If applydiff() raises an exception, the opened patch file is kept alive in the
exception context. If it is a temporary file (for instance supplied by import
command with stdin input), Windows cannot clean it up.
The built-in patch implementation applied the hunks to the wrong lines of the
file if the file in the repo has been modified to skew the patch line numbers
and the file contains repetitive sequences of lines.
Most of the time, one can reverse a diff by swapping the revisions passed with
-r but it happens that if you use the global -R, and diff against the tip of
the current repo, you can't swap the revisions. One use-case for that is
reviewing changes from a bundle before unbundling. One could also pipe the
output of `hg diff` to a command line filter that reverses the diff, but that
would remove the benefit from color diffs. Therefore, having an option in
`hg diff` to reverse a diff is a good thing.
The option flag selection was tricky. GNU patch uses -R/--reverse but -R is
already used as a global option and --reverse would make --rev ambiguous.
Normally, diffs without any text insertions or deletions are reported
as having 0 lines changed by stock diffstat. Compatibility is
preserved with stock diffstat in this case, but when using --git,
binary files are marked with Bin as a means of clarification.
git diff --stat does something similar, though it also includes the
old and new file sizes.
Symlink creations and deletions were handled with a special symlinkhunk object,
working like a binary hunk. However, this model does not support symlink
updates or replacements, so we teach regular hunks how to handle symlinks.
The previous method of scaling had a tendency to include graph lines that
went past the output width when the file with the most changes had a very
large number of changes.
The intent is to fix many issues involving patching when win32ext is enabled.
With win32ext, the working directory and repository files EOLs are not the same
which means that patches made on a non-win32ext host do not apply cleanly
because of EOLs discrepancies. A theorically correct approach would be
transform either the patched file or the patch content with the
encoding/decoding filters used by win32ext. This solution is tricky to
implement and invasive, instead we prefer to address the win32ext case, by
offering a way to ignore input EOLs when patching and rewriting them when
saving the patched result.
In worst case, generating diff in upgrade mode can be two times more expensive
than generating it in git mode directly: we may have to regenerate the whole
diff again whenever a git feature is detected. Also, the first diff attempt is
completely buffered instead of being streamed. That said, even without having
profiled it yet, I am convinced we can fast-path the upgrade mode if necessary
were it to be used in regular diff commands, and not only in mq where avoiding
data loss is worth the price.
With eolmode set to 'lf' or 'crlf' we avoided the hunk duplication and
normalization by reading the input patch in text mode. Dropping this
optimization simplifies code expectations for a small overhead.
The change in test-mq-eol comes from a tolerance to CRLF instead of LF for last
lines without newlines being broken by this revision. This tolerance was only
partially supported and will be added again in a better way.
EOLs in patched files are restored to their original value after
patching. We use the first EOL found in the file, files with
inconsistent EOLs will thus be normalized during this process.
The old code mapped the value of eolmode ('strict', 'crlf' or 'lf') to
eol (None, '\r\n' or '\n') at the entry point in internalpatch. The
value of eol was then used directly as the desired EOL in patchfile.
We now delay the mapping and let patchfile do it instead. This allows
for more complicated behavior where it does not make sense to map
eolmode directly to the target EOLs.
The built-in None object is a singleton and it is therefore safe to
compare memory addresses with is. It is also faster, how much depends
on the object being compared. For a simple type like str I get:
| s = "foo" | s = None
----------+-----------+----------
s == None | 0.25 usec | 0.21 usec
s is None | 0.17 usec | 0.17 usec
The context variable is either True, False or None. Abbreviate it as C
and we get the following truth table where the second column is the
original expression and the third column is the new expression:
C | C or C == None | C is not False
True | True | True
False | False | False
None | True | True
