This is an archive of the discontinued LLVM Phabricator instance.

Full support for dllexport/dllimport attributes
AbandonedPublic

Authored by nrieck on Jul 4 2013, 10:25 AM.

Details

Reviewers
None
Summary

This patch series implements full support for dllexport/dllimport attributes

  • Adds support for class-wide export/import attributes (PR11170)
  • Handles dllexport/dllimport on inline functions and templates

Fixes PR11170, PR13450

Implemented semantics:

  1. dllexport and dllimport can be applied to functions, static data members and global variables. Both can also be applied to typedefs and enums (without effect) if -fms-extensions is specified, for all other cases the attributes are ignored and a warning is issued. It is an error to apply them to non-static data members.
  1. dllexport requires a definition. The extern keyword can be used to force a declaration.
  1. dllimport requires a declaration.
  1. dllexport exposes the function or variable with its decorated name. a) For C++, this includes name mangling. b) For C or extern "C" functions, this includes platform-specific decoration for the calling convention.
  1. dllexport takes precedence over dllimport. If both are specified, dllimport is ignored and a warning is issued.
  1. Classes a) When applied to a class, dllexport or dllimport affect all static and non-static member functions and all static data members, including all non-deleted special members. b) Everything that is accessible to the DLL's client according to C++ access rules is exported (including private data members referenced in inline functions). c) Members can be imported or exported selectively. d) Explicit dllexport/dllimport attributes on members of exported/imported classes are forbidden. e) Base classes of an exported class should be exported. Emits a warning otherwise. (-Wmissing-dllexport) f) Virtual tables and typeinfo follow the emission rules of the Itanium ABI, i.e. for an exported/imported dynamic class with key function, the needed globals are exported/imported. If there is no key function, they are emitted everywhere and are not exported/imported.
  1. Inline functions a) dllexport can be applied to inline function definitions. The function is always instantiated and exported, presumed to be imported by another program. b) dllimport can be applied to inline function definitions. Such functions can be expanded, but never instantiated.
  1. Templates a) If a class template is exported, explicit and implicit instantiations are also exported. b) Non-exported/imported class templates can be exported/imported by decorating an explicit instantiation. c) Base classes that are template specializations must be instantiated explicitly to export them. d) If a class is marked as dllexport, any specializations of class templates in the class hierarchy are implicitly marked as declspec(dllexport). This means that class templates are explicitly instantiated and the class's members must be defined.

Implementation notes:

Representing dllexport/dllimport as distinct linkage types prevents using them
on templates and inline functions.

Instead of introducing further mixed linkage types to include linkonce and
weak ODR, the old import/export linkage types are replaced with function
attributes (DLLExport, DLLImport) and an extension to global variables, similar
to how alignment is specified:

@Exp = global i32 1, align 4, dllexport
@Imp = external global i32, dllimport

Linkage for dllexported globals and functions is now equal to their linkage
without dllexport. All dllexported entities are decorated with UsedAttr to
ensure their emission. Imported globals and functions must be either
declarations, or definitions with AvailableExternallyLinkage.

In some cases a dllimport/dllexport attribute is ignored. If this happens on
a member of an exported/imported class, we need a way to drop the attribute.
During Sema, when checking such a class, every member is decorated with an
inherited attribute, which allows subsequent dropping. This requires a small
change (or possibly fix?) to table-gen so that the inherited status of an
attribute is preserved when instantiating template attributes.

Open questions/issues:

  1. With two linkage types removed, should the bitcode mapping of linkage types to integers be adjusted to remove the "holes"?
  1. Should aliases be exportable/importable? For now an alias is exported if the aliasee is exported.
  1. Currently dynamic classes with key function cannot have individual virtual members exported as this leads to missing exports for typeinfo.
  1. The CodeGen tests are extremely slow due to CHECK-DAG, and without CHECK-DAG the tests are very hard to understand because virtual/inline functions are emitted in a different order. Suggestions?

Please note: The uploaded patch is a combination of two git diffs. I prefixed
each one with the project-path from the SVN repo.

Diff Detail

Event Timeline

nrieck abandoned this revision.Jul 5 2013, 6:20 PM
hans added a subscriber: hans.May 15 2014, 11:49 AM

Regarding handling the dll linkage in AST or Codegen, I was kind of hoping doing it in AST would help in functions such as ASTContext::DeclMustBeEmitted (but it currently doesn't).

I don't have enough experience here to feel strongly either way, but I do suspect that it will become clearer which solution is best as we get further along the train of patches towards full support for class-level dllimport/dllexport.

I would like to go ahead and commit this now, and am totally open to moving the linkage adjusting code if it seems like a good idea as we go forward.

hans added a comment.May 15 2014, 11:51 AM

Regarding handling the dll linkage in AST or Codegen, I was kind of hoping doing it in AST would help in functions such as ASTContext::DeclMustBeEmitted (but it currently doesn't).

I don't have enough experience here to feel strongly either way, but I do suspect that it will become clearer which solution is best as we get further along the train of patches towards full support for class-level dllimport/dllexport.

I would like to go ahead and commit this now, and am totally open to moving the linkage adjusting code if it seems like a good idea as we go forward.

Sorry, this comment was for http://reviews.llvm.org/D3772.

cfe/trunk/test/Sema/dllexport.c