Personally - I know this is going to be unpopular - I suggest to not think at the grammar at all.
There has been a time in the past in which I was all around BNF and stuff. My last parser has been going for a while now and I still have no explicitly written grammar.
how do I make sure that for "return 3 + 5;" the proper rules (D and C_2) are used?
I'd do first a keyword match - just for stuff as for, which got pretty odd syntax for example. In this case we match return.
Then we find a literal. Looking ahead 1 token (the whole point of LR) we find a + token which is recognized as a operator. It's a binary op because we found it after a literal, so we fetch something else. The point is: don't be greedy! Don't switch just because you matched now.
At this point we have this expression parsed: the compiler will have to find what 3 and 5 are so it can emit proper ADD instruction.
I actually do expression assembly in the compiler, someone could say because of poor design.
Every time I think I came up with something I end up with a function so long and ugly (and non operational) that I can't help but think I'm missing something. Does anyone have some pointers?
Are you trying to parse and compile at the same time? This will end in tears in my experience. Pre-tokenization is a must in my opinion (and no, I don't care about what GCC/CompilerX does).
Object orientation might help you - in my experience this comes at a negligible cost for example you can have a loop "by keyword match" which is very compact and yet dispatches the correct syntax without visible ifs. You'll need to provide a set of basic compiler features such as type lookup. I've done this successfully with an base interface.
Performance wise, I once had... a problem with my data import routines. So I made a Notepad++ script which would encode all that data in a program... which turned out like 3000 lines long. It took a while to process (like 10 secs) but it was acceptable as a band-aid solution.
edit: two small clarifications.