Re: [PATCH GCC]Resolve compilation time known alias checks in vectorizer

["Bin.Cheng" <amker dot cheng at gmail dot com>]

On Tue, Aug 9, 2016 at 4:43 PM, Bin.Cheng <amker.cheng@gmail.com> wrote:
> On Sat, Aug 6, 2016 at 9:20 PM, Andreas Schwab <schwab@linux-m68k.org> wrote:
>> On Mi, Jul 13 2016, "Bin.Cheng" <amker.cheng@gmail.com> wrote:
>>
>>> Patch re-tested/applied on trunk as r238301.
>>
>> This breaks gcc.dg/vect/vect-117.c on powerpc.
> Hi Andreas,
> Sorry for the inconvenience, I will have a look.
Looks like the patch exposed another latent issue in vectorizer.
Before patching, the loop is vectorized under below alias check
condition:

  <bb 3>:
  _7 = (int[5] *) ivtmp.88_168;
  _20 = (unsigned int) n_15(D);
  _12 = _20 > 3;
  _180 = ivtmp.85_166 + 32;
  _181 = (ssizetype) _180;
  _182 = ivtmp.85_166 + 20;
  _183 = (ssizetype) _182;
  _32 = _181 <= _183;
  _184 = ivtmp.85_166 + 36;
  _185 = (ssizetype) _184;
  _38 = (ssizetype) ivtmp.85_166;
  _39 = _38 >= _185;
  _40 = _32 | _39;
  _41 = _12 & _40
  if (_41 != 0)
    goto <bb 4>;
  else
    goto <bb 13>;

Note the condition _40 = (_32 | _39) can never be true thus vectorized
loop will never be executed in practice.  This patch only bypasses the
known to be false condition at compile time.
After investigation, I believe root causes is in unalignment handling,
especially vectorizer tries to optimize unaligned load by scheme
dr_explicit_realign_optimized on powerpc.
The loop itself should be vectorized successfully:

static  int a[N][N] = {{ 1, 2, 3, 4, 5},
               { 6, 7, 8, 9,10},
               {11,12,13,14,15},
               {16,17,18,19,20},
               {21,22,23,24,25}};

volatile int foo;

__attribute__ ((noinline))
int main1 (int A[N][N], int n)
{

  int i,j;

  /* vectorizable */
  for (i = 1; i < N; i++)
  {
    for (j = 0; j < n; j++)
    {
      A[i][j] = A[i-1][j] + A[i][j];
    }
  }

  return 0;
}

But vectorizer tries to realign A[i-1][j] using realign_load, thus the
real memory address accessed is different to the original address in
each iteration.  Function vect_vfa_segment_size takes this into
consideration and generates false alias check condition.  IMO, The
unalignment support scheme should be improved in a way that allows
more vectorization.  In other words, it should fall back to
dr_unaligned_supported to enable vectorization, rather than choose
dr_explicit_realign_optimized which blocks vectorization.

I will file a PR for this.

Thanks,
bin
>
> Thanks,
> bin
>>
>> Andreas.
>>
>> --
>> Andreas Schwab, schwab@linux-m68k.org
>> GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
>> "And now for something completely different."