On 05/04/2012 03:05 PM, Thomas DuBuisson wrote:
Vincent uses gcc header files to get the AES instructions:
Header files of:
#include
#include And later calls of:
x = _mm_aesenc_si128(m, K1);
But currently you must know you have AESNI and use a flag:
cabal install cryptocipher -faesni
But if you are wrong:
Illegal instruction (core dumped) Of course that's expected as of now, since it's not finished and i had to push a new release (related to some significant performance improvement for RSA/DH/DSA), the code is there as a "technology preview".
But the goal is to turn unconditionally the AESNI "flag" when arch is x86 or x86_64, which in this case the implementation would rely on the runtime cpuid check to use the aesni fastpath or not.
This is a great place to be - now we just take the CPU checking from intel-aes, make a switch between Vincent's C and Gladman (in haskell or out, I doesn't matter to me), graft on Ctr mode as specified then it's all about matching the current 'random' API.
Please don't take the intel-aes test implementation. it's skewed to just support Intel, since it basically testing for the cpu string "GenuineIntel". The only necessary test is the cpuid 1 with ecx having the 25th bit set. It should just work providing cpus other than intel have matching cpuid 1 layout (which as far i'm concerned seems to be the case in most cases) -- Vincent