[Bug tree-optimization/87031] nios2 optimization for size - two cases of regression relatively to 5.3.0

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87031

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2018-08-22
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
          Component|target                      |tree-optimization
     Ever confirmed|0                           |1

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
size: 24-3, last_iteration: 19-3
  Loop size: 24
  Estimated size after unrolling: 24

so GCC estimates that unrolling doesn't change size.  It arrives here
by applying it's "unrolling saves 1/3 of insns" heuristic.

First of all this isn't controlled by a --param but hard-coded.  Second,
it wasn't reduced after we improved estimates of what is optimized away.
Third, we probably shouldn't apply this when optimizing the loop for size
(aka UL_NO_GROWTH).

Honza?  Any opinion on the above?  Did you ever try removing that * 2/3
completely?