std::barrier constructor is not constexpr

[jwakely]

See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118395 where this was first reported.

Everything below only applies when _LIBCPP_HAS_NO_TREE_BARRIER is undefined, but that seems to be always true? I'm not sure, see

llvm-project/libcxx/test/tools/clang_tidy_checks/internal_ftm_use.cpp

Lines 29 to 30 in 3def49c

	// TODO: Why does this macro even exist?
	"_LIBCPP_HAS_NO_TREE_BARRIER",

The std::barrier(ptrdiff_t, CompletionFunction) is supposed to be constexpr. It can't be, because it calls a non-inline function which allocates memory.

Also std::barrier<>(std::barrier<>::max()) should be valid, but max() just returns numeric_limits<ptrdiff_t>::max() and then the non-inline allocating function does:

llvm-project/libcxx/src/barrier.cpp

Lines 28 to 29 in 3def49c

	size_t const __count = (__expected + 1) >> 1;
	__state_ = unique_ptr<__state_t[]>(new __state_t[__count]);

so if __expected == max() then __expected + 1 overflows, with undefined behaviour.

Yay.

[jwakely]

CC @ogiroux

[jwakely]

C++20 semaphores were also non-constexpr (#49865) but that could be fixed by dropping the implementation using POSIX sem_t.

[jwakely]

I think the unique_ptr<__state_t[]> can be replaced by atomic<__state_t*> and then in the constructor, only call the non-inline initialization function if not is_constant_evaluated(). Then in the __barrier_base::arrive function do something like:

if (!__state.load(memory_order_relaxed)) [[unlikely]] {
  size_t const __count = (__expected + 1) >> 1;
  auto __p = make_unique<__state_t[]>(__count);
  __state_t* __val = nullptr;
  if (__state.compare_exchange_strong(__val, __p.get(), memory_order_seq_cst, memory_order_acquire))
    __p.release();
}

Maybe only do that when __old_phase == 0 so that on subsequent phases the relaxed load and null check can be skipped.

[ldionne]

Step 1 to clean this up: #122769