The DFG::shouldSpeculateXYZ methods have some hold-overs from the static speculation days, as well as some hold-overs from when nodes did not remember their predictions after propagation. They should now be simplified to just use the node's prediction, instead of also querying if it's a constant, what its generation info is, etc.
I don't know how or why but this is a performance progression: Benchmark report for SunSpider, V8, and Kraken. VMs tested: "TipOfTree" at /Volumes/Data/pizlo/quinary/OpenSource/WebKitBuild/Release/jsc "SimpleSpec" at /Volumes/Data/pizlo/quartary/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 SimpleSpec SunSpider: 3d-cube 7.4846+-0.1938 7.3904+-0.1589 might be 1.0127x faster 3d-morph 7.3235+-0.1097 ? 7.3891+-0.1346 ? 3d-raytrace 7.6127+-0.1929 ? 7.7139+-0.3390 ? might be 1.0133x slower access-binary-trees 1.7695+-0.0618 1.7312+-0.0529 might be 1.0221x faster access-fannkuch 6.3561+-0.1230 ? 6.4046+-0.1169 ? access-nbody 3.4766+-0.0878 3.4593+-0.0674 access-nsieve 2.6102+-0.0673 2.6090+-0.0636 bitops-3bit-bits-in-byte 1.6826+-0.0267 ? 1.6976+-0.0402 ? bitops-bits-in-byte 2.6937+-0.0781 ? 2.7330+-0.0736 ? might be 1.0146x slower bitops-bitwise-and 3.3954+-0.0767 3.3739+-0.1450 bitops-nsieve-bits 5.5211+-0.1235 5.4688+-0.1827 controlflow-recursive 2.0641+-0.0427 2.0219+-0.0362 might be 1.0209x faster crypto-aes 6.6308+-0.1933 6.4991+-0.1958 might be 1.0203x faster crypto-md5 2.8396+-0.0868 ? 2.8502+-0.0976 ? crypto-sha1 2.5092+-0.0877 2.4994+-0.0765 date-format-tofte 9.8208+-0.2289 9.6545+-0.1639 might be 1.0172x faster date-format-xparb 9.0871+-0.2624 ? 9.1752+-0.1927 ? math-cordic 6.6069+-0.1143 ? 6.7638+-0.1043 ? might be 1.0238x slower math-partial-sums 7.5417+-0.1169 ? 7.5588+-0.1163 ? math-spectral-norm 2.8797+-0.0740 2.7982+-0.0634 might be 1.0291x faster regexp-dna 10.7476+-0.1960 10.7336+-0.1182 string-base64 5.5433+-0.1854 5.4857+-0.1473 might be 1.0105x faster string-fasta 6.6210+-0.1331 6.5756+-0.1248 string-tagcloud 11.6307+-0.1731 11.4985+-0.2257 might be 1.0115x faster string-unpack-code 21.0739+-0.2537 21.0662+-0.2105 string-validate-input 6.2776+-0.1459 ? 6.3075+-0.1592 ? <arithmetic> * 6.2231+-0.0252 6.2100+-0.0290 <geometric> 5.1154+-0.0224 5.0997+-0.0362 <harmonic> 4.1997+-0.0283 4.1810+-0.0474 TipOfTree SimpleSpec V8: crypto 72.0093+-0.4413 ? 72.1674+-0.3549 ? deltablue 218.1208+-1.6580 216.5963+-1.1503 earley-boyer 87.0210+-0.3445 ? 87.3982+-0.2968 ? raytrace 59.0426+-0.9198 ^ 57.4898+-0.2626 ^ definitely 1.0270x faster regexp 103.3065+-0.4788 ? 103.4281+-0.3658 ? richards 208.0894+-1.3868 206.7795+-0.6596 splay 90.8671+-0.5588 ^ 89.7117+-0.4158 ^ definitely 1.0129x faster <arithmetic> 119.7795+-0.2368 ^ 119.0816+-0.2255 ^ definitely 1.0059x faster <geometric> * 106.7105+-0.2330 ^ 106.0308+-0.1675 ^ definitely 1.0064x faster <harmonic> 96.6827+-0.3646 ^ 95.9378+-0.1593 ^ definitely 1.0078x faster TipOfTree SimpleSpec Kraken: ai-astar 499.2093+-2.2369 ^ 492.1926+-2.0969 ^ definitely 1.0143x faster audio-beat-detection 191.4963+-1.1932 190.6489+-1.3148 audio-dft 271.2691+-2.8687 266.2812+-2.3493 might be 1.0187x faster audio-fft 125.6523+-0.2270 ? 126.5136+-0.7651 ? audio-oscillator 244.6052+-1.6766 ? 247.0725+-2.0585 ? might be 1.0101x slower imaging-darkroom 419.3021+-1.0464 ? 425.8465+-5.9273 ? might be 1.0156x slower imaging-desaturate 230.7208+-0.4309 ? 231.6531+-1.7522 ? imaging-gaussian-blur 583.7580+-1.5630 581.8522+-1.2108 json-parse-financial 48.6361+-0.2707 48.1809+-0.2972 json-stringify-tinderbox 69.4407+-0.4426 ^ 68.4374+-0.4581 ^ definitely 1.0147x faster stanford-crypto-aes 129.0581+-1.4862 128.3939+-1.4514 stanford-crypto-ccm 99.9955+-0.5684 ? 100.7834+-1.5850 ? stanford-crypto-pbkdf2 190.7082+-1.1493 ^ 187.6297+-0.5089 ^ definitely 1.0164x faster stanford-crypto-sha256-iterative 75.5406+-0.3227 75.1208+-0.4026 <arithmetic> * 227.0995+-0.4739 226.4719+-0.6494 <geometric> 176.4004+-0.4323 175.8015+-0.5336 <harmonic> 136.4659+-0.3668 135.7894+-0.4376 TipOfTree SimpleSpec All benchmarks: <arithmetic> 88.9287+-0.1346 88.6306+-0.2067 <geometric> 23.0881+-0.0641 23.0034+-0.1058 <harmonic> 7.3831+-0.0486 7.3499+-0.0813 TipOfTree SimpleSpec Geomean of preferred means: <scaled-result> 53.2280+-0.0818 ^ 53.0284+-0.1145 ^ definitely 1.0038x faster
Created attachment 110031 [details] the patch
Created attachment 110040 [details] the patch I realized that I had to make the patch bigger by adding more code. (I forgot about JITCodeGenerator64.)
Comment on attachment 110040 [details] the patch "at" might be a little too concise. I think "node" might be a better name. shouldSpeculate vs canSpeculate is still a bit mind blowing in the bad way. r=me
Landed in r96871.