In projects like dav1d, developers intentionally bypass standard operating system calling conventions to implement custom ones. This allows for more efficient data sharing between functions by avoiding unnecessary register saving and loading, leading to significant performance gains, especially in highly optimized, frequently called code. This approach requires a deep understanding of CPU architecture and is a testament to the art of low-level programming.

Impact: High. This technique unlocks critical performance improvements in high-demand software, demonstrating that even established conventions can be creatively circumvented for speed. It highlights the ongoing relevance of assembly-level optimization.

In the source video, this keypoint occurs from 02:12:01 to 02:16:24.

Sources in support: Kieran Kunhya (FFmpeg Contributor, Developer of FFmpeg X account)