Schol-R-LEA wrote:
Actually, in terms of the time spent, the dominant part of most compilers is the lexical analyzer - which in most cases, amounts to a Deterministic Finite State Automaton, whether formally constructed or as an ad-hoc equivalent. Either way, I would have a lot of trouble seeing that as a graph coloring problem.
That's only true if you either use an existing backend or don't care about optimization. Lexical analysis is solved and there are so many algorithms to do it easily: recursive descend, PEG, LR parsers generators, or even
Earley parsers that can handle any CFG.
mariuszp wrote:
It's literally just lots of graph problems; graph coloring, vertex reachability, etc. Prove me wrong.
A LLVM-style Greedy RA is superior to graph coloring RAs in practice, prove me wrong. Graph coloring just cannot easily handle non-homogeneuos register files (and of course, it's also quite expensive). When I wrote an
implementation (only a bit over 1k SLOC, can do live range splitting but not register spilling right now) inspired by the LLVM one, I was surprised (i) how natural it is, and (ii) how good the results are if you tune it a bit (for smallish functions with only a few basic blocks/variables, the result is often optimal). It's incredibly easy to tune -- when you see a missed optimization, it's usually straightforward to see how to extend the RA to handle it.