Rule for mixed-signedness operands to integer div, rem, less-than, greater-than, less-or-equal, greater-or-equal #707

In SPIR-V the rule on integer operations is that the signedness-semantics is determined by the opcode, not the signedness of the type of the operands

This is why, for example, it has both OpSLessThan, OpULessThan (for scalar integer values %a and %b of the same bit width):

• %result = OpSLessThan %bool %a %b
  • Ignore the signedness of operands %a and %b compare them as if they are both signed.
• %result = OpULessThan %bool %a %b
  • Ignore the signedness of operands %a and %b compare them as if they are both unsigned.

Similarly, there are "U" and "S" variants for Div, Rem, > , <=, >=

I think we might have touched on this issue in an earlier meeting but I don't recall a firm resolution or record in the WGSL spec.

I think the favoured option was:

• For these binary operations, it's only valid to have operands with consistent signedness. The programmer has to insert bit-level (reinterpretation) "as" casts in order to achieve signedness-consistency.