https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77445

--- Comment #12 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Created attachment 40526
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40526&action=edit
Patch I am testing

The profile is quite inconsistent since thread1.
The problem is that duplicate_thread_path does
  if (total_count)                                                              
    {                                                                           
      scale_bbs_frequencies_gcov_type (region, n_region,                        
                                       total_count - entry_count,               
                                       total_count);                            
      scale_bbs_frequencies_gcov_type (region_copy, n_region, entry_count,      
                                       total_count);                            
    }                                                                           
  else                                                                          
    {                                                                           
      scale_bbs_frequencies_int (region, n_region, total_freq - entry_freq,     
                                 total_freq);                                   
      scale_bbs_frequencies_int (region_copy, n_region, entry_freq,
total_freq);
    }                                                                           

This is wrong when the duplicated path have extra entries that are not
duplicated which happens in the testcase.

The attached patch fixes it.  It is still not 100% correct as the outgoing
probabilities of the original path should be updated based on the path being
copied out, but I am not sure it is worth the trouble (and risk of roundoff
errors). I will think a bit more about this tomorrow.

Reply via email to