normalization-bench VS ghc-whole-program-compiler-project

I benchmarked lambda term normalization, which is heavy on closures (HOAS) and small allocation, and there V8 RTS performance was clearly inferior, and I even had stability issues with deep stack usage.

Another thing you can try is the ghc-wpc project which has an interpreter which supports breakpoints, though you may need to hack little a bit to achieve your goals.

If I understand the new GHCJS backend correctly, it diverges from the normal compilation pipeline at STG (see the new StgToJS module subtree). Wouldn't that make it an excellent candidate to use GHC-WPC's machine-readable external STG representation, instead of hooking it directly into the GHC source tree?

Here the goal is to build a high level, easy to understand model for all GHC backend features. Validations is also required. Once we know the semantics of GHC primops and RTS features then it becomes possible to figure out how to compile Haskell programs to GRIN. I started the GHC-WPC project for this reason. GHC-WPC exports the STG intermediate representation for the whole Haskell program, and I wrote an STG interpreter from scratch in Haskell that can run any Haskell program. (i.e. GHC itself) The STG interpreter is the high level model for the GHC primop and RTS semantics. It implements all these in pure Haskell, it does not depend on GHC RTS at all.

I can easily debug any Haskell program with the external STG interpreter. https://www.youtube.com/watch?v=DkDUEd3pUyM https://github.com/grin-compiler/ghc-whole-program-compiler-project

Hello, I'm using the external STG interpreter to observe the runtime behaviour of Haskell programs. Lately I've added an initial call-graph construction feature that I plan to refine further. https://twitter.com/csaba_hruska/status/1417486380536582151 Is there anyone who has dynamic analysis related research ambitions and wants to study Haskell program runtime behaviour in detail? If so then it would be great to talk.

I haven't gotten my hands dirty yet, but really excited hearing GHC-WPC is going on!

Why didn't you mention GHC-WPC? It is also a backend sample. It exports enough information (STG + linker opts + c bits) to interpret the program or to generate a binary executable via the regular GHC codegen system. https://github.com/grin-compiler/ghc-whole-program-compiler-project

You could use the GHC codegen and RTS via the external STG IR. https://github.com/grin-compiler/ghc-whole-program-compiler-project

Remarks: 1. Strict functional languages can be expressed in STG without overhead, because STG has explicit liftedness control. In a strict language every data is unlifted or unboxed. 2. Supporting all GHC primops is not unrealistic. See the primop implementation in the external STG interpreter source code. Here is the implementation of the threading primops.