A note on mask registers (AVX-512)

Yeah, I would think most mask registers values just "pop into existence" as already nonzero based on a full write to the register, e.g., from a comparison, a load from memory, move from GP reg, etc.

I can't see all that many scenarios where you want a zeroed register to begin with. Many of the cases that happens in the GP world (e.g., zeroing out some accumulator or pointer or whatever) just don't come up in mask registers.

I am not sure about the cost to implement it. I don't think it's all that complicated to describe how it would work, but it does occur in the renamer, which is a critical path on modern chips, so probably the less complication there, the better. We already know that not all zeroing idioms are recognized (like sub same, same) on some chips, so that is some indication that it's not super cheap (or else you'd think they'd have just implemented all the idioms the first time around).

Is zeroing a mask register a common operation? I'd suspect that setting all mask bits would be much more common.

I mean, if you are going to zero the register, the real optimization would be to completely skip emitting any instructions for the operation you are going to perform, since it probably isn't going to do anything if the mask is cleared.

Great article as usual! Nice analysis. btw, just wondering, were you a processor architect in your previous roles?