This fixes up a number of problems flagged by valgrind and also
cleans up the internal `git_submodule` allocation handling
overall with a simpler model.
* `git_buf_rfind` (with tests and tests for `git_buf_rfind_next`)
* `git_buf_puts_escaped` and `git_buf_puts_escaped_regex` (with tests)
to copy strings into a buffer while injecting an escape sequence
(e.g. '\') in front of particular characters.
GProf shows `git_text_gather_stats` as the most expensive call
in large diffs. The function calculates a lot of information
that is not actually used and does not do so in a optimal
order. This introduces a tuned `git_buf_is_binary` function
that executes the same algorithm in a fraction of the time.
The goal of this work is to rewrite git_status_file to use the
same underlying code as git_status_foreach.
This is done in 3 phases:
1. Extend iterators to allow ranged iteration with start and
end prefixes for the range of file names to be covered.
2. Improve diff so that when there is a pathspec and there is
a common non-wildcard prefix of the pathspec, it will use
ranged iterators to minimize excess iteration.
3. Rewrite git_status_file to call git_status_foreach_ext
with a pathspec that covers just the one file being checked.
Since ranged iterators underlie the status & diff implementation,
this is actually fairly efficient. The workdir iterator does
end up loading the contents of all the directories down to the
single file, which should ideally be avoided, but it is pretty
good.
This adds a `git_pool` object that can do simple paged memory
allocation with free for the entire pool at once. Using this,
you can replace many small allocations with large blocks that
can then cheaply be doled out in small pieces. This is best
used when you plan to free the small blocks all at once - for
example, if they represent the parsed state from a file or data
stream that are either all kept or all discarded.
There are two real patterns of usage for `git_pools`: either
for "string" allocation, where the item size is a single byte
and you end up just packing the allocations in together, or for
"fixed size" allocation where you are allocating a large object
(e.g. a `git_oid`) and you generally just allocation single
objects that can be tightly packed. Of course, you can use it
for other things, but those two cases are the easiest.
There was a bug in git_buf_join_n when the contents of the
original buffer were joined into itself and the realloc
moved the pointer to the original buffer.
This converts blob.c, fileops.c, and all of the win32 files.
Also, various minor cleanups throughout the code. Plus, in
testing the win32 build, I cleaned up a bunch (although not
all) of the warnings with the 64-bit build.
This also includes droping `git_buf_lasterror` because it makes no sense
in the new system. Note that in most of the places were it has been
dropped, the code needs cleanup. I.e. GIT_ENOMEM is going away, so
instead it should return a generic `-1` and obviously not throw
anything.
Initial implementation. The relevant code is in `blob.c`: the blob write
function has been split into smaller functions.
- Directly write a file to the ODB in streaming mode
- Directly write a symlink to the ODB in direct mode
- Apply a filter, and write a file to the ODB in direct mode
When trying to write a file, we first call `git_filter__load_for_file`,
which populates a filters array with the required filters based on the
filename.
If no filters are resolved to the filename, we can write to the ODB in
streaming mode straight from disk. Otherwise, we load the whole file in
memory and use double-buffering to apply the filter chain. We finish
by writing the file as a whole to the ODB.
This makes so much sense that I can't believe it hasn't been done
before. Kill the old `git_fbuffer` and read files straight into
`git_buf` objects.
Also: In order to fully support 4GB files in 32-bit systems, the
`git_buf` implementation has been changed from using `ssize_t` for
storage and storing negative values on allocation failure, to using
`size_t` and changing the buffer pointer to a magical pointer on
allocation failure.
Hopefully this won't break anything.
This create a new git_iterator type of object that provides a
uniform interface for iterating over the index, an arbitrary
tree, or the working directory of a repository.
As part of this, git ignore support was extended to support
push and pop of directory-based ignore files as the working
directory is being traversed (so the array of ignores does
not have to be recreated at each directory during traveral).
There are a number of other small utility functions in buffer,
path, vector, and fileops that are included in this patch
that made the iterator implementation cleaner.
This contains fixes for several issues discovered by MSVC and
by valgrind, including some bad data access, some memory
leakage (in where certain files were not being successfully
added to the cache), and some code simplification.
Adds support for .gitignore files to git_status_foreach() and
git_status_file(). This includes refactoring the gitattributes
code to share logic where possible. The GIT_STATUS_IGNORED flag
will now be passed in for files that are ignored (provided they
are not already in the index or the head of repo).
It was not safe for git_buf_joinpath to be used with a pointer
into the buf itself because a reallocation could invalidate
the input parameter that pointed into the buffer. This patch
makes it safe to self join, at least for the leading input to
the join, which is the common "append" case for self joins.
Also added unit tests to explicitly cover this case.
This should actually fix#511
This converts virtually all of the places that allocate GIT_PATH_MAX
buffers on the stack for manipulating paths to use git_buf objects
instead. The patch is pretty careful not to touch the public API
for libgit2, so there are a few places that still use GIT_PATH_MAX.
This extends and changes some details of the git_buf implementation
to add a couple of extra functions and to make error handling easier.
This includes serious alterations to all the path.c functions, and
several of the fileops.c ones, too. Also, there are a number of new
functions that parallel existing ones except that use a git_buf
instead of a stack-based buffer (such as git_config_find_global_r
that exists alongsize git_config_find_global).
This also modifies the win32 version of p_realpath to allocate whatever
buffer size is needed to accommodate the realpath instead of hardcoding
a GIT_PATH_MAX limit, but that change needs to be tested still.
This streamlines git_buf_join and removes the join-append behavior,
opting instead for a very compact join-replace of the git_buf contents.
The unit tests had to be updated to remove the join-append tests and
have a bunch more exhaustive tests added.
Taking a page from core git's strbuf, this introduces git_buf_initbuf
which is an empty string that is used to initialize the git_buf ptr
value even for new buffers. Now the git_buf ptr will always point to
a valid NUL-terminated string.
This change required jumping through a few hoops for git_buf_grow
and git_buf_free to distinguish between a actual allocated buffer
and the global initial value. Also, this moves the allocation
related functions to be next to each other near the top of buffer.c.
At a tiny cost of 1 extra byte per allocation, this makes
git_buf_cstr into basically a noop, which simplifies error
checking when trying to convert things to use dynamic allocation.
This patch also adds a new function (git_buf_copy_cstr) for copying
the cstr data directly into an external buffer.
This commit addresses two of the comments:
* renamed existing n-input git_buf_join to git_buf_join_n
* added new git_buf_join that always takes two inputs
* moved some parameter error checking to asserts
* extended unit tests to cover new version of git_buf_join
Add new functions to git_buf for:
* initializing a buffer from a string
* joining one or more strings onto a buffer with separators
* swapping two buffers in place
* extracting data from a git_buf (leaving it empty)
Also, make git_buf_free leave a git_buf back in its initted state,
and slightly tweak buffer allocation sizes and thresholds.
Finally, port unit tests to clay and extend with lots of new tests
for the various git_buf functions.
1. The license header is technically not valid if it doesn't have a
copyright signature.
2. The COPYING file has been updated with the different licenses used in
the project.
3. The full GPLv2 header in each file annoys me.
DIRECT WRITES ARE BACK AND FASTER THAN EVER. The streaming writer to the
ODB was an overkill for the smaller objects like Commit and Tags; most
of the streaming logic was taking too long.
This commit makes Commits, Tags and Trees to be built-up in memory, and
then written to disk in 2 pushes (header + data), instead of streaming
everything.
This is *always* faster, even for big files (since the git_filebuf class
still does streaming writes when the memory cache overflows). This is
also a gazillion lines of code smaller, because we don't have to
precompute the final size of the object before starting the stream (this
was kind of defeating the point of streaming, anyway).
Blobs are still written with full streaming instead of loading them in
memory, since this is still the fastest way.
A new `git_buf` class has been added. It's missing some features, but
it'll get there.
