reduce MAX_TUPLETYPE_LEN

[JeffBezanson]

A less ambitious version of #16628. The overhead of this in the compiler is quite significant, and not really worth the tradeoff of better-performing 42-dimensional arrays. This parameter controls how far inference will follow recursion, so for example a recursive variadic function might be analyzed 42 times. Eventually we will need a smarter, more surgical fix for this issue, so that we can specialize code for high-dimensional arrays while still avoiding non-termination.

[timholy]

👍 Thanks for the eagle eye here in terms of performance implications.

[andyferris]

@JeffBezanson +1 for eventually thinking about a "smarter, more surgical fix". I've been wondering if these parameters may one-day be controlled with a :meta expression or similar on a per-method, per-instruction or per-type basis? Similar to @inbounds and @inline, but something like @infer_the_crap_out_of_this (joking).

My current case of interest is TypedTables.jl where it's difficult to get great code-gen for tables with more than 8 columns (or is it 42, or 16? 😛 ). I'm currently trying to get a better understanding of inference.jl and although I have gained quite a lot of insight on matching my experimentation with the magic of inference and constant propagation, I'm still having trouble seeing how it all comes together (so that I might perhaps contribute one day).