[Bug tree-optimization/88464] AVX-512 vectorization of masked scatter failing with "not suitable for scatter store"

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88464

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2018-12-12
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
          Component|c                           |tree-optimization
             Blocks|                            |53947
     Ever confirmed|0                           |1
      Known to fail|                            |9.0

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  The issue here is the target doesn't advertise scatter:

3901              if (targetm.vectorize.builtin_scatter)
(gdb) n
3902                decl = targetm.vectorize.builtin_scatter (vectype, offtype,
scale);
(gdb) 
3905          if (!decl)
(gdb) p decl
$5 = (tree) 0x0
(gdb) p decl
$5 = (tree) 0x0
(gdb) p debug_generic_expr (vectype)
vector(4) double
$6 = void
(gdb) p debug_generic_expr (offtype)
int
$7 = void
(gdb) p scale
$8 = 8

probably because of the "mixup" of 4-byte index and 8 byte store element type.

I'm not sure whether Jakub implemented widening of the index at all.

OTOH when using long for idx it doesn't help.  Using -march=knl we get
a little bit further but ultimately fail the same way.


Referenced Bugs:

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