[PATCH] libstdc++: Optimize operator+(string/char*, string/char*) equally
Jonathan Wakely
jwakely@redhat.com
Wed Aug 24 23:02:55 GMT 2022
On Wed, 24 Aug 2022 at 23:47, Jonathan Wakely <jwakely@redhat.com> wrote:
>
> On Wed, 24 Aug 2022 at 23:39, Alexandre Oliva <oliva@gnu.org> wrote:
> >
> > On Aug 24, 2022, Jonathan Wakely via Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
> >
> > > * include/bits/basic_string.h (operator+(const string&,
> > > const char*)):
> > > Remove naive implementation.
> > > * include/bits/basic_string.tcc (operator+(const string&,
> > > const char*)):
> > > Add single-allocation implementation.
> >
> > ISTM this requires the following additional tweak:
> >
> > diff --git a/libstdc++-v3/src/c++11/string-inst.cc b/libstdc++-v3/src/c++11/string-inst.cc
> > index bfae6d902a1dd..2ec0e9d85f947 100644
> > --- a/libstdc++-v3/src/c++11/string-inst.cc
> > +++ b/libstdc++-v3/src/c++11/string-inst.cc
> > @@ -58,6 +58,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
> >
> > template class basic_string<C>;
> > template S operator+(const C*, const S&);
> > + template S operator+(const S&, const C*);
> > template S operator+(C, const S&);
> > template S operator+(const S&, const S&);
> >
> >
> > Without this, I'm getting undefined references to this specialization in
> > libstdc++.so, that I tracked down to a std::system_error ctor in
> > cxx11-ios_failure.o. I got this while testing another patch that might
> > be the reason why the template instantiation doesn't get inlined, but...
> > we can't depend on its being inlined, can we?
>
> Right. But adding that will cause another symbol to be exported,
> probably with the wrong symbol version.
Oh, but those symbols aren't exported ... they're just needed because
we compile with -fno-implicit-templatesand other symbols in
libstdc++.so require them.
So that probably is the right fix. I have another change for operator+
in mind now anyway, which optimizes operator(const string&, char) the
same way, and removes the duplication in those five operator+
overloads.
>
> To fix https://gcc.gnu.org/PR106735 I'm just going to revert it for
> now, and revisit in the morning.
More information about the Libstdc++
mailing list