Is `string.assign(string.data(), 5)` well-defined or UB?

Question

A coworker wanted to write this:

std::string_view strip_whitespace(std::string_view sv);

std::string line = "hello  ";
line = strip_whitespace(line);

I said that returning string_view made me uneasy a priori, and furthermore, the aliasing here looked like UB to me.

I can say with certainty that line = strip_whitespace(line) in this case is equivalent to line = std::string_view(line.data(), 5). I believe that will call string::operator=(const T&) [with T=string_view], which is defined to be equivalent to line.assign(const T&) [with T=string_view], which is defined to be equivalent to line.assign(line.data(), 5), which is defined to do this:

Preconditions: [s, s + n) is a valid range.
Effects: Replaces the string controlled by *this with a copy of the range [s, s + n).
Returns: *this.

But this doesn't say what happens when there's aliasing.

I asked this question on the cpplang Slack yesterday and got mixed answers. Looking for super authoritative answers here, and/or empirical analysis of real library vendors' implementations.

---

I wrote test cases for string::assign, vector::assign, deque::assign, list::assign, and forward_list::assign.

• Libc++ makes all of these test cases work.
• Libstdc++ makes them all work except for forward_list, which segfaults.
• I don't know about MSVC's library.

The segfault in libstdc++ gives me hope that this is UB; but I also see both libc++ and libstdc++ going to great effort to make this work at least in the common cases.

Answer 1

Barring a couple of exceptions of which yours is not one, calling a non-const member function (i.e. assign) on a string invalidates [...] pointers [...] to its elements. This violates the precondition on assign that [s, s + n) is a valid range, so this is undefined behavior.

Note that string::operator=(string const&) has language specifically to make self-assignment a no-op.