-Basic update algorithm:
-
- 1. recurse to all of our direct dependencies and
- update their bases and tips
-
- 2. Update our base.
-
- i. Compute our base's set of desired included deps:
- The set of desired included deps for a base
- is the union of the desired included deps for each
- branch named in the base's branch's direct deps, plus
- the name of every direct external dep.
- The set of desired included deps for a branch
- is the set of desired included deps for the branch's
- base plus the branch name itself.
-
- ii. For each source in the best order, do the following merge:
- (Our base has sources:
- - the branch for each direct dep
- - the remote base)
-
- Find the common ancestor.
-
- Check for unwanted dependency removals.
- An unwanted dependency removal is
- A branch in the desired included deps
- Which exists in the common ancestor's actual included deps
- but which is missing in the source's actual included deps
- (NB that this definition excludes dependency removals
- which have already occurred on our base; these will be
- reverted later.)
- For each unwanted dependency removal (ie for each such
- branch), find the most recent commit which unwantedly removed
- the dep from the source's actual included deps ("relevant
- unwanted removal commit"). (Abort if any such commit is a
- merge.) Select the earliest relevant unwanted removal commit
- (from the set of relevant unwanted removal commits
- corresponding to the unwanted dependency removals).
- Merge from the ancestor of the relevant unwanted removal commit.
- Merge from the relevant unwanted removal commit using -s ours.
-
- (The purpose of this, and the result, is that the unwanted
- dependency removal has gone away. Doing things in this order
- tries to keep the unwanted dependency removal's reversions as
- close as possible to their originating points. The
- recursion, which processes dependencies before their clients,
- tries to keep the reversion churn localised: client patches
- of a patch affected by an unwanted removal will benefit from
- that client's resolution of the situation.)
-
- Now continue to the next unwanted dependency removal.
-
- If there are no (more) unwanted dependency removals, merge
- from the source.
-
- iii.
- Check for missing or unwanted dependency inclusions. Compare
- our base's desired included deps with our base's actual
- included deps. In exceptional conditions, they will not
- be identical. This can happen, for example, because a
- dependency removal was incorporated into our base branch but
- the removed branch was introduced as an explicit dependency.
- This will also happen if we remove a dependency ourselves.
-
- Do the unwanted inclusions first, and then the missing ones.
- (This is because that the missing ones cannot depend on the
- unwanted ones, as otherwise the unwanted ones would be in the
- desired inclusions. So removing the unwanted ones first
- reduces the chances of conflicts.)
-
- So, repeatedly:
- * Do the comparison between desired and actual included
- * Pick a missing inclusion, or failing that an unwanted one
- (call this the "relevant" branch)
- * Depth first search all of the dependencies of the
- relevant branch; if any of these is also a missing
- (resp. unwanted) inclusion, start again processing it
- instead.
- * Attempt to apply the appropriate diff to add (resp. remove)
- the contents of the relevant patch (adjusted appropriately
- for metadata, XXX??? particularly the actual inclusion list)
- * Go round again looking for another discrepancy.
-
- 3. Update our branch.
- Our branch has sources:
- - our base
- - the remote for our branch
- For each source in the best order, do the merge.
-
- Double-check the actual dependency inclusions. In
- particular, if we just upgraded to actual dependency tracking
- we may need to explicitly add our branch name to the actual
- dependency inclusion list.
-
-The "best order" for merges is in order of recency of common
-ancestor, most recent first, and if that does not distinguish,
-merging from local branches first.
-
-"Recency" refers to the order from git-rev-list --date-order.
-
-Actual included deps:
-
- This is tracked explicitly in .topbloke/included, one branch per
- line. For compatibility with older versions, every time we think
- about a base, branch or source above, we check whether
- .topbloke/included is present.
-
- If it isn't then we calculate a child commit which has a
- .topbloke/included. In the case of a remote branch or base, we
- substitute this child commit for the relevant remote ref but do
- not record it in the remote ref; in the case of a local branch or
- base, we advance the local branch or base accordingly.
-
- When .topbloke/included is calculated in this way, it always gets
- the list of desired included deps. (topgit,
- which does not support dependency deletion, always has exactly the
- desired deps actually included.)
-
- Foreign branches cannot be removed from included and cannot
- therefore be removed from dependency lists.
-
-
Patch removal:
- removed patch must be removed from the deps of its
Foreign branches:
When merging from a foreign dependency, check that it
- does not have .topbloke metadata; LATER if it
- does, could produce a new commit which has .topbloke removed
- and merge from that
+ does not have .topbloke metadata
Patch naming:
needs to be globally unique
Operations like "go up the stack", goes towards leaf. Hopefully unique.
"Down" the stack, uses a "conventional" linearisation
Stack reordering op ? auto adjust deps
+
+
+When merging, we need to DTRT with our metadata.
+Do this by running write-tree/read-tree etc. ourselves ?
+For a source we're merging from, we make a version where the
+metadata we shouldn't be merging is removed ?
+Or something.
+Have discovered that specifying a custom merge driver for a file does
+not have any effect if the three-way-merge looks trivial based
+on looking at the file contents - at least, if you use git-merge.
+
+Only thing which knows how to do all the gitattributes stuff and
+conflict markers and what have you is git-merge. (git-merge-tree does
+too but the output format is unsuitable.) "git-merge-index
+... git-merge-one-file" does not honour the merge drivers and is,
+contrary to what the git docs seem to suggest but don't actually
+state, not actualy used by git-merge.
+
+OK so here is a plan:
+ Use git-merge --no-commit
+ Perhaps on a HEAD, and/or against a tree, which have been massaged
+ to make the metadata suitable as input.
+ Filtering out the "merge was successful but we aren't committing"
+ message. Use a single pipe for stdout/stderr to get interleaving
+ right; the message from git-merge is not i18n'd.
+ Afterwards we:
+ Check for merge success in the index and compare to exit status
+ from git-merge (which is 1 if the merge failed).
+ Adjust the metadata.
+ Print appropriate big fat warnings if we have merge conflicts in our
+ metadata.
+ Commit, adjusting the parents of the new commit to the original
+ parents if we made the merge with special massaged parents.
+ We may still need to have custom merge drivers for metadata.
+
+
+Strategies for each metadata file merge:
+
+ in base/tip same patch & branch dep -> base base -> tip
+
+ msg T textual merge rm from src not in src
+ patch BT must be same rm from src must be same
+ base T compute specially rm from src not in src
+ deps B list merge rm from src rm from src
+ deleted T std existence merge rm from src not in src
+ [^+]*- ?? textual merge rm from src rm from src
+ +included BT list merge rm from non-tb src list merge
+ +ends BT compute specially compute specially
+ +*- ?? textual merge rm from non-tb src textual merge
+ *[^-] ?? die, aborting all ops, if found in any tb src or branch
~ian / topbloke.git / blobdiff