AMD Optimizing C/C++ and Fortran Compilers (AOCC)

Overview

The AOCC compiler is designed for high-performance x86 CPU compilation of C, C++, and Fortran programming languages. It offers target-dependent and target-independent optimizations, with a particular focus on AMD "Zen" processors. These optimizations provide performance advantage for time-critical applications, especially in the field of high-performance computing (HPC). The AOCC compiler environment streamlines and speeds up the development and tuning process for x86 applications written in C, C++, and Fortran.

AOCC Features

• Platform and Foundation
  Built for 32-bit and 64-bit Linux® platforms, leveraging the robust LLVM™ infrastructure (based on LLVM 17.0.6, Nov 2023) with Clang as the default front-end for C/C++ and Flang for Fortran.
• Architecture Optimization
  Tuned for AMD processors across all “Zen” generations (Zen, Zen2, Zen3, Zen4, Zen5).
• Language Standards and Compliance
  Supports C17 (default for C), C++17 (default for C++), and Fortran F2008 with Real128 features (coarrays not supported);
  OpenMP 5.0 for C/C++ and OpenMP 4.5 for Fortran; DWARFv4 debugging by default with DWARFv5 available for C, C++, and Fortran.
• Advanced Features and Optimizations
  Includes inter/intraprocedural analysis, SLP and loop vectorization, loop optimizations, OpenMP Debugging Interface (OMPD), and CPU offload support for Fortran OpenMP.
• Additional Enhancements
  Optimization level -O2 and -fPIC/-fPIE options are default since 4.1; supports Spack for flexible package management and AMD optimized application recipes.

Notes

1. AOCC compiler binaries depend on Linux® systems having Glibc version 2.28 or later.
2. Since AOCC 4.1, the -fPIC and -fPIE options are enabled by default, which means location-independent code is generated automatically. For AOCC 4.1 or later, any static libraries that were previously built without -fPIC must be rebuilt using these default options; otherwise, linker errors may occur during static linking. Alternatively, you can disable position-independent code generation by specifying -fno-PIC or -fno-PIE when compiling with AOCC.

What’s new in AOCC 5.2

• Based on LLVM 17.0.6 release (llvm.org, Nov 2023)
• Zen5 tuned AOCL-LibM 5.3 (AMD Math Library)
• Bug fixes impacting C, C++, and Fortran applications
• Added DO CONCURRENT parallelization feature in Fortran
• -finline-aggressive has been deprecated in AOCC 5.1. Alternatively, use -mllvm -inline-threshold
• -freroll-loops has been deprecated in AOCC 5.1

AOCC 6.0 Announcement:

Introducing a New Fortran Compiler in AOCC 6.0
AOCC 6.0 is planned to introduce a new Fortran compiler, aof, based on llvm-flang. This compiler is expected to be packaged alongside the existing classic Flang compiler, which will continue to serve as the default Fortran compiler. The new toolchain is planned to have aof (Fortran compiler), aoc (soft link to clang) and aoc++ (soft link to clang++).

Migration to DWARFv5 Debug Format (Default in AOCC 6.0)
Starting with the AOCC 6.0 release, DWARFv5 becomes the default debug information version. When the -g option is specified, both the Fortran and C/C++ compilers will emit DWARFv5 debug information by default. To continue generating DWARFv4 debug information, explicitly specify the -gdwarf-4 option during compilation.

Users currently on AOCC 4.x or 5.x can evaluate DWARFv5 compatibility by rebuilding their workloads with the -gdwarf-5 option before upgrading.

Download with End User License Agreement