view mercurial/helptext/internals/mergestate.txt @ 46184:cb8b2ee89a5d

copies: stop attempt to avoid extra dict copies around branching

In the python code, we attempt to avoid unnecessary dict copies when gathering
copy information. However that logic is wobbly and I keep running into case
where independent branches affects each others.

With the current code we can't ensure we are the only "user" of dict when
dealing with merge.

This caused havoc in the next series on tests I am about to introduce.

So for now I am disabling the faulty optimisation. I believe we will need a
dedicated overlay to deal with the "copy on write logic" to have something
correct. I am also hoping to find time to build dedicated test case for this
category of problem instead of relying on side effect in other tests. However
for now I am focussing on another issue.

Differential Revision: https://phab.mercurial-scm.org/D9608

The active mergestate is stored in ``.hg/merge`` when a merge is triggered
by commands like ``hg merge``, ``hg rebase``, etc. until the merge is
completed or aborted to track the 3-way merge state of individual files.

The contents of the directory are:

Conflicting files
-----------------

The local version of the conflicting files are stored with their
filenames as the hash of their paths.

state
-----

This mergestate file record is used by hg version prior to 2.9.1
and contains less data than ``state2``. If there is no contradiction
with ``state2``, we can assume that both are written at the same time.
In this case, data from ``state2`` is used. Otherwise, we use ``state``.
We read/write both ``state`` and ``state2`` records to ensure backward
compatibility.

state2
------

This record stores a superset of data in ``state``, including new kinds
of records in the future.

Each record can contain arbitrary content and has an associated type. This
`type` should be a letter. If `type` is uppercase, the record is mandatory:
versions of Mercurial that don't support it should abort. If `type` is
lowercase, the record can be safely ignored.

Currently known records:

| * L: the node of the "local" part of the merge (hexified version)
| * O: the node of the "other" part of the merge (hexified version)
| * F: a file to be merged entry
| * C: a change/delete or delete/change conflict
| * P: a path conflict (file vs directory)
| * f: a (filename, dictionary) tuple of optional values for a given file
| * X: unsupported mandatory record type (used in tests)
| * x: unsupported advisory record type (used in tests)
| * l: the labels for the parts of the merge.

Merge record states (indexed by filename):

| * u: unresolved conflict
| * r: resolved conflict
| * pu: unresolved path conflict (file conflicts with directory)
| * pr: resolved path conflict

The resolve command transitions between 'u' and 'r' for conflicts and
'pu' and 'pr' for path conflicts.

This format is a list of arbitrary records of the form:

[type][length][content]

`type` is a single character, `length` is a 4 byte integer, and
`content` is an arbitrary byte sequence of length `length`.

Mercurial versions prior to 3.7 have a bug where if there are
unsupported mandatory merge records, attempting to clear out the merge
state with hg update --clean or similar aborts. The 't' record type
works around that by writing out what those versions treat as an
advisory record, but later versions interpret as special: the first
character is the 'real' record type and everything onwards is the data.