stack32 out_stack
int32 out
stack32 u_stack
int32 u
int32 eax
int32 ebx
int32 esi
int32 edi
int32 ebp
float80 out0
float80 out1
float80 out2
float80 out3
float80 out4
float80 carry0
float80 carry1
float80 carry2
float80 carry3
float80 carry4
float80 u22
float80 u32
float80 u42
stack64 u22_stack
stack64 u32_stack
float80 u0u0
float80 u0u1
float80 u0u2
float80 u0u3
float80 u1u1
float80 u1u2
float80 u1u3
float80 u2u2
float80 u2u3
float80 u3u3

enter gaudry_pm_square
input out_stack
input u_stack

caller eax
caller ebx
caller esi
caller edi
caller ebp

  u = u_stack
  out = out_stack

  out3 = *(float64 *) (u + 32)
  out3 *= out3

u22 = *(float64 *) (u + 16)
u22 += u22

u32 = *(float64 *) (u + 24)
u32 += u32

u42 = *(float64 *) (u + 32)
u42 += u42

  out3 *= *(float64 *) &gaudry_pm_scale

  u1u2 = *(float64 *) (u + 8)
  u1u2 *= u22
  out3 += u1u2

  u0u3 = *(float64 *) (u + 0)
  u0u3 *= u32
  out3 += u0u3

            carry3 = *(float64 *) &gaudry_pm_alpha102
            carry3 += out3

            carry3 -= *(float64 *) &gaudry_pm_alpha102

*(float64 *) &u22_stack = u22
*(float64 *) &u32_stack = u32


    out4 = *(float64 *) (u + 0)
    out4 *= u42
  
    u1u3 = *(float64 *) (u + 8)
    u1u3 *= u32
    out4 += u1u3
  
    u2u2 = *(float64 *) (u + 16)
    u2u2 *= u2u2
    out4 += u2u2


            out4 += carry3
            out3 -= carry3


      out0 = *(float64 *) (u + 8)
      out0 *= u42

      u2u3 = *(float64 *) (u + 16)
      u2u3 *= *(float64 *) &u32_stack
      out0 += u2u3


              carry4 = *(float64 *) &gaudry_pm_alpha127
              carry4 += out4
              carry4 -= *(float64 *) &gaudry_pm_alpha127
            
              out0 += carry4
              out4 -= carry4
            
              out0 *= *(float64 *) &gaudry_pm_scale
            
              u0u0 = *(float64 *) (u + 0)
              u0u0 *= u0u0
              out0 += u0u0



        out1 = *(float64 *) (u + 16)
        out1 *= u42
      
        u3u3 = *(float64 *) (u + 24)
        u3u3 *= u3u3
        out1 += u3u3
      
        out1 *= *(float64 *) &gaudry_pm_scale
      
        u0u1 = *(float64 *) (u + 0)
        u0u1 += u0u1
        u0u1 *= *(float64 *) (u + 8)
        out1 += u0u1

          out2 = *(float64 *) (u + 24)
          out2 *= u42
        
          out2 *= *(float64 *) &gaudry_pm_scale
        
          u0u2 = *(float64 *) (u + 0)
          u0u2 *= *(float64 *) &u22_stack
          out2 += u0u2
        
          u1u1 = *(float64 *) (u + 8)
          u1u1 *= u1u1
          out2 += u1u1

                carry0 = *(float64 *) &gaudry_pm_alpha26
                carry0 += out0

                carry0 -= *(float64 *) &gaudry_pm_alpha26

                out1 += carry0
                out0 -= carry0
                *(float64 *) (out + 0) = out0

                carry1 = *(float64 *) &gaudry_pm_alpha51
                carry1 += out1

                carry1 -= *(float64 *) &gaudry_pm_alpha51

                  out2 += carry1

                  carry2 = *(float64 *) &gaudry_pm_alpha77
                  carry2 += out2

                  carry2 -= *(float64 *) &gaudry_pm_alpha77

                    out3 += carry2

                    carry3 = *(float64 *) &gaudry_pm_alpha102
                    carry3 += out3

                    carry3 -= *(float64 *) &gaudry_pm_alpha102

                      out1 -= carry1
                      *(float64 *) (out + 8) = out1
                      out2 -= carry2
                      *(float64 *) (out + 16) = out2
                      out3 -= carry3
                      *(float64 *) (out + 24) = out3
                      out4 += carry3
                      *(float64 *) (out + 32) = out4



leave