[patch] PR21847 (wrong-code with -fnon-call-exceptions)

[Richard Kenner]

    I don't really like this.  I'd prefer the default to be that, while
    exceptions are propagated when they occur, they're not a hard side
    effect of the operation.  I'd very much prefer that the front end
    arrange for things that shouldn't die be marked volatile or something.

How would you make an FP divide be marked volatile?

But I do agree with you: in most languages if the only purpose of a
computation would be to produce a predefined exception or trap, it need
not be computed.