[dpdk-dev] [PATCH v3] mempool: replace c memcpy code semantics with optimized rte_memcpy

[Ferruh Yigit]

On 7/6/2016 5:21 PM, Ferruh Yigit wrote:
> On 7/5/2016 3:09 PM, Ferruh Yigit wrote:
>> On 7/5/2016 2:13 PM, Jerin Jacob wrote:
>>> On Tue, Jul 05, 2016 at 07:32:46PM +0800, Yuanhan Liu wrote:
>>>> On Tue, Jul 05, 2016 at 09:43:03AM +0100, Ferruh Yigit wrote:
>>>>> On 6/30/2016 6:28 PM, Thomas Monjalon wrote:
>>>>>> 2016-06-30 17:46, Jerin Jacob:
>>>>>>> Signed-off-by: Jerin Jacob <jerin.jacob at caviumnetworks.com>
>>>>>>> Acked-by: Olivier Matz <olivier.matz at 6wind.com>
>>>>>>
>>>>>> Applied, thanks
>>>>>>
>>>>>
>>>>> Hi Jerin,
>>>>>
>>>>> This commit cause a compilation error on target i686-native-linuxapp-gcc
>>>>> with gcc6.
>>>>
>>>> Besides that, I'm more curious to know have you actually seen any
>>>> performance boost?
>>>
>>> let me first address your curiosity,
>>> http://dpdk.org/dev/patchwork/patch/12993/( check the second comment)
>>> http://dpdk.org/ml/archives/dev/2016-June/042701.html
>>>
>>> Ferruh,
>>
>> Hi Jerin,
>>
>>>
>>> I have tested on a x86 machine with gcc 6.1. I could n't see any issues
>>> with i686-native-linuxapp-gcc target
>> Thanks for investigating the issue.
>>
>>>
>>> Steps following to create gcc 6.1 toolchain
>>> https://sahas.ra.naman.ms/2016/05/31/building-gcc6-1-on-fedora-23/
>>> (removed --disable-multilib to have support for -m32)
>>>
>>> ➜ [dpdk-master] $ gcc -v
>>> Using built-in specs.
>>> COLLECT_GCC=gcc
>>> COLLECT_LTO_WRAPPER=/opt/gcc-6.1.0/libexec/gcc/x86_64-pc-linux-gnu/6.1.0/lto-wrapper
>>> Target: x86_64-pc-linux-gnu
>>> Configured with: ../gcc-6.1.0/configure --prefix=/opt/gcc-6.1.0
>>> --enable-languages=c,c++ --enable-libmudflap --with-system-zlib
>>> Thread model: posix
>>> gcc version 6.1.0 (GCC)
>> I am using Fedora24, which has gcc6 (6.1.1) as default.
>>
>>>
>>> More over this issue seems like an issue from x86 rte_memcpy implementation.
>> You are right. But i686 compilation starts failing with this commit.
>> And reverting this commit in the current HEAD solves the compilation
>> problem.
>> I am not really clear about reason of the compilation error.
> 
> The compile error is because compiler is so smart now and at the same
> time not enough smart.
> 
> Call stack is as following:
> 
> virtio_xmit_pkts_simple
>   virtio_xmit_cleanup
>     rte_mempool_put_bulk
>       rte_mempool_generic_put
>         __mempool_generic_put
> 	  rte_memcpy
> 
> The array used as source buffer in virtio_xmit_cleanup (free) is a
> pointer array with 32 elements, in 32bit this makes 128 bytes.
> 
> in rte_memcpy() implementation, there a code piece as following:
> if (size > 256) {
>     rte_move128(...);
>     rte_move128(...); <--- [1]
>     ....
> }
> 
> The compiler traces the array all through the call stack and knows the
> size of array is 128 and generates a warning on above [1] which tries to
> access beyond byte 128.
> But unfortunately it ignores the "(size > 256)" check.
> 
> In 64bit, this warning is not generated because array size becomes 256
> bytes.
> 
> So this warning is a false positive. Although I am working on it, if
> anybody has a suggestion to prevent warning, quite welcome. At worst I
> will disable this compiler warning for the file.

I have sent a patch:
http://dpdk.org/ml/archives/dev/2016-July/043492.html

Giving a hint to compiler that variable "size" is related to the size of
the source buffer fixes compiler warning.

- This update is in virtio fast path, can you please review it from
point of performance effect.

- Isn't this surprisingly smart of compiler, or am I missing something J

Thanks,
ferruh