__assert_fail should not be __attribute__((__noreturn__))

[Bruno Haible]

Richard Henderson wrote:
> Actually, the most important thing is that GCC can prove that the
> assert condition is false after the assertion.  This knowledge gets
> propagated into following statements.

Yes, it does, but with two major limitations:

  1) As Johannes Sixt mentioned, when NDEBUG is used, glibc's assert()
     expansion does not contain the condition, therefore the compiler
     cannot optimize anything since it doesn't see the condition.
     Also there is no way to tell the compiler "this condition is true,
     you don't need to evaluate it, you can just assume it is true".
     A special built-in, say, __builtin_assert_without_verification(condition),
     could do this, no?

  2) The information propagation in the compiler is quite limited. It cannot
     even deduce that if x >= 10, then also x >= 5: In the following snippet
     gcc doesn't eliminate the second test for x, regardless of the NDEBUG
     setting.

        #include <assert.h>
        int foo (int x, int *p) {
          assert (x >= 10);
          return (x >= 5 ? p[1] : 3*p[0]);
        }

Bruno