[cfe-commits] r83816 - in /cfe/trunk: lib/CodeGen/CGCXX.cpp test/CodeGenCXX/virtual-function-calls.cpp

[Sebastian Redl]

Douglas Gregor wrote:
> The type of the expression could be a record type even if the dynamic  
> type of the object isn't known at compile time. If I tweak your  
> example like this:
>
> void f(A &a) {
>    a.f('c');
> }
>
> I now get a direct call, when I should get a virtual call. To do this  
> devirtualization, I suggest abstracting the "find the dynamic type of  
> this expression" logic into a separate function. Then, we can teach it  
> about common cases, such as referring to a VarDecl of (non-reference)  
> type, constructing a temporary of a specific type, etc. I expect that  
> this logic will get smarter over time, as we think of more analyses  
> that can help make this distinction.
>   
Devirtualization is something we should somehow make LLVM do, when we do
other optimizations. For example, take your 'f' above and do this:

void g() {
  A a;
  f(a);
}

If LLVM decides to inline the call to f here, the type of the call to
A::f becomes statically known in the inlined version, and the call
should be direct.

Sebastian