LLVM/Clang 3.2 SVN Compiler Performance

Printable View

10-13-2012, 12:30 AM
phoronix

LLVM/Clang 3.2 SVN Compiler Performance

Phoronix: LLVM/Clang 3.2 SVN Compiler Performance

With LLVM 3.2 set for release in mid-December, the time to benchmark this next major compiler infrastructure release paired with the Clang C/C++ compiler is approaching. Well, that time has already come; up this weekend are some benchmarks of the Intel Core i7 3960X "Sandy Bridge Extreme Edition" on Ubuntu 12.04 when comparing LLVM/Clang 3.1 to their latest SVN development code for LLVM/Clang 3.2 as of this week.

http://www.phoronix.com/vr.php?view=17983
10-13-2012, 05:56 PM
ryao

Did Michael seriously just publish another set of benchmarks without setting the optimization level? These are useless.

If Michael is not willing to learn how a compiler is meant to be used, he should not bother benchmarking them.
10-14-2012, 08:12 AM
XorEaxEax

Quote:

Originally Posted by ryao

Did Michael seriously just publish another set of benchmarks without setting the optimization level? These are useless.

If Michael is not willing to learn how a compiler is meant to be used, he should not bother benchmarking them.

Indeed, I've been harping about this for as long as I can remember, just as I've asked him to configure/compile x264 with --disable-asm so that the compilers can actually have something to optimize and compare. Instead we are seeing the exact same hand-written assembly routines being benchmarked against eachother with the usual comment from michael: -'no difference on x264 with these compilers', I mean really?

While having regular automated tests being done on these compiler toolchains is very interesting, Michael's lack of understanding or plain uninterest in presenting valid data unfortunately makes the results little more than a curiosity. Atleast of late he has conducted some of the tests with a declared -On option, so perhaps there is hope.
10-14-2012, 04:55 PM
wizard69

Compilers have other important qualities beyond generated code.

What really needs to be tested is the compliancy with C++11. C++11 and a compilers ability to support it, will become very important very quickly. Of course the compilers need to generate good code while supporting C++11 and that should be tested, but significance here isn't that great in comparison.

To put it another way, testing really should be helping people use and transition to the new conventions. As can be seen in this report generated code isn't changing significantly so why bother fretting over a few percentage here and there.
10-14-2012, 05:16 PM
ryao

Quote:

Originally Posted by wizard69

What really needs to be tested is the compliancy with C++11. C++11 and a compilers ability to support it, will become very important very quickly. Of course the compilers need to generate good code while supporting C++11 and that should be tested, but significance here isn't that great in comparison.

To put it another way, testing really should be helping people use and transition to the new conventions. As can be seen in this report generated code isn't changing significantly so why bother fretting over a few percentage here and there.

I doubt that Phoronix possesses the ability to do standard compliance testing. Few organizations do.