Version 1.9.15, 2011-03-21
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

* First release candidate for 1.10.0

* Modify how message expansion is done in SHA-256 and SHA-512.
  Instead of expanding the entire message at the start, compute them
  in the minimum number of registers. Values are computed 15 rounds
  before they are needed. On a Core i7-860, GCC 4.5.2, went from 143
  to 157 MiB/s in SHA-256, and 211 to 256 MiB/s in SHA-512.

* Pipe will delete empty output queues as soon as they are no longer
  needed, even if earlier messages still have data unread. However an
  (empty) entry in a deque of pointers will remain until all prior
  messages are completely emptied.

* Avoid reading the SPARC ``%tick`` register on OpenBSD as unlike the
  Linux and NetBSD kernels, it will not trap and emulate it for us,
  causing a illegal instruction crash.

* Improve detection and autoconfiguration for ARM processors. Thanks
  go out to the the `Tahoe-LAFS Software Foundation
  <http://tahoe-lafs.org>`_, who donated a Sheevaplug that I'll be
  using to figure out how to make the cryptographic primitives
  Tahoe-LAFS relies on faster, particularly targeting the ARMv5TE.