On Wed Sep 21, 2022 at 8:16 AM AEST, Segher Boessenkool wrote:
> Hi!
>
> On Tue, Sep 20, 2022 at 12:01:47AM +1000, Nicholas Piggin wrote:
> > Update the 64s GENERIC_CPU option. POWER4 support has been dropped, so
> > make that clear in the option name.
>
> AFAIR the minimum now is POWER4+ (ISA 2.01), not POWER5 (ISA 2.02).
It's POWER5 now, because of commit 471d7ff8b5 ("powerpc/64s: Remove
POWER4 support"), which is misguided about POWER4+ and also introduced
the -mcpu=power5 bug on 970 builds :\
Not sure it's worth adding POWER4+ support back but if someone has a
POWER4+ or adds it to QEMU TCG, I will do the patch.
> > -mtune= before power8 is dropped because the minimum gcc version
> > supports power8, and tuning is made consistent between big and little
> > endian.
>
> Tuning for p8 on e.g. 970 gives quite bad results. No idea if anyone
> cares, but this is a serious regression if so.
It's for "generic" kernel so we set low minimum but higher tune,
assuming that people would usually have newer, so it was already
doing -mtune=power7.
We could make a specific 970/G5 entry though, since those still
have users.
> > Big endian drops -mcpu=power4 in favour of power5. Effectively the
> > minimum compiler version means power5 was always being selected here,
> > so this should not change anything. 970 / G5 code generation does not
> > seem to have been a problem with -mcpu=power5, but it's possible we
> > should go back to power4 to be really safe.
>
> Yes, -mcpu=power5 code does *not* run on 970, if you are unlucky enough
> that the compiler does something smart with popcntb (the sole non-float
> insn new on p5, not counting hrfid).
>
> > +# -mcpu=power5 should generate 970 compatible kernel code
>
> It doesn't. Even if it did, it would need more explanation!
Okay, sounds like we should go back to power4.
Thanks,
Nick
