hoisting loop invariant loads?

[Jeff Epler]

I tested with gcc version 4.9.2 (Debian 4.9.2-10) targeting x86_64.

The snippet you posted compiles into a single 'ret' instruction under
most optimization levels since sum has no visible effect.

When I change the return type of sum to int and return s, most
optimization levels (even just -O) seem to do the hoisting you desire.
The loop became
    
.L3:
        addl    (%rdx), %eax
        addq    $4, %rdx
        cmpq    %rcx, %rdx
        jne     .L3
        rep ret

which I read roughly as
        do { eax += *rdx; rdx += 4; } while(rcx != rdx)

I do not have an aarch64 compiler handy, but I do have an arm compiler,
so I checked it.  Again under -Os,
.L3:
        ldr     r2, [r3], #4
        add     r0, r0, r2
        cmp     r3, r1
        bne     .L3
        bx      lr

which I read the same as the x86_64 code above.

Jeff