Some Fresh Clang Optimization Benchmarks

[phoronix]

Phoronix: Some Fresh Clang Optimization Benchmarks

As some extra benchmarks to toss out there this weekend are some Clang 3.8 SVN compiler benchmarks when trying out different optimization levels...

http://www.phoronix.com/scan.php?page=news_item&px=Nov2015-Fresh-Clang-O

[Guest]

Hmm, is there a reason for no -0fast, or does that not exist with Clang?

I compile some software daily and currently use -0s, and prior to that, -02, but would really be interested in seeing if there's any real benefits to -0fast in benchmarks, mainly with GCC (but wouldn't mind also with Clang).

[Guest]

Hmm, is there a reason for no -0fast, or does that not exist with Clang?

I compile some software daily and currently use -0s, and prior to that, -02, but would really be interested in seeing if there's any real benefits to -0fast in benchmarks, mainly with GCC (but wouldn't mind also with Clang).

GCC is unable to ensure that -O3 and -Ofast are always better than -O2.

[Guest]

The real benefits do come with -march=native and profile guided optimizations.

Just some notes:

- I am avoiding -march=native because Valgrind is unable to handle certain AMD instructions.

- In my opinion, C/C++ profile guided optimizations won't be generally used until the recompilation is fully automated (or at least reasonably automated).

[quaz0r]

the main difference you would see with -Ofast is that it turns on ffast-math

[Marc Driftmeyer]

You guys whined about no OpenMP support in LLVM/Clang. It's now well supported and you don't leverage it with GraphicsMagick and more. What gives?

[carewolf]

GCC is unable to ensure that -O3 and -Ofast are always better than -O2.

So is clang and ANY compiler with an -O3 level. If an optimization is trivially better it would be included in -O2, -O3 is specifically for things that are most likely faster, but is likely to make binaries larger and if the program has no large inner-loop the compiler can improve, it might overall become slightly slower.

On another topic, seems clang has two minor issues they might want to look into, there was a -O1 to -O2 performance hit and a -march=native performance hit.