Re: [V2] Adding new CRC32C implementation for IBM S390X

On Tue, Jul 8, 2025 at 6:46 PM Eduard Stefes <Eduard(dot)Stefes(at)ibm(dot)com> wrote:
>
> Hi,
>
> here is V3 of the patch. Changes from V2:
>
> - removed IBM copyright
> - removed GETAUXVAL check in favor of the already provided check
> - moved runtime selection code from pg_crc32c_s390x_choose.c to
> pg_crc32c_s390x.c and removed _choose.c file
> - removed the config time compiler check and let the buildsystem fall
> back to sb8
> - changed buffer limit back to 32 bytes before calling s390x specific
> implementation

Hi Eduard, I look a brief look at v3 and it seems mostly okay at a
glance. There is just one major thing that got left out:

On Wed, Jul 2, 2025 at 3:27 PM Eduard Stefes <Eduard(dot)Stefes(at)ibm(dot)com> wrote:
> On Wed, 2025-06-11 at 13:48 +0700, John Naylor wrote:
> > As I alluded to before, I'm not in favor of having both direct-call
> > and runtime-check paths here. The reason x86 and Arm have it is
> > because their hardware support works on any length input. Is there
> > any
> > reason to expect auxv.h to be unavailable?
>
> I tried to find a reason but did not find any. So I'll remove it.

v3 still has direct-call and runtime-check paths. Let's keep only
USE_S390X_CRC32C_WITH_RUNTIME_CHECK and discard the direct call
configure checks. Once that's done I'll take a closer look and test as
well. The rest should be small details.

--
John Naylor
Amazon Web Services