Re: GSoC Git Proposal Draft - ZheNing Hu

From: Jeff King <peff@peff.net>
Date: Apr 2, 2021, 15:39 UTC
Message-ID: <YGc6ybE1wD1ck0uB@coredump.intra.peff.net>
In-Reply-To: <CAOLTT8RfE4nn5NnjZh7xuF09-5=+K+_j_2kP0327HVdR4x_wAQ@mail.gmail.com>

On Fri, Apr 02, 2021 at 05:03:17PM +0800, ZheNing Hu wrote:

Show 18 quoted lines

> * Because part of the feature of `git for-each-ref` is very similar to
> that of `git cat-file`, I think `git cat-file` can learn some feasible
> solutions from it.
> 
> #### My possible solutions:
> 
> 1. Same [solution](https://github.com/git/git/pull/568/commits/cc40c464e813fc7a6bd93a01661646114d694d76)
> as Olga, add member `struct ref_format format` in `struct
> batch_options`.
> 2. Use the function
> [`verify_ref_format()`](https://github.com/gitgitgadget/git/blob/84d06cdc06389ae7c462434cb7b1db0980f63860/ref-filter.c#L904)
> to replace the first `expand_format()` for parsing format strings.
> 3. Write a function like
> [`format_ref_array_item()`](https://github.com/gitgitgadget/git/blob/84d06cdc06389ae7c462434cb7b1db0980f63860/ref-filter.c#L2392),
> get information about objects, and use `get_object()` to grub the
> information which we prefer (or just use `grab_common_value()`).
> 4. The migration of `%(rest)` may require learning the handling of
> `%(if)` ,`%(else)`.

I think one thing to keep an eye on here is the performance of cat-file. The formatting code used by for-each-ref is rather slow (it may load more of the object details than is necessary, it is too eager to allocate intermediate strings, and so on). That's usually not _too_ big a problem for ref-filter, because the number of refs tends to be much smaller than the number of total objects. But I'd expect that moving to the ref-filter code would make something like:

 git cat-file --batch-all-objects --batch-check='%(objectname) %(objecttype)'

measurably slower.

IMHO the right path forward is not to try porting cat-file to use the ref-filter code, but to start first with writing a universal formatting module that takes the best of both implementations (and the commit pretty-printer):

  - separate the format-parsing stage from formatting actual items, as
    ref-filter does. This lets us have more complex formats without
    paying a per-item runtime cost while formatting. This should also
    allow us to handle multiple syntaxes for the same thing (e.g.,
    ref-filter %(authorname) vs pretty.c %an).

  - figure out which data will be needed for each item based on the
    parsed format, and then do the minimum amount of work to get that
    data (using "oid_object_info_extended()" helps here, because it
    likewise tries to do as little work as possible to satisfy the
    request, but there are many elements that it doesn't know about)

  - likewise avoid doing any intermediate work we can; as much as
    possible, format the result directly into a result strbuf, rather
    than allocating many sub-strings and assembling them (as cat-file
    does).

  - handle formats where the necessary item data may or may not be
    present. E.g., if we're given a refname, then "%(refname)" makes
    sense. But in cat-file we'd not have a refname, and just an object.
    We should still be able to use the same formatting code to handle
    "%(objecttype)", etc. Likewise for formats which require a specific
    type (say %(authorname) for a commit, but the object is a blob).
    Ref-filter does this to some degree for things like authorname, but
    we'd be extending it to the case that we don't even have a refname.

-Peff