[Bug tree-optimization/83253] -ftree-slsr causes performance regression

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2017-12-04
                 CC|                            |amker at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org,
                   |                            |wschmidt at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
It looks more like a backend issue to me.  I can reproduce it on x86_64 but
only with -march=nocona, not with generic or say, core-avx2.

Note that I think it is odd that SLSR ever does this replacement since
(on GIMPLE...) no instruction is saved.

I guess the expmed costing stuff doesn't consider the computations in
address context (though it must consider using lea on x86 for x + C*y?).

Eventually the IVO costing would be more accurate for address uses, but then
SLSR might want to create TARGET_MEM_REFs ...