lifetimebound attribute on std::move/std::forward

[higher-performance]

It would be great if the lifetimebound attribute could be added to std::move/std::forward for compilers that support it.

[philnik777]

Could you give some details on what this attribute does? i.e. specific examples where it finds bugs. I guess that's it's purpose?

[higher-performance]

Yeah. lifetimebound indicates that the lifetime of the argument is expected to extend beyond the function call itself. Here's an example of what it can catch:

#include <iostream>
#include <string>
#include <vector>

std::string get_default() { return "-"; }

template<class T>
std::remove_reference_t<T> &&safe_move(
#if defined(__clang__)
	[[clang::lifetimebound]]
#endif
	T &&value)
{
	return static_cast<std::remove_reference_t<T> &&>(value);
}

int main(int argc, char *argv[])
{
	std::vector<std::string> v(argv, argv + argc);
	// warning: temporary bound to local reference 's1'
	// will be destroyed at the end of the full-expression
	auto&& s1 = safe_move(v.empty() ? get_default() : v.back());
	std::cout << s1 << std::endl;
}

[alexander-shaposhnikov]

https://reviews.llvm.org/D141744

[zygoloid]

Perhaps this should be added to as_const and addressof too?

[philnik777]

Perhaps this should be added to as_const and addressof too?

Does this add any additional value? The rvalue overloads of these functions are already deleted.

[higher-performance]

That sounds like a great idea. const & can bind to r-values as well. If there are any other similar functions (e.g., std::forward_like? std::get?) it might be worth adding the same for them too.

[philnik777]

That sounds like a great idea. const & can bind to r-values as well. If there are any other similar functions (e.g., std::forward_like? std::get?) it might be worth adding the same for them too.

The rvalue overloads are specifically deleted. See https://en.cppreference.com/w/cpp/utility/as_const for example. I've already added the attribute to std::forward_like and std::move_if_noexcept.

[higher-performance]

Sorry, to clarify, I meant something like this is still possible.
I've had to do static_cast<const T &> in template code in the past, so it seems plausible to me that it would come up on rare occasions, though I definitely agree it wouldn't be too common.

[zygoloid]

Perhaps this should be added to as_const and addressof too?

Does this add any additional value? The rvalue overloads of these functions are already deleted.

The lvalue cases are still potentially problematic:

struct A {
  int n;
  int &get() [[clang::lifetimebound]] { return n; }
}
// Warns
int *p = &A().get();
// Does not warn without `[[lifetimebound]]`
int *q = std::addressof(A().get());

But Clang trunk automatically infers [[lifetimebound]] on addressof (as well as move, forward, etc.) these days, so I'm not sure there's any value in also adding it to libc++.

[higher-performance]

But Clang trunk automatically infers [[lifetimebound]] on addressof (as well as move, forward, etc.) these days, so I'm not sure there's any value in also adding it to libc++.

Oh wow, this was added just yesterday? 31b0be4

I guess that one can't cover non-Clang, and this one can't cover non-libc++?