GCC 14 vs. LLVM Clang 18 Compiler Performance On Fedora 40

[vladpetric]

The performce bugs I have filed has been adressed, especially if you debug it down to general cases and assembler. Though I also fixed a few of them myself. The only one I can remember that wasn't fixed, was because it was caused by a specific micro-ops limit in Sandybridge processors, so slightly more agressive optimization caused a regression, but they didn't want to hardcore such specific limits for all processors. clang also doesnt have that.

E.g. https://bugs.llvm.org/show_bug.cgi?id=20014

[mrg666]

At this point I am not expecting any compiler to beat GCC after averaged in a set of benchmarks like Michael did here. GCC 14 is also great with building the latest kernel version. I have been using with Fedora 40 on 8 (or so) computers and there are no problems at all. I tried LLVM before for building Linux kernel, but I did not see any performance miracles there. It is just as good as GCC. Just take your pick.

[junkbustr]

Is that clock speed on the CPU (7.79Ghz) a typo? I thought the vendor offered 5.1Ghz on the 7980X.

ALL RIGHT! WHO USED UP ALL THE LIQUID NITROGEN??

[imaami]

Yawn. I'm already on clang-19.

[Kjell]

Yawn. I'm already on clang-19.

Gentoo?

[Anux]

Good we have 2 OS compilers sort of competing with each other, more benefits and choices for us.

Not a lot of sense in comparing video codecs because they include hand written assembly for most important/heavy parts of processing, so the compiler's role is quite minimal.

Since we see up to 5% difference in video codecs it certainly is valuable to compare them and assembly clearly isn't the only relevant factor.

[avis]

Good we have 2 OS compilers sort of competing with each other, more benefits and choices for us.

Since we see up to 5% difference in video codecs it certainly is valuable to compare them and assembly clearly isn't the only relevant factor.

Um, where?

SVT-AV1: ~2%
x265: ~2%

ugv266 is a dead encoder without too much assembly where yeah, up to 5%, but no one cares. VVC is hated and destroyed here on Phoronix, why would you use it? vvenc is tons better, why would you use ugv266?

What's the practical use of seeing the results of something that absolutely nobody uses? It's like the dude earlier in the thread who requested ages old Quake and Doom games to be tested. Why? This will extremely unlikely to translate to any modern workloads.

[marios]

Although those benchmarks are a nice start, we need to see the effects of LTO, graphite vs poly etc. PGO and bolt, albeit useful, are too much of a hassle, since they are not just a flag.

[Anux]

Um, where?

Exactly the places that you avoid looking because the results are an inconvenience for your strange argumentation. Kvazaar ugv266

What's the practical use of seeing the results of something that absolutely nobody uses?

How do you know that no one uses it? Are you working for the NSA?

[imaami]

Gentoo?

Just Debian, tracking the unstable repo + specific parts of experimental (e.g. the clang/llvm stuff).