[Python-Dev] gcc 4.2 exposes signed integer overflows

[Daniel Berlin]

Jack Howarth wrote:
> Guido,
>     You'll never win that argument with the gcc developers. If you
> rely on undefined behavior in the c language standard, they have
> in the past, and will continue to, feel free to ignore those cases.
> If you plan on ignoring this issue, just be prepared to see a 
> testcase failure in the python testsuite when python is built with
> gcc 4.2.

In addition, i'm surprised that you don't see these failures on other
compilers.  I know XLC, for example, would treat signed integer overflow
as undefined at all opt levels as well.
I would not be surprised to find Intel's compiler doing the same thing.

Speaking as a gcc optimization person, if we were to treat signed
integer overflow as wrapping, you would inhibit a very large number of
loop optimizations on a very large class of loops (due to a number of
reasons, from no longer being able to prove termination of loops or
estimation of number of iterations, to other things).

Heck, XLC at -O3+ will ignore *unsigned* integer wraparound of loop
indices unless you use -qstrict_induction.
As the docs for this option say, using that option can cause severe
performance degradation (I happen to know the numbers from when i worked
at IBM, and they are surprisingly high).

GCC is willing to take the performance hit relative to other compilers
on things like that to be a standard conforming compiler, just like we
are willing to optimize code based on the assumption that you do not
invoke undefined behavior.

--Dan