> On Sept. 15, 2013, 10:29 a.m., Frank Reininghaus wrote: > > Thanks for your cool work on QCollator! It will be interesting to see how > > much we can gain by using QCollatorSortKey for sorting large sets of > > QStrings :-) > > > > I'm not really familiar with most of the affected code, and I couldn't test > > it yet (frameworks currently does not build for me, but it's most likely an > > issue with my system which can fixed by doing a clean build), but here are > > some things that I noticed: > > > > (a) Is there a reason why you use a helper class to wrap the QCollator in > > kfile/kurlnavigatorbutton.cpp, but pass the QCollator directly to the sort > > function in other places? > > > > (b) I'm wondering how cheap it is to initialize a QCollator. Could existing > > code outside of kdelibs that uses KStringHandler::naturalCompare() for > > sorting become slow if that happens N*log(N) times? > > > > (c) Something that was not clear to me at all when I first heard about > > QCollator is that its behavior will depend on whether ICU headers were > > installed when Qt was built or not, and that things like > > setNumericMode(true) will simply be ignored (with a warning printed to the > > terminal) if ICU was not available then. Even worse: > > QCollator::numericMode() returns true in that case, but it does not use > > "numeric mode" for sorting at all! > > > > I just found out about that when I tried to write a simple test program > > that sorts strings using QCollator. It did not work, and after some > > research I found out that this is because I only have the ICU libs, but not > > the headers installed on my system. > > > > Now the Qt 5 packages that end up on our users' systems will probably be > > compiled with ICU support, but still, I have a very uneasy feeling about > > using a class that may or may not do what you expect, and that provides no > > good way to find out if it will do what you expect (as I said, > > QCollator::numericMode() from qcollator_posix.cpp always returns true). I'm > > already seeing the bug reports coming from people who built Qt from source > > and "forgot" (like me) to install the ICU headers before. > > > > I don't see a good solution for that problem. Even if we made the ICU > > headers a hard dependency for frameworks, it could still be that Qt was > > built without ICU support. > > > > Probably the best solution would be to try to get something like our > > KStringHandler::naturalCompare() function into qcollator_posix.cpp, to make > > sure that the fallback that is used if ICU isn't available actually uses > > "numeric mode" if it's told to? > > > > Aleix Pol Gonzalez wrote: > a) Well, I tried to minimize the number initializations, but I also tried > to reduce the code changes. > b) I don't have such data. It's a possibility, that it's slower. Either > way, the less we do, the faster it will work. > c) When you configure Qt, if ICU is found it will be used. I think it's > ok to assume that Dolphin on linux users will all have ICU available. > > I wouldn't hack around the posix backends, personally. > > Frank Reininghaus wrote: > "I think it's ok to assume that Dolphin on linux users will all have ICU > available." > > If you build Qt from source, you have to install ICU headers manually in > order to "have ICU available" (at least if the ICU-devel package is not > installed by the distro by default). This means that it's very easy to end up > with a QCollator that does not support "numeric mode". Considering that many > people who contribute to KDE do build Qt from source, it will most likely go > wrong for some of them, so I tend to strongly disagree with the "it's ok to > assume..." statement. These people will notice that things don't work as > expected and either waste time analyzing the problem or file a bug report. > > I experienced this myself yesterday: I noticed that QCollator did not > work for me, and I was surprised about that because, according to the API > docs, "setNumericMode(true)" causes the sorting to be "natural", and it does > not mention any conditions that have to be fulfilled. At least I saw the > warning message in Konsole, but if a user/developer doesn't even see that > (e.g., because it gets lost in .xsession-errors), how is he/she supposed to > know what the cause of the problem is? > > This is all my personal opinion, and I don't actually maintain any of the > affected code, but I tend to think that using a class that may or may not do > what it actually pretends to do, depending on things that are out of our > control, might not be a good idea. > > John Layt wrote: > The plan for Qt is in 5.3 to have ICU is a hard-ish dependency on Linux. > QLocale will use ICU for all localization on Linux, and only provide a simple > POSIX fallback for embedded platforms that don't want ICU. Distro's will be > told that they should always build with ICU support enabled. (We were going > to make ICU a hard dependency on all platforms in 5.2, but Windows devs > objected too much). > > This is not that different from needing the Cups headers installed if you > want to do any Qt printing under Linux, Qt will build without them, the API > won't change, and there's no way to query if you have Cups support built, but > when you try to print all you'll get is the option for PDF. > > The problem really comes from Qt configure not giving very good feedback > on what's missing and the consequences there-of. On Linux it auto-detects of > ICU headers are installed and then enables ICU support, but you can force it > on by using the -icu configure flag which will cause it to complain if they > are not found. I think we will have to document in the build instructions > that KF5 works best with ICU support explicitly enabled with -icu.
Thanks for the info! I fully agree with Aurélien that this should be fixed in Qt (i.e., either the build should fail without ICU, or QCollator should do what its API docs say it does even if ICU is not available). However, I don't see why it is so urgent to start using QCollator in frameworks now, before the problem is fixed. Even if it helps with splitting, one could at least leave KStringHandler::naturalCompare() as it is and only use QCollator in code which should not depend on kcoreaddons any more. But if I'm the only one who thinks that it can be problematic to throw away <150 lines of working and well-tested code in favor of something that only works if some requirements that we have no control over are met, then I will stop complaining now ;-) I'm not really involved in the frameworks splitting anyway, but I had been added to this review request, so I thought that I could share my opinion. Maybe I'm a bit overcautious because of my frustration that is caused by the endless stream of incoming bug reports which are mostly not really about bugs in Dolphin, but bugs in kdelibs, Qt, various kioslaves, thumbnailers, context menu plugins and quite a few other things (one user even reported a problem with the GTK+ file dialog as a Dolphin bug). This is why I don't really like the idea of adding code which makes us vulnerable to known problems in external libraries. - Frank ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: http://git.reviewboard.kde.org/r/112717/#review40054 ----------------------------------------------------------- On Sept. 13, 2013, 5:55 p.m., Aleix Pol Gonzalez wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > http://git.reviewboard.kde.org/r/112717/ > ----------------------------------------------------------- > > (Updated Sept. 13, 2013, 5:55 p.m.) > > > Review request for KDE Frameworks, Frank Reininghaus and Aurélien Gâteau. > > > Description > ------- > > Now that QCollator is public API, I wanted to give it a try, so I decided to > port all uses KStringHandler::naturalCompare() to QCollator. > > All the porting was quite straightforward I'd say, the only weird part is > that I removed some tests of the KStringHandler tests. There are 2 kind of > tests disabled: > - The ones that say that "A" == "a" in case of Qt::CaseInsensitive. ICU is > deterministic and it will decide itself which one goes in, so the test > doesn't make sense anymore. > - There's a mention to the 237788 bug where in some cases our former > algorithm wouldn't be deterministic. This doesn't apply anymore, but also now > ICU takes care about it now, so there's little point of keeping unit testing > it. > I decided to leave the unit test because it might be useful eventually > (although note that it was not being compiled so far). In any case we > probably want it out. > > In any case, the rest seems straightforward enough. I didn't concentrate on > performance though, in some cases we'll want to use the QCollatorSortKey. > > > Diffs > ----- > > KDE5PORTING.html 1287de7 > kfile/kdirsortfilterproxymodel.cpp 7c7aa5f > kfile/kurlnavigatorbutton.cpp d2c27fd > staging/itemviews/src/CMakeLists.txt 353a413 > staging/itemviews/src/kcategorizedsortfilterproxymodel.h a21e7ca > staging/itemviews/src/kcategorizedsortfilterproxymodel.cpp c8b652d > staging/itemviews/src/kcategorizedsortfilterproxymodel_p.h eb1a67b > staging/kcompletion/src/kcompletion.cpp 5f60a6c > staging/xmlgui/src/kshortcutsdialog_p.h ab102bc > staging/xmlgui/src/kshortcutseditoritem.cpp e89a8aa > tier1/kcoreaddons/autotests/CMakeLists.txt 19227dc > tier1/kcoreaddons/autotests/kstringhandlertest.cpp d12f086 > tier1/kcoreaddons/src/lib/text/kstringhandler.h 1b08f6f > tier1/kcoreaddons/src/lib/text/kstringhandler.cpp 2f192aa > > Diff: http://git.reviewboard.kde.org/r/112717/diff/ > > > Testing > ------- > > Builds, affected tests pass. > > > Thanks, > > Aleix Pol Gonzalez > >
_______________________________________________ Kde-frameworks-devel mailing list Kde-frameworks-devel@kde.org https://mail.kde.org/mailman/listinfo/kde-frameworks-devel
