std::string add nullptr attribute

[Gabriel Ravier]

On 2/20/23 11:54, Xi Ruoyao via Gcc-help wrote:
> On Mon, 2023-02-20 at 10:37 +0000, Jonathan Wakely wrote:
>> On Mon, 20 Feb 2023 at 10:26, Xi Ruoyao <xry111@xry111.site> wrote:
>>> On Sun, 2023-02-19 at 21:33 +0000, Jonny Grant wrote:
>>>
>>>> I noticed -Wanalyzer-null-dereference reports at build time a
>>>> dereference. Also works if a function parameter. I wondered why
>>>> std::string isn't detected by this static analyser option.
>>> Because the analyzer does not know the C++ standard disallows to use
>>> NULL here.  It just analyzes the code.  The code in libstdc++ reads:
>>>
>>>        basic_string(const _CharT* __s, const _Alloc& __a = _Alloc())
>>>        : _M_dataplus(_M_local_data(), __a)
>>>        {
>>>          // NB: Not required, but considered best practice.
>>>          if (__s == 0)
>>>            std::__throw_logic_error(__N("basic_string: "
>>>                                         "construction from null is not valid"));
>>>          const _CharT* __end = __s + traits_type::length(__s);
>>>          _M_construct(__s, __end, forward_iterator_tag());
>>>        }
>>>
>>> As you can see yourself, though the standard implies using NULL here is
>>> a UB, libstdc++ does not really code a UB here.  So the analyzer will
>>> consider the code absolutely valid.
>> Right, it's defined behaviour in libstdc++, as an extension.
>>
>>> Note that throwing a C++ exception is not a programming error.  It's
>>> perfectly legal to catch the exception elsewhere.  It's also perfectly
>>> legal not to catch it and treat it as an abort() (calling abort is also
>>> not a programming error).
>>>
>>>
>>>> It's not pretty, but this wrapper catches NULL passed at compile time:
>>>>
>>>> std::string make_std_string(const char * const str)
>>>> {
>>>>      // This line ensures: warning: dereference of NULL '0' [CWE-476]
>>>> [-Wanalyzer-null-dereference]
>>>>      char b = *str;
>>> You are invoking an undefined behavior here if str is NULL, so it's
>>> essentially same as using a nonnull attribute for make_std_string.
>> And turned defined behaviour back into UB. The warning isn't reliable
>> (only if the compiler can see the point is null, which isn't the case
>> without optimization, or if the pointer comes from some non-inline
>> function), the exception is. You're trading guaranteed exception for a
>> not guaranteed warning and unbounded misoptimization due to undefined
>> behaviour.
> Well, maybe we should have a warning here with -Wpedantic (or something)
> as the standard does not allow people to pass NULL and expect a
> logic_error.  But "deliberately making a UB to raise the warning" is not
> good.

This is the kind of thing that makes me wonder why there isn't some kind 
of `__builtin_unreachable_do_not_optimize()` builtin that allows one to 
mark places in code that should never be reached and should thus be 
warned about if such a thing happens while at the same time never doing 
any optimization on the basis of the presence of the call.