LLVM Clang vs. GCC Compiler Comparison On Ubuntu 16.04 LTS

[phoronix]

Phoronix: LLVM Clang vs. GCC Compiler Comparison On Ubuntu 16.04 LTS

When getting access to an assortment of new Intel Xeon E3 "Skylake" processors one of the first testing thoughts that came to mind were some fresh GCC vs. Clang benchmarks. So using the $600+ Xeon E3-1280 v5 processor running up to 4.0GHz, I carried out a comparison of the GCC and Clang compilers using the packaged versions being offered by Ubuntu 16.04, the Xenial Xerus.

http://www.phoronix.com/vr.php?view=22887

[Marc Driftmeyer]

Run against Bullet Physics 2.83. You know, something that has been out for a year. When we get to Bullet 3.0 OpenCL that should be real interesting.

[Michael]

Run against Bullet Physics 2.83. You know, something that has been out for a year. When we get to Bullet 3.0 OpenCL that should be real interesting.

Last time I tried pulling down a newer version of Bullet, I ran into problems: trying to recall, either the benchmark functionality I was using was no longer in the code-base or it was broken when trying to build it.

[Steffo]

Is there a regression in LLVM clang 3.8?

[mlau]

The scimark2 results are suspicios. My haswell laptop (2.9 GHz) scores this (gcc-5.3):

Composite Score:         2589.21
FFT             Mflops:  1989.04    (N=1024)
SOR             Mflops:  2203.60    (100 x 100)
MonteCarlo:     Mflops:  1118.20
Sparse matmult  Mflops:  2798.64    (N=1000, nz=5000)
LU              Mflops:  4836.57    (M=100, N=100)

It's faster than your skylake xeon except for LU Matrix Factorization, and the system wasn't even idle when the tests ran.
That doesn't make sense.

[cb88]

It would be interesting to see binary size for each complier / set of flags used...

[Michael]

It would be interesting to see binary size for each complier / set of flags used...

Wouldn't be too hard for someone interested to script up for PTS so it could be done automatically and universally.

[octoploid]

You must configure gcc with --enable-checking=release if you want to benchmark compile times.
It defaults to "‘assert,misc,tree,gc,rtlflag,runtime" for snapshots. This slows down the compiler immensely.

If you're unable to build your own compiler, at least use -fno-checking when compiling, to turn off some expensive checking.

[Michael]

You must configure gcc with --enable-checking=release if you want to benchmark compile times.
It defaults to "‘assert,misc,tree,gc,rtlflag,runtime" for snapshots. This slows down the compiler immensely.

If you're unable to build your own compiler, at least use -fno-checking when compiling, to turn off some expensive checking.

When I do build my own GCC builds, I do enable-checking=release, but this was a benchmark of what's packaged by Ubuntu 16.04.

[carewolf]

When I do build my own GCC builds, I do enable-checking=release, but this was a benchmark of what's packaged by Ubuntu 16.04.

Was the prerelease version in Ubuntu configured with or without checking? It is very relevant if you want to compare compiling times and look for regressions.