untyped param is typed when there's an overload with a typed param in same position

[timotheecour]

I strongly suspect this explains a few of other issues related to untped ...
EDIT: indeed, it's root cause for #14827 plus other issues listed there.

• an untyped param is incorrectly reported as typed (nnkSym) when the caller is overloaded
• callsite on the other hand correctly reports the AST as untyped

this is the root cause for the CI failure in #17156 when (undone in b81a987) when let checked = conditions was attempted in place of let checked = callsite()[1] (ie, callsite was required to make it work for the reduced example refs #17156 (comment)).

Example

when defined case9:
  import macros

  macro check2(body: untyped): untyped =
    echo "---"
    echo callsite().repr
    let a1 = body
    let a2 = callsite()[1]
    echo a1.treeRepr
    echo a2.treeRepr

  block:
    let check2 = 123
    var b1 = 1
    check2(b1)

  block:
    var b2 = 1
    check2(b2)

  block:
    proc check2(a: float, b: string) = discard
    var b3 = 1
    check2(b3)

  block:
    macro check2(a: typed, b: string) = discard
    var b4 = 1
    check2(b4)

  block:
    macro check2(a: untyped, b: string) = discard
    var b5 = 1
    check2(b5)

Current Output

---
check2(b1)
Sym "b1"
Ident "b1"
---
check2(b2)
Ident "b2"
Ident "b2"
---
check2(b3)
Sym "b3"
Ident "b3"
---
check2(b4)
Sym "b4"
Ident "b4"
---
check2(b5)
Ident "b5"
Ident "b5"

Expected Output

all Sym should be Ident

Additional Information

1.5.1 c274e67

this is a bit tricky. What's probably happening is that when compiler semchecks check2(b3), since check2 is overloaded, it runs sigmatch to pick the correct overload, which (currently) involves sem-checking the params (b3); by that point b3 is a typed AST.

For some reason callsite is correct, so the compiler must be caching the original AST somewhere.

proposed solution

make sigmatch smarter so that it avoids semchecking a param that's untyped if it can determine it's ok to do so;
TODO: elaborate on details...

[Araq]

Well known sigmatch bug, hard to fix, unfortunately. sigmatch could really see some refactorings...