[Bug target/105197] [12 Regression] SVE: wrong code with -O -ftree-vectorize

[tnfchris at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105197

Tamar Christina <tnfchris at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |tnfchris at gcc dot gnu.org
                 CC|                            |tnfchris at gcc dot gnu.org
   Last reconfirmed|                            |2022-04-10
     Ever confirmed|0                           |1

--- Comment #2 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #1)
> the GIMPLE doesn't look wrong.  We're using an EXTRACT_LAST, so that might
> be the special thing.  Vectorization of the first loop is probably not
> necessary to trigger the failure.

Hmmm looks like the GIMPLE is wrong, the masks it combines creates a
contradiction

At GIMPLE we have

  mask__44.14_114 = vect__4.13_112 != { 0, ... };
  mask__26.22_128 = vect__6.17_121 == { 0, ... };
  mask_patt_65.24_130 = mask__44.14_114 & mask__26.22_128;
  mask__43.26_135 = vect__4.13_112 == { 0, ... };
  mask__25.18_123 = vect__6.17_121 != { 0, ... };
  _137 = mask__43.26_135 & loop_mask_111;
  _163 = mask_patt_65.24_130 & _137;

where _163 demands vect__4.13_112 != 0 && vect__4.13_112 == 0

  _163 should have been _163 = mask_patt_65.24_130 & loop_mask_111;

So it looks like the wrong loop masks are combined.

Mine.