What about dynamic utilization?

Article: ARM Goes 64-bit
By: Exophase (exophase.delete@this.gmail.com), August 17, 2012 9:30 am
Room: Moderated Discussions
Kenneth Jonsson (kj.delete@this.localhost.org) on August 17, 2012 2:41 am wrote:
> Its true one can make use of a large number of GPRs in
> hand-written assembler, I've done a lot of assembler programming on 68k and
> while the 8 addresser register was often enough, having just 8 data register was
> often a limitation. I've also done some assembler optimizations for ARM and my
> handwritten functions definitely uses more GPRs compared to the code generated
> by GCC.
>
> It has been suggested that the load/store nature of a RISC
> architecture will lead to usage of more GPRs compared to x86, but I cannot
> really see that when looking into code generated by compilers and registers are
> faster compared to pushing things to the stack even on the x86 so a optimizing
> compiler want to keep things i registers if possible.
>
> So why isn't more
> register used in code generated by compilers. I looked at fairly large code-base
> and got this register usage
>
> ARM.
> Number of instructions that refers to any
> of the r0-r12 registers: ~170k
> r10-r12 is not used
> r4-r7 are all used 5-6k
> times
> r8,r9 ~10k times
> r3 16k
> r2 23k
> r1 38k
> r0 69k
>
> So it seems that
> there is "enough" register in every single case in this program.
>
> I did the
> same thing on PPC and MIPS using mostly the same source code just to see what
> the compiler would do when there is more registers available. This source do
> include a small embedded OS-kernel and drivers so it is not a true
> apple-to-apple as the PPC card has more devices, hence will be
> bigger.
>
> PPC
> Number of instructions that refer to any register: ~300k
> Every
> single register is actually used.
> r2,r14-r18,r21-r23 are all used 1k-2k
> times
> r8,r10,r12-r13,r19-r20,r24-r26 are all used 3k-6k times
> r6-r7,r9,r27-r29
> are all used ~10k times
> r4-r5,r30-r31 are all used 20k-30k times
> r1 65k
> r0
> 81k
> r3 85k
>
> MIPS is similar, 4 of the temporary registers are used very
> sparsely while t0 and a0 are the most frequently used register (not too
> surprising).

These counts only tell part of the story because code isn't uniformly distributed during execution. Often a large part of the runtime is dominated by a small part of code. A lot of the code may consist of many branches and function calls that are difficult to allocate registers across, but some of the more performance critical sections may contain tight loops that can be more aggressively optimized. The ABIs can also inhibit register allocation when a lot of function calls are going on, especially to library functions. Although this is another strength for more registers: less need to pass part of a function's parameters on the stack.

That assembly coders can often make use of a lot of registers isn't just because they're finding more use than a compiler would (or are wasting registers) but because they've chosen to target critical functions.

> Number of registers is something that people tend to bring up
> quite a lot and suggest more is always better. IA32 is often frowned upon
> because it has so few GPRs.
>
> So to spin the question a bit: what are the
> downsides of a large number of GPRs? It has to be fairly big trade-offs to be
> made somewhere as very few architectures seem to move beyond 32 GPRs and AMD did
> for some reason settle at 16 GPRs when they designed x86_64. Not a perfect
> comparison, but the performance difference between IA32 and x86_64 is very small
> so IA32 cannot be held back too much by its lack of GPRs.
>
> One point of
> reference to compare IA32 vs x86_64 is . Looking at the single core results for
> C and Java show that some things are faster and some things are slower on IA32,
> but the results are with very few exceptions VERY similar.
>
> Sorry for a long
> post, interesting topic i.m.h.o :)

On the other hand, I've encountered programs that see a 15-20% performance delta from x86-32 to x86-64, for instance the emulator cores in bsnes. AFAIK this code doesn't use a lot of pointer heavy data structures so it's less penalized by the increase in pointer size. However it does heavily use software context switches for co-routines.

That's just one example, although it should be pretty representative of most emulators in general and emulators are often performance critical. I'd give it at least as much weight as the stuff they throw on the language shootout.

If the extra registers in x86-64 wasn't worth anything I'm sure AMD wouldn't have done it (and push come to shove, Intel may not have gone along with it). I'm sure they've researched this, just like I'm sure ARM has done its research.

Of course there are real disadvantages, encoding bits leave out other things you can do with the instructions and the registers have a real hardware cost. For a number of OoO implementations that cost will be in the RAT, not in the register file, since the physical register file won't necessarily be increased in size to accommodate more architectural registers (high end x86 implementations have used 128 physical registers for instance, even before x86-64)
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
New Article: ARM Goes 64-bitDavid Kanter08/14/12 12:04 AM
  New Article: ARM Goes 64-bitnone08/14/12 12:44 AM
    New Article: ARM Goes 64-bitDavid Kanter08/14/12 01:04 AM
    MIPS MT-ASEPaul A. Clayton08/14/12 09:01 AM
      MONITOR/MWAITEduardoS08/14/12 10:08 AM
        MWAIT not specifically MTPaul A. Clayton08/14/12 10:36 AM
          MWAIT not specifically MTEduardoS08/15/12 03:16 PM
        MONITOR/MWAITanonymou508/14/12 11:07 AM
          MONITOR/MWAITEduardoS08/15/12 03:20 PM
      MIPS MT-ASErwessel08/14/12 10:14 AM
  New Article: ARM Goes 64-bitSHK08/14/12 02:01 AM
  New Article: ARM Goes 64-bitanon08/14/12 02:37 AM
    New Article: ARM Goes 64-bitRichard Cownie08/14/12 03:57 AM
      New Article: ARM Goes 64-bitanon08/14/12 04:29 AM
      New Article: ARM Goes 64-bitnone08/14/12 04:44 AM
        New Article: ARM Goes 64-bitanon08/14/12 05:28 AM
          New Article: ARM Goes 64-bitanon08/14/12 05:32 AM
            New Article: ARM Goes 64-bitEduardoS08/14/12 06:06 AM
          New Article: ARM Goes 64-bitnone08/14/12 05:40 AM
            AArch64 select better than cmovPaul A. Clayton08/14/12 06:08 AM
            New Article: ARM Goes 64-bitanon08/14/12 06:12 AM
              New Article: ARM Goes 64-bitnone08/14/12 06:25 AM
                Predicated ld/store are usefulPaul A. Clayton08/14/12 06:48 AM
                  Predicated ld/store are usefulnone08/14/12 06:56 AM
                    Predicated ld/store are usefulanon08/14/12 07:07 AM
                    Predicated stores might not be that badPaul A. Clayton08/14/12 07:27 AM
                      Predicated stores might not be that badDavid Kanter08/15/12 01:14 AM
                        Predicated stores might not be that badMichael S08/15/12 11:41 AM
                        Predicated stores might not be that badR Byron08/17/12 04:09 AM
                New Article: ARM Goes 64-bitanon08/14/12 06:54 AM
                  New Article: ARM Goes 64-bitnone08/14/12 07:04 AM
                    New Article: ARM Goes 64-bitanon08/14/12 07:43 AM
          New Article: ARM Goes 64-bitEduardoS08/14/12 06:07 AM
            New Article: ARM Goes 64-bitanon08/14/12 06:20 AM
              New Article: ARM Goes 64-bitnone08/14/12 06:29 AM
                New Article: ARM Goes 64-bitanon08/14/12 07:00 AM
            New Article: ARM Goes 64-bitMichael S08/14/12 03:43 PM
        New Article: ARM Goes 64-bitRichard Cownie08/14/12 06:53 AM
          OT: Conrad's "Youth"Richard Cownie08/14/12 07:20 AM
      New Article: ARM Goes 64-bitEduardoS08/14/12 06:04 AM
        New Article: ARM Goes 64-bitmpx08/14/12 08:59 AM
          New Article: ARM Goes 64-bitAntti-Ville Tuunainen08/14/12 09:16 AM
        New Article: ARM Goes 64-bitanonymou508/14/12 11:03 AM
          New Article: ARM Goes 64-bitname9911/17/12 03:31 PM
            Microarchitecting a counter registerPaul A. Clayton11/17/12 07:37 PM
    New Article: ARM Goes 64-bitbakaneko08/14/12 04:21 AM
      New Article: ARM Goes 64-bitname9911/17/12 03:40 PM
        New Article: ARM Goes 64-bitEduardoS11/17/12 04:52 PM
        New Article: ARM Goes 64-bitDoug S11/17/12 05:48 PM
        New Article: ARM Goes 64-bitbakaneko11/18/12 05:40 PM
          New Article: ARM Goes 64-bitWilco11/19/12 07:59 AM
            New Article: ARM Goes 64-bitEduardoS11/19/12 08:23 AM
              New Article: ARM Goes 64-bitWilco11/19/12 09:31 AM
                Downloading µarch-specific binaries?Paul A. Clayton11/19/12 11:21 AM
                New Article: ARM Goes 64-bitEduardoS11/19/12 11:41 AM
                  New Article: ARM Goes 64-bitWilco11/21/12 07:44 AM
                    JIT vs. static compilation (Was: New Article: ARM Goes 64-bit)VMguy11/22/12 03:21 AM
                      JIT vs. static compilation (Was: New Article: ARM Goes 64-bit)David Kanter11/22/12 12:12 PM
                        JIT vs. static compilation (Was: New Article: ARM Goes 64-bit)Gabriele Svelto11/23/12 03:50 AM
                    New Article: ARM Goes 64-bitEduardoS11/23/12 10:09 AM
                      New Article: ARM Goes 64-bitEBFE11/26/12 01:24 AM
                        New Article: ARM Goes 64-bitGabriele Svelto11/26/12 03:33 AM
                          New Article: ARM Goes 64-bitEBFE11/27/12 11:17 PM
                            New Article: ARM Goes 64-bitGabriele Svelto11/28/12 02:32 AM
                        New Article: ARM Goes 64-bitEduardoS11/26/12 12:16 PM
                          New Article: ARM Goes 64-bitEBFE11/28/12 12:33 AM
                            New Article: ARM Goes 64-bitEduardoS11/28/12 05:53 AM
                              New Article: ARM Goes 64-bitMichael S11/28/12 06:15 AM
                                New Article: ARM Goes 64-bitEduardoS11/28/12 07:33 AM
                                  New Article: ARM Goes 64-bitMichael S11/28/12 09:16 AM
                                    New Article: ARM Goes 64-bitEduardoS11/28/12 09:53 AM
                                    New Article: ARM Goes 64-bitEugene Nalimov11/28/12 05:58 PM
                                      Amazing!EduardoS11/28/12 07:25 PM
                                        Amazing! (non-italic response)EduardoS11/28/12 07:25 PM
                                        Amazing!EBFE11/28/12 08:20 PM
                                          Undefined behaviour doubles downEduardoS11/28/12 09:10 PM
                              New Article: ARM Goes 64-bitEBFE11/28/12 07:54 PM
                                New Article: ARM Goes 64-bitEduardoS11/28/12 09:21 PM
                Have you heard of Transmeta?David Kanter11/19/12 03:47 PM
            New Article: ARM Goes 64-bitbakaneko11/19/12 09:08 AM
            New Article: ARM Goes 64-bitDavid Kanter11/19/12 03:40 PM
              Semantic Dictionary EncodingRay11/19/12 10:37 PM
              New Article: ARM Goes 64-bitRohit11/20/12 04:48 PM
                New Article: ARM Goes 64-bitDavid Kanter11/20/12 11:07 PM
                  New Article: ARM Goes 64-bitWilco11/21/12 06:41 AM
                    New Article: ARM Goes 64-bitDavid Kanter11/21/12 10:12 AM
                    A JIT exampleMark Roulo11/21/12 10:30 AM
                      A JIT exampleWilco11/21/12 07:04 PM
                        A JIT examplerwessel11/21/12 09:05 PM
                        A JIT exampleGabriele Svelto11/23/12 03:53 AM
                        A JIT exampleEduardoS11/23/12 10:13 AM
                          A JIT exampleWilco11/23/12 01:41 PM
                            A JIT exampleEduardoS11/23/12 02:06 PM
                            A JIT exampleGabriele Svelto11/23/12 04:09 PM
                              A JIT exampleSymmetry11/26/12 05:58 AM
            New Article: ARM Goes 64-bitRay11/19/12 10:27 PM
    New Article: ARM Goes 64-bitDavid Kanter08/14/12 09:11 AM
  v7-M is Thumb-onlyPaul A. Clayton08/14/12 06:58 AM
  Minor suggested correctionPaul A. Clayton08/14/12 08:33 AM
    Minor suggested correctionanon08/14/12 08:57 AM
  New Article: ARM Goes 64-bitExophase08/14/12 08:33 AM
    New Article: ARM Goes 64-bitDavid Kanter08/14/12 09:16 AM
      New Article: ARM Goes 64-bitjigal08/15/12 01:49 PM
  Correction re ARM and BBC MicroPaul08/14/12 08:59 PM
    Correction re ARM and BBC MicroPer Hesselgren08/15/12 03:27 AM
  Memory BW so lowPer Hesselgren08/15/12 03:14 AM
    Memory BW so lownone08/15/12 11:16 AM
  New Article: ARM Goes 64-bitdado08/15/12 10:25 AM
  Number of GPRsKenneth Jonsson08/16/12 02:35 PM
    Number of GPRsExophase08/16/12 02:52 PM
      Number of GPRsKenneth Jonsson08/17/12 02:41 AM
        Ooops, missing link...Kenneth Jonsson08/17/12 02:44 AM
        64-bit pointers eat some performancePaul A. Clayton08/17/12 06:19 AM
          64-bit pointers eat some performancebakaneko08/17/12 08:37 AM
            Brute force seems to workPaul A. Clayton08/17/12 10:08 AM
              Brute force seems to workbakaneko08/17/12 11:15 AM
          64-bit pointers eat some performanceRichard Cownie08/17/12 08:46 AM
            Pointer compression is atypicalPaul A. Clayton08/17/12 10:43 AM
              Pointer compression is atypicalRichard Cownie08/17/12 12:57 PM
                Pointer compression is atypicalHoward Chu08/22/12 10:17 PM
                  Pointer compression is atypicalRichard Cownie08/23/12 04:48 AM
                    Pointer compression is atypicalHoward Chu08/23/12 06:51 AM
              Pointer compression is atypicalWilco08/17/12 02:41 PM
                Pointer compression is atypicalRichard Cownie08/17/12 04:13 PM
                  Pointer compression is atypicalRicardo B08/19/12 10:44 AM
                  Pointer compression is atypicalHoward Chu08/22/12 10:08 PM
                    Unified libraries?Paul A. Clayton08/23/12 07:49 AM
                    Pointer compression is atypicalRichard Cownie08/23/12 08:44 AM
                      Pointer compression is atypicalHoward Chu08/23/12 05:17 PM
                        Pointer compression is atypicalanon08/23/12 08:15 PM
                          Pointer compression is atypicalHoward Chu08/23/12 09:33 PM
            64-bit pointers eat some performanceFoo_08/18/12 12:09 PM
              64-bit pointers eat some performanceRichard Cownie08/18/12 05:25 PM
                64-bit pointers eat some performanceRichard Cownie08/18/12 05:32 PM
            Page-related benefit of small pointersPaul A. Clayton08/23/12 08:36 AM
        Number of GPRsWilco08/17/12 06:31 AM
          Number of GPRsKenneth Jonsson08/17/12 11:54 AM
            Number of GPRsExophase08/17/12 12:44 PM
              Number of GPRsKenneth Jonsson08/17/12 01:22 PM
                Number of GPRsWilco08/17/12 02:53 PM
        What about dynamic utilization?Exophase08/17/12 09:30 AM
          Compiler vs. assembly aliasing knowledge?Paul A. Clayton08/17/12 10:20 AM
            Compiler vs. assembly aliasing knowledge?Exophase08/17/12 11:09 AM
            Compiler vs. assembly aliasing knowledge?anon08/18/12 02:23 AM
              Compiler vs. assembly aliasing knowledge?Ricardo B08/19/12 11:02 AM
                Compiler vs. assembly aliasing knowledge?anon08/19/12 06:07 PM
                  Compiler vs. assembly aliasing knowledge?Ricardo B08/19/12 07:26 PM
                    Compiler vs. assembly aliasing knowledge?anon08/19/12 10:03 PM
                      Compiler vs. assembly aliasing knowledge?anon08/20/12 01:59 AM
        Number of GPRsDavid Kanter08/17/12 12:46 PM
          RAT issues as part of reason 1Paul A. Clayton08/17/12 02:18 PM
        Number of GPRsname9911/17/12 06:37 PM
          Large ARFs increase renaming costPaul A. Clayton11/17/12 09:23 PM
    Number of GPRsDavid Kanter08/16/12 03:31 PM
    Number of GPRsRichard Cownie08/16/12 05:17 PM
    32 GPRs ~2-3%Paul A. Clayton08/16/12 06:27 PM
      Oops, Message-ID: aaed6e38-c7bd-467e-ba41-f40cf1020e5e@googlegroups.com (NT)Paul A. Clayton08/16/12 06:29 PM
      32 GPRs ~2-3%Exophase08/16/12 10:06 PM
        R31 as SP/zero is kind of neat (NT)Paul A. Clayton08/17/12 06:23 AM
        32 GPRs ~2-3%rwessel08/17/12 08:24 AM
          32 GPRs ~2-3%Exophase08/17/12 09:16 AM
            32 GPRs ~2-3%Max08/17/12 04:19 PM
      32 GPRs ~2-3%name9911/17/12 07:43 PM
    Number of GPRsmpx08/17/12 01:11 AM
      Latency and powerPaul A. Clayton08/17/12 06:54 AM
    Number of GPRsbakaneko08/17/12 03:09 AM
  New Article: ARM Goes 64-bitSteve08/17/12 02:12 PM
    New Article: ARM Goes 64-bitDavid Kanter08/19/12 12:42 PM
      New Article: ARM Goes 64-bitDoug S08/19/12 02:02 PM
      New Article: ARM Goes 64-bitAnon08/19/12 07:16 PM
      New Article: ARM Goes 64-bitSteve08/30/12 07:51 AM
  Scalar vs Vector registersRobert David Graham08/19/12 05:19 PM
    Scalar vs Vector registersDavid Kanter08/19/12 05:29 PM
  New Article: ARM Goes 64-bitBaserock ARM servers08/21/12 04:13 PM
    Baserock ARM serversSysanon08/21/12 04:14 PM
    A-15 virtualization and LPAE?Paul A. Clayton08/21/12 06:13 PM
      A-15 virtualization and LPAE?Anon08/21/12 07:13 PM
        Half-depth advantages?Paul A. Clayton08/21/12 08:42 PM
          Half-depth advantages?Anon08/22/12 03:33 PM
            Thanks for the information (NT)Paul A. Clayton08/22/12 04:04 PM
      A-15 virtualization and LPAE?C. Ladisch08/23/12 11:12 AM
        A-15 virtualization and LPAE?Paul08/23/12 03:17 PM
        Excessive pessimismPaul A. Clayton08/23/12 04:08 PM
          Excessive pessimismDavid Kanter08/23/12 05:05 PM
    New Article: ARM Goes 64-bitMichael S08/22/12 07:12 AM
      BTW, Baserock==product, Codethink==company (NT)Paul A. Clayton08/22/12 08:56 AM
  New Article: ARM Goes 64-bitReinoud Zandijk08/21/12 11:27 PM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell blue?