none/tests/amd64/fma.c

   1 #include <stdio.h>
   2 #include <string.h>
   3
   4 #define N 64
   5 struct float_test {
   6    float x[N], y[N], z[N], expected[N], res[N];
   7 } ft __attribute__((aligned (32)));
   8
   9 struct double_test {
  10    double x[N], y[N], z[N], expected[N], res[N];
  11 } dt __attribute__((aligned (32)));
  12
  13 float plus_zero, plus_infty, minus_infty, nan_value;
  14
  15 static int testf( float x, float y )
  16 {
  17    unsigned int a, b;
  18    memcpy( &a, &x, sizeof (a) );
  19    memcpy( &b, &y, sizeof (b) );
  20    if ((a & 0x7fc00000U) == 0x7fc00000U)
  21       return (b & 0x7fc00000U) != 0x7fc00000U;
  22    return memcmp( &a, &b, sizeof (a) ) != 0;
  23 }
  24
  25 static int test_fmaf( void )
  26 {
  27    int res = 0, i, j;
  28    float w;
  29    for (i = 0; i < N; i++) {
  30       int thisres = 0;
  31       __asm __volatile__ ("vfmadd132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  32       thisres |= testf( w, ft.expected[i] );
  33       __asm __volatile__ ("vfmadd132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "m" (ft.y[i]), "x" (ft.z[i]));
  34       thisres |= testf( w, ft.expected[i] );
  35       __asm __volatile__ ("vfmadd213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  36       thisres |= testf( w, ft.expected[i] );
  37       __asm __volatile__ ("vfmadd213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "m" (ft.z[i]));
  38       thisres |= testf( w, ft.expected[i] );
  39       __asm __volatile__ ("vfmadd231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "x" (ft.y[i]), "0" (ft.z[i]));
  40       thisres |= testf( w, ft.expected[i] );
  41       __asm __volatile__ ("vfmadd231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "m" (ft.y[i]), "0" (ft.z[i]));
  42       thisres |= testf( w, ft.expected[i] );
  43       if (thisres)
  44          printf( "Failure 1 %d %a %a\n", i, w, ft.expected[i] );
  45       res |= thisres;
  46       thisres = 0;
  47       __asm __volatile__ ("vfnmsub132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  48       thisres |= testf( -w, ft.expected[i] );
  49       __asm __volatile__ ("vfnmsub132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "m" (ft.y[i]), "x" (ft.z[i]));
  50       thisres |= testf( -w, ft.expected[i] );
  51       __asm __volatile__ ("vfnmsub213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  52       thisres |= testf( -w, ft.expected[i] );
  53       __asm __volatile__ ("vfnmsub213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "m" (ft.z[i]));
  54       thisres |= testf( -w, ft.expected[i] );
  55       __asm __volatile__ ("vfnmsub231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "x" (ft.y[i]), "0" (ft.z[i]));
  56       thisres |= testf( -w, ft.expected[i] );
  57       __asm __volatile__ ("vfnmsub231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "m" (ft.y[i]), "0" (ft.z[i]));
  58       thisres |= testf( -w, ft.expected[i] );
  59       if (thisres)
  60          printf( "Failure 2 %d %a %a\n", i, w, ft.expected[i] );
  61       res |= thisres;
  62    }
  63    for (i = 0; i < N; i++)
  64       ft.z[i] = -ft.z[i];
  65    for (i = 0; i < N; i++) {
  66       int thisres = 0;
  67       __asm __volatile__ ("vfmsub132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  68       thisres |= testf( w, ft.expected[i] );
  69       __asm __volatile__ ("vfmsub132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "m" (ft.y[i]), "x" (ft.z[i]));
  70       thisres |= testf( w, ft.expected[i] );
  71       __asm __volatile__ ("vfmsub213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  72       thisres |= testf( w, ft.expected[i] );
  73       __asm __volatile__ ("vfmsub213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "m" (ft.z[i]));
  74       thisres |= testf( w, ft.expected[i] );
  75       __asm __volatile__ ("vfmsub231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "x" (ft.y[i]), "0" (ft.z[i]));
  76       thisres |= testf( w, ft.expected[i] );
  77       __asm __volatile__ ("vfmsub231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "m" (ft.y[i]), "0" (ft.z[i]));
  78       thisres |= testf( w, ft.expected[i] );
  79       if (thisres)
  80          printf( "Failure 3 %d %a %a\n", i, w, ft.expected[i] );
  81       res |= thisres;
  82       thisres = 0;
  83       __asm __volatile__ ("vfnmadd132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  84       thisres |= testf( -w, ft.expected[i] );
  85       __asm __volatile__ ("vfnmadd132ss %2, %3, %0" : "=x" (w) : "0" (ft.x[i]), "m" (ft.y[i]), "x" (ft.z[i]));
  86       thisres |= testf( -w, ft.expected[i] );
  87       __asm __volatile__ ("vfnmadd213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "x" (ft.z[i]));
  88       thisres |= testf( -w, ft.expected[i] );
  89       __asm __volatile__ ("vfnmadd213ss %3, %2, %0" : "=x" (w) : "0" (ft.x[i]), "x" (ft.y[i]), "m" (ft.z[i]));
  90       thisres |= testf( -w, ft.expected[i] );
  91       __asm __volatile__ ("vfnmadd231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "x" (ft.y[i]), "0" (ft.z[i]));
  92       thisres |= testf( -w, ft.expected[i] );
  93       __asm __volatile__ ("vfnmadd231ss %2, %1, %0" : "=x" (w) : "x" (ft.x[i]), "m" (ft.y[i]), "0" (ft.z[i]));
  94       thisres |= testf( -w, ft.expected[i] );
  95       if (thisres)
  96          printf( "Failure 4 %d %a %a\n", i, w, ft.expected[i] );
  97       res |= thisres;
  98    }
  99    for (i = 0; i < N; i++)
 100       ft.z[i] = -ft.z[i];
 101    for (i = 0; i < N; i += 4) {
 102       int thisres = 0;
 103       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
 104                           "vfmadd132ps %%xmm7, %%xmm8, %%xmm9;"
 105                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 106                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 107       for (j = 0; j < 4; j++)
 108          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 109       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
 110                           "vfmadd132ps (%2), %%xmm8, %%xmm9;"
 111                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 112                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 113       for (j = 0; j < 4; j++)
 114          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 115       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
 116                           "vfmadd213ps %%xmm7, %%xmm8, %%xmm9;"
 117                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 118                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 119       for (j = 0; j < 4; j++)
 120          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 121       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
 122                           "vfmadd213ps (%3), %%xmm8, %%xmm9;"
 123                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 124                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 125       for (j = 0; j < 4; j++)
 126          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 127       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
 128                           "vfmadd231ps %%xmm7, %%xmm8, %%xmm9;"
 129                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 130                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 131       for (j = 0; j < 4; j++)
 132          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 133       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
 134                           "vfmadd231ps (%2), %%xmm8, %%xmm9;"
 135                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 136                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 137       for (j = 0; j < 4; j++)
 138          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 139       if (thisres) {
 140          printf( "Failure 5 %d", i );
 141          for (j = 0; j < 4; j++)
 142             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 143          printf( "\n" );
 144       }
 145       res |= thisres;
 146       thisres = 0;
 147       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
 148                           "vfnmsub132ps %%xmm7, %%xmm8, %%xmm9;"
 149                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 150                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 151       for (j = 0; j < 4; j++)
 152          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 153       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
 154                           "vfnmsub132ps (%2), %%xmm8, %%xmm9;"
 155                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 156                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 157       for (j = 0; j < 4; j++)
 158          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 159       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
 160                           "vfnmsub213ps %%xmm7, %%xmm8, %%xmm9;"
 161                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 162                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 163       for (j = 0; j < 4; j++)
 164          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 165       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
 166                           "vfnmsub213ps (%3), %%xmm8, %%xmm9;"
 167                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 168                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 169       for (j = 0; j < 4; j++)
 170          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 171       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
 172                           "vfnmsub231ps %%xmm7, %%xmm8, %%xmm9;"
 173                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 174                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 175       for (j = 0; j < 4; j++)
 176          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 177       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
 178                           "vfnmsub231ps (%2), %%xmm8, %%xmm9;"
 179                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 180                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 181       for (j = 0; j < 4; j++)
 182          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 183       if (thisres) {
 184          printf( "Failure 6 %d", i );
 185          for (j = 0; j < 4; j++)
 186             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 187          printf( "\n" );
 188       }
 189       res |= thisres;
 190    }
 191    for (i = 0; i < N; i++)
 192       ft.z[i] = -ft.z[i];
 193    for (i = 0; i < N; i += 4) {
 194       int thisres = 0;
 195       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
 196                           "vfmsub132ps %%xmm7, %%xmm8, %%xmm9;"
 197                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 198                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 199       for (j = 0; j < 4; j++)
 200          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 201       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
 202                           "vfmsub132ps (%2), %%xmm8, %%xmm9;"
 203                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 204                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 205       for (j = 0; j < 4; j++)
 206          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 207       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
 208                           "vfmsub213ps %%xmm7, %%xmm8, %%xmm9;"
 209                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 210                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 211       for (j = 0; j < 4; j++)
 212          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 213       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
 214                           "vfmsub213ps (%3), %%xmm8, %%xmm9;"
 215                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 216                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 217       for (j = 0; j < 4; j++)
 218          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 219       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
 220                           "vfmsub231ps %%xmm7, %%xmm8, %%xmm9;"
 221                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 222                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 223       for (j = 0; j < 4; j++)
 224          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 225       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
 226                           "vfmsub231ps (%2), %%xmm8, %%xmm9;"
 227                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 228                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 229       for (j = 0; j < 4; j++)
 230          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 231       if (thisres) {
 232          printf( "Failure 7 %d", i );
 233          for (j = 0; j < 4; j++)
 234             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 235          printf( "\n" );
 236       }
 237       res |= thisres;
 238       thisres = 0;
 239       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
 240                           "vfnmadd132ps %%xmm7, %%xmm8, %%xmm9;"
 241                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 242                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 243       for (j = 0; j < 4; j++)
 244          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 245       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
 246                           "vfnmadd132ps (%2), %%xmm8, %%xmm9;"
 247                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 248                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 249       for (j = 0; j < 4; j++)
 250          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 251       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
 252                           "vfnmadd213ps %%xmm7, %%xmm8, %%xmm9;"
 253                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 254                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 255       for (j = 0; j < 4; j++)
 256          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 257       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
 258                           "vfnmadd213ps (%3), %%xmm8, %%xmm9;"
 259                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 260                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 261       for (j = 0; j < 4; j++)
 262          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 263       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
 264                           "vfnmadd231ps %%xmm7, %%xmm8, %%xmm9;"
 265                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 266                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 267       for (j = 0; j < 4; j++)
 268          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 269       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
 270                           "vfnmadd231ps (%2), %%xmm8, %%xmm9;"
 271                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 272                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 273       for (j = 0; j < 4; j++)
 274          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 275       if (thisres) {
 276          printf( "Failure 8 %d", i );
 277          for (j = 0; j < 4; j++)
 278             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 279          printf( "\n" );
 280       }
 281       res |= thisres;
 282    }
 283    for (i = 1; i < N; i += 2)
 284       ft.z[i] = -ft.z[i];
 285    for (i = 0; i < N; i += 4) {
 286       int thisres = 0;
 287       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
 288                           "vfmaddsub132ps %%xmm7, %%xmm8, %%xmm9;"
 289                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 290                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 291       for (j = 0; j < 4; j++)
 292          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 293       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
 294                           "vfmaddsub132ps (%2), %%xmm8, %%xmm9;"
 295                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 296                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 297       for (j = 0; j < 4; j++)
 298          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 299       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
 300                           "vfmaddsub213ps %%xmm7, %%xmm8, %%xmm9;"
 301                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 302                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 303       for (j = 0; j < 4; j++)
 304          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 305       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
 306                           "vfmaddsub213ps (%3), %%xmm8, %%xmm9;"
 307                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 308                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 309       for (j = 0; j < 4; j++)
 310          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 311       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
 312                           "vfmaddsub231ps %%xmm7, %%xmm8, %%xmm9;"
 313                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 314                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 315       for (j = 0; j < 4; j++)
 316          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 317       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
 318                           "vfmaddsub231ps (%2), %%xmm8, %%xmm9;"
 319                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 320                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 321       for (j = 0; j < 4; j++)
 322          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 323       if (thisres) {
 324          printf( "Failure 9 %d", i );
 325          for (j = 0; j < 4; j++)
 326             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 327          printf( "\n" );
 328       }
 329       res |= thisres;
 330    }
 331    for (i = 0; i < N; i++)
 332       ft.z[i] = -ft.z[i];
 333    for (i = 0; i < N; i += 4) {
 334       int thisres = 0;
 335       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%3), %%xmm8;"
 336                           "vfmsubadd132ps %%xmm7, %%xmm8, %%xmm9;"
 337                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 338                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 339       for (j = 0; j < 4; j++)
 340          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 341       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm8;"
 342                           "vfmsubadd132ps (%2), %%xmm8, %%xmm9;"
 343                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 344                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 345       for (j = 0; j < 4; j++)
 346          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 347       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%3), %%xmm7; vmovaps (%2), %%xmm8;"
 348                           "vfmsubadd213ps %%xmm7, %%xmm8, %%xmm9;"
 349                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 350                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 351       for (j = 0; j < 4; j++)
 352          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 353       __asm __volatile__ ("vmovaps (%1), %%xmm9; vmovaps (%2), %%xmm8;"
 354                           "vfmsubadd213ps (%3), %%xmm8, %%xmm9;"
 355                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 356                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 357       for (j = 0; j < 4; j++)
 358          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 359       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%2), %%xmm7; vmovaps (%1), %%xmm8;"
 360                           "vfmsubadd231ps %%xmm7, %%xmm8, %%xmm9;"
 361                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 362                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 363       for (j = 0; j < 4; j++)
 364          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 365       __asm __volatile__ ("vmovaps (%3), %%xmm9; vmovaps (%1), %%xmm8;"
 366                           "vfmsubadd231ps (%2), %%xmm8, %%xmm9;"
 367                           "vmovaps %%xmm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 368                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 369       for (j = 0; j < 4; j++)
 370          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 371       if (thisres) {
 372          printf( "Failure 10 %d", i );
 373          for (j = 0; j < 4; j++)
 374             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 375          printf( "\n" );
 376       }
 377       res |= thisres;
 378    }
 379    for (i = 1; i < N; i += 2)
 380       ft.z[i] = -ft.z[i];
 381    for (i = 0; i < N; i += 8) {
 382       int thisres = 0;
 383       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%3), %%ymm8;"
 384                           "vfmadd132ps %%ymm7, %%ymm8, %%ymm9;"
 385                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 386                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 387       for (j = 0; j < 8; j++)
 388          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 389       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm8;"
 390                           "vfmadd132ps (%2), %%ymm8, %%ymm9;"
 391                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 392                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 393       for (j = 0; j < 8; j++)
 394          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 395       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm7; vmovaps (%2), %%ymm8;"
 396                           "vfmadd213ps %%ymm7, %%ymm8, %%ymm9;"
 397                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 398                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 399       for (j = 0; j < 8; j++)
 400          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 401       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm8;"
 402                           "vfmadd213ps (%3), %%ymm8, %%ymm9;"
 403                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 404                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 405       for (j = 0; j < 8; j++)
 406          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 407       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%1), %%ymm8;"
 408                           "vfmadd231ps %%ymm7, %%ymm8, %%ymm9;"
 409                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 410                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 411       for (j = 0; j < 8; j++)
 412          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 413       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%1), %%ymm8;"
 414                           "vfmadd231ps (%2), %%ymm8, %%ymm9;"
 415                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 416                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 417       for (j = 0; j < 8; j++)
 418          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 419       if (thisres) {
 420          printf( "Failure 11 %d", i );
 421          for (j = 0; j < 8; j++)
 422             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 423          printf( "\n" );
 424       }
 425       res |= thisres;
 426       thisres = 0;
 427       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%3), %%ymm8;"
 428                           "vfnmsub132ps %%ymm7, %%ymm8, %%ymm9;"
 429                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 430                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 431       for (j = 0; j < 8; j++)
 432          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 433       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm8;"
 434                           "vfnmsub132ps (%2), %%ymm8, %%ymm9;"
 435                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 436                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 437       for (j = 0; j < 8; j++)
 438          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 439       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm7; vmovaps (%2), %%ymm8;"
 440                           "vfnmsub213ps %%ymm7, %%ymm8, %%ymm9;"
 441                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 442                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 443       for (j = 0; j < 8; j++)
 444          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 445       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm8;"
 446                           "vfnmsub213ps (%3), %%ymm8, %%ymm9;"
 447                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 448                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 449       for (j = 0; j < 8; j++)
 450          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 451       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%1), %%ymm8;"
 452                           "vfnmsub231ps %%ymm7, %%ymm8, %%ymm9;"
 453                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 454                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 455       for (j = 0; j < 8; j++)
 456          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 457       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%1), %%ymm8;"
 458                           "vfnmsub231ps (%2), %%ymm8, %%ymm9;"
 459                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 460                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 461       for (j = 0; j < 8; j++)
 462          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 463       if (thisres) {
 464          printf( "Failure 12 %d", i );
 465          for (j = 0; j < 8; j++)
 466             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 467          printf( "\n" );
 468       }
 469       res |= thisres;
 470    }
 471    for (i = 0; i < N; i++)
 472       ft.z[i] = -ft.z[i];
 473    for (i = 0; i < N; i += 8) {
 474       int thisres = 0;
 475       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%3), %%ymm8;"
 476                           "vfmsub132ps %%ymm7, %%ymm8, %%ymm9;"
 477                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 478                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 479       for (j = 0; j < 8; j++)
 480          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 481       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm8;"
 482                           "vfmsub132ps (%2), %%ymm8, %%ymm9;"
 483                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 484                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 485       for (j = 0; j < 8; j++)
 486          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 487       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm7; vmovaps (%2), %%ymm8;"
 488                           "vfmsub213ps %%ymm7, %%ymm8, %%ymm9;"
 489                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 490                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 491       for (j = 0; j < 8; j++)
 492          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 493       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm8;"
 494                           "vfmsub213ps (%3), %%ymm8, %%ymm9;"
 495                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 496                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 497       for (j = 0; j < 8; j++)
 498          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 499       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%1), %%ymm8;"
 500                           "vfmsub231ps %%ymm7, %%ymm8, %%ymm9;"
 501                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 502                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 503       for (j = 0; j < 8; j++)
 504          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 505       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%1), %%ymm8;"
 506                           "vfmsub231ps (%2), %%ymm8, %%ymm9;"
 507                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 508                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 509       for (j = 0; j < 8; j++)
 510          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 511       if (thisres) {
 512          printf( "Failure 13 %d", i );
 513          for (j = 0; j < 8; j++)
 514             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 515          printf( "\n" );
 516       }
 517       res |= thisres;
 518       thisres = 0;
 519       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%3), %%ymm8;"
 520                           "vfnmadd132ps %%ymm7, %%ymm8, %%ymm9;"
 521                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 522                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 523       for (j = 0; j < 8; j++)
 524          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 525       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm8;"
 526                           "vfnmadd132ps (%2), %%ymm8, %%ymm9;"
 527                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 528                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 529       for (j = 0; j < 8; j++)
 530          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 531       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm7; vmovaps (%2), %%ymm8;"
 532                           "vfnmadd213ps %%ymm7, %%ymm8, %%ymm9;"
 533                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 534                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 535       for (j = 0; j < 8; j++)
 536          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 537       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm8;"
 538                           "vfnmadd213ps (%3), %%ymm8, %%ymm9;"
 539                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 540                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 541       for (j = 0; j < 8; j++)
 542          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 543       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%1), %%ymm8;"
 544                           "vfnmadd231ps %%ymm7, %%ymm8, %%ymm9;"
 545                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 546                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 547       for (j = 0; j < 8; j++)
 548          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 549       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%1), %%ymm8;"
 550                           "vfnmadd231ps (%2), %%ymm8, %%ymm9;"
 551                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 552                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 553       for (j = 0; j < 8; j++)
 554          thisres |= testf( -ft.res[i+j], ft.expected[i+j] );
 555       if (thisres) {
 556          printf( "Failure 14 %d", i );
 557          for (j = 0; j < 8; j++)
 558             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 559          printf( "\n" );
 560       }
 561       res |= thisres;
 562    }
 563    for (i = 1; i < N; i += 2)
 564       ft.z[i] = -ft.z[i];
 565    for (i = 0; i < N; i += 8) {
 566       int thisres = 0;
 567       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%3), %%ymm8;"
 568                           "vfmaddsub132ps %%ymm7, %%ymm8, %%ymm9;"
 569                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 570                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 571       for (j = 0; j < 8; j++)
 572          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 573       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm8;"
 574                           "vfmaddsub132ps (%2), %%ymm8, %%ymm9;"
 575                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 576                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 577       for (j = 0; j < 8; j++)
 578          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 579       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm7; vmovaps (%2), %%ymm8;"
 580                           "vfmaddsub213ps %%ymm7, %%ymm8, %%ymm9;"
 581                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 582                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 583       for (j = 0; j < 8; j++)
 584          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 585       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm8;"
 586                           "vfmaddsub213ps (%3), %%ymm8, %%ymm9;"
 587                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 588                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 589       for (j = 0; j < 8; j++)
 590          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 591       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%1), %%ymm8;"
 592                           "vfmaddsub231ps %%ymm7, %%ymm8, %%ymm9;"
 593                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 594                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 595       for (j = 0; j < 8; j++)
 596          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 597       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%1), %%ymm8;"
 598                           "vfmaddsub231ps (%2), %%ymm8, %%ymm9;"
 599                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 600                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 601       for (j = 0; j < 8; j++)
 602          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 603       if (thisres) {
 604          printf( "Failure 15 %d", i );
 605          for (j = 0; j < 8; j++)
 606             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 607          printf( "\n" );
 608       }
 609       res |= thisres;
 610    }
 611    for (i = 0; i < N; i++)
 612       ft.z[i] = -ft.z[i];
 613    for (i = 0; i < N; i += 8) {
 614       int thisres = 0;
 615       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%3), %%ymm8;"
 616                           "vfmsubadd132ps %%ymm7, %%ymm8, %%ymm9;"
 617                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 618                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 619       for (j = 0; j < 8; j++)
 620          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 621       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm8;"
 622                           "vfmsubadd132ps (%2), %%ymm8, %%ymm9;"
 623                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 624                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 625       for (j = 0; j < 8; j++)
 626          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 627       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%3), %%ymm7; vmovaps (%2), %%ymm8;"
 628                           "vfmsubadd213ps %%ymm7, %%ymm8, %%ymm9;"
 629                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 630                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 631       for (j = 0; j < 8; j++)
 632          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 633       __asm __volatile__ ("vmovaps (%1), %%ymm9; vmovaps (%2), %%ymm8;"
 634                           "vfmsubadd213ps (%3), %%ymm8, %%ymm9;"
 635                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 636                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 637       for (j = 0; j < 8; j++)
 638          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 639       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%2), %%ymm7; vmovaps (%1), %%ymm8;"
 640                           "vfmsubadd231ps %%ymm7, %%ymm8, %%ymm9;"
 641                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 642                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 643       for (j = 0; j < 8; j++)
 644          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 645       __asm __volatile__ ("vmovaps (%3), %%ymm9; vmovaps (%1), %%ymm8;"
 646                           "vfmsubadd231ps (%2), %%ymm8, %%ymm9;"
 647                           "vmovaps %%ymm9, (%0)" : : "r" (&ft.res[i]), "r" (&ft.x[i]),
 648                                                      "r" (&ft.y[i]), "r" (&ft.z[i]) : "xmm7", "xmm8", "xmm9");
 649       for (j = 0; j < 8; j++)
 650          thisres |= testf( ft.res[i+j], ft.expected[i+j] );
 651       if (thisres) {
 652          printf( "Failure 16 %d", i );
 653          for (j = 0; j < 8; j++)
 654             printf( " %a %a", ft.res[i+j], ft.expected[i+j] );
 655          printf( "\n" );
 656       }
 657       res |= thisres;
 658    }
 659    for (i = 1; i < N; i += 2)
 660       ft.z[i] = -ft.z[i];
 661    return res;
 662 }
 663
 664 static int test( double x, double y )
 665 {
 666    unsigned long long a, b;
 667    memcpy( &a, &x, sizeof (a) );
 668    memcpy( &b, &y, sizeof (b) );
 669    if ((a & 0x7ff8000000000000ULL) == 0x7ff8000000000000ULL)
 670       return (b & 0x7ff8000000000000ULL) != 0x7ff8000000000000ULL;
 671    return memcmp( &a, &b, sizeof (a) ) != 0;
 672 }
 673
 674 static int test_fma( void )
 675 {
 676    int res = 0, i, j;
 677    double w;
 678    for (i = 0; i < N; i++) {
 679       int thisres = 0;
 680       __asm __volatile__ ("vfmadd132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 681       thisres |= test( w, dt.expected[i] );
 682       __asm __volatile__ ("vfmadd132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "m" (dt.y[i]), "x" (dt.z[i]));
 683       thisres |= test( w, dt.expected[i] );
 684       __asm __volatile__ ("vfmadd213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 685       thisres |= test( w, dt.expected[i] );
 686       __asm __volatile__ ("vfmadd213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "m" (dt.z[i]));
 687       thisres |= test( w, dt.expected[i] );
 688       __asm __volatile__ ("vfmadd231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "x" (dt.y[i]), "0" (dt.z[i]));
 689       thisres |= test( w, dt.expected[i] );
 690       __asm __volatile__ ("vfmadd231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "m" (dt.y[i]), "0" (dt.z[i]));
 691       thisres |= test( w, dt.expected[i] );
 692       if (thisres)
 693          printf( "Failure 1 %d %a %a\n", i, w, dt.expected[i] );
 694       res |= thisres;
 695       thisres = 0;
 696       __asm __volatile__ ("vfnmsub132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 697       thisres |= test( -w, dt.expected[i] );
 698       __asm __volatile__ ("vfnmsub132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "m" (dt.y[i]), "x" (dt.z[i]));
 699       thisres |= test( -w, dt.expected[i] );
 700       __asm __volatile__ ("vfnmsub213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 701       thisres |= test( -w, dt.expected[i] );
 702       __asm __volatile__ ("vfnmsub213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "m" (dt.z[i]));
 703       thisres |= test( -w, dt.expected[i] );
 704       __asm __volatile__ ("vfnmsub231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "x" (dt.y[i]), "0" (dt.z[i]));
 705       thisres |= test( -w, dt.expected[i] );
 706       __asm __volatile__ ("vfnmsub231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "m" (dt.y[i]), "0" (dt.z[i]));
 707       thisres |= test( -w, dt.expected[i] );
 708       if (thisres)
 709          printf( "Failure 2 %d %a %a\n", i, w, dt.expected[i] );
 710       res |= thisres;
 711    }
 712    for (i = 0; i < N; i++)
 713       dt.z[i] = -dt.z[i];
 714    for (i = 0; i < N; i++) {
 715       int thisres = 0;
 716       __asm __volatile__ ("vfmsub132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 717       thisres |= test( w, dt.expected[i] );
 718       __asm __volatile__ ("vfmsub132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "m" (dt.y[i]), "x" (dt.z[i]));
 719       thisres |= test( w, dt.expected[i] );
 720       __asm __volatile__ ("vfmsub213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 721       thisres |= test( w, dt.expected[i] );
 722       __asm __volatile__ ("vfmsub213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "m" (dt.z[i]));
 723       thisres |= test( w, dt.expected[i] );
 724       __asm __volatile__ ("vfmsub231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "x" (dt.y[i]), "0" (dt.z[i]));
 725       thisres |= test( w, dt.expected[i] );
 726       __asm __volatile__ ("vfmsub231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "m" (dt.y[i]), "0" (dt.z[i]));
 727       thisres |= test( w, dt.expected[i] );
 728       if (thisres)
 729          printf( "Failure 3 %d %a %a\n", i, w, dt.expected[i] );
 730       res |= thisres;
 731       thisres = 0;
 732       __asm __volatile__ ("vfnmadd132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 733       thisres |= test( -w, dt.expected[i] );
 734       __asm __volatile__ ("vfnmadd132sd %2, %3, %0" : "=x" (w) : "0" (dt.x[i]), "m" (dt.y[i]), "x" (dt.z[i]));
 735       thisres |= test( -w, dt.expected[i] );
 736       __asm __volatile__ ("vfnmadd213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "x" (dt.z[i]));
 737       thisres |= test( -w, dt.expected[i] );
 738       __asm __volatile__ ("vfnmadd213sd %3, %2, %0" : "=x" (w) : "0" (dt.x[i]), "x" (dt.y[i]), "m" (dt.z[i]));
 739       thisres |= test( -w, dt.expected[i] );
 740       __asm __volatile__ ("vfnmadd231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "x" (dt.y[i]), "0" (dt.z[i]));
 741       thisres |= test( -w, dt.expected[i] );
 742       __asm __volatile__ ("vfnmadd231sd %2, %1, %0" : "=x" (w) : "x" (dt.x[i]), "m" (dt.y[i]), "0" (dt.z[i]));
 743       thisres |= test( -w, dt.expected[i] );
 744       if (thisres)
 745          printf( "Failure 4 %d %a %a\n", i, w, dt.expected[i] );
 746       res |= thisres;
 747    }
 748    for (i = 0; i < N; i++)
 749       dt.z[i] = -dt.z[i];
 750    for (i = 0; i < N; i += 2) {
 751       int thisres = 0;
 752       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
 753                           "vfmadd132pd %%xmm7, %%xmm8, %%xmm9;"
 754                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 755                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 756       for (j = 0; j < 2; j++)
 757          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 758       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
 759                           "vfmadd132pd (%2), %%xmm8, %%xmm9;"
 760                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 761                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 762       for (j = 0; j < 2; j++)
 763          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 764       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
 765                           "vfmadd213pd %%xmm7, %%xmm8, %%xmm9;"
 766                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 767                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 768       for (j = 0; j < 2; j++)
 769          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 770       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
 771                           "vfmadd213pd (%3), %%xmm8, %%xmm9;"
 772                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 773                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 774       for (j = 0; j < 2; j++)
 775          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 776       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
 777                           "vfmadd231pd %%xmm7, %%xmm8, %%xmm9;"
 778                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 779                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 780       for (j = 0; j < 2; j++)
 781          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 782       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
 783                           "vfmadd231pd (%2), %%xmm8, %%xmm9;"
 784                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 785                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 786       for (j = 0; j < 2; j++)
 787          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 788       if (thisres) {
 789          printf( "Failure 5 %d", i );
 790          for (j = 0; j < 2; j++)
 791             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
 792          printf( "\n" );
 793       }
 794       res |= thisres;
 795       thisres = 0;
 796       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
 797                           "vfnmsub132pd %%xmm7, %%xmm8, %%xmm9;"
 798                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 799                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 800       for (j = 0; j < 2; j++)
 801          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 802       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
 803                           "vfnmsub132pd (%2), %%xmm8, %%xmm9;"
 804                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 805                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 806       for (j = 0; j < 2; j++)
 807          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 808       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
 809                           "vfnmsub213pd %%xmm7, %%xmm8, %%xmm9;"
 810                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 811                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 812       for (j = 0; j < 2; j++)
 813          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 814       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
 815                           "vfnmsub213pd (%3), %%xmm8, %%xmm9;"
 816                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 817                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 818       for (j = 0; j < 2; j++)
 819          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 820       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
 821                           "vfnmsub231pd %%xmm7, %%xmm8, %%xmm9;"
 822                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 823                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 824       for (j = 0; j < 2; j++)
 825          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 826       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
 827                           "vfnmsub231pd (%2), %%xmm8, %%xmm9;"
 828                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 829                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 830       for (j = 0; j < 2; j++)
 831          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 832       if (thisres) {
 833          printf( "Failure 6 %d", i );
 834          for (j = 0; j < 2; j++)
 835             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
 836          printf( "\n" );
 837       }
 838       res |= thisres;
 839    }
 840    for (i = 0; i < N; i++)
 841       dt.z[i] = -dt.z[i];
 842    for (i = 0; i < N; i += 2) {
 843       int thisres = 0;
 844       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
 845                           "vfmsub132pd %%xmm7, %%xmm8, %%xmm9;"
 846                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 847                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 848       for (j = 0; j < 2; j++)
 849          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 850       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
 851                           "vfmsub132pd (%2), %%xmm8, %%xmm9;"
 852                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 853                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 854       for (j = 0; j < 2; j++)
 855          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 856       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
 857                           "vfmsub213pd %%xmm7, %%xmm8, %%xmm9;"
 858                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 859                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 860       for (j = 0; j < 2; j++)
 861          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 862       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
 863                           "vfmsub213pd (%3), %%xmm8, %%xmm9;"
 864                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 865                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 866       for (j = 0; j < 2; j++)
 867          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 868       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
 869                           "vfmsub231pd %%xmm7, %%xmm8, %%xmm9;"
 870                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 871                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 872       for (j = 0; j < 2; j++)
 873          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 874       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
 875                           "vfmsub231pd (%2), %%xmm8, %%xmm9;"
 876                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 877                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 878       for (j = 0; j < 2; j++)
 879          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 880       if (thisres) {
 881          printf( "Failure 7 %d", i );
 882          for (j = 0; j < 2; j++)
 883             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
 884          printf( "\n" );
 885       }
 886       res |= thisres;
 887       thisres = 0;
 888       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
 889                           "vfnmadd132pd %%xmm7, %%xmm8, %%xmm9;"
 890                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 891                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 892       for (j = 0; j < 2; j++)
 893          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 894       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
 895                           "vfnmadd132pd (%2), %%xmm8, %%xmm9;"
 896                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 897                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 898       for (j = 0; j < 2; j++)
 899          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 900       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
 901                           "vfnmadd213pd %%xmm7, %%xmm8, %%xmm9;"
 902                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 903                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 904       for (j = 0; j < 2; j++)
 905          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 906       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
 907                           "vfnmadd213pd (%3), %%xmm8, %%xmm9;"
 908                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 909                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 910       for (j = 0; j < 2; j++)
 911          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 912       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
 913                           "vfnmadd231pd %%xmm7, %%xmm8, %%xmm9;"
 914                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 915                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 916       for (j = 0; j < 2; j++)
 917          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 918       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
 919                           "vfnmadd231pd (%2), %%xmm8, %%xmm9;"
 920                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 921                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 922       for (j = 0; j < 2; j++)
 923          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
 924       if (thisres) {
 925          printf( "Failure 8 %d", i );
 926          for (j = 0; j < 2; j++)
 927             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
 928          printf( "\n" );
 929       }
 930       res |= thisres;
 931    }
 932    for (i = 1; i < N; i += 2)
 933       dt.z[i] = -dt.z[i];
 934    for (i = 0; i < N; i += 2) {
 935       int thisres = 0;
 936       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
 937                           "vfmaddsub132pd %%xmm7, %%xmm8, %%xmm9;"
 938                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 939                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 940       for (j = 0; j < 2; j++)
 941          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 942       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
 943                           "vfmaddsub132pd (%2), %%xmm8, %%xmm9;"
 944                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 945                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 946       for (j = 0; j < 2; j++)
 947          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 948       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
 949                           "vfmaddsub213pd %%xmm7, %%xmm8, %%xmm9;"
 950                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 951                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 952       for (j = 0; j < 2; j++)
 953          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 954       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
 955                           "vfmaddsub213pd (%3), %%xmm8, %%xmm9;"
 956                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 957                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 958       for (j = 0; j < 2; j++)
 959          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 960       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
 961                           "vfmaddsub231pd %%xmm7, %%xmm8, %%xmm9;"
 962                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 963                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 964       for (j = 0; j < 2; j++)
 965          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 966       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
 967                           "vfmaddsub231pd (%2), %%xmm8, %%xmm9;"
 968                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 969                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 970       for (j = 0; j < 2; j++)
 971          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 972       if (thisres) {
 973          printf( "Failure 9 %d", i );
 974          for (j = 0; j < 2; j++)
 975             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
 976          printf( "\n" );
 977       }
 978       res |= thisres;
 979    }
 980    for (i = 0; i < N; i++)
 981       dt.z[i] = -dt.z[i];
 982    for (i = 0; i < N; i += 2) {
 983       int thisres = 0;
 984       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%3), %%xmm8;"
 985                           "vfmsubadd132pd %%xmm7, %%xmm8, %%xmm9;"
 986                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 987                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 988       for (j = 0; j < 2; j++)
 989          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 990       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm8;"
 991                           "vfmsubadd132pd (%2), %%xmm8, %%xmm9;"
 992                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 993                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
 994       for (j = 0; j < 2; j++)
 995          thisres |= test( dt.res[i+j], dt.expected[i+j] );
 996       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%3), %%xmm7; vmovapd (%2), %%xmm8;"
 997                           "vfmsubadd213pd %%xmm7, %%xmm8, %%xmm9;"
 998                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
 999                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1000       for (j = 0; j < 2; j++)
1001          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1002       __asm __volatile__ ("vmovapd (%1), %%xmm9; vmovapd (%2), %%xmm8;"
1003                           "vfmsubadd213pd (%3), %%xmm8, %%xmm9;"
1004                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1005                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1006       for (j = 0; j < 2; j++)
1007          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1008       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%2), %%xmm7; vmovapd (%1), %%xmm8;"
1009                           "vfmsubadd231pd %%xmm7, %%xmm8, %%xmm9;"
1010                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1011                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1012       for (j = 0; j < 2; j++)
1013          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1014       __asm __volatile__ ("vmovapd (%3), %%xmm9; vmovapd (%1), %%xmm8;"
1015                           "vfmsubadd231pd (%2), %%xmm8, %%xmm9;"
1016                           "vmovapd %%xmm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1017                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1018       for (j = 0; j < 2; j++)
1019          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1020       if (thisres) {
1021          printf( "Failure 10 %d", i );
1022          for (j = 0; j < 2; j++)
1023             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
1024          printf( "\n" );
1025       }
1026       res |= thisres;
1027    }
1028    for (i = 1; i < N; i += 2)
1029       dt.z[i] = -dt.z[i];
1030    for (i = 0; i < N; i += 4) {
1031       int thisres = 0;
1032       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%3), %%ymm8;"
1033                           "vfmadd132pd %%ymm7, %%ymm8, %%ymm9;"
1034                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1035                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1036       for (j = 0; j < 4; j++)
1037          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1038       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm8;"
1039                           "vfmadd132pd (%2), %%ymm8, %%ymm9;"
1040                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1041                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1042       for (j = 0; j < 4; j++)
1043          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1044       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm7; vmovapd (%2), %%ymm8;"
1045                           "vfmadd213pd %%ymm7, %%ymm8, %%ymm9;"
1046                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1047                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1048       for (j = 0; j < 4; j++)
1049          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1050       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm8;"
1051                           "vfmadd213pd (%3), %%ymm8, %%ymm9;"
1052                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1053                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1054       for (j = 0; j < 4; j++)
1055          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1056       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%1), %%ymm8;"
1057                           "vfmadd231pd %%ymm7, %%ymm8, %%ymm9;"
1058                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1059                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1060       for (j = 0; j < 4; j++)
1061          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1062       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%1), %%ymm8;"
1063                           "vfmadd231pd (%2), %%ymm8, %%ymm9;"
1064                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1065                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1066       for (j = 0; j < 4; j++)
1067          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1068       if (thisres) {
1069          printf( "Failure 11 %d", i );
1070          for (j = 0; j < 4; j++)
1071             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
1072          printf( "\n" );
1073       }
1074       res |= thisres;
1075       thisres = 0;
1076       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%3), %%ymm8;"
1077                           "vfnmsub132pd %%ymm7, %%ymm8, %%ymm9;"
1078                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1079                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1080       for (j = 0; j < 4; j++)
1081          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1082       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm8;"
1083                           "vfnmsub132pd (%2), %%ymm8, %%ymm9;"
1084                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1085                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1086       for (j = 0; j < 4; j++)
1087          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1088       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm7; vmovapd (%2), %%ymm8;"
1089                           "vfnmsub213pd %%ymm7, %%ymm8, %%ymm9;"
1090                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1091                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1092       for (j = 0; j < 4; j++)
1093          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1094       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm8;"
1095                           "vfnmsub213pd (%3), %%ymm8, %%ymm9;"
1096                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1097                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1098       for (j = 0; j < 4; j++)
1099          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1100       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%1), %%ymm8;"
1101                           "vfnmsub231pd %%ymm7, %%ymm8, %%ymm9;"
1102                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1103                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1104       for (j = 0; j < 4; j++)
1105          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1106       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%1), %%ymm8;"
1107                           "vfnmsub231pd (%2), %%ymm8, %%ymm9;"
1108                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1109                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1110       for (j = 0; j < 4; j++)
1111          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1112       if (thisres) {
1113          printf( "Failure 12 %d", i );
1114          for (j = 0; j < 4; j++)
1115             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
1116          printf( "\n" );
1117       }
1118       res |= thisres;
1119    }
1120    for (i = 0; i < N; i++)
1121       dt.z[i] = -dt.z[i];
1122    for (i = 0; i < N; i += 4) {
1123       int thisres = 0;
1124       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%3), %%ymm8;"
1125                           "vfmsub132pd %%ymm7, %%ymm8, %%ymm9;"
1126                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1127                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1128       for (j = 0; j < 4; j++)
1129          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1130       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm8;"
1131                           "vfmsub132pd (%2), %%ymm8, %%ymm9;"
1132                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1133                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1134       for (j = 0; j < 4; j++)
1135          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1136       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm7; vmovapd (%2), %%ymm8;"
1137                           "vfmsub213pd %%ymm7, %%ymm8, %%ymm9;"
1138                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1139                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1140       for (j = 0; j < 4; j++)
1141          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1142       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm8;"
1143                           "vfmsub213pd (%3), %%ymm8, %%ymm9;"
1144                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1145                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1146       for (j = 0; j < 4; j++)
1147          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1148       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%1), %%ymm8;"
1149                           "vfmsub231pd %%ymm7, %%ymm8, %%ymm9;"
1150                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1151                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1152       for (j = 0; j < 4; j++)
1153          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1154       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%1), %%ymm8;"
1155                           "vfmsub231pd (%2), %%ymm8, %%ymm9;"
1156                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1157                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1158       for (j = 0; j < 4; j++)
1159          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1160       if (thisres) {
1161          printf( "Failure 13 %d", i );
1162          for (j = 0; j < 4; j++)
1163             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
1164          printf( "\n" );
1165       }
1166       res |= thisres;
1167       thisres = 0;
1168       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%3), %%ymm8;"
1169                           "vfnmadd132pd %%ymm7, %%ymm8, %%ymm9;"
1170                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1171                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1172       for (j = 0; j < 4; j++)
1173          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1174       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm8;"
1175                           "vfnmadd132pd (%2), %%ymm8, %%ymm9;"
1176                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1177                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1178       for (j = 0; j < 4; j++)
1179          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1180       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm7; vmovapd (%2), %%ymm8;"
1181                           "vfnmadd213pd %%ymm7, %%ymm8, %%ymm9;"
1182                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1183                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1184       for (j = 0; j < 4; j++)
1185          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1186       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm8;"
1187                           "vfnmadd213pd (%3), %%ymm8, %%ymm9;"
1188                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1189                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1190       for (j = 0; j < 4; j++)
1191          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1192       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%1), %%ymm8;"
1193                           "vfnmadd231pd %%ymm7, %%ymm8, %%ymm9;"
1194                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1195                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1196       for (j = 0; j < 4; j++)
1197          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1198       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%1), %%ymm8;"
1199                           "vfnmadd231pd (%2), %%ymm8, %%ymm9;"
1200                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1201                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1202       for (j = 0; j < 4; j++)
1203          thisres |= test( -dt.res[i+j], dt.expected[i+j] );
1204       if (thisres) {
1205          printf( "Failure 14 %d", i );
1206          for (j = 0; j < 4; j++)
1207             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
1208          printf( "\n" );
1209       }
1210       res |= thisres;
1211    }
1212    for (i = 1; i < N; i += 2)
1213       dt.z[i] = -dt.z[i];
1214    for (i = 0; i < N; i += 4) {
1215       int thisres = 0;
1216       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%3), %%ymm8;"
1217                           "vfmaddsub132pd %%ymm7, %%ymm8, %%ymm9;"
1218                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1219                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1220       for (j = 0; j < 4; j++)
1221          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1222       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm8;"
1223                           "vfmaddsub132pd (%2), %%ymm8, %%ymm9;"
1224                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1225                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1226       for (j = 0; j < 4; j++)
1227          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1228       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm7; vmovapd (%2), %%ymm8;"
1229                           "vfmaddsub213pd %%ymm7, %%ymm8, %%ymm9;"
1230                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1231                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1232       for (j = 0; j < 4; j++)
1233          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1234       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm8;"
1235                           "vfmaddsub213pd (%3), %%ymm8, %%ymm9;"
1236                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1237                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1238       for (j = 0; j < 4; j++)
1239          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1240       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%1), %%ymm8;"
1241                           "vfmaddsub231pd %%ymm7, %%ymm8, %%ymm9;"
1242                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1243                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1244       for (j = 0; j < 4; j++)
1245          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1246       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%1), %%ymm8;"
1247                           "vfmaddsub231pd (%2), %%ymm8, %%ymm9;"
1248                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1249                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1250       for (j = 0; j < 4; j++)
1251          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1252       if (thisres) {
1253          printf( "Failure 15 %d", i );
1254          for (j = 0; j < 4; j++)
1255             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
1256          printf( "\n" );
1257       }
1258       res |= thisres;
1259    }
1260    for (i = 0; i < N; i++)
1261       dt.z[i] = -dt.z[i];
1262    for (i = 0; i < N; i += 4) {
1263       int thisres = 0;
1264       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%3), %%ymm8;"
1265                           "vfmsubadd132pd %%ymm7, %%ymm8, %%ymm9;"
1266                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1267                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1268       for (j = 0; j < 4; j++)
1269          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1270       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm8;"
1271                           "vfmsubadd132pd (%2), %%ymm8, %%ymm9;"
1272                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1273                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1274       for (j = 0; j < 4; j++)
1275          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1276       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%3), %%ymm7; vmovapd (%2), %%ymm8;"
1277                           "vfmsubadd213pd %%ymm7, %%ymm8, %%ymm9;"
1278                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1279                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1280       for (j = 0; j < 4; j++)
1281          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1282       __asm __volatile__ ("vmovapd (%1), %%ymm9; vmovapd (%2), %%ymm8;"
1283                           "vfmsubadd213pd (%3), %%ymm8, %%ymm9;"
1284                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1285                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1286       for (j = 0; j < 4; j++)
1287          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1288       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%2), %%ymm7; vmovapd (%1), %%ymm8;"
1289                           "vfmsubadd231pd %%ymm7, %%ymm8, %%ymm9;"
1290                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1291                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1292       for (j = 0; j < 4; j++)
1293          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1294       __asm __volatile__ ("vmovapd (%3), %%ymm9; vmovapd (%1), %%ymm8;"
1295                           "vfmsubadd231pd (%2), %%ymm8, %%ymm9;"
1296                           "vmovapd %%ymm9, (%0)" : : "r" (&dt.res[i]), "r" (&dt.x[i]),
1297                                                      "r" (&dt.y[i]), "r" (&dt.z[i]) : "xmm7", "xmm8", "xmm9");
1298       for (j = 0; j < 4; j++)
1299          thisres |= test( dt.res[i+j], dt.expected[i+j] );
1300       if (thisres) {
1301          printf( "Failure 16 %d", i );
1302          for (j = 0; j < 4; j++)
1303             printf( " %a %a", dt.res[i+j], dt.expected[i+j] );
1304          printf( "\n" );
1305       }
1306       res |= thisres;
1307    }
1308    for (i = 1; i < N; i += 2)
1309       dt.z[i] = -dt.z[i];
1310    return res;
1311 }
1312
1313 int main( )
1314 {
1315    int res = 0;
1316    int i = 0;
1317    plus_zero = 0.0;
1318    __asm __volatile__ ("" : : "r" (&plus_zero) : "memory");
1319    nan_value = plus_zero / plus_zero;
1320    plus_infty = 3.40282346638528859812e+38F * 16.0F;
1321    minus_infty = -plus_infty;
1322 #define TEST_F( a, b, c, d ) \
1323    do {                         \
1324       ft.x[i] = a;              \
1325       ft.y[i] = b;              \
1326       ft.z[i] = c;              \
1327       ft.expected[i] = d;       \
1328       i++;                      \
1329    } while (0)
1330    TEST_F( 1.0, 2.0, 3.0, 5.0 );
1331    TEST_F( nan_value, 2.0, 3.0, nan_value );
1332    TEST_F( 1.0, nan_value, 3.0, nan_value );
1333    TEST_F( 1.0, 2.0, nan_value, nan_value );
1334    TEST_F( plus_infty, 0.0, nan_value, nan_value );
1335    TEST_F( minus_infty, 0.0, nan_value, nan_value );
1336    TEST_F( 0.0, plus_infty, nan_value, nan_value );
1337    TEST_F( 0.0, minus_infty, nan_value, nan_value );
1338    TEST_F( plus_infty, 0.0, 1.0, nan_value );
1339    TEST_F( minus_infty, 0.0, 1.0, nan_value );
1340    TEST_F( 0.0, plus_infty, 1.0, nan_value );
1341    TEST_F( 0.0, minus_infty, 1.0, nan_value );
1342    TEST_F( plus_infty, plus_infty, minus_infty, nan_value );
1343    TEST_F( minus_infty, plus_infty, plus_infty, nan_value );
1344    TEST_F( plus_infty, minus_infty, plus_infty, nan_value );
1345    TEST_F( minus_infty, minus_infty, minus_infty, nan_value );
1346    TEST_F( plus_infty, 3.5L, minus_infty, nan_value );
1347    TEST_F( minus_infty, -7.5L, minus_infty, nan_value );
1348    TEST_F( -13.5L, plus_infty, plus_infty, nan_value );
1349    TEST_F( minus_infty, 7.5L, plus_infty, nan_value );
1350    TEST_F( 1.25L, 0.75L, 0.0625L, 1.0L );
1351    TEST_F( -3.40282346638528859812e+38F, -3.40282346638528859812e+38F, minus_infty, minus_infty );
1352    TEST_F( 3.40282346638528859812e+38F / 2, 3.40282346638528859812e+38F / 2, minus_infty, minus_infty );
1353    TEST_F( -3.40282346638528859812e+38F, 3.40282346638528859812e+38F, plus_infty, plus_infty );
1354    TEST_F( 3.40282346638528859812e+38F / 2, -3.40282346638528859812e+38F / 4, plus_infty, plus_infty );
1355    TEST_F( plus_infty, 4, plus_infty, plus_infty );
1356    TEST_F( 2, minus_infty, minus_infty, minus_infty );
1357    TEST_F( minus_infty, minus_infty, plus_infty, plus_infty );
1358    TEST_F( plus_infty, minus_infty, minus_infty, minus_infty );
1359    TEST_F( 0x1.7ff8p+13, 0x1.000002p+0, 0x1.ffffp-24, 0x1.7ff802p+13 );
1360    TEST_F( 0x1.fffp+0, 0x1.00001p+0, -0x1.fffp+0, 0x1.fffp-20 );
1361    TEST_F( 0x1.9abcdep+127, 0x0.9abcdep-126, -0x1.f08948p+0, 0x1.bb421p-25 );
1362    TEST_F( 0x1.9abcdep+100, 0x0.9abcdep-126, -0x1.f08948p-27, 0x1.bb421p-52 );
1363    TEST_F( 0x1.fffffep+127, 0x1.001p+0, -0x1.fffffep+127, 0x1.fffffep+115 );
1364    TEST_F( -0x1.fffffep+127, 0x1.fffffep+0, 0x1.fffffep+127, -0x1.fffffap+127 );
1365    TEST_F( 0x1.fffffep+127, 2.0, -0x1.fffffep+127, 0x1.fffffep+127 );
1366
1367    res |= test_fmaf( );
1368    i = 0;
1369 #define TEST( a, b, c, d ) \
1370    do {                         \
1371       dt.x[i] = a;              \
1372       dt.y[i] = b;              \
1373       dt.z[i] = c;              \
1374       dt.expected[i] = d;       \
1375       i++;                      \
1376    } while (0)
1377    TEST( 1.0, 2.0, 3.0, 5.0 );
1378    TEST( nan_value, 2.0, 3.0, nan_value );
1379    TEST( 1.0, nan_value, 3.0, nan_value );
1380    TEST( 1.0, 2.0, nan_value, nan_value );
1381    TEST( plus_infty, 0.0, nan_value, nan_value );
1382    TEST( minus_infty, 0.0, nan_value, nan_value );
1383    TEST( 0.0, plus_infty, nan_value, nan_value );
1384    TEST( 0.0, minus_infty, nan_value, nan_value );
1385    TEST( plus_infty, 0.0, 1.0, nan_value );
1386    TEST( minus_infty, 0.0, 1.0, nan_value );
1387    TEST( 0.0, plus_infty, 1.0, nan_value );
1388    TEST( 0.0, minus_infty, 1.0, nan_value );
1389    TEST( plus_infty, plus_infty, minus_infty, nan_value );
1390    TEST( minus_infty, plus_infty, plus_infty, nan_value );
1391    TEST( plus_infty, minus_infty, plus_infty, nan_value );
1392    TEST( minus_infty, minus_infty, minus_infty, nan_value );
1393    TEST( plus_infty, 3.5L, minus_infty, nan_value );
1394    TEST( minus_infty, -7.5L, minus_infty, nan_value );
1395    TEST( -13.5L, plus_infty, plus_infty, nan_value );
1396    TEST( minus_infty, 7.5L, plus_infty, nan_value );
1397    TEST( 1.25L, 0.75L, 0.0625L, 1.0L );
1398    TEST( -1.79769313486231570815e+308L, -1.79769313486231570815e+308L, minus_infty, minus_infty );
1399    TEST( 1.79769313486231570815e+308L / 2, 1.79769313486231570815e+308L / 2, minus_infty, minus_infty );
1400    TEST( -1.79769313486231570815e+308L, 1.79769313486231570815e+308L, plus_infty, plus_infty );
1401    TEST( 1.79769313486231570815e+308L / 2, -1.79769313486231570815e+308L / 4, plus_infty, plus_infty );
1402    TEST( plus_infty, 4, plus_infty, plus_infty );
1403    TEST( 2, minus_infty, minus_infty, minus_infty );
1404    TEST( minus_infty, minus_infty, plus_infty, plus_infty );
1405    TEST( plus_infty, minus_infty, minus_infty, minus_infty );
1406    TEST( 0x1.7fp+13, 0x1.0000000000001p+0, 0x1.ffep-48, 0x1.7f00000000001p+13 );
1407    TEST( 0x1.fffp+0, 0x1.0000000000001p+0, -0x1.fffp+0, 0x1.fffp-52 );
1408    TEST( 0x1.0000002p+0, 0x1.ffffffcp-1, 0x1p-300, 1.0 );
1409    TEST( 0x1.0000002p+0, 0x1.ffffffcp-1, -0x1p-300, 0x1.fffffffffffffp-1 );
1410    TEST( 0x1.deadbeef2feedp+1023, 0x0.deadbeef2feedp-1022, -0x1.a05f8c01a4bfbp+1, 0x1.0989687bc9da4p-53 );
1411    TEST( 0x1.deadbeef2feedp+900, 0x0.deadbeef2feedp-1022, -0x1.a05f8c01a4bfbp-122, 0x1.0989687bc9da4p-176 );
1412    TEST( 0x1.fffffffffffffp+1023, 0x1.001p+0, -0x1.fffffffffffffp+1023, 0x1.fffffffffffffp+1011 );
1413    TEST( -0x1.fffffffffffffp+1023, 0x1.fffffffffffffp+0, 0x1.fffffffffffffp+1023, -0x1.ffffffffffffdp+1023 );
1414    TEST( 0x1.fffffffffffffp+1023, 2.0, -0x1.fffffffffffffp+1023, 0x1.fffffffffffffp+1023 );
1415    TEST( 0x1.6a09e667f3bccp-538, 0x1.6a09e667f3bccp-538, 0.0, 0.0 );
1416    TEST( 0x1.deadbeef2feedp-495, 0x1.deadbeef2feedp-495, -0x1.bf86a5786a574p-989, 0x0.0000042625a1fp-1022 );
1417    TEST( 0x1.deadbeef2feedp-503, 0x1.deadbeef2feedp-503, -0x1.bf86a5786a574p-1005, 0x0.0000000004262p-1022 );
1418    TEST( 0x1p-537, 0x1p-538, 0x1p-1074, 0x0.0000000000002p-1022 );
1419    TEST( 0x1.7fffff8p-968, 0x1p-106, 0x0.000001p-1022, 0x0.0000010000001p-1022 );
1420    TEST( 0x1.4000004p-967, 0x1p-106, 0x0.000001p-1022, 0x0.0000010000003p-1022 );
1421    TEST( 0x1.4p-967, -0x1p-106, -0x0.000001p-1022, -0x0.0000010000002p-1022 );
1422    TEST( -0x1.19cab66d73e17p-959, 0x1.c7108a8c5ff51p-107, -0x0.80b0ad65d9b64p-1022, -0x0.80b0ad65d9d59p-1022 );
1423    TEST( -0x1.d2eaed6e8e9d3p-979, -0x1.4e066c62ac9ddp-63, -0x0.9245e6b003454p-1022, -0x0.9245c09c5fb5dp-1022 );
1424    TEST( 0x1.153d650bb9f06p-907, 0x1.2d01230d48407p-125, -0x0.b278d5acfc3cp-1022, -0x0.b22757123bbe9p-1022 );
1425    TEST( -0x1.fffffffffffffp-711, 0x1.fffffffffffffp-275, 0x1.fffffe00007ffp-983, 0x1.7ffffe00007ffp-983 );
1426
1427    res |= test_fma( );
1428    if (res == 0)
1429       printf( "Testing successful\n");
1430    return 0;
1431 }