All current callers already pass in the repo root.
This normalizes things a bit. Now all repo types take a direct path to
repo root and only the command line interface (or the dispatcher)
searches from the current directory.
Right now, there are some situations in which localrepo.filecommit can
create filelog entries even though they're not needed. For example:
- permissions for a file have changed;
- qrefresh can create a filelog entry identical to its parent (see the
added test);
- convert-repo creates extra filelog entries in every merge where the
first parent has added files (for example, changeset ebebe9577a1a of
the kernel repo added extra filelog entries to files in the
arch/blackfin directory, even though the merge should only touch the
drivers/ata directory). This makes "hg log file" in a converted repo
less useful than it could be, since it may mention many merges that
don't actually touch that specific file.
They all come from the same basic problem: localrepo.commit (through
filecommit) creates new filelog entries for all files passed to it
(except for some cases during a merge).
Patch and test case provided by Benoit.
This should fix issue351.
These changes don't actually need a new filelog entry (see next patch),
but if we don't mention the files in the changelog, it becomes much
harder to find all changesets that touch a file (not even a
"hg log --removed file" will work).
Every time util.pathto is called, we have to pass the repo root and the
repo cwd.
dirstate.pathto is a simple convenience function that knows about the
root and the cwd arguments. It's still possible to pass the cwd as an
optimization.
localrepo.pathto is a convenience function that just calls
dirstate.pathto, just like localrepo.getcwd.
dirstate.pathto becomes a single point that converts most (all?) paths
from the internal representation to some OS-specific relative path for
display purposes.
Rationale for not aborting instead:
The first parent is usually more important as it is the local branch
during a merge and commands like 'hg diff' and 'hg diff -r.' behave still
identically (except for the warning of course).
Added a test for log -r. with one and two parents.
- make appending code proper part of changelog with delayupdate/finalize
- use simplified appender that tracks pending data in memory
- eliminate old appendfile and helper classes
- update addchangegroup to use new interface and reuse the existing changelog
