Imperative compilers (and the lack of it)

Article: Bulldozer Benchmarks, What Do They Mean?
By: Vincent Diepeveen (diep.delete@this.xs4all.nl), April 10, 2011 6:20 am
Room: Moderated Discussions
Paul (no@thanks.com) on 4/10/11 wrote:
---------------------------
>? (0xe2.0x9a.0x9b@gmail.com) on 4/10/11 wrote:
>---------------------------
>>Bottom line: if any of the compilers would just *blindly* follow the instructions
>>given by the programmer, it would actually generate optimal code in the case of
>>this particular function. But no, that's not the way how compilers work - they are
>>trying to be "smart" - but unfortunately, using those "smart" algorithms can be
>worse than using no algorithm at all !!!
>
>However most of the time compilers are asked to compile low quality code where
>blindly following what was asked for would lead to horrendously slow code. For example
>I recently encountered some code which performed the same multiplication 6 times
>in side a loop. The compiler was smart enough to notice this this, precompute it and pull it out of the loop.
>
>Now you can argue as much as you like about programmers writing optimal code to
>start off with but the fact of the matter is that most code is crap which is why
>we have compilers which try to make the most of the code and honking great big dynamic
>frontends on processors which try to make the most of what the compiler spits out.

There is no excuse for optimizing a program in a manner that slows it down when PGO gets used.

Even brain dead parameter tuning will figure out how to efficiently use which optimization.

The PGO has been broken nearly forever in GCC, with exception of 1 or 2 snapshots in 2005 and then it was broken again.

That was the case with GCC 2006-2010.

PGO didn't work at all, and i'm not the only one posting that. It's getting ignored in a systematic manner by the GCC team.

Also Linus dropped some lines on similar subjects. For example he wrote in 2007 that there is no excuse to not use CMOV type optimizations now that both core2 as well as AMD have fast ways to execute CMOV.

The response to Linus was that: "it would be slow on a P4 otherwise".

The GCC team has systematically ignored that remark from Linus and everyone and even today it's rewriting branches to total crap.

In fact it's also rewriting them in a manner which is very slow for AMD. If you get outside of the lookahead and jump to there with for example a JE (jump if equal) then such branchprediction will be horrible slow of course as it's outside of the lookahead of AMD. It was inside lookahead of intel (in fact intel rewrites it if i understand well).

yet objectively generating CMOV's is a lot faster in the generic case of course.

As this rewrite happens *prior* to pgo stage, it is obvious gonna go wrong.

The correct way to do things is to have a pgo stage decide when to apply which optimization and allow this pgo stage to decide it for ALL optimizatoins when pgo gets used.

And exactly THAT crucial thing has been broken in GCC.

If you do not want to call that clear sabotage, then it is really huge incompetence.

5 line examples are not relevant. Use some real world codes and see for yourself. Now you might not have Adobe source codes at home maybe.

But you can easily do next yourself.

Start 2 threads. Let 1 thread idle and the other thread have it execute something.

Find me 1 serious software program that nowadays isn't multithreaded. My chessprogram uses in the first proces like 2 or 3 threads form which 1 thread is the calculating thread, the others are just input/command threads and idle all time. The rest works multiprocessed.

To PGO i start a single instance, which means just 1 proces having 2 or 3 threads, in fact just 2 threads for the pgo run with GCC.

Then i apply the PGO and it speeds up 1.5% in GCC.

They should be ashamed of themselves. I've reported this for 7 years now and only 1 snapshot in 2005 seemed to do it ok, yet 0 releases ever did.

GCC is so so behind.

Vincent
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
New article: Bulldozer benchmark analysisDavid Kanter2011/03/30 02:27 AM
  New article: Bulldozer benchmark analysisNaN2011/03/30 07:47 AM
    New article: Bulldozer benchmark analysisDavid Kanter2011/03/30 09:25 AM
      New article: Bulldozer benchmark analysisdc2011/03/30 12:07 PM
      New article: Bulldozer benchmark analysisFoo_2011/03/30 12:17 PM
        New article: Bulldozer benchmark analysisDavid Kanter2011/03/30 01:34 PM
          BD won't "kick in" until 2.5GHzBaronMAtrix2011/03/31 08:37 AM
            BD won't "kick in" until 2.5GHzDavid Kanter2011/03/31 09:21 AM
              He is an AMDZone lunaticNIKOLAS2011/04/01 06:40 AM
                See how he behaves hereDavid Kanter2011/04/01 08:08 AM
                He is an AMDZone lunatichorsefly2011/04/02 05:46 AM
                  This is not relevantDavid Kanter2011/04/02 09:16 AM
                  Post removedDavid Kanter2011/04/02 12:39 PM
          New article: Bulldozer benchmark analysisanonymous2011/04/04 05:08 AM
      New article: Bulldozer benchmark analysisinterested2011/04/01 03:28 PM
        New article: Bulldozer benchmark analysisDavid Kanter2011/04/01 03:53 PM
          New article: Bulldozer benchmark analysisNaN2011/04/02 07:12 AM
            New article: Bulldozer benchmark analysisDavid Kanter2011/04/02 09:05 AM
              New article: Bulldozer benchmark analysisNaN2011/04/02 09:24 AM
                New article: Bulldozer benchmark analysisDavid Kanter2011/04/02 12:36 PM
                  AMD's Open64 compiler ?Alex2011/04/04 02:23 AM
                    AMD's Open64 compiler ?David Kanter2011/04/04 08:19 AM
                      AMD's Open64 compiler ?Richard Cownie2011/04/04 09:52 AM
                        AMD's Open64 compiler ?David Kanter2011/04/04 01:20 PM
                          AMD's Open64 compiler ?none2011/04/04 02:25 PM
                        AMD's Open64 compiler ?gallier22011/04/04 11:06 PM
                          thanks, very interesting (NT)Richard Cownie2011/04/05 03:44 AM
                          AMD's Open64 compiler ?Andreas Stiller2011/04/05 08:23 AM
                        AMD's Open64 compiler ?Vincent Diepeveen2011/04/06 01:11 AM
                      AMD's Open64 compiler ?Vincent Diepeveen2011/04/06 01:01 AM
                        AMD's Open64 compiler ?none2011/04/06 01:41 AM
                          AMD's Open64 compiler ?Vincent Diepeveen2011/04/07 02:02 PM
                            AMD's Open64 compiler ?Brett2011/04/07 04:55 PM
                              AMD's Open64 compiler ?Megol2011/04/09 11:03 AM
                                Clang+LLVMBrett2011/04/11 10:51 PM
                                  Clang+LLVMRob Thorpe2011/04/12 12:52 AM
                                    Clang+LLVMBrett2011/04/12 08:27 PM
                                      I prefer my dollars exponential, not cubic (NT)David Kanter2011/04/12 11:34 PM
                                        I prefer cubic meters of euros over cubic feet of dollarsJouni Osmala2011/04/13 01:13 AM
                                          Life is not about money, but about power (and informational asymetry)?2011/04/18 01:14 AM
                                      Clang+LLVMRob Thorpe2011/04/13 10:37 AM
                                        Clang+LLVMLinus Torvalds2011/04/13 11:47 AM
                                          Clang+LLVMRob Thorpe2011/04/13 01:46 PM
                                            Clang+LLVMLinus Torvalds2011/04/13 04:40 PM
                                              Clang+LLVManon2011/04/13 07:31 PM
                                                Clang+LLVMLinus Torvalds2011/04/14 06:11 AM
                                              Clang+LLVMRob Thorpe2011/04/14 12:33 PM
                                                Clang+LLVMLinus Torvalds2011/04/14 12:48 PM
                                                  Clang+LLVMAndi Kleen2011/04/14 06:38 PM
                                                    Clang+LLVMJouni Osmala2011/04/14 09:54 PM
                                                  Clang+LLVM?2011/04/14 11:56 PM
                                                    Clang+LLVMLinus Torvalds2011/04/15 06:23 AM
                                                      Clang+LLVM?2011/04/15 10:11 PM
                                                        Clang+LLVMLinus Torvalds2011/04/16 08:10 AM
                                                          Clang+LLVMLinus Torvalds2011/04/16 08:32 AM
                                                            Competing for speed?2011/04/16 11:47 PM
                                                            Clang+LLVMEduardoS2011/04/17 12:38 PM
                                                              Clang+LLVMLinus Torvalds2011/04/17 02:00 PM
                                                                Competing for speed?2011/04/17 10:57 PM
                                                                  Competing for speedLinus Torvalds2011/04/18 09:07 AM
                                                                    Competing for speedanon2011/04/18 05:38 PM
                                                                      Competing for speedLinus Torvalds2011/04/18 07:24 PM
                                                                        Competing for speedanon2011/04/19 05:05 AM
                                                                          Competing for speedanon2011/04/21 01:35 PM
                                                                            Competing for speedLinus Torvalds2011/04/22 11:09 AM
                                                                    Competing for speed?2011/04/18 11:45 PM
                                                                      Competing for speedLinus Torvalds2011/04/19 06:41 AM
                                                                        Competing for speedLinus Torvalds2011/04/19 07:46 AM
                                                                        Oh ... the irony?2011/04/19 10:36 AM
                                                                          Oh ... the ironyLinus Torvalds2011/04/19 01:49 PM
                                                                            What's the speedup?iz2011/04/19 10:54 PM
                                                                              What's the speedup?Linus Torvalds2011/04/20 06:48 AM
                                                                            Oh ... the irony?2011/04/23 05:52 AM
                                                                          Oh ... the ironyDougal2011/04/20 06:23 AM
                                                                            Oh ... the irony?2011/04/25 12:05 PM
                                                                              Oh ... the ironyDougal2011/04/26 05:22 AM
                                                                                Oh ... the irony?2011/04/27 12:14 AM
                                                  Clang+LLVMRob Thorpe2011/04/16 01:23 PM
                                                    Clang+LLVManon2011/04/16 03:01 PM
                                                      Clang+LLVMRob Thorpe2011/04/18 04:53 AM
                                                        Clang+LLVManon2011/04/18 05:28 PM
                                        Clang+LLVMPaul2011/04/14 10:43 AM
                                  Clang+LLVMFoo_2011/04/12 05:36 AM
                              AMD's Open64 compiler ?Paul2011/04/10 02:50 AM
                                C++ front-endsDavid Kanter2011/04/10 06:29 AM
                                  C++ front-endsPaul2011/04/10 07:19 AM
                                    C++ front-endsanonymous2011/04/10 07:28 AM
                                      C++ front-endsnone2011/04/10 07:36 AM
                                      C++ front-endsPaul2011/04/10 07:44 AM
                                    C++ front-endsMichael S2011/04/10 07:47 AM
                                      C++ front-endsnone2011/04/10 07:51 AM
                                      C++ front-endsRob Thorpe2011/04/10 11:13 AM
                                        C++ front-endsPaul2011/04/11 10:07 AM
                                AMD's Open64 compiler ?Konrad Schwarz2011/04/14 01:06 AM
                              Testing Clang + LLVMVincent Diepeveen2011/04/20 06:31 AM
                                Does your program crash when compiled with -O0? (NT)Michael S2011/04/20 10:11 AM
                                  Does your program crash when compiled with -O0? (NT)Vincent Diepeveen2011/04/20 11:33 AM
                                Testing Clang + LLVMMegol2011/04/20 11:05 AM
                                  Testing Clang + LLVMVincent Diepeveen2011/04/20 11:43 AM
                                    Testing Clang + LLVMstubar2011/04/20 12:23 PM
                                      Testing Clang + LLVMVincent Diepeveen2011/04/20 12:42 PM
                                        Testing Clang + LLVMRob Thorpe2011/04/20 02:30 PM
                                          Testing Clang + LLVMahnoun2011/04/20 03:13 PM
                                            Testing Clang + LLVMMark Roulo2011/04/20 04:05 PM
                                              Testing Clang + LLVMMegol2011/04/22 04:30 AM
                                    Testing Clang + LLVMahnoun2011/04/20 02:31 PM
                                      Testing Clang + LLVMVincent Diepeveen2011/04/21 02:17 AM
                                        Testing Clang + LLVMnone2011/04/21 02:36 AM
                                        Testing Clang + LLVMahnoun2011/04/21 11:02 AM
                                          Testing Clang + LLVMMichael S2011/04/21 12:01 PM
                                            Testing Clang + LLVMMichael S2011/04/21 03:29 PM
                                              Testing Clang + LLVMBrett2011/04/21 04:49 PM
                                              Testing Clang + LLVMEduardoS2011/04/22 02:22 PM
                                                Testing Clang + LLVMMichael S2011/04/23 09:59 AM
                                            Testing Clang + LLVMDavid Ball2011/04/21 06:43 PM
                                Testing Clang + LLVMBrett2011/04/20 12:54 PM
                            AMD's Open64 compiler ?Megol2011/04/09 10:49 AM
                              Imperative compilers (and the lack of it)?2011/04/10 12:36 AM
                                Imperative compilers (and the lack of it)?2011/04/10 12:43 AM
                                Imperative compilers (and the lack of it)Paul2011/04/10 02:37 AM
                                  Imperative compilers (and the lack of it)?2011/04/10 05:30 AM
                                  Imperative compilers (and the lack of it)Vincent Diepeveen2011/04/10 06:20 AM
                                  Imperative compilers (and the lack of it)Gabriele Svelto2011/04/11 08:24 AM
                                  Imperative compilers (and the lack of it)Michael S2011/04/12 12:12 AM
                                    Imperative compilers (and the lack of it)?2011/04/12 01:41 AM
                                      Imperative compilers (and the lack of it)EduardoS2011/04/12 03:13 PM
                                        U+269BMichael S2011/04/13 12:45 AM
                                          U+269B?2011/04/13 03:52 AM
                                            U+269BBrendan2011/04/20 07:18 AM
                                              U+269BMichael S2011/04/20 09:11 AM
                                                U+269BHeath Provost2011/04/20 10:26 AM
                                                  You are right, Heath. Sorry, Brannon.Michael S2011/04/20 10:42 AM
                                                    Sorry again, Brendan (NT)Michael S2011/04/20 10:43 AM
                                                    Testing western non-us letters.Jouni Osmala2011/04/20 09:41 PM
                                                      Those show up fine to me (NT)David Kanter2011/04/21 07:48 AM
                                                      Testing western non-us letters.Heath Provost2011/04/21 11:15 AM
                                                        Testing western non-us letters.EduardoS2011/04/22 02:14 PM
                                                          Testing western non-us letters.Heath Provost2011/04/22 03:21 PM
                                                            Testing western non-us letters.EduardoS2011/04/22 04:24 PM
                                                            We actually upgraded to 2008 R2David Kanter2011/04/22 10:23 PM
                                                              We actually upgraded to 2008 R2EduardoS2011/04/22 10:33 PM
                                                                We actually upgraded to 2008 R2rwessel2011/04/22 11:11 PM
                                                                  We actually upgraded to 2008 R2David Kanter2011/04/22 11:28 PM
                                                                    We actually upgraded to 2008 R2Ricardo B2011/04/23 09:43 AM
                                                                      We actually upgraded to 2008 R2David Kanter2011/04/23 10:31 AM
                                                                    We actually upgraded to 2008 R2sparkplug2011/04/23 08:25 PM
                                                                      We actually upgraded to 2008 R2David Kanter2011/04/24 03:36 PM
                                                                        We actually upgraded to 2008 R2anon2011/04/24 06:32 PM
                                                                          We actually upgraded to 2008 R2David Kanter2011/04/24 11:12 PM
                                                                        We actually upgraded to 2008 R2EduardoS2011/04/24 08:10 PM
                                                                          We actually upgraded to 2008 R2David Kanter2011/04/24 11:04 PM
                                                                        We actually upgraded to 2008 R2anonymous2011/04/24 08:44 PM
                                                                          We actually upgraded to 2008 R2David Kanter2011/04/24 10:57 PM
                                                                        File size optimizationsjrl2011/04/25 02:03 AM
                                                                          File size optimizationsRicardo B2011/04/25 07:14 AM
                                                                            File size optimizationsjrl2011/04/25 05:52 PM
                                                                          File size optimizationsDavid Kanter2011/04/25 07:49 AM
                                                                            File size optimizationsRicardo B2011/04/25 08:22 AM
                                                                              ContinuationRicardo B2011/04/25 08:25 AM
                                                                              File size optimizationsChristian Packmann2011/04/25 09:56 AM
                                                                                File size optimizationsEduardoS2011/04/25 11:57 AM
                                                                                  File size optimizationsRicardo B2011/04/25 01:01 PM
                                                                            File size optimizationsEduardoS2011/04/25 12:01 PM
                                                                            Javascript is at the root of the latency problemjrl2011/04/25 06:03 PM
                                                                          File size optimizationsEduardoS2011/04/25 12:14 PM
                                                                            File size optimizationsrwessel2011/04/25 12:49 PM
                                                                              File size optimizationsEduardoS2011/04/25 12:57 PM
                                                                            File size optimizationsjrl2011/04/25 05:54 PM
                                                                              A simple approachSilent2011/04/25 09:24 PM
                                                                                Show Thread exists but the default is Yes (NT)Silent2011/04/25 09:26 PM
                                                                          File size optimizationsHeath Provost2011/04/25 02:02 PM
                                    Imperative compilers (and the lack of it)Azazel2011/04/12 08:23 PM
                                Imperative compilers (and the lack of it)none2011/04/10 06:00 AM
                                  Imperative compilers (and the lack of it)dmsc2011/04/10 06:39 AM
                                  Imperative compilers (and the lack of it)EduardoS2011/04/10 12:55 PM
                                compiler failure rateshobold2011/04/10 10:55 AM
                                  compiler failure rates?2011/04/10 11:13 PM
                                    compiler failure rateshobold2011/04/11 05:07 AM
                                Imperative compilers (and the lack of it)EduardoS2011/04/10 12:55 PM
                                  Imperative compilers (and the lack of it)?2011/04/10 11:13 PM
                                    Imperative compilers (and the lack of it)Konrad Schwarz2011/04/14 01:28 AM
                                      Imperative compilers (and the lack of it)gallier22011/04/14 04:47 AM
                                      Imperative compilers (and the lack of it)EduardoS2011/04/14 02:04 PM
                                        Imperative compilers (and the lack of it)Konrad Schwarz2011/04/15 03:43 AM
                                Hard register problemsRob Thorpe2011/04/11 12:14 AM
                                  Hard register problemsMax2011/04/11 03:05 PM
                                    Hard register problemsRob Thorpe2011/04/11 03:28 PM
                                Imperative compilers (and the lack of it)gallier22011/04/11 09:26 AM
                                  I wonder why.j2011/04/11 10:39 PM
                                    I wonder why.?2011/04/11 11:44 PM
                                Imperative compilers (and the lack of it)nksingh2011/04/12 11:00 PM
                                  Imperative compilers (and the lack of it)?2011/04/12 11:23 PM
                              AMD's Open64 compiler ?Vincent Diepeveen2011/04/10 06:06 AM
                                AMD's Open64 compiler ?none2011/04/10 07:23 AM
                                  AMD's Open64 compiler ?Eugene Nalimov2011/04/11 03:57 PM
                                    AMD's Open64 compiler ?EduardoS2011/04/12 02:55 PM
                                AMD's Open64 compiler ?Megol2011/04/10 11:05 PM
                                  AMD's Open64 compiler ?Vincent Diepeveen2011/04/20 12:52 PM
                                    Yap transform? What is it? (NT)Michael S2011/04/21 09:04 AM
                                      Yap transform? What is it? (NT)Vincent Diepeveen2011/04/22 02:56 AM
                    AMD's Open64 compiler ?Vincent Diepeveen2011/04/05 12:02 PM
    New article: Bulldozer benchmark analysisNaN2011/03/30 05:24 PM
      New article: Bulldozer benchmark analysisChris2011/04/05 03:33 AM
  New article: Bulldozer benchmark analysisdc2011/03/30 08:09 AM
    New article: Bulldozer benchmark analysisMS2011/03/30 08:53 AM
      New article: Bulldozer benchmark analysisDaniel Bizo2011/03/31 12:20 AM
        New article: Bulldozer benchmark analysisanun2011/03/31 02:48 AM
    New article: Bulldozer benchmark analysisAlex2011/03/30 09:18 AM
      New article: Bulldozer benchmark analysisIntelUser20002011/03/30 01:19 PM
    New article: Bulldozer benchmark analysisGroo2011/03/31 06:14 AM
      New article: Bulldozer benchmark analysisanon2011/03/31 08:29 AM
        AMD's memory controllers are fineDavid Kanter2011/03/31 09:27 AM
          AMD's memory controllers are fineanon2011/04/05 11:24 AM
            AMD's memory controllers are fineanon2011/04/05 04:52 PM
              AMD's memory controllers are fineVincent Diepeveen2011/04/06 01:36 AM
                AMD's memory controllers are fineanon2011/04/06 06:59 AM
                  AMD's memory controllers are fineVincent Diepeveen2011/04/07 02:15 PM
                    AMD's memory controllers are fineVincent Diepeveen2011/04/07 02:27 PM
                      AMD's memory controllers are fineanon2011/04/07 02:57 PM
                    AMD's memory controllers are fineanon2011/04/07 02:54 PM
        New article: Bulldozer benchmark analysisAzazel2011/03/31 09:54 AM
        New article: Bulldozer benchmark analysisAlex2011/03/31 11:38 PM
      New article: Bulldozer benchmark analysisanon2011/04/01 03:23 AM
    Streaming conflicts can lead to performance degradationsgc2011/04/08 06:35 AM
  AMD's probe filterAndreas Stiller2011/04/05 08:37 AM
  a waste of timeMoritz2011/04/09 07:01 AM
    a waste of timehorsefly2011/04/09 03:34 PM
      a waste of timeDavid Kanter2011/04/09 06:56 PM
        a waste of timehorsefly2011/04/10 05:13 AM
          InvestorsDavid Kanter2011/04/10 06:27 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell tangerine? 🍊