[PATCH] Improved __builtin_constant_p

[Richard Smith]

This makes Clang's constant expression evaluator and its IR generation disagree about the value of `__builtin_constant_p`. That seems dangerous to me, so I'd like to understand the justification for this a bit better. Random example of why this is dangerous:

  int f(int n) {
    if (!__builtin_constant_p(n))
      return 0;
  }
  int k = f(0);

Here, we'll suppress the `-Wreturn-type` warning because the frontend believes that it's impossible to reach the `}` of `f`. But when we generate IR, we'll fall off the function and experience nasal demons.

What's the motivation behind this change?

http://reviews.llvm.org/D4492