Re: [tree-ssa] Implement non-call exceptions

[Richard Henderson <rth at redhat dot com>]

On Thu, Dec 04, 2003 at 11:30:11AM +0000, Andrew Haley wrote:
> Hmm.  Why do we need to treat instructions that can fault, but cannot
> throw, in a special way?

You can't move an instruction that may trap to a place that isn't
protected by a previous conditional.  Hmm, that's awkward.

	if (test)
	  g = a+b+c;

=>

	t1 = a + b;
	if (test)
	  {
	    t2 = t1 + c;
	    g = t2;
	  }

might be a win if test is expected to be true, since the latency
for the addition is partially shared with the test and branch.
But if that's FP, and we're not using whichever -ffast-math subset
supressed concern about traps, then the transformed code may trap
when the original won't.


r~