Changeset 203318 in webkit


Ignore:
Timestamp:
Jul 15, 2016 7:12:31 PM (8 years ago)
Author:
commit-queue@webkit.org
Message:

[JSC] Change some parameters based on a random search
​https://bugs.webkit.org/show_bug.cgi?id=158514

Patch by Benjamin Poulain <​bpoulain@apple.com> on 2016-07-15
Reviewed by Saam Barati.

Source/JavaScriptCore:

  • bytecode/CodeBlock.cpp:

(JSC::CodeBlock::optimizationThresholdScalingFactor):

  • runtime/Options.h:

Tools:

  • Scripts/run-jsc-stress-tests:
Location:
trunk
Files:
5 edited

Legend:

Unmodified
Added
Removed

  • trunk/Source/JavaScriptCore/ChangeLog

    r203315 r203318  
     12016-07-15  Benjamin Poulain  <bpoulain@apple.com>
     2
     3        [JSC] Change some parameters based on a random search
     4        https://bugs.webkit.org/show_bug.cgi?id=158514
     5
     6        Reviewed by Saam Barati.
     7
     8        * bytecode/CodeBlock.cpp:
     9        (JSC::CodeBlock::optimizationThresholdScalingFactor):
     10        * runtime/Options.h:
     11
    1122016-07-15  Mark Lam  <mark.lam@apple.com>
    213

  • trunk/Source/JavaScriptCore/bytecode/CodeBlock.cpp

    r203239 r203318  
    36943694double CodeBlock::optimizationThresholdScalingFactor()
    36953695{
    3696     // This expression arises from doing a least-squares fit of
    3697     //
    3698     // F[x_] =: a * Sqrt[x + b] + Abs[c * x] + d
    3699     //
    3700     // against the data points:
    3701     //
    3702     //    x       F[x_]
    3703     //    10       0.9          (smallest reasonable code block)
    3704     //   200       1.0          (typical small-ish code block)
    3705     //   320       1.2          (something I saw in 3d-cube that I wanted to optimize)
    3706     //  1268       5.0          (something I saw in 3d-cube that I didn't want to optimize)
    3707     //  4000       5.5          (random large size, used to cause the function to converge to a shallow curve of some sort)
    3708     // 10000       6.0          (similar to above)
    3709     //
    3710     // I achieve the minimization using the following Mathematica code:
    3711     //
    3712     // MyFunctionTemplate[x_, a_, b_, c_, d_] := a*Sqrt[x + b] + Abs[c*x] + d
    3713     //
    3714     // samples = {{10, 0.9}, {200, 1}, {320, 1.2}, {1268, 5}, {4000, 5.5}, {10000, 6}}
    3715     //
    3716     // solution =
    3717     //     Minimize[Plus @@ ((MyFunctionTemplate[#[[1]], a, b, c, d] - #[[2]])^2 & /@ samples),
    3718     //         {a, b, c, d}][[2]]
    3719     //
    3720     // And the code below (to initialize a, b, c, d) is generated by:
    3721     //
    3722     // Print["const double " <> ToString[#[[1]]] <> " = " <>
    3723     //     If[#[[2]] < 0.00001, "0.0", ToString[#[[2]]]] <> ";"] & /@ solution
    3724     //
    3725     // We've long known the following to be true:
    3726     // - Small code blocks are cheap to optimize and so we should do it sooner rather
    3727     //   than later.
    3728     // - Large code blocks are expensive to optimize and so we should postpone doing so,
    3729     //   and sometimes have a large enough threshold that we never optimize them.
    3730     // - The difference in cost is not totally linear because (a) just invoking the
    3731     //   DFG incurs some base cost and (b) for large code blocks there is enough slop
    3732     //   in the correlation between instruction count and the actual compilation cost
    3733     //   that for those large blocks, the instruction count should not have a strong
    3734     //   influence on our threshold.
    3735     //
    3736     // I knew the goals but I didn't know how to achieve them; so I picked an interesting
    3737     // example where the heuristics were right (code block in 3d-cube with instruction
    3738     // count 320, which got compiled early as it should have been) and one where they were
    3739     // totally wrong (code block in 3d-cube with instruction count 1268, which was expensive
    3740     // to compile and didn't run often enough to warrant compilation in my opinion), and
    3741     // then threw in additional data points that represented my own guess of what our
    3742     // heuristics should do for some round-numbered examples.
    3743     //
    3744     // The expression to which I decided to fit the data arose because I started with an
    3745     // affine function, and then did two things: put the linear part in an Abs to ensure
    3746     // that the fit didn't end up choosing a negative value of c (which would result in
    3747     // the function turning over and going negative for large x) and I threw in a Sqrt
    3748     // term because Sqrt represents my intution that the function should be more sensitive
    3749     // to small changes in small values of x, but less sensitive when x gets large.
    3750    
    3751     // Note that the current fit essentially eliminates the linear portion of the
    3752     // expression (c == 0.0).
    3753     const double a = 0.061504;
    3754     const double b = 1.02406;
    3755     const double c = 0.0;
    3756     const double d = 0.825914;
    3757    
     3696    // We want a good threshold based on the instruction count.
     3697    // Here, we are trying to optimize the following formula:
     3698    //     F[x_] =: a * Sqrt[x + b] + Abs[c * x] + d
     3699    // The parameters were chosen by testing random values
     3700    // between 1 and 2 and keeping the best combination.
     3701    const double a = Options::optimizationThresholdScalingFactorA();
     3702    const double b = Options::optimizationThresholdScalingFactorB();
     3703    const double c = Options::optimizationThresholdScalingFactorC();
     3704    const double d = Options::optimizationThresholdScalingFactorD();
     3705
    37583706    double instructionCount = this->instructionCount();
    37593707   

  • trunk/Source/JavaScriptCore/runtime/Options.h

    r203142 r203318  
    9898    unsigned m_highLimit;
    9999};
     100
     101#if CPU(X86_64)
     102constexpr int32_t archThresholdForJITAfterWarmUp = 610;
     103constexpr int32_t archThresholdForJITSoon = 89;
     104constexpr int32_t archThresholdForOptimizeAfterWarmUp = 864;
     105constexpr int32_t archThresholdForOptimizeAfterLongWarmUp = 1489;
     106constexpr int32_t archThresholdForOptimizeSoon = 864;
     107constexpr int32_t archExecutionCounterIncrementForLoop = 2;
     108constexpr int32_t archExecutionCounterIncrementForEntry = 16;
     109constexpr int32_t archThresholdForFTLOptimizeAfterWarmUp = 109160;
     110constexpr int32_t archThresholdForFTLOptimizeSoon = 825;
     111constexpr int32_t archFtlTierUpCounterIncrementForLoop = 8;
     112constexpr int32_t archFtlTierUpCounterIncrementForReturn = 27;
     113constexpr unsigned archFtlOSREntryRetryThreshold = 109;
     114constexpr double archOptimizationThresholdScalingFactorA = 0.0258587392234135;
     115constexpr double archOptimizationThresholdScalingFactorB = 1.2428014544978696;
     116constexpr double archOptimizationThresholdScalingFactorC = 0.0013276440240339;
     117constexpr double archOptimizationThresholdScalingFactorD = 1.3130654609331458;
     118constexpr int32_t archEvalThresholdMultiplier = 12;
     119#else
     120constexpr int32_t archThresholdForJITAfterWarmUp = 500;
     121constexpr int32_t archThresholdForJITSoon = 100;
     122constexpr int32_t archThresholdForOptimizeAfterWarmUp = 1000;
     123constexpr int32_t archThresholdForOptimizeAfterLongWarmUp = 1000;
     124constexpr int32_t archThresholdForOptimizeSoon = 1000;
     125constexpr int32_t archExecutionCounterIncrementForLoop = 1;
     126constexpr int32_t archExecutionCounterIncrementForEntry = 15;
     127constexpr int32_t archThresholdForFTLOptimizeAfterWarmUp = 100000;
     128constexpr int32_t archThresholdForFTLOptimizeSoon = 1000;
     129constexpr int32_t archFtlTierUpCounterIncrementForLoop = 1;
     130constexpr int32_t archFtlTierUpCounterIncrementForReturn = 15;
     131constexpr unsigned archFtlOSREntryRetryThreshold = 100;
     132constexpr double archOptimizationThresholdScalingFactorA = 0.061504;
     133constexpr double archOptimizationThresholdScalingFactorB = 1.02406;
     134constexpr double archOptimizationThresholdScalingFactorC = 0.0;
     135constexpr double archOptimizationThresholdScalingFactorD = 0.825914;
     136constexpr int32_t archEvalThresholdMultiplier = 10;
     137#endif
    100138
    101139typedef OptionRange optionRange;
    … …  
    254292    v(double, jitPolicyScale, 1.0, Normal, "scale JIT thresholds to this specified ratio between 0.0 (compile ASAP) and 1.0 (compile like normal).") \
    255293    v(bool, forceEagerCompilation, false, Normal, nullptr) \
    256     v(int32, thresholdForJITAfterWarmUp, 500, Normal, nullptr) \
    257     v(int32, thresholdForJITSoon, 100, Normal, nullptr) \
    258     \
    259     v(int32, thresholdForOptimizeAfterWarmUp, 1000, Normal, nullptr) \
    260     v(int32, thresholdForOptimizeAfterLongWarmUp, 1000, Normal, nullptr) \
    261     v(int32, thresholdForOptimizeSoon, 1000, Normal, nullptr) \
    262     v(int32, executionCounterIncrementForLoop, 1, Normal, nullptr) \
    263     v(int32, executionCounterIncrementForEntry, 15, Normal, nullptr) \
    264     \
    265     v(int32, thresholdForFTLOptimizeAfterWarmUp, 100000, Normal, nullptr) \
    266     v(int32, thresholdForFTLOptimizeSoon, 1000, Normal, nullptr) \
    267     v(int32, ftlTierUpCounterIncrementForLoop, 1, Normal, nullptr) \
    268     v(int32, ftlTierUpCounterIncrementForReturn, 15, Normal, nullptr) \
     294    v(int32, thresholdForJITAfterWarmUp, archThresholdForJITAfterWarmUp, Normal, nullptr) \
     295    v(int32, thresholdForJITSoon, archThresholdForJITSoon, Normal, nullptr) \
     296    \
     297    v(int32, thresholdForOptimizeAfterWarmUp, archThresholdForOptimizeAfterWarmUp, Normal, nullptr) \
     298    v(int32, thresholdForOptimizeAfterLongWarmUp, archThresholdForOptimizeAfterLongWarmUp, Normal, nullptr) \
     299    v(int32, thresholdForOptimizeSoon, archThresholdForOptimizeSoon, Normal, nullptr) \
     300    v(int32, executionCounterIncrementForLoop, archExecutionCounterIncrementForLoop, Normal, nullptr) \
     301    v(int32, executionCounterIncrementForEntry, archExecutionCounterIncrementForEntry, Normal, nullptr) \
     302    \
     303    v(int32, thresholdForFTLOptimizeAfterWarmUp, archThresholdForFTLOptimizeAfterWarmUp, Normal, nullptr) \
     304    v(int32, thresholdForFTLOptimizeSoon, archThresholdForFTLOptimizeSoon, Normal, nullptr) \
     305    v(int32, ftlTierUpCounterIncrementForLoop, archFtlTierUpCounterIncrementForLoop, Normal, nullptr) \
     306    v(int32, ftlTierUpCounterIncrementForReturn, archFtlTierUpCounterIncrementForReturn, Normal, nullptr) \
    269307    v(unsigned, ftlOSREntryFailureCountForReoptimization, 15, Normal, nullptr) \
    270     v(unsigned, ftlOSREntryRetryThreshold, 100, Normal, nullptr) \
    271     \
    272     v(int32, evalThresholdMultiplier, 10, Normal, nullptr) \
     308    v(unsigned, ftlOSREntryRetryThreshold, archFtlOSREntryRetryThreshold, Normal, nullptr) \
     309    \
     310    v(double, optimizationThresholdScalingFactorA, archOptimizationThresholdScalingFactorA, Normal, nullptr) \
     311    v(double, optimizationThresholdScalingFactorB, archOptimizationThresholdScalingFactorB, Normal, nullptr) \
     312    v(double, optimizationThresholdScalingFactorC, archOptimizationThresholdScalingFactorC, Normal, nullptr) \
     313    v(double, optimizationThresholdScalingFactorD, archOptimizationThresholdScalingFactorD, Normal, nullptr) \
     314    \
     315    v(int32, evalThresholdMultiplier, archEvalThresholdMultiplier, Normal, nullptr) \
    273316    v(unsigned, maximumEvalCacheableSourceLength, 256, Normal, nullptr) \
    274317    \

  • trunk/Tools/ChangeLog

    r203284 r203318  
     12016-07-15  Benjamin Poulain  <bpoulain@apple.com>
     2
     3        [JSC] Change some parameters based on a random search
     4        https://bugs.webkit.org/show_bug.cgi?id=158514
     5
     6        Reviewed by Saam Barati.
     7
     8        * Scripts/run-jsc-stress-tests:
     9
    1102016-07-15  Jon Davis  <jond@apple.com>
    211

  • trunk/Tools/Scripts/run-jsc-stress-tests

    r203067 r203318  
    430430BASE_OPTIONS = ["--useFTLJIT=false", "--useFunctionDotArguments=true", "--maxPerThreadStackUsage=1572864"]
    431431EAGER_OPTIONS = ["--thresholdForJITAfterWarmUp=10", "--thresholdForJITSoon=10", "--thresholdForOptimizeAfterWarmUp=20", "--thresholdForOptimizeAfterLongWarmUp=20", "--thresholdForOptimizeSoon=20", "--thresholdForFTLOptimizeAfterWarmUp=20", "--thresholdForFTLOptimizeSoon=20", "--maximumEvalCacheableSourceLength=150000"]
    432 NO_CJIT_OPTIONS = ["--useConcurrentJIT=false", "--thresholdForJITAfterWarmUp=100"]
     432NO_CJIT_OPTIONS = ["--useConcurrentJIT=false", "--thresholdForJITAfterWarmUp=100", "--thresholdForJITSoon=100", "--thresholdForOptimizeAfterWarmUp=1000", "--thresholdForOptimizeAfterLongWarmUp=1000", "--thresholdForOptimizeSoon=1000", "--executionCounterIncrementForLoop=1", "--executionCounterIncrementForEntry=15", "--thresholdForFTLOptimizeAfterWarmUp=100000", "--thresholdForFTLOptimizeSoon=1000", "--ftlTierUpCounterIncrementForLoop=1", "--ftlTierUpCounterIncrementForReturn=15", "--evalThresholdMultiplier=10", "--optimizationThresholdScalingFactorA=0.061504", "--optimizationThresholdScalingFactorB=1.02406", "--optimizationThresholdScalingFactorC=0.0", "--optimizationThresholdScalingFactorD=0.825914"]
    433433FTL_OPTIONS = ["--useFTLJIT=true"]
    434434
Note: See TracChangeset for help on using the changeset viewer.