Re: Move cmpi to decodetree causing regression

[David Gibson]

On Fri, Jun 18, 2021 at 02:46:59AM +0200, BALATON Zoltan wrote:
> Hello,

CCing Richard Henderson, who's the person who most needs to see this,
I think.

> Commit 8f0a4b6a9 (target/ppc: Move cmp/cmpi/cmpl/cmpli to decodetree) breaks
> AROS on sam460ex (see http://zero.eik.bme.hu/~balaton/qemu/amiga/#aros ) as
> it seems to have a cmpi instruction with the L bit set that real hardware
> apparently ignores. Now I'm getting an exception (can be seen on serial when
> selecting Debug option from AROS boot menu):
> 
> [KRN] Exception 6 handler. Context @ ff7fb628, SysBase @ 00a60420, KernelBase 
> @ 0100017c
> [KRN] SRR0=0124b6c0, SRR1=000af000 DEAR=00000000 ESR=08000000
> [KRN] CTR=ff850a24 LR=0124c3dc XER=00000007 CCR=44022844
> [KRN] DAR=00000000 DSISR=00000000
> [KRN] GPR00=0124f9ec GPR01=012e0f40 GPR02=00a60420 GPR03=0123bd80
> [KRN] GPR04=00000001 GPR05=00000000 GPR06=012e0f33 GPR07=00000008
> [KRN] GPR08=0123beb0 GPR09=0123274c GPR10=00000004 GPR11=012aa7d8
> [KRN] GPR12=01230000 GPR13=00000000 GPR14=0000001e GPR15=01230000
> [KRN] GPR16=01232560 GPR17=0123272c GPR18=00000000 GPR19=80000a55
> [KRN] GPR20=0123df2c GPR21=0123df3c GPR22=0123df34 GPR23=01232778
> [KRN] GPR24=0123de6c GPR25=010f9998 GPR26=01230000 GPR27=00000000
> [KRN] GPR28=0105e0b0 GPR29=00000000 GPR30=00000001 GPR31=0123bc60
> [KRN] Instruction dump:
> [KRN] 0124b6c0: 2c240000 40820008 38800001 80230004
> 
> This -----------^^^^^^^^ seems to be cmpi with L bit set.
> 
> The commit message mentions this and we've found before that some guests
> rely on invalid bits being ignored by real hardware, such as 27a3ea7e and
> 0123d3cb. I think IBM cores generally ignore reserved bits while some other
> implementations (e.g. from Motorola) may raise exceptions. It would be
> better to do what hardware does so guest code running on real hardware would
> also run on QEMU, then have an option for enforcing reserved bits if this is
> useful for debugging (but just logging with -d guest_errors might be enough
> for that). However since e500 for example seems to raise exceptions for
> invalid bits then there should be a switch so that could be connected to a
> debug option as well.
> 
> Regards,
> BALATON Zoltan
> 

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature