Port of Robin's work from I0a53aa4581e25b351b9cb5033415b5163d05fe71
on top of the new qHash patches (the original commit just introduced
lots of conflicts, so I redid it from scratch).
This is based on the work done in the QHash benchmark over the past
few months experimenting with the performance of the string hashing
algorithm used by Java.
The Java algorithm, in turn, appears to have been based off a
variant of djb's work at http://cr.yp.to/cdb/cdb.txt.
This commit provides a performance boost of ~12-33% on the
QHash benchmark.
Unfortunately, the rcc test depends on QHash ordering.
Randomizing QHash or changing qHash will cause the test to fail
(see QTBUG-25078), so for now the testdata is changed as well.
Done-with: Robin Burchell
Change-Id: Ie05d8e21588d1b2d4bd555ef254e1eb101864b75
Reviewed-by: João Abecasis <joao.abecasis@nokia.com>
Reviewed-by: Robin Burchell <robin+qt@viroteck.net>
Avoid the conversion to a temporary QString -- just hash the address
as a byte array.
Change-Id: Ic35cdbbc3ee66c32a28d911bd27de0092395979f
Done-with: Shane Kearns
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Shane Kearns <shane.kearns@accenture.com>
When sorting a model recursively, the children of a QFileSystemNode
are extracted from their parent in a QHash order; then filtered,
then sorted (using a stable sort) depending on the sorting column.
This means that the order of the children comparing to equal for
the chosen sort are shown in the order they were picked from the
iteration on the QHash, which isn't reliable at all.
Moreover, the criteria used in QFileSystemModelSorter for sorting
are too loose: when sorting by any column but the name, if the result
is "equality", then the file names should be used to determine
the sort order.
This patch removes the stable sort in favour of a full sort,
and fixes the criteria of soring inside QFileSystemModelSorter.
Change-Id: Idd9aece22f2ebbe77ec40d372b43cde4c200ff38
Reviewed-by: Stephen Kelly <stephen.kelly@kdab.com>
Same as with QtCore, remove the #ifdef and #ifndef and select the side
with STL.
Change-Id: If1440080328c7c51afe35f5944a19dafc4761ee5
Reviewed-by: Stephen Kelly <stephen.kelly@kdab.com>
Reviewed-by: Morten Johan Sørvig <morten.sorvig@nokia.com>
QT_NO_STL is now no longer defined, so remove the conditionals and
select the STL side.
Change-Id: Ieedd248ae16e5a128b4ac287f850b3ebc8fb6181
Reviewed-by: João Abecasis <joao.abecasis@nokia.com>
Two equal QByteArrays must return the same hash.
Change-Id: Iddd45b0c420213ca2b82bbcb164367acb6104ec8
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Two equal strings / stringrefs must return the same hash.
Change-Id: I2af9a11ab721ca25f4039048a7e5f260e6ff0148
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
It was confusing DataLocation and GenericDataLocation, and the same
for CacheLocation and GenericCacheLocation. The test was passing in
the api_changes branch because these were giving the same result
(empty app name), but the QCoreApplication::applicationName fix in master
makes these different, so the bug in the test showed up after merging.
Change-Id: I80ef6883c96cfd02b8c277d9d686717028d396bb
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This uses an alternative approach to the testing formerly introduced
in 4ef5a626. Zero-termination tests are injected into all QCOMPARE/QTEST
invocations. This makes such testing more thorough and widespread, and
gets seamlessly extended by future tests.
It also fixes an issue uncovered by the test where using a past-the-end
position with QString::insert(pos, char), could move uninitialized data
and clobber the null-terminator.
Change-Id: I7392580245b419ee65c3ae6f261b6e851d66dd4f
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@nokia.com>
The approach used to verify for zero-termination is too intrusive and
requires additional maintenance work to ensure new zero-termination
tests are added with new functionality.
Zero-termination testing will be re-established in a subsequent commit.
This reverts commit 4ef5a6269c.
Change-Id: I862434a072f447f7f0c4bbf8f757ba216212db3c
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@nokia.com>
This enables easier updating of those structs, by reducing the amount of
code that needs to be fixed. The common (and known) use cases are
covered by the two macros being introduced in each case.
Change-Id: I44981ca9b9b034f99238a11797b30bb85471cfb7
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
There were two constuctors offering essentially the same functionality.
One taking the QStatic*Data<N> struct, the other what essentially
amounts to a pointer wrapper of that struct. The former was dropped and
the latter untemplatized and kept, as that is the most generic and
widely applicable. The template parameter in the wrapper was not very
useful as it essentially duplicated information that already maintained
in the struct, and there were no consistency checks to ensure they were
in sync.
In this case, using a wrapper is preferred over the use of naked
pointers both as a way to make explicit the transfer of ownership as
well as to avoid unintended conversions. By using the reference count
(even if only by calling deref() in the destructor), QByteArray and
QString must own their Data pointers.
Const qualification was dropped from the member variable in these
wrappers as it causes some compilers to emit warnings on the lack of
constructors, and because it isn't needed there.
To otherwise reduce noise, QStatic*Data<N> gained a member function to
directly access the const_cast'ed naked pointer. This plays nicely with
the above constructor. Its use also allows us to do further changes in
the QStatic*Data structs with fewer changes in remaining code. The
function has an assert on isStatic(), to ensure it is not inadvertently
used with data that requires ref-count operations.
With this change, the need for the private constructor taking a naked
Q*Data pointer is obviated and that was dropped too.
In updating QStringBuilder's QConcatenable specializations I noticed
they were broken (using data, instead of data()), so a test was added to
avoid this happening again in the future.
An unnecessary ref-count increment in QByteArray::clear was also
dropped.
Change-Id: I9b92fbaae726ab9807837e83d0d19812bf7db5ab
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
QTBUG-23059 only affects 2 test functions, not the whole test. XFAIL the
2 failing tests.
Change-Id: I87086a9ec573362625bc090038dfd7c79aeb9426
Reviewed-by: Jason McDonald <jason.mcdonald@nokia.com>
This tests that we get the windowModalityChanged() signal as needed, but
not unnecessarily either.
Change-Id: I2232fa9d45c72e472b324b681859b4b0d574b467
Reviewed-by: João Abecasis <joao.abecasis@nokia.com>
Reviewed-by: Robin Burchell <robin+qt@viroteck.net>
Algorithmic complexity attacks against hash tables have been known
since 2003 (cf. [1, 2]), and they have been left unpatched for years
until the 2011 attacks [3] against many libraries /
(reference) implementations of programming languages.
This patch adds a qHash overload taking two arguments: the value to
be hashed, and a uint to be used as a seed for the hash function
itself (support the global QHash seed was added in a previous patch).
The seed itself is not used just yet; instead, 0 is passed.
Compatibility with the one-argument qHash(T) implementation is kept
through a catch-all template.
[1] http://www.cs.rice.edu/~scrosby/hash/CrosbyWallach_UsenixSec2003.pdf
[2] http://perldoc.perl.org/perlsec.html#Algorithmic-Complexity-Attacks
[3] http://www.ocert.org/advisories/ocert-2011-003.html
Task-number: QTBUG-23529
Change-Id: I1d0a84899476d134db455418c8043a349a7e5317
Reviewed-by: João Abecasis <joao.abecasis@nokia.com>
tst_rcc and tst_qdom rely on specific QHash orderings inside
rcc and QDom respectively (see QTBUG-25078 and QTBUG-25071).
A workaround is added to make them succeed: QDom checks for
all possible orderings, and rcc initializes the hash seed to 0
if the QT_RCC_TEST environment variable is set.
Change-Id: I5ed6b50602fceba731c797aec8dffc9cc1d6a1ce
Reviewed-by: João Abecasis <joao.abecasis@nokia.com>
Reviewed-by: Robin Burchell <robin+qt@viroteck.net>
It is an extension coming from the use case when you, for instance, need to
implement a countdown timer in client codes, and manually maintain a dedicated
variable for counting down with the help of yet another Timer. There might be
other use cases as well. The returned value is meant to be in milliseconds, as
the method documentation says, since it is reasonable, and consistent with the
rest (ie. the interval accessor).
The elapsed time is already being tracked inside the event dispatcher, thus the
effort is only exposing that for all platforms supported according to the
desired timer identifier, and propagating up to the QTimer public API. It is
done by using the QTimerInfoList class in the glib and unix dispatchers, and the
WinTimeInfo struct for the windows dispatcher.
It might be a good idea to to establish a QWinTimerInfo
(qtimerinfo_win{_p.h,cpp}) in the future for resembling the interface for
windows with the glib/unix management so that it would be consistent. That would
mean abstracting out a base class (~interface) for the timer info classes.
Something like that QAbstractTimerInfo.
Test: Build test only on (Arch)Linux, Windows and Mac. I have also run the unit
tests and they passed as well.
Change-Id: Ie37b3aff909313ebc92e511e27d029abb070f110
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Bradley T. Hughes <bradley.hughes@nokia.com>
The string from the server should begin with "* OK" and end
with "\r\n" according to the IMAP specification.
Still have a check for "server ready" as this does not change between
cyrus versions.
Change-Id: Ia01ed8aa054e5726bba8b411d30edc6205cc8465
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
According to documentation, SQLite doesn't have a separate Boolean
storage class. Instead, values are stored as integers 0(false) and
1(true). In QSqlQuery::bindValue(), if a boolean value is bound
to a placeholder, it is converted to text true and false. This fix
converts boolean value to integer 0 and 1.
Task-number: QTBUG-23895
Change-Id: I4945971172f0b5e5819446700390033a1a4ce301
Reviewed-by: Michael Goddard <michael.goddard@nokia.com>
Reviewed-by: Mark Brand <mabrand@mabrand.nl>
While QArrayDataPointer offers generic detach() functionality, this is
only useful for operations that may modify data, but don't otherwise
affect the container itself, such as non-const iteration, front() and
back().
For other modifying operations, users of the API typically need to
decide whether a detach is needed based on QArrayData's requirements
(is data mutable? is it currently shared?) and its own (do we have
spare capacity for growth?).
Now that data may be shared, static or otherwise immutable (e.g.,
fromRawData) it no longer suffices to check the ref-count for
isShared().
This commit adds needsDetach() which, from the point-of-view of
QArrayData(Pointer), answers the question: 'Can contained data and
associated metadata be changed?'.
This fixes QArrayDataPointer::setSharable for static data (e.g.,
Q_ARRAY_LITERAL), previously it only catered to shared_null.
SimpleVector is also fixed since it wasn't checking Mutability and it
needs to because it supports fromRawData().
Change-Id: I3c7f9c85c83dfd02333762852fa456208e96d5ad
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This enables a truncating resize() to be implemented. It is similar to
destroyAll(), but updates the size() as it goes, so it is safe to use
outside a container's destructor (and doesn't necessarily destroy all
elements).
The appendInitialize test was repurposed and now doubles as an
additional test for QArrayDataOps as well as exercising SimpleVector's
resize().
Change-Id: Iee94a685c9ea436c6af5b1b77486734a38c49ca1
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This follows QArrayData::detachFlags's lead. Given the (known) size for
a detached container, the function helps determine capacity, ensuring
the capacityReserved flag is respected.
This further helps aggregating behaviour on detach in QArrayData itself.
SimpleVector was previously using qMax(capacity(), newSize), but there's
no reason to pin the previous capacity value if reserve() wasn't
requested. It now uses detachCapacity().
Change-Id: Ide2d99ea7ecd2cd98ae4c1aa397b4475d09c8485
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Adds given number of default-initialized elements at end of array. For
POD types, initialization is reduced to a single memset call. Other
types get default constructed in place.
As part of adding a test for the new functionality the arrayOps test was
extended to verify objects are being constructed and assigned as
desired.
Change-Id: I9fb2afe0d92667e76993313fcd370fe129d72b90
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Way back in the mists of time, someone added SO_REUSEPORT to socket binding,
which was great, because otherwise it meant that multiple UDP sockets couldn't
share the same port on OS X (as platforms with SO_REUSEPORT apparently don't
support rebinding with SO_REUSEADDR).
However: SO_REUSEPORT also means that *any* bind on a port will succeed, which
is most definitely not wanted in the case of TCP sockets, so check the socket
type before performing the actual bind.
Also test that multiple listens don't take effect.
Change-Id: I2f8d450bcfb8a7f3abd8918a4e789a850281dd13
Done-with: Thiago Macieira
Done-with: Shane Kearns
Task-number: QTBUG-6305
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Shane Kearns <shane.kearns@accenture.com>
On OS X and Windows, this was not working, because the socket was being bound
in v6 mode (due to ::Any being for dual mode), but the address passed was a v4
address, meaning it took the wrong codepath. Linux, strangely, apparently works
anyway.
This is fixable in OS X (by using the v6 join path when bound in v6/dual mode),
but the same fix doesn't work on Windows, failing with WSAEADDRNOTAVAIL.
Don't allow this behaviour, and provide a sane error message telling the user
what to do instead.
Done-with: Shane Kearns
Task-number: QTBUG-25047
Change-Id: Iaf5bbee82e13ac92e11b60c558f5af9ce26f474b
Reviewed-by: Shane Kearns <shane.kearns@accenture.com>
QPlatformInputContext now gets notified on changed focus and
has inputMethodAccepted() telling whether current focus object
accepts input method events.
Also adapted IBus plugin to use this. Key event filtering for
focused objects without input method support got fixed by the
change.
Change-Id: I6910aa6af2459d752a5763f0ae88fa8c34e5b165
Reviewed-by: Joona Petrell <joona.t.petrell@nokia.com>
Unit test to override mask delay value so running it is not dependent
on platform style hint.
Change-Id: Ic5cc12d32cf97e64729b3af54250bdc05c0c95ad
Reviewed-by: Joona Petrell <joona.t.petrell@nokia.com>
When FTP login fails we fail to remove the entry from the cache.
This is because the cache key is created from the url with the
userInfo. So this needs to be set again to match the key used
when inserted.
Task-number: QTBUG-11824
Change-Id: Ib3fd2d737581653ae59c56d0810d42e2d8dc2176
Reviewed-by: Shane Kearns <shane.kearns@accenture.com>
Do not append blank character if it is the default.
Task-number: QTBUG-20834
Change-Id: I17f6ac4058f295f25ff49f33c41bd9ee40b75811
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@nokia.com>
Callers should just call the standard allocation functions directly.
Adding an extra function call onto all basic memory management for the sake of
making it instrumentable in rare cases isn't really fair to everyone else.
What's more, this wasn't completely reliable, as not everything was using them
in a number of places. Memory management can still be overridden using tricks
like LD_PRELOAD if needed.
Their aligned equivilents cannot be deprecated, as no standard equivilents
exist, although investigation into posix_memalign(3) is a possibility
for the future.
Change-Id: Ic5f74b14be33f8bc188fe7236c55e15c36a23fc7
Reviewed-by: Lars Knoll <lars.knoll@nokia.com>
Postgres async notifications can contain a payload parameter
that is currently discarded. This patch provides the QSqlDriver
api change necessary to deliver a payload with each emitted
notification by adding a QVariant parameter to the notification
signal. It also provides the implementation for the qsqlpsql driver.
The qsql_ibase driver has been updated to reflect the change to the
notification signal signature.
The eventNotificationPSQL test in the qsqldatabase test has
been expanded to test proper payload sending and receiving.
All tests/auto/sql/kernel tests have been run with sqllite and
postgres with no regressions.
Task-number: QTBUG-13500
Change-Id: I9137f6acc8cfca93f45791ca930e0287d93d5d0d
Reviewed-by: Mark Brand <mabrand@mabrand.nl>
Tests added that cover boundValues with positional binding,
and boundValueName.
Change-Id: I2962d76607b716d19d3e0be958109be2f032f2d9
Reviewed-by: Mark Brand <mabrand@mabrand.nl>
This patch fixes a critical bug in the qsqlpsql driver where
notifications aren't delivered when received. Any blocking libpq
function(specifically PQexec) will read all the incoming data
from the socket, including any pending notifications. This would
cause the socket notifier to never be fired for incoming
notifications that are already queued inside libpq. The qsqldriver
test case was skipping the postgres notification test because of
this bug, now its enabled and passing. In order to fix this
bug I made a wrapper function for PQexec in QPSQLDriverPrivate
that calls _q_handleNotification via QMetaObject::callMethod
QueuedConnection in order to deliver pending notifications
when control returns to the event loop. I also added a flag
to ensure only one call is made each time the event loop is
entered.
Change-Id: I19f5297094ae7ae46bfb0717e4fca744d69f7b92
Reviewed-by: Honglei Zhang <honglei.zhang@nokia.com>
Reviewed-by: Mark Brand <mabrand@mabrand.nl>
QSqlTableModel::headerData() generates a crash if an invalid filter
is set. QSqlQueryModel::indexInQuery() should check the index value
before applied to d->colOffsets[].
QSqlQueryModel::initRecordAndPrimaryIndex() is updated to sync the
size of rec and colOffsets.
Task-number: QTBUG-23879
Change-Id: Ic9f88bb288592aa6fb3c1415cc818632dadaab56
Reviewed-by: Michael Goddard <michael.goddard@nokia.com>
Reviewed-by: Mark Brand <mabrand@mabrand.nl>
For data allocated and maintained by QByteArray, there's a guarantee
that data() is null-terminated. This holds true even for null and empty,
where logically the terminating character should never be dereferenced.
For tests that modify or generate QByteArrays, this ensures the
invariant is kept.
In the toFromHex() text, const-ness of temporary variables was dropped
to enable the test macro to be used, as the qualification didn't add
much to the test otherwise.
Change-Id: I7ee52e79e3a9df7de18c743f3698dab688e6bf0e
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@nokia.com>
Prepared queries should be able to use a name parameter more than
once. Currently this will result in undefined behavior and crashes.
This patch fixes the bug and implements the needed test case.
Task-number: QTBUG-6420
Change-Id: I07d6537e432a9b2781e9ef3d9f597bceb054527e
Reviewed-by: Andy Shaw <andy.shaw@digia.com>
Reviewed-by: Honglei Zhang <honglei.zhang@nokia.com>
There are probably lots of places that rely on that behaviour, so go
back to what it was.
Change-Id: I4d1503a0ee105a50cdfaab52d9a5862a02c70757
Reviewed-by: David Faure <faure@kde.org>
I don't know if the bug is in moc or in qmake. But it bails out trying
to parse the .cpp file after the
tst_QUrlInternal::nameprep_testsuite_data function. If the #include is
placed above, it works. If it's placed below, it doesn't.
Change-Id: Ide554aa5aa3f1999e29604ba6d25ccdb09f6ef28
Reviewed-by: Marius Storm-Olsen <marius.storm-olsen@nokia.com>
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@nokia.com>
Don't crash when either side is null but not both sides.
Also make sure operator< is working properly and satisfies the basic
conditions of a type (such as that if A < B, then !(B < A)).
Change-Id: Idd9e9fc593e1a7781d9f4f2b13a1024b643926fd
Reviewed-by: Giuseppe D'Angelo <dangelog@gmail.com>
Reviewed-by: Lars Knoll <lars.knoll@nokia.com>
The strict mode check is now implemented after the tolerant parser has
finished, and only if the tolerant parser has not found any errors. We
catch the use of disallowed characters (control characters plus a few
not permitted anywhere) and broken percent encodings.
We do not catch the use of Unicode characters, as they are permitted
in IRIs.
In the tests, remove the old errorString test since it makes little
sense.
Change-Id: I8261a2ccad031ad68fc6377a206e59c9db89fb38
Reviewed-by: Lars Knoll <lars.knoll@nokia.com>
Note that QUrl can only remember one error. If the URL contains more
than one error condition, only the latest (in whichever parsing order
URL decides to use) will be reported.
I don't want too keep too much data in QUrlPrivate for validation, so
let's use 4 bytes only.
Change-Id: I2afbf80734d3633f41f779984ab76b3a5ba293a2
Reviewed-by: Lars Knoll <lars.knoll@nokia.com>
The use of any broken-down components of the query now needs
QUrlQuery.
The QUrl constructor and toString() are now rehabilitated and the
preferred forms. Use toEncoded() and fromEncoded() now only when we
need to store data in a QByteArray or the data comes from a QByteArray
anyway. Change to toString() or the constructor if the data was in a
QString.
Change-Id: I9d761a628bef9c70185a48e927a61779a1642342
Reviewed-by: Lars Knoll <lars.knoll@nokia.com>
