read!() and readbytes!() mutate their second argument

[nalimilan]

I thought the convention was that in-place versions of functions mutate their first argument. It looks like we could easily change the argument order and print deprecations if we want to enforce this convention.

Though the rand*! family of functions also doesn't follow this rule, so maybe we should just give up on consistency.

[ivarne]

It's true that in-place functions mutate their first argument, but IO methods takes an optional ::IO object as the first argument. The question is which convention that should win?

(The ::AbstractRNG interfaces tries to use the same convention as ::IO)

[ScottPJones]

I think @ivarne is correct, which convention takes precedence.
In this case, I personally would prefer to keep the IO as the first argument, even if that causes a little inconsistency with the mutating versions usually modifying the first argument.
It also makes sense if you wanted to have read!(arr) read from STDIN, for example.
(for some reason that short form doesn't seem to exist, though)

[stevengj]

I agree that io should continue to be the first argument.

[vtjnash]

the mutation of the first argument is common (since it often appears when used for container-semantics and oo-semantics), but I don't think it was ever intended that should be used as a rule. for example, i believe most of the linalg ! functions mutate/assign/overwrite/destroy one or more arguments, but not necessarily the first argument. (unhelpfully, all of the IO functions actually mutate their first argument also, since IO is a form of mutation).

is there a place this needs to be documented better (e.g. when ! is first described in the manual)?

[nalimilan]

To be honest, I haven't read this anywhere (or I don't recall it). No worries.