| | 1 | The first routine I worked on was memcpy(). |
| | 2 | |
| | 3 | Comparison of stock memcpy() of a single page on FreeBSD vs Linux on a Westmere (values are TSC deltas): |
| | 4 | |
| | 5 | {{{ |
| | 6 | x fbsd/westmere/builtin |
| | 7 | + linux/builtin |
| | 8 | N Min Max Median Avg Stddev |
| | 9 | x 1000 336 18444 340 361.628 573.11483 |
| | 10 | + 1000 276 9996 280 288.924 307.34136 |
| | 11 | Difference at 95.0% confidence |
| | 12 | -72.704 +/- 40.3074 |
| | 13 | -20.1046% +/- 11.1461% |
| | 14 | (Student's t, pooled s = 459.847) |
| | 15 | }}} |
| | 16 | |
| | 17 | ||= Idea =||= Westmere =||= Sandy Bridge =||= Ivy Bridge =|| |
| | 18 | || Replace `dec` with `sub` || none || none || none || |
| | 19 | || Use movsd instead of movsq || slightly slower || slightly slower || 6% faster || |
| | 20 | || Simple `movdqa` loop || 138% slower || 58% slower || 46% slower || |
| | 21 | || `movdqa` 32 at a time (old) || 27% slower || 14% faster || 17% faster || |
| | 22 | || `movdqa` 32 at a time (new) || 27% slower || 15% faster || 18% faster || |
| | 23 | || `movdqa` 64 at a time || 224% slower || 131% slower || 116% slower || |