https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120378
--- Comment #3 from Robin Dapp <rdapp at gcc dot gnu.org> --- vnclipu is basically a scaling (narrowing), rounding shift with subsequent "clip" i.e. saturation. Its input and output is unsigned, though, so for the function above we first need to "clip" the negative values to 0 and then shift twice from unsigned int to unsigned char. So far I'm only aware of the vector insn but I think there are discussions about a scalar one for a future extension. There's also vnclip (signed -> signed). An alternative to vnclipu would be vmax (vmin (...)) but then we'd still need to truncate the result 2x. Truncations are narrowing shifts as well, meaning we'd need 4 instructions instead of 3.
