Filter out constraint tuple class constructors in hasNoTypeClasses

[Mikolaj]

A rough sketch. I don't know what I'm doing. See #74.

[nomeata]

Thanks, decent start, I can clean up it a bit before merging. But I’ll wait for the discussion in #74, to see if all your issues can be resolved.

[Mikolaj]

I'm additionally filtering out heqTyConName and eqTyConName, because both appear in my code, I think. Now inspection-testing no longer complains about ~ classes spotted, so it seems to work. Let me know if that's a bad idea.

[Mikolaj]

I've reverted the last change --- it works great with ''(~) --- I had no idea to try something like that.

[nomeata]

Can you do it for the tuples as well? (Mostly curious, it might still be good to do so by default)

[Mikolaj]

Any idea how to try? I did try adding ''(%,,,,,,,,%) to the list and ''$d(%,,,,,,,,%) and ''d(%,,,,,,,,%) and each is rejected with "parse error on input ‘,’".

[nomeata]

I would have expected the first to work. Nevermind, not that important.

[Mikolaj]

I don't think I will have any more actionable input for now.

Your tool, after patching, helped me manually specialize my code much more, to the point I've discovered a new bug in GHC and/or plugins (https://gitlab.haskell.org/ghc/ghc/-/issues/23866), which makes GHC 9.4 unusable for me. I still have one GHC version that works correctly, GHC 9.2, though it probably doesn't specialize as well --- I can't tell, because I needed to add as exceptions all classes that constrain existential variables, so I can't easily detect failures to eliminate these dictionaries in non-existential contests (the lack of an explicit exists quantifier in Core and in Haskell surface language doesn't help, either):

inspect $ hasNoTypeClassesExcept 'revRankedListProd [''GoodScalar, ''KnownNat, ''OS.Shape, ''AstSpan, ''Show, ''Ord, ''Numeric, ''Num, ''RowSum, ''Typeable, ''IfDifferentiable, ''NFData, ''RealFrac, ''Integral, ''Fractional, ''RealFloat, ''Floating]  

BTW, I've also just discovered, using your tool, than GHC 9.8.1-alpha1 specializes worse than GHC 9.4, even in its unsound -fpolymorphic-specialisation mode (https://gitlab.haskell.org/ghc/ghc/-/issues/23874). [Edit: I was wrong, this is not a regression; GHC 9.4 gave up specialization even earlier, which is why there was no mention of IsPrimal class in the benchmark file, which is why your tool did not report problems --- this is the non-whole-program nature of this analysis showing its limitations. I'd need to add hasNoTypeClassesExcept to ~5 modules to make sure I catch non-specialized IsPrimal regardless of where GHC gets stuck specializing it.]

[nomeata]

Well, that's at least something :-)

[Mikolaj]

So, does it have a chance to get merged? :)

[Mikolaj]

Oh, yes, the CI fails, because I cheated for the old GHCs. :)

[Mikolaj]

Thank you!