GCC is unable to emit the SEH metadata about the stack aligning that is
required to execute AVX aligned instructions (VMOVDQA, VMOVAPS, etc.),
so it just doesn't align the stack. That causes crashes on a 50/50
chance every time the compiler attempts to address a stack-aligned
variable. In a debug-mode build, because it always loads & saves
everything on the stack, the chance of a crash happening is a near
certainty.
So we hack around it by going behind the compiler's back and instructing
the assembler to emit the unaligned counterparts of the instructions
every time the compiler wished to emit the aligned one. There's no
performance penalty: if the variable is actually aligned, the unaligned
instruction executes in the exact same time.
Change-Id: Ib42b3adc93bf4d43bd55fffd16c29cac0da18972
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Emscripten only supports
SSE1, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, and 128-bit AVX instruction
sets at this time.
https://emscripten.org/docs/porting/simd.html
Browsers might need to enable simd support in the advanced
configurations
about: config or chrome:flags
Enable by configuring Qt with -sse2
Pick-to: 6.2
Fixes: QTBUG-63924
Change-Id: Ifeafae20e199dee0d19689802ad20fd0bd424ca7
Reviewed-by: Morten Johan Sørvig <morten.sorvig@qt.io>
The separate_debug_info configure test uses the CMake variable
CMAKE_OBJCOPY. CMakeFindBinUtils in the test project finds the host's
objcopy despite CMAKE_TOOLCHAIN_FILE being correctly set.
We now add CMAKE_OBJCOPY to the list of variables that are passed to
configure test projects and remove the CMakeFindBinUtils include, which
looks rather internal anyways.
Pick-to: 6.2
Fixes: QTBUG-96798
Change-Id: I164c6bd1771e8789e9dd19b50573b33b8866bd3b
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>
Needed for subsequent change that will check and error out if the
version is lower than 3.16. We do that to ensure all policies
introduced by CMake up to version 3.16 have their behavior set to
NEW.
Pick-to: 6.2
Task-number: QTBUG-95018
Change-Id: Ieaf82c10987dd797d86a3fd4a986a67e72de486a
Reviewed-by: Kai Koehne <kai.koehne@qt.io>
This configure test always failed, and its result was never used.
Pick-to: 6.2
Change-Id: I5112464b247efb5327ef5c23c96ef27168c11afc
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>
Remove target specific flags from static_link_order.
Move the check to the common config.tests folder.
Amends 5fb99e3860
Pick-to: 6.2
Task-number: QTBUG-93002
Task-number: QTBUG-94528
Change-Id: I1368075ec6bd1e743b2b89fd93143df38a278ec2
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>
Adds runtime CPU detection for Windows and macOS, and switches feature
detection of AES to runtime like for x86,
So far only on ARM64, since gcc doesn't do function versioning on ARM32,
but clang can, so it could be added later.
Change-Id: Ibe5d60f48cdae3e366a8ecd6263534ba2b09b131
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@qt.io>
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>
CMake 3.18 introduced the file(ARCHIVE_CREATE) API that we use with
COMPRESSION Zstd for compressing corelib's mimedatabase.
It's possible to build CMake without proper zstd support, and we have
encountered such builds in the wild where the file(ARCHIVE_CREATE) call
crashes.
Add a configure test to determine whether CMake properly supports the
Zstd compression method.
Fixes: QTBUG-89108
Change-Id: I37e389c878845162b6f18457984d4f73a265b604
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>
Remove the qmake project files for most of Qt.
Leave the qmake project files for examples, because we still test those
in the CI to ensure qmake does not regress.
Also leave the qmake project files for utils and other minor parts that
lack CMake project files.
Task-number: QTBUG-88742
Change-Id: I6cdf059e6204816f617f9624f3ea9822703f73cc
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Kai Koehne <kai.koehne@qt.io>
For this, we have to uninline the separate_debug_info configure test,
because supporting the conversion of this in configurejson2cmake is not
worth the hassle.
Separate debug information can be turned on for a target by calling the
function qt_enable_separate_debug_info. For Qt's shared libraries and
tools separate debug information is generated if the
'separate_debug_info' feature is manually turned on.
Change-Id: Ic2ffc15efef3794dc0aa42f3d853ef6d651a751c
Reviewed-by: Leander Beernaert <leander.beernaert@qt.io>
Reviewed-by: Alexandru Croitor <alexandru.croitor@qt.io>
Tested locally with the following configurations:
- iOS device builds (arm64)
- iOS simulator builds (x86_64)
- iOS simulator_and_device builds (fat arm64 and x86_64 archives)
All iOS builds currently require a custom vcpkg fork which contains
fixes for building the required 3rd party libraries.
qtsvg, qtdeclarative, qtgraphicaleffects and qtquickcontrols2
have also been tested to build successfully.
simulator_and_device builds are also supported, but require an umerged
patch in upstream CMake as well as further patches to vcpkg.
Task-number: QTBUG-75576
Change-Id: Icd29913fbbd52a60e07ea5253fd9c7af7f8ce44c
Reviewed-by: Cristian Adam <cristian.adam@qt.io>
Reviewed-by: Qt CMake Build Bot
Reviewed-by: Leander Beernaert <leander.beernaert@qt.io>
The Intel whitepaer[1] recommends using the RDSEED over RDRAND whenever
present. libstdc++ from GCC 10 will also use it in std::random_device.
[ChangeLog][QtCore][QRandomGenerator] The system() random generator will
now use the RDSEED instruction on x86 processors whenever available as
the first source of random data. It will fall back to RDRAND and then to
the system functions, in that order.
[1] https://software.intel.com/en-us/articles/intel-digital-random-number-generator-drng-software-implementation-guide
Change-Id: I907a43cd9a714da288a2fffd15bab176e54e1975
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Multi arch build in one go is need to support the new .aab packaging format.
By default the users apps are built for all Android ABIs: arm64-v8a armeabi-v7a x86_64 x86
The user can pass ANDROID_ABIS to qmake to filter the ABIs during development,
e.g. qmake ANDROID_ABIS="arm64-v8a armeabi-v7a" will build only for arm ABIs.
[ChangeLog][Android] Android multi arch build in one go,
needed to support the new .aab packaging format.
Change-Id: I3a64caf9621c2a195863976a62a57cdf47e6e3b5
Reviewed-by: Eskil Abrahamsen Blomfeldt <eskil.abrahamsen-blomfeldt@qt.io>
If the host architecture is different from -platform (canadian cross
build with -external-hostbindir) then we cannot use QMAKE_HOST.os to
deduce the executable extension for that platform, because this value
comes from the qmake binary that was pointed to by
-external-hostbindir.
Move the target name deduction mechanism to the actual configure test
.pro files to make sure the right scopes are available, and write the
deduced target name to a text file. That text file is read by
qtConfTest_architecture to get the right binary to analyze.
Fixes: QTBUG-77286
Change-Id: I68b844dd51dbfda6432a4b0dca6331899c82255f
Reviewed-by: Kai Koehne <kai.koehne@qt.io>
Android is also unix, so can pick up the host 'arch' binary when
rerunning configure. This patch splits the names so we don't end up
confusing target and host binaries.
Task-number: QTBUG-76445
Change-Id: Ib65251a514e45ad8873f523d71c17e13e56ea58a
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
In qmake there are at least 2 things to know regarding
sub-architectures and instruction sets.
Which instruction sets does the compiler know to compile for,
represented by the various config.tests and features in
qtbase/configure.json.
And which instructions sets are enabled by the compiler by default,
represented by the configure.json "architecture" test and accessed
via QT_CPU_FEATURES.$$arch qmake argument.
Before this patch there was some mishandling of the above concepts
in CMake code.
The former can now be checked in CMake with via TEST_subarch_foo and
QT_FEATURE_foo (where foo is sse2, etc).
The latter can now be checked by
TEST_arch_${TEST_architecture_arch}_subarch_foo
(where foo is sse2, etc and the main arch is dynamyicall evaluated).
The configurejson2cmake script was adjusted to take care of the above
changes, and the cmake files were regenerated as well.
Change-Id: Ifbf558242e320cafae50da388eee56fa5de2a50c
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
GCC for 64-bit Windows has a bug that it fails to properly re-align the
stack pointer for use with 256-bit memory addresses (AVX). Therefore,
there's about a 50/50 chance that any function using AVX will have an
improperly-aligned stack. In release mode, stack accesses should be
rare, but in debug mode they happen frequently. Either way, this is a
ticking time bomb, so we disable.
Clang is not affected.
32-bit MinGW is not affected.
64-bit in other OSes with GCC are not affected.
Fixes: QTBUG-73539
Change-Id: Id061f35c088044b69a15fffd1580967808f31671
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
use library objects for all variants, and inline the tests.
Change-Id: I029f9a6655a783dab4a22abf601aadbb484c03af
Reviewed-by: Joerg Bornemann <joerg.bornemann@qt.io>
... and use that to inline the xlocalescanprint test.
Change-Id: I0973133d7f9ecc9a38b70dc4b83df174a35b2b1f
Reviewed-by: Joerg Bornemann <joerg.bornemann@qt.io>
these were new on dev while the original migration happened on 5.9, or
came from new changes which hadn't adapted yet.
Change-Id: I5e48437061a97e6df6e93881c98471455e177631
Reviewed-by: Joerg Bornemann <joerg.bornemann@qt.io>
Some Linux distributions patch OpenSSL's soname, making builds
on such distributions not deployable elsewhere. The problem is that
the code loading OpenSSL symbols would attempt to use the soname
of the build machine, and therefore not finding the OpenSSL
libraries on the deploy system.
The binary builds of Qt for Linux are affected by this problem,
as they build under RHEL7.4 which changes to soname of OpenSSL to
a non-standard string. This makes the binary builds not pick up
OpenSSL 1.0 from the machine where the build gets installed on.
Given that in the pre-1.1 versions only the 1.0 series is supported,
bump the minimum requirement of Qt to that. The 1.0.x releases
(up to 1.0.2, at the time of this writing) have kept binary
compatibility, and advertise a soname of "1.0.0", which is used
by most distributions.
So, if loading of OpenSSL with the build-time soname fails,
try to load them with the "1.0.0" hardcoded soname.
[ChangeLog][QtNetwork][SSL] OpenSSL >= 1.0 is now required to build
Qt with OpenSSL support.
Task-number: QTBUG-68156
Change-Id: Ieff1561a3c1d278b511f09fef06580f034f188c6
Reviewed-by: Timur Pocheptsov <timur.pocheptsov@qt.io>
_mm512_mask_cvtepi32_storeu_epi8 is VPMOVDB (convert from 32-bit to 8-bit
with truncation) where the destination is a memory address, with an
OpMask register used to indicate which of the lanes in the vector to
store. Similarly, _mm512_mask_cvtepi16_storeu_epi8 is VPMOVWB (convert
from 16-bit o 8-bit), which is useful for UTF-16 to Latin1 conversion.
Change-Id: I8f261579aad648fdb4f0fffd15542ea306841ce6
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
"AVX512MIC" (Many Integrated Cores) is the set of AVX-512 features found
on the Intel Xeon Phi coprocessors (codename "Knights Landing"), which
is an unlikely architecture for Qt to run on.
The two profiles with VL came from study of early GCC code and are no
longer applicable. GCC source code now shows both VBMI and IFMA as part
of the -march=cannonlake feature set.
Change-Id: Iff4151c519c144d580c4fffd153a0f268919fe2c
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
This adds detection for: VAES, GFNI, AVX512VBMI2, AVX512VNNI,
AVX512BITALG, AVX512VPOPCNTDQ, AVX512_4NNIW, AVX512_4FMAPS. These
features were found in the "Intel® Architecture Instruction Set
Extensions and Future Features" manual, revision 30. This commit also
adds support for RDPID (already in the main manual) and the Control-flow
Enforcement Technology, which appears in a separate Intel paper.
This new support was done by adding a new generator script so we don't
have to maintain two tables in sync, one in qsimd.cpp with the feature
names, and the other in qsimd_p.h.
Since we now need a lot more bits, it's no longer worth keeping the two
halves of the qt_cpu_features variable mostly similar to the main two
CPUID results. This commit goes back to keeping things in order, like we
used to prior to commit 6a8251a89b (Qt 5.6)
At the time of this commit, GCC 8 has macros for AVX512VPOPCNTDQ,
AVX512_4NNIW, AVX512_4FMAPS, AVX512VBMI2 and GFNI.
Change-Id: I938b024e38bf4aac9154fffd14f7afae50faaa96
Reviewed-by: Edward Welbourne <edward.welbourne@qt.io>
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Since the x86_simd/main.cpp file already has all the source for each and
every test anyway, just reuse it.
Change-Id: I938b024e38bf4aac9154fffd14f779f450827fb9
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
This has two main benefits:
1) introduces a qmake CONFIG we can use in .pro/.pri/.prf files
2) removes the need to keep an up-to-date list of which compilers
support the feature
The test is implemented as trying to compile every single SIMD test we
currently have, but without passing the -mXXX option. The reason for
trying all of them is that some people may have modified their mkspecs
to add -mXXX options or -march=XXX, which could enable the particular
feature we tried, resulting in a false positive outcome.
Change-Id: I938b024e38bf4aac9154fffd14f7784dc8d1f020
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
It seems the compiler supports /arch:AVX512 and /arch:AVX512F but none
of the other switches (and neither are documented). And when you pass
those, you also get Conflict Detection (CD), Double & Quad (DQ), Byte &
Word (BW) and Vector Length (VL), which matches the ICC switch
"-xCORE-AVX512". Unlike ICC, there doesn't seem to be an option to
enable only the common part of AVX-512.
Support for Intel Xeon Phi's current features (Exponential &
Reciprocation and Prefetch) and future ones (IFMA, VBMI, 4FMAPS, 4VNNI
and VPOPCNTDQ) seems to be missing altogether.
See https://blogs.msdn.microsoft.com/vcblog/2017/07/11/microsoft-visual-studio-2017-supports-intel-avx-512/
Change-Id: I98105cd9616b8097957db680d73eb1f86e487e6d
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
Convert QSysInfo/QOperatingSystemVersion to __builtin_available where
required or possible, or to QOperatingSystemVersion where
__builtin_available cannot be used and is not needed (such as negated
conditions, which are not supported by that construct).
Change-Id: I83c0e7e777605b99ff4d24598bfcccf22126fdda
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
The use as in the code:
futimesat(fd, NULL, &tv)
is not documented to work. The file descriptor should be a directory's
one, not an open file (though the Linux source code seems to handle that
case). This call was done as a fallback to futimes, so it's very
unlikely a system would have futimesat and not futimes.
Both the Linux and the FreeBSD man pages say it's deprecated anyway.
Change-Id: I8d96dea9955d4c749b99fffd14cd94068dc7668a
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
Availability of D3D11_QUERY_DATA_TIMESTAMP_DISJOINT depends on the used
MinGW version so that the check for MINGW is not sufficient. The newly
added configure test can be used for every toolset.
Task-number: QTBUG-57916
Change-Id: Ia9cb48f3e673841101a93cbc8ea23aff9547f639
Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@qt.io>
Reviewed-by: Kai Koehne <kai.koehne@qt.io>
