[Bug tree-optimization/103797] Clang vectorized LightPixel while GCC does not

["hubicka at kam dot mff.cuni.cz" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103797

--- Comment #4 from hubicka at kam dot mff.cuni.cz ---
> -E and remove not needed code.
> 
> > The
> > declaratoins are quite convoluted, but the function is well isolated and
> > easy to inspect from full one...
> 
> Do we speak about:
> https://github.com/mozilla/gecko-dev/blob/bd25b1ca76dd5d323ffc69557f6cf759ba76ba23/gfx/2d/FilterNodeSoftware.cpp#L3670-L3691
> ?
Yes.
> 
> It should be possible creating a synthetical test that does the same (and lives
> in a loop, right?).

Well, I tried that for a while and got bit lost (either code got
vectorized by both gcc and clang or by neither).  There are more issues
where we have over 50% regression wrt clang build at gfx code, so I
think I will first try to reproduce those locally and perf them to see
if there is more pattern here.

The releavant code is:

uint32_t mozilla::gfx::{anonymous}::SpecularLightingSoftware::LightPixel
(struct SpecularLightingSoftware * const this, const struct Point3D & aNormal,
const struct Point3D & aVectorToLight, uint32_t aColor)
{

  <bb 2> [local count: 118111600]:
  _48 = MEM[(const struct BasePoint3D
*)aVectorToLight_25(D)].D.75826.D.75829.z;
  _49 = _48 + 1.0e+0;
  _50 = MEM[(const struct BasePoint3D
*)aVectorToLight_25(D)].D.75826.D.75829.y;
  _51 = _50 + 0.0;
  _52 = MEM[(const struct BasePoint3D
*)aVectorToLight_25(D)].D.75826.D.75829.x;
  _53 = _52 + 0.0;
  _80 = _53 * _53;
  _82 = _51 * _51;
  _83 = _80 + _82;
  _85 = _49 * _49;
  _86 = _83 + _85;
  if (_86 u>= 0.0)
    goto <bb 3>; [99.95%]
  else
    goto <bb 4>; [0.05%]

  <bb 3> [local count: 118052545]:
  _87 = .SQRT (_86);
  goto <bb 5>; [100.00%]

  <bb 4> [local count: 59055]:
  _29 = __builtin_sqrtf (_86);

  <bb 5> [local count: 118111600]:
  # _30 = PHI <_29(4), _87(3)>
  _88 = _53 / _30;
  _89 = _51 / _30;
  _90 = _49 / _30;
  _41 = MEM[(const struct BasePoint3D *)aNormal_26(D)].D.75826.D.75829.x;
  _39 = _41 * _88;
  _37 = MEM[(const struct BasePoint3D *)aNormal_26(D)].D.75826.D.75829.y;
  _33 = _37 * _89;
  _27 = _33 + _39;
  _45 = MEM[(const struct BasePoint3D *)aNormal_26(D)].D.75826.D.75829.z;
  _46 = _45 * _90;
  _47 = _27 + _46;
  if (_47 >= 0.0)
    goto <bb 12>; [59.00%]
  else
    goto <bb 6>; [41.00%]


With -Ofast it gets bit more streamlined:


  <bb 2> [local count: 118111600]:
  _48 = MEM[(const struct BasePoint3D
*)aVectorToLight_25(D)].D.75826.D.75829.z;
  _49 = _48 + 1.0e+0;
  _50 = MEM[(const struct BasePoint3D
*)aVectorToLight_25(D)].D.75826.D.75829.y;
  _51 = MEM[(const struct BasePoint3D
*)aVectorToLight_25(D)].D.75826.D.75829.x;
  powmult_78 = _51 * _51;
  powmult_80 = _50 * _50;
  _81 = powmult_78 + powmult_80;
  powmult_83 = _49 * _49;
  _84 = _81 + powmult_83;
  _85 = __builtin_sqrtf (_84);
  _86 = _51 / _85;
  _87 = _50 / _85;
  _88 = _49 / _85;
  _41 = MEM[(const struct BasePoint3D *)aNormal_26(D)].D.75826.D.75829.x;
  _39 = _41 * _86;
  _37 = MEM[(const struct BasePoint3D *)aNormal_26(D)].D.75826.D.75829.y;
  _33 = _37 * _87;
  _27 = _33 + _39;
  _45 = MEM[(const struct BasePoint3D *)aNormal_26(D)].D.75826.D.75829.z;
  _46 = _45 * _88;
  _47 = _27 + _46;
  if (_47 >= 0.0)
    goto <bb 3>; [59.00%]
  else
    goto <bb 9>; [41.00%]

But I do not quite see in the slp dump why this is not considered for
vectorization.

I attach the dump.
Honza