Notices by Paul Khuong (pkhuong@discuss.systems)

@mcc movzx/sx for widening. regular mov for narrowing, but be mindful of partial dependencies on 8 and 16 bit destinations (16b in particular).

The size suffix is an at&t special and not part of the ISA (which also has movd/movq…)

@cr1901 @mcc 16 bit ops are extra slow nowadays (no one uses them).

So, the MOAR LAYERS folks like log-sum-exp… do we know if one tried to replace log with a bitcast of float to int, and exp with the reverse? In a float world, the cost is int -> float after bitcast, and float -> int before bitcast.

Let's say you know the input is an exact multiple of your integer divisor. Does that let you come up with a smaller div-by-mul constant?

EDIT: I don't mean a reciprocal inverse mod 2^w, but literally a fixed point div by mul (which also avoids the need to ensure the divisor is odd).

https://discuss.systems/@pkhuong/113663255154088174