The Itanium C++ ABI mandates that for a unique (complete) C++ type a single
unique symbol for the type's RTTI name is used across a process's dynamic
objects (so type equivalence can be determined via pointer comparison on the
RTTI names).
GCC nowadays deviates from that, using strcmp to determine equivalence, so it is
resilient to RTTI names being bound locally within dynamic objects (which has
performance benefits, but also makes it impossible to have unrelated types that
happen to have the same name "encapsulated" in individual dynamic objects---
whether or not that would violate the ODR would be open to interpretation of how
dynamic objects fit into the C++ Standard).
LLVM sticks to the Itanium ABI, which becomes notable in at least two places:
For one, libc++abi's __dynamic_cast uses strict checking. It still has a
_LIBCXX_DYNAMIC_FALLBACK for now that additionally uses strcmp checking and
syslogs visibility violations. Mac OS X uses libc++abi with
_LIBCXX_DYNAMIC_FALLBACK enabled, and running LO routinely logs dynamic_cast
errors to the Console there.
For another, RTTI-based UBSan checks unconditionally only use strict checking
(cf. isDerivedFromAtOffset in lib/ubsan/ubsan_type_hash.cc). This causes false
positives from Clang -fsanitize=function and -fsanitize=vptr even on Linux not
using libc++abi.
Therefore, introduce SAL_DLLPUBLIC_RTTI to mark types for which RTTI needs to
have default visibility under the Itanium/LLVM semantics. There is
unfortunately no way to mark only the (implicitly generated) RTTI symbols for
default visibility, but at least with the cases where SAL_DLLPUBLIC_RTTI is used
for now that is no real problem---any class type marked SAL_DLLPUBLIC_RTTI only
has inline (covered by -fvisibility-inlines-hidden) or undefined pure virtual
functions. It appears that even the vtables of those classes remain hidden, at
least with Mach-O on Mac OS X. (That also means there is no need for a
SAL_DLLPRIVATE_RTTI marker analoguous to the---also superfluous in retrospect---
CPPU_GCC_DLLPRIVATE one.)
Nevertheless, the number of exported symbols of course increases when
SAL_DLLPUBLIC_RTTI is "active." For a full-blown --enable-dbgutil build on Mac
OS X,
find instdir/LibreOffice.app/Contents -name \*.dylib\* -exec nm -gU {} \; \
wc -l
increased from 125541 to 139239. For Linux, an option might be to "activate"
SAL_DLLPUBLIC_RTTI only for __clang__ plus !ENABLE_RUNTIME_OPTIMIZATIONS.
The set of types marked SAL_DLLPUBLIC_RTTI with this patch (wholesale cppumaker-
generated UNO enum, struct, and interface types; plus some IEmbeddedHelper and
IUndoManager) is chosen so that a full "make check" on Mac OS X no longer
syslogs any dynamic_cast errors to the Console.
Change-Id: I42fa6ec01c2503ec24bcd9c0518abb112afa3235
Signed-off-by: Riccardo Magliocchetti <riccardo.magliocchetti@gmail.com>
Signed-off-by: Stephan Bergmann <sbergman@redhat.com>, undid one remove that was
detrimental to loplugin:unreffun
Change-Id: I18d8252084d828f94ef7a954e1dbfb45743d7970
This reverts commit 05050cdb23de586870bf479a9df5ced06828d498,
not all places that use e.g. OStringToOUString to convert potential UTF-8
are guaranteed to fulfil the prerequisites necessary to use fromUtf8 (and
some places like e.g. in codemaker are happy with the best-effort effect
of OStringToOUString's OSTRING_TO_OUSTRING_CVTFLAGS).
It turns out that almost none of them were necessary.
Change-Id: I1311ed28409c682b57ea8d149bcbaf2c49133e83
Reviewed-on: https://gerrit.libreoffice.org/12133
Reviewed-by: Noel Grandin <noelgrandin@gmail.com>
Tested-by: Noel Grandin <noelgrandin@gmail.com>
Find places where we are returning a pointer to something, where we can
be returning a reference.
e.g.
class A {
struct X x;
public X* getX() { return &x; }
}
which can be:
public X& getX() { return x; }
Change-Id: I796fd23fd36a18aedf6e36bc28f8fab4f518c6c7
after commit
c5d47c327a57df55
"add default value for Context param in uno::Exception constructors"
Change-Id: Ib429b58c559ecd48764a8df2f49d5470057c2c63
and all it's subtypes, which is almost never used, so this allows us to
simplify lots of call sites.
Change-Id: I0b05793ea2bdd1027679f63252d42ce4af89433b
Detect arguments larger than 64 chars passed by value.
Change-Id: I9b0ea9ccb99d115984a26eab67c9cf6afd5f6cae
Signed-off-by: Stephan Bergmann <sbergman@redhat.com>
It appears that the C++ standard allows overriding destructors to be marked
"override," but at least some MSVC versions complain about it, so at least make
sure such destructors are explicitly marked "virtual."
Change-Id: I0e1cafa7584fd16ebdce61f569eae2373a71b0a1
...mostly done with a rewriting Clang plugin, with just some manual tweaking
necessary to fix poor macro usage.
Change-Id: I71fa20213e86be10de332ece0aa273239df7b61a
