I observed in some profiles that calls from Mutex::lock() to Mutex::lock(Thread *) are not being inlined, at least not by gcc on Linux. By outlining the contended loop, we see better inlining and less instructions spent when using mutexes in hotspot[1], at little to no static footprint cost.

A more thorough examination should be considered: MutexLocker could possibly inline through a call to Mutex::try_lock, which reduce the uncontended fast-path to a call down to the platform primitive, e.g., pthread_mutex_trylock.

[1] http://cr.openjdk.java.net/~redestad/scratch/mutex_contended.00/