[_mhp_]

I have got experience of DSP assembly (mostly 320C25 stuff, a little 'C50), and I initially doubted the story. DSPs like the 'C25 were pretty easy to work with since the instructions were all the same length (ignoring memory wait states for the moment). But with the 'C50 (and the 'C30 from the article), delayed branching was introduced. Whilst you can optimise for that in your head, it becomes a pain and often results in hard-to-maintain code. Adding a couple of NOPs makes it much easier!

The article highlighted common sub-expression elimination too, which again can be done manually. But (and this is the crux IMHO) writing optimal code continuously at this level is mentally tiring (leading to bugs) and often gives brittle unmaintainable code. A compiler never needs to worry about this as it generates fresh code every time rather than maintaining the code from the last build. But a human will learn to avoid brittle code and hence won't be able to take advantage of the possible optimisations.