Yeah, I'm calling bullshit on this one. At least, it doesn't line up with my experience. In my experience, games are optimized _just enough_ for a decent playing experience (and not always then). Games devs, as a whole, are the worst offenders of expecting their users to just throw more money (hardware) at the software to achieve usable/enjoyable experiences. There are, of course, exceptions. But, for every Carmack, there's 10s of thousands of developers scrambling to make their deadline, doing just enough to ship.
The games themselves though are a different story. Outside of open sources games (which are usually less demanding than commercial ones), you don't have the source code to rebuild it. Even if you did, enabling optimizations beyond what the developer used risks breakage so you will have to do your own testing. Even then, simply rebuilding the software wouldn't address the quality of the code created by those developers who are scrambling to meet a deadline with as little effort as possible.
Back in the day, I spent a fair amount of time working on gutting the DLL because frankly, it was atrocious. My memory is a little fuzzy as it's been +10 years since I've looked at it, but things I remember seeing:
* over use of C/C++ preprocessor macros where an inline function would have been more appropriate to say, get the array/list of CvPlots (tiles) that needed to be iterated over all the time. * lack of hoisting loop invariants out of loops. It is common to see usages of the non-trivial macros above in the bodies of tight loops, often tight nested loops. Optimizing compilers are great, but they're not _that_ great. * the exposure of the C++ types to Python was done...poorly. It was done using Boost Python (which, while a great library for its day had a _huge_ learning curve). For every Cv type in the C++ DLL, there was a corresponding Cy type to expose it to Python. Collection types were _copied_ every call into Python code, which was quite frequent. The collections should have been done as proxies into the underlying C++ collections, instead of as copies.
Most of the changes I made were quite small, but over the course of a month of part-time hacking on it, I'd gotten late game turns down to a couple of minutes from 30-minutes and memory usage was extremely reduced; and I never did get around to fixing the Python wrapper as it would have too intrusive to fix it properly. I could have made more aggressive changes if I had full access to the source, but being constrained by DLL boundaries and C++ types being exported limited what could be done w/o breaking the ABI (had to be extremely careful about not changing object sizes, affecting vtable layout, etc).
Frankly, I doubt the developers spent very much time at all, if any, with a profiler during the course of development with the game.
Civ had been a 2d game until then, it was their first 3d title.
Not to mention that it was turn based strategy, and the main performance problem was AI turn length in the endgame.
Unless you run javascript 'applications'. Games are already optimized.