https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70102
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> --- This happens with reductions in BBs as well - the "reduction" part gets pulled into the chain: c.21_93 = c; _112 = _111 * _110; _95 = _92 + _112; c.22_114 = _95 + c.21_93; <--- _132 = _131 * _130; c.22_134 = c.22_114 + _132; _152 = _151 * _150; c.22_154 = c.22_134 + _152; _55 = _54 * _53; c.22_57 = c.22_154 + _55; c = c.22_57; while before reassoc we have _92 = _90 * _91; c.21_93 = c; c.22_94 = _92 + c.21_93; _112 = _110 * _111; c.22_114 = c.22_94 + _112; ... the rank of _93 is 1 while that of _112 is 655366. This is because the load of c doesn't have any SSA operands and thus gets rank 0 + 1. Other loads get sth biased by BB rank because ultimatively the array index used depends on a loop phi. I don't see how we can (or should) reasonably fix this in reassociation. Instead what we want to do is re-org SLP detection to cover these cases which requires some changes to its IL.
