[PATCH] D10272: Fix alignment issues in Clang.

[Reid Kleckner]

I think there were issues with trailing arrays, which is why Clang ended up settling on this pattern in the first place. Barring significant changes, I think wrapping this up with some library code is the way to go. That's kind of what we were thinking with getTrailingObjects(). :-)

I think for the oddball cases like multiple trailing arrays, we can do stuff like this:

  FunctionTemplateDecl * const *getTemplates() const {
    return getTrailingObjects<FunctionTemplateDecl*const>(&getTemplateArgs()[NumArgs-1]);
  }

In general, we probably want to lay out trailing arrays in order of higher alignments to lower alignments to avoid the extra alignment code, and if we get it wrong, the wrapper will tell us.

For the DeclRefExpr::getFoundDeclInternal() case, I could see adding some kind of wrapper struct with 8 byte alignment around the found found Decl pointer.

As far as documentation goes, I'd rather update the Doxygen to indicate which classes have trailing data. Most of the commonly used classes like DeclRefExpr are already annotated as such.

Another benefit of the getTrailingObjects wrapper is that we can static_assert that the class with trailing data is `final`.

In conclusion, I guess I would like to push for this wrapper. It seems like it moves incrementally closer towards the better end state that you want, which is less pointer arithmetic and casts sprinkled across the AST. Sound reasonable?


http://reviews.llvm.org/D10272