64 bit operations

GCC uses the NEON unit for 64 bit operations when tuning for the Cortex-A9. Not all of the basic operations are implemented which mean some routines ping pong values back and forth between the NEON unit and core registers.

At least multiple, negate, {left,right} shift, not, and bitwise not are missing. See:
 http://lists.linaro.org/pipermail/linaro-toolchain/2011-October/001713.html

See:
 https://bazaar.launchpad.net/~linaro-toolchain-wg/linaro-toolchain-benchmarks/private-runs/view/head:/mixed/2011-11-15-a9-4.6.2-vs-2011.11-michaelh-1/NOTES.rst

for a list of benchmarks which regress. Five from EEMBC and DENBench significantly regress. 186.crafty is dominated by 64 bit operations and regresses. 171.swim and 164.gzip also regress.

This may explain why -mtune=cortex-a8 is sometimes faster when running on an A9.