Benchmarking LLVM's Clang OpenMP Support Against GCC

[phoronix]

Phoronix: Benchmarking LLVM's Clang OpenMP Support Against GCC

With it looking like LLVM Clang 3.5 might finally have OpenMP support, I tested out Intel's latest out-of-tree LLVM/Clang OpenMP code to see how the performance compares to GCC for this multi-processing API. Overall, the Clang results increase the level of competition against GCC.

http://www.phoronix.com/vr.php?view=20488

[Lockal]

Benchmarking LLVM's Clang OpenMP Support

Timed Apache Compilation

Seriously.

[Daktyl198]

Seriously.

Can't read the article? That's alright, I'll quote it for you:

Outside of OpenMP, the LLVM Clang vs. GCC results continue to vary:

[sunweb]

Can't read the article? That's alright, I'll quote it for you:

What he meant is that we don't care the time of complation, at the end of the day we care for fast binaries. Faster compilation is marketing move, once they'll reach GCC build binaries speed and stability than we can watch at how fast it compiles.

[ctwise]

Faster compilation is marketing move

Production build compiles are run once. Compiles during development are run thousands of times. The speed of compilation isn't a marketing move, it's a desperately needed advancement. If you don't build it in from the start, you'll play hell getting it later on - witness GCC's efforts to improve compilation time.

[Veerappan]

Production build compiles are run once. Compiles during development are run thousands of times. The speed of compilation isn't a marketing move, it's a desperately needed advancement. If you don't build it in from the start, you'll play hell getting it later on - witness GCC's efforts to improve compilation time.

Agreed. For our windows developers at work, running a full build of the java web app we produce takes upwards of 3-4 minutes to build/package/deploy. Just think of how little you would get done if you had to wait 4 minutes to see the results of every code change you made. Compilation speed during development workflows can be a huge deal.

Of course, our linux/mac devs have managed to cut that in half due to various optimizations (through a lot of hard work by myself), but the point still stands. You can try to get around this with tools like JRebel, but compilation/installation speed will always be important to a developer.

[jcownie]

Serial/Parallel?

You publish only one gcc time, which seems to be a serial time. (There is no -openmp in the compilation flags you quote).

For completeness (and simplicity of comparison), it would be nice to have all four sets of results (gcc, gcc -openmp, clang, clang -openmp).

[curaga]

Michael, people were complaining on slashdot that this bench had -O0 for clang.

Was it really so, or did PTS just fail to wrap clang to capture the arguments? The benches only display GCC args.