Re: gcc-12 miscompilation

[Jerry James]

On Sat, Dec 24, 2022 at 3:27 PM Camm Maguire <camm@maguirefamily.org> wrote:
> Greetings, and thank you so much for pinning down the offending
> optimization flag, reducing the bug report, and forwarding the response!
>
> I'm not convinced (yet at least) that this code being classified as
> 'undefined' means gcc is not in error.  Signed and unsigned integer
> arithmetic of any precision is defined in C to loop around extremal
> values, i.e. to act as modular operations.  This code acknowledges this
> and relies upon it -- runtime 'overflows' are not a mistake.  Perhaps
> there is some better syntax to make this clear?

My understanding is that *unsigned* integer overflow is well-defined,
and simply wraps around as you describe.  *Signed* integer overflow,
on the other hand, is undefined behavior.  This article supports that
view: https://en.wikipedia.org/wiki/Integer_overflow (see the section
on "Inconsistent behavior").  This also means that, while
-fno-tree-vrp may cause the symptoms to disappear, the correct flag to
use is -fwrapv.  Better would be to make sure that overflow only
happens for unsigned integer operations.

Regards,
-- 
Jerry James
http://www.jamezone.org/