GCC can't optimize out the mask (which is required to avoid UB if n may be equal or greater than 64):

    #include <stdint.h>

    uint64_t rol(uint64_t n, uint64_t val)
    {
        n &= 63;
        return (val << n) | (val >> 64-n);
    }

Compiles to

    rol(unsigned long, unsigned long):
      mov rcx, rdi
      mov rax, rsi
      and ecx, 63
      rol rax, cl
      ret

Clang and ICC do get it right.