We were using direct loading and operator=, which for everything except
std::atomic was very relaxed. But std::atomic<T *> defines the direct
access to actually be the least relaxed possible, under the idea that if
you didn't know any better to use a member function, you probably need
the most protection.
So use Ops::load and Ops::store.
Change-Id: Id5480807d25e49e78b79ffff144a06a2e6398576
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
I wrote a script to help find the files, but I reviewed the
contributions manually to be sure I wasn't claiming copyright for search
& replace, adding Q_DECL_NOTHROW or adding "We mean it" headers.
Change-Id: I7a9e11d7b64a4cc78e24ffff142b506368fc8842
Reviewed-by: Lars Knoll <lars.knoll@theqtcompany.com>
[ChangeLog][Important Behavior Changes] Starting with Qt 5.7, Qt
requires a C++11 compiler with support for C++11 atomics. This affects
user code too: Qt headers no longer compile with a C++98 compiler. The
minimum compiler versions for this release are:
* Clang 3.4 (found in XCode 5.1)
* GCC 4.7
* Microsoft Visual Studio 2012
Change-Id: Ib056b47dde3341ef9a52ffff13ef1f496ea9363f
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Dmitry Shachnev <mitya57@gmail.com>
From Qt 5.7 -> LGPL v2.1 isn't an option anymore, see
http://blog.qt.io/blog/2016/01/13/new-agreement-with-the-kde-free-qt-foundation/
Updated license headers to use new LGPL header instead of LGPL21 one
(in those files which will be under LGPL v3)
Change-Id: I046ec3e47b1876cd7b4b0353a576b352e3a946d9
Reviewed-by: Lars Knoll <lars.knoll@theqtcompany.com>
The "previous" value is always 1 when the compare-and-swap succeeded,
instead of the previous value.
Instead of fixing this, let's just remove this file a bit earlier than
the rest. All of them will be removed in Qt 5.7 anyway, so let's leave
MIPS atomics to the compiler.
Task-number: QTBUG-49168
Change-Id: Idba8c29717f34c70a58fffff14133304595165f5
Reviewed-by: Dmitry Shachnev <mitya57@gmail.com>
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@theqtcompany.com>
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Lisandro Damián Nicanor Pérez Meyer <perezmeyer@gmail.com>
And add tests for the GCC intrinsics and for std::atomic.
Task-number: QTBUG-43794
Change-Id: Ic5d393bfd36e48a193fcffff13b9b2dbaee80469
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Qt copyrights are now in The Qt Company, so we could update the source
code headers accordingly. In the same go we should also fix the links to
point to qt.io.
Outdated header.LGPL removed (use header.LGPL21 instead)
Old header.LGPL3 renamed to header.LGPL3-COMM to match actual licensing
combination. New header.LGPL-COMM taken in the use file which were
using old header.LGPL3 (src/plugins/platforms/android/extract.cpp)
Added new header.LGPL3 containing Commercial + LGPLv3 + GPLv2 license
combination
Change-Id: I6f49b819a8a20cc4f88b794a8f6726d975e8ffbe
Reviewed-by: Matti Paaso <matti.paaso@theqtcompany.com>
We map the Apple Clang versions to upstream, so that we have one
define to compare against.
Fixes build break on iOS due to qbasicatomic.h not defining
QT_BASIC_ATOMIC_HAS_CONSTRUCTORS on Apple Clang versions, which
is needed after 1e9db9f5e1
Change-Id: I17493c0187c20abc5d22e71944d62bfd16afbad2
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
qatomic_sparc.h was removed previously as part of general cleanup of
old atomic code for Qt 5.3. Unfortunately SPARC include reference was
not removed resulting in build failure on that platform.
Task-number: QTBUG-41384
Change-Id: Ic6e31b32324b0e5dd3700a6a21515a8eea5668bd
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
[ChangeLog][QtCore][Atomic support]Added more operations to the atomic
classes, including operator T(), operator=(T), operator++,
operator--. For the QAtomicInteger, bit-manipulation operations are
also provided, both in operator and in fetchAndXxxYyyyyy modes.
Change-Id: I39c07be74e15e0a48f9e931f4342b182004dee1a
Reviewed-by: David Faure <david.faure@kdab.com>
This is extremely useful, since the most common action after a failed
compare-and-swap is to loop around, trying again with the current
value as found in memory.
Code currently written as:
do {
Type value = atomic.load();
...
} while (!atomic.testAndSetRelaxed(value, desired));
Becomes:
Type value = atomic.load();
do {
...
} while (!atomic.testAndSetRelaxed(value, desired, value));
In most CPU architectures, the value that was found in memory is known
to the compare-and-swap code, so this is more efficient than the
previous code. In architectures where the value is not known, the new
code is no worse than before.
The implementation sometimes modified an existing function, sometimes
it added a new one, depending on whether more registers were needed in
the assembly (like ARMv6-7), the code became more complex (ARMv5), the
optimizer failed (C++11), or it was just plain equivalent (MIPS).
Change-Id: I7d6d200ea9746ec8978a0c1e1969dbc3580b9285
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@digia.com>
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This simplifies the code a lot and avoids silly mistakes where a
specific integer type is missing (such as char16_t).
Change-Id: Id91dfd1919e783e0a9af7bfa093ca560a01b22d1
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
The new atomic code was introduced in Qt 5.0. The platforms that did not
get ported were announced as deprecated in Qt 5.2. The code is now
removed in Qt 5.3.
The status for the platform/compiler/OS combinations affected is:
* Linux with GCC or Clang: still compiles on all platforms
(via qatomic_cxx11.h or qatomic_gcc.h)
* INTEGRITY with Green Hills compiler: no longer compiles
* Solaris on UltraSPARC, with Sun Studio: no longer compiles
* AIX on POWER5 or 6, with IBM Visual Age: no longer compiles
(probably did not compile Qt 5.0 either)
* VxWorks in kernel mode: no longer compiles
[ChangeLog][General] Support for the following platforms has been
removed, due to lack of interest in updating support: INTEGRITY,
VxWorks, Solaris on UltraSPARC (with the Sun Studio compiler suite), AIX
on POWER processors (with IBM Visual Age compiler suite).
Change-Id: I8a961385fd95011c016b2b1eec52034794dae3e1
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
Reviewed-by: Konstantin Ritt <ritt.ks@gmail.com>
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Before, we'd get just an error message that the size of the array was
negative. Now, for C++11 compilers, we get a better error message:
qbasicatomic.h:117:5: error: static assertion failed: Template parameter is not a supported integer on this platform
qbasicatomic.h:119:24: error: invalid use of incomplete type ‘struct QAtomicOps<long long unsigned int>’
Change-Id: I6b0792254c0dc6103a4a7608f2942d59cda07c00
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
The public documentation for load() and store() says it's atomic,
and it is:
* using _q_value.store(newValue, std::memory_order_relaxed) in the C++11
implementation
* using a simple assignment otherwise, which is atomic (and relaxed, no
memory barriers) on all the existing C++ ABIs.
Change-Id: I40faa47120163225bd11c3a32514ac97ef8bbbd4
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The comments in this file suggested that the bug was only present in clang/3.1.
However, qRegisterMetaType was failing even on clang 3.2
("tags/RELEASE_32/final"). The clang's bugzilla says [1] that the problem was
fixed two days before the release of 3.2, but apparently it didn't make it to
the release branch.
[1] http://llvm.org/bugs/show_bug.cgi?id=13470
Change-Id: I37db8f6f6b22ab939110e79240d92313c1786d6a
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
don't test for building specific tools when we actually only want to
know whether we are bootstrapping. so far, this was only redundant; with
the upcoming change of not bootstrapping unnecessarily it would be
outright broken.
Change-Id: I7600d8ebb14a4194640c50035e35a04263f2ccce
Reviewed-by: Joerg Bornemann <joerg.bornemann@digia.com>
The macro was made empty in ba3dc5f3b5
and is no longer necessary or used.
Discussed-on: http://lists.qt-project.org/pipermail/development/2013-January/009284.html
Change-Id: Id2bb2e2cabde059305d4af5f12593344ba30f001
Reviewed-by: Laszlo Papp <lpapp@kde.org>
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@digia.com>
Reviewed-by: hjk <hjk121@nokiamail.com>
qatomic.h includes qbasicatomic.h, so break the cyclic inclusion
problem.
Change-Id: If72bacbcfa36993336786f31e8980fbd43df8bdc
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
This triggers a bug in clang (before version 3.2). See also the bug
report: http://llvm.org/bugs/show_bug.cgi?id=12670 .
Change-Id: I9e0bc1cc39059ffa5e062652b932bc01e453ef98
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Morten Johan Sørvig <morten.sorvig@digia.com>
As of Qt5, this macro is defined to be empty; simply get rid of these leftovers.
Change-Id: I167ccb4c9e92ec9b5e4faeb02bf9c5ef5d982b50
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Change copyrights and license headers from Nokia to Digia
Change-Id: If1cc974286d29fd01ec6c19dd4719a67f4c3f00e
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Reviewed-by: Sergio Ahumada <sergio.ahumada@digia.com>
This macro is defined when the QBasicAtomic* types have constructors.
It will be used to simplify maintenance of QBasicAtomic subclasses,
as well as classes that aggregate Q(Basic)Atomics and wish to be
literal types if possible.
Change-Id: I7ab181c58a13ec2195449bbae5b3476b9017a8f2
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This allows one to write code that depends on these values at
compile-time.
Change-Id: I7d78524ed9c70d4141360496d1d764dcbfa92e62
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
These functions are inline (on most architectures) but they contain
inline assembly which the compiler could not understand. In any case,
if it decides not to inline them, it would need to generate exception
handlers.
Change-Id: If9d50793d715c51781e76d0a539da03a6d83e255
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Lars Knoll <lars.knoll@nokia.com>
Based on reviewer comments on
https://codereview.qt-project.org/#change,25992, split
Q_COMPILER_DEFAULT_DELETE_MEMBERS into
Q_COMPILER_DEFAULT_MEMBERS and Q_COMPILER_DELETE_MEMBERS. Clang provides
2 separate feature checks for these, all other compilers always define
both.
Keep the Q_COMPILER_DEFAULT_DELETE_MEMBERS define, though, for any
existing code that may be using it.
Change-Id: I8969e10989c776a822da153152cfecb99dda8cb9
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
It's almost exactly the same code in both files, so let's have one file
only. That means we need an #ifdef for the special case of 64-bit types
on i386.
Also take the opportunity to add a comment explaining how this works.
Change-Id: I50d274fa026806ae511b1045aa8a5c25daaa0edc
Reviewed-by: Bradley T. Hughes <bradley.hughes@nokia.com>
This is the first step in merging the i386 and x86-64
architectures. The next commit will bring i386 support into
qatomic_x86.h.
Change-Id: I24105ea70f3fc29b3fb779a70053f99117440573
Reviewed-by: Bradley T. Hughes <bradley.hughes@nokia.com>
IBM's POWER and the PowerPC architecture have been merged into a single
ISA, the Power ISA (see http://www.power.org). Use this unified name in
Qt.
Change-Id: Ia41492b0031d890843e43c5f7ecd1e60c65bb75b
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
This provides a fallback implementation on UNIX when the Q_PROCESSOR_*
and Q_CC_*/Q_COMPILER_* checks fail to find an implementation.
Note that we always compile qatomic_unix.cpp, but code is only included
when QATOMIC_UNIX_H is defined (meaning the checks above did not find an
implementation).
Change-Id: I8ce047847206003b4fa96eb3fb76b1c2ffbc2dfc
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Make qbasicatomic.h include the OS/compiler/processor dependent
implementation.
For implementations that have not yet been ported to declare a
QAtomicOps, they need to #include <QtCore/qoldbasicatomic.h>, and the
new QBasicAtomicInteger and QBasicAtomicPointer should not be declared.
Change-Id: Ia951834484c9f8dfa75131592e5e716b68ff989b
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Use the new Q_PROCESSOR_* macros to decide which headers to include in
the atomic implementation. This also removes qatomic_arm.h, which isn't
needed anymore, just select the correct qatomic_armv*.h from
qbasicatomic.h
Change-Id: I954848feafb8c420949d066ffcee1dd2b271e13b
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Most of these headers are either forwarding headers, or we explicitly
stop syncqt so that it doesn't generate class includes for the atomic
implementation. Either way, syncqt doesn't see the QT_END_* (and
sometimes not QT_BEGIN_*), which this commit fixes.
Change-Id: Icc8da6f384f38b1ff4eb265c731ce2f2ed92a1a3
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The QT_ARCH_WINDOWS(CE) define(s) aren't used to control which
header to include anymore, so just remove it. We also do not need
the empty src/corelib/arch/windows/arch.pri.
Change-Id: I5400fc852af31907e533d0278540b8cd3da391cb
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@nokia.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
First, we do not need to have the QT_INTERLOCKED_REMOVE_VOLATILE(a)
macro anymore either, since the value stored in QBasicAtomicInteger is
not volatile.
Windows provides Interlocked*Pointer() functions in some configurations,
so we provide a partial specialization of QAtomicOps for pointer types.
For integer types, QAtomicOps selects an implementation based on the
size of the type. At the moment, we only support 32-bit types, but it
will be possible to add 64-bit later.
Note that the 32-bit specialization of QAtomicOpsBySize declares the
Type typedef as long, not int, since the Windows Interlocked*() API
takes parameters as longs and long pointers. Since this typedef differs
from the type given to QBasicAtomicInteger<T> by the QBasicAtomicInt
typedef, we need to templatise the _q_value parameter separately from
the other arguments in QGenericAtomicOps.
This templatisation would be necessary to port other architectures, such
as PA_RISC, where we need to have an int[4] array in the atomic type
while the arguments do not need this array.
Change-Id: Id71fa1ae334da2023553cb402b45e6c285f1d344
Reviewed-by: Friedemann Kleint <Friedemann.Kleint@nokia.com>
Reviewed-by: João Abecasis <joao.abecasis@nokia.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
As in the past, to avoid rewriting various autotests that contain
line-number information, an extra blank line has been inserted at the
end of the license text to ensure that this commit does not change the
total number of lines in the license header.
Change-Id: I311e001373776812699d6efc045b5f742890c689
Reviewed-by: Rohan McGovern <rohan.mcgovern@nokia.com>
The C++11 std::atomic type is very close to our API, to the point one
has to wonder if the committe was inspired by it. It provides all of
the memory semantics that Qt requires and more, plus some
compare-and-swap operations that we don't use.
The idea of returning the actual value in the event of a failed
compare-and-swap is actually quite good, as often we'll retry with
it. We just couldn't come up with a good name (fetchAndTestAndSet?).
The C++11 atomics require that the compiler support constexpr as well,
since std::atomic itself isn't required by the standard to be
trivially-constructible (in fact, it has a constexpr constructor in
the standard). For that reason, we need constexpr so we can add a
constructor to QBasicAtomic too.
Change-Id: I12c51455ba73350a6f7501aacc2ca7681c4369dd
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
With this implementation, we can have Qt run on any architecture that
GCC supports without having to write specialised code. However, on
some architectures, the code that GCC generates is not optimal: it
uses locking on ARMv5 and it's always fully-ordered. For that reason,
it appears after the Qt native assembly implementations (it's a
fallback, not an override).
Since they all have fully-ordered memory semantics, we define only the
xxxRelaxed functions. The exception is __sync_lock_and_test, which has
acquire semantics, so we need to define the Release and Ordered
versions too.
On some architectures, GCC can support atomics on types different than
32-bit and pointer-sized (like x86, x86-64, ARM and even
MIPS). However, there's no standardised way of telling: GCC seems to
define __GCC_HAVE_SYNC_COMPARE_AND_SWAP_{1,2,4,8} if those operations
are present, but I couldn't find it on the ARM compiler (it was there
for i386, x86-64, IA-64 and MIPS).
Change-Id: I55ff7a7c0cfc6388b7ad8e2c0dedecffdf2a3e01
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Reviewed-by: Olivier Goffart <ogoffart@woboq.com>
The IA-64 architecture supports the actual memory ordering semantics
in many instructions, but not all. We actually implement the functions
for all operations, so we get the best possible output.
It does support proper load-acquire and store-release semantics, but
we don't need instructions for it: the ABI requires that a volatile
load be acquire and a volatile store be release.
The Intel and HP compiler codepaths are rewritten, but untested.
Change-Id: I7aa62a4ec65f63a97d1bbd8418bb2492c2be465f
Reviewed-by: Bradley T. Hughes <bradley.hughes@nokia.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The LL/SC instructions are only present on MIPS II and up, so don't
pretend to support MIPS I. The previous implementation emitted the
instructions by telling the assembler to change instruction sets. Now,
the user must pass an -march= option to GCC telling it which
architecture or processor is being targetted.
On MIPS64, the 64-bit implementation allows supporting for long long
too.
Change-Id: I6dae6f8f61e563aba6a663227d91c5ddf554aa6a
Reviewed-by: Bradley T. Hughes <bradley.hughes@nokia.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The implementation is divided in two files, as it used to be in the
previous implementation: one for ARMv5, one for ARMv6 and up.
For the ARMv5 implementation:
Drop the non-Linux EABI version of the atomics, as it's not
ABI-compatible with the ARMv6 and ARMv7 implementations. This
means this ARMv5 implementation only works on Linux. If other
systems implement kernel helpers like Linux, they can be added
too.
We use the __kernel_cmpxchg located at 0xffff0fc0 to implement the
operations, except for fetchAndStore, for which we use the SWP
instruction.
Also introduce the use of __kernel_dmb (at 0xffff0fa0) for the
memory barrier. Now this code is SMP-safe even when built with
ARMv5.
The kernel cmpxchg helper was introduced in Linux 2.6.12, whereas
the dmb helper was introduced in 2.6.15. That means 2.6.15 is the
minimum version now.
For ARMv6 and up:
Introduce byte, half-word and doubleword atomics that work on
ARMv6K and up.
For ARMv6 specifically, the memory barrier instruction (DMB) isn't
present, so we need to accomplish the same with the MCR
coprocessor instruction.
Change-Id: Ife7f9b920bcc7d1eef7611761f6c59ea940ec7df
Reviewed-by: Bradley T. Hughes <bradley.hughes@nokia.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Liang Qi
Thiago Macieira
Jani Heikkinen
Liang Qi
Tor Arne Vestbø
Matti Paaso
Adam Majer
David Faure
Jan Kundrát
Oswald Buddenhagen
Sergio Ahumada
Erik Verbruggen
Konstantin Ritt
Iikka Eklund
Marc Mutz
Bradley T. Hughes
Stephen Kelly
Jason McDonald
Thiago Macieira
Denis Dzyubenko