Use intrinsic functions for bit counting to reduce footprint

Use intrinsic functions for bit counting to reduce footprint

The clz/bsr instructions produce the same result as the lookup table "jpeg_nbits_table" with equivalent performance. Removing the table saves 64kB. It is not much, but for mobile devices we need to save all memory that is possible.

This change is for gcc and clang compilers.

BUG=

[Olle Liljenzin, 2014-03-20 11:15:41]

Upstreaming to sourceforge:

https://sourceforge.net/p/libjpeg-turbo/patches/57/

[noel chromium, 2014-03-21 01:06:07]

Thanks Olle. Let's await news from upstream.

[Olle Liljenzin, 2014-04-07 15:11:28]

On 2014/03/21 01:06:07, noel chromium wrote:
> Thanks Olle. Let's await news from upstream.

The patch landed upstream in r1221 with intrinsic code enabled for arm, but not
for x86 since a performance regression was found on some chips. For x86 there is
another patch in r1222 that moves the lookup table from bss to ro-data to avoid
duplication. Both patches should be useful in chromium without breaking
anything.

[noel chromium, 2014-04-12 07:18:11]

Ok thanks Olle, I intend to update libjpeg_turbo to 1.3.1 stable (r1219)
to bring our turbo code revision closer to your upstream commits.

First step was https://codereview.appspot.com/87270043, more to come.

[Olle Liljenzin, 2014-05-22 15:11:29]

Created new review that cherry-picks the upstream patch:
https://codereview.appspot.com/97690043/

[noel chromium, 2014-05-23 12:22:06]

On 2014/05/22 15:11:29, Olle Liljenzin wrote:
> Created new review that cherry-picks the upstream patch:
> https://codereview.appspot.com/97690043/

Thanks. Let's continue to work there (and close this issue if you no longer need
it).