[PATCH] Use GEPs correctly when adjusting "this" in MicrosoftCXXABI

[Richard Smith]

================
Comment at: lib/CodeGen/MicrosoftCXXABI.cpp:636-638
@@ +635,5 @@
+    if (ML.VBase) {
+      // Non-virtual adjustment might result in a pointer outside the allocated
+      // object, e.g. if the final overrider class is laid out after the virtual
+      // base that declares a method in the most derived class.
+      // FIXME: Update the code that emits this adjustment in thunks prologues.
----------------
Richard Smith wrote:
> How can this give an offset that takes you outside the complete object?
OK, I think I see. During construction or destruction, we have a static offset based on an assumption that vbases are laid out as if we are the most-derived type, and then an additional offset to correct that assumption. And the vbase layout can change between this class and the most-derived class such that the location where we thought the vbase would be is no longer within the object. Something like:

  struct A { int x, y; };
  struct B : virtual A { int n; };
  struct C : virtual A, virtual B {};

  |-- complete-object layout of B --|
  |-- B members --| |-- A members --|

  |-- complete-object layout of C --|
  |-- A members --| |-- B members --|

During the construction of B (within the construction of a C complete object), if we want to access the 'y' member, we add an offset of 8 to get to 'y' assuming we're using B's complete-object layout, then subtract off 12 to adjust for the A vbase not being at the offset where we expected. And that transiently takes us outside the object. Is that it?

Ugh.


http://llvm-reviews.chandlerc.com/D1977