gcc/testsuite/gcc.dg/vect/vect-over-widen-17.c

   1 /* { dg-require-effective-target vect_int } */
   2 /* { dg-require-effective-target vect_shift } */
   3 /* { dg-require-effective-target vect_pack_trunc } */
   4 /* { dg-require-effective-target vect_unpack } */
   5
   6 #include "tree-vect.h"
   7
   8 #define N 1024
   9
  10 /* This should not be treated as an over-widening pattern, even though
  11    "(b[i] & 0xef) | 0x80)" could be done in unsigned chars.  */
  12
  13 void __attribute__ ((noipa))
  14 f (unsigned short *restrict a, unsigned short *restrict b)
  15 {
  16   for (__INTPTR_TYPE__ i = 0; i < N; ++i)
  17     {
  18       unsigned short foo = ((b[i] & 0xef) | 0x80) + (a[i] << 4);
  19       a[i] = foo;
  20     }
  21 }
  22
  23 int
  24 main (void)
  25 {
  26   check_vect ();
  27
  28   unsigned short a[N], b[N];
  29   for (int i = 0; i < N; ++i)
  30     {
  31       a[i] = i;
  32       b[i] = i * 3;
  33       asm volatile ("" ::: "memory");
  34     }
  35   f (a, b);
  36 #pragma GCC novector
  37   for (int i = 0; i < N; ++i)
  38     if (a[i] != ((((i * 3) & 0xef) | 0x80) + (i << 4)))
  39       __builtin_abort ();
  40
  41   return 0;
  42 }
  43
  44 /* { dg-final { scan-tree-dump-not {vect_recog_over_widening_pattern: detected} "vect" } } */
  45 /* On Power, if there is no vect_hw_misalign support, unaligned vector access
  46    adopts realign_load scheme.  It requires rs6000_builtin_mask_for_load to
  47    generate mask whose return type is vector char.  */
  48 /* { dg-final { scan-tree-dump-not {vector[^\n]*char} "vect" { target vect_hw_misalign } } } */
  49 /* { dg-final { scan-tree-dump-not {vector[^ ]* int} "vect" } } */
  50 /* { dg-final { scan-tree-dump-times "vectorized 1 loop" 1 "vect" } } */