[Bug tree-optimization/103999] New: Vectorizer failed to reduce sum with conversion.

[Bug tree-optimization/103999] New: Vectorizer failed to reduce sum with conversion.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103999

            Bug ID: 103999
           Summary: Vectorizer failed to reduce sum with conversion.
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: crazylht at gmail dot com
  Target Milestone: ---

Vectorizer failed for sum_reduce_cd, but ok for sum_reduce/sum_reduce_cint

float sum_reduce_cd (int n, float* array)
{
    float ans = 0.0f, x;
    for (int i = 0; i != n; i++)
      {
          x = array[i];
          ans += x + 1.0;
      }
    return ans;
}

float sum_reduce (int n, float* array)
{
    float ans = 0.0f, x;
    for (int i = 0; i != n; i++)
      {
          x = array[i];
          ans += x + 1.0f;
      }
    return ans;
}

float sum_reduce_cint (int n, float* array)
{
    int ans = 0;
    float x;
    for (int i = 0; i != n; i++)
      {
          x = array[i];
          ans += (int)(x + 1.0);
      }
    return ans;
}

https://godbolt.org/z/q3McP6d15

[Bug tree-optimization/103999] Vectorizer failed to reduce sum with conversion.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103999

--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
the difference between sum_reduce_cd and sum_reduce is one is 1.0 which is
double and another is 1.0f which doesn't need any conversion and truncation.

[Bug tree-optimization/103999] Vectorizer failed to reduce sum with conversion.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103999

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Blocks|                            |53947
           Keywords|                            |missed-optimization
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2022-01-13
             Status|UNCONFIRMED                 |NEW

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
t.c:4:23: note:   reduction path: ans_15 _7 _9 _4 ans_18
t.c:4:23: note:   reduction: unknown pattern

We do handle integer sign conversions by "ignoring" them but with float
truncations/extensions we have to be more careful.

We start with

  ((double)(float)((double)(float)(array[0] + 1.) + array[1] + 1.) + array[2] +
1.) ...

and would need to associate that in some way.  Ideally we'd arrange for
doing the reduction as 'double' and only truncate the final value getting
us "only" extra precision.  But that would require major surgery in
reduction handling.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations