SSE2: pxor vs subps

[dioxin]

not that I really believed in align 16, but many here do

You don't believe in aligning code?
But it makes a big difference if you align it. I get the following results if I sprinkle a few ALIGN 16s in the code:

Code Select Expand

5748 ms for psubb
5729 ms for psubq
5730 ms for xorps
5730 ms for xorpd
5729 ms for pxor
5990 ms for subps
That's much better than the original.

[jj2007]

Putting align 16 in front of the innermost loop has absolutely NO effect on my Celeron M... but it's true that there are more sensitive CPUs around.

[Rockoon]

Branch target aligning is not so effective on Intel (currently) due to the trace cache, but it doesnt usually hurt to align anyways.