Spinlocks for GC on multiprocessor PCs?

[dsimcha]

Since resource contention for memory allocations seems to be a big issue in D,
at least for me, and memory allocation is usually pretty fast, would it make
sense to make the lock used for GC.malloc a spinlock?  Since memory allocation
takes very little time relative to a full timeslice, there is likely to be
very little contention for malloc on a single-processor machine.  This means
that a thread will very seldom be left spinning until its timeslice is up.  I
believe (correct me if I'm wrong) that the kernel call, etc. for a full lock
takes longer than a memory allocation from the GC pool.  This makes the
locking in truly parallel multiprocessor code a significant bottleneck.  Of
course, if the GC needs to free memory, not just allocate from its available
pool, the spinlock should ideally inflate to a full lock.  Does this sound
reasonable/implementable?