"Regardless, the whole point of the greedy allocator is to throw out the fancy list of active intervals that linear scan maintains in favor of doing something simpler that allows more flexibility in spilling/splitting, right?"

Kinda. It's not that simple, either :)

Linear scan is actually just bad all around, IMHO. It's meant for JITs that want to value speed above all else. At the time, I expect Chris was hopeful better spilling algorithms would take care of the fact that Linear Scan sucks.

It's not effective at solving the coloring problem quickly and accurately either :). It's kinda a worst-of-both worlds algorithms

Greedy is closer to what other compilers do: optimistic coloring, splitting if necessary, then remat/spilling.

So it's more accurate to say "LLVM made a bad choice in linear scan. They later fixed that mistake". I"m not away of any other high performance compiler that went down the linear scan path ever, because they didn't think it would work out.