git patches may require copies to be handled out-of-order. For instance, take
the following sequence:
* modify a
* copy a into b
Here, we have to generate b from a before its modification. To do so,
applydiff() was scanning for copy metadata and performing the copies before
processing the other changes in-order. While smart and efficient, this approach
complicates things by handling file copies and file creations at different
places and times. While a new file must not exist before being patched a copied
file already exists before applying the first hunk.
Instead of copying the files at their final destination before patching, we
store them in a temporary file location and retrieve them when patching. The
filestore always stores file content in real files but nothing prevents adding
a cache layer. The filestore class was kept separate from fsbackend for at
least two reasons:
- This class is likely to be reused as a temporary result store for a future
repository patching call (entries just have to be extended to contain copy
sources).
- Delegating this role to backends might be more efficient in a repository
backend case: the source files are already available in the repository itself
and do not need to be copied again. It also means that third-parties backend
would have to implement two other methods. If we ever decide to merge the
filestore feature into backend, a minimalistic approach would be to compose
with filestore directly. Keep in mind this copy overhead only applies for
copy/rename sources, and may even be reduced to copy sources which have to
handled ahead of time.
The patcher has to know if a file is being created or removed to check if the
target already exists, or to actually unlink the file when a hunk emptying it
is applied. This was done by embedding the creation/removal information in the
first (and only) hunk attached to the file.
There are two problems with this approach:
- creation/removal is really a property of the file being patched and not its
hunk.
- for regular patches, file creation cannot be deduced at parsing time: there
are case where the *stripped* file paths must be compared. Modifying hunks
after their creation is clumsy and prevent further refactorings related to
copies handling.
Instead, we delegate this job to selectfile() which has all the relevant
information, and remove the hunk createfile() and rmfile() methods.
Restore the previous diffstat behaviour of scaling by the maximum number of
changes to a single file. Changeset 7bb0e22a7988 modified the diffstat to be
scaled by the total number of changes. This seems to have been unintentional.
gitpatch objects emitted by iterhunks() were referencing file paths unmodified
from the input patch. _applydif() made them usable by modifying the gitpatch
objects in-place with specified path strip level. The same modified objects
were then reused by iterhunks() generator. _applydiff() now copies and update
the paths which completely decouples both routines.
As a side effect, the "git" event now receives only metadata about
copies/renames to perform the necessary copies ahead of time. Other actions are
handled in the "file" event.
Patch changes are emitted by iterhunks() in two separate events: 'file' when
hunks have to be applied and 'git' to describe other modifications like copies
or mode changes. Note that a file which mode is changed and which content is
modified by the same patch will be emitted in both events. It is more
convenient to handle all file modifications in a single event. This patch
"zips" git actions with regular changes so both kinds can be emitted at the
same place.
git afile/bfile are extracted twice, once when reading a 'diff --git', and
again when reading a unified hunk. The problem is not all git blocks have
unified hunks (renames just have metadata) and they were not extracted the same
way. This is what this patch unifies.
gitpatch objects emitted by iterhunks() are modified in place by applydiff().
Processing them earlier improves iterhunks() isolation. applydiff() modifying
them should still be fixed though.
_updatedir() is no longer used by internalpatch()
The change in test-mq-missingfiles.t comes from workingbackend not considering
the missing 'b' file as changed, thus not calling addremove() on it.
This introduces a performance regression for large files, as they will be
copied just to be clobbered afterwards since binary patching does not use
deltas. But it simplifies the code and the previous optimization will be
reintroduced later in a better way.
_applydiff() patcher argument was added to help hgsubversion like extension
monkeypatching the patching process. While it could be removed at this point, I
prefer to leave it until patch.py is completely refactored and there is a valid
and tested alternative.
Most filesystem calls are already isolated in patchfile but this is not enough:
renames are performed before patchfile is available and some chmod calls are
even done outside of the applydiff call. Once all these calls are extracted
into a backend class, we can provide cleaner APIs to write to a working
directory context directly into the repository.
The local err variable would be bound to PatchError thrown and it
keeps this value even after the except block is executed.
The whole thing worked anyway since the rejected variable is set in
the except block and this makes the function return -1 when a
PatchError is thrown.
Why?
- Mercurial internal patcher works correctly for regular patches and git
patches, is much faster at least on Windows and is more extensible.
- In theory, the external patcher can be used to handle exotic patch formats. I
do not know any and have not heard about any such use in years.
- Most patch programs cannot handle git format patches, which makes the API
caller to decide either to ignore ui.patch by calling patch.internalpatch()
directly, or take the risk of random failures with valid inputs.
- One thing a patch program could do Mercurial patcher cannot is applying with
--reverse. Apparently several shelve like extensions try to use that,
including passing the "reverse" option to Mercurial patcher, which has been
removed mid-2009. I never heard anybody complain about that, and would prefer
reimplementing it anyway.
And from the technical perspective:
- The external patcher makes everything harder to maintain and implement. EOL
normalization is not implemented, and I would bet file renames, if supported
by the patcher, are not correctly recorded in the dirstate.
- No tests.
How?
- Remove related documentation
- Clearly mark patch.externalpatch() as private
- Remove the debuginstall check. This deprecation request was actually
triggered by this last point. debuginstall is the only piece of code patching
without a repository. When migrating to an integrated patch() + updatedir()
call, this was really a showstopper, all workarounds were either ugly or
uselessly complicated to implement. If we do not support external patcher
anymore, the debuginstall check is not useful anymore.
- Remove patch.externalpatch() after 1.9 release.
