Rather than allocating a frame on every call, we keep an already allocated list
of frames ready to go. This trades lightly more memory usage for faster calls.