Announcement

Collapse
No announcement yet.

LLVM Clang 3.4 Is Running Very Well On AMD's High-End APU

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

  • LLVM Clang 3.4 Is Running Very Well On AMD's High-End APU

    Phoronix: LLVM Clang 3.4 Is Running Very Well On AMD's High-End APU

    After earlier this week running GCC 4.8.2 vs. GCC 4.9 development snapshot benchmarks on the AMD A10-7850K Kaveri APU, up for testing today are new compiler tests from this new high-end APU comparing GCC 4.9 in its current development form to LLVM Clang 3.4. This GCC 4.9 vs. LLVM Clang 3.4 compiler performance comparison is more competitive than some of the past compiler comparisons and does hold a few surprises.

    http://www.phoronix.com/vr.php?view=19797

  • #2
    First thing:
    I see that you are using -O3. Clang has -O4, which uses -O3 and LTO. Is it possible to use this flag?

    Second thing:
    I would like to see a benchmark with more "distribution like" compiler flags like:

    -march=x86-64 -mtune=generic -O2 -pipe -fstack-protector --param=ssp-buffer-size=4
    (from ArchLinux)

    It would be nice to see, what are the differences for distributions between Clang and GCC. I think most people are using software compiled with O2 and march=x86-64.

    Comment


    • #3
      Originally posted by -MacNuke- View Post
      First thing:
      I see that you are using -O3. Clang has -O4, which uses -O3 and LTO. Is it possible to use this flag?

      Second thing:
      I would like to see a benchmark with more "distribution like" compiler flags like:

      -march=x86-64 -mtune=generic -O2 -pipe -fstack-protector --param=ssp-buffer-size=4
      (from ArchLinux)

      It would be nice to see, what are the differences for distributions between Clang and GCC. I think most people are using software compiled with O2 and march=x86-64.
      -O4 is the same as -O3 since Clang 3.4. (http://llvm.org/releases/3.4/tools/c...compiler-flags)

      Comment


      • #4
        Originally posted by -MacNuke- View Post
        First thing:
        I see that you are using -O3. Clang has -O4, which uses -O3 and LTO. Is it possible to use this flag?

        Second thing:
        I would like to see a benchmark with more "distribution like" compiler flags like:

        -march=x86-64 -mtune=generic -O2 -pipe -fstack-protector --param=ssp-buffer-size=4
        (from ArchLinux)

        It would be nice to see, what are the differences for distributions between Clang and GCC. I think most people are using software compiled with O2 and march=x86-64.
        I'm using -march=bdver2 and seeing as this is a bdver3 (steamroller) I would expect the march and tuning flagged specifically for it.

        Comment


        • #5
          As far as I know LLVM doesn't have any optimizations based on AMD extensions. So they might be a little handicapped compared to GCC which can use XOP and other AMD extensions. Then again maybe LLVM should hurry up and fix that

          Comment


          • #6
            I like that the results are all over the place.

            It provides plenty of room for improvement for both platforms.

            Comment


            • #7
              There is a strong bias towards LLVM Clang...

              Comment


              • #8
                scimark Jacobi SOR test

                Jakub Jelinek has improved GCC's jacobi scores at 10th of January. http://gcc.gnu.org/ml/gcc-patches/2014-01/msg00646.html
                I tried to reproduce the results on bdver1 machine and get 896 from LLVM (today checkout), 693 from gcc-4.8 and 1239 from today checkout of GCC.
                So it seems to me that GCC used was not GCC 4.9.0 20140126 or it is some really architecture specivic bdver3 thing.

                Comment


                • #9
                  It looks like not even concrete results where Clang performs better than GCC will convince people here that Clang is already in pair with GCC. No, that is just impossible; no body can beat the almighty GCC. It is all a conspiracy...
                  They will continue to say that Clang produces "vastly inferior binaries".

                  Comment


                  • #10
                    Originally posted by Sergio View Post
                    It looks like not even concrete results where Clang performs better than GCC will convince people here that Clang is already in pair with GCC. No, that is just impossible; no body can beat the almighty GCC. It is all a conspiracy...
                    They will continue to say that Clang produces "vastly inferior binaries".
                    Read the results again. Maybe you confused the graphs. GCC is the one that is faster in 90% of the tests.

                    Comment


                    • #11
                      Originally posted by carewolf View Post
                      Read the results again. Maybe you confused the graphs. GCC is the one that is faster in 90% of the tests.
                      90%? It seems you are the one confused...

                      Comment


                      • #12
                        Originally posted by Sergio View Post
                        90%? It seems you are the one confused...
                        I find the current clang results pretty impressive. But then i think someone said that clang is using link time optimizations by default now in -O3? That probably explains it right there. If GCC had -flto added to the flags it would probably jump out in front again.

                        Comment


                        • #13
                          Originally posted by smitty3268 View Post
                          I find the current clang results pretty impressive. But then i think someone said that clang is using link time optimizations by default now in -O3? That probably explains it right there. If GCC had -flto added to the flags it would probably jump out in front again.
                          I haven't said that if you're referring to my post. -flto needs to be manually enabled, it isn't with any -O levels.

                          Comment

                          Working...
                          X