According to Intel's recommendation, 'rep;nop; should be called before
testing if the lock variable was modified (i.e. rep nop;cmp;jcc). The
current implementation does it the wrong way around: first test, then
wait, then branch. I've asked Asit Mallik from Intel, and he recommended
to change it.

It should be at least consistent: Right now, spinlock uses
'cmp;rep nop;jcc', rwlock uses 'rep nop;cmp;jcc'