[Bug ipa/91374] [Missed optimization] Versioning opportunities to improve performance

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91374

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
            Version|tree-ssa                    |10.0
           Keywords|                            |missed-optimization
   Last reconfirmed|                            |2019-08-07
          Component|tree-optimization           |ipa
                 CC|                            |jamborm at gcc dot gnu.org,
                   |                            |marxin at gcc dot gnu.org
     Ever confirmed|0                           |1
           Severity|normal                      |enhancement

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
So you ask for main to be converted to

 if (idx == 0)
   foo_32_16 ();
 else /* idx == 1 */
   foo_16_8 ();

correct?  It shoulds like an interesting idea for an IPA-CP cloning
opportunity.  But I wonder how much real-world testcases exists
that are not benchmarks ;)  One could tell by implementing the
analysis into the param analysis stage (look for defs that are
loads from a constant initialized entity, possibly multiple of them
and (some of them) reusing the same memory index).  A related thing
would be to value-profile the indexes since when the array doesn't
only contain two values it isn't feasible to clone for all possible
values.