Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
308 Python 3 preparation [2021/03/03] [Infernio, Utumno]
308 is a release almost entirely focused on getting us closer to running on Python 3. Nearly all refactoring here has the end goal of making that transition easier, even if that may not be immediately apparent. All commits mentioned here were authored by @Infernio or @Utumno. ----------------------------------------------------------------------- ### Records (#480) 308 development began with the merge of Part '2.5', which was already finished in 307, but had not been merged yet for lack of testing. This was done in bfae14e. 7e67f63 introduced a significant refactoring of subrecord handling, bringing all sorts of improvements like usage of AOT unpacker construction via struct.Struct, centralized strings handling (see also #460 below), removal of lots of special cases, etc. It also acts as a basis for most other records work that happened in 308 and will happen in 309+. A huge improvement came in the form of 0848872, which finally got rid of the horrible ModFile.__getattr__. This hack (let's call it what it is) caused fun, surprising behavior when debugging and trying to extend ModFile's APIs every time. It also stood in the way of #460, because py3 does not like getattr with a bytestring. d7bd007 was a merge that sprang from the goal of refactoring getNumRecords, but soon grew to encompass larger refactorings under #312, #460 and #480. One final key merge in 308 was af16a5d, which focused on the setDefault API. Wrye Bash mostly copied defaults from xEdit, but it turns out that in many cases we don't need or want to have defaults (e.g. for performance - see also next section). Dropping them entirely is not feasible, however, since we do need to write out some new records (e.g. the TES4 header, GMSTs, etc.). This merge is a step in the direction of making defaults None unless otherwise specified. ----------------------------------------------------------------------- ### Performance (mostly #563) One of my personal pet peeves is how slow many of Wrye Bash's operations still are. 309 will tackle more in this direction. These are not just nice to haves either, they can cause serious usability problems. See issue #563 for some information on the '0.1s threshold' and how important it is to usability of an application. Specific improvements in 308 are: - c0b638b improved performance when viewing INI tweaks and target INIs on the INI Edits tab. - 9946a83 cached the target INIs on the INI Edits tab to reduce system calls. - 3d09aeb was the main merge, significantly improving the performance of most load order operations. See the merge and its commit message for details on the improvements. - 8507350 was a small commit that fixed one of my biggest problems with 307: the incessant flicker every time you tried using the global menu. - 4a48ae9 dropped backwards compatibility with 306 settings (see #376), speeding up BAIN exit significantly. - bd73ea6 made CIstr construction much faster, which made many BAIN operations (e.g. dragging and dropping, startup, exit) faster. ----------------------------------------------------------------------- ### wx (#190) Some more de-wx'ing took place in f707b59 and 7645e40. In 5dbda51, we upgraded to wxPython 4.1, mostly for the high DPI code that was added in wxWidgets 3.1.x. We were originally OK with keeping bash.py was a file that directly used wxPython APIs, but high DPI (see #555 section below) forced us to find a way to de-wx even it. This was accomplished in e899dc8, which also laid the groundwork for showing our app icons on boot popups (which we did in 538ff2a). One major open problem with de-wx'ing was the wizards API. The problem is that the wx.adv.Wizard class wants instances of wx.adv.WizardPage to be passed along, but we want to encapsulate those instances entirely. In addition, our hacky usage of 'dummy pages' did not at all fit with wx's intended use case of the wizards API. All this came to a head when wxPython 4.1 broke the Back button in our wizards, likely because of our hacky usage with 'dummy' pages. Instead of monkeypatching and trying to fix it, we threw out the entire wx.adv.Wizard API and wrote our own, based on de-wx'd components, in a3e6c4d. With this merge, the biggest roadblock to fully de-wx'ing Wrye Bash is now gone. ----------------------------------------------------------------------- ### Patchers (#312) The first massive and extremely impactful merge of 308 was b25727b, known as 'tweak pooling'. This is an entirely different implementation of tweaks in the Bashed Patch that brings all sorts of improvements: - Thousands of lines of boilerplate code are gone - The pickle files we used to store injection FormIDs are gone - Tweaks are now almost entirely static classes (finished off a couple merges later) - Several tweakers have been ported to newer games - Significant performance improvements because much fewer records now have to be copied into the BP The second merge focusing on patchers was ac72709, in which the names of patcher classes were decoupled from the keys used to store and load them from/to BP config pickles. This allowed us to rename them to all fit one consistent scheme. Another notable commit that was part of this merge is eaeb831, which brings us much closer to being able to absorb Import Cells into _APreserver as well as fixing several issues with Import Cells. 8a9b581 was a small followup commit that unified scanOrder and editOrder into a single patcher_order variable. f3b6578 followed up on tweak pooling to make all tweaks entirely static. This brings with it huge benefits like potential for deduplication of similar tweaks, making it much easier to add new ones, etc. 9a1cff7 was a relatively small merge under #312, #460 and #480 that focused on renames and using some of the APIs introduced in earlier #312 and #480 merges to refactor more code. ----------------------------------------------------------------------- ### CBash (#530) After having been deprecated in 307, CBash was removed entirely in 120234c. See that commit for reasoning on why, but the short version is that it was unmaintained, cint was poorly designed and the CBash patcher implementations were extremely buggy. ----------------------------------------------------------------------- ### GPLv3 (#531) Mostly a formality, but we used the 'or any later version' clause of our GPLv2 or later license to upgrade to GPLv3 or later in 3721668. We were effectively already on GPLv3 because of the dependencies we used, some of which would only work if the combination of WB source code + dependency is licensed under the GPLv3. This is just making it official. ----------------------------------------------------------------------- ### Morrowind (#479) a5d40e1 was a merge that introduced decoded record classes for Morrowind. We're still nowhere close to actually reading and writing Morrowind plugins, but this is a step in the right direction. The main reason of course is so that all the records refactoring has to take into account Morrowind support when designing APIs. ----------------------------------------------------------------------- ### 64bit (#481) In 30dd357, we finally upgraded Wrye Bash to 64bit after it had sat around on nightly for a long, long time. The main reason for the hesitation was that some people reported cryptic tracebacks on launch (python-lz4 was failing to import). It turned out in the end that the fix for this was installing the MSVC 2010 x64 redistributable, which we now do as part of our installer. 64bit Python 2 gives us somewhere between a 10%-30% speedup for most operations that we tested, most notably the Bashed Patch. ----------------------------------------------------------------------- ### Python 3 (#460) As mentioned above, Python 3 preparation is the focus of this release. The main hurdle here turned out to be the amount of low-level bytestring handling that Wrye Bash does. As a result, refactoring these parts of the code was a key goal, which resulted in patchers (#312) and records (#480) refactoring becoming the most important open issues. Everything described in those sections above applies to this section too. Furthermore, we performed many targeted changes to make a Python 3 upgrade more feasible and to catch errors early. Several commits here were originally created by @GandaG in 307: - 4b4cc01 cherry-picked a bunch of backwards-compatible changes made by 2to3. - 8e0fa26 dropped Path usages in bash.ini settings handling (see also the section on #543 below). - 567f153 and 09e1633 replaced several usages of keys() and iterkeys() with equivalent expressions that work identically in py3 (e.g. next(iter(a)) intead of a.keys()[0]). - 83162d7 was a minor merge focusing mostly on getting rid of __getattr__ and __setattr__ and its cousins. They are slower on Python 3 than simple getattr/setattr. - b337e98 was a major merge focusing on cooperating with 2to3 to help it do better at its job. We also created a fork of 2to3 to push this concept even further with specialized fixers. - a75dbc6 was originally a large commit by @GandaG that shrank over the course of 308's development. It focused on dropping map/filter/range usages in favor of list comprehensions or xrange. Prefixing of strings was also an ongoing task, worked on in many commits: - cb6b9ce - 2486397 - 83162d7 - 1a29a01 - 20432e6 - 42e5795 - 52ba47b ----------------------------------------------------------------------- ### DataStore keys (#543) With Python 3 on the way and after a bunch of performance profiling (see the performance section above), it became clear that bolt.Path stood in the way of both. Thus, #543 was born - a project to make DataStore (our class for handling file collections like mods, BSAs, INI tweaks, etc.) use strings instead of Paths as keys. This turned out to be a gigantic rabbit hole, leading to a ton of refactoring: - 74e0524 was a merge containing a bunch of housekeeping, but the most important change was a reduction in the usage of the '.s' property on Paths. - 129e098 renamed tons of 'path' occurences to make tracking down Path instances easier. - 5d81acd was a big merge focusing on following the DataStore key rabbit hole down to its event horizon. The resulting refactoring improved the names of many variables, locals, etc. and brought us much closer to absorbing BAIN structures, Paths instances, etc. ----------------------------------------------------------------------- ### BSA load order (#534, #546) We had several scattered attempts to discern the order in which BSAs load all over the codebase. 6aec0ce centralized this handling and gave us several key improvements in the process, including the ability for BAIN to show conflicts with BSAs loaded via INIs, e.g. the vanilla BSAs (#546) and correct handling of the special sVrResourceArchiveList INI setting in Skyrim VR (#534). ----------------------------------------------------------------------- ### The People tab (#548) Was removed in 308. We used to keep it around for the purpose of keeping the data structures generic enough, but with the ongoing refactoring, it turned out to be too much of a maintenance burden, so it had to go. ----------------------------------------------------------------------- ### High DPI (#555) With the upgrade to wxPython 4.1, nothing was stopping us from finally adding high DPI support to Wrye Bash, which we did in e899dc8. This is still not perfect, the most notable issue is that we do not have high resolution versions of our icons, but it's better than the blurry mess Windows displays us as otherwise. 12571a6 was a small followup commit that marked our installer as high DPI-aware too. ----------------------------------------------------------------------- ### env (#258) Running Wrye Bash on Linux is a long-term goal that's very important to us - not just for personal usage, but also for code quality. One important part of this is abstracting over OS differences, for which we have our env module. Unfortunately, env in 307 was in effect a completely Windows-specific module that could only run on Linux if you commented out large chunks of it. All that was fixed in 19461b3, a merge that split env.py and windows.py into an env package. Wrye Bash can now launch of Linux and even build a BP - though it is still far from usable, especially BAIN. ----------------------------------------------------------------------- ### Renaming (#580) Renaming is a highly complex thing. Near the end of 308's development, one big branch focusing on this area was merged in 5a985d8. It fixed issues with renaming screenshots and installers, significantly improved our rename APIs and added the ability to rename .bak files on the Saves tab. We're not quite done with refactoring here, which is why #580 is not yet closed, but this will hopefully be the last time we have to deal with renaming issues (plus renaming is not really a good idea in many cases - e.g. renaming a mod will cause serious problems if you're using it on an existing save and power users can do it anyways by just opening the Data folder and editing the mod in there). ----------------------------------------------------------------------- Massive thanks to everyone who contributed to this release, including: @Infernio, @Utumno, @GandaG, @lojack5, @Gavvers, @LordNyriox, @Sharlikran, @Arthmoor and many more that GitHub's contribution tracker doesn't list.
- Loading branch information
