Regression from 2.0 to devel with dirty template

[tersec]

Description

a.nim

import "."/b

type
  B = object
  L = object

template F(T: type B): type = F(Json, B)
var j = F(B).init()
var f: L
s(j, f)

b.nim

type P = object

template d(Name: untyped) {.dirty.} =
  type Name* = object

import "."/c

d(Json)

type K[Flavor = P] = object
  lex: V

template F*(T: type Json, F: distinct type = P): type = K[F]

proc init*(T: type K): T = discard

proc s*[T](r: var K, value: var T) =
  x(r.lex)

c.nim

type
  U* = enum
    errNone
  V* = object
    err: U

template x*(lex: V) {.dirty.} =
  lex.err = errNone

Nim Version

Compiles:

Nim Compiler Version 2.0.5 [Linux: amd64]
Compiled at 2024-05-15
Copyright (c) 2006-2023 by Andreas Rumpf

git hash: 2f399807cb45ea073e1c1bb640d16956e58fcae0
active boot switches: -d:release

Does not compile:

Nim Compiler Version 2.1.1 [Linux: amd64]
Compiled at 2024-05-15
Copyright (c) 2006-2024 by Andreas Rumpf

git hash: b42f1ca8a4d574b087af9d86b7627a5817dad9b0
active boot switches: -d:release

Current Output

a.nim(10, 2) template/generic instantiation of `s` from here
b.nim(18, 4) template/generic instantiation of `x` from here
c.nim(8, 13) Error: undeclared identifier: 'errNone'
candidates (edit distance, scope distance); see '--spellSuggest': 
 (4, 5): 'NimNode'
 (4, 5): 'range'

Expected Output

No response

Possible Solution

No response

Additional Information

No response

[metagn]

Simplified:

# a.nim
import "."/b

s[int]()

# b.nim
import "."/c

type K = object
  lex: V

proc s*[T]() =
  var r: K
  x(r.lex)

# c.nim
type
  U* = enum
    errNone
  V* = object
    err: U

template x*(lex: V) {.dirty.} =
  lex.err = errNone

Removing the lex: V parameter from x gives no error, but changing lex.err = to let y = or errNone to a const instead of an enum symbol gives the same error on 2.0. dirty doesn't bind anything by default, so errNone was never accessible as a symbol through lookup.

So this very likely fails now because of #23588, which again, means this only worked since errNone is specifically an enum member in a context where its type can be inferred. I don't know if it's worth making this specific case work by adding back awkward behavior to the compiler. Is it too much of a problem to either bind errNone or not use dirty here?

[jangko]

We need the dirty template because sometimes we pass in statement such as return. But using bind also works.

[tersec]

To the extent status-im/nim-json-serialization#90 solves this, I'm fine with that. I reported this as a regression fairly mechanically, not so much prescriptively as descriptively. But I tend to agree it accidentally hit a weird edge case.