> Is a single MOV instruction still fast when the 8 bytes begin on an odd address?

On x86, yes. There is no performance penalty for misaligned loads, except when the misaligned load also happens to straddle a cache line boundary, in which case it is slower, but only marginally so.