14 Mar
2024
14 Mar
'24
2:12 p.m.
On Thu Mar 14 10:49:00 2024 +0000, Giovanni Mascellani wrote:
Mmh, given your earlier patch I guess those are meant to be the signed and unsigned version and then the backend is expected to extract the right sign from the operand type. However, it seems that DXIL always emits unsigned integers, doesn't it? So the typing information here is lost anyway. To be honest I don't even understand why here signed vs unsigned is important, given that we're truncating to the lower bits anyway. Do you have an idea of what is meant to happen here?
I have no idea why there are signed and unsigned instructions, since it has no effect on the multiply or the add.