Quoting an internal variable causes an error when splicing

{-# LANGUAGE TemplateHaskell #-}
module A where

import B

main = $$f

{-# LANGUAGE TemplateHaskell #-}
module B(f) where

d = 0

f = [|| d ||]

Note that d is not exported from B.

[1 of 2] Compiling B                ( B.hs, B.o )
[2 of 2] Compiling A                ( A.hs, A.o )

A.hs:6:8: error:
    • Can't find interface-file declaration for variable B.d
        Probable cause: bug in .hi-boot file, or inconsistent .hi file
        Use -ddump-if-trace to get an idea of which file caused the error
    • In the expression: B.d
      In the result of the splice:
        $f
      To see what the splice expanded to, use -ddump-splices
      In the Template Haskell splice $$f
  |
6 | main = $$f
| ^^^

This doesn't seem to happen for untyped quotes/splices.

Trac metadata

Trac field	Value
Version	8.6.1
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture

changed weight to 5

added Tbug Trac import labels

Trac metadata

Trac field	Value
Component	Compiler → Template Haskell

Curiously, this doesn't happen when loaded into GHCi, only when compiled with --make.

Also, if you compile these files one at a time with -c, you get something... interesting:

$ /opt/ghc/8.6.1/bin/ghc -c B.hs
$ /opt/ghc/8.6.1/bin/ghc -c A.hs
ghc: panic! (the 'impossible' happened)
  (GHC version 8.6.1 for x86_64-unknown-linux):
        Loading temp shared object failed: /tmp/ghc6527_0/libghc_1.so: undefined symbol: templatezmhaskell_LanguageziHaskellziTHziLib_varE_closure

Another curious thing. If you compile B.hs with -ddump-simpl, you'll get this:

$ /opt/ghc/8.6.1/bin/ghc B.hs -fforce-recomp -ddump-simpl
[1 of 1] Compiling B                ( B.hs, B.o )

==================== Tidy Core ====================
Result size of Tidy Core
  = {terms: 24, types: 9, coercions: 0, joins: 0/0}

-- RHS size: {terms: 9, types: 1, coercions: 0, joins: 0/0}
f :: Language.Haskell.TH.Syntax.Q
       (Language.Haskell.TH.Syntax.TExp Integer)
[GblId]
f = Language.Haskell.TH.Syntax.unsafeTExpCoerce
      @ Integer
      (Language.Haskell.TH.Lib.Internal.varE
         (Language.Haskell.TH.Syntax.mkNameG_v
            (GHC.CString.unpackCString# "main"#)
            (GHC.CString.unpackCString# "B"#)
            (GHC.CString.unpackCString# "d"#)))

<Typeable stuff elided>

Now change the typed quote to an untyped one (i.e., change f = [|| d ||] to f = [| d |]) and recompile it:

$ /opt/ghc/8.6.1/bin/ghc B.hs -fforce-recomp -ddump-simpl
[1 of 1] Compiling B                ( B.hs, B.o )

==================== Tidy Core ====================
Result size of Tidy Core
  = {terms: 25, types: 7, coercions: 0, joins: 0/0}

-- RHS size: {terms: 8, types: 0, coercions: 0, joins: 0/0}
f :: Language.Haskell.TH.Lib.Internal.ExpQ
[GblId]
f = Language.Haskell.TH.Lib.Internal.varE
      (Language.Haskell.TH.Syntax.mkNameG_v
         (GHC.CString.unpackCString# "main"#)
         (GHC.CString.unpackCString# "B"#)
         (GHC.CString.unpackCString# "d"#))

-- RHS size: {terms: 1, types: 0, coercions: 0, joins: 0/0}
d :: Integer
[GblId, Caf=NoCafRefs, Unf=OtherCon []]
d = 0

<Typeable stuff elided>

Notice that d actually makes an appearance this time! This makes me wonder if d is being mistakenly removed under the pretense of it being dead code.

I figured out what's causing this. See D5248 for a fix.

It turns out that occurrence analysis was dropping the binding for d during desugaring (this explains why this bug didn't surface in GHCi, as occurrence analysis does not run there). So why was occurrence analysis dropping bindings for top-level things referenced from typed TH quotes, but not from untyped TH quotes? It turns out that two functions called checkCrossStageLifting are to blame.

1. ..No seriously, there are two completely separate functions named checkCrossStageLifting in GHC. One exists in RnSplice, and only handles untyped quotes, and the other exists in TcExpr, and only handles typed quotes. And wouldn't you know it, RnSplice.checkCrossStageLifting was doing something that TcExpr.checkCrossStageLifting wasn't doing. In particular, these lines of RnSplice.checkCrossStageLifting are crucial:

  | isTopLevel top_lvl
        -- Top-level identifiers in this module,
        -- (which have External Names)
        -- are just like the imported case:
        -- no need for the 'lifting' treatment
        -- E.g.  this is fine:
        --   f x = x
        --   g y = [| f 3 |]
  = when (isExternalName name) (keepAlive name)
    -- See Note [Keeping things alive for Template Haskell]

That call to keepAlive ensures that the binding for f (from the example in the comments) doesn't get discarded during occurrence analysis. TcExpr.checkCrossStageLifting wasn't doing anything like this, which explains why this bug exists.

Trac metadata

Trac field	Value
Differential revisions	- → D5248

mentioned in commit bb835c96

closed

changed milestone to %8.8.1

Trac metadata

Trac field	Value
Resolution	Unresolved → ResolvedFixed
Test case	- → th/T15783

added Pnormal label

mentioned in issue #19662