[silvestrov]

Wild guess: this is a left-over from the old days of Sun SPARC processors which could only access doubles on 8-byte aligned memory addresses.

For Intel processors you just paid a performance penality for unaligned accesses but SPARC processors would generate a TRAP for unaligned accesses.

I'd really like to see some performance testing of alignment for modern Intel/ARM CPUs. My guess is that tightly packing would be better as more data would fit into caches and you would have less i/o.

Modern CPUs are very different from back in the 90's when PostgreSQL was designed. Posgresql has traditionally been very slow to stop supporting old OSes that are no longer used.