Is glibc_file::needs_flush correct

[YohnWang]

I find this->file_->_IO_write_ptr - this->file_->_IO_write_end in arm and aarch64 will be negative some times.
if I does'nt define NDEBUG, to_unsigned will assert failed and abort.

[vitaut]

Could you provide a repro?

[YohnWang]

It may my code bug.
This problem is caused by setvbuf, ISO C says setvbuf must be called before all other functions to operate FILE *.
the minimal code to reappear is

int main()
{
    printf("111\n");
    setvbuf(stdout, nullptr, _IOLBF, 0);
    fmt::print(stdout, "------\n");
}

and then run this code as

./a.out > /dev/null

the actual situation is complex, we use two shared libraries, the one is

// libsetlinebuf.so
__attribute__((constructor(101))) static void setlinebuf()
{
    setvbuf(stdout, nullptr, _IOLBF, 0);
}

else is an extern shared library, which use like

__attribute__((constructor)) static void print()
{
    printf("a string\n");
}

the main is

int main()
{
    fmt::print(stdout, "------\n");
}

run as

LD_PRELOAD=$PWD/libsetlinebuf.so ./a.out > /dev/null

the function print called before setlinebuf, so it violation of ISO standards

but if I use printf (or fmt::println, or unuse redirect, or call fflush(stdout) after setvbuf), the program will run correct.

should we support this situation

[vitaut]

Repro: https://www.godbolt.org/z/x7PjxPb4c

[vitaut]

Fixed in 300ce75. Thanks for reporting.