A Quick Look At GCC 4.9 vs. LLVM Clang 3.5

[phoronix]

Phoronix: A Quick Look At GCC 4.9 vs. LLVM Clang 3.5

Following my most recent GCC 4.9 benchmarks for the open-source compiler that should be officially released next week, I ran some benchmarks of the GCC compiler results against LLVM's Clang 3.5 compiler in its latest SVN state. Here's the data for those curious how the very latest compiler code is comparing between GCC and LLVM/Clang.

http://www.phoronix.com/vr.php?view=20189

[Szzz]

Usually applications with hand tuned assembly can be compiled without these optimizations; this would allow compliers to show their power. This way we can compare compilers not only with each other but also with humans. It would be interesting.

[toguro123]

Uh oh!

Flame war! 3 2 1 GO!!

[gamerk2]

Still really not a fair comparison until OpenMP gets mainlined into LLVM. Do we have any new news on that front?

[dimko]

Thanks!

Phoronix: A Quick Look At GCC 4.9 vs. LLVM Clang 3.5

Following my most recent GCC 4.9 benchmarks for the open-source compiler that should be officially released next week, I ran some benchmarks of the GCC compiler results against LLVM's Clang 3.5 compiler in its latest SVN state. Here's the data for those curious how the very latest compiler code is comparing between GCC and LLVM/Clang.

http://www.phoronix.com/vr.php?view=20189

Thanks for bunch of interesting news recently Michael.
For me, a Gentoo user, this is very interesting.

[Michael]

Still really not a fair comparison until OpenMP gets mainlined into LLVM. Do we have any new news on that front?

Only a few tests were even using OpenMP and again it was clearly noted that LLVM Clang still lacks mainline OpenMP support. It's fair considering that OpenMP is widely used.

[Vim_User]

Very nice, indeed. So for projects not using OpenMP we can say that LLVM is up to par with GCC, at least in most cases, and compiles much faster.
All we need now is OpenMP support and the ability to compile the Linux kernel with it, and on both subjects people are already working.
Once it works would be nice to see kernel compile times with both compilers.

[oleid]

Michael, thanks for testing! Is Clang still much slower on AMD processors, as it used to be? Or to put it differently: is it still optimized for Intel?

Also, considering the *BSDs are switching to Clang as default compiler, how does -mtune=generic compare between gcc and clang? And how does it compare to -march=native? I'd find that rather interesting.

Best,
Olaf

[rmeakins]

Typo

Nice info.

I think there is a typo in the ebizzy benchmark. The unit says "Seconds, more is better" but it should say "Records per second, more is better".

[Marc Driftmeyer]

Only a few tests were even using OpenMP and again it was clearly noted that LLVM Clang still lacks mainline OpenMP support. It's fair considering that OpenMP is widely used.

Your logic is flawed.

``Here we have a test that can be compiled with or without OpenMP. One compiler supports, the other presently does not. We've decided to enable it to show how GCC kicks the crap out of LLVM who cannot leverage OpenMP. OpenMP is widely used, so it is a fair race.''

A fair comparison is to show the test with both compilers having disabled OpenMP support. We can determine from this just how dependent the test is on OpenMP versus the present state of compilers, but we feel this isn't fair to GCC who is already getting stomped too often.