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

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