Improving ticket spinlocks

How expensive are inter-processor interrupts? If they're cheaper than 50 cycles and a cache-line bounce then the thing to do would be to have the lock owner CPU look at the next ticket in line and send it an interrupt while the waiting CPU just sits in an idle. From what I could pick up in a fast Google an IPI is about the same as a cache line bounce, it is the register saves and all that make it more expensive. But since the waiting CPU is in a known state the code could probably skip all that and just resume with the lock acquire.