Attached is the latest version of my "pure C" (modulo aliasing issues)
memcpy implementation. Compiled with -O3 on arm, it matches the
performance of the assembly language memcpy from Bionic for aligned
copies, and is only 25% slower than the asm for misaligned copies. And
it's only mildly larger. It uses the same principle as the Bionic
code: large block copies as aligned 32-bit units for aligned copies,
and aligned-load, bitshift-then-or, aligned-store for misaligned
copies. This should, in principle, work well on typical risc archs
that have plenty of registers but no misaligned load or store support.

Unfortunately it only works on little-endian (I haven't though much
yet about how it could be adapted to big-endian), but testing it on
qemu-ppc with the endian check disabled (thus wrong behavior)
suggested that this approach would work well on there too if we could
adapt it. Of course tests under qemu are not worth much; the ARM tests
were on real hardware and I'd like to see real-hardware results for
others archs (mipsel?) too.

This is not a replacement for the ARM asm (which is still better), but
it's a step towards avoiding the need to have written-by-hand assembly
for every single new arch we add as a prerequisite for tolerable
performance.

Rich