#3283 made an optimization to the parser to avoid checking for _s in numeric literals twice, and got a nice 1% speed-up on parser benchmarks.

It occurs to me that we could repeat the trick for the common case of simple decimal literals (e.g. 0, 1, 123). str::parse::<f64>() is quite a complicated function which checks for all kinds of things (exponent, decimal point, valid input etc). All of this is repeated work, since we've already determined in the lexer what the input is.

Perhaps we could have a flag which lexer sets if the input is a simple case, and then use a hand-rolled parse_simple_decimal function to convert to f64, like the ones we have for binary and octal?

(we should base implementation on std's str::parse::<f64> as it's probably very optimized already, but just cut out all the extraneous checks)

Given that #3283 got a 1% speed-up, maybe this can get the same again, or maybe more?

@DonIsaac Would you be interested in investigating?