New std::string implementation

On 28/11/14 15:24 +0000, Jonathan Wakely wrote:
>Tested x86_64-linux with:
>
>  --target_board=unix\{-m32,\}\{,-D_GLIBCXX_USE_CXX11_ABI=0\}
>
>Also tested on powerpc64-linux (where this patch doesn't actually
>bootstrap due to PR 63573, there's a patch in the PR that's needed)
>and i686-linux.

Oh how silly, the attached patch is needed for powerpc. I had fixed
this once, but must have lost the change on the compile farm machine I
was using and then not committed the fix in my local repo. The
attached patch also includes the PR63573 fix which isn't committed
yet.

I've committed the libstdc++ parts (but not the PR fix)  to my repo
now, so it won't get lost again.

The powerpc64 ICE is fixed, so I'm committing the std::string ABI
transition patch.

This replaces our venerable Copy-on-Write std::string with a
C++11-conforming* Small-String Optimized std::__cxx11::string.

(* It's not quite 100% conforming, as it's missing some allocator
features.  Of course. Always with the allocators. But that's only a
small fix to make next week.)

Since the last version of this patch Jason has added support for
attribute((abi_tag)) on namespaces, so the new types can be defined in
an inline namespace that has the tag, rather than co-existing in
namespace std with the un-tagged types. This makes overloading and
exporting symbols a bit more straightforward as the different linkage
names can be expressed with a fully-qualified name without needing the
attribute.

I've also added regex templates to the new __cxx11 namespace, as they
depend on std::string.

There are now two new configure options, --disable-libstdcxx-dual-abi
and --with-default-libstdcxx-abi. The former disables all support for
the new __cxx11 ABI pieces. The latter is only relevant when the dual
ABI is enabled and can be used to make the old definitions the
default ones you get without defining the _GLIBCXX_USE_CXX11_ABI
macro.

As described previously, the locale facets that depend on the
std::string ABI exist twice as "twinned facets" and if users replace
a twinned facet the library generates a shim to replace its twin.

Also as described previously, the types in <stdexcept> continue to use
the COW string, to ensure nothrow copying and ABI stability.

The Python pretty printer problems I was seeing with the new string
turned out to be a GDB bug (17728).

Target maintainers will see a *lot* of new exports at the latest
symbol version if they generate a new baseline-symbols.txt file. I
suggest waiting and doing that nearer the end of stage 3 in case there
are any fixes needed after this change.

Bootstrapped and tested in various configurations on x86_64-linux and
powerpc64-linux.

Committed to trunk!

Phew.

Jonathan Wakely <jwakely@redhat.com> writes:

> The powerpc64 ICE is fixed, so I'm committing the std::string ABI
> transition patch.
>
> This replaces our venerable Copy-on-Write std::string with a
> C++11-conforming* Small-String Optimized std::__cxx11::string.
>
> (* It's not quite 100% conforming, as it's missing some allocator
> features.  Of course. Always with the allocators. But that's only a
> small fix to make next week.)
[...]
> Target maintainers will see a *lot* of new exports at the latest
> symbol version if they generate a new baseline-symbols.txt file. I
> suggest waiting and doing that nearer the end of stage 3 in case there
> are any fixes needed after this change.
>
> Bootstrapped and tested in various configurations on x86_64-linux and
> powerpc64-linux.

Unfortunately, this patch broke Solaris bootstrap with /bin/ld:
libstdc++.so fails to link with

ld: fatal: libstdc++-symbols.ver-sun: 5383: symbol 'std::locale::name[abi:cxx11]() const': symbol version conflict

l.5383 has

    ##_ZNKSt6locale4nameB5cxx11Ev (glob)
    _ZNKSt6locale4nameB5cxx11Ev;

i.e.

    # std::locale::name() returning new std::string

in GLIBCXX_3.4.21 vs.

      ##std::locale::[A-Zn-z]* (cxx)
      _ZNKSt6locale4nameB5cxx11Ev;

in GLIBCXX_3.4.

	Rainer