Bug 68227

Summary:

DFG JIT should inline Math.abs

Product:

WebKit

Reporter:

Filip Pizlo <fpizlo>

Component:

JavaScriptCore

Assignee:

Nobody <webkit-unassigned>

Status:

RESOLVED FIXED

Severity:

Normal

CC:

webkit.review.bot

Priority:

Version:

528+ (Nightly build)

Hardware:

All

OS:

All

Attachments:

Description	Flags
the patch	gyuyoung.kim: commit-queue-
the patch - fix a lot of stuff	none
the patch - refactored	oliver: review+

Filip Pizlo

Reported 2011-09-16 00:03:26 PDT

Math.abs is a simple function to inline. We know exactly what it does. DFG already has the ability to speculate that it is definitely calling a particular function, thanks to CheckMethod. Implementing this feature will be a great first step for supporting inlining in general, and other sorts of intrinsic functions in particular.

Attachments
the patch (31.18 KB, patch) 2011-09-16 00:09 PDT, Filip Pizlo	gyuyoung.kim: commit-queue-	Details Formatted Diff Diff
the patch - fix a lot of stuff (37.00 KB, patch) 2011-09-16 01:08 PDT, Filip Pizlo	no flags	Details Formatted Diff Diff
the patch - refactored (39.72 KB, patch) 2011-09-16 01:54 PDT, Filip Pizlo	oliver: review+	Details Formatted Diff Diff
Show Obsolete (2) View All Add attachment proposed patch, testcase, etc.

Filip Pizlo

Comment 1 2011-09-16 00:05:10 PDT

This is a 13% win on Kraken, due to a 61% win on imaging-gaussian-blur. Benchmark report for SunSpider, V8, and Kraken. VMs tested: "TipOfTree" at /Volumes/Data/pizlo/quinary/OpenSource/WebKitBuild/Release/jsc "CheckMethod" at /Volumes/Data/pizlo/quartary/OpenSource/WebKitBuild/Release/jsc "MathDotAbs" at /Volumes/Data/pizlo/septenary/OpenSource/WebKitBuild/Release/jsc Collected 12 samples per benchmark/VM, with 4 VM invocations per benchmark. 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 CheckMethod MathDotAbs MathDotAbs v. TipOfTree SunSpider: 3d-cube 7.8771+-0.2118 ? 7.9524+-0.1532 7.7974+-0.2673 might be 1.0102x faster 3d-morph 7.4824+-0.1647 7.3816+-0.1877 ? 7.3948+-0.1308 might be 1.0118x faster 3d-raytrace 7.9141+-0.1938 ? 7.9445+-0.1934 ^ 7.5801+-0.1588 might be 1.0441x faster access-binary-trees 2.3728+-0.0569 2.2900+-0.0530 ? 2.2988+-0.0655 might be 1.0322x faster access-fannkuch 11.7252+-0.2695 ? 11.7583+-0.1997 11.5732+-0.2085 might be 1.0131x faster access-nbody 4.1582+-0.0611 ? 4.3343+-0.1218 ? 4.8783+-0.8866 ? might be 1.1732x slower access-nsieve 2.6476+-0.0877 2.6017+-0.0657 2.5885+-0.0507 might be 1.0228x faster bitops-3bit-bits-in-byte 1.6404+-0.0357 ? 1.6442+-0.0437 ? 1.7019+-0.0660 ? might be 1.0375x slower bitops-bits-in-byte 2.7977+-0.0822 2.6993+-0.0352 ? 2.7583+-0.0610 might be 1.0143x faster bitops-bitwise-and 3.5206+-0.1040 ? 3.6103+-0.0862 ? 3.7339+-0.1119 ? might be 1.0606x slower bitops-nsieve-bits 5.2149+-0.0636 ? 5.3302+-0.1189 ? 5.3482+-0.0815 ? might be 1.0256x slower controlflow-recursive 1.9950+-0.0440 ? 2.0208+-0.0498 ? 2.0828+-0.0918 ? might be 1.0440x slower crypto-aes 7.2947+-0.3756 6.9862+-0.3805 6.9653+-0.3419 might be 1.0473x faster crypto-md5 2.8767+-0.1047 2.8361+-0.1007 2.7773+-0.0539 might be 1.0358x faster crypto-sha1 2.2561+-0.0575 2.1765+-0.0356 ? 2.2091+-0.0720 might be 1.0213x faster date-format-tofte 10.5105+-0.2760 10.3554+-0.1631 ? 10.3874+-0.2410 might be 1.0118x faster date-format-xparb 8.6902+-0.2079 ? 9.0536+-0.2347 8.8852+-0.2032 ? might be 1.0224x slower math-cordic 6.2988+-0.1205 6.1587+-0.1171 ? 6.2109+-0.1638 might be 1.0141x faster math-partial-sums 7.2658+-0.1211 ? 7.3612+-0.1526 7.3564+-0.1448 ? might be 1.0125x slower math-spectral-norm 2.8941+-0.4114 ? 3.1316+-0.6769 2.7383+-0.0906 might be 1.0569x faster regexp-dna 10.7131+-0.2434 ? 10.9977+-0.1918 10.9197+-0.2360 ? might be 1.0193x slower string-base64 5.8277+-0.1519 5.7318+-0.1803 ? 6.0018+-0.2913 ? might be 1.0299x slower string-fasta 7.0203+-0.1499 ? 7.0551+-0.2153 6.9808+-0.1341 string-tagcloud 11.9234+-0.3495 ? 12.2052+-0.4943 11.9608+-0.3251 ? string-unpack-code 19.0595+-0.4566 18.9328+-0.3451 18.3859+-0.2872 might be 1.0366x faster string-validate-input 6.6001+-0.2032 6.4331+-0.1183 ? 6.5165+-0.1550 might be 1.0128x faster <arithmetic> 6.4837+-0.0365 ? 6.4993+-0.0494 6.4627+-0.0509 <geometric> 5.3517+-0.0358 ? 5.3518+-0.0542 ? 5.3537+-0.0428 ? <harmonic> 4.3701+-0.0436 4.3537+-0.0536 ? 4.3803+-0.0349 ? TipOfTree CheckMethod MathDotAbs MathDotAbs v. TipOfTree V8: crypto 83.3726+-0.6100 ? 83.3745+-0.5430 82.8878+-0.3526 deltablue 246.6145+-3.1631 ^ 240.5071+-1.2818 ? 243.8680+-2.7147 might be 1.0113x faster earley-boyer 95.5398+-0.6002 94.6500+-0.4064 ? 95.3739+-0.3684 raytrace 75.6773+-6.1967 72.7885+-0.4288 72.7452+-0.5102 might be 1.0403x faster regexp 108.8515+-0.3325 108.4031+-0.5479 ^ 106.5182+-0.3822 ^ definitely 1.0219x faster richards 219.0579+-0.7349 ? 219.3573+-2.8656 218.7951+-1.8159 splay 100.3839+-0.9403 99.6274+-0.6926 ^ 98.2193+-0.3959 ^ definitely 1.0220x faster <arithmetic> 132.7853+-1.1595 131.2440+-0.5493 131.2011+-0.6375 might be 1.0121x faster <geometric> 119.7848+-1.3477 118.4509+-0.3559 118.1263+-0.3993 might be 1.0140x faster <harmonic> 110.1116+-1.5381 108.8954+-0.2899 108.4592+-0.2939 might be 1.0152x faster TipOfTree CheckMethod MathDotAbs MathDotAbs v. TipOfTree Kraken: ai-astar 633.9435+-4.1433 632.6228+-4.4504 ? 638.0119+-2.3449 ? audio-beat-detection 472.9599+-1.8222 ^ 468.2434+-1.5094 ! 477.2912+-7.3407 ? audio-dft 424.4090+-1.8641 ? 430.5855+-4.3855 ? 458.9972+-42.7620 ? might be 1.0815x slower audio-fft 369.4709+-1.7994 ^ 363.4016+-1.2242 ? 365.5753+-3.3093 might be 1.0107x faster audio-oscillator 319.4655+-1.6668 ^ 316.7981+-0.6281 ^ 314.1501+-1.7790 ^ definitely 1.0169x faster imaging-darkroom 418.6947+-2.0732 ^ 415.0992+-1.2111 414.6453+-0.6410 ^ definitely 1.0098x faster imaging-desaturate 209.1284+-0.7906 ^ 207.6594+-0.6454 ? 209.4528+-2.0826 ? imaging-gaussian-blur 1736.4665+-7.2265 ^ 1574.9463+-7.2155 ^ 1077.3633+-4.2862 ^ definitely 1.6118x faster json-parse-financial 48.9103+-0.2949 ! 50.7933+-1.0150 50.1767+-0.5739 ! definitely 1.0259x slower json-stringify-tinderbox 69.6547+-0.4141 69.3233+-0.1353 69.0610+-0.3808 stanford-crypto-aes 145.1025+-0.9297 ? 145.1324+-0.8121 144.1862+-2.0463 stanford-crypto-ccm 112.5759+-0.5658 ^ 111.2905+-0.4928 ? 112.6970+-1.1181 ? stanford-crypto-pbkdf2 398.3721+-2.9153 395.1204+-2.0538 ? 400.3701+-5.5710 ? stanford-crypto-sha256-iterative 149.4793+-1.5592 ? 149.6702+-0.6328 ? 149.7595+-0.2238 ? <arithmetic> 393.4738+-0.6612 ^ 380.7633+-0.9488 ^ 348.6955+-2.5650 ^ definitely 1.1284x faster <geometric> 262.9734+-0.4518 ^ 260.7038+-0.4816 ^ 255.3935+-1.2807 ^ definitely 1.0297x faster <harmonic> 176.2926+-0.4149 ? 177.1632+-0.8505 176.5071+-0.6388 ? TipOfTree CheckMethod MathDotAbs MathDotAbs v. TipOfTree All benchmarks: <arithmetic> 140.5682+-0.2888 ^ 136.5612+-0.2688 ^ 126.9825+-0.7800 ^ definitely 1.1070x faster <geometric> 27.1244+-0.1400 27.0094+-0.1561 26.8389+-0.1312 ^ definitely 1.0106x faster <harmonic> 7.7144+-0.0761 7.6856+-0.0928 ? 7.7308+-0.0601 ?

Filip Pizlo

Comment 2 2011-09-16 00:09:35 PDT

Created attachment 107613 [details] the patch This is still a work in progress, as it needs some build magic for other platforms. I also plan on refactoring the intrinsic handling in DFGByteCodeParser, since I anticipate adding a lot more intrinsics.

WebKit Review Bot

Comment 3 2011-09-16 00:11:31 PDT

Attachment 107613 [details] did not pass style-queue: Failed to run "['Tools/Scripts/check-webkit-style', '--diff-files', u'Source/JavaScriptCore/ChangeLog', u'Source..." exit_code: 1 Source/JavaScriptCore/runtime/JSGlobalData.h:34: Alphabetical sorting problem. [build/include_order] [4] Source/JavaScriptCore/runtime/JSGlobalData.h:223: The parameter name "intrinsic" adds no information, so it should be removed. [readability/parameter_name] [5] Total errors found: 2 in 19 files If any of these errors are false positives, please file a bug against check-webkit-style.

Gyuyoung Kim

Comment 4 2011-09-16 00:23:26 PDT

Comment on attachment 107613 [details] the patch Attachment 107613 [details] did not pass efl-ews (efl): Output: http://queues.webkit.org/results/9711076

Filip Pizlo

Comment 5 2011-09-16 01:08:08 PDT

Created attachment 107617 [details] the patch - fix a lot of stuff This is still a bit of a work in progress because the changes are pretty extensive and I'm not sure I've gotten it right.

Filip Pizlo

Comment 6 2011-09-16 01:54:05 PDT

Created attachment 107625 [details] the patch - refactored Refactored the intrinsic handling in DFG::ByteCodeParser to be more reasonable, particularly for when we add more intrinsics.

Filip Pizlo

Comment 7 2011-09-16 11:49:41 PDT

Landed in r95310.

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