GSoC21 W6: LFortran ASR and Values

Enriching ASR nodes at compile time

Background

Serialized update for the 2021 Google Summer of Code under the fortran-lang organization, mentored by Ondrej Certik.

Logistics

• Met with Ondrej on Monday and Wednesday
  • Discussed the design choices w.r.t. class hierarchies (and the lack thereof)

Overview

Note that the title is rather misleading, this post has nothing to do with the values of the LFortran project (which are by the way, fantastic), but instead is about adding more detail to the ASR nodes.

New Merge Requests

Implement expr values for Integer Binops (1045)

First example of using value-enriched nodes for operating during the ast2asr pass

Misc Logs

False Starts

Coming from a “throw it in a debugger” mindset, the ASR codebase and design of LFortran can be a bit disconcerting. This is because it is by design implemented mostly in terms of C structs which are cast explicitly throughout the code. Ondrej mentioned that the overhead of having a virtual function table for each instance of LFortran::ASR was prohibitive.

Why bother with such speed at the cost of the C++ guarantees?

Well for one thing, we are building a compiler, and one with an interactive kernel at that, the overhead must be kept low. No one wants to grab a coffee everytime they execute a Python jupyter cell and the goal should be the same for LFortran too.

Why not just add value to the expr_t structure?

This was a bit of a head scratcher to me, especially my debugger (lldb) offered no information on the value stored even in Constant* structures. The real solution, which required more of an understanding of the code (offered freely and kindly by my mentor) involves casting with down_cast.

Essentially, as far as the code base goes, we have a class based hierarchy, which we manipulate with casts.

Code duplication is a non-squinter though, since only integer and real have similar implementations, and Ondrej pointed out that over-engineering a design simply because two if-esle branches are similar is effectively yak shaving.

Conclusions

real was initially implemented as a string, which has now been replaced by Ondrej in 1066, making an extension to it trivial. I expect to focus on getting started on functions which will have the added benefit of easily adding functionality to the backend (since the LLVM backend can choose to simply use value). I intend to refocus my efforts towards getting minidftatom to work as well.