..this turned out to be more complicated than I expected.

my compiler seems to generate 40.  are you using clang?

instead of 128-bit cmpxchg16.  minor speedup, even though the profiler said
that cmpxchg16 was the top offender.

get rid of gc_epoch_lock.

to help measure just the map/unmap cost.

the kernel spends about 35% of its time in mfence.
commented-out #define to get rid of __sync_synchronize for testing.
maybe having many atomics is going to be expensive.