On 09/05/2011 05:38 PM, Phil Endecott wrote:
I've just done some quick benchmarks of Beman's proposed byte-swapping code using the following test program:
...
What do people see on other platforms?
I added a test case that uses Beman's integer types that yields much better performance (~2.5X) than the reorder functions in the conversion portion of the library. I.e. I added: #elif USE_BIG return *reinterpret_cast<const ubig32_t*>(&src); #elif USE_LITTLE return *reinterpret_cast<const ulittle32_t*>(&src); #elif USE_NATIVE return src; and on a Intel Core2 with gcc-4.4.3, I get the following results at -O2 (except where noted, -O3 and -O4 produced nearly identical results): USE_BUILTIN 1.42 USE_HTONL 1.39 USE_BEMAN 7.81 USE_BIG 3.09 USE_LITTLE 2.95 USE_NATIVE 0.86 (0.23 for -O3) So it looks like the mechanism used in the "integers" part of the library is superior to the mechanism used in the "conversion" part of the library and maybe the latter should be adapted to use the former. It seems like for that to work the the "endianness" enum would have to be extended to include "nonnative". Then the reorder<T> functions could then be implemented in terms of endian<T, nonnative>.