Bug 160539

Summary: [JSC] Speed up InPlaceAbstractState::endBasicBlock()
Product: WebKit Reporter: Benjamin Poulain <benjamin>
Component: New BugsAssignee: Benjamin Poulain <benjamin>
Status: RESOLVED FIXED    
Severity: Normal CC: commit-queue, keith_miller, mark.lam, msaboff, saam
Priority: P2    
Version: WebKit Nightly Build   
Hardware: Unspecified   
OS: Unspecified   
Attachments:
Description Flags
Patch
none
Patch for landing none

Benjamin Poulain
Reported 2016-08-04 00:09:26 PDT
[JSC] Speed up InPlaceAbstractState::endBasicBlock()
Attachments
Patch (15.18 KB, patch)
2016-08-04 00:23 PDT, Benjamin Poulain 		no flags
DetailsFormatted DiffDiff
Patch for landing (14.90 KB, patch)
2016-08-04 11:46 PDT, Benjamin Poulain 		no flags
DetailsFormatted DiffDiff
Show Obsolete (1) View All Add attachment proposed patch, testcase, etc.
Benjamin Poulain
Comment 1 2016-08-04 00:23:39 PDT
Created attachment 285305 [details] Patch
Benjamin Poulain
Comment 2 2016-08-04 00:24:25 PDT
24 runs: Conf#1 Conf#2 SunSpider: 3d-cube 4.7019+-0.0445 ? 4.7260+-0.0561 ? 3d-morph 5.0836+-0.0799 ^ 4.9649+-0.0244 ^ definitely 1.0239x faster 3d-raytrace 4.9661+-0.1071 4.9161+-0.0559 might be 1.0102x faster access-binary-trees 2.0023+-0.0263 1.9813+-0.0175 might be 1.0106x faster access-fannkuch 5.7993+-0.1762 5.6978+-0.0431 might be 1.0178x faster access-nbody 2.3969+-0.0276 2.3558+-0.0251 might be 1.0174x faster access-nsieve 2.9599+-0.0235 2.9555+-0.0190 bitops-3bit-bits-in-byte 1.0751+-0.0229 ? 1.0854+-0.0299 ? bitops-bits-in-byte 2.5775+-0.0164 ? 2.5817+-0.0210 ? bitops-bitwise-and 1.9970+-0.0298 1.9916+-0.0243 bitops-nsieve-bits 3.0999+-0.0508 ? 3.1225+-0.1006 ? controlflow-recursive 2.3036+-0.0123 ? 2.3185+-0.0339 ? crypto-aes 4.4593+-0.0247 4.4255+-0.0157 crypto-md5 2.6770+-0.0292 2.6587+-0.0351 crypto-sha1 2.7302+-0.0260 ? 2.7562+-0.0388 ? date-format-tofte 6.5233+-0.0668 6.4830+-0.0387 date-format-xparb 4.7152+-0.0269 ? 4.7308+-0.0337 ? math-cordic 2.7434+-0.0169 ? 2.7471+-0.0198 ? math-partial-sums 3.9862+-0.0359 3.9834+-0.0454 math-spectral-norm 2.0678+-0.0431 2.0581+-0.0369 regexp-dna 6.4471+-0.1295 6.4428+-0.0842 string-base64 4.0172+-0.0486 3.9858+-0.0242 string-fasta 5.5386+-0.0332 5.5166+-0.0451 string-tagcloud 8.4564+-0.1596 8.4525+-0.1690 string-unpack-code 18.2313+-0.2031 18.1489+-0.2162 string-validate-input 4.0572+-0.0358 ? 4.0676+-0.0298 ? <arithmetic> 4.4467+-0.0138 4.4290+-0.0144 might be 1.0040x faster Conf#1 Conf#2 Octane: encrypt 0.15532+-0.00134 ? 0.15603+-0.00141 ? decrypt 2.76892+-0.02251 2.75079+-0.00556 deltablue x2 0.13315+-0.00364 0.13298+-0.00372 earley 0.28542+-0.00091 ? 0.28626+-0.00112 ? boyer 4.94042+-0.04671 ? 4.97826+-0.03856 ? navier-stokes x2 4.94975+-0.01413 4.94368+-0.00819 raytrace x2 0.79568+-0.00331 0.79556+-0.00277 richards x2 0.08226+-0.00068 ? 0.08235+-0.00046 ? splay x2 0.33844+-0.00093 ? 0.33932+-0.00111 ? regexp x2 16.62279+-0.26727 ? 16.70929+-0.29884 ? pdfjs x2 38.90697+-0.17084 ^ 38.37963+-0.14737 ^ definitely 1.0137x faster mandreel x2 42.89519+-0.23220 42.77201+-0.13249 gbemu x2 29.87420+-0.47615 29.46608+-0.09044 might be 1.0139x faster closure 0.48411+-0.00113 0.48294+-0.00134 jquery 6.45182+-0.01299 6.44828+-0.01829 box2d x2 9.24477+-0.02668 ? 9.24608+-0.03331 ? zlib x2 365.30700+-4.35224 357.26474+-5.09561 might be 1.0225x faster typescript x2 604.52266+-3.40624 597.89421+-3.47951 might be 1.0111x faster <geometric> 5.03481+-0.01501 5.01664+-0.01060 might be 1.0036x faster Conf#1 Conf#2 Kraken: ai-astar 86.698+-0.545 ? 86.725+-0.791 ? audio-beat-detection 38.933+-0.145 38.907+-0.161 audio-dft 98.381+-1.161 97.826+-0.890 audio-fft 30.242+-0.024 ? 30.326+-0.109 ? audio-oscillator 49.023+-0.509 48.819+-0.436 imaging-darkroom 61.247+-0.278 ? 61.384+-0.509 ? imaging-desaturate 44.062+-0.566 43.958+-0.326 imaging-gaussian-blur 59.879+-0.743 ? 60.623+-0.889 ? might be 1.0124x slower json-parse-financial 34.682+-0.344 34.153+-0.319 might be 1.0155x faster json-stringify-tinderbox 23.239+-0.199 ^ 22.415+-0.249 ^ definitely 1.0367x faster stanford-crypto-aes 37.081+-0.433 ? 37.185+-0.300 ? stanford-crypto-ccm 33.517+-0.969 ? 34.784+-0.780 ? might be 1.0378x slower stanford-crypto-pbkdf2 93.333+-0.490 92.594+-0.319 stanford-crypto-sha256-iterative 30.490+-0.119 ? 30.523+-0.121 ? <arithmetic> 51.486+-0.145 51.445+-0.130 might be 1.0008x faster Conf#1 Conf#2 AsmBench: bigfib.cpp 425.7745+-0.8215 425.4747+-1.2422 cray.c 387.3546+-1.3674 384.7876+-1.4579 dry.c 466.1051+-33.8684 ^ 428.9013+-3.2271 ^ definitely 1.0867x faster FloatMM.c 735.3686+-14.8127 721.2791+-7.1189 might be 1.0195x faster gcc-loops.cpp 3594.9277+-7.0192 ? 3600.2280+-9.0302 ? n-body.c 800.5333+-1.3423 799.7137+-1.3151 Quicksort.c 398.6741+-2.8060 395.8523+-1.9722 stepanov_container.cpp 3295.6699+-10.0760 3279.1960+-10.5901 Towers.c 264.8358+-0.5208 ? 265.1645+-0.8941 ? <geometric> 725.8621+-6.1604 ^ 717.1344+-1.2986 ^ definitely 1.0122x faster Conf#1 Conf#2 Geomean of preferred means: <scaled-result> 30.2426+-0.0712 ^ 30.0891+-0.0366 ^ definitely 1.0051x faster
Mark Lam
Comment 3 2016-08-04 10:15:56 PDT
Comment on attachment 285305 [details] Patch View in context: https://bugs.webkit.org/attachment.cgi?id=285305&action=review r=me. > Source/JavaScriptCore/ChangeLog:17 > + One key insight is that using HashMap to map Nodes > + to Value in valuesAtTail is too inefficient at the scale > + we use it. > + > + We need some mapping from node to value, but HashMap > + does not cut it. Instead, I reuse our existing mapping > + from every Node to its value, abstracted by forNode(). The "We need some mapping from node to value, but HashMap does not cut it." part is redundant with (and not as informative as) the previous sentence. I suggest removing it > Source/JavaScriptCore/ChangeLog:19 > + Since we are not gonna use the mapping after endBasicBlock() nit: /gonna/going to/. > Source/JavaScriptCore/ChangeLog:23 > + In endBasicBlock(), valuesAtTail is now a vector all values live typo: /vector all/vector of all/? Also, you capitalize Vector below (as in a reference to the type). Do you want to capitalize it here? > Source/JavaScriptCore/dfg/DFGInPlaceAbstractState.cpp:179 > - bool changed = checkAndSet(block->cfaStructureClobberStateAtTail, m_structureClobberState); > + checkAndSet(block->cfaStructureClobberStateAtTail, m_structureClobberState); You're changing this to not use the "check" part (because the changed bool is unused). Why not just change this to an assignment? block->cfaStructureClobberStateAtTail = m_structureClobberState; > Source/JavaScriptCore/dfg/DFGInPlaceAbstractState.cpp:295 > + changed |= entry.value.merge(forNode(node)); Above this line, can you add the following? ASSERT(from->ssa->valuesAtTail.find(node)->value == forNode(node)); If I read your patch correctly, this is the crux of the patch i.e. you also stash the from->ssa->valuesAtTail.find(node)->value in forNode(node) back in InPlaceAbstractState::endBasicBlock() when you computed from->ssa->valuesAtTail.find(node)->value. This assertion documents this expectation of equivalence and ensures that nothing breaks it silently.
Benjamin Poulain
Comment 4 2016-08-04 11:46:40 PDT
Created attachment 285342 [details] Patch for landing
WebKit Commit Bot
Comment 5 2016-08-04 12:34:43 PDT
Comment on attachment 285342 [details] Patch for landing Clearing flags on attachment: 285342 Committed r204130: <http://trac.webkit.org/changeset/204130>
WebKit Commit Bot
Comment 6 2016-08-04 12:34:47 PDT
All reviewed patches have been landed. Closing bug.
Note You need to log in before you can comment on or make changes to this bug.