Regression from v2.0.8 to v2.0.10/version-2-0/2.2/devel: Error: request to generate code for .compileTime proc: $

[tersec]

Description

proc d(_: static[string]) = discard $0
proc m(_: static[string]) = d("")
iterator v(): int {.closure.} =
  template a(n: untyped) =
    when typeof(n) is void:
      n
    else:
      n
  a(m(""))
let _ = v

Nim Version

Builds:

Nim Compiler Version 2.0.8 [Linux: amd64]
Compiled at 2024-10-03
Copyright (c) 2006-2023 by Andreas Rumpf

git hash: 5935c3bfa9fec6505394867b23510eb5cbab3dbf
active boot switches: -d:release

Does not build:

Nim Compiler Version 2.0.10 [Linux: amd64]
Compiled at 2024-10-03
Copyright (c) 2006-2023 by Andreas Rumpf

git hash: e941ee15be775fe3c46db1bed9b4f41c7dfb1334
active boot switches: -d:release

Nim Compiler Version 2.2.0 [Linux: amd64]
Compiled at 2024-10-03
Copyright (c) 2006-2024 by Andreas Rumpf

git hash: 78983f1876726a49c69d65629ab433ea1310ece1
active boot switches: -d:release

Nim Compiler Version 2.2.1 [Linux: amd64]
Compiled at 2024-10-03
Copyright (c) 2006-2024 by Andreas Rumpf

git hash: d6a71a10671b66ee4f5be09f99234b3d834e7fce
active boot switches: -d:release

Current Output

/tmp/h.nim(1, 37) Error: request to generate code for .compileTime proc: $

Expected Output

No response

Known Workarounds

No response

Additional Information

No response

[metagn]

Simplified

proc d[T]() = discard $0
proc v() =
  when typeof(d[string]()) is void:
    d[string]()
v()

The original d called as d("") has the same effect, just signaling that d has to be generic

No idea what the root issue is but worst case we can remove these compileTime overloads for $ for int literals:

Nim/lib/system/dollars.nim

Lines 29 to 34 in 7dfadb8

	template gen(T) =
	# xxx simplify this by supporting this in compiler: int{lit} | uint64{lit} | int64{lit}
	func `$`*(x: T{lit}): string {.compileTime.} = result.addInt(x)
	gen(int)
	gen(uint64)
	gen(int64)

[metagn]

The issue is that, when we call typeof, d is instantiated with the belief that we are still in typeof, so the compile time folding of $0 is disabled in the body of d. But then we actually call d, which means the leftover unfolded $0 call tries to get compiled which fails.

We can make the compiler stop believing we're in typeof if a generic proc is being instantiated but that sounds too special cased. I would really prefer reverting the typeof special case entirely from #22022. To make declval work with the rest of #22022, we could split compileTime into another flag which enables/disables the eager folding, but this is a new feature and would still break one of either stew or other packages.

At the very least we should revert #22022 from the 2.0 branch.