Why is a float being converted to an int before adding?

I was messing around on http://ift.tt/1GF9raf when I observed something peculiar. Consider the following function:

#include <algorithm>
#include <cstdlib>
#include <functional>

float dot(float src1[], float src2[], int size) {
  float* vecmul = static_cast<float*>(malloc(size * sizeof (float)));
  float dotprod = 0;
  std::transform(src1, src1+size, src2, vecmul, std::multiplies<float>());
  dotprod = std::accumulate(vecmul, vecmul+size, 0);
  free(vecmul);
  return dotprod;
}

With flags -O3 -std=c++11 on x86 gcc 4.9.2 this gets compiled down to:

dot(float*, float*, int):
// load args, do multiplication from std::transform (with mulss)
.L22:
    pxor    %xmm0, %xmm0
    addq    $4, %rcx
    cvtsi2ss    %edx, %xmm0 *
    addss   -4(%rcx), %xmm0 *
    cmpq    %rcx, %rsi      *
    cvttss2si   %xmm0, %edx *
    jne .L22              
    pxor    %xmm0, %xmm0    
    cvtsi2ss    %edx, %xmm0 
.L4:
// pop arguments, free, etc.

I'm curious as to why we have the float-to-int conversion, then an int addition, and then a conversion back (asterisked lines). Why would this be faster than a direct fadd?