The subprocess module in Python 2.6 has a CalledProcessError class,
but its constructor doesn't take a third (output) argument. So don't
only pass the constructor two arguments.
Handle the automatic merges at the vertex of a block more carefully,
by merging them two ways and verifying that the two resulting trees
are identical.
When autofilling the vertex of a block, it is possible to do the merge
three ways:
1. Using the commits directly above and to the left as parents
2. Using the commits at [i1,0] and [i1-1,i2] as parents (i.e., as a
continuation of the bottom edge)
3. Using the commits at [i1,i2-1] and [0,i2] as parents (i.e., as a
continuation of the right edge)
We want the final history to be *as if* the parents were as in (1).
Formerly we were doing (1) and simply using the result. This is
incorrect because the neighboring commits do not have correct
histories. Therefore, the merge base for this merge is [0,0], and the
merge can have unnecessary conflicts.
Change to trying both (2) and (3). If both of those commits succeed
*and* if they give identical trees, then create a synthetic commit
with the parents as in (1) and use that; otherwise, fail.
MergeFrontier.remove_failure() was not doing the right thing. If
there is a conflict like this:
*******
*......
*.....X
*......
then it should attempt to split the conflicting block into two blocks
(one of which might be empty) like this:
*****||
-----++
*....|X
-----+.
but instead it is doing this:
*****|*
*----+.
*.....X
*......
which is nonsense. This could lead to an infinite loop (GitHub issue
Some environments might have a old and/or broken version of python2
installed in /usr/bin but still wish to use git-imerge. The non-broken
version of python might be loaded separately from the base operating
system image via for example environment modules [1].
So use the currently active python2 from the environment instead of
hard coding the path to /usr/bin/python2.
[1] http://modules.sourceforge.net/
These files are in the public domain, except for iepngfix.htc, which
is under LGPL 2.1+ (see notice in file).
This command adds the pristine versions of the files, so that the
small changes that I made in the files in the next commit will be
documented.
The test scripts used to assume that they were being run from a
strange place. Instead, teach them to derive the main project
directory "BASE" from $0 and
* always place the temporary directory at $BASE/t/tmp
* always run git-imerge from $BASE/git-imerge
* be more careful before deleting an old test directory.
Keep track of which commits block frontier blocker blocks. (Hmmm,
maybe the word "block" is being overused...)
This makes it possible to preserve a persistent record of which of the
blocks on the current frontier have no hope of progressing without
user interaction, and allow us to avoid some checks at every program
start.
I also think that with some extra code we could use an *unblocked*
blocker when bisecting and outlining and thereby avoid the need to
outline unblocked blocks as two blocks in many cases; i.e., instead of
filling such a block like this:
*******
*@.....
*?????.
*?????.
*......
we could fill it like this:
*******
*@????.
*?????.
*?????.
*......
and thereby avoid a lot of bookkeeping commits.
Rename MergeFrontier.add_success() to get_affected_blocker_block() and
change it to return the block that was freed up by the new merge. If
there is no such merge, raise an exception.
Generate diagram output in two stages:
* Write one string for each commit to a 2D array
* Write the array to output
Then give access to the former functionality via new fill_diagram()
methods.
