Performance hit in qemu-system-ppc

Performance hit in qemu-system-ppc

[Howard Spoelstra]

[-- Attachment #1: Type: text/plain, Size: 1452 bytes --]

 Hi,

I noticed a considerable (~20%) slowdown in the cpu performance of
qemu-system-ppc.
Bisecting led me to this commit:

d03f140804b345a85973976506492027f703d82d is the first bad commit
commit d03f140804b345a85973976506492027f703d82d
Author: Richard Henderson <richard.henderson@linaro.org>
Date:   Mon Dec 9 13:49:58 2019 -0800

    cputlb: Move body of cpu_ldst_template.h out of line

    With the tracing hooks, the inline functions are no longer
    so simple.  Once out-of-line, the current tlb_entry lookup
    is redundant with the one in the main load/store_helper.

    This also begins the introduction of a new target facing
    interface, with suffix *_mmuidx_ra.  This is not yet
    official because the interface is not done for user-only.

    Use abi_ptr instead of target_ulong in preparation for
    user-only; the two types are identical for softmmu.

    What remains in cpu_ldst_template.h are the expansions
    for _code, _data, and MMU_MODE<N>_SUFFIX.

    Tested-by: Philippe Mathieu-Daudé <philmd@redhat.com>
    Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

 accel/tcg/cputlb.c               | 116 ++++++++++++++++++++++++++++++++++++
 include/exec/cpu_ldst.h          |  25 +++++++-
 include/exec/cpu_ldst_template.h | 125
++++++++-------------------------------
 3 files changed, 166 insertions(+), 100 deletions(-)

Thanks for looking into this issue,
Howard

[-- Attachment #2: Type: text/html, Size: 1928 bytes --]

Re: Performance hit in qemu-system-ppc

[BALATON Zoltan]

[-- Attachment #1: Type: text/plain, Size: 1924 bytes --]

Hello,

cc-ing the ppc list, maintainer and original author of the patch could 
increase the chances that they actually see your message which they might 
miss on qemu-devel. David explicitely said he usually cannot read all 
mails on qemu-devel so likely should be cc-d on all PPC issues (or also cc 
the qemu-ppc list besides qemu-devel).

Regards,
BALATON Zoltan

On Sun, 26 Jan 2020, Howard Spoelstra wrote:
> Hi,
>
> I noticed a considerable (~20%) slowdown in the cpu performance of
> qemu-system-ppc.
> Bisecting led me to this commit:
>
> d03f140804b345a85973976506492027f703d82d is the first bad commit
> commit d03f140804b345a85973976506492027f703d82d
> Author: Richard Henderson <richard.henderson@linaro.org>
> Date:   Mon Dec 9 13:49:58 2019 -0800
>
>    cputlb: Move body of cpu_ldst_template.h out of line
>
>    With the tracing hooks, the inline functions are no longer
>    so simple.  Once out-of-line, the current tlb_entry lookup
>    is redundant with the one in the main load/store_helper.
>
>    This also begins the introduction of a new target facing
>    interface, with suffix *_mmuidx_ra.  This is not yet
>    official because the interface is not done for user-only.
>
>    Use abi_ptr instead of target_ulong in preparation for
>    user-only; the two types are identical for softmmu.
>
>    What remains in cpu_ldst_template.h are the expansions
>    for _code, _data, and MMU_MODE<N>_SUFFIX.
>
>    Tested-by: Philippe Mathieu-Daudé <philmd@redhat.com>
>    Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
>
> accel/tcg/cputlb.c               | 116 ++++++++++++++++++++++++++++++++++++
> include/exec/cpu_ldst.h          |  25 +++++++-
> include/exec/cpu_ldst_template.h | 125
> ++++++++-------------------------------
> 3 files changed, 166 insertions(+), 100 deletions(-)
>
> Thanks for looking into this issue,
> Howard
>

Re: Performance hit in qemu-system-ppc

[Richard Henderson]

On 1/25/20 3:30 PM, Howard Spoelstra wrote:
> I noticed a considerable (~20%) slowdown in the cpu performance of qemu-system-ppc.

ENOINFO.

For what test case?  This should not have been on any hot path.


r~

Re: Performance hit in qemu-system-ppc

[Howard Spoelstra]

[-- Attachment #1: Type: text/plain, Size: 850 bytes --]

On Tue, Jan 28, 2020 at 7:30 PM Richard Henderson <
richard.henderson@linaro.org> wrote:

> On 1/25/20 3:30 PM, Howard Spoelstra wrote:
> > I noticed a considerable (~20%) slowdown in the cpu performance of
> qemu-system-ppc.
>
> ENOINFO.
>
> For what test case?  This should not have been on any hot path.
>
>
Hi,
I tested Mac OS 9.2 and 10.3/4 with qemu-system-ppc built from 4.2 and
current master like this:

Command line:
./qemu-system-ppc -L pc-bios -boot c -M mac99,via=pmu -m 512 \
-serial stdio -sdl \
-hda 9.2.img (10.3.img|10.4.img)

Mac OS 9.2 MacBench 5.0 reports processor speed:
4.2: ~190%
Master: ~160%

Mac OSX 10.3.9 Skidmarks 4.0.1 test: specific drop in performance of
MPEG(INT) test:
4.2: 118%
Master: 100%

Mac OS 10.4.11 Skidmarks 4.0.1 test: specific drop in performance of
MPEG(INT) test:
4.2: 117%
Master: 99%

Best,
Howard

[-- Attachment #2: Type: text/html, Size: 1336 bytes --]