Patch from Nicolas Pitre

This improves on what I did with patch #1362/1 by adding preloads for 
architectures that support it.  On an XScale PXA255 this provides a 20% 
performance gain.

Tested with all combinations of sizes and alignments.