[Bug tree-optimization/94792] New: Missed SLP optimization in pr65930-2.c variation

[Bug tree-optimization/94792] New: Missed SLP optimization in pr65930-2.c variation

[iii at linux dot ibm.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94792

            Bug ID: 94792
           Summary: Missed SLP optimization in pr65930-2.c variation
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: iii at linux dot ibm.com
  Target Milestone: ---

gcc commit cf3a909cf455. Consider the following variation of pr65930-2.c:

$ cat pr65930-2b.c
#include "tree-vect.h"

int __attribute__((noipa))
bar (unsigned int *x, int n)
{
  unsigned int sum = 4;
  x = __builtin_assume_aligned (x, __BIGGEST_ALIGNMENT__);
  for (int i = 0; i < n; ++i)
    sum += x[i*4+0]+ x[i*4 + 1] + x[i*4 + 2] + x[i*4 + 3];
  return sum;
}

int
main ()
{
  static int a[16] __attribute__((aligned(__BIGGEST_ALIGNMENT__)))
    = { 1, 3, 5, 8, 9, 10, 17, 18, 23, 29, 30, 55, 42, 2, 3, 1 };
  check_vect ();
  if (bar (a, 4) != 260)
    abort ();
  return 0;
}

This differs from pr65930-2.c only in that sum type is unsigned int, which
should be on cast less. And yet:

$ gcc pr65930-2b.c -fno-diagnostics-show-caret
-fno-diagnostics-show-line-numbers -fdiagnostics-color=never
-fdiagnostics-urls=never -msse2 -ftree-vectorize
-fno-tree-loop-distribute-patterns -fno-vect-cost-model -fno-common -O2
-fdump-tree-vect-details -lm -o ./pr65930-2.exe ; grep SLP
pr65930-2b.c.161t.vect | wc -l
0

whereas for the original version:

$ gcc pr65930-2.c -fno-diagnostics-show-caret
-fno-diagnostics-show-line-numbers -fdiagnostics-color=never
-fdiagnostics-urls=never -msse2 -ftree-vectorize
-fno-tree-loop-distribute-patterns -fno-vect-cost-model -fno-common -O2
-fdump-tree-vect-details -lm -o ./pr65930-2.exe ; grep SLP
pr65930-2.c.161t.vect | wc -l
33

The resulting assembly is also noticeably larger and uses regular adds for at
least part of the data.

[Bug tree-optimization/94792] Missed SLP optimization in pr65930-2.c variation

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94792

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2020-04-27
             Blocks|                            |53947
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Yeah, that's reassoc associating the sum in a way for non-SLP reduction
detection.  We're currently not trying to associate things differently in the
vectorizer.


Referenced Bugs:

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