Bug 153583

Summary: LowerToAir::preferRightForResult() should resolve use count ties by selecting the child that is closest in an idom walk
Product: WebKit Reporter: Filip Pizlo <fpizlo>
Component: JavaScriptCoreAssignee: Filip Pizlo <fpizlo>
Status: RESOLVED FIXED    
Severity: Normal CC: barraclough, benjamin, commit-queue, fpizlo, ggaren, keith_miller, mark.lam, mhahnenb, msaboff, oliver, saam, sam
Priority: P2    
Version: WebKit Nightly Build   
Hardware: All   
OS: All   
Bug Depends on: 153477    
Bug Blocks: 150279    
Attachments:
Description Flags
the patch benjamin: review+

Filip Pizlo
Reported 2016-01-27 21:46:49 PST
Patch forthcoming.
Attachments
the patch (3.26 KB, patch)
2016-01-27 21:48 PST, Filip Pizlo 		benjamin: review+
DetailsFormatted DiffDiff
View All Add attachment proposed patch, testcase, etc.
Filip Pizlo
Comment 1 2016-01-27 21:48:26 PST
Created attachment 270092 [details] the patch
Filip Pizlo
Comment 2 2016-01-27 22:20:27 PST
Performance is great. Benchmark report for SunSpider, V8Spider, Octane, Kraken, and AsmBench on shakezilla (MacBookPro11,3). VMs tested: "LLVMToT" at /Volumes/Data/secondary/OpenSource/WebKitBuild/Release/jsc (r195683) "B3ToT" at /Volumes/Data/quartary/OpenSource/WebKitBuild/Release/jsc (r195683) "IntRangeResultTie" at /Volumes/Data/tertiary/OpenSource/WebKitBuild/Release/jsc (r195727) Collected 6 samples per benchmark/VM, with 6 VM invocations per benchmark. Emitted a call to gc() between sample measurements. Used 1 benchmark iteration per VM invocation for warm-up. Used the jsc-specific preciseTime() function to get microsecond-level timing. Reporting benchmark execution times with 95% confidence intervals in milliseconds. LLVMToT B3ToT IntRangeResultTie IntRangeResultTie v. LLVMToT SunSpider: 3d-cube 4.7868+-0.2815 ? 4.8253+-0.2163 4.7780+-0.1649 3d-morph 5.6364+-0.2738 5.4150+-0.0928 5.4127+-0.0806 might be 1.0413x faster 3d-raytrace 5.5544+-0.2719 5.5263+-0.1225 ? 5.5907+-0.3660 ? access-binary-trees 2.1321+-0.0821 ? 2.3525+-0.3584 2.1240+-0.1099 access-fannkuch 5.7092+-0.1440 5.6554+-0.0604 ? 5.7390+-0.1997 ? access-nbody 2.7085+-0.1070 ? 2.7121+-0.0783 ? 2.8502+-0.3419 ? might be 1.0523x slower access-nsieve 3.5424+-0.2419 3.2070+-0.1591 3.1936+-0.1696 might be 1.1092x faster bitops-3bit-bits-in-byte 1.1918+-0.0365 ? 1.2418+-0.0750 1.1825+-0.0229 bitops-bits-in-byte 3.1667+-0.0612 ? 3.1761+-0.0907 3.1346+-0.0460 might be 1.0102x faster bitops-bitwise-and 1.9809+-0.0357 1.9589+-0.0194 ? 2.0785+-0.2312 ? might be 1.0493x slower bitops-nsieve-bits 3.0190+-0.0810 2.9652+-0.0485 ? 2.9932+-0.0770 controlflow-recursive 2.3630+-0.0406 2.3073+-0.0352 ? 2.4025+-0.1975 ? might be 1.0167x slower crypto-aes 4.0763+-0.1320 3.9845+-0.0495 ? 4.0228+-0.0513 might be 1.0133x faster crypto-md5 2.5489+-0.1657 2.4900+-0.0233 2.4760+-0.0468 might be 1.0294x faster crypto-sha1 2.5167+-0.1554 ? 2.6619+-0.0732 ^ 2.4764+-0.1106 might be 1.0163x faster date-format-tofte 7.4203+-0.1415 ^ 6.8451+-0.1873 ? 6.9632+-0.1901 ^ definitely 1.0656x faster date-format-xparb 4.6837+-0.0694 ? 4.8920+-0.4579 ? 4.9088+-0.5330 ? might be 1.0481x slower math-cordic 2.9353+-0.0171 ? 2.9582+-0.0705 2.9404+-0.0834 ? math-partial-sums 5.0071+-0.2618 4.8955+-0.0623 4.8215+-0.0394 might be 1.0385x faster math-spectral-norm 2.0692+-0.0338 ? 2.1110+-0.1384 2.0952+-0.0570 ? might be 1.0126x slower regexp-dna 6.4076+-0.2203 6.3205+-0.2380 6.1802+-0.1799 might be 1.0368x faster string-base64 4.6141+-0.1047 ? 4.6174+-0.0840 4.6077+-0.1342 string-fasta 5.8393+-0.1244 5.7441+-0.0525 ? 5.7911+-0.1190 string-tagcloud 8.0772+-0.5305 ? 8.2496+-0.5266 7.8110+-0.0823 might be 1.0341x faster string-unpack-code 19.2227+-1.3339 ? 19.5211+-0.9737 18.6553+-0.6031 might be 1.0304x faster string-validate-input 4.3283+-0.0854 4.2504+-0.0842 4.2223+-0.0712 might be 1.0251x faster <arithmetic> 4.6745+-0.0637 4.6494+-0.0270 4.5943+-0.0352 might be 1.0175x faster LLVMToT B3ToT IntRangeResultTie IntRangeResultTie v. LLVMToT V8Spider: crypto 50.6172+-0.8063 ^ 38.0944+-0.4680 37.8385+-1.0320 ^ definitely 1.3377x faster deltablue 78.7497+-2.2153 ^ 52.9686+-1.9318 ? 54.0748+-1.7604 ^ definitely 1.4563x faster earley-boyer 43.8992+-2.1352 42.1202+-0.7371 41.8734+-0.8442 might be 1.0484x faster raytrace 30.4988+-1.3855 ^ 20.8573+-0.8064 20.4883+-0.7649 ^ definitely 1.4886x faster regexp 63.5347+-1.5673 62.5897+-1.1853 ? 62.6914+-1.8203 might be 1.0135x faster richards 52.7556+-1.1944 ^ 40.2718+-0.7943 ? 40.4253+-1.4797 ^ definitely 1.3050x faster splay 38.2197+-0.7033 36.8014+-1.3184 ? 36.8274+-1.8142 might be 1.0378x faster <geometric> 49.0414+-0.6832 ^ 40.0124+-0.5709 39.9853+-0.5778 ^ definitely 1.2265x faster LLVMToT B3ToT IntRangeResultTie IntRangeResultTie v. LLVMToT Octane: encrypt 0.15205+-0.00039 ! 0.16076+-0.00385 0.15910+-0.00345 ! definitely 1.0464x slower decrypt 2.90081+-0.00599 ^ 2.81434+-0.00715 ^ 2.76677+-0.00440 ^ definitely 1.0484x faster deltablue x2 0.13712+-0.00197 0.13482+-0.00331 0.13391+-0.00188 might be 1.0240x faster earley 0.28233+-0.00328 0.27799+-0.00359 0.27796+-0.00302 might be 1.0157x faster boyer 4.23605+-0.01324 ! 4.29561+-0.02226 ? 4.30663+-0.00993 ! definitely 1.0167x slower navier-stokes x2 4.80442+-0.01421 ? 4.80731+-0.02071 4.78451+-0.00512 ^ definitely 1.0042x faster raytrace x2 0.85049+-0.00423 ! 0.87279+-0.00517 0.87060+-0.00305 ! definitely 1.0236x slower richards x2 0.08571+-0.00104 ^ 0.07783+-0.00105 ? 0.07875+-0.00115 ^ definitely 1.0884x faster splay x2 0.35215+-0.00335 ? 0.35240+-0.00397 ? 0.35381+-0.00423 ? regexp x2 24.44420+-0.38045 24.07823+-0.45033 ? 24.26558+-0.29596 pdfjs x2 37.74735+-0.33441 ? 37.80085+-0.59317 37.58011+-0.33760 mandreel x2 42.59685+-0.27649 ? 43.08369+-0.41976 43.03390+-0.49042 ? might be 1.0103x slower gbemu x2 30.07276+-1.97624 ? 32.31426+-2.45155 31.58574+-0.16929 ? might be 1.0503x slower closure 0.56629+-0.00258 0.55992+-0.00424 ? 0.56022+-0.00195 ^ definitely 1.0108x faster jquery 7.36849+-0.05980 ^ 7.25221+-0.02702 ? 7.31534+-0.06379 box2d x2 9.14968+-0.04354 9.04354+-0.07781 ? 9.13952+-0.03874 zlib x2 390.75338+-6.96674 383.96236+-15.33596 ? 389.82506+-9.95550 typescript x2 666.12174+-8.54683 ? 674.75081+-27.22033 666.94397+-14.59913 ? <geometric> 5.28699+-0.02790 5.27420+-0.04054 5.27020+-0.01419 might be 1.0032x faster LLVMToT B3ToT IntRangeResultTie IntRangeResultTie v. LLVMToT Kraken: ai-astar 128.430+-1.538 ^ 93.769+-2.474 93.432+-0.897 ^ definitely 1.3746x faster audio-beat-detection 49.458+-0.166 ! 59.269+-0.802 ^ 54.922+-0.286 ! definitely 1.1105x slower audio-dft 95.576+-1.890 ? 96.949+-3.010 96.005+-2.287 ? audio-fft 36.022+-1.572 ! 47.503+-0.813 ^ 43.272+-0.212 ! definitely 1.2013x slower audio-oscillator 57.931+-1.198 ^ 51.428+-1.854 51.017+-1.087 ^ definitely 1.1355x faster imaging-darkroom 59.509+-0.100 ! 60.686+-0.057 ? 61.000+-1.960 ? might be 1.0251x slower imaging-desaturate 48.273+-0.264 ^ 43.311+-0.391 ? 43.741+-0.679 ^ definitely 1.1036x faster imaging-gaussian-blur 86.639+-0.920 ^ 67.517+-0.479 ? 68.136+-0.345 ^ definitely 1.2716x faster json-parse-financial 37.665+-1.001 ? 37.739+-1.344 37.515+-0.910 json-stringify-tinderbox 22.157+-0.605 ? 22.726+-0.644 ? 23.162+-0.353 ! definitely 1.0454x slower stanford-crypto-aes 42.631+-1.200 41.913+-1.579 ? 42.838+-2.336 ? stanford-crypto-ccm 38.123+-1.328 ? 39.390+-2.078 37.795+-2.042 stanford-crypto-pbkdf2 97.494+-0.674 ! 98.928+-0.514 ? 99.142+-0.348 ! definitely 1.0169x slower stanford-crypto-sha256-iterative 37.540+-1.184 ? 38.207+-1.662 ? 38.643+-0.824 ? might be 1.0294x slower <arithmetic> 59.818+-0.107 ^ 57.095+-0.388 56.473+-0.437 ^ definitely 1.0592x faster LLVMToT B3ToT IntRangeResultTie IntRangeResultTie v. LLVMToT AsmBench: bigfib.cpp 449.6363+-8.0252 ? 459.7618+-14.9895 454.5553+-8.9742 ? might be 1.0109x slower cray.c 390.4478+-2.1556 387.2905+-2.9131 ^ 374.8730+-1.8157 ^ definitely 1.0415x faster dry.c 413.2968+-8.9830 ? 414.1680+-12.4379 409.9243+-4.1435 FloatMM.c 680.6209+-0.6985 ! 735.0622+-3.6272 734.9834+-2.9606 ! definitely 1.0799x slower gcc-loops.cpp 3446.3645+-51.2775 ? 3495.3138+-5.9967 ? 3500.0423+-10.3657 ? might be 1.0156x slower n-body.c 847.5238+-65.3937 ! 1170.9742+-2.6420 ^ 851.1848+-1.5462 ? Quicksort.c 404.0680+-4.0243 397.1890+-3.5025 395.3180+-4.7062 ^ definitely 1.0221x faster stepanov_container.cpp 3486.7481+-19.2881 ^ 3211.4794+-19.9022 ? 3224.5616+-18.1876 ^ definitely 1.0813x faster Towers.c 233.1697+-3.0000 ! 249.3888+-0.8874 249.2753+-0.4977 ! definitely 1.0691x slower <geometric> 712.4230+-6.3717 ! 744.8021+-3.6816 ^ 714.5944+-1.6391 ? might be 1.0030x slower LLVMToT B3ToT IntRangeResultTie IntRangeResultTie v. LLVMToT Geomean of preferred means: <scaled-result> 34.8819+-0.0873 ^ 33.4249+-0.1811 ^ 32.9885+-0.1670 ^ definitely 1.0574x faster
Benjamin Poulain
Comment 3 2016-01-28 00:07:59 PST
Comment on attachment 270092 [details] the patch View in context: https://bugs.webkit.org/attachment.cgi?id=270092&action=review > Source/JavaScriptCore/ChangeLog:3 > + LowerToAir::preferRightForResult() should resolve use count ties by selecting the child that is closest in an idom walk I thought about picking the closest but that seemed overkill at the time. I guess it is not that crazy :)
Filip Pizlo
Comment 4 2016-01-28 09:30:29 PST
Landed in http://trac.webkit.org/changeset/195744
Note You need to log in before you can comment on or make changes to this bug.