In 99% of cases you start out with an unoptimized version, find the areas that need optimizing, and improve the automated output by hand.
This strongly depends on the scenario.
If people are identifying and optimizing a hot spot, it tends to happen this way, yes.
But what I am thinking about is the common situation where a company produces a product in a low level language, and then later on a competitor produces a competing product in a high level language. The second entrant is optimized by an automated process, the first was done by hand, and can't easily be rewritten.
In these situations - and I have encountered several - the second implementation frequently winds up doing more and doing it faster than the first implementation.
This strongly depends on the scenario.
If people are identifying and optimizing a hot spot, it tends to happen this way, yes.
But what I am thinking about is the common situation where a company produces a product in a low level language, and then later on a competitor produces a competing product in a high level language. The second entrant is optimized by an automated process, the first was done by hand, and can't easily be rewritten.
In these situations - and I have encountered several - the second implementation frequently winds up doing more and doing it faster than the first implementation.