Clang's Competition For GCC On Intel Haswell

[Pseus]

Does that mean that the GCC devs tuned the performance for certain benchmarks to get better results than what would normally come out of the compiler...?

It probably means they found a way to speed up certain computations by analyzing the code of the benchmark in question.

[tuubi]

Does that mean that the GCC devs tuned the performance for certain benchmarks to get better results than what would normally come out of the compiler...?

It's more likely that they optimized the functionality these benchmarks happen to test, which is likely to be beneficial in at least some real-world use cases. Any hacks to speed up specific benchmarks like the ones in proprietary graphics drivers from AMD and NVidia would be difficult to hide in an open source code base, after all. Not to mention an idiotic thing to do with no benefit to the project.

[curaga]

For C-ray, it was about making the inlining more aggressive. C-ray has a static function called a lot, gcc wasn't inlining it, now it does. It helps more code than just the bench.

[FourDMusic]

And now please proof that it really is GCC that got slower, and not llvm/clang that got faster. Beacause IMHO it is the second one that happened.

http://www.phoronix.com/scan.php?pag..._pentium&num=1

GCC 4.7, 4.8, 4.9 benchmarks for Haswell dual core shows GCC getting faster or being approximately stable across releases.

[s_j_newbury]

llvm/clang auto-vectorizes by default, right? I thought gcc did too on relevant targets at -O3, via "-ftree-vectorize", but according to "gcc -O3 -march=native -Q --help=optimizers" it doesn't!

Tested with amdfam10 (AMD Phenom(tm) II X4 965 Processor) and i7-3840QM, gcc-4.9 git checkout from a few weeks ago.

[Herem]

Impressive... If Clang had openmp support it would be faster overall...

Anyone know how close the OpenMP support is to being merged? I notice on the OpenMP LLVM site it mentions:

Support for OpenMP 3.1 in Clang is in the process of being promoted into the Clang mainline, and can be found at OpenMP/Clang.

But as far as I can see there are no estimates for when this will be complete and which version of LLVM will contain the code.

[Guest]

llvm/clang auto-vectorizes by default, right? I thought gcc did too on relevant targets at -O3, via "-ftree-vectorize", but according to "gcc -O3 -march=native -Q --help=optimizers" it doesn't!

Tested with amdfam10 (AMD Phenom(tm) II X4 965 Processor) and i7-3840QM, gcc-4.9 git checkout from a few weeks ago.

So, FUD goes on from Phoronix. It's at least second time when Phoronix makes unfair benchmarks in favor to clang.

[Michael]

So, FUD goes on from Phoronix. It's at least second time when Phoronix makes unfair benchmarks in favor to clang.

No it's not, it's testing the defaults. It's a choice Clang chose to make for -O3 and afaik, -ftree-vectorize is on GCC for -O3.

[computerquip]

Test in favor of LLVM

Test in favor of GCC

I don't get what the big deal is. One tests compile-time optimization. The other tests run-time optimization. A compiler can be slow as hell but output greatly optimized code.

[Michael]

So, FUD goes on from Phoronix. It's at least second time when Phoronix makes unfair benchmarks in favor to clang.

It is indeed a reporting bug with GCC - http://gcc.gnu.org/ml/gcc/2014-02/msg00054.html