[Bug tree-optimization/58143] [4.8/4.9 regression] wrong code at -O3

[rguenther at suse dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58143

--- Comment #25 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 15 Oct 2013, bernd.edlinger at hotmail dot de wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58143
> 
> --- Comment #24 from Bernd Edlinger <bernd.edlinger at hotmail dot de> ---
> Wall,
> 
> the logic behind the loop niter is really strange
> and the results are simply insane.
> 
> When you consider this example:
> 
> int a, b, c, d, e;
> 
> int
> main ()
> {
>   for (b = 4; b > -30; b--)
>     {
>     e = 2147483647 - b;
>     for (; c;)
>       for (;;)
>         {
>           if (d)
>             break;
>         }
>     }
>   return 0;
> }
> 
> then the reasoning is that an "undefined" value will
> be written to e in loop iteration #4.
> And therefore it does not matter what happens afterwards.
> But for the result of main() which will ultimately be 0
> the undefined value of e does not matter.
> 
> Why should a undefined value that is not used for anything
> create such a problem?

Because undefined behavior includes random effects (like
formatting your hard drive).  In this case the undefined
behavior is simply assuming we never reach loop iteration #4
and thus the loop exit test is never true.

Richard.