A less safe, but very small base62 decoder:

function(a,b,c,d){for(b=c=0;d=a.charCodeAt(c++);b=b*62+d-[,48,29,87][d>>5]);return b}

@atk : thanks, the b=b*62+d-[,48,29,87][d>>5] part is awesome, exactly what I needed !

You're welcome, @sebastien-p :-)

I would really like to see more comments in the annotated part. I have only just worked out what the first part of the for loop does (leading to the false-y charCodeAt(-1) in the second for-loop-part if a is not string or not base62).

@atk : is it better now ? :) Can you explain the line I kept from your code as I don't always understand well those bitwise tricks ? Thanks.

Thanks, much better now. Of course I can explain this line (except of the b*62 part, which seems pretty self-explaining to me):

// We are expecting either numbers (ascii 48-58), uppercase letter (ascii 65-90) and lowercase letters (ascii 97-122)
// and want to map the numbers to 0-9, the uppercase letters to 36-61 and the lowercase letters to 10-35; therefore,
// we integer divide the character code by 32 using shift right 5, so numbers will yield 1, uppercase letters 2 and lowercase 3
// and let this result select the corresponding substractor from a sparse array (because the extra comma is smaller than -1):
[,48,29,87][d>>5]

@atk : thank you, I understand it a little more now :) but still don't manage to apply this kind of hack to the encoding function :(

it does not apply there, because we'd need another formula than d>>5 to filter 0-9, 10-35 and 36-61 to 0, 1 and 2. If I find one, I'll tell you.

@atk : I would be very interested to know, thank you.

no formula I could think of would be shorter than x>35?87:x>10?29:48.

@atk : ok, thanks. If you want to help, I released the new version of the encoding function, still 4 bytes to go :) https://gist.github.com/1168420

a === (/\W|_/.test(a += "") || a) could be great instead of a === (a += "") && /^[a-z\d]+$/i.test(a) but sadly doesn't prevent passing an empty string as argument.

how about saving at least 1 byte with a===(/^[a-z\d]+$/i.exec(a)||0)[0] which should take care of empty strings? Or even shorter: a===(/\W|_|^$/.test(a)||a) in which we use the property of /^$/ to match empty strings?

@atk : thanks, base62 decode is now 4 bytes shorter :)

@sebastien-p: test coerces its argument to string, so a+='' has no function there. You may want to use a===(/\W|_|^$/.test(a)||a+'')

@atk : yep, I know, but a needs to be a string for a.charCodeAt(c++) to work, so I force coercion here on purpose.

For a not to be a string and to be coerced to a string that coerced into a valid base62 string, it would have to be an object with a toString method, so handling non-strings is rather pointless, isn't it?

@atk : you're right, input value validation is optional and the whole entry would be shorter without it. I added it because it fits and mostly because I needed the function to be designed this way.

true // -1
false // -1

I found this useful and converted it into C# if anyone ever needs it. Not intended to be 140 bytes small.