Re: [Qemu-devel] [PATCH v2 0/7] TCG global variables clean-up

[Evgeny Voevodin <e.voevodin@samsung.com>]

On 10/27/2012 06:34 PM, Blue Swirl wrote:
> On Fri, Oct 26, 2012 at 6:32 AM, Evgeny Voevodin <e.voevodin@samsung.com> wrote:
>> Today I made more precise testing with usage of --enable-profiler.
>>
>> Here is the test procedure:
>> 1. Boot Linux Kernel 5 times.
>> 2. For each iteration wait while "JIT cycles" is stable for ~10 seconds
>> 3. Write down the "cycles/op"
>>
>> Here are the results:
>>
>> Before clean-up:
>> min: 731.9
>> max: 735.8
>> avg: 734.3
>> standard deviation: ~2 = 0.3%
>> Avarage cycles/op = 734 +- 2
>>
>> After clean-up:
>> min: 747.2
>> max: 751.7
>> avg: 750.5
>> standard deviation: ~2 = 0.3%
>> Avarage cycles/op = 750 +- 2
>> Slow-down of TCG code generation = 2.2%
>>
>>
>> After clean-up with TCGContext *const tcg_cur_ctx:
>> min: 730.6
>> max: 733.2
>> avg: 728.7
>> standard deviation: ~2 = 0.3%
>> Avarage cycles/op = 729 +- 2
>> Slow-down of TCG code generation = 0%
>>
>> I suggest to define tcg_cur_ctx as TCGContext *const.
>> Then we will get rid of TCG code generation slow-down and also
>> will have no usage of global variables.
> How does this compare with the original version without pointers? I
> think that version may be safer to be assumed to be optimized by the
> compiler.

I did more testing with different gcc versions and different patch series:

gcc verion   v1 clean-up, no pointer       v2 clean-up, const pointer   
    master
gcc-4.4        754.3                                   752.1 
                                   769.8
gcc-4.5        770.8                                   779.8 
                                   774.8
gcc-4.6        731.8                                   729.8 
                                   737

Conclusion:
  - First clean-up series without pointer operates faster than master in 
all cases. It's probably because
    data is cached more efficiently.
  - Second clean-up series with constant pointer operates faster than 
master in the case of gcc-4.4 and gcc-4.6.
    In the case of gcc-4.5 it seems that const pointer is not optimised 
as I assumed.

I think that it's worth to generate third series without pointer and 
with code clean-up included in second.

How do you think?

>>
>> On 10/25/2012 10:45 AM, Evgeny Voevodin wrote:
>>> Here are the results of tests before and after this patch series was
>>> applied:
>>>
>>> * EEMBC CoreMark (before -> after)
>>>     - Guest: Exynos4210 ARMv7, Linux (Custom buildroot image)
>>>     - Host: Intel(R) Core(TM) i5 CPU 750 @ 2.67GHz, 4GB RAM, Linux
>>>     - Results: 1148.105626 -> 1161.440186 (+1.16%)
>>>
>>> * nbench (before -> after)
>>>     - Guest: Exynos4210 ARMv7, Linux (Custom buildroot image)
>>>     - Host: Intel(R) Core(TM) i5 CPU 750 @ 2.67GHz, 4GB RAM, Linux
>>>     - Results
>>>       . MEMORY INDEX: 1.864 -> 1.862 (-0.11%)
>>>       . INTEGER INDEX: 2.518 -> 2.523 (+0.2%)
>>>       . FLOATING-POINT INDEX: 0.385 -> 0.394 (+2.34%)
>>>
>>>
>>> Those tests show that it became even faster :))
>>>
>>> But I'm quite sceptical about such results.
>>> The thing is that in case of nbench it prints the warning if results are
>>> not 95% statistically accurate.
>>> So we can be sure that nbench result is 95% accurate.
>>> And it's obvious that result shown above are in the scope of this
>>> accuracy.
>>> I don't know the accuracy of CoreMark.
>>>
>>> So, the main decision we can make that this patch series didn't
>>> introduce any slow-down comparable to inaccuracy of the measurement.
>>>
>>> Is this enough?
>>>
>>> On 10/23/2012 10:21 AM, Evgeny Voevodin wrote:
>>>> This set of patches moves global variables to tcg_ctx:
>>>> gen_opc_ptr
>>>> gen_opparam_ptr
>>>> gen_opc_buf
>>>> gen_opparam_buf
>>>>
>>>> Build tested for all targets.
>>>> Execution tested on ARM.
>>>>
>>>> I didn't notice any slow-down of kernel boot after this set was applied.
>>>>
>>>> Changelog:
>>>> v1->v2:
>>>> Introduced TCGContext *tcg_cur_ctx global to use in those places where
>>>> we don't
>>>> have an interface to pass pointer to tcg_ctx.
>>>> Code style clean-up
>>>>
>>>> Evgeny (2):
>>>>     tcg/tcg.h: Duplicate global TCG variables in TCGContext
>>>>     TCG: Remove unused global variables
>>>>
>>>> Evgeny Voevodin (5):
>>>>     translate-all.c: Introduce TCGContext *tcg_cur_ctx
>>>>     TCG: Use gen_opc_ptr from context instead of global variable.
>>>>     TCG: Use gen_opparam_ptr from context instead of global variable.
>>>>     TCG: Use gen_opc_buf from context instead of global variable.
>>>>     TCG: Use gen_opparam_buf from context instead of global variable.
>>>>
>>>>    gen-icount.h                  |    2 +-
>>>>    target-alpha/translate.c      |   10 +-
>>>>    target-arm/translate.c        |   10 +-
>>>>    target-cris/translate.c       |   13 +-
>>>>    target-i386/translate.c       |   10 +-
>>>>    target-lm32/translate.c       |   13 +-
>>>>    target-m68k/translate.c       |   10 +-
>>>>    target-microblaze/translate.c |   13 +-
>>>>    target-mips/translate.c       |   11 +-
>>>>    target-openrisc/translate.c   |   13 +-
>>>>    target-ppc/translate.c        |   11 +-
>>>>    target-s390x/translate.c      |   11 +-
>>>>    target-sh4/translate.c        |   10 +-
>>>>    target-sparc/translate.c      |   10 +-
>>>>    target-unicore32/translate.c  |   10 +-
>>>>    target-xtensa/translate.c     |    8 +-
>>>>    tcg/optimize.c                |   62 ++++----
>>>>    tcg/tcg-op.h                  |  324
>>>> ++++++++++++++++++++---------------------
>>>>    tcg/tcg.c                     |   85 ++++++-----
>>>>    tcg/tcg.h                     |   11 +-
>>>>    translate-all.c               |    4 +-
>>>>    21 files changed, 328 insertions(+), 323 deletions(-)
>>>>
>>>
>>
>> --
>> Kind regards,
>> Evgeny Voevodin,
>> Technical Leader,
>> Mobile Group,
>> Samsung Moscow Research Center,
>> e-mail: e.voevodin@samsung.com


-- 
Kind regards,
Evgeny Voevodin,
Technical Leader,
Mobile Group,
Samsung Moscow Research Center,
e-mail: e.voevodin@samsung.com