Which build has the best decoding performance?

All builds have in common that the GCC compiler is used for building libavcodec, which is ffdshow's main library for decoding. This library contains a lot of hand-optimized assembly code (which auto-detects the CPU's supported feature and instruction sets). Using additional compiler optimizations or advanced instruction sets (such as SSE/SSE2) doesn't result in any significant or noticeable performance increases. So when comparing pure decoding performance, there is almost no difference between the builds.