See this post from Ray Dillinger: https://bitcointalk.org/index.php?topic=819656.msg9170781#msg9170781

I remember this discussion, actually. 

Finney, Satoshi, and I discussed how divisible a Bitcoin ought to be.  Satoshi had already more or less decided on a 50-coin per block payout with halving every so often to add up to a 21M coin supply.  Finney made the point that people should never need any currency division smaller than a US penny, and then somebody (I forget who) consulted some oracle somewhere like maybe Wikipedia and figured out what the entire world's M1 money supply at that time was. 

We debated for a while about which measure of money Bitcoin most closely approximated; but M2, M3, and so on are all for debt-based currencies, so I agreed with Finney that M1 was probably the best measure. 

21Million, times 10^8 subdivisions, meant that even if the whole word's money supply were replaced by the 21 million bitcoins the smallest unit (we weren't calling them Satoshis yet)  would still be worth a bit less than a penny, so no matter what happened -- even if the entire economy of planet earth were measured in Bitcoin -- it would never inconvenience people by being too large a unit for convenience.

Now that I think of it, we did talk about the floating point format in that discussion.  8-decimal divisibility was the maximum Satoshi would consider, for that reason (although he was a fanatic about doing everything with unsigned integers).   Hal's point about the smallest division being less than a penny, and that being possible even if the whole world's money supply were denominated in Bitcoin, meant no extraordinary measures were necessary.