regression from 2.2.10 to `devel` with `lent` and type aliasing

[tersec]

Nim Version

Pre-regression:

Nim Compiler Version 2.2.10 [Linux: amd64]
Compiled at 2026-05-06
Copyright (c) 2006-2026 by Andreas Rumpf

git hash: bfeb3146d1638b39f69007a4ae5a23e23ae4e5ef
active boot switches: -d:release

Post-regression:

Nim Compiler Version 2.3.1 [Linux: amd64]
Compiled at 2026-05-06
Copyright (c) 2006-2026 by Andreas Rumpf

git hash: f0077a12b20a6cbf3358eaeb09e528ec65e9eca9
active boot switches: -d:release

Description

type A = object
  c: int
type H = proc(): lent A {.nimcall.}
const u = A(c: 0)
proc e(T: typedesc): lent A = u
proc y(T: typedesc): H =
  proc(): lent A {.nimcall.} = T.e
discard y(int)

Current Output

/tmp/v.nim(8, 10) template/generic instantiation of `y` from here
/tmp/v.nim(7, 7) Error: type mismatch: got 'proc (): lent A{.nimcall, noSideEffect, gcsafe.}' for 'proc (): lent A {.nimcall.} = result = e(T)' but expected 'H = proc (): lent A{.nimcall.}'

Expected Output

Known Workarounds

No response

Additional Information

No response

[tersec]

type A = object
  c: int
type H = proc(): lent A {.nimcall.}
let u = A(c: 0)
proc y(_: int | int): H =
  proc(): lent A {.nimcall.} = u
discard y(0)

No typedesc, and either let u = A(c: 0) or const u = A(c: 0) works.

f2e4ae0 does not exhibit this issue, while 568eccd does. This suggests

• fixes #25617; handle backend type aliasing in procParamTypeRel #25692

as a proximate trigger.

[tersec]

No generics and no nested proc, effectively the same issue (a bit different because it's not noSideEffect anymore):

type A = object
  c: int
type H = proc(): lent A {.nimcall.}
let u = A()
let _: H = proc(): lent A {.nimcall.} = u

This one produces the variant:

/tmp/v.nim(5, 16) Error: type mismatch: got 'proc (): lent A{.nimcall, gcsafe.}' for 'proc (): lent A {.nimcall.} = result = u' but expected 'H = proc (): lent A{.nimcall.}'

[ringabout]

I will strip lent/var/... when comparing proc types, which are treated as different types

[Araq]

That is not sound, proctypes must agree on all type modifiers.

[ringabout]

Yeah, it only strips the lent/var/ for sameBackendTypePickyAliases after typeRel agrees that they are the same param type