why is CONFIG_FRAME_POINTER=y so expensive?

Greetings,

tbench 8 throughput on my i4790 box:
1 2 3 avg

master                  3769.95  3759.28  3762.83  3764.02  1.000
master-framepointer     3476.73  3453.52  3460.62  3463.62   .920

Does anyone know why the performance impact is this large?

The only thing I found googling was a reference to 3 arg lea using rbp or r13 being slow, which wasn't particularly helpful. Nor was staring at asm deltas with non-expert eyeballs.

-Mike