In Mauchly's "Preparation of Problems for EDVAC-Type Machines", quoted in part in the blog post, he writes:
> The total number of operations for which instructions must be provided will usually be exceedingly large, so that the instruction sequence would be far in excess of the internal memory capacity. However, such an instruction sequence is never a random sequence, and can usually be synthesized from subsequences which frequently recur.
> By providing the necessary subsequences, which may be utilized as often as desired, together with a master sequence directing the use of these subsequences, compact and easily set up instructions for very complex problems can be achieved.
The verbs he uses here for subroutine calls are "utilize" and "direct". Later in the paper he uses the term "subroutine" rather than "subsequence", and does say "called for" but not in reference to the subroutine invocation operation in the machine:
> For these, magnetic tapes containing the series of orders required for the operation can be prepared once and be made available for use when called for in a particular problem. In order that such subroutines, as they can well be called, be truly general, the machine must be endowed with the ability to modify instructions, such as placing specific quantities into general subroutines. Thus is created a new set of operations which might be said to form a calculus of instructions.
Of course nowadays we do not pass arguments to subroutines by modifying their code, but index registers had not yet been invented, so every memory address referenced had to be contained in the instructions that referenced it. (This was considered one of the great benefits of keeping the program in the data memory!)
A little lower down he says "initiate subroutines" and "transferring control to a subroutine", and talks about linking in subroutines from a "library", as quoted in the post.
He never calls subroutines "functions"; I'm not sure where that usage comes from, but certainly by BASIC and LISP there were "functions" that were at least implemented by subroutines. He does talk about mathematical functions being computed by subroutines, including things like matrix multiplication:
> If the subroutine is merely to calculate a function for a single argument, (...)