Kiran Chandramohan

Flang is the new Fortran frontend of LLVM. While the frontend is not yet ready for general use, significant progress has been made. In this talk I summarise the progress that has been achieved and the major development thrusts.

Flang supports compilation of most of Fortran 95. Significant progress is being made on Fortran 2003 features. Flang can compile Spec and other important benchmarks. Some work has been done to improve performance. Further work is under way, this includes Alias Analysis, improved codegen for Array Expressions and Assumed shape arrays. A new dialect called HLFIR (High Level FIR) is also under construction. This dialect will capture more higher level Fortran info in the IR. Flang is mostly OpenMP 1.1 compliant. Significant progress has been made with the Driver as well. The presentation will cover these points.

This talk introduces Flang (F18), the new Fortran frontend of LLVM being written in modern C++. The talk will provide a brief introduction to Flang, motivation for writing this new compiler, design principles, architecture, status, and an invitation to contribute.

F18 project started at PGI/Nvidia as a new Fortran frontend designed to work with LLVM. The aim of the project is to create a modern Fortran frontend (Fortran 2018 standard) in modern C++. In April of this year, it was accepted as an LLVM project (https://lists.llvm.org/pipermail/llvm-dev/2019-April/131703.html).

The parser and semantic analysis are implemented in a way that provides a strong correspondence to the standards document. It is hoped that this correspondence will help in the development of new features and will become the testbed for deciding future Fortran standard features. The frontend also embraces the newly open-source MLIR framework for language-specific optimisations. This will be through a new dialect call FIR (https://www.youtube.com/watch?v=ff3ngdvUang). MLIR will also be used for creating an OpenMP dialect. The project also hopes to share code with the Clang frontend. While the parser/AST will not be shared, code will be shared in the Driver, OpenMP codegen etc.

In this presentation, we hope to cover the technical details mentioned in the paragraph above, the status of implementation and also give an invitation to contribute.