------- Comment #3 from rakdver at gcc dot gnu dot org 2009-08-07 08:44 ------- (In reply to comment #1) > The tree optimizers canonicalize the loop to > > <bb 3>: > # i_5 = PHI <i_3(4), 0(2)> > # ivtmp.23_1 = PHI <ivtmp.23_4(4), 10(2)> > f2 (); > i_3 = i_5 + 1; > ivtmp.23_4 = ivtmp.23_1 - 1; > if (ivtmp.23_4 != 0) > goto <bb 4>; > else > goto <bb 5>; > > <bb 4>: > goto <bb 3>; > > But then IVOPTs chooses i as the induction variable again.
This is what I would expect it to do (I am somewhat surprised that 3.3 did something else). Ivopts at the moment do not know that comparing with 0 is more efficient than comparing with any other expression. In all other aspects, i and ivtmp.23 have the same cost, so ivopts prefers to preserve the original induction variable. Altering determine_use_iv_cost_condition to take the cost of the comparison into account should fix this. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40886
