gcc/testsuite/gcc.target/powerpc/recip-4.c

   1 /* { dg-do compile { target { powerpc*-*-* } } } */
   2 /* { dg-require-effective-target powerpc_fprs } */
   3 /* { dg-options "-O3 -ftree-vectorize -mrecip -ffast-math -mdejagnu-cpu=power7 -fno-unroll-loops" } */
   4 /* { dg-final { scan-assembler-times "xvrsqrtedp" 1 } } */
   5 /* { dg-final { scan-assembler-times "xvmuldp" 2 } } */
   6 /* { dg-final { scan-assembler-times "xvnmsub.dp" 2 } } */
   7 /* { dg-final { scan-assembler-times "xvmadd.dp" 3 } } */
   8 /* { dg-final { scan-assembler-times "xvrsqrtesp" 1 } } */
   9 /* { dg-final { scan-assembler-times "xvmulsp" 2 } } */
  10 /* { dg-final { scan-assembler-times "xvnmsub.sp" 1 } } */
  11 /* { dg-final { scan-assembler-times "xvmadd.sp" 1 } } */
  12
  13 #define SIZE 1024
  14
  15 extern double a_d[SIZE] __attribute__((__aligned__(32)));
  16 extern double b_d[SIZE] __attribute__((__aligned__(32)));
  17
  18 void
  19 vectorize_rsqrt_d (void)
  20 {
  21   int i;
  22
  23   for (i = 0; i < SIZE; i++)
  24     a_d[i] = 1.0 / __builtin_sqrt (b_d[i]);
  25 }
  26
  27 extern float a_f[SIZE] __attribute__((__aligned__(32)));
  28 extern float b_f[SIZE] __attribute__((__aligned__(32)));
  29
  30 void
  31 vectorize_rsqrt_f (void)
  32 {
  33   int i;
  34
  35   for (i = 0; i < SIZE; i++)
  36     a_f[i] = 1.0f / __builtin_sqrtf (b_f[i]);
  37 }