[rwmj]

One reason is that you just want to call a single instruction, and the overhead of making a function call to another file would be too much. Picking an example from the Linux kernel pretty much at random:

    #define mb()    asm volatile("mfence":::"memory")
    #define rmb()   asm volatile("lfence":::"memory")
    #define wmb()   asm volatile("sfence" ::: "memory")

Those are memory barriers, so they essentially must be inline (they'd be too slow and possibly even change their meaning if they were located in a separate source file and you had to call them).