Please decide for _flag access on 64-bit machine, if 32-bit access is better or not.
If 64-bit access is better, update version of {Diff1255} is needed, instead.
unsigned long is not good for LLP64 system. Use unsigned long long for 64-bit system.

@gniibe thanks for figuring this out. This is rather mundane, I'm somewhat astonished that I did not think of that :(.

Do you know why this is only sometimes causing problems? I have never once seen a crash on my s390 emulator, yet @dkg seems to have no problem reproducing it.

Wouldn't a nicer and simpler fix be to use a machine-word-sized integer for _flag? Do you see any downside to that?

I think that there are two archs: s390 and s390x. Latter is 64-bit and supports 32-bit version as well.
Use of machine word size (32-bit for 32-bit machine, 64-bit for 32-bit machine) is good. That will be update of Diff 1255.
But I don't know how it is achieved easily.
(If we can ignore LLP64, we can use unsigned long.)
Please go ahead that way.

In D421#2968, @gniibe wrote:

I think that there are two archs: s390 and s390x. Latter is 64-bit and supports 32-bit version as well.

I believe I have an emulated s390x:

$ uname -a
Linux s390 4.9.0-2-s390x #1 SMP Debian 4.9.13-1 (2017-02-27) s390x GNU/Linux

(gdb) print sizeof (void *)
$1 = 8
(gdb) print sizeof (int)
$2 = 4

Use of machine word size (32-bit for 32-bit machine, 64-bit for 32-bit machine) is good. That will be update of Diff 1255.
But I don't know how it is achieved easily.
(If we can ignore LLP64, we can use unsigned long.)

uintptr_t should do the trick.

I understand your emulator is s390x. Perhaps, on the emulator, memory layout is different?
I now see that C99 is OK for GnuPG, or at least no problem for gpgscm.