* using [::file] instead of [file] in snit methods to get the core file command * taking care of files that only differ in case in the Attic and the repo on case insensitive FS * passing platform specific ... (check-in: 7208c7ac user: mjanssen tags: trunk)

Speed. This is now mostly acceptable. The one exception is pass 'BreakAllCsetCycles'. The reason is that the limit computation it does for backward branches still uses the inefficient file-level dependency computation. This will be tackled in short order. For the other passes the file spent is 'CsetsDeps' is recouped by the much faster graph setup.

Memory. This seems to be mostly acceptable as well, with the exceptions of 'BreakAllCsetCycles' (again, for reasons see above), and 'InitCsets'. It seems to happen while the pass breaks internal dependencies, but there is no hard data. I have to measure using a memory-debug enabled tclsh. I suspect either the actual internal dependencies, or the pseudo-dependencies. Maybe combined with a bad choice of data structures. Well, measuring first. ... (check-in: e8c374f6 user: aku tags: trunk)

Created a new pass 'CsetDeps' which is run between 'InitCsets' and 'BreakRevCsetCycles' (i.e. changeset creation and first breaker pass). It computes the changeset dependencies from the file-level dependencies once and saves the result in the state, in the new table 'cssuccessor'. Now the breaker and sort passes can get the information quickly, with virtually no effort. The dependencies are recomputed incrementally when a changeset is split by one of the breaker passes, for its fragments and its predecessors.

The loop check is now trivial, and integrated into the successor computation, with the heavy lifting for the detailed analysis and reporting moved down into the type-dependent SQL queries. The relevant new method is 'loops'. Now that the loop check is incremental the pass based checks have been removed from the integrity module, and the option '--loopcheck' has been eliminated. For paranoia the graph setup and modification code got its loop check reinstated as an assert, redusing the changeset report code.

Renumbered the breaker and sort passes. A number of p... (check-in: 00bf8c19 user: aku tags: trunk)

Background: In the last few days I repeatedly ran against the wall of an assertion in pass 9, last of the cycle breakers, with the revision changesets coming in out of order when the symbols were added to the dependency graph.

While walking to the office, and later re-reading the relevant parts of cvs2svn again I had several insights. Trigger was the realization that giving the tag changesets successor dependencies was wrong. Tags describe a state, they have no successors. This caused the re-read, and I recognized that my handling of the symbol changesets was completely wrong, that with using revisions as their data. It should have been the tags and branches. From there their actual dependencies (versus my reuse of revision dependencies) fell out naturally.

I have decided to commit my rewrite of the internals incrementally to make it easier to follow them, despite leaving the sourcebase in an unusable state during the series. One big commit would be much more difficult to understand.

The... (check-in: 27b15b70 user: aku tags: trunk)