No `XDeclaredButNotUsed` in overloaded function with generics

[tersec]

Description

proc n() = discard
proc n(g: int|int) = n()

Nim Version

Nim Compiler Version 1.6.20 [Linux: amd64]
Compiled at 2024-03-20
Copyright (c) 2006-2023 by Andreas Rumpf

git hash: 8f9fde0615c686357774736c0bce6f3c2075a94d
active boot switches: -d:release

Nim Compiler Version 2.0.4 [Linux: amd64]
Compiled at 2024-03-20
Copyright (c) 2006-2023 by Andreas Rumpf

git hash: d4b58b0b068475e937e9e25af4f11f649285274c
active boot switches: -d:release

Nim Compiler Version 2.1.1 [Linux: amd64]
Compiled at 2024-03-20
Copyright (c) 2006-2024 by Andreas Rumpf

git hash: 6c4c60eade66e45448a220f4c8a91f866b76baf7
active boot switches: -d:release

Current Output

No XDeclaredButNotUsed hints

Expected Output

r.nim(2, 6) Hint: 'n' is declared but not used [XDeclaredButNotUsed]

Possible Solution

No response

Additional Information

No response

[Araq]

But it really is not used as you don't instantiate the generic.

[tersec]

But it really is not used as you don't instantiate the generic.

Exactly. And yet there's no XDeclaredButNotUsed hint.

[Araq]

Ah, but it is used in the generic lookup pass. :P In fact, the overload is what causes it to be an "open symchoice". Screwed if you emit the warning, screwed if you don't.

[metagn]

This mixes in both n symbols and soft marks them as used, meaning only the warning is omitted but things like {.error.} don't trigger.

Otherwise this would warn that every n is unused if no code uses foo:

proc n(x: int) = discard
proc n(x: float) = discard
proc foo*(x: int | float) =
  n(x)

The issue is marking the symbol of the proc itself used in general, this also doesn't give an unused warning:

proc foo() =
  foo()

[tersec]

The issue is marking the symbol of the proc itself used in general, this also doesn't give an unused warning:

proc foo() =
  foo()

Sure, https://github.com/nim-lang/Nim/issues/20264 https://github.com/nim-lang/Nim/issues/22274 also an issue, but there's the higher-level semantic argument there (which I disagree with, but it's at least therefore a distinct case) that in some sense recursive functions are always used, by themselves.

This is different because the n(int | int) is unambugously not used, that Nim does this analysis by soft-marking symbols of the procs notwithstanding, so it's a more unambiguous case.

[Araq]

The real solution here is to type-check generics so that the "pre pass" is not required anymore and it's clear what is going on within a generic (nothing special at all). But lacking this, the overload is not unused since you cannot just remove it and get equivalent behavior in all the edge cases. So a warning would be more harmful then helpful.