sdist ships (duplicate) generated modules in the source directories

[andreasabel]

Originating from https://github.com/haskell/cabal/issues/7251#issuecomment-767487524 : cabal sdist includes useless, possibly generated modules in the package Scenario:

• Say I have in my modules in src/ a module M that is generated from M.x (alex lexer generator source file).
• I run alex M.x in src, generating M.hs there.
• Now module M is formally ambiguous, but cabal build will ignore the M.hs and use M.x.
• However, sdist will include M.hs in the package.
• I can even put arbitrary garbage into M.hs that will ignore in the build, but still be shipped with sdist.

Enhancement request:

• If a module resolves to several source files, only sdist those that will actually be used by build.

To Reproduce Steps to reproduce the behavior (uses BNFC and PR #7251): [Without PR #7251, you might have to fix up the generated .cabal file.]

bnfc --version || cabal install BNFC
mkdir -p myparser
cd myparser
echo "Foo. Bar ::= Integer" > Foo.cf
bnfc -o src -d -m Foo.cf
make -C src
mv src/Foo/Test.hs Main.hs
cabal init --libandexe --application-dir=.
cabal sdist
find . -name "*.tar.gz" -exec tar tf {} \;

Notice the content of the generated archive with the redundant .hs files:

myparser-0.1.0.0/src/Foo/Lex.hs
myparser-0.1.0.0/src/Foo/Lex.x
myparser-0.1.0.0/src/Foo/Par.hs
myparser-0.1.0.0/src/Foo/Par.y

System information

• cabal 3.5.0.0
• ghc 8.10.3

[andreasabel]

In the wild: cabal install java-adt-0.2018.11.4 fails. The tarball contains a generated Lexer.hs and Parser.hs adjacent to Lexer.x and Parser.y: https://hackage.haskell.org/package/java-adt-0.2018.11.4/src/

• dist/build/java-adt/java-adt-tmp/
  • Lexer.hs
  • Parser.hs
• src/
  • Lexer.hs
  • Lexer.x
  • Main.hs
  • Options.hs
  • Parser.hs
  • Parser.y