Re: [Qemu-ppc] [Qemu-devel] regression: target/ppc: convert VSX logical operations to vector operations broke X for ppc64le guest

[Mark Cave-Ayland]

On 26/02/2019 22:25, Andrew Randrianasulu wrote:

(adding qemu-ppc, Richard and David - please make sure you add the relevant
maintainer on bug reports, as otherwise due to the high volume of mails to the 
list
it's very easy to miss things)

> Hello.
> 
> I bisected this problem with fonts (and multicolored vertical stripes) in 
> qemu 
> git (ppc64-softmmu)
> 
> address@hidden:/dev/shm/qemu$ git bisect good
> 7b8fe477e12b164dda97f79e27b55b805d90384f is the first bad commit
> commit 7b8fe477e12b164dda97f79e27b55b805d90384f
> Author: Richard Henderson <address@hidden>
> Date:   Fri Feb 15 10:00:46 2019 +0000
> 
>     target/ppc: convert VSX logical operations to vector operations
> 
>     Signed-off-by: Richard Henderson <address@hidden>
>     Acked-by: David Gibson <address@hidden>
>     Message-Id: <address@hidden>
>     Signed-off-by: David Gibson <address@hidden>
> 
> :040000 040000 da3024ad2c9dfc3b7170516a8b321ef8c5d5bdf8 
> a0257b9f5880ecc8e7001a59ccaa10407084623f M      target
> 
> 
> address@hidden:/dev/shm/qemu$ git bisect log
> git bisect start
> # good: [32a1a94dd324d33578dca1dc96d7896a0244d768] Update version for v3.1.0 
> release
> git bisect good 32a1a94dd324d33578dca1dc96d7896a0244d768
> # bad: [86c7e2f4a93322a76afea5ee6806a83420d1dfea] Merge remote-tracking 
> branch 'remotes/berrange/tags/authz-core-pull-request' into staging
> git bisect bad 86c7e2f4a93322a76afea5ee6806a83420d1dfea
> # good: [95ebd99dcd37b8574426c876502bfcc7c299584b] target/arm: Decode PAuth 
> within disas_data_proc_1src
> git bisect good 95ebd99dcd37b8574426c876502bfcc7c299584b
> # good: [268dfefa690b2bdee1f8c5090d2343871cf3467c] 
> hw/microblaze/Makefile.objs: 
> Create configs for petalogix and xilinx boards
> git bisect good 268dfefa690b2bdee1f8c5090d2343871cf3467c
> # good: [f5117fd28552fe3fe32ef0495582b1caaef7a28d] hw/mips_cpc: kick a VP 
> when 
> putting it into Run statewq
> git bisect good f5117fd28552fe3fe32ef0495582b1caaef7a28d
> # bad: [2e68b8620637a4ee8c79b5724144b726af1e261b] Merge remote-tracking 
> branch 'remotes/dgibson/tags/ppc-for-4.0-20190219' into staging
> git bisect bad 2e68b8620637a4ee8c79b5724144b726af1e261b
> # good: [4c668f4a3d684ec133a52d936314379f6edd672e] target/ppc: Remove some 
> #if 
> 0'ed code
> git bisect good 4c668f4a3d684ec133a52d936314379f6edd672e
> # bad: [9b5b74da0a07a89ef71c7f7da0b36560a3bac521] target/ppc: Split out 
> VSCR_SAT 
> to a vector field
> git bisect bad 9b5b74da0a07a89ef71c7f7da0b36560a3bac521
> # good: [444d6ca301d97de141a502851940943b09a9ebee] spapr/irq: Use the 
> "simple" 
> ICS class for KVM
> git bisect good 444d6ca301d97de141a502851940943b09a9ebee
> # bad: [9bb0048ec6f8f3bcc144b2c5769d9301e824f946] target/ppc: convert xxspltw 
> to 
> vector operations
> git bisect bad 9bb0048ec6f8f3bcc144b2c5769d9301e824f946
> # good: [471ff3d0257135b938d0a5f2181f22cd753d50de] target/ppc: convert 
> vspltis[bhw] to use vector operations
> git bisect good 471ff3d0257135b938d0a5f2181f22cd753d50de
> # bad: [7b8fe477e12b164dda97f79e27b55b805d90384f] target/ppc: convert VSX 
> logical operations to vector operations
> git bisect bad 7b8fe477e12b164dda97f79e27b55b805d90384f
> # good: [0f6a6d5db853c0cbe438c1831c70710bfb6530ee] target/ppc: convert 
> vsplt[bhw] to use vector operations
> git bisect good 0f6a6d5db853c0cbe438c1831c70710bfb6530ee
> # first bad commit: [7b8fe477e12b164dda97f79e27b55b805d90384f] target/ppc: 
> convert VSX logical operations to vector operations
> address@hidden:/dev/shm/qemu$                                                
> 
> configure line:
> setarch i686 ./configure --target-list=ppc64-softmmu             
> 
> launch line:
> ppc64-softmmu/qemu-system-ppc64 -display 
> sdl,gl=on -hda /mnt/sdb1/PPC-img/alpine_disk.img
> 
> where alpine_disk.img is HDD installed Alpine 3.7 for ppc64le with xfce4 
> desktop.

AFAICT the vector instructions converted here are independent of endian, so I 
can't
see why this patch on its own would have any effect.

Maybe it could be because the "setarch i686" part forces use of some lesser-used
32-bit paths in the vector code - is this required? And do you see the same 
issue on
a x86_64 build?


ATB,

Mark.