[Bug tree-optimization/103144] vectorizer failed to recognize shift>>=1 in loop as shift>>i

[crazylht at gmail dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103144

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
Another issue is for SLP, when trip count is small and loop is completely
unrolled. SLP failed to generate vlshr_optab.

#include<stdint.h>
void
foo (uint64_t* __restrict pdst, uint64_t* psrc, uint64_t shift)
{
    for (int64_t i = 0; i != 4; i++)
    {
      pdst[i] = psrc[i] + shift;
      shift >>= 1;
    }
}

.175t.slp1 dump:

  _8 = *psrc_12(D);
  _28 = _8 + shift_10(D);
  shift_30 = shift_10(D) >> 1;
  _39 = psrc_12(D) + 8;
  _40 = *_39;
  _41 = pdst_13(D) + 8;
  _42 = shift_30 + _40;
  shift_44 = shift_30 >> 1;
  _53 = psrc_12(D) + 16;
  _54 = *_53;
  _55 = pdst_13(D) + 16;
  _56 = shift_44 + _54;
  shift_58 = shift_44 >> 1;
  _3 = psrc_12(D) + 24;
  _4 = *_3;
  _5 = pdst_13(D) + 24;
  _6 = _4 + shift_58;
  _15 = {_28, _42, _56, _6};
  vectp.8_18 = pdst_13(D);
  MEM <vector(4) long unsigned int> [(uint64_t *)vectp.8_18] = _15;