[mpweiher]

Maybe I should call the backend for Objective-S Fran :-)

> graph coloring register allocation

I knew some people working on linear optimization, at the time one of the most performance-intensive applications around (with real money/competitiveness in many industries riding on it). The compiler that produced the best code, by quite a bit, was the IBM FORTRAN compiler (3090 was the preferred target at the time), which also didn't do graph coloring. It just allocated the registers by loops from innermost to outermost.

¯\_(ツ)_/¯\

For a simple compiler, we should also look at Wirth's works, and Wirth's rule: an optimization has to pay for itself, that is, the compiler compiling itself with the optimization both in the executable and in the source code has to be faster than without it in both places.

[scott_s]

> For a simple compiler, we should also look at Wirth's works, and Wirth's rule: an optimization has to pay for itself, that is, the compiler compiling itself with the optimization both in the executable and in the source code has to be faster than without it in both places.

That will result in a compiler best at optimizing programs that behave like compilers. Which may be fine for simple compilers, but it's worth keeping in mind that will not be a good compiler for domains where the programs don't behave like compilers (such as scientific computing, or even ML these days) that also want high performance.

[MrBuddyCasino]

So that is were this Golang rule comes from.