.text
        .global dot_product
dot_product:
        push    %ebp
        mov     %esp, %ebp
        push    %ebx
        mov     8(%ebp), %eax
        mov     12(%ebp), %ebx
        mov     16(%ebp), %edx
        movss   $0, %xmm2
        cmpl    $4, %eax
        jge     more_than_four
last_point:
        testl   %eax, %eax
        je      is_zero
        movss   (%ebx), %xmm0
        movss   (%edx), %xmm1
        mulss   %xmm0, %xmm1
        addss   %xmm1, %xmm2
        addl    $4, %edx
        addl    $4, %ebx
        decl    %eax
        jmp     last_point
more_than_four:
        movupd  (%ebx), %xmm0
        movupd  (%edx), %xmm1
        mulps   %xmm1, %xmm0
        movss   $0, %xmm1
        subl    $16, %esp
        movupd  %xmm0, %esp
        addss   (%esp), %xmm2
        addss   4(%esp), %xmm2
        addss   8(%esp), %xmm2
        addss   12(%esp), %xmm2
        addl    $16, %esp
        addl    $16, %ebx
        addl    $16, %edx
        subl    $4, %eax
        cmpl    $4, %eax
        jl      last_point
        jmp     more_than_four
is_zero:
        movss   %xmm2, %xmm0
        subl    $4, %esp
        movss   %xmm0, (%esp)
        flds    (%esp)
        addl    $4, %esp
        popl    %ebx
        movl    %ebp, %esp
        popl    %ebp
        ret