Optimize normalization of == and !=

[sjakobi]

dhall-haskell/dhall/src/Dhall/Eval.hs

Lines 370 to 379 in 5f3b05a

	BoolEQ t u -> case (evalE t, evalE u) of
	(VBoolLit True, u) -> u
	(t, VBoolLit True) -> t
	(t, u) | conv env t u -> VBoolLit True
	(t, u) -> VBoolEQ t u
	BoolNE t u -> case (evalE t, evalE u) of
	(VBoolLit False, u) -> u
	(t, VBoolLit False) -> t
	(t, u) | conv env t u -> VBoolLit False
	(t, u) -> VBoolNE t u

In both cases we currently quickly pull out what I consider to be the rather big hammer of conv. I suspect that we should try to cover all combinations of BoolLits before we resort to conv. I haven't benchmarked this yet though!

[ocharles]

But conv on two VBoolLits has no recursion or anything and just does ==. That is, what I'm reading as the "fast path" is already fast in conv. What would we achieve by adding more custom pattern matches?

[sjakobi]

Oh, good point! I guess I confused conv with judgmentallyEqual here. That means we could optimize normalizeWithM, but that is less important, I guess.

dhall-haskell/dhall/src/Dhall/Core.hs

Lines 1565 to 1578 in 5f3b05a

	BoolEQ x y -> decide <$> loop x <*> loop y
	where
	decide (BoolLit True ) r = r
	decide l (BoolLit True ) = l
	decide l r
	| judgmentallyEqual l r = BoolLit True
	| otherwise = BoolEQ l r
	BoolNE x y -> decide <$> loop x <*> loop y
	where
	decide (BoolLit False) r = r
	decide l (BoolLit False) = l
	decide l r
	| judgmentallyEqual l r = BoolLit False
	| otherwise = BoolNE l r