I'm getting race conditions when using the stl wide char library functions. Here's a minimal reproducible example in c, where I mention the compilers I've tried, the library version, and the expected and received outputs:

// temp.cpp

#include <wchar.h>
#include <assert.h>
#include <locale.h>

int main() {
    setlocale(LC_ALL, "");
    char utf8[] = {'a'};
    wchar_t output[1];
    mbstate_t ps;
    size_t ret = mbrtowc(output, utf8, 1, &ps);
    assert(ret == 1);
    assert(output[0] == L'a');
    return 0;

    // let me define BAD output as:

        // most of the time (~9/10), I get the following error:
        //      a.out: ../iconv/loop.c:457: utf8_internal_loop_single: Assertion `inptr - bytebuf > (state->__count & 7)' failed.
        //      Aborted (core dumped)

        // the other (~1/10 times) the program ends correctly (exit code 0)

    // and let me define GOOD output as:

        // the program ends correctly (exit code 0), always
    
    // compiler     output
    //
    // gcc-7        BAD
    // gcc-9        BAD
    // gcc-10       BAD
    // clang-10     GOOD
    // clang-12     GOOD

    // compile command: gcc temp.cpp
    //                  clang temp.cpp


    // gcc (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
    // glibc version: 2.31
}

My questions are as follow:

• Can anyone else reproduce this, or is it just me?
• What is the meaning of the assertion being tripped?
• Why is this happening?

Edit: ah, c default ctors have spoiled me. (also, in my original source I was working with std::mbrtowc, which threw me off.

Introspection: I will indeed use valgrind more in these cases.

CodePudding user response：

You forgot to initialize ps.

man mbrtowc:

In all of the above cases, if ps [fourth argument] is NULL, a static anonymous state known only to the mbrtowc() function is used instead. Otherwise, *ps must be a valid mbstate_t object. An mbstate_t object a can be initialized to the initial state by zeroing it, for example using

       memset(&a, 0, sizeof(a));

Your object ps is uninitialized and therefore cannot in general be a valid mbstate_t object.

valgrind detects this bug.