GCC 12 Profile Guided Optimization Benchmarks With The AMD Threadripper 3990X

[phoronix]

Phoronix: GCC 12 Profile Guided Optimization Benchmarks With The AMD Threadripper 3990X

Last month I ran a number of GCC 12 compiler optimization benchmarks for this latest-stable compiler atop an AMD Ryzen Threadripper 3990X workstation. Those tests included various optimization levels as well as link-time optimizations (LTO). Some Phoronix Premium supporters also requested to see some fresh GCC 12 Profile Guided Optimization (PGO) benchmarks, so here in this article are those PGO benchmark results...

https://www.phoronix.com/news/GCC-12-PGO-TR-3990X-AMD

[polluktus]

maybe one day we'll see pgo for linux kernel and some amazing benchmarks with lto+pgo

[coder]

That's some nice benefit.

I do wonder just how much of it could be gained by manually adding a few function attributes[1] and builtin functions[2], like __builtin_expect(), in well-chosen spots. Otherwise, GCC has no idea how hot a loop, function, or condition is.

1. https://gcc.gnu.org/onlinedocs/gcc-1...ion-Attributes
2. https://gcc.gnu.org/onlinedocs/gcc-1...Other-Builtins

[espi]

That's some nice benefit.

I do wonder just how much of it could be gained by manually adding a few function attributes[1] and builtin functions[2], like __builtin_expect(), in well-chosen spots. Otherwise, GCC has no idea how hot a loop, function, or condition is.

1. https://gcc.gnu.org/onlinedocs/gcc-1...ion-Attributes
2. https://gcc.gnu.org/onlinedocs/gcc-1...Other-Builtins

Lets make a C JIT compiler so that it can find out in real time what code is hot amirite? it could also apply optimizations for the system its running on.

[coder]

Lets make a C JIT compiler so that it can find out in real time what code is hot amirite? it could also apply optimizations for the system its running on.

I guess there's a bit of a packaging problem, since you'd need all the build dependencies. Not to mention the lack of buildsystem standards. And how fast a C or C++ codebase compiles is hugely variable, ranging from almost-JIT to not-remotely-close.

It's not as if you can just concatenate all the source files, either. That would break static, plus various things people do with the preprocessor.

[kozman]

maybe one day we'll see pgo for linux kernel and some amazing benchmarks with lto+pgo

Has anyone ever attempted it recently? I feel like I remember back in the early gcc 9 or 10 days someone tried it and there were problems or it just didn't work.

[polluktus]

Has anyone ever attempted it recently? I feel like I remember back in the early gcc 9 or 10 days someone tried it and there were problems or it just didn't work.

i was using https://gitlab.com/torvic9/linux513-...pgo-kees.patch with 5.13 and cfi. Didn't do any benchmarks, but it worked. But currently i dont think anyone is still working on it.

[V1tol]

Has anyone ever attempted it recently? I feel like I remember back in the early gcc 9 or 10 days someone tried it and there were problems or it just didn't work.

We should have LTO with GCC first

[cj.wijtmans]

Yeah at the moment kernel can only be compiled lto with clang. Which is funny because for a long time clang didnt compile kernel. Also we have gcc extensions being used in the kernel. So they would not work with clang? Weird situation.
Anyway.
is the primary optimization of pgo perhaps branch prediction? Because c++ supports branch prediction hints now. And C does as well. What other major optimizations would pgo have that cant be done manually. I think pgo is DOA and we just need more standardized attributes to give the compiler hints. Oh theres cache lining hints as well i believe in c++ now.

[joshx1]

I recently went down this rabbithole with Rust, and it's interesting!

See their blog post at https://blog.rust-lang.org/inside-ru...-compiler.html and their documentation https://doc.rust-lang.org/rustc/prof...imization.html