[GSoC 2021] Some ideas & questions about the Enzyme GSoC project

Hello, I’m working on the GSoC project “Integrate Enzyme into Rust to provide high-performance differentiation in Rust”. I have skimmed over the codebase and now I have some preliminary ideas about the project.

As you have said, the project can be divided to two main parts:

  1. passing and parsing Rust type metadata into LLVM / Enzyme type analysis, and

  2. integrating the Enzyme API/pass into rust.

Integrating Enzyme API to rust is trivial by using proc_macros in Rust. However, passing Rust type metadata to LLVM and integrating the Enzyme pass into rust is difficult if we don’t modify the Rust compiler. The Rust compiler seems to lack ability to emit metadata to LLVM, and if we want to add the Enzyme pass without modifying the compiler, the only way is, I think, using a command like

which is not user-friendly. I sought hard for a nonintrusive way to implement them but got nothing, so it seems that the only option is modifying the Rust compiler to let it

  1. emit extra debuginfo to the generated IR code, and

  2. use lld as its linker and add the Enzyme pass to lld

with a specific option, say “–enzyme”

However, is it appropriate to modify the compiler to just port a plugin to it, though the modification won’t be big or impact the rest of the compiler?

PS: Another question is about differentiating closures in Rust, but that can be postponed to the following discussions.

Chuyang Chen

MSc Student

At Department of Computer Science and Technology, Nanjing University

A thing I forgot to mention is that a Rust feature named “compiler plugin” do satisfy the requirement of “emitting extra debuginfo to the generated IR code”, but it’s unstable and won’t be stabilized in the foreseeable future. Moreover, as to “using lld as its linker and add the Enzyme pass to lld”, compiler plugins help nothing.

There’s certainly utility to be gained by hooking into the rust compiler. At minimum this allows us to effectively integrate high level functions for calling into Enzyme as an API without going through FFI.

As you say, one of the biggest issues here is essentially parsing Rust types. This actually may not require a modification to rustc as I believe this is available in LLVM as LLVM metadata (see https://github.com/tiberiusferreira/oxide-enzyme/issues/1#issuecomment-735634144 for an earlier discussion on this topic specifically).

It turns out that one doesn’t need to piggyback off LLD if rustc itself can call the custom LLVM pass which runs Enzyme. This could be done by say loading a plugin to rust, or otherwise calling the API. Of course in general enabling better integration of LLVM internals/pass plugins into rust directly is potentially desirable. I recently proposed a patch to rustc that would allow importing LLVM plugins that may be worth a look (see https://github.com/rust-lang/compiler-team/issues/419 https://github.com/rust-lang/rust/pull/82734). That said such support isn’t a blocker, as much as we need to start parsing this metadata. In theory we can even fork the “nice interface” and “parsing metadata” into two concurrent projects/pieces.

There’s certainly utility to be gained by hooking into the rust compiler. At minimum this allows us to effectively integrate high level functions for calling into Enzyme as an API without going through FFI.

As you say, one of the biggest issues here is essentially parsing Rust types. This actually may not require a modification to rustc as I believe this is available in LLVM as LLVM metadata (see https://github.com/tiberiusferreira/oxide-enzyme/issues/1#issuecomment-735634144 for an earlier discussion on this topic specifically).

It turns out that one doesn’t need to piggyback off LLD if rustc itself can call the custom LLVM pass which runs Enzyme. This could be done by say loading a plugin to rust, or otherwise calling the API. Of course in general enabling better integration of LLVM internals/pass plugins into rust directly is potentially desirable. I recently proposed a patch to rustc that would allow importing LLVM plugins that may be worth a look (see https://github.com/rust-lang/compiler-team/issues/419 https://github.com/rust-lang/rust/pull/82734). That said such support isn’t a blocker, as much as we need to start parsing this metadata. In theory we can even fork the “nice interface” and “parsing metadata” into two concurrent projects/pieces.