A lot of one- or two-count arena frees occur in rapid succession while the current arena is only about ~40% utilization. This happens as part of our normal build procedure, and relates to bug 558971.

The fundamental difficulty here is that thrashing can occur at any stride -- if you account for a n-arena-allocation followed by n-arena-deallocation pattern, you are always foiled by the regexp that requires n+1 arenas each execution. Because of this, you can use time locality to let the arenas sit and age until the next garbage collection. For imacro_asm, this seems particularly useful in regular expression evaluation. For our arena implementation we can add a "soft release" mechanism that sets the |avail| pointers back to their base values. I'm also going to go nuts and say we can spring for a |prev| pointer in the arena structure to facilitate this, especially given that we're going for page-sized arena chunks nowadays.

Resolving with a dependency on bug 564060.