7 ms·

Summarizing the interesting bit for non-compiler-specialists: This is an improvement to a part of GCC's compiler optimizations which will make compiled-with-GCC

Summarizing the interesting bit for non-compiler-specialists: This is an improvement to a part of GCC's compiler optimizations which will make compiled-with-GCC programs slightly faster. The SPEC INT benchmark mean score went from 3717 to 3730 (0.5%), and SPEC FP from 4752 to 4775 (0.34%). Which doesn't seem like much, but making a good compiler in practice tends to mean stacking lots and lots of improvements of this magnitude on top of each other.

chrisseaton 11y ago
But is 0.5% statistically significant? How do I know it isn't natural variation or measurement error? He must have the most stable setup in the world if he has confidence in those results.
- DannyBee 11y ago
  SPEC is a geometric mean, so 0.5% is very significant. (Assuming you are accurately measuring the parts)
  - chrisseaton 11y ago
    Does using the geometric mean do something to the size of the error? If the geometric mean of the original implementation varies 1% under normal conditions then this is still not a significant result is it? I have benchmarks where my geomean easily varies 1% even in the most stable conditions I know to give it.
    - exDM69 11y ago
      It's a benchmark engineered to measure the performance of compilers and CPUs. It's healthy to be skeptical but in general if you get a better score, it's a real improvement. You should be able to safely exclude measurement errors and other mundane issues. These benchmarks have a slight issue that they might not represent real world performance, but compilers and CPUs tend to be optimized against them - because their timing and measurement is generally more reliable than ad-hoc real world measurement. It might lead to slight "over-optimization" which doesn't reflect real performance but at least it's reliable and repeatable.
  - gus_massa 11y ago
    To know if it's statically significant,it's not important if the benchmark use the arithmetic or the geometric mean. Did they repeat the benchmark? How many times? Is that information available? (Perhaps it's available for the master branch and not for the new commit).
    - DannyBee 11y ago
      "To know if it's statically significant,it's not important if the benchmark use the arithmetic or the geometric mean. " Of course, but i was answering whether 0.5% matters, not whether it is valid. This is explicitly why i said "assuming you are accurately measuring the parts". It seems you are just making a side point for the heck of it, despite completely agreeing with what i said.
      - chrisseaton 11y ago
        I'm not sure what you mean by 'valid' or 'matters' or the difference between them though. I'm asking if it is more likely that the 0.5% difference is due to the change in the implementation than normal variation in the original implementation. I don't know how you can say that this is the case based only on the information that it's a geometric mean. You need to know the variation in the reference implementation, and you don't. The facts that he measured an increase of performance of 0.5% and he used a geometric mean are not enough information to say that it's statistically significant.
        DannyBee 11y ago
        I understand exactly what you are saying. I'm saying you are being super pedantic here, since it's pretty obvious I meant valid to mean statistically significant, and matters took mean whether improving spec by 0.5% translates into real world application performance benefit. I'm not sure why you want to be so pedantic, but it doesn't seem worthwhile to me since nobody was confused.
        zodiac 11y ago
- the8472 11y ago
  I assume that after years of optimization work you're not going to get more out of a small tweak to one part of the compiler unless you harness new CPU instructions.
- iheartmemcache 11y ago
  You know way way more about compilers than I do, but the key is in this: Furthermore, there exists a path from h to a use of v which does not go through d. For every node p in the loop, since the loop is strongly connected and node is a component of the CFG, there exists a path, consisting only of nodes of L from p to h. Concatenating these two paths proves that v is live-in and live-out of p. I have no idea how to formally prove that assertion (it's logically sound, afaik, someone like Walter Bright will have to jump in, he probably knows GCC inside-out), but if correct, the code change is correct semantically. The computational gain can be trivially exhibited (again, no idea how to 'prove' it) by throwing code at it and seeing the RTL emitted. (I agree it lacks 'rigor' in the academic sense, but based on the list to which he submitted his code-change and the fact that he's at Xilinx and sent it to his colleagues makes me confident enough that the change is sound.) Edit: Fair enough. You've convinced me. (Though the SPEC2000 resulting numbers still have me in the 'maybe' category - your argument below is more sound than my ignorance.) Anyone reading this should defer to Danny's posts rather than mine for the time being (after a day or two, read the resulting responses to see the expert analyses on the mailing list). Keeping this post intact solely for continuity for the reader. RE: Walter - I'd assume he'd keep up with the competition and know enough about the RTL as it hasn't changed extensively in the last decade to make a valid opinion w/r/t changes. Either way, gracefully upvoting Danny's two posts ;)
  - DannyBee 11y ago
    "someone like Walter Bright will have to jump in, he probably knows GCC inside-out" FWIW, walter has not contributed a single patch to GCC i can find in the past 10 years (It may be longer, i stopped looking) :) Not that this makes him bad in any way, mind you, he's just not who i'd go to for gcc expertise. That would be someone like Richard Henderson, who is pretty much never talked about, but has touched pretty much all parts of the compiler and always does great work. Also note that embedded developers in general do not have a good reputation among compiler people for doing "sound work". Again, not that this is a characterization of their engineering skills - They often have very tight deadlines, so the amount of time they can spend figuring out things instead of applying bandaids tends to be pretty limited. So they patch something, submit it, and move on.
johnbender 11y ago
To add just a bit more: When assignments to variables happen in the context of a procedure call it's preferable to use registers for performance reasons since memory access (even cache) is much slower by comparison. Unfortunately registers are a limited resource, so we would like to reuse them when a variable, though in scope, isn't "live" or in use. This whole area of optimization is called register allocation and the optimal solution reduces to the graph coloring problem which is NP-complete. As a result there are a myriad of approximation algorithms with knobs to turn to spend a bit more complexity for a bit more accuracy. This patch appears to be the turning of one such knob.
- DannyBee 11y ago
  Graph coloring is a small part of register allocation, and not the hard part. It's also not NP complete on interference graphs built from SSA form, which most modern compilers (though not GCC) use in the backend. In fact, it's linear time optimally solvable if you build ssi form, which can be done in linear time. The optimal spilling and rematerialization problems have not yet shown to be solvable in polynomial time on these forms, but folks are working on it.
  - andrewflnr 11y ago
    The bit about SSI form is news to me. I have more research to do.
    - DannyBee 11y ago
      See http://web.cs.ucla.edu/~palsberg/paper/PereiraPalsberg08.pdf http://web.cs.ucla.edu/~palsberg/paper/PereiraPalsberg08.pdf Look at page 10. (Fernando was once my intern, back in the day ;P)
  - Someone 11y ago
    Also, something being NP complete does not say much about whether it is hard to do in the real world. Exponential is fine, as long as the constant, the value of n, or the exponent aren't too large. Here, real world cases likely will have only a fairly limited number of registers and only a limited (but potentially very long) length of code to optimize across. It would be interesting (and quite an accomplishment, even if doesn't give significant performance gains) to combine this with whole-program optimization (imagine optimizing the calling conventions on a per-function basis, and considering generating code for a function twice, with different calling conventions) and/or with the code that decides what functions to inline, but AFAIK, nobody does that.
amelius 11y ago
Could this also be improved in LLVM? Or is this specific to GCC?

chrisseaton 11y ago

But is 0.5% statistically significant? How do I know it isn't natural variation or measurement error? He must have the most stable setup in the world if he has confidence in those results.

DannyBee 11y ago
SPEC is a geometric mean, so 0.5% is very significant. (Assuming you are accurately measuring the parts)
- chrisseaton 11y ago
  Does using the geometric mean do something to the size of the error? If the geometric mean of the original implementation varies 1% under normal conditions then this is still not a significant result is it? I have benchmarks where my geomean easily varies 1% even in the most stable conditions I know to give it.
  - exDM69 11y ago
    It's a benchmark engineered to measure the performance of compilers and CPUs. It's healthy to be skeptical but in general if you get a better score, it's a real improvement. You should be able to safely exclude measurement errors and other mundane issues. These benchmarks have a slight issue that they might not represent real world performance, but compilers and CPUs tend to be optimized against them - because their timing and measurement is generally more reliable than ad-hoc real world measurement. It might lead to slight "over-optimization" which doesn't reflect real performance but at least it's reliable and repeatable.
- gus_massa 11y ago
  To know if it's statically significant,it's not important if the benchmark use the arithmetic or the geometric mean. Did they repeat the benchmark? How many times? Is that information available? (Perhaps it's available for the master branch and not for the new commit).
  - DannyBee 11y ago
    "To know if it's statically significant,it's not important if the benchmark use the arithmetic or the geometric mean. " Of course, but i was answering whether 0.5% matters, not whether it is valid. This is explicitly why i said "assuming you are accurately measuring the parts". It seems you are just making a side point for the heck of it, despite completely agreeing with what i said.
    - chrisseaton 11y ago
      I'm not sure what you mean by 'valid' or 'matters' or the difference between them though. I'm asking if it is more likely that the 0.5% difference is due to the change in the implementation than normal variation in the original implementation. I don't know how you can say that this is the case based only on the information that it's a geometric mean. You need to know the variation in the reference implementation, and you don't. The facts that he measured an increase of performance of 0.5% and he used a geometric mean are not enough information to say that it's statistically significant.
      - DannyBee 11y ago
        I understand exactly what you are saying. I'm saying you are being super pedantic here, since it's pretty obvious I meant valid to mean statistically significant, and matters took mean whether improving spec by 0.5% translates into real world application performance benefit. I'm not sure why you want to be so pedantic, but it doesn't seem worthwhile to me since nobody was confused.
        zodiac 11y ago
the8472 11y ago
I assume that after years of optimization work you're not going to get more out of a small tweak to one part of the compiler unless you harness new CPU instructions.
iheartmemcache 11y ago
You know way way more about compilers than I do, but the key is in this: Furthermore, there exists a path from h to a use of v which does not go through d. For every node p in the loop, since the loop is strongly connected and node is a component of the CFG, there exists a path, consisting only of nodes of L from p to h. Concatenating these two paths proves that v is live-in and live-out of p. I have no idea how to formally prove that assertion (it's logically sound, afaik, someone like Walter Bright will have to jump in, he probably knows GCC inside-out), but if correct, the code change is correct semantically. The computational gain can be trivially exhibited (again, no idea how to 'prove' it) by throwing code at it and seeing the RTL emitted. (I agree it lacks 'rigor' in the academic sense, but based on the list to which he submitted his code-change and the fact that he's at Xilinx and sent it to his colleagues makes me confident enough that the change is sound.) Edit: Fair enough. You've convinced me. (Though the SPEC2000 resulting numbers still have me in the 'maybe' category - your argument below is more sound than my ignorance.) Anyone reading this should defer to Danny's posts rather than mine for the time being (after a day or two, read the resulting responses to see the expert analyses on the mailing list). Keeping this post intact solely for continuity for the reader. RE: Walter - I'd assume he'd keep up with the competition and know enough about the RTL as it hasn't changed extensively in the last decade to make a valid opinion w/r/t changes. Either way, gracefully upvoting Danny's two posts ;)
- DannyBee 11y ago
  "someone like Walter Bright will have to jump in, he probably knows GCC inside-out" FWIW, walter has not contributed a single patch to GCC i can find in the past 10 years (It may be longer, i stopped looking) :) Not that this makes him bad in any way, mind you, he's just not who i'd go to for gcc expertise. That would be someone like Richard Henderson, who is pretty much never talked about, but has touched pretty much all parts of the compiler and always does great work. Also note that embedded developers in general do not have a good reputation among compiler people for doing "sound work". Again, not that this is a characterization of their engineering skills - They often have very tight deadlines, so the amount of time they can spend figuring out things instead of applying bandaids tends to be pretty limited. So they patch something, submit it, and move on.

johnbender 11y ago

To add just a bit more: When assignments to variables happen in the context of a procedure call it's preferable to use registers for performance reasons since memory access (even cache) is much slower by comparison. Unfortunately registers are a limited resource, so we would like to reuse them when a variable, though in scope, isn't "live" or in use. This whole area of optimization is called register allocation and the optimal solution reduces to the graph coloring problem which is NP-complete. As a result there are a myriad of approximation algorithms with knobs to turn to spend a bit more complexity for a bit more accuracy. This patch appears to be the turning of one such knob.

DannyBee 11y ago
Graph coloring is a small part of register allocation, and not the hard part. It's also not NP complete on interference graphs built from SSA form, which most modern compilers (though not GCC) use in the backend. In fact, it's linear time optimally solvable if you build ssi form, which can be done in linear time. The optimal spilling and rematerialization problems have not yet shown to be solvable in polynomial time on these forms, but folks are working on it.
- andrewflnr 11y ago
  The bit about SSI form is news to me. I have more research to do.
  - DannyBee 11y ago
    See http://web.cs.ucla.edu/~palsberg/paper/PereiraPalsberg08.pdf http://web.cs.ucla.edu/~palsberg/paper/PereiraPalsberg08.pdf Look at page 10. (Fernando was once my intern, back in the day ;P)
- Someone 11y ago
  Also, something being NP complete does not say much about whether it is hard to do in the real world. Exponential is fine, as long as the constant, the value of n, or the exponent aren't too large. Here, real world cases likely will have only a fairly limited number of registers and only a limited (but potentially very long) length of code to optimize across. It would be interesting (and quite an accomplishment, even if doesn't give significant performance gains) to combine this with whole-program optimization (imagine optimizing the calling conventions on a per-function basis, and considering generating code for a function twice, with different calling conventions) and/or with the code that decides what functions to inline, but AFAIK, nobody does that.

amelius 11y ago

Could this also be improved in LLVM? Or is this specific to GCC?