Announcement

Collapse
No announcement yet.

GCC 4.9.2 vs. GCC 5 Benchmarks On An Intel Xeon Haswell

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • GCC 4.9.2 vs. GCC 5 Benchmarks On An Intel Xeon Haswell

    Phoronix: GCC 4.9.2 vs. GCC 5 Benchmarks On An Intel Xeon Haswell

    For those craving some more GCC 5 compiler benchmark numbers following last week's release of GCC 5.1, here's some new comparison numbers between GCC 4.9.2 stable and the near-final release candidate of GCC 5.1...

    http://www.phoronix.com/scan.php?pag...enchmarks-Xeon

  • #2
    Looks like there was a noticeable regression on C-Ray but other than that no real difference.

    Comment


    • #3
      it looks like the ones where the old gcc wins the most, it is being passed march=native and the new gcc isnt? thats a big difference right there

      Comment


      • #4
        Originally posted by kenjitamura View Post
        Looks like there was a noticeable regression on C-Ray but other than that no real difference.
        cray is tracked on AMD hardware at http://gcc.opensuse.org/c++bench-frescobaldi/c-ray/
        and given the fact that internal loop is almost trivial once one specific function (ray_sphere) is inlined (it is not real raytracer, only on balls), I would guess this is more probably a noise, too.

        Comment


        • #5
          shouldn't gcc5 be tested with new optimizations it offers instead of using exact same optimization flags as in gcc-4.9 ? what is the point of this benchmark ?

          Comment


          • #6
            Originally posted by yoshi314 View Post
            shouldn't gcc5 be tested with new optimizations it offers instead of using exact same optimization flags as in gcc-4.9 ? what is the point of this benchmark ?
            Even with -O2 or -Ofast the performance should improve between GCC versions. I think the problem is bit more with choice of the benchmarks that often have quite small hand optimized internal loops that are relatively easy to generate code for and thus there is just small margin for improvement from GCC 4.9 (or earlier)

            Main areas of improvements (beside hardware enablement and implementing new language features) is improving codegen with LTO, reducing abstraction penalty in modern C++ programs, improving loop optimization (that shows a lot i.e. for numeric programs) and adding support for parallelization (better vectorization, openMP etc.). So if you do benchmarks stressing these you are likely seeing more of difference from release to release

            Comment


            • #7
              I'd like to see the results w/o LTO.

              Comment


              • #8
                I recompiled my whole system with gcc 5.1 with -Ofast then -O2

                I encountered a few problems - the main ones were with Qt5 & Plasma 5 which wouldn't launch correctly - so I've switched back to gcc 4.9

                Qt5/Plasma5 seems to have issues with -Ofast too

                Comment


                • #9
                  what about clang vs gcc ?

                  Comment


                  • #10
                    Originally posted by FireBurn View Post
                    I recompiled my whole system with gcc 5.1 with -Ofast then -O2

                    I encountered a few problems - the main ones were with Qt5 & Plasma 5 which wouldn't launch correctly - so I've switched back to gcc 4.9

                    Qt5/Plasma5 seems to have issues with -Ofast too
                    Interesting that you avoided most problems that gentoo found so far : https://bugs.gentoo.org/showdependen...ide_resolved=1

                    Comment

                    Working...
                    X