...to ease the implementation in C++ code of constructor-based <implementation>s
that shall be single-instance (and which would have been implemented with e.g.
cppu::creaetOneInstanceFactory for non--constructor-based <implementation>s).
See e.g. 6e35794cad "terminate XDesktop properly
in unit tests" and 6362ebab29 "dbaccess: create
instances with uno constructors" for the clumsy approach used until now, where
the C++ constructor function uses a static instance that is cleared in
dispose(), adding fake <singleton> entries to <implementation>s where necessary
so that the ServiceManager will call those XComponent::dispose() functions when
it itself gets disposed.
For every <implementation>, the ServiceManager already holds an Implementation
data structure, so it can easily hold a singleInstance there and clear it when
the ServiceManager gets disposed. (One consequence is that single-instance
implementations are now created with their Instance.mutex locked, but that
should not cause problems in practice, given that the construction of a single-
instance implementation should not recursively request its own construction
anyway.)
The new single-instance="true" attribute is mostly useful in combination with
the constructor attribute (see above), but it can also be used for non--
constructor-based <implementation>s, at least in theory.
(The single-instance="true" attribute is orthogonal to <singleton> elements.
There are existing single-instance services---even if those should arguably have
been defined as singletons in the first place---that can benefit from
single-instance="true". And there are <implementation>s that support one or
more <singleton>s alongside one or more <service>s, where the latter are not
single-instance.)
This new single-instance="true" attribute in *.component files should not
interfere with the lo_get_constructor_map machinery in the DISABLE_DYNLOADING
branch of cppuhelper::detail::loadSharedLibComponentFactory
(cppuhelper/source/shlib.cxx) used by Android, iOS and some fuzzer code.
AFAIU, the lo_get_constructor_map machinery should only ever come into play
after the ServiceManager has decided whether or not to create a new instance
based on the single-instance attributes in the *.component files.
This commit only provides the new single-instance="true" attribute. Actual
changes of <implementation>s and their C++ constructor functions are delegated
to a series of follow-up commits. (So that they can easily be reverted
individually should the need arise.)
Change-Id: Iea6c0fc539d74477b7a536dc771b198df6b0510e
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/103734
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
...from which an OUString can cheaply be instantiated. This is the OUString
equivalent of 4b9e440c51 "Turn OStringLiteral into
a consteval'ed, static-refcound rtl_String". Most remarks about that commit
apply here too (this commit is just substantially bigger and a bit more
complicated because there were so much more uses of OUStringLiteral than of
OStringLiteral):
The one downside is that OUStringLiteral now needs to be a template abstracting
over the string length. But any uses for which that is a problem (e.g., as the
element type of a container that would no longer be homogeneous, or in the
signature of a function that shall not be turned into a template for one reason
or another) can be replaced with std::u16string_view, without loss of efficiency
compared to the original OUStringLiteral, and without loss of expressivity.
The new OUStringLiteral ctor code would probably not be very efficient if it
were ever executed at runtime, but it is intended to be only executed at compile
time. Where available, C++20 "consteval" is used to statically ensure that.
The intended use of the new OUStringLiteral is in all cases where an
object that shall itself not be an OUString (e.g., because it shall be a
global static variable for which the OUString ctor/dtor would be detrimental at
library load/unload) must be converted to an OUString instance in at least one
place. Other string literal abstractions could use std::u16string_view (or just
plain char16_t const[N]), but interestingly OUStringLiteral might be more
efficient than constexpr std::u16string_view even for such cases, as it should
not need any relocations at library load time. For now, no existing uses of
OUStringLiteral have been changed to some other abstraction (unless technically
necessary as discussed above), and no additional places that would benefit from
OUStringLiteral have been changed to use it.
Global constexpr OUStringLiteral variables defined in an included file would be
somewhat suboptimal, as each translation unit that uses them would create its
own, unshared instance. The envisioned solution is to turn them into static
data members of some class (and there may be a loplugin coming to find and fix
affected places). Another approach that has been taken here in a few cases
where such variables were only used in one .cxx anyway is to move their
definitions from the .hxx into that one .cxx (in turn causing some files to
become empty and get removed completely)---which also silenced some GCC
-Werror=unused-variable if a variable from a .hxx was not used in some .cxx
including it.
To keep individual commits reasonably manageable, some consumers of
OUStringLiteral in rtl/ustrbuf.hxx and rtl/ustring.hxx are left in a somewhat
odd state for now, where they don't take advantage of OUStringLiteral's
equivalence to rtl_uString, but just keep extracting its contents and copy it
elsewhere. In follow-up commits, those consumers should be changed
appropriately, making them treat OUStringLiteral like an rtl_uString or
dropping the OUStringLiteral overload in favor of an existing (and cheap to use
now) OUString overload, etc.
In a similar vein, comparison operators between OUString and std::u16string_view
have been added to the existing plethora of comparison operator overloads. It
would be nice to eventually consolidate them, esp. with the overloads taking
OUStringLiteral and/or char16_t const[N] string literals, but that appears
tricky to get right without introducing new ambiguities. Also, a handful of
places across the code base use comparisons between OUString and OUStringNumber,
which are now ambiguous (converting the OUStringNumber to either OUString or
std::u16string_view). For simplicity, those few places have manually been fixed
for now by adding explicit conversion to std::u16string_view.
Also some compilerplugins code needed to be adapted, and some of the
compilerplugins/test cases have become irrelevant (and have been removed), as
the tested code would no longer compile in the first place.
sal/qa/rtl/strings/test_oustring_concat.cxx documents a workaround for GCC bug
<https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96878> "Failed class template
argument deduction in unevaluated, parenthesized context". That place, as well
as uses of OUStringLiteral in extensions/source/abpilot/fieldmappingimpl.cxx and
i18npool/source/localedata/localedata.cxx, which have been replaced with
OUString::Concat (and which is arguably a better choice, anyway), also caused
failures with at least Clang 5.0.2 (but would not have caused failures with at
least recent Clang 12 trunk, so appear to be bugs in Clang that have meanwhile
been fixed).
Change-Id: I34174462a28f2000cfeb2d219ffd533a767920b8
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/102222
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
This is a prerequisite for making conversion from OUStringLiteral to OUString
more efficient at least for C++20 (by replacing its internals with a constexpr-
generated sal_uString-compatible layout with a SAL_STRING_STATIC_FLAG refCount,
conditionally for C++20 for now).
For a configure-wise bare-bones build on Linux, size reported by `du -bs
instdir` grew by 118792 bytes from 1155636636 to 1155755428.
In most places just a u"..." string literal prefix had to be added. In some
places
char const a[] = "...";
variables have been changed to char16_t, and a few places required even further
changes to code (which prompted the addition of include/o3tl/string_view.hxx
helper function o3tl::equalsIgnoreAsciiCase and the additional
OUString::createFromAscii overload).
For all uses of macros expanding to string literals, the relevant uses have been
rewritten as
u"" MACRO
instead of changing the macro definitions. It should be possible to change at
least some of those macro definitions (and drop the u"" from their call sites)
in follow-up commits.
Change-Id: Iec4ef1a057d412d22443312d40c6a8a290dc6144
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/101483
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
> cppuhelper/source/typemanager.cxx:1571:47: error: field 'entity_' is uninitialized when used here [-Werror,-Wuninitialized]
> ServiceBasedSingletonDescription_Base(entity_->isPublished()),
> ^
The code was like that ever since it was introduced in
fa559f1c41 "WIP: Experimental new binary type.rdb
format". Odd that that never caused an issue, but the code appears to at least
never be executed during `make check`.
Change-Id: I19fe78162c7c4a9211c380648595df9aa54bf4d7
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/100554
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
Use the same code as for Linux on aarch64, with minor
additional hacks. But that will not actually work in all cases, as
there are slight differences in the ABI. See
https://developer.apple.com/library/archive/documentation/Xcode/Conceptual/iPhoneOSABIReference/Articles/ARM64FunctionCallingConventions.html
Thus we can drop the use of the lo_mobile_throwException() hack that
was very temporarily used.
The run-time code generation requires use of a new API on macOS to
work: See the use of pthread_jit_write_protect_np() in
bridges/source/cpp_uno/shared/vtablefactory.cxx.
For some reason, with the Xcode 12 betas, when compiling for
arm64-apple-macos, the symbols for the type_infos for the UNO
exception types (_ZTIN3com3sun4star3uno16RuntimeExceptionE etc) end up
as "weak private external" in the object file, as displayed by "nm -f
darwin". We try to look them up with dlsym(), but that then fails. So
use a gross hack: Introduce separate real variables that point to
these typeinfos, and look up and dereference them instead. If this
hack ends up needing to be permanent, instead of having a manually
edited set of such pointer variables, we should teach codemaker to
generate corresponding functions, and look up and invoke them to get
the std::type_info pointer.
When compiling for x86_64-apple-macos, the type_info symbols end up as
"weak external" which is fine.
With this, LibreOffice starts and seems to work to some extent, and
many unit tests succeed.
Change-Id: I05f46a122a51ade1ac7dccd57cb90e594547740e
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/100408
Tested-by: Jenkins
Reviewed-by: Tor Lillqvist <tml@collabora.com>
...that was apparently meant to guard the following
&context_values[0]
that got changed to
context_values.data()
in c18e2ea2a5 "New loplugin:data"
Change-Id: Ic63d8e775f3b75cca7df2b08665cccc0b4c3ec3e
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/99316
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
to look for the
x.get() != null
pattern, which can be simplified to
x
I'll do the
x.get() == nullptr
pattern in a separate patch, to reduce the chances of a mistake
Change-Id: I45e0d178e75359857cdf50d712039cb526016555
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/95354
Tested-by: Jenkins
Reviewed-by: Noel Grandin <noel.grandin@collabora.co.uk>
This basically extends what commit
eb15ac837e ("android: Avoid
throwing exceptions through the bridges.", 2019-11-29), amended
by f2f8a2ec10 ("android:
Actually let's avoid the exceptions via bridges only on aarch4.",
2019-12-02) did for the online-based Android app to cover the Android
Viewer case as well.
This makes Android Viewer work on aarch64, which previously crashed
whenever trying to load a document.
Change-Id: Id8b30e1dced583159cc4d81aaa18f4c25b901cff
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/90047
Tested-by: Jenkins
Reviewed-by: Michael Weghorn <m.weghorn@posteo.de>
With --enable-pch=full there's not much difference between a "public"
header in <module>/inc and a private one in <module>/src/somewhere/inc .
And since the script searches recursively, this apparently helps to
find even more headers for lower pch levels.
Change-Id: I8483d0aa5b4fea5a59107c20a8aa5f1ef694af0a
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/87799
Tested-by: Jenkins
Reviewed-by: Luboš Luňák <l.lunak@collabora.com>
"Find explicit casts from signed to unsigned integer in comparison against
unsigned integer, where the cast is presumably used to avoid warnings about
signed vs. unsigned comparisons, and could thus be replaced with
o3tl::make_unsigned for clairty." (compilerplugins/clang/unsignedcompare.cxx)
o3tl::make_unsigned requires its argument to be non-negative, and there is a
chance that some original code like
static_cast<sal_uInt32>(n) >= c
used the explicit cast to actually force a (potentially negative) value of
sal_Int32 to be interpreted as an unsigned sal_uInt32, rather than using the
cast to avoid a false "signed vs. unsigned comparison" warning in a case where
n is known to be non-negative. It appears that restricting this plugin to non-
equality comparisons (<, >, <=, >=) and excluding equality comparisons (==, !=)
is a useful heuristic to avoid such false positives. The only remainging false
positive I found was 0288c8ffec "Rephrase cast
from sal_Int32 to sal_uInt32".
But which of course does not mean that there were no further false positivies
that I missed. So this commit may accidentally introduce some false hits of the
assert in o3tl::make_unsigned. At least, it passed a full (Linux ASan+UBSan
--enable-dbgutil) `make check && make screenshot`.
It is by design that o3tl::make_unsigned only accepts signed integer parameter
types (and is not defined as a nop for unsigned ones), to avoid unnecessary uses
which would in general be suspicious. But the STATIC_ARRAY_SELECT macro in
include/oox/helper/helper.hxx is used with both signed and unsigned types, so
needs a little oox::detail::make_unsigned helper function for now. (The
ultimate fix being to get rid of the macro in the first place.)
Change-Id: Ia4adc9f44c70ad1dfd608784cac39ee922c32175
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/87556
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
instead of fighting the fact that Sequence is intended to be immutable
getSequence() will become a little slower, but since getSequence()
is called 5k times during 'make check', while add/remove are called
3M times, I think it's safe to say mutation is dominant.
Change-Id: I23d127c93c589050d6c06c0da47871e1ebe7b52d
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/86929
Tested-by: Jenkins
Reviewed-by: Noel Grandin <noel.grandin@collabora.co.uk>
this fixes commit 3a0c457444, by
converting the assert in removeInterface() to SAL_WARN_IF, and similarly
for the related addInterface() method
Comment from sberg on the original commit:
This leaves OInterfaceContainerHelper::add-/removeInterface in somewhat
inconsistent state (as addInterface OSL_ASSERTs non-null but then also
explicitly checks for null, i.e., effectively supports null).
css.lang.XComponent::add-/removeEventListener do not explicitly specify
whether or not their arg may be null. I /think/ we traditionally
supported null; asserting non-null thus might cause issues for 3rd-party
code.
Change-Id: I1e28cc03aa7dfe662125c573d60b447309009448
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/86827
Tested-by: Jenkins
Reviewed-by: Noel Grandin <noel.grandin@collabora.co.uk>
It got introduced with b6cebf4a3e "Allow UNO
component libraries to have each implementation in its own function", then
ae3a0c8da5 "Add .component <implementation
constructor='...' feature" removed all of its uses and noted: "[...] (which had
previously been adapted to use the prefix="direct" feature, which may become
unnecessary again in the end, depending on how to handle single-instance
services/singletons)." Handling of single-instance services/singletons has been
added with 997d211833 "Support for singleton
constructor functions" and the feature has really become unnecessary again.
Change-Id: I1ab910805d61421f892a244313c8778cd453f98c
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/86765
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
found using 'git grep', I tried using clang-tidy, but it only
successfully found a tiny fraction of these
Change-Id: I61c7d85105ff7a911722750e759d6641d578da33
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/86526
Tested-by: Jenkins
Reviewed-by: Noel Grandin <noel.grandin@collabora.co.uk>
...and rename constructor members to constructorName/constructorFn for
disambiguation.
"Splitting ImplementationInfo out of Implementation has become necessary to
avoid circular references" in c16d0dd846
"fdo#49291 Postpone actual factory instantiation as long as possible", but was
no longer necessary since 997d211833 "Support for
singleton constructor functions" (which also removed the need to hold
ImplemenationInfo by shared_ptr, which has already been cleaned up with
fc11bce491 "ImplementationInfo can be inlined").
Change-Id: I51467aa7b1fce56d6549738efd1f34f91925dac5
Reviewed-on: https://gerrit.libreoffice.org/c/core/+/86013
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
...following up on 314f15bff0 "Extend
loplugin:external to warn about enums".
Cases where free functions were moved into an unnamed namespace along with a
class, to not break ADL, are in:
filter/source/svg/svgexport.cxx
sc/source/filter/excel/xelink.cxx
sc/source/filter/excel/xilink.cxx
svx/source/sdr/contact/viewobjectcontactofunocontrol.cxx
All other free functions mentioning moved classes appear to be harmless and not
give rise to (silent, even) ADL breakage. (One remaining TODO in
compilerplugins/clang/external.cxx is that derived classes are not covered by
computeAffectedTypes, even though they could also be affected by ADL-breakage---
but don't seem to be in any acutal case across the code base.)
For friend declarations using elaborate type specifiers, like
class C1 {};
class C2 { friend class C1; };
* If C2 (but not C1) is moved into an unnamed namespace, the friend declaration
must be changed to not use an elaborate type specifier (i.e., "friend C1;"; see
C++17 [namespace.memdef]/3: "If the name in a friend declaration is neither
qualified nor a template-id and the declaration is a function or an
elaborated-type-specifier, the lookup to determine whether the entity has been
previously declared shall not consider any scopes outside the innermost
enclosing namespace.")
* If C1 (but not C2) is moved into an unnamed namespace, the friend declaration
must be changed too, see <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71882>
"elaborated-type-specifier friend not looked up in unnamed namespace".
Apart from that, to keep changes simple and mostly mechanical (which should help
avoid regressions), out-of-line definitions of class members have been left in
the enclosing (named) namespace. But explicit specializations of class
templates had to be moved into the unnamed namespace to appease
<https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92598> "explicit specialization of
template from unnamed namespace using unqualified-id in enclosing namespace".
Also, accompanying declarations (of e.g. typedefs or static variables) that
could arguably be moved into the unnamed namespace too have been left alone.
And in some cases, mention of affected types in blacklists in other loplugins
needed to be adapted.
And sc/qa/unit/mark_test.cxx uses a hack of including other .cxx, one of which
is sc/source/core/data/segmenttree.cxx where e.g. ScFlatUInt16SegmentsImpl is
not moved into an unnamed namespace (because it is declared in
sc/inc/segmenttree.hxx), but its base ScFlatSegmentsImpl is. GCC warns about
such combinations with enabled-by-default -Wsubobject-linkage, but "The compiler
doesn’t give this warning for types defined in the main .C file, as those are
unlikely to have multiple definitions."
(<https://gcc.gnu.org/onlinedocs/gcc-9.2.0/gcc/Warning-Options.html>) The
warned-about classes also don't have multiple definitions in the given test, so
disable the warning when including the .cxx.
Change-Id: Ib694094c0d8168be68f8fe90dfd0acbb66a3f1e4
Reviewed-on: https://gerrit.libreoffice.org/83239
Tested-by: Jenkins
Reviewed-by: Stephan Bergmann <sbergman@redhat.com>
E.g. #ifdef LIBO_INTERNAL_ONLY is always true for code that builds
with our PCHs.
Change-Id: I3cf311ea3621b909105754cfea2cb0116b8b67f5
Reviewed-on: https://gerrit.libreoffice.org/80961
Tested-by: Jenkins
Reviewed-by: Luboš Luňák <l.lunak@collabora.com>
Caolán McNamara