The ARM 6.6.3 says "Flowing off the end of a function is equivalent to
a return with no value; this is illegal in a value-returning
function".  If they're "equivalent", that implies to me that it should
give a compile time error rather than just returning an undefined
value.

But Stroustrup himself uses the following example in CPL2 4.6.4:

 int fac(int n) {
    if (n>1) return n*fac(n-1);
    else return 1;
 }

Does that mean that a compiler must do flow analysis and allow a
function body that couldn't possibly flowing off the end?  How
complicated does the analysis have to be?  For example, does a
compiler have to allow:

 int g = 0;

 int foo() {
    switch (g) {
    case 4:
       break;
    default:
       return 7;
    };
    if (g==4)    // yeah, I know this is silly
       return 7;
 }

Does it make any difference if the function returns a simple type as
opposed to a class object with a constructor/destructor?

If a compiler doesn't *have* to give an error for cases that would
allow flowing off the end, *may* it?  For example, may a (future)
ANSI/ISO C++ -conforming compiler refuse to compile:

 string foo(int i) {
    if (i<kMax) return "hello";
    abort();
 }

Thanks,
Jamshid Afshar
jamshid@ses.com