Bug 74682

Summary:

Value profiling should distinguished between NaN and non-NaN doubles

Product:

WebKit

Reporter:

Filip Pizlo <fpizlo>

Component:

JavaScriptCore

Assignee:

Nobody <webkit-unassigned>

Status:

RESOLVED FIXED

Severity:

Normal

CC:

dglazkov, webkit.review.bot

Priority:

Keywords:

InRadar

Version:

528+ (Nightly build)

Hardware:

All

OS:

All

Attachments:

Description	Flags
the patch	barraclough: review+, webkit.review.bot: commit-queue-

Filip Pizlo

Reported 2011-12-15 20:55:17 PST

We can do some optimizations if we know that a double is non-NaN. For example, we can check that it is a double by simply comparing it against itself. Currently this would be hazardous since the value profiler makes no distinction between NaN and non-NaN doubles.

Attachments
the patch (4.21 KB, patch) 2011-12-15 21:02 PST, Filip Pizlo	barraclough: review+ webkit.review.bot: commit-queue-	Details Formatted Diff Diff
View All Add attachment proposed patch, testcase, etc.

Filip Pizlo

Comment 1 2011-12-15 21:02:32 PST

Created attachment 119554 [details] the patch As expected it's performance neutral. Benchmark report for SunSpider, V8, and Kraken on nitroflex (MacBookPro8,2). VMs tested: "TipOfTree" at /Volumes/Data/pizlo/quinary/OpenSource/WebKitBuild/Release/jsc (r103019) "PredictDoubleNaN" at /Volumes/Data/pizlo/quartary/OpenSource/WebKitBuild/Release/jsc (r103018) Collected 12 samples per benchmark/VM, with 4 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 PredictDoubleNaN SunSpider: 3d-cube 5.4840+-0.1330 5.3382+-0.0660 might be 1.0273x faster 3d-morph 7.4073+-0.1111 ? 7.4246+-0.0961 ? 3d-raytrace 6.9559+-0.1634 ? 7.0458+-0.1572 ? might be 1.0129x slower access-binary-trees 1.4580+-0.0310 ! 1.5520+-0.0530 ! definitely 1.0645x slower access-fannkuch 6.0871+-0.0773 ? 6.1192+-0.1244 ? access-nbody 3.1432+-0.0693 3.1330+-0.0459 access-nsieve 2.5735+-0.0658 ? 2.6017+-0.0665 ? might be 1.0109x slower bitops-3bit-bits-in-byte 1.2333+-0.0313 ? 1.2434+-0.0274 ? bitops-bits-in-byte 2.3625+-0.0654 2.3457+-0.0599 bitops-bitwise-and 3.3340+-0.0585 ? 3.3454+-0.0785 ? bitops-nsieve-bits 5.3881+-0.1048 5.3104+-0.0930 might be 1.0146x faster controlflow-recursive 2.0178+-0.0476 1.9701+-0.0325 might be 1.0242x faster crypto-aes 7.2700+-0.1402 7.1048+-0.1512 might be 1.0233x faster crypto-md5 2.4290+-0.1085 2.3045+-0.0370 might be 1.0540x faster crypto-sha1 2.1217+-0.0528 2.0630+-0.0252 might be 1.0285x faster date-format-tofte 9.8704+-0.1792 9.7020+-0.1486 might be 1.0174x faster date-format-xparb 8.9539+-0.1590 ? 8.9686+-0.1146 ? math-cordic 6.2687+-0.0764 ? 6.3139+-0.0836 ? math-partial-sums 7.5351+-0.1942 7.4479+-0.1355 might be 1.0117x faster math-spectral-norm 2.3118+-0.0427 ? 2.3248+-0.0184 ? regexp-dna 7.6980+-0.1357 ? 7.8306+-0.1395 ? might be 1.0172x slower string-base64 4.2284+-0.0946 4.2016+-0.0984 string-fasta 6.6104+-0.1455 6.5745+-0.1342 string-tagcloud 10.8547+-0.2017 10.8213+-0.1787 string-unpack-code 18.5251+-0.2388 ? 18.7376+-0.2351 ? might be 1.0115x slower string-validate-input 5.1609+-0.0778 ? 5.3114+-0.1091 ? might be 1.0292x slower <arithmetic> * 5.6647+-0.0115 5.6591+-0.0251 might be 1.0010x faster <geometric> 4.6001+-0.0154 4.5926+-0.0224 might be 1.0016x faster <harmonic> 3.6840+-0.0201 ? 3.6850+-0.0206 ? might be 1.0003x slower TipOfTree PredictDoubleNaN V8: crypto 67.9820+-0.6179 ! 69.7032+-0.9241 ! definitely 1.0253x slower deltablue 147.3645+-0.6929 ? 148.0295+-1.2605 ? earley-boyer 79.1573+-0.9040 79.0512+-0.8209 raytrace 46.8626+-0.4293 46.7946+-0.5416 regexp 101.8191+-0.5797 101.0834+-0.4287 richards 117.5731+-1.0009 117.2163+-1.3162 splay 71.2523+-0.9562 ? 72.8645+-1.3583 ? might be 1.0226x slower <arithmetic> 90.2873+-0.2690 ? 90.6775+-0.3451 ? might be 1.0043x slower <geometric> * 84.8563+-0.2995 ? 85.3231+-0.3419 ? might be 1.0055x slower <harmonic> 79.6867+-0.3411 ? 80.1875+-0.3707 ? might be 1.0063x slower TipOfTree PredictDoubleNaN Kraken: ai-astar 482.9699+-5.6795 478.8014+-5.7710 audio-beat-detection 170.2342+-1.3675 ? 171.0469+-2.2224 ? audio-dft 281.0433+-10.6172 272.1817+-2.9181 might be 1.0326x faster audio-fft 105.3395+-0.6960 ? 105.4496+-0.6958 ? audio-oscillator 252.2022+-5.6441 248.1389+-2.0256 might be 1.0164x faster imaging-darkroom 273.3609+-9.3956 266.7807+-6.2242 might be 1.0247x faster imaging-desaturate 223.3151+-1.0393 ? 224.7058+-1.6644 ? imaging-gaussian-blur 482.5333+-3.2677 481.7591+-3.2505 json-parse-financial 57.6880+-0.1772 ^ 56.9333+-0.4492 ^ definitely 1.0133x faster json-stringify-tinderbox 70.9564+-0.5107 ? 71.9561+-0.5235 ? might be 1.0141x slower stanford-crypto-aes 96.0367+-0.5502 96.0310+-0.6714 stanford-crypto-ccm 97.1362+-0.6539 ? 97.8263+-1.1311 ? stanford-crypto-pbkdf2 191.7424+-1.7969 189.9253+-1.7164 stanford-crypto-sha256-iterative 81.8126+-0.5954 ? 82.1507+-0.4898 ? <arithmetic> * 204.7408+-1.3838 203.1205+-0.6283 might be 1.0080x faster <geometric> 164.6821+-0.7232 163.9029+-0.3866 might be 1.0048x faster <harmonic> 133.6188+-0.3542 133.3949+-0.3179 might be 1.0017x faster TipOfTree PredictDoubleNaN All benchmarks: <arithmetic> 77.5673+-0.4335 77.1397+-0.1916 might be 1.0055x faster <geometric> 20.6141+-0.0611 20.5831+-0.0493 might be 1.0015x faster <harmonic> 6.4825+-0.0348 ? 6.4847+-0.0351 ? might be 1.0003x slower TipOfTree PredictDoubleNaN Geomean of preferred means: <scaled-result> 46.1690+-0.1460 46.1158+-0.0783 might be 1.0012x faster

WebKit Review Bot

Comment 2 2011-12-15 21:31:23 PST

Comment on attachment 119554 [details] the patch Attachment 119554 [details] did not pass chromium-ews (chromium-xvfb): Output: http://queues.webkit.org/results/10902542 New failing tests: media/event-attributes.html

Filip Pizlo

Comment 3 2011-12-15 21:41:00 PST

(In reply to comment #2) > (From update of attachment 119554 [details]) > Attachment 119554 [details] did not pass chromium-ews (chromium-xvfb): > Output: http://queues.webkit.org/results/10902542 > > New failing tests: > media/event-attributes.html Assuming that this is a fluke since I didn't touch any code that chromium uses.

Filip Pizlo

Comment 4 2011-12-15 21:41:31 PST

<rdar://problem/10591560>

Filip Pizlo

Comment 5 2011-12-15 21:42:12 PST

Landed in http://trac.webkit.org/changeset/103023

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