There's a documented compiler switch -unroll0 which should turn off unrolling (beyond what is inherent in vectorization). For an individual loop, $pragma unroll(0) or other pragmas may be useful.
In general, -O2 is already fairly aggressive in most situations, and we've had to drop back to it frequently in order to avoid additional optimizations presented by -O3.
If you mean loop distribution and fusion (altering the boundaries of for loops), besides considering -O2, you could consider placing
#pragma distribute point
immediately after the for(), so as to prevent the compiler from splitting the loop, or between loops, to prevent fusion.
