improve method specificity algorithm

[JeffBezanson]

This code needed some weeding and strengthening. First, I had the existing code write out a trace of all non-subtype specificity results computed during the system image build (45MB, and took so long just to parse that it prompted me to do #22161). Then I added assertions to ensure that morespecific(a,b) => !morespecific(b, a) during the algorithm, and otherwise tried to clean up the logic as much as I could. I tested that against the trace, fixed the cases that needed fixing, and extracted interesting cases into a new specificity test suite.

This makes many changes to the specificity relation, but mostly involving disjoint types, so it does not seem to affect any code behavior, at least in the test suite. However it is very hard to know whether this will break anything.

[timholy]

Thanks! Perhaps check if it fixes issues #22002, #20995, #20714, #20672, #17278, #10174, #6383, #3025.

[vtjnash]

Could we please get rid of all variable names with < 6 characters in this file? I hate to impose some sort of arbitrary limit, but I've been having trouble following along with the recent PRs to this file since I can't keep straight what these acronyms mean. The aggressive compression of variable names in this PR (env -> e, akind -> ak, for example) is only making it worse. (a little more whitespace, so the structure of the code is visible at a glance rather than requiring actual parsing of the code, would also be much appreciated 🙂)

[JeffBezanson]

Can I keep i?

[JeffBezanson]

I will also have trouble thinking of > 6 character names for a, b, and n.

[vtjnash]

yes, you can keep i :)

[JeffBezanson]

Ok I tried to do a few style improvements. Most importantly I think the tuple specificity algorithm itself is now far simpler.

[JeffBezanson]

@timholy I didn't expect much, but lo and behold this does fix #22002! That might have been a similar issue to #22164. Thanks for the very good test case there.

[timholy]

Woot! Thanks for checking (always nice to add to the "this closes n issues" stats 😉). And of course I'm especially grateful for the fix; convert is particularly scary when it comes to specificity, because there are so many methods, so it's awesome that you got it on the way to other things.

[timholy]

Is this backportable?

[JeffBezanson]

The fix for #22002 was backported in #22206.