I only ever found one of their journals of any value whatsoever (Computing Surveys). Their "collected algorithms" was lousy. If I were interested in representation of polynomial equations in Fortran it would sometimes be useful...but I haven't done that since college...decades ago.
Occasionally I'll follow a link that ends up in the ACM members only section. Sometimes it looks interesting, but back in the time I could follow it into the article only once was it really at all interesting, and that time it still wasn't useful.
If you've got a set of Knuth's books, then I don't think the ACM has anything to offer.
WRT ACM articles linked from Google: They are there, if only as indirect links [not sure], because every once in a while I end up on one of their "you can only read the abstract" pages. I never regret not being able to read further, because I *was* a member and *could* read the linked article for awhile. Every single one was worthless (for my purposes).
The only useful thing I've ever gotten out of the ACM site that I didn't find in Knuth was a date algorithm. And I already had most of it down. And their version still didn't deal with pre-Gregorian dates (except as if they had been Gregorian dates). (To be fair, Julian dates are rather different. Still...) Also it didn't properly handle dates BC, even in Gregorian terms.
Well, they guy that wrote the algorithm was still really tight on conserving RAM usage. It *was* a very concise algorithm. And it worked without problems (in Gregorian) back to 1AD. It was also (IIRC) nigh unintelligible because of embedded magic numbers. When unpacked it basically just said skip leap years for centuries unless the century divided by 400 is an integer. But he did it in one line of fortran.