[Bug middle-end/83023] branch probabilities pessimize malloc

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83023

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2017-11-19
                 CC|                            |mliska at suse dot cz
     Ever confirmed|0                           |1

--- Comment #2 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
We get:

Predictions for bb 2                                                            
  DS theory heuristics: 53.5%                                                   
  combined heuristics: 53.5%                                                    
  pointer (on trees) heuristics of edge 2->3: 70.0%                             
  call heuristics of edge 2->3: 33.0%                                           

We do not know that malloc is likely returning non-NULL, but we predict with
70% that pointer is non-NULL but in the testcase this prediction is overwritten
with call heuristics.

builtin_expect code can be easily extended to handle builtins that likely
return given value. I am adding Martin to CC as this may be part of predictor
retuning this time.