Oops, I see Richard King has that.
Richard's implementation is faster than using shifts because the self-addition doesn't require carry set-up.

This type of decomposition reminds me of a contact I did in 1986 with a 4Mhz Z-80. I had a laser barcode reader scanning labels of microfiche jackets passing by at speeds of 100 feet per second. Each barcode had 23 characters of base43 information in it. The hopper loaded about 7 units a second so I had to read each laser pass over the moving jacket and when I got enough matching reads, validate it and convert it into plaintext and pass along its route to the next subprocessor before the next one came along.

Base43 was very ugly to covert and the number of digits couldn't be done in the remaining time window.

I thought about it, figured out that by switching them from base43 to base40, I could quickly compute the conversion with only 4 or 5 partials term. After calculating them, I merely had to sum into the correct columns to get plaintext on the fly.

-Jay C. Box