Often having everything in one compilation unit results in a smaller and/or faster binary. I’ve seen this happen even when using LTO. It’s a common practice when developing firmware for small micros.
It's stupid. It messes up `using`. It doesn't do incremental rebuilds. But it does clean builds fast for the release script, and it saves me from setting up cross-compilation on a faster computer.
It's also a really great idea, for the reasons you give. This is how the Firefox tree is built, and it gives massive compile and link time speedups. (We call it "unified builds".)
Some runtime speedups too. I guess that's from better cross-file inlining?
Another highlight, no main loop. Everything was in an interrupt. This allowed the code to run despite a very broken i2c driver because it was given a low enough priority that other tasks could preempt it while it was hanging indefinitely.
Most of the manufacturers eventually just wrote gcc backends for their architectures, but the era of crappy proprietary compilers went on far too long in many cases.