gfortran or tree-ssa bug?

[Paul Brook]

> I assumed that gfortran implicitly set the target
> to 0 due to the underflow.  At least, my test showed
> that it did.

Nope. The frontend simplification routines assume any error means "I can't 
touch this, so use the original value". The subtraction gets folded by the 
backend folders later on, but we need to handle this in the frontend.

The output of -fdump-parse-tree shows the subtraction is still there.

I'm basically suggesting we remove ARITH_UNDERFLOW altogether.

Paul