Jan,

> On x86 processors (perhaps others?), it seems possible you
> could write the cache probe code to use conditional move
> instructions instead of branches (which might be mispredicted).

IIRC, then conditional moves are not supported by all processors of the x86
family. This might cause a problem unless you specifically test for the
presence of conditional moves and I don't think todays C compilers are smart
enough to generate both, the conditional move and a "regular" version of the
code.

Cheers,
  - Andreas