Bug 127639

Summary:

FTL should do polyvariant GetById inlining

Product:

WebKit

Reporter:

Filip Pizlo <fpizlo>

Component:

JavaScriptCore

Assignee:

Filip Pizlo <fpizlo>

Status:

RESOLVED FIXED

Severity:

Normal

CC:

barraclough, ggaren, mark.lam, mhahnenberg, mmirman, msaboff, oliver, sam

Priority:

Version:

528+ (Nightly build)

Hardware:

All

OS:

All

Bug Depends on:

127646

Bug Blocks:

127325

Attachments:

Description	Flags
it begins	none
almost there	none
the patch	oliver: review+

Filip Pizlo

Reported 2014-01-25 20:04:19 PST

Patch forthcoming.

Attachments
it begins (4.90 KB, patch) 2014-01-25 22:44 PST, Filip Pizlo	no flags	Details Formatted Diff Diff
almost there (15.36 KB, patch) 2014-01-26 10:50 PST, Filip Pizlo	no flags	Details Formatted Diff Diff
the patch (15.81 KB, patch) 2014-01-26 12:45 PST, Filip Pizlo	oliver: review+	Details Formatted Diff Diff
Show Obsolete (2) View All Add attachment proposed patch, testcase, etc.

Filip Pizlo

Comment 1 2014-01-25 22:44:10 PST

Created attachment 222257 [details] it begins

Filip Pizlo

Comment 2 2014-01-26 10:50:56 PST

Created attachment 222279 [details] almost there It inlined a GetById.

Filip Pizlo

Comment 3 2014-01-26 12:16:09 PST

This, combined with polyvariant Call/Construct inlining, is a 15-16% speed-up on Raytrace.

Oliver Hunt

Comment 4 2014-01-26 12:20:13 PST

(In reply to comment #3) > This, combined with polyvariant Call/Construct inlining, is a 15-16% speed-up on Raytrace. ooh, nice

Filip Pizlo

Comment 5 2014-01-26 12:35:02 PST

(In reply to comment #4) > (In reply to comment #3) > > This, combined with polyvariant Call/Construct inlining, is a 15-16% speed-up on Raytrace. > > ooh, nice Yup.

Filip Pizlo

Comment 6 2014-01-26 12:39:12 PST

Here's the combined impact of polyvariant call, construct, and get_by_id inlining, as well as our loosening of the restrictions on recursive inlining. Benchmark report for SunSpider, Octane, Kraken, and AsmBench on oldmac (MacPro4,1). VMs tested: "TipOfTree" at /Volumes/Data/pizlo/cStack/OpenSource/WebKitBuild/Release/jsc (r162775) "Polyvariant" at /Volumes/Data/fromMiniMe/cStack/OpenSource/WebKitBuild/Release/jsc (r162802) Collected 7 samples per benchmark/VM, with 7 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. TipOfTree Polyvariant SunSpider: 3d-cube 8.2543+-0.0749 8.2333+-0.1511 3d-morph 9.2846+-0.2250 9.2691+-0.1157 3d-raytrace 9.7788+-0.2722 9.7150+-0.0529 access-binary-trees 2.7546+-0.0173 ! 2.9426+-0.0121 ! definitely 1.0683x slower access-fannkuch 8.1062+-0.1008 ? 8.1651+-0.0838 ? access-nbody 4.6833+-0.1771 4.4671+-0.1296 might be 1.0484x faster access-nsieve 5.5225+-0.2111 5.4650+-0.1239 might be 1.0105x faster bitops-3bit-bits-in-byte 1.9982+-0.0312 ? 2.0090+-0.0279 ? bitops-bits-in-byte 6.6814+-0.0321 6.6639+-0.0190 bitops-bitwise-and 3.0589+-0.0673 3.0476+-0.0139 bitops-nsieve-bits 5.8736+-0.0810 5.7569+-0.1250 might be 1.0203x faster controlflow-recursive 3.9936+-0.0468 ^ 3.3387+-0.0278 ^ definitely 1.1962x faster crypto-aes 6.2213+-0.0279 ? 6.2234+-0.0255 ? crypto-md5 3.8075+-0.0526 ^ 3.6968+-0.0291 ^ definitely 1.0299x faster crypto-sha1 3.7536+-0.0368 ? 3.7608+-0.0212 ? date-format-tofte 12.3529+-0.0818 ! 12.6441+-0.1839 ! definitely 1.0236x slower date-format-xparb 8.9583+-0.0575 ^ 8.7369+-0.0866 ^ definitely 1.0253x faster math-cordic 4.8908+-0.0517 4.8886+-0.0249 math-partial-sums 10.3546+-0.1096 ? 10.4173+-0.1506 ? math-spectral-norm 3.2556+-0.0364 ? 3.2887+-0.0389 ? might be 1.0102x slower regexp-dna 13.0241+-0.1945 ? 13.0838+-0.1657 ? string-base64 5.9731+-0.0482 ? 5.9746+-0.0953 ? string-fasta 11.1139+-0.2209 11.1097+-0.1173 string-tagcloud 15.8351+-0.1706 15.6483+-0.2020 might be 1.0119x faster string-unpack-code 33.6274+-0.6313 ? 33.8924+-0.1845 ? string-validate-input 7.8653+-0.1073 7.8404+-0.0598 <arithmetic> * 8.1163+-0.0253 8.0877+-0.0243 might be 1.0035x faster <geometric> 6.6451+-0.0221 ^ 6.5917+-0.0170 ^ definitely 1.0081x faster <harmonic> 5.5843+-0.0167 ^ 5.5303+-0.0152 ^ definitely 1.0098x faster TipOfTree Polyvariant Octane and V8v7: encrypt 0.41947+-0.00054 ? 0.42183+-0.00353 ? decrypt 7.59646+-0.05317 7.56945+-0.04025 deltablue x2 0.50688+-0.00434 0.50290+-0.00519 earley 0.98659+-0.02694 0.96930+-0.00832 might be 1.0178x faster boyer 12.25735+-0.09908 12.23577+-0.04478 raytrace x2 4.08674+-0.11312 ^ 3.46633+-0.03018 ^ definitely 1.1790x faster regexp x2 32.47881+-0.10753 ? 32.58171+-0.17481 ? richards x2 0.21225+-0.00624 0.21141+-0.00340 splay x2 0.61052+-0.00422 ^ 0.58871+-0.00244 ^ definitely 1.0370x faster navier-stokes x2 7.79889+-0.00486 ? 7.80587+-0.03070 ? closure 0.74937+-0.01671 ? 0.76622+-0.00554 ? might be 1.0225x slower jquery 10.81645+-0.07306 ! 11.03254+-0.02684 ! definitely 1.0200x slower gbemu x2 69.14262+-0.79314 ^ 62.30284+-0.78711 ^ definitely 1.1098x faster mandreel x2 97.96271+-2.83544 97.44957+-0.68272 pdfjs x2 101.85877+-0.27247 ? 102.26682+-0.30497 ? box2d x2 30.28804+-0.64902 30.23497+-0.16425 V8v7: <arithmetic> 7.04050+-0.01758 ^ 6.96939+-0.02719 ^ definitely 1.0102x faster <geometric> * 2.12872+-0.00887 ^ 2.07182+-0.00711 ^ definitely 1.0275x faster <harmonic> 0.75943+-0.00996 0.74952+-0.00610 might be 1.0132x faster Octane including V8v7: <arithmetic> 27.79685+-0.25170 ^ 27.22375+-0.10254 ^ definitely 1.0211x faster <geometric> * 6.30850+-0.01485 ^ 6.16355+-0.01226 ^ definitely 1.0235x faster <harmonic> 1.14888+-0.01537 1.13631+-0.00882 might be 1.0111x faster TipOfTree Polyvariant Kraken: ai-astar 495.650+-0.810 495.638+-0.390 audio-beat-detection 221.692+-1.536 ? 222.753+-1.646 ? audio-dft 301.221+-5.226 299.042+-6.949 audio-fft 130.039+-0.116 ? 131.248+-2.567 ? audio-oscillator 592.556+-9.839 ^ 578.174+-3.019 ^ definitely 1.0249x faster imaging-darkroom 298.977+-0.823 298.561+-1.151 imaging-desaturate 105.858+-0.138 ? 105.867+-0.526 ? imaging-gaussian-blur 202.816+-35.705 189.334+-1.013 might be 1.0712x faster json-parse-financial 82.880+-0.263 82.461+-0.503 json-stringify-tinderbox 106.786+-2.419 104.643+-0.793 might be 1.0205x faster stanford-crypto-aes 160.896+-2.225 ^ 157.776+-0.839 ^ definitely 1.0198x faster stanford-crypto-ccm 108.134+-3.806 106.794+-0.659 might be 1.0125x faster stanford-crypto-pbkdf2 272.840+-3.271 271.732+-5.495 stanford-crypto-sha256-iterative 116.669+-0.647 ? 117.601+-2.366 ? <arithmetic> * 228.358+-2.643 225.830+-0.827 might be 1.0112x faster <geometric> 189.878+-2.074 188.136+-0.744 might be 1.0093x faster <harmonic> 162.401+-1.325 161.188+-0.647 might be 1.0075x faster TipOfTree Polyvariant AsmBench: bigfib.cpp 1194.8333+-19.1009 1187.6722+-11.3044 cray.c 56.2320+-0.2700 ? 56.7966+-1.3339 ? might be 1.0100x slower dry.c 881.8271+-56.9067 ? 898.1438+-62.3191 ? might be 1.0185x slower FloatMM.c 1777.0444+-28.7809 1762.6009+-1.0136 gcc-loops.cpp 2177.6453+-1.1674 ? 2177.7537+-1.4576 ? n-body.c 3069.9110+-1.4905 ? 3081.2242+-28.9799 ? Quicksort.c 83.4500+-0.4061 83.0268+-0.3604 stepanov_container.cpp 9341.3638+-31.3451 9304.7015+-39.0187 Towers.c 69.8374+-0.5867 ? 70.6055+-0.3869 ? might be 1.0110x slower <arithmetic> 2072.4605+-5.4676 2069.1694+-7.9309 might be 1.0016x faster <geometric> * 695.8181+-4.7783 ? 697.3273+-6.1953 ? might be 1.0022x slower <harmonic> 189.4043+-0.7306 ? 190.5190+-1.8482 ? might be 1.0059x slower TipOfTree Polyvariant All benchmarks: <arithmetic> 303.7720+-0.8340 302.6966+-1.0118 might be 1.0036x faster <geometric> 21.3231+-0.0741 ^ 21.0623+-0.0531 ^ definitely 1.0124x faster <harmonic> 2.7351+-0.0305 2.7059+-0.0180 might be 1.0108x faster TipOfTree Polyvariant Geomean of preferred means: <scaled-result> 53.4062+-0.2108 ^ 52.9316+-0.1753 ^ definitely 1.0090x faster

Filip Pizlo

Comment 7 2014-01-26 12:45:50 PST

Created attachment 222282 [details] the patch

Filip Pizlo

Comment 8 2014-01-26 14:35:33 PST

Landed in http://trac.webkit.org/changeset/162811

Note You need to log in before you can comment on or make changes to this bug.