On 9/22/17, 8:00 AM, "linux-kernel-ow...@vger.kernel.org on behalf of Colin 
King" <linux-kernel-ow...@vger.kernel.org on behalf of 
colin.k...@canonical.com> wrote:
> From: Colin Ian King <colin.k...@canonical.com>
> 
> Don't populate const array rtbTable on the stack, instead make it
> static. Also split overly long line to clean a chechkpach warning.
> Makes the object code smaller by nearly 500 bytes:
> 
> Before:
>    text          data     bss     dec     hex filename
>   13297           104       0   13401    3459 lib/zstd/fse_compress.o
> 
> After:
>    text          data     bss     dec     hex filename
>   12742           160       0   12902    3266 lib/zstd/fse_compress.o
> 
> (gcc 6.3.0, x86-64)
> 
> Signed-off-by: Colin Ian King <colin.k...@canonical.com>

I tested your patch with gcc-7.1 on x86, and benchmarked the speed on
upstream zstd. There isn't a noticeable speed difference, since it isn't a
particularly hot piece of code. Would you be able to submit the same patch
upstream [1], or would you be okay with me porting it back upstream, so it
doesn't get lost on an update?

I didn't expect gcc to leave constant arrays on the stack, that seems
silly. Clang makes it static, but gcc loads it onto the stack, and in 6.3+
it saves the data statically, and then uses vector instructions to load it
onto the stack [2].

Tested-by: Nick Terrell <terre...@fb.com>

[1] https://github.com/facebook/zstd
[2] https://godbolt.org/g/fvTcED


Reply via email to