Project Stage 3 - Function Multi-Versioning Enhancement Project


This term, our team has been focused on extending the GCC compiler's capabilities by implementing Automatic Function Multi-Versioning (AFMV) cloning functionality. This technology enables the compiler to generate multiple optimized variants of functions specifically tailored to different CPU architectures, enhancing performance without requiring developers to manually add attributes in their code.

Our research investigated two potential implementation strategies:

  1. Core Compiler Modification: Altering GCC's fundamental processing to natively support automatic function multi-versioning.

  3. Automated Attribute Implementation: Developing a compilation pass that would automatically insert target_clones attributes into functions early in the compilation process.

After significant investigation into the first approach with guidance from our professor, we encountered substantial technical hurdles—particularly segmentation faults triggered by specific edge cases (like non-local goto statements) within GCC's codebase. Considering these challenges and our project timeline constraints, we've decided to pivot to the second approach.

Project Direction4:

For Phase III of our project, we're now prioritizing strengthening the stability and dependability of our compiler modifications through comprehensive testing protocols. While automatic function multi-versioning remains our ultimate objective, establishing a thoroughly validated compiler foundation will better position us for future enhancements.

Comments

Post a Comment

Popular posts from this blog

Project Stage 3: Enhancing GCC Pass for Multiple Function Clone Handling: Progress Update

In my earlier work, I developed a basic GCC pass for identifying and comparing function clones. For this latest phase, I focused on improving the pass to handle programs containing multiple cloned functions. Additionally, I wanted to enhance code organization by consolidating data storage, replacing multiple parallel vectors and maps with a structured approach using a dedicated struct to store characteristics of clone variants. Implementation Process Phase 1: Code Revision for Multi-Function Support I modified the pass to properly handle programs with multiple cloned functions in x86 server environments. The updated code includes all critical support files. The revised implementation's core is a specialized pass that processes GIMPLE representation data. This pass systematically identifies resolver functions and creates appropriate tracking structures for each function being cloned. The implementation uses maps and vectors to store function characteristics, including basic block co...
Read more

Lab - 1: Exploring 6502 Assembly

Image
Exploring 6502 Assembly: A Journey into Low-Level Programming  Introduction Assembly language has always been both fascinating and challenging for programmers. Unlike high-level languages where abstractions make coding easier, assembly requires a deep understanding of how the CPU processes instructions. In this lab, we explore the basics of 6502 Assembly Language through bitmap displays, code modifications, performance enhancements, and using the 6502 Emulator. This experience is the first step towards learning more complex architectures such as x86_64 and AArch64.
Read more

Project Stage 1: Troubleshooting a bit - Coming soon

  Procedure for Troubleshooting Debugging using the following method: First Error Analysis: I carefully reviewed the error logs after my initial build attempt failed. Copy ./pass-instances.def:47:23: error: expected ';' at end of member declaration 47 | NEXT_PASS_WITH_ARG (pass_mypass, 1, 1) This suggested an issue with my pass registration syntax. Examining Pass Declaration Conventions: I ran the following to examine the passes that were already included in the GCC source code: Copy cd ~/git/gcc/gcc grep -r "NEXT_PASS" passes.def This showed that NEXT_PASS (pass_name, number) was the proper format instead of NEXT_PASS_WITH_ARG. Function name Pattern Analysis: I looked at various pass implementations to learn more about the name conventions: Copy cd ~/git/gcc/gcc find . -name "tree-*.cc" | xargs grep "make_pass_" This demonstrated to me that the pattern make_pass_X, where X is the pass name without the "pass_" prefix, is what GCC ...
Read more