Re-fix r51907 in a way which doesn't affect valid code. This essentially

moves the check for the invalid construct to a point where it doesn't 
affect other uses of isIntegerConstantExpr, and we can warn properly 
when the extension is used.  This makes it a bit more complicated, but 
it's a lot cleaner.

Steve, please tell me if this check is sufficient to handle the 
relevant system header.  I know it's enough to handle the testcase, but 
I don't know what exactly the original looks like.

llvm-svn: 51918