ÿØÿà JFIF    ÿÛ „ !.%+&8&+/1555$;@;4?.451 4,$,44444444444414444444444444444444444444444444444444ÿÀ  á á" ÿÄ     ÿÄ ?    !1AQaq"2‘¡±ÁðBRbrÑá#‚’¢²3S CñÿÄ   ÿÄ !    !1QAa‘2ÿÚ   ? 5˜Z¯V¦cø)›t/? z¨±>Õ5€¶‹Á¤·¼z¼Ü¬+ñ®v¤¨_ˆR­BFn©—˜ý®ç̝P8gýt·ÉSTŦˆìät?þé¼íìN/Þa)ì–í6ô… Ï¿øÃj´¿KÇü]ÿ ªô¹-eKànëÕHTx}ýSÜ›ÿ ”7Ø×&µ<¦  ¥ÑO¶[Ù¯ä¨ÞÃÿ PZ-¬;#õ|•oaÿ ©CìÞz3˜öː/¤­ñTûIØ}š^ mÓ%ªxˆ¥ÉŸu=Z+ISe¿45™¼u;ú&WØ÷€æßQ™®{|íx*TC“#ZŠìZ§²‹ 6pv…³¿¡äª*áZÐ%ÒOáˆo"x«OHk w±æ+¬V(kMúŸ5Vö«$ ÁrÏbàb57/luR ¸ÑÛj Òµì`Ðœq­û žICÀÊ•©4€Âcà¨Ï€O´<èÐ:›ù(Ë^L8þ‘ÍÌ#¸Ð_Ì©ÙK(Öz 4¬û+¸;ü’V’84‘¬ÃŽ:[â‡ÔÌáõp¢~§ªlæ£ö{®G>J¼"°‡7¯ÆÉèßû ‹É‹§ÁòÃýâßî ^ƾÙõ‹×óH#«LP½ïX=xÑÍ$|W?•~• îëÔ©ª‹ {ÝT…Kÿ ”hûâá)J*ö˜–ÔU;iÇ€/ ÆþjóZ\ýwØ=Ìm ºèËL9 ýèÆð/¨’¥öo=nË.%Îì ŽÕ¯È|{Oj²ƒE6e/ßdÄõ²Ìâ1O®ò×TsəԸhOMýíMˆ¿¼H˜l²,7Â¥#MF/Úf°Ö½± ¸–dr‹NýÊ íjqx{œÉ ä-È ¦ øÄër¨q°ð †nцýÑÄÆ’mä…n<0È™;ÁÝá¯ÁZƒ7FÀmì­ É&9ˆîéi¶ùN§Y• ÃZãAâ?•‡©‰ , ó¾IŸŠc1 4â&y­&pŠ­6;M À 0¹qç»p.á …ŸÅáK@%6·y6ƒ‰3?”úºŽ‰éX5ªPT §µ!=Mž«Ú½‹ÅgÂSâÉaþÓoö–¯ÁÔìR>5éÿ üs¶ÆUcÌ kÇR ]ÿ ù¬¼«VŽ;Â|‡~¢¦”ÏÅ°æ {L™Õ°Óv¹ò¸írŽ¡×¢CÃ!íVÕ {¶»sŒNPg/ "uÕbkm²“$ďå¿é¹§°½æz¯6 †s¿!s–wÚÝ“™Œ °.ûj>·+™Òa…©Œ&rÝÎtÛë긪Ît’LAVp%c Úý[ÄzJ¾ÇàXXç@˜ó<êL]·T˜¾¥1Ó©V‡g´æ½¦Ý@¹óø!_@´ÞâSÁ —S3™•& ]@JHÚý©ZŽ €×æÔr»Áf!‡yÞ4Mv*èÓã_{‘åóUuљØ«Oïé*®EvÑ Œ÷‡U \"㪒ÍK+À 4“M¡ï:0¥5í!'<@î´”>Ç»&Z–ïCCV˜Ì5Šo&îhè.žû |ÓK©h$s6KìŒëã)¹hI¦GïOåóI;ììü#É$Š0…Ææ¥TØ.5­¾gn´ “ÂÖ\:hœ89G)J@„}œ:’Ò{/Š"¦_Æ×7Æ3VÇŠÊa]ÚŒÙ€Ä–=®uÁßâACZƒ§§£ Qnâ:«,×{tyø¬iÛcœÜÄ€H½ÄÍCk´÷šß .W'b¤Íåh]÷€=,Žv×cÚEÚHXJX¶îo¨FÒtèöŸ>ªª6[J®Fµ£sGÁeqõfe\íjÒÐïÄÐGˆe1Ø‹.Ø”‘Ëuø Y­ˆÜ ŽG|zùªüMpDnQWÄ”%JŠ™)â*p@Örš«ÕT2Ð%ˆG#ª„ ·¤!°ŸOTÂT¸aÚ%4&h™LµšØüÐ.F¿²ÐÞ_Ç‚¾ÅÃaÜ÷09Æ q€öy˜v‡85õN÷]¬äѼóS{°_MŽ¬úÔ#°Ç¸0åÞè2ëôPcvÆw9®ií1Ä8F™˜à‰´+‰Ik1òÝ7“Ñ×ÒsÝ\x‚h`ÞÑ`ó"|µEcý£n˜h`}GÞ !±ù²Ápü²ß6 0ïi󜵩SÈÇ7˜-ÕURO˜¦´f$ªž-Í6(œ}<„ éc øs]ŽŽ„*—¾ ìdŽ„)méª\¿êÎIg¾ØÞ~I#C/¼¼´EÁÈŽi8“©õådô·>euä ƒ'Ê×लR1ÉJE1ÐAát`t;ÇР%Ý<‡¥„ÍÆ`×Oyó)õiI€ñQaŸ4Ûù\áàaÃÔ¹HÃu¹*k€¦<„e S‡&õÏ B!ŽhüÞ`yj}mªf×\¿ Ç~æ­9‡û\՞Ǖg²1Žû5V7 !àöšm° c`ܬøÇìµÒ'P"?…´Ö,"§^•õŽ¨sÔ)6˜sæéÍR¼ ò|Sl”‹7 nPW Gòú÷½§O¯‡„l¡kSÞŒr½PÊ@æ¢pŽ-mÿ #Ÿ˜Àº¶Áä¦;ïÔæ$1££`“Õ>„—·ž)ßð³ñ#Ï Ô$¶œ‰ÊE‹À;÷º ¯«P:Ñ”8–IÊtpÞ3ª“>ê“þës4ò2OÏÕ­±zô†Õ§‰.÷ä¸;¿˜“'œ›žª}«Œ{ª±Ì 9ÔóÞÕ‡0 $íWV3Üì¬ —@kÝ4@¿r¼±½¬™›?øØæ´'Áé®CË3-g$˜ö‡×auÚi´Žp/êÛ æF›Ú2v‹ã¿¿,nB1̨ƃqÞa5͝@&Æû“él÷ \C²½UÍc ¯k×¢U ÖéQå™—-r wô ÞÏ<Ò=&=ÿ Ôê Òêˈt,i—;LîÜ á¸*ÚÃ1$êL•LÍ <É)ýÐà’ ;F™{ƒ™˜€&'}‚ãÄK`¡ÞT@I;®žZóè‚s’7®°›+§O­Åq©é»²9<Ô J ¼9O’HL»Ùïì¸rk¼Ž_ý‘TŸu[²ßÚŒ·ü÷B%¯E ŸÔX5êO´ Ç•€’I0 ÉJX` ñ¹õ%;µŸD‘«´€àwÒ™U ûئžÖö\×®×´8 ½‡ºÐÆÓ§?Àkmœ=;d5*@-ì0F Rªýš[Ü6âö̃ڸr*KA9· u*µæ£?U¸Âêí†8@¦X4 e-ò„0s{ HâUpU?¼mñRa°®a%Ð'tÉ×’\¾ÊÉ]t›h>·(Ë@R¼¡Ãt h}’O÷au<+nT…Ö…MӐ??Óe95 q>í/;&JSû °¯ÊéÞ øƒ*Ã2½Ài&:nôUl=¾¿5eˆ3”ñc|Ú2V”>„»&eE;«ÚäC p¢Û úy 9š[ŒÌx¼擼A&DåÒ¯ˆ¤ÀÌ;"˜ ÏQä¸åhÊ}Ûq«Û0WžÒ|»€ø®öCm5•\ÇÀ§Pe3£]0ÃàLDÉ‰1øªxjgwT‚÷¿LΨK‹›ùs—xˆÜ±µ kæ¸f‰‰ÜGk/LÛØ6d9ò¶ùA{ƒA3š/¬D¬khÓk‰`˜"㯒r¿±Óã jx‡°e}<Ñø\3y:'À•/h½Í€Ç4~g ?Û(¼]v‘ªlKÎâ~?O‚W%{Ì:“'©úNq¾›úo(X’¥¯ˆ nFê{Ç€ü?º'ë ø‹ì Þ09ŒÌç9Æ —ËC`j@ÓÄ(+a‹un¸#ÂꟋ{K`‘ÑÍÍ'à´»/Û,KW;Þ4²þð ï Nm|~fGÏ(…³Ã)«1ö­Õ ¥‡¨©ƒÃ™ü-s=à=U66Ï«Ýc蓦W¹íž®›nÔ%êÇìŒ<#Ü×84ån®Ð ÒåOC` ñânÑs‡¢ç 1õ%Îhì½Ã½® e:ݼUZo™`  ÅZŸŒÊ«ê1ÏÄo$q¹Þ€©ˆhÐÉä¯ñ[!…Ú˜àJ:x2$Íß&PåT£6ç— ‡Í*4Ýšçjÿ ‰É nófÐ ó(L5C•åÆ\rMÒ@ò }y-W}™üýVù—ú¢=Ù”c®‘< M ž ´Phr ¦©TD ‘ù.$´÷O‡‘V2Æò.=IUŒ=ž‡â¬i™aþÓåÙ?òUø'ØÖ•.~* šTŒ!•-×áºTâ®ä#õü'´ eýlYÅÓeÕKÂrT"CÚ@u!Óxƒ{š3€}1¿(r}%«nËamjÑ%ÑNEò v ˜à  σöK³,*º.àzù¨™Ó ÚçâU¦*¿ 9{%Ö¹ njûdaXöb) kÛƱûÓ\°M7ˆÂ=û›ç¿Ã‚­V»Cg–8ÙêE- j)k$º`Ã-ùEýeBÆÇ]c¡°ñty&Òd0nõ'¡W+ƒ*|–øµFa\GQªEAÔp5\Ǽ·¼Ç8·õ -â§Ú[ ‡ uZeÖ 3}×d'+¹:ð+K†Û®s!Ï$úe€<Û”x)1»a­¡LC]¸µík…ÚàA»AYº{†ªS[¦5HÒ7ù --,ísòDØ€èk ÞÀîÜ ò@â( ËNˆë›4ô½•/¦o‡€Û7 ê•ÆêòðÜy'Án½µ á˜ݦ ndeo…[ì¶Ê,¥R³Ä=À±—–ß;£™´ñSâ*g§”ïaið‘Jå~™ÓÞ ß³Õ¢»8x埒²52>AÊb&-÷\7´éÄù€T˜,w;3{ï˜k…à¹ÄqÀ«œ{€\ ˆ¾[´¨јr &Úé„Ívˆ±8†¿]|¬ņ4I×pÞS1ÈÖz‰#Ìv‡G!YNògñ:màTz¢Ý1ô©^O=~ë|5Bã™ç•¼µõ•bÆ@úÕS¬ÈŒ#¬zünrŸ û” Z²•ï›èðV"ÁHÚý©wÝ €7¼Ìu1hÑa3Éä û f$o¿É ™Ú›ÝçnpÒ3äÌ3†Í§,Äï]$‰/pê †«À¼¸e9­Æê_C]žƒ·ý·frÁN«, E=›Çq -‰öŒ:aÏ¿±í&£Í:-} 84‘ÿ eƒQÑeëSsuiA ³g㟥ú£?ÿ ʼn*”“÷aühe:ÊWa@ÒÞk±eØ] F Ô—r.åä˜ @ö¥ªZoÐýYL·¥S²G/‡ñ <~*ZÆ´è>JlòàÛƽÿ 窘ìGN¢:I®KšJp/`íIÁÀõ#Ä-€ö­šµŒoF4|ÆQØÆ@Ì|£Ô…¢À{9˜è½Üó›€ôYÒÎYsið;ís¤€à²ˆ‚4qÉVŒI$ ‰"° æµ8cXGjœˏ¡Aâý•ËÜ¢ûï e·çLx']á"oÅÎê3¯Ç—¹”ó0nå‚âg{Œñ> S´˜îè°g238‚ãköÝfÚd´6Ò€;ò÷±¢™¼›º ¢Æ'¥Ðx'e¬ç ]bÈÆV¢ó‹kýBO ðÊâ$Ÿ!×T 3Mýמ žìٍàÌü‘8÷€àæØ8æ©6‰©L´«…oãpð„~Çk‰!ñ;‹”ÛžÍ àž±z Ÿôû øŸÝužÏ;ÿ #|u6™Þ¬ÚˆÐõA4¶â|ôl|Ê2ŽÇ¤ÝÅÇY.<#Aí.k§hóF‚”Y; M½Ö4hŸ4&›­¿tès´%FìL¥£Ãk‰ÇT¤haÁ¤ÚxfÉ`ÑìË›>i 3t‚:,–+^÷´–{Û–Nxi"x‘Ûg î¨>¥Õ܁ùZH,2Û“:8xÊ¢Çí9.É-Ìâã-=çjwµS˜dütžçwýGòú®®ûº_ˆýx$–¡ãøO EÚÛÏ÷R„×w+3£Á£öUMyR²¹âŒ°š›¸Ñãò9§Ó_Dl+Ùßc›úšGÅÌc†Ž!Ko=¶.‘Îÿ c²(2®V mª.ÿ ¹B›¹å ù„öŸSV>™ü¯$y:G¢Z×àøúdî¹û­·ýÇ´:•c LÍõi_‹ö+ÎæGÊè>OŠ•äž´§Þ{X}¨1ÚTc›»Qþ•êô°t¿OP?eæ~É{5]•ÙR£r5†nZ\ã@ &îJõ ¾àC°þV>fé¥/ü5ñÊIº_é5 ;e­h<@ Ä&æÃëE%;X,ÒãÆÞ`Oò¦kŸm#˜!ÀyÄ¢| óLšò¥Ä` ¶R=|ÈCâh5ò3DˆïF†ðÒ#ÅìÛœ?¸yhBãœí ZxßÎÄhºRK„`Þödvײ™ÀÈÑÒgŒuY w³%†ƒÓzõ ÖÏp‚dH®¦A´ù§»ÓÇMæ~)ˆð‡û:ù&Ä •vGD´À n ݇¼Ö8Fö óáà£~Ë¥x`oK|Ä?fxiØü%pìR>éò+Û±éÎ>núlFŤ'tq8LZÏvÃ?„¡ß±È⽆¯³íü@x|PöUäèØã¡ð‚ŒAìÏ"vÍwóŸÍ{ ý0.z È•Ö{,N¡£¡ŸKÕÙž>Ýœþ ÍÀ°<×EA!Å‚D™IúOÍ¡>ôG}Â` ÍßkÜL™Ž Þð™ {IøF²¹òQ3&!ÃÂÞz.d&Ï-sH¸,Ôõ˜ŽP€ 77ˆÝ¼ÊëÜw =cÕ Ú,ØÐ5ÎYÐ)ì´öœgŒ[¤ßv㙑8心>h]§µháYš£²ºÑ.{Ï7Sð•?´~×SÃKýJÛ˜ ™Íäiúu<µX¶1õ^kâçIÑ£sZ4h>j*ÔšD:4­¿_ ÷¸ Õxæÿ ¸?Mù _•­ÊÐ ä ÷ý ÑwL œ­ïnTkÛUÍN©ë:¦fV ¶ÜÔÜMªÅâA½–¿R×TXš-%iTÊT•‡Ù‚JôϐZxWÑè‰f‰òG º ×Õû2aZ7OU3[“×AT–ÞŒ…-‘¤”Ì ì&(ˆ¿­•ƒkï’:ðY¦W‘ Å)“†‘˜³Åtcø˜ñTÂwÚÇ4|üLÇªí–v- qˆèU qPE.†â‘˜µ Æ,ÐÅs]8¾„oúÑ i>ÜxxÈó)ƒ ´æÁâØ$À‰vžŸf$Ž |ãw;ÀÁIJ»b` {¦Ó¤Ú$©YÀ‘n@Óïž«9J¼êG m¤ ܯ¹ÌW4€ÐÒÅÛ‡#褕Ÿn-?í|с¥÷Ú¹¬'´ÞÜ9ÓK `hê£SÄSà?7—Wí_´…óB›»:=Ãïq`<8ñÓŒÑlú2d¬ê³£hÖ[l|$vÝro~'R®‰§°ñmY ͧäP |PUª¹·:3Œ[Û{Xÿ ºâ@‚W–Äé u‚ ¯´*=íή.pûÒdt @G‰¬ s¸ ëÉücr ÞæѨÊ@>¤¢Ö±. Þ'¯°ÌME[YéïĵÂCå½ Ué©Áû'Ê9%eÔðNU”ë‘ÌsD3/®+UI˜9h.WC”빓$#:pz:YÓ ¿xž* ³$Í +$kñAŠ‹†¢ Uê>¸)_š¬÷©ßAÂÔb9ÇU ¯¾á•9¯ÏÏ÷O÷¼¼Fähal1‰3Ì[Ïr•´UCksNÐ] R‘¸¥H+§Šé†c©vÖÞ0iÓ76s†î!§=ß ¼~Ô'°Ãmäoäš³ªøi1úÉ)³yV8 CLÄØÁ‘WYïi€H6ÖÑiámø^ÈY´°Ñ7¥Û*—Ñ©L«Qƒï—Ùrÿ ›£Ð*š¸ˆL©ˆ$ˆ ÷¾D§9È®«qbqC)–ˆïv´çñsÑVT­Ø, <àïºÀO«Jý·õ àfPìð .wFšir´þ’2_Y *Æ€x\« ì€9š@ Ž|F⇥ˆkZ@hÖÄ0t¿-<“‹qµ¾*ZL¤Ú)&BJpÓF5=$„at*Zš$’ÑtdûÝRI1 2Ž‰$€$I$#‰SÞ’Hë¬ï;Á$¡t$’`<(ñÇt)$‡Ð.Êf¢X’Kt=Éé$‚ˆªè¢oÝëòI%Rgcª÷ŠyI%¡‰ÿ !ñ)´õ $¤ Ô’IIGÿÙfrom test import support, seq_tests import unittest import gc import pickle # For tuple hashes, we normally only run a test to ensure that we get # the same results across platforms in a handful of cases. If that's # so, there's no real point to running more. Set RUN_ALL_HASH_TESTS to # run more anyway. That's usually of real interest only when analyzing, # or changing, the hash algorithm. In which case it's usually also # most useful to set JUST_SHOW_HASH_RESULTS, to see all the results # instead of wrestling with test "failures". See the bottom of the # file for extensive notes on what we're testing here and why. RUN_ALL_HASH_TESTS = False JUST_SHOW_HASH_RESULTS = False # if RUN_ALL_HASH_TESTS, just display class TupleTest(seq_tests.CommonTest): type2test = tuple def test_getitem_error(self): t = () msg = "tuple indices must be integers or slices" with self.assertRaisesRegex(TypeError, msg): t['a'] def test_constructors(self): super().test_constructors() # calling built-in types without argument must return empty self.assertEqual(tuple(), ()) t0_3 = (0, 1, 2, 3) t0_3_bis = tuple(t0_3) self.assertTrue(t0_3 is t0_3_bis) self.assertEqual(tuple([]), ()) self.assertEqual(tuple([0, 1, 2, 3]), (0, 1, 2, 3)) self.assertEqual(tuple(''), ()) self.assertEqual(tuple('spam'), ('s', 'p', 'a', 'm')) self.assertEqual(tuple(x for x in range(10) if x % 2), (1, 3, 5, 7, 9)) def test_keyword_args(self): with self.assertRaisesRegex(TypeError, 'keyword argument'): tuple(sequence=()) def test_truth(self): super().test_truth() self.assertTrue(not ()) self.assertTrue((42, )) def test_len(self): super().test_len() self.assertEqual(len(()), 0) self.assertEqual(len((0,)), 1) self.assertEqual(len((0, 1, 2)), 3) def test_iadd(self): super().test_iadd() u = (0, 1) u2 = u u += (2, 3) self.assertTrue(u is not u2) def test_imul(self): super().test_imul() u = (0, 1) u2 = u u *= 3 self.assertTrue(u is not u2) def test_tupleresizebug(self): # Check that a specific bug in _PyTuple_Resize() is squashed. def f(): for i in range(1000): yield i self.assertEqual(list(tuple(f())), list(range(1000))) # We expect tuples whose base components have deterministic hashes to # have deterministic hashes too - and, indeed, the same hashes across # platforms with hash codes of the same bit width. def test_hash_exact(self): def check_one_exact(t, e32, e64): got = hash(t) expected = e32 if support.NHASHBITS == 32 else e64 if got != expected: msg = f"FAIL hash({t!r}) == {got} != {expected}" self.fail(msg) check_one_exact((), 750394483, 5740354900026072187) check_one_exact((0,), 1214856301, -8753497827991233192) check_one_exact((0, 0), -168982784, -8458139203682520985) check_one_exact((0.5,), 2077348973, -408149959306781352) check_one_exact((0.5, (), (-2, 3, (4, 6))), 714642271, -1845940830829704396) # Various tests for hashing of tuples to check that we get few collisions. # Does something only if RUN_ALL_HASH_TESTS is true. # # Earlier versions of the tuple hash algorithm had massive collisions # reported at: # - https://bugs.python.org/issue942952 # - https://bugs.python.org/issue34751 def test_hash_optional(self): from itertools import product if not RUN_ALL_HASH_TESTS: return # If specified, `expected` is a 2-tuple of expected # (number_of_collisions, pileup) values, and the test fails if # those aren't the values we get. Also if specified, the test # fails if z > `zlimit`. def tryone_inner(tag, nbins, hashes, expected=None, zlimit=None): from collections import Counter nballs = len(hashes) mean, sdev = support.collision_stats(nbins, nballs) c = Counter(hashes) collisions = nballs - len(c) z = (collisions - mean) / sdev pileup = max(c.values()) - 1 del c got = (collisions, pileup) failed = False prefix = "" if zlimit is not None and z > zlimit: failed = True prefix = f"FAIL z > {zlimit}; " if expected is not None and got != expected: failed = True prefix += f"FAIL {got} != {expected}; " if failed or JUST_SHOW_HASH_RESULTS: msg = f"{prefix}{tag}; pileup {pileup:,} mean {mean:.1f} " msg += f"coll {collisions:,} z {z:+.1f}" if JUST_SHOW_HASH_RESULTS: import sys print(msg, file=sys.__stdout__) else: self.fail(msg) def tryone(tag, xs, native32=None, native64=None, hi32=None, lo32=None, zlimit=None): NHASHBITS = support.NHASHBITS hashes = list(map(hash, xs)) tryone_inner(tag + f"; {NHASHBITS}-bit hash codes", 1 << NHASHBITS, hashes, native32 if NHASHBITS == 32 else native64, zlimit) if NHASHBITS > 32: shift = NHASHBITS - 32 tryone_inner(tag + "; 32-bit upper hash codes", 1 << 32, [h >> shift for h in hashes], hi32, zlimit) mask = (1 << 32) - 1 tryone_inner(tag + "; 32-bit lower hash codes", 1 << 32, [h & mask for h in hashes], lo32, zlimit) # Tuples of smallish positive integers are common - nice if we # get "better than random" for these. tryone("range(100) by 3", list(product(range(100), repeat=3)), (0, 0), (0, 0), (4, 1), (0, 0)) # A previous hash had systematic problems when mixing integers of # similar magnitude but opposite sign, obscurely related to that # j ^ -2 == -j when j is odd. cands = list(range(-10, -1)) + list(range(9)) # Note: -1 is omitted because hash(-1) == hash(-2) == -2, and # there's nothing the tuple hash can do to avoid collisions # inherited from collisions in the tuple components' hashes. tryone("-10 .. 8 by 4", list(product(cands, repeat=4)), (0, 0), (0, 0), (0, 0), (0, 0)) del cands # The hashes here are a weird mix of values where all the # variation is in the lowest bits and across a single high-order # bit - the middle bits are all zeroes. A decent hash has to # both propagate low bits to the left and high bits to the # right. This is also complicated a bit in that there are # collisions among the hashes of the integers in L alone. L = [n << 60 for n in range(100)] tryone("0..99 << 60 by 3", list(product(L, repeat=3)), (0, 0), (0, 0), (0, 0), (324, 1)) del L # Used to suffer a massive number of collisions. tryone("[-3, 3] by 18", list(product([-3, 3], repeat=18)), (7, 1), (0, 0), (7, 1), (6, 1)) # And even worse. hash(0.5) has only a single bit set, at the # high end. A decent hash needs to propagate high bits right. tryone("[0, 0.5] by 18", list(product([0, 0.5], repeat=18)), (5, 1), (0, 0), (9, 1), (12, 1)) # Hashes of ints and floats are the same across platforms. # String hashes vary even on a single platform across runs, due # to hash randomization for strings. So we can't say exactly # what this should do. Instead we insist that the # of # collisions is no more than 4 sdevs above the theoretically # random mean. Even if the tuple hash can't achieve that on its # own, the string hash is trying to be decently pseudo-random # (in all bit positions) on _its_ own. We can at least test # that the tuple hash doesn't systematically ruin that. tryone("4-char tuples", list(product("abcdefghijklmnopqrstuvwxyz", repeat=4)), zlimit=4.0) # The "old tuple test". See https://bugs.python.org/issue942952. # Ensures, for example, that the hash: # is non-commutative # spreads closely spaced values # doesn't exhibit cancellation in tuples like (x,(x,y)) N = 50 base = list(range(N)) xp = list(product(base, repeat=2)) inps = base + list(product(base, xp)) + \ list(product(xp, base)) + xp + list(zip(base)) tryone("old tuple test", inps, (2, 1), (0, 0), (52, 49), (7, 1)) del base, xp, inps # The "new tuple test". See https://bugs.python.org/issue34751. # Even more tortured nesting, and a mix of signed ints of very # small magnitude. n = 5 A = [x for x in range(-n, n+1) if x != -1] B = A + [(a,) for a in A] L2 = list(product(A, repeat=2)) L3 = L2 + list(product(A, repeat=3)) L4 = L3 + list(product(A, repeat=4)) # T = list of testcases. These consist of all (possibly nested # at most 2 levels deep) tuples containing at most 4 items from # the set A. T = A T += [(a,) for a in B + L4] T += product(L3, B) T += product(L2, repeat=2) T += product(B, L3) T += product(B, B, L2) T += product(B, L2, B) T += product(L2, B, B) T += product(B, repeat=4) assert len(T) == 345130 tryone("new tuple test", T, (9, 1), (0, 0), (21, 5), (6, 1)) def test_repr(self): l0 = tuple() l2 = (0, 1, 2) a0 = self.type2test(l0) a2 = self.type2test(l2) self.assertEqual(str(a0), repr(l0)) self.assertEqual(str(a2), repr(l2)) self.assertEqual(repr(a0), "()") self.assertEqual(repr(a2), "(0, 1, 2)") def _not_tracked(self, t): # Nested tuples can take several collections to untrack gc.collect() gc.collect() self.assertFalse(gc.is_tracked(t), t) def _tracked(self, t): self.assertTrue(gc.is_tracked(t), t) gc.collect() gc.collect() self.assertTrue(gc.is_tracked(t), t) @support.cpython_only def test_track_literals(self): # Test GC-optimization of tuple literals x, y, z = 1.5, "a", [] self._not_tracked(()) self._not_tracked((1,)) self._not_tracked((1, 2)) self._not_tracked((1, 2, "a")) self._not_tracked((1, 2, (None, True, False, ()), int)) self._not_tracked((object(),)) self._not_tracked(((1, x), y, (2, 3))) # Tuples with mutable elements are always tracked, even if those # elements are not tracked right now. self._tracked(([],)) self._tracked(([1],)) self._tracked(({},)) self._tracked((set(),)) self._tracked((x, y, z)) def check_track_dynamic(self, tp, always_track): x, y, z = 1.5, "a", [] check = self._tracked if always_track else self._not_tracked check(tp()) check(tp([])) check(tp(set())) check(tp([1, x, y])) check(tp(obj for obj in [1, x, y])) check(tp(set([1, x, y]))) check(tp(tuple([obj]) for obj in [1, x, y])) check(tuple(tp([obj]) for obj in [1, x, y])) self._tracked(tp([z])) self._tracked(tp([[x, y]])) self._tracked(tp([{x: y}])) self._tracked(tp(obj for obj in [x, y, z])) self._tracked(tp(tuple([obj]) for obj in [x, y, z])) self._tracked(tuple(tp([obj]) for obj in [x, y, z])) @support.cpython_only def test_track_dynamic(self): # Test GC-optimization of dynamically constructed tuples. self.check_track_dynamic(tuple, False) @support.cpython_only def test_track_subtypes(self): # Tuple subtypes must always be tracked class MyTuple(tuple): pass self.check_track_dynamic(MyTuple, True) @support.cpython_only def test_bug7466(self): # Trying to untrack an unfinished tuple could crash Python self._not_tracked(tuple(gc.collect() for i in range(101))) def test_repr_large(self): # Check the repr of large list objects def check(n): l = (0,) * n s = repr(l) self.assertEqual(s, '(' + ', '.join(['0'] * n) + ')') check(10) # check our checking code check(1000000) def test_iterator_pickle(self): # Userlist iterators don't support pickling yet since # they are based on generators. data = self.type2test([4, 5, 6, 7]) for proto in range(pickle.HIGHEST_PROTOCOL + 1): itorg = iter(data) d = pickle.dumps(itorg, proto) it = pickle.loads(d) self.assertEqual(type(itorg), type(it)) self.assertEqual(self.type2test(it), self.type2test(data)) it = pickle.loads(d) next(it) d = pickle.dumps(it, proto) self.assertEqual(self.type2test(it), self.type2test(data)[1:]) def test_reversed_pickle(self): data = self.type2test([4, 5, 6, 7]) for proto in range(pickle.HIGHEST_PROTOCOL + 1): itorg = reversed(data) d = pickle.dumps(itorg, proto) it = pickle.loads(d) self.assertEqual(type(itorg), type(it)) self.assertEqual(self.type2test(it), self.type2test(reversed(data))) it = pickle.loads(d) next(it) d = pickle.dumps(it, proto) self.assertEqual(self.type2test(it), self.type2test(reversed(data))[1:]) def test_no_comdat_folding(self): # Issue 8847: In the PGO build, the MSVC linker's COMDAT folding # optimization causes failures in code that relies on distinct # function addresses. class T(tuple): pass with self.assertRaises(TypeError): [3,] + T((1,2)) def test_lexicographic_ordering(self): # Issue 21100 a = self.type2test([1, 2]) b = self.type2test([1, 2, 0]) c = self.type2test([1, 3]) self.assertLess(a, b) self.assertLess(b, c) # Notes on testing hash codes. The primary thing is that Python doesn't # care about "random" hash codes. To the contrary, we like them to be # very regular when possible, so that the low-order bits are as evenly # distributed as possible. For integers this is easy: hash(i) == i for # all not-huge i except i==-1. # # For tuples of mixed type there's really no hope of that, so we want # "randomish" here instead. But getting close to pseudo-random in all # bit positions is more expensive than we've been willing to pay for. # # We can tolerate large deviations from random - what we don't want is # catastrophic pileups on a relative handful of hash codes. The dict # and set lookup routines remain effective provided that full-width hash # codes for not-equal objects are distinct. # # So we compute various statistics here based on what a "truly random" # hash would do, but don't automate "pass or fail" based on those # results. Instead those are viewed as inputs to human judgment, and the # automated tests merely ensure we get the _same_ results across # platforms. In fact, we normally don't bother to run them at all - # set RUN_ALL_HASH_TESTS to force it. # # When global JUST_SHOW_HASH_RESULTS is True, the tuple hash statistics # are just displayed to stdout. A typical output line looks like: # # old tuple test; 32-bit upper hash codes; \ # pileup 49 mean 7.4 coll 52 z +16.4 # # "old tuple test" is just a string name for the test being run. # # "32-bit upper hash codes" means this was run under a 64-bit build and # we've shifted away the lower 32 bits of the hash codes. # # "pileup" is 0 if there were no collisions across those hash codes. # It's 1 less than the maximum number of times any single hash code was # seen. So in this case, there was (at least) one hash code that was # seen 50 times: that hash code "piled up" 49 more times than ideal. # # "mean" is the number of collisions a perfectly random hash function # would have yielded, on average. # # "coll" is the number of collisions actually seen. # # "z" is "coll - mean" divided by the standard deviation of the number # of collisions a perfectly random hash function would suffer. A # positive value is "worse than random", and negative value "better than # random". Anything of magnitude greater than 3 would be highly suspect # for a hash function that claimed to be random. It's essentially # impossible that a truly random function would deliver a result 16.4 # sdevs "worse than random". # # But we don't care here! That's why the test isn't coded to fail. # Knowing something about how the high-order hash code bits behave # provides insight, but is irrelevant to how the dict and set lookup # code performs. The low-order bits are much more important to that, # and on the same test those did "just like random": # # old tuple test; 32-bit lower hash codes; \ # pileup 1 mean 7.4 coll 7 z -0.2 # # So there are always tradeoffs to consider. For another: # # 0..99 << 60 by 3; 32-bit hash codes; \ # pileup 0 mean 116.4 coll 0 z -10.8 # # That was run under a 32-bit build, and is spectacularly "better than # random". On a 64-bit build the wider hash codes are fine too: # # 0..99 << 60 by 3; 64-bit hash codes; \ # pileup 0 mean 0.0 coll 0 z -0.0 # # but their lower 32 bits are poor: # # 0..99 << 60 by 3; 32-bit lower hash codes; \ # pileup 1 mean 116.4 coll 324 z +19.2 # # In a statistical sense that's waaaaay too many collisions, but (a) 324 # collisions out of a million hash codes isn't anywhere near being a # real problem; and, (b) the worst pileup on a single hash code is a measly # 1 extra. It's a relatively poor case for the tuple hash, but still # fine for practical use. # # This isn't, which is what Python 3.7.1 produced for the hashes of # itertools.product([0, 0.5], repeat=18). Even with a fat 64-bit # hashcode, the highest pileup was over 16,000 - making a dict/set # lookup on one of the colliding values thousands of times slower (on # average) than we expect. # # [0, 0.5] by 18; 64-bit hash codes; \ # pileup 16,383 mean 0.0 coll 262,128 z +6073641856.9 # [0, 0.5] by 18; 32-bit lower hash codes; \ # pileup 262,143 mean 8.0 coll 262,143 z +92683.6 if __name__ == "__main__": unittest.main()