoding, but it's still a much larger size than you would want Git to handle.
Make our overflow checking look more like gcc and clang's, so that we can substitute it out with the compiler instrinsics on platforms that support it. This means dropping the ability to pass `NULL` as an out parameter. As a result, the macros also get updated to reflect this as well.