From a maintenance perspective, it's really appealing to have a lot of small clearly defined passes so that you can have good separation of concerns. But over time, they often end up needing to interact in complex ways anyway.
For example, you might think you can do lexical identifier resolution and type checking in separate passes. But then the language gets extension methods and now it's possible for a bare identifier inside a class declaration to refer to an extension method that can only be resolved once you have some static type information available.
Or maybe you want to do error reporting separately from resolution and type checking. But in practice, a lot of errors come from resolution failure, so the resolver has to do almost all of the work to report that error, stuff that resulting data somewhere the next pass can get to it, and then the error reporter looks for that and reports it.
Instead of nice clean separate passes, you sort of end up with one big tangled pass anyway, but architected poorly.
Also, the performance cost of multiple passes is not small. This is especially true if each pass is actually converting to a new representation.
> For example, you might think you can do lexical identifier resolution and type checking in separate passes. But then the language gets extension methods and now it's possible for a bare identifier inside a class declaration to refer to an extension method that can only be resolved once you have some static type information available.
Some of this is language design though. If you make it a requirement that scope analysis can be done in isolation (so that it's parallelizable), then you design things like imports and classes so that you never have identifiers depend on types or other libraries.
I've been working on a language lately and thinking about passes - mostly where they run and what previous state they depend on - has been a big help in designing language so the compiler can be parallel, cache, and incremental friendly.
I miss getting to talk to you about music!
> Some of this is language design though.
Totally, but it's no fun to be stuck between users wanting some eminently useful feature and not being able to ship it because of your compiler architecture.
Nanopass had a DSL for describing what forms you delete from each intermediate language and which forms you add. Each pass was generally pretty small then, usually just replacing one form with some set of other forms.
Be cause each pass was really small it was pretty easy to reorder them. Sometimes we figured out if we moved one optimization sooner then later passes worked better. Or we realized some analysis was useful somewhere else so we rearranged the passes again. The pass ordering made it really clear which analysis results are valid at each point in the compilation process.
There's also a question of data about the trees (like, a flow graph) being recomputed for each nanopass. Also expensive.
The Nanopass dsl just gives the user a nicer syntax to specify the transformations.
Admittedly, this framework has extensive metaprogramming (it's a Racket DSL), so it probably has much less redundancy, but is probably even harder to debug.
To an extent, it's impossible to implement nanopass in a way that's neither redundant nor confusing, without a projectional editor. Because either each pass's AST is fully redefined, which adds redundancy, or most pass's ASTs are defined in many places, which adds confusion. But my wild speculation is that one day projectional editors will gain popularity, and someone will make a compiler where one observe and debug individual passes.
I'm creating a language/compiler now, and I'm quite certain that I did not have enough passes initially, but I hope I'm at a good spot now - but time will tell.
https://andykeep.com/pubs/dissertation.pdf
Also see the this text:
Bottlenecks are changing and it's pretty interesting.
Nanopass is compiler passes, which each have their own IR, and run once in a fixed sequence.
Egraphs also require the IR to be defined a specific way, which prevents some optimizations. My understanding of https://github.com/bytecodealliance/rfcs/blob/main/accepted/... is that cranelift’s egraph optimizations are pure expression reordering/duplication/deduplication and rewrite rules.
https://blog.sigplan.org/2021/04/06/equality-saturation-with...
[1] The acyclic e-graph: Cranelift's mid-end optimizer https://news.ycombinator.com/item?id=47717192