[derefr]

A 30-year-old compiler would still have come out in 1990, well after SSA-based optimizations were discovered. But maybe I'm expecting too much in thinking that the author of a flagship optimizing compiler should be up-to-date on CS research in compiler optimization, before deciding the architecture of their compiler.

(Also, plenty of the compilers and/or JITs that I'm talking about are far newer. The first attempt to get a JVM to use SSA optimization during JIT — within SafeTSA — only occurred in 2000. Such an approach was copied by pretty much every JVM implementation by 2005, suggesting that a legacy of incompatible architecture was never the problem, but rather that JVM implementors just didn't think it was a worthwhile technique until they saw it demonstrated for their particular workloads.)

CPS is closely theoretically related to SSA (it carries equivalent information across lexical boundaries) but CPS isn't a basis for optimization transforms in the same way that SSA is. You can't hoist loop invariants using CPS... as far as I know, at least.

[titzer]

> But maybe I'm expecting too much in thinking that the author of a flagship optimizing compiler should be up-to-date on CS research in compiler optimization, before deciding the architecture of their compiler.

In 1990 I don't think SSA was broadly considered to be the basis of a good optimizing compiler the way it is today. So the "author of a flagship optimizing compiler" would be gambling on an unknown that looked good in a couple new papers. The Cytron paper on how to efficiently compute it came out in 1991. So SSA was still a WIP.

It wasn't really until the late 1990s IMO, that SSA became widespread and into the 2000s when it became the standard.

[aidenn0]

Someone else already replied to your first paragraph, so I'll reply to your second:

The safe TSA papers from 2000-2001 were the papers that showed that SSA was practical in a JIT environment. The fact that this approach was adopted less than 5 years after those papers come out suggest that compiler authors are in fact reading research papers.

Rearchitecting a production JIT to use a different IR for theoretical gains when nobody's demonstrated those gains on a toy JIT is high risk and demonstrating those gains on a toy JIT is research, not development.

[p1necone]

I think a lot of now big, "professional" projects started as people's non serious hobby projects.

If you just want to write a compiler for fun, I can understand not reading up on state of the art theory beforehand. And then it's accidentally actually useful to other people and it's too far into development to change the fundamental architecture.