Optimize some alpha blending modes.

Optimize some alpha blending modes.

- Inline the component calculation helpers.
- Reorder the computation in plus_modeproc.

This doubles the performance of plus_modeproc.

[deanm, 2009-07-03 10:40:53]

Seems like for the ARGB vs BGRA etc ordering, it would be nice to generate the
code based on the compiled color ordering.  I don't know an easy way to do that
though.

[reed, 2009-07-03 12:47:31]

I am quite surprised that adding 'inline' to the static functions helped, but
hooray.

I'm also surprised that reordering the calls to saturated_add helped, but hooray
for that too. In this case, on machines with a different order for colorpriv
32bit, the code order won't match, but I think that's ok.

LGTM

[deanm, 2009-07-03 12:53:25]

The inline keyword definitely makes a difference.

In a perfect world it wouldn't.  The compiler would realize this is a small
function and that performance is important.  However, it can't really know that
without things like profile guided optimization.

If it doesn't know it's important, it's actually a better decision to not-inline
it, as it decreases the generated code size.

On 2009/07/03 12:47:31, reed wrote:
> I am quite surprised that adding 'inline' to the static functions helped, but
> hooray.
> 
> I'm also surprised that reordering the calls to saturated_add helped, but
hooray
> for that too. In this case, on machines with a different order for colorpriv
> 32bit, the code order won't match, but I think that's ok.
> 
> LGTM