[libcxx] Reinstate <string.h> and fix overload sets to be const-correct wherever possible

[Richard Smith via cfe-commits]

Ping.

On Mon, Nov 23, 2015 at 6:55 PM, Richard Smith <richard at metafoo.co.uk>
wrote:

> Ping.
>
> On Thu, Nov 5, 2015 at 6:32 PM, Richard Smith <richard at metafoo.co.uk>
> wrote:
>
>> Ping.
>>
>> On Thu, Oct 29, 2015 at 5:21 PM, Richard Smith <richard at metafoo.co.uk>
>> wrote:
>>
>>> Hi,
>>>
>>> The attached patch undoes the revert of r249929, and adds an extension
>>> to allow <string.h> (and <wchar.h>) to work properly even in environments
>>> such as iOS where the underlying libc does not provide C++'s const-correct
>>> overloads of strchr and friends.
>>>
>>> This works as follows:
>>>
>>>  * The macro _LIBCPP_PREFERRED_OVERLOAD is, where possible, defined by
>>> <__config> to an attribute that provides the following semantics:
>>>    - A function declaration with the attribute declares a different
>>> function from a function declaration without the attribute.
>>>    - Overload resolution prefers a function with the attribute over a
>>> function without.
>>>  * For each of the functions that has a "broken" signature in C, if we
>>> don't believe that the C library provided the C++ signatures, and we have a
>>> _LIBCPP_PREFERRED_OVERLOAD, then we add the C++ declarations and mark them
>>> as preferred over the C overloads.
>>>  * The overloads provided in namespace std always exactly match those in
>>> ::.
>>>
>>>
>>> This results in the following changes in cases where the underlying libc
>>> provides the C signature not the C++ one, compared to the status quo:
>>>
>>>
>>> <string.h>:
>>>
>>>       char *strchr(const char*, int) // #1
>>>       char *strchr(char*, int) // #2
>>>       const char *strchr(const char*, int) // #3
>>>
>>> We used to provide #1 and #2 in namespace std (in <cstring>) and only #1
>>> in global namespace (in <string.h>).
>>>
>>> For a very old clang or non-clang compiler, we now have only #1 in both
>>> places (note that #2 is essentially useless). This is unlikely to be a
>>> visible change in real code, but it's slightly broken either way and we
>>> can't fix it.
>>>
>>> For newer clang (3.6 onwards?), we now have correct signatures (#2 and
>>> #3) in :: and std (depending on header). Taking address of strchr requires
>>> ~trunk clang (but it didn't work before either, so this is not really a
>>> regression).
>>>
>>>
>>> <wchar.h>:
>>>
>>>       wchar_t *wcschr(const wchar_t *, wchar_t) // #1
>>>       const wchar_t *wcschr(const wchar_t *, wchar_t) // #2
>>>       wchar_t *wcschr(wchar_t *, wchar_t) // #3
>>>
>>> We used to provide #1 in global namespace, and #2 and #3 in namespace
>>> std. This broke code that uses 'using namespace std;'.
>>>
>>> For a very old clang or non-clang compiler, we now have #1 in global
>>> namespace and namespace std. This fixes the ambiguity errors, but decreases
>>> const-correctness in this case. On the whole, this seems like an
>>> improvement to me.
>>>
>>> For newer clang, we now have correct signatures (#2 and #3) in :: and
>>> std (depending on header). As above, taking address doesn't work unless
>>> you're using very recent Clang (this is not a regression in ::, but is a
>>> regression in namespace std).
>>>
>>>
>>> To summarize, we previously had ad-hoc, inconsistent, slightly broken
>>> rules for <cstring> and <cwchar>, and with this patch we fix the overload
>>> set to give the exact C++ semantics where possible (for all recent versions
>>> of Clang), and otherwise leave the C signatures alone.
>>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20151208/f07d7f6e/attachment-0001.html>