GCC 13 vs. Clang 17 Compiler Benchmarks, Early Clang 18 & GCC 14 Development Benchmarks

[phoronix]

Phoronix: GCC 13 vs. Clang 17 Compiler Benchmarks, Early Clang 18 & GCC 14 Development Benchmarks

As it's been a while since last delivering any competitive GCC versus LLVM Clang compiler competitive analysis and with the year quickly drawing a close, here's a fresh look at the GCC vs. Clang C/C++ compiler performance of various resulting application binaries tested on x86_64. GCC 13 vs. Clang 17 were tested as what's readily available on Ubuntu 23.10 Linux plus a look ahead in using the latest GCC 14 and LLVM Clang 18 development snapshots as of this week.

https://www.phoronix.com/review/gcc-clang-eoy2023

[jeisom]

Those are some wild differences that get pretty close by the geometric mean. And that is just on Zen4. Would be curious to see results on various AMD64 generations from Intel and AMD as well as on non-AMD64 arches like aarch64. Does gcc "win" on Zen3? How about pi5 or AppleSi. Don't you love when you have more questions after seeing results like these.

[mathletic]

Nice to see some numbers! I hope until the release of GCC 14 they fix some of the performance regression they have piled in the issue tracker. Thankless task, much appreciated!

[Kjell]

We need Clang vs GCC gaming benchmarks

[user1]

We need Clang vs GCC gaming benchmarks

I don't think it would really bring meaningful difference for proprietary games, especially if Wine / Proton are also not compiled from source. Though I think it would be interesting to do such comparison for open source games like Xonotic or Red Eclipse, and see the performance difference between the game compiled with GCC vs the game compiled with Clang.

[ms178]

I don't think it would really bring meaningful difference for proprietary games, especially if Wine / Proton are also not compiled from source. Though I think it would be interesting to do such comparison for open source games like Xonotic or Red Eclipse, and see the performance difference between the game compiled with GCC vs the game compiled with Clang.

Can Proton even be compiled with Clang? At least Wine cannot, if I am not mistaken. But I'd love to see that happening in my life time, on that whish list is also getting games compiled for the architecture the user runs with and regular re-compiles of popular titles with newer toolchains and optimization techniques, such as LTO + PGO and BOLT.

[user1]

Can Proton even be compiled with Clang? At least Wine cannot, if I am not mistaken.

Interesting. Idk about Proton, but I know that OpenMandriva is pretty much entirely compiled with Clang and it has Wine in its repository. I heard the only thing that's still built with GCC in OpenMandriva is glibc, but Idk about other packages. Maybe some are still built with GCC?

[ms178]

Interesting. Idk about Proton, but I know that OpenMandriva is pretty much entirely compiled with Clang and it has Wine in its repository. I heard the only thing that's still built with GCC in OpenMandriva is glibc, but Idk about other packages. Maybe some are still built with GCC?

What a coincidence, there is an experimental glibc-git repo out there maintained by a Linaro dev for making it compatible with Clang and I've been using it for some days now. It works as advertised. That would make wine one of the last holdouts for gcc-mingw.

[lowflyer]

It strikes me quite odd that so many significant differences are "equaled out" in the geometric means. IMHO this means that *both* compilers are not yet doing the maximum possible - i.e. there is still quite some room for improvement on both sides.

[discordian]

It strikes me quite odd that so many significant differences are "equaled out" in the geometric means. IMHO this means that *both* compilers are not yet doing the maximum possible - i.e. there is still quite some room for improvement on both sides.

both the choice of the "geometric mean", aswell as the result is weird. the only use i could think of is multiplicating the ratios (like gcc has 1.2 times more throughput here, 0.9 times there).
yet the "mean" is in the 3000s instead of close to 1.