gcc/testsuite/gcc.target/powerpc/swaps-p8-14.c

   1 /* { dg-do compile } */
   2 /* { dg-require-effective-target powerpc_p8vector_ok } */
   3 /* { dg-options "-mdejagnu-cpu=power8 -O3" } */
   4 /* { dg-final { scan-assembler "lxvd2x" } } */
   5 /* { dg-final { scan-assembler "stxvd2x" } } */
   6 /* { dg-final { scan-assembler "stxsdx" } } */
   7 /* { dg-final { scan-assembler-times "xxpermdi" 1 } } */
   8
   9 /* The only xxpermdi expected is for the vec_splats.  */
  10
  11 #include <altivec.h>
  12 void abort ();
  13
  14 #define N 4096
  15 long long ca[N] __attribute__((aligned(16)));
  16 long long cb[N] __attribute__((aligned(16)));
  17 long long cc[N] __attribute__((aligned(16)));
  18 long long cd[N] __attribute__((aligned(16)));
  19 long long x;
  20
  21 __attribute__((noinline)) void foo ()
  22 {
  23   int i;
  24   vector long long va, vb, vc, vd, tmp;
  25   volatile unsigned long long three = 3;
  26   vector unsigned long long threes = vec_splats (three);
  27   for (i = 0; i < N; i+=2) {
  28     vb = vec_vsx_ld (0, (vector long long *)&cb[i]);
  29     vc = vec_vsx_ld (0, (vector long long *)&cc[i]);
  30     vd = vec_vsx_ld (0, (vector long long *)&cd[i]);
  31     tmp = vec_add (vb, vc);
  32     tmp = vec_sub (tmp, vd);
  33     tmp = vec_sra (tmp, threes);
  34     x = vec_extract (tmp, 0);
  35     vec_vsx_st (tmp, 0, (vector long long *)&ca[i]);
  36   }
  37 }
  38
  39 int main ()
  40 {
  41   foo ();
  42   return 0;
  43 }