https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128348
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- Confirmed (note this is at -O2). -induct2.f90:2073:78: optimized: basic block part vectorized using 16 byte vectors Here we manage to do additional operator swapping thanks to the better representation. This results in a larger SLP graph and it becoming part of a larger graph during partitioning that is determined as not profitable: +induct2.f90:2227:77: note: Cost model analysis for part in loop 6: + Vector cost: 4 + Scalar cost: 0 this will be possibly addressed by some pending patches. -induct2.f90:2084:77: optimized: basic block part vectorized using 16 byte vectors -induct2.f90:2216:78: optimized: basic block part vectorized using 16 byte vectors -induct2.f90:2227:77: optimized: basic block part vectorized using 16 byte vectors +induct2.f90:2509:42: optimized: basic block part vectorized using 16 byte vectors +induct2.f90:2560:42: optimized: basic block part vectorized using 16 byte vectors +induct2.f90:2702:42: optimized: basic block part vectorized using 16 byte vectors +induct2.f90:2845:42: optimized: basic block part vectorized using 16 byte vectors +induct2.f90:2896:42: optimized: basic block part vectorized using 16 byte vectors +induct2.f90:3038:42: optimized: basic block part vectorized using 16 byte vectors
