[Bug tree-optimization/68050] SLP vectorization should negate constants to match up + vs -

[Bug tree-optimization/68050] SLP vectorization should negate constants to match up + vs -

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68050

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2015-10-22
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Mine.

[Bug tree-optimization/68050] New: SLP vectorization should negate constants to match up + vs -

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68050

            Bug ID: 68050
           Summary: SLP vectorization should negate constants to match up
                    + vs -
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
            Blocks: 53947
  Target Milestone: ---

double x[1024], y[1024], z[1024];
void foo (double w)
{
  int i;
  for (i = 0; i < 1023; i+=2)
    {
      z[i] = x[i] + 3.;
      z[i+1] = x[i+1] + -3.;
    }
}


Referenced Bugs:

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

[Bug tree-optimization/68050] SLP vectorization should negate constants to match up + vs -

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68050

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|2015-10-22 00:00:00         |2021-8-28

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
WE do this now:

  <bb 3> [local count: 1063004409]:
  # ivtmp.13_20 = PHI <ivtmp.13_15(3), 0(2)>
  vect__1.5_14 = MEM <vector(2) double> [(double *)&x + ivtmp.13_20 * 1];
  vect__5.7_12 = vect__1.5_14 + { -3.0e+0, -3.0e+0 };
  vect__2.6_13 = vect__1.5_14 + { 3.0e+0, 3.0e+0 };
  _7 = VEC_PERM_EXPR <vect__2.6_13, vect__5.7_12, { 0, 3 }>;
  MEM <vector(2) double> [(double *)&z + ivtmp.13_20 * 1] = _7;
  ivtmp.13_15 = ivtmp.13_20 + 16;
  if (ivtmp.13_15 != 8192)
    goto <bb 3>; [99.00%]
  else
    goto <bb 4>; [1.00%]


But if we use long instead of double, we get:
  <bb 3> [local count: 1063004409]:
  # ivtmp.12_20 = PHI <ivtmp.12_15(3), 0(2)>
  vect__1.5_14 = MEM <vector(2) long int> [(long int *)&x + ivtmp.12_20 * 1];
  vect__2.6_13 = vect__1.5_14 + { 3, -3 };
  MEM <vector(2) long int> [(long int *)&z + ivtmp.12_20 * 1] = vect__2.6_13;
  ivtmp.12_15 = ivtmp.12_20 + 16;
  if (ivtmp.12_15 != 8192)
    goto <bb 3>; [99.00%]
  else
    goto <bb 4>; [1.00%]


Note clang has worse code for the VECT_PERM_EXPR but does the same thing as GCC
here.

So the difference between long and double for GCC at least comes from having IR
difference in that the "a - 3" in the long case is actually "a + -3".

[Bug tree-optimization/68050] SLP vectorization should negate constants to match up + vs -

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68050

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
We don't really do it as you show - instead of

  vect__5.7_12 = vect__1.5_14 + { -3.0e+0, -3.0e+0 };
  vect__2.6_13 = vect__1.5_14 + { 3.0e+0, 3.0e+0 };
  _7 = VEC_PERM_EXPR <vect__2.6_13, vect__5.7_12, { 0, 3 }>;

we'd like to see

  _7 = vect__1.5_14 + { 3.0e+0, -3.0e+0 };

which is what we should cost as well.