gcc/testsuite/gcc.dg/vect/fast-math-vect-call-2.c

   1 /* { dg-require-effective-target vect_double } */
   2 /* { dg-additional-options "-ffast-math" } */
   3 /* { dg-additional-options "--param vect-epilogues-nomask=0" } */
   4
   5 #include "tree-vect.h"
   6
   7 extern long int lrint (double);
   8 extern void abort (void);
   9 long int a[64];
  10 double b[64];
  11
  12 __attribute__((noinline, noclone)) void
  13 f1 (int n)
  14 {
  15   int i;
  16   for (i = 0; i < n; i++)
  17     {
  18       a[4 * i + 0] = lrint (b[4 * i + 0]) + 1;
  19       a[4 * i + 1] = lrint (b[4 * i + 1]) + 2;
  20       a[4 * i + 2] = lrint (b[4 * i + 2]) + 3;
  21       a[4 * i + 3] = lrint (b[4 * i + 3]) + 4;
  22     }
  23 }
  24
  25 __attribute__((noinline, noclone)) void
  26 f2 (int n)
  27 {
  28   int i;
  29   for (i = 0; i < 2 * n; i++)
  30     {
  31       a[2 * i + 0] = lrint (b[2 * i + 0]) + 1;
  32       a[2 * i + 1] = lrint (b[2 * i + 1]) + 2;
  33     }
  34 }
  35
  36 __attribute__((noinline, noclone)) void
  37 f3 (void)
  38 {
  39   int i;
  40   for (i = 0; i < 64; i++)
  41     a[i] = lrint (b[i]) + 1;
  42 }
  43
  44 __attribute__((noinline, noclone)) void
  45 f4 (int n)
  46 {
  47   int i;
  48   for (i = 0; i < n; i++)
  49     {
  50       a[4 * i + 0] = lrint (b[4 * i + 0]);
  51       a[4 * i + 1] = lrint (b[4 * i + 1]);
  52       a[4 * i + 2] = lrint (b[4 * i + 2]);
  53       a[4 * i + 3] = lrint (b[4 * i + 3]);
  54     }
  55 }
  56
  57 __attribute__((noinline, noclone)) void
  58 f5 (int n)
  59 {
  60   int i;
  61   for (i = 0; i < 2 * n; i++)
  62     {
  63       a[2 * i + 0] = lrint (b[2 * i + 0]);
  64       a[2 * i + 1] = lrint (b[2 * i + 1]);
  65     }
  66 }
  67
  68 __attribute__((noinline, noclone)) void
  69 f6 (void)
  70 {
  71   int i;
  72   for (i = 0; i < 64; i++)
  73     a[i] = lrint (b[i]);
  74 }
  75
  76 __attribute__((noinline, noclone)) int
  77 main1 ()
  78 {
  79   int i;
  80
  81   for (i = 0; i < 64; i++)
  82     {
  83       asm ("");
  84       b[i] = ((i & 1) ? -4 * i : 4 * i) + 0.25;
  85     }
  86   f1 (16);
  87 #pragma GCC novector
  88   for (i = 0; i < 64; i++)
  89     if (a[i] != ((i & 1) ? -4 * i : 4 * i) + 1 + (i & 3))
  90       abort ();
  91     else
  92       a[i] = 131.25;
  93   f2 (16);
  94 #pragma GCC novector
  95   for (i = 0; i < 64; i++)
  96     if (a[i] != ((i & 1) ? -4 * i : 4 * i) + 1 + (i & 1))
  97       abort ();
  98     else
  99       a[i] = 131.25;
 100   f3 ();
 101 #pragma GCC novector
 102   for (i = 0; i < 64; i++)
 103     if (a[i] != ((i & 1) ? -4 * i : 4 * i) + 1)
 104       abort ();
 105     else
 106       a[i] = 131.25;
 107   f4 (16);
 108 #pragma GCC novector
 109   for (i = 0; i < 64; i++)
 110     if (a[i] != ((i & 1) ? -4 * i : 4 * i))
 111       abort ();
 112     else
 113       a[i] = 131.25;
 114   f5 (16);
 115 #pragma GCC novector
 116   for (i = 0; i < 64; i++)
 117     if (a[i] != ((i & 1) ? -4 * i : 4 * i))
 118       abort ();
 119     else
 120       a[i] = 131.25;
 121   f6 ();
 122 #pragma GCC novector
 123   for (i = 0; i < 64; i++)
 124     if (a[i] != ((i & 1) ? -4 * i : 4 * i))
 125       abort ();
 126   return 0;
 127 }
 128
 129 int
 130 main ()
 131 {
 132   check_vect ();
 133   return main1 ();
 134 }
 135
 136 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 6 "vect" { target vect_call_lrint } } } */
 137 /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 6 "vect" { target vect_call_lrint } } } */