Various improvements of data race analysis in Archer / ThreadSanitizer

llvm-doe-org / llvm-project

The LLVM DOE Fork is a fork of upstream LLVM (https://github.com/llvm/llvm-project/) that hosts multiple DOE-funded projects. Contact information and other details for each project are described in the wiki.

https://github.com/llvm-doe-org/llvm-project/wiki/

Other

22 stars 13 forks source link

Various improvements of data race analysis in Archer / ThreadSanitizer #25

Closed jprotze closed 9 months ago

jprotze commented 9 months ago

Detailed description available in the individual commit messages.

Improvements:

reduce runtime overhead for concurrent reads from high number of threads
task-centric data race analysis (ARCHER_OPTIONS=tasking=1)
loop-level data race analysis (e.g., ARCHER_OPTIONS=dispatch=13)
enables data race detection for vectorized code (e.g., AVX2/AVX512)
enable detection of certain data races in combination with OpenMP reductions