http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49749
--- Comment #11 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2011-07-20 19:01:30 UTC --- I forgot to mention some justification for the value of PHI_LOOP_BIAS, and I notice it has a misleading comment by it at the moment. The value is a constant that should be larger than the depth of the largest expected reassociation tree, but generally smaller than twice the rank of the basic block. The intent is to limit the effect to single-block loops for now. One of the questions I have is whether the bias should be allowed for more complex innermost loops. PHI_LOOP_BIAS is a crude method and something more refined is probably needed, but it serves to demonstrate the viability of the general approach.
