There were two ways this could fail:

1. division by zero when sysconf(_SC_PAGESIZE)==0

2. mmap() failure because the size parameter is not aligned to page
   size

Neither ever happened: sysconf() never fails, and the only caller
passes a size that is already aligned.  Phew.

Necessary to preserve type information.  The try/catch sequence didn't
work previously.

Same fix as in commit 1c904000

Similar to boost::scope_exit, but fewer include dependencies.

Some optimized implementations of memcmp() may not start from the
beginning of the string, and may thus segfault.