This will be generated by glsl_to_nir, and it turns out that this is a
more code-efficient path than the floating point math, anyway.

No change on shader-db, but drops an instruction in piglit's
glsl-fs-frontfacing.