Pure Macros

Nobody has claimed this yet.

Hello,

I hope you having a wonderful day, First off let me thank you for your awesome projects; I'm a fan of your work.

I was wondering have you ever tried to cache the `TokenStream` based on the input? So maybe we can mark a procedural macro as `#[pure]` and get the results cached based on the input?

I have recently forked proc-macro2 and serde (to remove its dependency on proc-macro2) and bootstrapped a simple serializable TokenStream, But It means that I would still spend a lot of time compiling and running a noop macro that just reads the cache and returns it.

Then I came around to this article a few days ago https://www.coderemote.dev/blog/faster-rust-compiler-macro-expansion-caching/

It sounded promising to me, So that's when I started thinking about using Watt as both the runtime and cache server to prevent the redundant time spent on the macro execution by short-circuiting from the cache auto-magically.

So what I want to ask you is, Have you ever tried this? What are your thoughts? Do you think Watt can be a good place to do this kind of caching? Or do you prefer me to not submit such a PR and instead work on a standalone solution?

I also would like to know about the current situation with the `Compiler` and `Fallback` variants of the token stream and whether are they going to both be kept in or one is a temporary transition period solution. But I think that's another question for another repo.

#### Note: while this text is written with @dtolnay as the recipient the question remains open for everyone to discuss. Sorry if it felt more like an email than an issue, I just noticed the tone as I'm about to submit😅

The issue identifies no repository files, tests, or entry points to inspect. Start by reviewing Watt's current runtime and token-stream handling, then compare the proposed cache-server approach with the linked article and clarify whether caching belongs in Watt; done requires an agreed scope and design rather than an implementation target.

Written by the indexing model from the issue text.