[Bug tree-optimization/114166] New: word_mode vectorization still relies on vector lowering

[Bug tree-optimization/114166] New: word_mode vectorization still relies on vector lowering

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114166

            Bug ID: 114166
           Summary: word_mode vectorization still relies on vector
                    lowering
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

When the vectorizer decides to use emulated vectors for operations the target
doesn't support and the vector types do not have integer mode but the
vectorizer simply emits vector code nevertheless which it does for
bitwise operations, we fail RTL expansion.

This can be seen with g++.dg/opt/pr82929.C when moving vector lowering
before vectorization.  We then vectorize V1SImode bitwise and even
though the optab check discovered that's not supported.

Vectorizer code generation should, like it now does for PLUS/MINUS and NEGATE,
use integer types when code generating emulated vectors.

[Bug tree-optimization/114166] word_mode vectorization still relies on vector lowering

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114166

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
g++.dg/opt/pr82929.C with -m32, that is.

[Bug tree-optimization/114166] word_mode vectorization still relies on vector lowering

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114166

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ice-on-valid-code,
                   |                            |missed-optimization

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
g++.dg/opt/vectcond-1.C also with -m32 is another testcase, this time for
BIT_XOR_EXPR.

SLP vectorization is also a bit too happy converting scalar code to
vector<1> code, we should filter those out.

[Bug tree-optimization/114166] word_mode vectorization still relies on vector lowering

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114166

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2024-02-29
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed. I accidently came across this today via PR 114178 .