llvm/llvm-project

OCUDU benchmarks notably slower on clang22 than gcc16 on avx512 target

Open

#198106 opened on May 16, 2026

View on GitHub
 (8 comments) (1 reaction) (1 assignee)C++ (26,378 stars) (10,782 forks)batch import
good first issueperformance

Description

Reported here: https://www.phoronix.com/review/gcc-16-vs-clang-22/2

Running the OCUDU benchmarks on a Threadripper 9980X (-march=znver5) is almost 50% faster when built with gcc16 vs clang22: https://openbenchmarking.org/innhold/92454b6c98c3ad8beb08a5084a278386c6b06a46

This might be a good first issue for somebody with access to both compilers and an interest in codegen profiling:

1 - build and run the benchmarks on gcc16 and clang22 - preferably on a avx512 machine, but I suspect older machines will show a perf difference as well.

2 - profile to identify any hot code sections on either build and compare the builds and code quality (e.g. missing compiler flags/attributes? vectorisation width? performant use of particular vector instructions? use of slow gather/scatter instructions?)

3 - identify any missing llvm optimisations and raise suitable issue(s)

Contributor guide

Tech stack
cpp
Domain
performance
Issue type
performance
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
5
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
blocked
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
C++ programmingCompiler toolchain knowledgeBenchmarking and profiling skillsAccess to AVX512 hardware
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
First, build and run the OCUDU benchmarks from the llvm test suite with both gcc16 and clang22 on an AVX512 machine (e.g., Threadripper 9980X). Use profiling tools like perf or llvm xray to identify hot code sections and compare the generated assembly. Look for differences in vectorization width, use of gather/scatter instructions, or missing compiler attributes. Based on findings, file upstream LLVM issues for any missing optimizations. The Phoronix article linked provides baseline numbers and hardware details.