func f(a []int) int {
	s := 0
	for _, x := range a {
		s += x
	}
	return s
}

Generates for the inner loop

		MOVQ	(AX)(DX*8), SI
		INCQ	DX
		ADDQ	SI, BX
		CMPQ	CX, DX
		JGT	16

There's no reason we couldn't do ADDQ (AX)(DX*8), BX instead of the MOVQ/ADDQ combo.

There's actually code that does this combining optimization in the compiler, but it doesn't trigger. The difficulty is that the combined ADDQ instruction will clobber its input, so we don't combine when the non-load argument to the ADDQ is used somewhere else (see rewrite.go:canMergeLoadClobber). In this case the other use of BX is at the return statement. But that use shouldn't stop the optimization from happening, as it is in a separate branch. At the actual point where we're doing the optimization, the old value of BX is dead.