Created attachment 823 [details]
SDL_blit_N.patch

There is only one Altivec accelerated blit function (ConvertAltivec32to32_prefetch() or ConvertAltivec32to32_noprefetch(), depending on the CPU used) that is supposed to handle all alpha combinations. This works as follows for every pixel line:
1. Blit single pixels until an aligned address is reached
2. Accelerated blit as far as possible
3. Blit single remaining pixels
Part 2. is set up correctly to handle different combinations of the alpha channels of the participating surfaces. Parts 1. and 3. only do a simple copy of all the pixel's components from souce to destination. But when the source surface has no alpha channel (Amask is 0, e.g. the video surface) the surface's alpha value must be used instead. Otherwise crap (uninitialized data) is being copied to the destiniation's alpha channel.

The attached patch is a quick'n'dirty solution to the problem. A more sophisticated solution might require separate functions for different combinations of the alpha channels of the participating surfaces.

The problem seems not to come up when the width of the destination surface enforces all pixel lines to start at an aligned address. This can easily be accomplished by using a suitable pitch value for all surfaces, but SDL_CreateRGBSurface doesn't make use of this right now.