Hi,
This is my draft for my proposal on "Convert Scripts to builtins" for GSoC.
Please review and provide feedback.
Cheers,
Pratik Karki
Convert Scripts to builtins
===========================
Abstract
--------
Many components of Git are still in the form of shell and Perl scripts.
This has certain advantages of being extensible but causes problems in
production code on multiple platforms like Windows.
I propose to rewrite a couple of shell and perl scripts into portable
and performant C code, making them built-ins. The major advantage of
doing this is improvement in efficiency and performance.
Much more scripts like `git-am` , `git-pull`, `git-branch` have already
been rewritten in C. Much more scripts like `git-rebase`, `git-stash`,
`git-add --interactive` are still present in shell and perl scripts.
I propose to work in `git-rebase` and `git-stash`.
Shell Scripts:
--------------
Although shell scripts are more faster can be extensible in
functionality and can be more easier to write, they introduce certain
disadvantages.
1. Dependencies:
The scripting languages and shell scripts make more productive code
but there is an overhead of dependencies. The shell scripts are
lighter and simpler and call other executables to perform
non-trivial tasks. Taking `git-stash` shell script for example.
`sed`, `rm`, `echo`, `test` are constantly present in `git-stash`.
These look common to POSIX platforms but for non-POSIX platforms
there needs some extra work for porting these commands. For example,
in Git for Windows, the workaround for these commands in non-POSIX
platform adds some extra utilities and adds MSYS2 shell commands and
needs to pack language runtime for Perl. This increases the
installation size and requires more disk space. Again, adding more
batteries again needs implementation in all of the dependency
libraries and executables.
2. Inefficiency:
Git has internal caches for configuration values, the repository
index and repository objects. The porcelain commands do not have
access to git's internal API and so they spawn git processes to
perform git operations. For every git invocation, git would re-read
the user's configuration files, repository index, repopulate the
filesystem cache, etc. This leads to overhead and unnecessary I/O.
Windows is known to have worse I/O performance compared to Linux.
There is also slower I/O performance of HDD compared to SSD. This
unnecessary I/O operations causes runtime overhead and becomes
slower in poor I/O performance setups. Now, writing the porcelain
into C built-ins leverages the git API and there is no need of
spawning separate git processes, caching can be used to reduce
unnecessary I/O processes.
3. Spawing processes is less performant:
Shell scripts usually spawn a lot of processes. Shell scripts are
very lighter and hence have limited functionalites. For
`git-stash.sh` to work it needs to perform lots of git operations
like `git rev-parse` `git config` and thus spawns git executable
processes for performing these operations. Again for invoking
`git config` and providing configuration values, it spawn new
processes to handle that. Spawning is implemented by `fork()` and
`exec()` by shells. Now, on systems that do not support
copy-on-write semantics for `fork()`, there is duplication of the
memory of the parent process for every `fork()` call which turns out
to be an expensive process. Now, in Windows, Git uses MSYS2
exclusively to emulate `fork()` but since, Windows doesnot support
forking semantics natively, the workaround provided by MSYS2
emulates `fork()` without [copy-on-write
semantics](https://www.cygwin.com/faq.html#faq.api.fork). Doing this
creates another layer over Windows processes and thus slows git.
Rewriting C built-ins
---------------------
These above mentioned problems need to be fixed. The only fix for these
problems would be to write built-ins in C for all these shell scripts
leveraging the git API. Writing in built-in reduces the dependency
required by shell scripts. Since, Git is native executable in Windows,
doing this can make MSYS2 POSIX emulation obsolete. Then, using git's
internal API and C data types, built-in `git_config_get_value()` can be
used to get configuration value rather than spawning another git-config
process. This removes the necessary to re-read git configuration cache
everytime and reduces I/O. Furthermore, git-stash and git-rebase will be
more faster and show consistent behaviour as instead of spawing another
process and parsing command-line arguments manually, they can be
hardcoded to be built-in and leverage all the required git's internal
API's like `parse-options`.
To implement git-stash and git-rebase in C, I propose to avoid spawning
lots of external git processes and reduce redundant I/O by taking
advantage of the internal object, index and configuration cache. I
propose to use C data structures instead of needless parsing as C has
richer and performant data structures. Windows will definitely benefit
from these rewrites as the use of MSYS2 POSIX emulation would be
obsolete and unnecessary I/O processes would be reduced and there won't
be need of spawning external git processes.
But, while rewriting regressions should be totally avoided. The code
should not cause regression and introduce bugs not present in previous
scripts. Thought the re-write wouldn't look more simpler like previous
scripts, it should not have different behaviour or bugs compared to the
script and should not be hard to maintain compared to the script.
Potential Problems
------------------
### New Re-write code becomes less simpler and Bugs could get introduced:
C is a more verbose language than shell script and looks less simpler
than shell scripts. Doing things in less than 4 lines of shell script in
C can take up more than 12 lines to make it more simpler. For example,
in a previously used script `git-commit.sh` we can see,
```bash
run_status () {
# If TMP_INDEX is defined, that means we are doing
# "--only" partial commit, and that index file is used
# to build the tree for the commit. Otherwise, if
# NEXT_INDEX exists, that is the index file used to
# make the commit. Otherwise we are using as-is commit
# so the regular index file is what we use to compare.
if test '' != "$TMP_INDEX"
then
GIT_INDEX_FILE="$TMP_INDEX"
export GIT_INDEX_FILE
elif test -f "$NEXT_INDEX"
then
GIT_INDEX_FILE="$NEXT_INDEX"
export GIT_INDEX_FILE
fi
if test "$status_only" = "t" -o "$use_status_color" = "t"; then
color=
else
color=--nocolor
fi
git runstatus ${color} \
${verbose:+--verbose} \
${amend:+--amend} \
${untracked_files:+--untracked}
}
```
This script has been rewritten as builtin in C.
These lines of code have been replaced in C providing more
built-in performance but looks less simpler and
contains high multiline functions in `commit.c`.
```C
`static int run_status(FILE *fp, const char *index_file, const char *prefix,
int nowarn,
struct wt_status *s)
{
struct object_id oid;
if (s->relative_paths)
s->prefix = prefix;
if (amend) {
s->amend = 1;
s->reference = "HEAD^1";
}
s->verbose = verbose;
s->index_file = index_file;
s->fp = fp;
s->nowarn = nowarn;
s->is_initial = get_oid(s->reference, &oid) ? 1 : 0;
if (!s->is_initial)
hashcpy(s->sha1_commit, oid.hash);
s->status_format = status_format;
s->ignore_submodule_arg = ignore_submodule_arg;
wt_status_collect(s);
wt_status_print(s);
return s->commitable;
}
```
We can see the C version requires more lines of code and seem complex
than the script. But, this C code is not depending on external libraries
and dependencies hence provides good performance and can be optimized
further. The C codes after re-writing can be further documented,
optimized and designed to make it more simpler for maintaining it. The
functions can be documented with well-defined inputs, outputs and
behaviour. The internal git API can be used and the recurring codes can
be modified which can shrink the codebase and make git codebase smaller
and resulting in smaller installation size in different platforms.
Now, while doing this re-write there is high possibility of introducing
bugs. Though, the test-suite can catch obvious bugs, other less-obvious
bugs can escape from test scenarios. To fix this problem, new tests with
the co-ordination of other maintainers and contributors should be
written for the cases where the test fails to catch bugs. Other code
coverage tools can also be used to ensure that the test suite for the
command tests all code in builtin. Though it may take more time, careful
rewriting and review of code must be done seriously by doing one to one
line translations of shell scripts to C. The re-written code after tests
must go through rigorous review process to find and squash bugs present
in it.
> "Given enough eyeballs, all bugs are shallow."
> - Linus Torvalds
Timeline and Development Cycle
------------------------------
- Apr 23: Accepted student proposals announced.
- Apr 23 onwards: Researching of all the test suites. Discussion of
possible test improvements in for `git-stash` and `git-rebase`.
- May 1: Rewriting skeleton begins.
- May 13: Making `builtin/stash.c` ready for review process. (This
goes on for some time.)
- May 26: Making `builtin/rebase.c` ready for review process. (This
goes on for some time.)
- June 10: Make second versions with more improvements and more
batteries ready for next review cycle.
- June 20: Writing new tests and using more code-coverage tools to
squash bugs present.
- June 25 - Jul 20: Start optimizing `builtin/stash.c` and
`builtin/rebase.c`. Benchmarking and profiling is done. They are
exclusively compared to their original shell scripts, to check
whether they are more performant or not and the results are
published in the mailing list for further discussion.
- Jul 20 - Aug 5: More optimizing and polishing of `builtin/stash.c`
and `builtin/rebase.c` and running of new tests series written and
send them for code review.
- Aug 14: Submit final patches.
- Aug 22: Results announced.
- Apr 24 - Aug 14: Documentation is written. "What I'm working on" is
written and posted in my blog regarding GSoC with Git.
About me
--------
I'm Pratik Karki and am studing bachelors in Computer Engineering in
Advanced College of Engineering and Management (Affiliated to Institue
of Engineering, Tribhuvan University). I've been writing C, C++, Ruby,
Perl, Clojure, Lisp, Java, JS, Erlang, Rust for 3 years, and contributed
to some projects which can be seen in
[Github](https://github.com/prertik). I've been doing independent
contract works for small upcoming start-ups in Nepal.
I have been planning to contribute in Git for a long time. Thanks, to
GSoC I've submitted small patch as a microproject: [test: avoid pipes in
git related commands for test
suite](https://public-inbox.org/git/20180319173204.31952-1-predatoram...@gmail.com/).
I am looking forward to submitting more patches to Git on a long time
basis.
