Thanks for this, both __resolver_fn and marking the clones unsafe is a big improvement.

I think you're right about Relaxed ordering, especially since the pointer can only be one of two values and either are correct at any point in time.

After sleeping on it I realized the reasoning behind one of those wrappers. After my commit, it is not possible to write this code:

#[target_clones("[x86|x86_64]+avx")]
pub fn bad_unsafe() {
    let x: Vec<u32> = std::mem::uninitialized();
}

I'll make a new PR that re-introduces one of the wrappers without causing worse generated output.

@TethysSvensson I think the is_x86_feature_detected! macros of the standard library could benefit from a similar optimization to remove a branch. If you enjoy this a good place to start would be the rust-lang/stdarch repository, and in there, the std_detect sub crate.

It uses an AtomicUXY as a bitset, and there is a function that checks whether the bitset has already been initialized every time that one attempts to detect a feature, and if it isn't, it then initializes it. It should probably be possible to replace that function with a function pointer that, on the first call, initializes the cache, and then modifies the function pointer to one that does not do initialization, such that this is then always called all the time.