[Bug tree-optimization/19701] [4.0 regression] Way too many IVs

pinskia at gcc dot gnu dot org gcc-bugzilla@gcc.gnu.org
Sun Jan 30 01:56:00 GMT 2005


------- Additional Comments From pinskia at gcc dot gnu dot org  2005-01-30 01:55 -------
(In reply to comment #4)
> Also happens on x86-64, but not on x86. 

Well the IV selection on x86 is not good either.
  *(long long unsigned int *) d.18 = 0;
  D.1187 = (long long unsigned int *) ivtmp.4 + (long long unsigned int *) d;
  *(D.1187 + 8B) = 0;
...
  ivtmp.4 = ivtmp.4 + 64B;
  if ((int) ((unsigned int) w - (unsigned int) ivtmp.4) > 63) goto <L10>; else goto <L2>;

Notice how we use ivtmp.4 for "d+8".  And then we do w - ivtmp.4 when ivtmp.4 should go from w and 
go down like what it is in the source.  So we select the same IVs as in the source.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19701



More information about the Gcc-bugs mailing list