Initialize std::string from a possibly NULL char pointer

Question

Initializing std::string from a NULL char pointer is undefined behaviour, I believe. So, here are alternative versions of a constructor, where mStdString is a member variable of type std::string:

void MyClass::MyClass(const char *cstr) :
    mStdString( cstr ? cstr : "")
{}

void MyClass::MyClass(const char *cstr) :
    mStdString(cstr ? std::string(cstr) : std::string())
{}

void MyClass::MyClass(const char *cstr)
{
    if (cstr) mStdString = cstr;
    // else keep default-constructed mStdString
}

Edit, constructor declaration inside class MyClass:

MyClass(const char *cstr = NULL);

Which of these, or possibly something else, is the best or most proper way to initialize std::string from a possibly NULL pointer, and why? Is it different for different C++ standards? Assume normal release build optimization flags.

I'm looking for an answer with explanation of why a way is the right way, or an answer with a reference link (this also applies if answer is "doesn't matter"), not just personal opinions (but if you must, at least make it just a comment).

Answer 1

The last one is silly because it doesn't use initialization when it could.

The first two are completely identical semantically (think of the c_str() member function), so prefer the first version because it is the most direct and idiomatic, and easiest to read.

(There would be a semantic difference if std::string had a constexpr default constructor, but it doesn't. Still, it's possible that std::string() is different from std::string(""), but I don't know any implementations that do this, since it doesn't seem to make a lot of sense. On the other hand, popular small-string optimizations nowadays mean that both versions will probably not perform any dynamic allocation.)

---

Update: As @Jonathan points out, the two string constructors will probably execute different code, and if that matters to you (though it really shouldn't), you might consider a fourth version:

: cstr ? cstr : std::string()

Both readable and default-constructing.

---

Second update: But prefer cstr ? cstr : "". As you can see below, when both branches call the same constructor, this can be implemented very efficiently using conditional moves and no branches. (So the two versions do indeed generate different code, but the first one is better.)

---

For giggles, I've run both versions through Clang 3.3, with -O3, on x86_64, for a struct foo; like yours and a function foo bar(char const * p) { return p; }:

Default constructor (std::string()):

    .cfi_offset r14, -16
    mov     R14, RSI
    mov     RBX, RDI
    test    R14, R14
    je      .LBB0_2
    mov     RDI, R14
    call    strlen
    mov     RDI, RBX
    mov     RSI, R14
    mov     RDX, RAX
    call    _ZNSt3__112basic_stringIcNS_11char_traitsIcEENS_9allocatorIcEEE6__initEPKcm
    jmp     .LBB0_3
.LBB0_2:
    xorps   XMM0, XMM0
    movups  XMMWORD PTR [RBX], XMM0
    mov     QWORD PTR [RBX + 16], 0
.LBB0_3:
    mov     RAX, RBX
    add     RSP, 8
    pop     RBX
    pop     R14
    ret

Empty-string constructor (""):

    .cfi_offset r14, -16
    mov     R14, RDI
    mov     EBX, .L.str
    test    RSI, RSI
    cmovne  RBX, RSI
    mov     RDI, RBX
    call    strlen
    mov     RDI, R14
    mov     RSI, RBX
    mov     RDX, RAX
    call    _ZNSt3__112basic_stringIcNS_11char_traitsIcEENS_9allocatorIcEEE6__initEPKcm
    mov     RAX, R14
    add     RSP, 8
    pop     RBX
    pop     R14
    ret

.L.str:
    .zero    1
    .size    .L.str, 1

In my case, it would even appear that "" generates better code: Both versions call strlen, but the empty-string version doesn't use any jumps, only conditional moves (since the same constructor is called, just with two different arguments). Of course that's a completely meaningless, non-portable and non-transferable observation, but it just goes to show that the compiler doesn't always need as much help as you might think. Just write the code that looks best.