At least the xlc version on AIX 7.1 has aggressive optimizations even
with -O:

==================================================================== 
#include <stdio.h>
#include <stdlib.h>


static int
f(long *a, long long *b)
{
     int t = *a;
     *b = 0;          // cannot change *a
     return *a - t;   // can be folded to zero
}

int
main(void)
{
    long a = 10;

    printf("%d\n", f(&a, (long long *)&a));

    return 0;
}
====================================================================



$ xlc -O -o alias alias.c
$ ./alias
0
$ 
$ xlc -O -qalias=noansi -o alias alias.c
$ ./alias
-10

We would also need -fwrapv to be safe, but I cannot find an equivalent
option for xlc.

The Intel compiler had a similar failure to #40244 that was resolved
by -fwrapv (see #40223).

-qmaxmem=-1 is added to disable verbose remarks.

New changeset 84a7917b4c9afec07575065cffa143b91fe98c14 by Stefan Krah in branch 'master':
bpo-41721: Add xlc options (GH-22096)
https://github.com/python/cpython/commit/84a7917b4c9afec07575065cffa143b91fe98c14

The code example is for 64-bit Linux with sizeof(long) == sizeof(long long).

It also works on 32-bit xlc with int/long.

New changeset fb050d0d60f38dc9b6c30df1864020a92981be5b by Miss Islington (bot) in branch '3.9':
bpo-41721: Add xlc options (GH-22097)
https://github.com/python/cpython/commit/fb050d0d60f38dc9b6c30df1864020a92981be5b