agda / agda

Reported by @plt-amy in #7156 (comment)

open import Agda.Builtin.Nat

record T : Set where
  field
    f : Nat → Nat

foo : T
foo .T.f = suc

test : Nat
test = foo 123

Agda 2.5.3 said (good):

T should be a function type, but it isn't
when checking that 123 is a valid argument to a function of type T

Since 2.5.4 Agda says (bad):

T !=< Nat of type Set
when checking that the inferred type of an application
  T
matches the expected type
  Nat

@plt-amy wrote:

I suspect that this would be fixable by checking that length tel >= sArgsLen in Application.hs.

It seems that telViewUpTo' here is not sufficient, correct would be a "telViewExactly" that asks for exactly sArgsLen visible arguments:

agda/src/full/Agda/TypeChecking/Rules/Application.hs

Line 658 in d959b19

TelV tel tgt <- telViewUpTo' sArgsLen visible sFun

So, yes, we need to check length tel >= sArgsLen otherwise the target type is bogus and checking it is moot.

Possibly this regression originates from e49d1ca. Unfortunately, we didn't seem to have a test like the OP back then, so the regression went unnoticed.

I wonder whether one could find a meaner counterexample, e.g. where the checking against the bogus target instantiates some meta variables which make type checking fail with an even more obscure error.

Fixed by 084dd4d.

That was speedy, thank you!

Well, I just implemented the fix you outlined...

Non-sensical error since 2.5.4 when applying a non-function