I changed the return stmt to "return sum2" so the for-loop calculations aren't optimized away and fed the code to godbolt.org (compiler explorer).

gcc-trunk at -O3 for x64 will vectorize the loops but there's no register pressure, so the register allocator wasn't taxed much.

No niche optimization pass to convert to using Gauss's shortcut - https://physicsdb.com/sum-natural-numbers/