Re: Libjava failures status

[Chris Lattner <sabre at nondot dot org>]

> > So, this method must throw an ArithmeticException, even though i2 is
> > never used (for that matter, i1 is only used in generating the throw),
> > because every reachable expression along a control flow path must be
> > evaluated:

Consider the code with explicit checks put into it:

void m() {
  int i1 = 0;
  int i2 = i1 / i1;
}

becomes:

void m() {
  int i1 = 0;
  if (i1 == 0) throw ...divide by zero...;
  int i2 = i1 / i1;
}

In this case, yes the division can be removed, and furthermore, the
function should be compiled to a simple:

void m() {
  throw ...divide by zero...;
}

Making the checks explicit decouples the divide from the check implicit in
the instruction.

> Every _reachable_?  Insane, I'm sure Sun specified this to make really
> sure, that no highly optimizing compiler for java is implementable :/

No, it would be perfectly reasonable to represent this (safely) with
explicit checks, which would not break any optimizations... Why does GCC
have the notion of a trapping and nontrapping instruction?  Perhaps it is
not the best approach for GCJ?

-Chris