gcc/testsuite/gcc.target/i386/sse4_1-round.h

   1 #include <smmintrin.h>
   2 #include <math.h>
   3
   4 #define NUM 64
   5
   6 static void
   7 init_round (FP_T *src)
   8 {
   9   int i, sign = 1;
  10   FP_T f = rand ();
  11
  12   for (i = 0; i < NUM; i++)
  13     {
  14       src[i] = (i + 1)* f * M_PI * sign;
  15       if (i < (NUM / 2))
  16         {
  17           if ((i % 6) == 0)
  18             f = f * src[i];
  19         }
  20       else if (i == (NUM / 2))
  21         f = rand ();
  22       else if ((i % 6) == 0)
  23         f = 1 / (f * (i + 1) * src[i] * M_PI *sign);
  24       sign = -sign;
  25     }
  26 }
  27
  28 static FP_T
  29 do_round (FP_T f, int type)
  30 {
  31   short saved_cw, new_cw, clr_mask;
  32   FP_T ret;
  33
  34   if ((type & 4))
  35     {
  36       type = 0;
  37       clr_mask = 0xFFFF;
  38     }
  39   else
  40     {
  41       type = 0x003F | ((type & 3) << 10);
  42       clr_mask = ~0x0C3F;
  43     }
  44
  45   __asm__ ("fld" ASM_SUFFIX " %0" : : "m" (*&f));
  46
  47   __asm__ ("fstcw %0" : "=m" (*&saved_cw));
  48   new_cw = saved_cw & clr_mask;
  49   new_cw |= type;
  50   __asm__ ("fldcw %0" : : "m" (*&new_cw));
  51
  52   __asm__ ("frndint\n"
  53            "fstp" ASM_SUFFIX " %0\n" : "=m" (*&ret));
  54   __asm__ ("fldcw %0" : : "m" (*&saved_cw));
  55   return ret;
  56 }
  57
  58 static void
  59 sse4_1_test (void)
  60 {
  61   int i;
  62   FP_T f;
  63   union
  64     {
  65       VEC_T x[NUM / LOOP_INCREMENT];
  66       FP_T f[NUM];
  67     } dst, src;
  68
  69   init_round (src.f);
  70
  71   for (i = 0; i < NUM / LOOP_INCREMENT; i++)
  72     dst.x[i] =  ROUND_INTRIN (src.x[i], ROUND_MODE);
  73
  74   for (i = 0; i < NUM; i += CHECK_LOOP_INCREMENT)
  75     {
  76       f = do_round (src.f[i], CHECK_ROUND_MODE);
  77      if (f != dst.f[i])
  78        abort ();
  79     }
  80
  81   if (_MM_FROUND_TO_NEAREST_INT != 0x00
  82       || _MM_FROUND_TO_NEG_INF != 0x01
  83       || _MM_FROUND_TO_POS_INF != 0x02
  84       || _MM_FROUND_TO_ZERO != 0x03
  85       || _MM_FROUND_CUR_DIRECTION != 0x04
  86       || _MM_FROUND_RAISE_EXC != 0x00
  87       || _MM_FROUND_NO_EXC != 0x08
  88       || _MM_FROUND_NINT != 0x00
  89       || _MM_FROUND_FLOOR != 0x01
  90       || _MM_FROUND_CEIL != 0x02
  91       || _MM_FROUND_TRUNC != 0x03
  92       || _MM_FROUND_RINT != 0x04
  93       || _MM_FROUND_NEARBYINT != 0x0C)
  94     abort ();
  95 }