constexpr std::string_view::find_last_of doesn't work on clang 8 with libstdc++ 9

Viewed 501

The following code compiles under g++ 9, with the flag -std=c++17, but not clang 8 with the same flag:

#include <string_view>
#include <cstdlib>


int main() {

        constexpr std::string_view hello = "hello";
        constexpr size_t last_l = hello.find_last_of("lo");
}

The error message is as follows:

test.cpp:8:19: error: constexpr variable 'last_l' must be initialized by a constant expression
        constexpr size_t last_l = hello.find_last_of("lo");
                         ^        ~~~~~~~~~~~~~~~~~~~~~~~~
/usr/bin/../lib/gcc/x86_64-linux-gnu/9/../../../../include/c++/9/bits/char_traits.h:349:9: note: cast from 'void *' is not allowed in a constant expression
        return static_cast<const char_type*>(__builtin_memchr(__s, __a, __n));
               ^
/usr/bin/../lib/gcc/x86_64-linux-gnu/9/../../../../include/c++/9/bits/string_view.tcc:150:12: note: in call to 'find(&"lo"[0], 2, "hello"[4])'
              if (traits_type::find(__str, __n, this->_M_str[__size]))
                  ^
/usr/bin/../lib/gcc/x86_64-linux-gnu/9/../../../../include/c++/9/string_view:402:22: note: in call to '&hello->find_last_of(&"lo"[0], 18446744073709551615, 2)'
      { return this->find_last_of(__str, __pos, traits_type::length(__str)); }
                     ^
test.cpp:8:34: note: in call to '&hello->find_last_of(&"lo"[0], 18446744073709551615)'
        constexpr size_t last_l = hello.find_last_of("lo");
                                        ^
test.cpp:10:16: error: static_assert expression is not an integral constant expression
        static_assert(last_l == 3);
                      ^~~~~~~~~~~
test.cpp:10:16: note: initializer of 'last_l' is not a constant expression
test.cpp:8:19: note: declared here
        constexpr size_t last_l = hello.find_last_of("lo");
                         ^
2 errors generated.

Looking at the error message, it looks like it calls std::char_traits<char>::find, which according to cppreference should be constexpr and not a problem. However, looking at the implementation in char_traits.h, while it it marked constexpr, it does not seem to follow constexpr rules:

static _GLIBCXX17_CONSTEXPR const char_type*
find(const char_type* __s, size_t __n, const char_type& __a)
{
  if (__n == 0)
    return 0;
#if __cplusplus >= 201703L
  if (__builtin_constant_p(__n)
      && __builtin_constant_p(__a)
      && __constant_char_array_p(__s, __n))
    return __gnu_cxx::char_traits<char_type>::find(__s, __n, __a);
#endif
  return static_cast<const char_type*>(__builtin_memchr(__s, __a, __n));
}

However, g++ does not seem to have a problem with that, even though it seems from the clang diagnostic message __builtin_memchr returns a void*, which is then casted to a char*, which seems to be disallowed (see bullet 14).

Following this same train of thought, the following sample also compiles with g++ but not clang:

#include <string>

int main() {
        constexpr const char* asdf = "asdf";
        constexpr const char* s_ptr = std::char_traits<char>::find(asdf, 2, 's');
}

With the error message:

test_traits.cpp:5:24: error: constexpr variable 's_ptr' must be initialized by a constant expression
        constexpr const char* s_ptr = std::char_traits<char>::find(asdf, 2, 's');
                              ^       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/bin/../lib/gcc/x86_64-linux-gnu/9/../../../../include/c++/9/bits/char_traits.h:349:9: note: cast from 'void *' is not allowed in a constant expression
        return static_cast<const char_type*>(__builtin_memchr(__s, __a, __n));
               ^
test_traits.cpp:5:32: note: in call to 'find(&"asdf"[0], 2, 's')'
        constexpr const char* s_ptr = std::char_traits<char>::find(asdf, 2, 's');
                                      ^
1 error generated.

This seems like both a stdc++ bug, by not having std::char_traits<_>::find be properly constexpr, and a g++ bug, for compiling it.

Is there something I'm missing? Also, is there a workaround? I came across this trying to use magic enum, and would like to use it in a clang environment.

Update (6/7/19)

As pointed out by @cpplearner, it seems like clang should be taking the return __gnu_cxx::char_traits<char_type>::find(__s, __n, __a); Interestingly enough, that doesn't seem to be the problem. The following code, which is the exact same function from std::char_traits<char>::find, except without that constant branch, still compiles fine in g++, which makes me wonder if g++ is even taking that path:

#include <cstddef>

static constexpr  const char*
find_noconst(const char* __s, size_t __n, const char& __a)
{
        if (__n == 0)
                return 0;
        return static_cast<const char*>(__builtin_memchr(__s, __a, __n));
}

int main() {
        constexpr const char* asdf = "asdf";
        constexpr const char* s_ptr = find_noconst(asdf, 2, 's');
}

This still compiles in g++, and fails in clang with the same error.

Update 2 (6/7/19)

More research confirms @cpplearner comment:

#include <cstddef>
#include <string>
#include <iostream>

static constexpr  const char*
branch_test(const char* __s, size_t __n, const char& __a)
{
        if (__n == 0)
                return 0;
        if (__builtin_constant_p(__n)
                        && __builtin_constant_p(__a)
                        && std::__constant_char_array_p(__s, __n))
                return "took branch";

        return "did not take branch";

}

int main() {
        constexpr const char* test = branch_test("asdf", 2, 's');

        std::cout << test << std::endl;

}

With gcc, this prints took branch, but with clang it prints did not take branch.

Refining it more:

#include <string>
#include <iostream>

int main() {
        constexpr bool test = std::__constant_char_array_p("test", 2);

        std::cout << test << std::endl;
}

Prints 0 on clang and 1 on g++.

After digging into that implementation, it seems that this is actually a clang bug. The most minimal example of difference I could find was this:

#include <iostream>

constexpr bool is_const_char(const char* a, size_t idx) {
        return __builtin_constant_p(a[idx]);
}

int main() {
        constexpr bool test = is_const_char("test", 0);

        std::cout << test << std::endl;
}

Prints 0 on clang and 1 on gcc. The most interesting part to me, is what actually works:

#include <iostream>

int main() {
        constexpr const char* str = "test";
        constexpr bool test1 = __builtin_constant_p(str[0]);
        constexpr bool test2 = __builtin_constant_p("test"[0]);


        std::cout << test1 << ", " << test2 << std::endl;
}

That program consistently prints 1, 1 on both clang and g++.

0 Answers
Related