[Bug tree-optimization/126376] aarch64: cunrolli unrolls small constant-trip inner loops and blocks dot-product vectorization

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126376

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu.org
   Last reconfirmed|                            |2026-07-24
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
There is a related bug that the partial-lane-reducing ops like SAD and DOT_PROD
are not specified enough on the optab level to be useable for the case the
vector result is not only used as input of a reduction to a scalar (aka
which input lanes are combined to which output lanes).  Otherwise this would
be a pattern recognizable in SLP pattern recognition, but there might be
initial issues with SLP discovery?

So IMO a good thing would be to nail down SAD/DOT_PROD and friends to make
them useable here.  Or alternatively expose the fact that we are reducing
to a scalar (but then it would need to be BB SLP reduction discovery, with
no intermediate live lanes - a constraint we currently have no code to
ensure validity).