On 06/06/2011 04:07 PM, malc wrote: > Depends on how bswap_16 is defined. If it is __builtin_bswap16 > then 4.5.0 and 4.6.0 generate byte reversed loads, and previous > versions lack that builtin, so i don't think this generic code > should go in.
It would continue to be defined as-is, without direct reference to the __builtin_bswap functions. But you're right that the generic code would depend on the bswap optimization pass that recognizes the mask/shift/or pattern and converts it to the builtins internally. What if we kept the ppc ifdefs, but converted the rest to the gcc generic code? r~
