I did some similar experiments myself a long time ago. Something mildly surprising for me at that time was that this got optimized to a single roll instruction:

    uint32_t roll(uint32_t const x) {
        return (x << 15) | (x >> 17);
    }

I haven't managed to make GCC generate a non-constant roll (i.e. (x << c) | (x >> (32-c)) ), probably because it couldn't infer that c is in the [0,31] range.