Intel Corp (INTC)

[mmoy]

> There may be other improvements in this area, but the big
> one I'm aware of is that Nehalem is once again able to do
> store-forwards to unaligned loads across 1,2,3,4,5,6,7,8,
> and 12 byte boundaries (ie forward data to unaligned loads
> without waiting for that store to complete its write to the
> cache).

That's pretty interesting given these results:

0 1: memcpy = 812 ms memcpy1 = 125 ms 15.39 compare = 0
0 2: memcpy = 672 ms memcpy1 = 109 ms 16.22 compare = 0
0 3: memcpy = 672 ms memcpy1 = 110 ms 16.37 compare = 0
0 4: memcpy = 671 ms memcpy1 = 94 ms 14.01 compare = 0
0 5: memcpy = 656 ms memcpy1 = 110 ms 16.77 compare = 0
0 6: memcpy = 672 ms memcpy1 = 109 ms 16.22 compare = 0
0 7: memcpy = 656 ms memcpy1 = 110 ms 16.77 compare = 0
0 8: memcpy = 343 ms memcpy1 = 94 ms 27.41 compare = 0
0 9: memcpy = 375 ms memcpy1 = 94 ms 25.07 compare = 0
0 10: memcpy = 375 ms memcpy1 = 109 ms 29.07 compare = 0
0 11: memcpy = 375 ms memcpy1 = 94 ms 25.07 compare = 0
0 12: memcpy = 234 ms memcpy1 = 94 ms 40.17 compare = 0
0 13: memcpy = 344 ms memcpy1 = 94 ms 27.33 compare = 0
0 14: memcpy = 343 ms memcpy1 = 110 ms 32.07 compare = 0
0 15: memcpy = 344 ms memcpy1 = 93 ms 27.03 compare = 0