Bug 70578

Summary: DFG should not try to predict argument types by looking at the values of argument registers at the time of compilation
Product: WebKit Reporter: Filip Pizlo <fpizlo>
Component: JavaScriptCoreAssignee: Nobody <webkit-unassigned>
Status: RESOLVED FIXED    
Severity: Normal    
Priority: P2    
Version: 528+ (Nightly build)   
Hardware: All   
OS: All   
Bug Depends on:    
Bug Blocks: 69996    
Attachments:
Description Flags
the patch oliver: review+

Filip Pizlo
Reported 2011-10-20 20:58:46 PDT
DFG is invoked from loops or when a function returns. By that point, the registers holding arguments may hold a different value than what the arguments would have held at the prologue. Hence it is unprofitable to try to augment argument predictions by looking at argument registers at time of compilation. This is also an opportunity to kill some code.
Attachments
the patch (15.67 KB, patch)
2011-10-20 21:00 PDT, Filip Pizlo 		oliver: review+
DetailsFormatted DiffDiff
View All Add attachment proposed patch, testcase, etc.
Filip Pizlo
Comment 1 2011-10-20 21:00:45 PDT
Created attachment 111892 [details] the patch
Filip Pizlo
Comment 2 2011-10-20 21:02:15 PDT
This blocks https://bugs.webkit.org/show_bug.cgi?id=69996 in the sense that inlining reveals some bugs in trying to read argument registers during return. I don't want to fix those bugs, since this code is pointless anyway.
Oliver Hunt
Comment 3 2011-10-20 21:04:13 PDT
Comment on attachment 111892 [details] the patch Any perf impact?
Filip Pizlo
Comment 4 2011-10-20 21:09:51 PDT
(In reply to comment #3) > (From update of attachment 111892 [details]) > Any perf impact? None whatsover. :-)
Filip Pizlo
Comment 5 2011-10-20 21:16:33 PDT
Benchmark report for SunSpider, V8, and Kraken. VMs tested: "TipOfTree" at /Volumes/Data/pizlo/quinary/OpenSource/WebKitBuild/Release/jsc "NoCompTimeArgs" at /Volumes/Data/pizlo/senary/OpenSource/WebKitBuild/Release/jsc 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 NoCompTimeArgs SunSpider: 3d-cube 7.3723+-0.1421 7.3718+-0.1683 3d-morph 7.5918+-0.1756 ? 7.6909+-0.1443 ? might be 1.0130x slower 3d-raytrace 7.4912+-0.1396 ? 7.5606+-0.1668 ? access-binary-trees 1.6921+-0.0328 ? 1.7380+-0.0622 ? might be 1.0271x slower access-fannkuch 6.5964+-0.1435 6.4074+-0.1125 might be 1.0295x faster access-nbody 3.3254+-0.0996 ? 3.3516+-0.0844 ? access-nsieve 2.5642+-0.0691 2.5588+-0.0626 bitops-3bit-bits-in-byte 1.7538+-0.0365 ? 1.7549+-0.0348 ? bitops-bits-in-byte 2.7935+-0.0715 2.7635+-0.0485 might be 1.0109x faster bitops-bitwise-and 3.4800+-0.0846 3.4311+-0.1203 might be 1.0143x faster bitops-nsieve-bits 5.4330+-0.1049 ? 5.6038+-0.1305 ? might be 1.0314x slower controlflow-recursive 2.1088+-0.0302 2.0993+-0.0372 crypto-aes 6.7030+-0.1589 ? 6.7458+-0.1328 ? crypto-md5 2.8391+-0.0752 2.8132+-0.0666 crypto-sha1 2.5540+-0.0592 2.4981+-0.0651 might be 1.0224x faster date-format-tofte 9.9222+-0.1246 ? 10.2834+-0.6119 ? might be 1.0364x slower date-format-xparb 8.8578+-0.1634 ? 8.9056+-0.1820 ? math-cordic 6.3843+-0.1158 ? 6.4119+-0.0593 ? math-partial-sums 7.6487+-0.1453 ? 7.6995+-0.1202 ? math-spectral-norm 2.9134+-0.0943 2.8859+-0.0645 regexp-dna 11.4696+-0.1283 ? 11.6820+-0.2139 ? might be 1.0185x slower string-base64 4.4170+-0.0872 4.3391+-0.0614 might be 1.0180x faster string-fasta 6.4018+-0.1135 6.3781+-0.1241 string-tagcloud 11.4927+-0.2361 ? 11.5964+-0.2181 ? string-unpack-code 20.5333+-0.3631 20.1617+-0.1621 might be 1.0184x faster string-validate-input 5.2815+-0.0860 5.2707+-0.0874 <arithmetic> * 6.1393+-0.0354 ? 6.1540+-0.0280 ? <geometric> 5.0460+-0.0270 ? 5.0513+-0.0221 ? <harmonic> 4.1623+-0.0263 4.1611+-0.0291 TipOfTree NoCompTimeArgs V8: crypto 72.9880+-0.3284 ? 73.0395+-0.3671 ? deltablue 223.6140+-1.4193 ? 225.3372+-1.1424 ? earley-boyer 91.8322+-1.7376 ? 92.5651+-1.9726 ? raytrace 58.4976+-0.9168 ? 58.7943+-0.7432 ? regexp 106.2676+-0.3634 ? 107.4226+-0.9603 ? might be 1.0109x slower richards 182.5372+-0.5436 ! 184.3530+-0.2627 ! definitely 1.0099x slower splay 97.2508+-0.6331 96.4778+-0.4111 <arithmetic> 118.9982+-0.4084 ? 119.7128+-0.3716 ? <geometric> * 107.4427+-0.4634 ? 107.9671+-0.4743 ? <harmonic> 98.1164+-0.5333 ? 98.5162+-0.5509 ? TipOfTree NoCompTimeArgs Kraken: ai-astar 502.6183+-3.6015 ? 506.2388+-4.7345 ? audio-beat-detection 192.8065+-0.8792 ? 193.9236+-1.3722 ? audio-dft 272.8947+-1.8728 272.4993+-3.5137 audio-fft 125.4737+-0.6517 ? 126.1636+-1.0248 ? audio-oscillator 255.4600+-1.6487 255.3413+-1.9681 imaging-darkroom 422.4700+-0.4584 ^ 419.7746+-1.4165 ^ definitely 1.0064x faster imaging-desaturate 221.6168+-0.4444 221.3184+-0.8106 imaging-gaussian-blur 562.8225+-0.9280 ? 563.5646+-1.4490 ? json-parse-financial 57.5406+-0.4912 57.4979+-0.2811 json-stringify-tinderbox 68.5846+-0.3008 ? 68.7109+-0.3577 ? stanford-crypto-aes 133.2974+-2.0460 ? 134.2106+-1.9429 ? stanford-crypto-ccm 102.7151+-0.7707 101.8204+-0.5318 stanford-crypto-pbkdf2 195.1052+-1.2864 ! 198.6756+-2.0544 ! definitely 1.0183x slower stanford-crypto-sha256-iterative 71.6124+-0.2288 ? 71.8927+-0.6898 ? <arithmetic> * 227.5013+-0.4851 ? 227.9737+-0.7908 ? <geometric> 178.6384+-0.4575 ? 179.0348+-0.5937 ? <harmonic> 140.4540+-0.4462 ? 140.7147+-0.4567 ? TipOfTree NoCompTimeArgs All benchmarks: <arithmetic> 88.8857+-0.1819 ? 89.1410+-0.2656 ? <geometric> 23.0240+-0.0935 ? 23.0693+-0.0827 ? <harmonic> 7.3236+-0.0456 7.3221+-0.0503 TipOfTree NoCompTimeArgs Geomean of preferred means: <scaled-result> 53.1401+-0.1801 ? 53.3057+-0.1747 ?
Filip Pizlo
Comment 6 2011-10-21 01:21:16 PDT
Landed in http://trac.webkit.org/changeset/98082
Note You need to log in before you can comment on or make changes to this bug.