localrepo: enable ellipsis flag on revlogs when repo is narrow If the narrow requirement is present, revlogs created for that repository will have the ellipsis flag enabled. This is the same behavior that the narrow extension exhibits. Except the ellipsis flag won't be enabled on repos/revlogs that don't have the narrow requirement. Differential Revision: https://phab.mercurial-scm.org/D4648

revlog: add opener option to enable ellipsis flag processor The ellipsis flag processor can now be registered by specifying an opener option when constructing a revlog instance. This allows us to enable ellipsis flags on a per-revlog basis. Differential Revision: https://phab.mercurial-scm.org/D4647

revlog: store flag processors per revlog Previously, revlog flag processing would consult a global dict when processing flags. This was simple. But it had the undesired side-effect that any extension could load flag processors once and those flag processors would be available to any revlog that was subsequent loaded in the process. e.g. in hgweb, if the narrow extension were loaded for repo A but not repo B, repo B would be able to decode ellipsis flags even though it shouldn't be able to. Making the flag processors dict per-revlog allows us to have per-revlog controls over what flag processors are available, thus preserving desired granular access to flag processors depending on the revlog's needs. If a flag processor is globally registered, it is still globally available. So this commit should not meaningfully change behavior. Differential Revision: https://phab.mercurial-scm.org/D4646

revlog: define ellipsis flag processors in core We will soon be teaching core to honor the ellipsis flag on revlogs. Moving the definition of the processor functions to core is the first step in this. The processor is still not registered unless the narrow extension is loaded. Differential Revision: https://phab.mercurial-scm.org/D4645

narrow: remove custom filelog type This functionality is now handled by core as of the previous commit. I wanted this to be a standalone commit because the deleted code makes a reference to remotefilelog's file type missing a node() method and this may have implications to narrow+remotefilelog usage. The code in core doesn't perform this check and therefore behavior may be subtly different and buggy. But I /think/ the check is merely a performance optimization and nothing more. So I'm optimistic this will continue to "just work." Differential Revision: https://phab.mercurial-scm.org/D4644

filelog: custom filelog to be used with narrow repos Narrow repos may have file revisions whose copy/rename metadata references files not in the store. This can pose problems when consumers attempt to access a missing referenced file revision. The narrow extension hacks around this problem by implementing a derived filelog type that provides custom implementations of renamed(), size(), and cmp() which handle renames against files not in the narrow spec by silently removing the rename metadata. While silently dropping metadata isn't the most robust solution, it is the easiest to implement. This commit ports the custom narrow filelog class to core. When a narrow repo is constructed, its ifilestorage creation function will automatically use the new filelog type. This means the extra logic is 0 cost for non-narrow repos and shouldn't interfere with their operation. Differential Revision: https://phab.mercurial-scm.org/D4643

localrepo: iteratively derive local repository type This commit implements the dynamic local repository type derivation that was explained in the recent commit bfeab472e3c0 "localrepo: create new function for instantiating a local repo object." Instead of a static localrepository class/type which must be customized after construction, we now dynamically construct a type by building up base classes/types to represent specific repository interfaces. Conceptually, the end state is similar to what was happening when various extensions would monkeypatch the __class__ of newly-constructed repo instances. However, the approach is inverted. Instead of making the instance then customizing it, we do the customization up front by influencing the behavior of the type then we instantiate that custom type. This approach gives us much more flexibility. For example, we can use completely separate classes for implementing different aspects of the repository. For example, we could have one class representing revlog-based file storage and another representing non-revlog based file storage. When then choose which implementation to use based on the presence of repo requirements. A concern with this approach is that it creates a lot more types and complexity and that complexity adds overhead. Yes, it is true that this approach will result in more types being created. Yes, this is more complicated than traditional "instantiate a static type." However, I believe the alternatives to supporting alternate storage backends are just as complicated. (Before I arrived at this solution, I had patches storing factory functions on local repo instances for e.g. constructing a file storage instance. We ended up having a handful of these. And this was logically identical to assigning custom methods. Since we were logically changing the type of the instance, I figured it would be better to just use specialized types instead of introducing levels of abstraction at run-time.) On the performance front, I don't believe that having N base classes has any significant performance overhead compared to just a single base class. Intuition says that Python will need to iterate the base classes to find an attribute. However, CPython caches method lookups: as long as the __class__ or MRO isn't changing, method attribute lookup should be constant time after first access. And non-method attributes are stored in __dict__, of which there is only 1 per object, so the number of base classes for __dict__ is irrelevant. Anyway, this commit splits up the monolithic completelocalrepository interface into sub-interfaces: 1 for file storage and 1 representing everything else. We've taught ``makelocalrepository()`` to call a series of factory functions which will produce types implementing specific interfaces. It then calls type() to create a new type from the built-up list of base types. This commit should be considered a start and not the end state. I suspect we'll hit a number of problems as we start to implement alternate storage backends: * Passing custom arguments to __init__ and setting custom attributes on __dict__. * Customizing the set of interfaces that are needed. e.g. the "readonly" intent could translate to not requesting an interface providing methods related to writing. * More ergonomic way for extensions to insert themselves so their callbacks aren't unconditionally called. * Wanting to modify vfs instances, other arguments passed to __init__. That being said, this code is usable in its current state and I'm convinced future commits will demonstrate the value in this approach. Differential Revision: https://phab.mercurial-scm.org/D4642