Am I stupid, or what? PSRLW weirdness/misunderstanding

[SansGrip]

I have a sum in mm5 and a counter in mm2. The counter can be either 0, 1 or 2. I want to divide the sum by the counter. Here's the code I have right now:

psrlw mm2, 1 // counter /= 2
psrlw mm5, mm2 // mm5 = average

In theory this should work because:

Two values in sum? Shift right once (divide by two).
One value in sum? Shift right zero times.
Zero in sum? Shift right zero times.

But I get some weird output, and when I run the filter in the debugger I see that following the second line, mm5 is all zeros.

I assume I'm misunderstanding how PSRLW works, but I've not seen this behaviour mentioned in any of the references I've checked.

My processor is an Athlon 2100+ XP.

Any ideas would be greatly appreciated.

[SansGrip]

I think I figured it out. It seems to use the entire source register as the shift count, instead of the individual words. This is extraordinarily non-useful. Perhaps there's another instruction that works the way I need...