[LLVMdev] Advice on CFG pre-analysis

[Joachim Durchholz]

Am Dienstag, den 27.05.2008, 20:41 -0700 schrieb Chris Lattner:
> On May 27, 2008, at 1:57 PM, Mike Stump wrote:
> 
> > On May 23, 2008, at 11:53 PM, Talin wrote:
> >> In the language I am working on, there are cases where the call flow
> >> graph can affect the type of a variable.
> >
> > :-)
> >
> > This reminds me of people that want to use CFG and the optimizer to
> > make:
> >
> > int i; int() { if (i) return 1; else return 0; }
> >
> > not warn/error that flow falls off the end.

Java does exactly this.

> > Extend that out and you have to do arbitrarily hard things and the
> > document the limit of what was done, which is messy and wrong

Correct in general, but not an issue if you draw a clear line.

In Java's case, the line is that the inference uses just structural
control flow information.
This is clearly spelled out in the language reference.

> [...] any  
> compiler that emits a false positive for that would be pretty useless  
> in practice.

I'd agree that such a language is more restrictive than necessary, but I
wouldn't say it is useless.
You can always rewrite examples like the above as

  int result;
  if (i) result = 1; else result = 0;
  return result;

If the resulting ugliness is a factor, the computation can be moved to a
function and we get

  return int_to_01 (i)

which even gives us an opportunity to name that piece of code.
(Returns that are buried in multiple levels of nesting aren't too
beautiful either...)

Regards,
Jo