[cfe-dev] libc++ supporting older compilers

[Eric Fiselier via cfe-dev]

On Mon, Dec 21, 2015 at 9:22 PM, Chandler Carruth <chandlerc at google.com>
wrote:

> Relatedly, I think you could require a C++11 implementation to build
> libc++ itself and merely support C++03 parsing of the headers (and
> subsequent linking of that object code with the library). That might allow
> you to maintain a single ABI of the built library even as you switch
> between C++03 and C++11 usage of the library.
>


This is the status-quo. You have always needed C++11 to build libc++ and
dylib produced is ABI compatible between Clang C++03/C++11 and GCC C++11.
GCC in C++03 doesn't provide char16_t and char32_t and this causes link
errors in locale code.


FWIW, for the few users I support who care about C++03 mode, it would
> actually be a significant feature to turn off all C++11 emulation in libc++
> and to present as close to a pure C++03 environment as possible. Those
> users have *some* environments they support which are functionally C++03
> and they are setting the C++03 mode to try and detect problems sooner when
> building and testing with the nice modern Clang+libc++ (or GCC+libc++)
> toolchain they have in their main development environment. I wonder if
> ceasing to emulate so much of C++11 library features (and language
> features!!! nullptr for example!) when building in C++03 would help with
> this.
>


I'm sympathetic to this case for C++11 onward. For newer dialects we
provide a mostly conforming implementation. However Libc++ was initially
developed to the C++11 standard and meaningful C++03 conformance is not in
the cards. We internally depend on C++11 library classes and headers when
implementing C++03. This is a fundamental part of the library design and
shouldn't be change. (The "#define" nullptr needs to die though).
If users want to find conformance issues in C++03 mode they should use a
library that implements the C++03 spec.

IMO offering C++11 library extensions in C++03 is an important part of
libc++ that users depend on.
My concern is that some of the emulation was never intended to be
permanent. Instead it came along with the need to support bad
C++0x compilers that would some day go away. However C++03 is not going
away soon.
I fear that supporting this emulation in perpetuity is detrimental to the
overall maintainability of libc++.

However much of the "language-emulation" is needless for Clang users
because Clang provides the required language features as extensions.
GCC in C++03 is the only configuration that would continue to need
"emulation". Dropping or limiting support for GCC in C++03 might be
unrealistic
but it would also be very beneficial to libc++.

Anyway, I feel I've been too hasty in raising this issue and I don't plan
to keep pushing it.

/Eric



> On Mon, Dec 21, 2015 at 5:32 PM Eric Fiselier via cfe-dev <
> cfe-dev at lists.llvm.org> wrote:
>
>> I think we have found consensus, or at least mild agreement, that nobody
>> needs libc++ to support GCC 4.6, 4.7 or 4.8.
>> Unless there are any late objections I'm going to proceed with
>>
>> * Requiring GCC 4.9+ to use libc++ in C++11 mode.
>>
>> While GCC 4.9 may not be optimal for everybody it is a compiler we can
>> meaningfully support.
>> Users of GCC 4.9+ should expect no test failures in C++11 mode and I will
>> set up a buildbot to enforce this.
>>
>> This new requirement also has an exciting implication: Libc++ no longer
>> has to configure for incomplete C++11 implementations.**
>> This means that libc++ no longer has to restrict its use of C++11
>> features such as "constexpr" and alias templates in C++11 mode.
>>
>> However I still don't know what to do about GCC in C++03 mode, which
>> still has over 600 test failures. But that is a separate
>> question and deserves it's separate thread. Expect to see a new thread in
>> cfe-dev tonight.
>>
>> /Eric
>>
>> ** I'll triple check this before I act on it.
>>
>> On Mon, Dec 21, 2015 at 4:49 PM, Eric Fiselier <eric at efcs.ca> wrote:
>>
>>>
>>>
>>> On Mon, Dec 21, 2015 at 8:19 AM, Craig, Ben via cfe-dev <
>>> cfe-dev at lists.llvm.org> wrote:
>>>
>>>> On 12/19/2015 2:44 PM, Renato Golin wrote:
>>>>
>>>>> On 18 December 2015 at 14:30, Craig, Ben via cfe-dev
>>>>> <cfe-dev at lists.llvm.org> wrote:
>>>>>
>>>>>> Ubuntu 12.04 LTS (precise), released Apr-26-2012: GCC 4.6.3
>>>>>> Ubuntu 14.04 LTS (trusty), released Apr-17-2014: GCC 4.8.2
>>>>>>
>>>>> Previous LTS is deprecated when a new LTS comes along, so no one
>>>>> should have to worry about 12.04 in this day and age.
>>>>>
>>>>> Is there any one that this is important?
>>>>
>>>>
>>> These configurations break enough that I doubt anybody is depending on
>>> them. I think I would have seen more bug reports if people were trying to
>>> use it.
>>>
>>> This is also the entirely wrong question to ask. What's important is
>>> that Libc++ is allowed to required C++11, and C++11 implementation should
>>> work.
>>> People who insist on using a ToT STL with an old compiler puzzle me. You
>>> can't expect new library features without additional language support.
>>>
>>> /Eric
>>>
>>>
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20151222/b793982e/attachment.html>