I really wish the idea that AF has no performance impact would die because it DOES have an effect on performance.
The main ressource used by AF is texture bandwith. AF works by doing many more sampling than trilinear filtering so that's more bandwith and potential cache thrashing for bigger textures.
When a shader samples a texture, there is a latency before the shader can actually use the result. To avoid having the GPU doing nothing while waiting for a texture sample, shader compilers will fill the gap with arithmetic operations from independant code branches if there are any available.
What it means is that if there is no arithmetic operations to be processed in between texture samples, the GPU will have to wait for them. That's what you would call being "texture bound". In these cases, if you add more samples (like AF does), GPU time will increase.
What people have to understand is that this is highly engine/shader dependant. In an engine where most shaders are ALU bound, adding more texture samples won't impact performance at all because there will always be stuff to do during the wait, but in the case where you are texture bound, it will.
Another thing to note is that PC shader compilers are pretty shitty, resulting in much less ALU optimization, so that's also more arithmetic operations to hide latency with.
That being said, there is nothing magic with implementing AF on PS4. It's only a flag to setup on the samplers like on DirectX. So that's only a dev's decision/mistake if it's missing in a game.