error FS0192: internal error: Error in pass3 for some F# code.

[sylvainouel]

Repro steps

Copy/paste the following program in F# interactive:

type A = {
    F: uint64  
}

type B = {
    C: uint64[]
}

let g x =
    let c = 
        [|
            for d in x.C do
                {
                    F = d
                }
        |]
    c

It produces the following error:

error FS0192: internal error: Error in pass3 for type FSI_0003, error: ldelem

Expected behavior

I think it should build fine.

Actual behavior

internal error

Known workarounds
changing the uint64 to int64 fixes it.

Related information

It used to work, it stopped working after upgrading to latest VS.

• Operating system
  Windows 11

• .NET Runtime kind (.NET Core, .NET Framework, Mono)

.NET 9

• Editing Tools (e.g. Visual Studio Version, Visual Studio)
  Microsoft Visual Studio Enterprise 2022 (64-bit) - Current
  Version 17.12.1

[brianrourkeboll]

This seems like it could be related to some of my changes. I will try to take a look tomorrow.

[T-Gro]

This a regression from NET8, I also suspect the lowering changes being applied here.
Even though the previous version was already optimized to Array.map, so it could be a different cause of the regression.

https://sharplab.io/#v2:DYLgZgzgPg9gDgUwHYAIDKBPCAXBBbAWACgBtAHgDkYBRARwFcBDYAS2wwG4qBhGPORgCcWEGEgB8AXWLtEKAIIoAvCgDexFJpQAxECnosk2AGwAWTcQC+xclTpNW7LjF78hIsVJkY5AIWVqGlrcegZGZiTSRNZExMAI2CgA5igAHspBmvGJAMYBmVqaJFAFhYVgMIIoACYohmkAdNw1MKVl7epE7d3d2gHVbT2aMT1QUYU5QA==

[brianrourkeboll]

Here's a more minimal repro:

[|for x in [|1UL|] -> x|]

The same thing applies for unativeint:

[|for x in [|1un|] -> x|]

The problem seems likely to be here

fsharp/src/Compiler/Optimize/LowerComputedCollections.fs

Lines 393 to 394 in 95fe045

	let ldelem = mkIlInstr g I_ldelem (fun ilTy -> I_ldelem_any (ILArrayShape.SingleDimensional, ilTy)) srcIlTy
	let stelem = mkIlInstr g I_stelem (fun ilTy -> I_stelem_any (ILArrayShape.SingleDimensional, ilTy)) destIlTy

and here

fsharp/src/Compiler/Optimize/LowerComputedCollections.fs

Lines 360 to 374 in 95fe045

	let private mkIlInstr (g: TcGlobals) specific any ilTy =
	if ilTy = g.ilg.typ_Int32 then specific DT_I4
	elif ilTy = g.ilg.typ_Int64 then specific DT_I8
	elif ilTy = g.ilg.typ_UInt64 then specific DT_U8
	elif ilTy = g.ilg.typ_UInt32 then specific DT_U4
	elif ilTy = g.ilg.typ_IntPtr then specific DT_I
	elif ilTy = g.ilg.typ_UIntPtr then specific DT_U
	elif ilTy = g.ilg.typ_Int16 then specific DT_I2
	elif ilTy = g.ilg.typ_UInt16 then specific DT_U2
	elif ilTy = g.ilg.typ_SByte then specific DT_I1
	elif ilTy = g.ilg.typ_Byte then specific DT_U1
	elif ilTy = g.ilg.typ_Char then specific DT_U2
	elif ilTy = g.ilg.typ_Double then specific DT_R8
	elif ilTy = g.ilg.typ_Single then specific DT_R4
	else any ilTy

I think we need to be emitting ldelem.i8 and ldelem.i for uint64 and unativeint, respectively. (And maybe likewise for stelem?)

Even though the previous version was already optimized to Array.map

That was an intermediate optimization introduced in #16948, but superseded in #17067 in part to fix debug stepping in the loop body. SharpLab still has a preview version of 9 that doesn't include #17067, #17419, #17711.

[brianrourkeboll]

Known workarounds
changing the uint64 to int64 fixes it.

@sylvainouel Another workaround for the time being would be to set <LangVersion>8.0</LangVersion> in your project file.

[brianrourkeboll]

I opened a PR that I believe should fix this: #18081.

[sylvainouel]

Thank you for this fix!