people online welcome to everybody herejust gonna be really quick veryefficient dyslexia us plus not going totalk very long I just want to say a fewthanks and the people who helped makethis possible that's of course thepeople in the back the channel nine teamstreaming live Denis Begley who made allthe logistics possible and what we'regonna do is invite herb Sutter up herein about 20 seconds he's going to make avery special announcement so I justwanted to say again thank you welcomeand her come up here manthank you Charles and thank youeverybody for coming and all those whoare watching even in the room we havefolks from 18 countries 23 states in theUnited States and a worldwide audienceso thank you for attending and before Iinvite our most distinguished and mostimportant speaker we wanted to dedicatethis conference to a very importantgentleman so if we could have hispicture on the screen please we'd liketo dedicate this- one of the most influential computerscientists in our young industry'shistory there are very few people on theface of the planet who have made asgreat a contribution and he's also thereason why we're here today when wethink about what happened at Bell Labswhere yarda and AT&T and c+ also grew upbefore there was a C with classes beforethere was C++ there was C the mostinfluential programming language in theworld every single modern language I wasa debt to it and is either directlyderived from it or borrows heavily fromit and has been influenced in all thecoding we do by what this man did 40years ago now he was also famous foranother thing the UNIX operating systemtogether with Ken Thompson which hasagain directly influenced or actuallydirectly led to almost every majoroperating system in use today inmainstream that's a very big set ofshoulders so we appreciate very much hiscontribution and we think of all theawards that Ritchie has received it'sbecause coming back to see it is such anempowering part of native code it iswhat has made our industry possiblebecause libraries that are portable thatare efficient are what this world isbuilt on this world runs on C and C++ itis the foundation of everythingincluding the manage languages it iswhat we all build on and so we'd like todedicate this event to the memory of therecently passed Dennis Ritchie anycreator of C co-creator of UNIX to whomwe all and everybody who uses softwareon this planet owes a great debt so withthat we'd like to dedicate that and I'dlike now to introduce our keynotespeaker the artist drew soup needs nointroduction he's the creator of scenecloth sauce and a friend of DennisRitchie and we would like now to see C++11 style a touch of class the Articthankyes thank you very much Dennis as muchmissed his he was a a real realgentleman and when I talk about memorymodel issues that's borrowed straightfrom C and that's a very important partof C++ okay so I have to talk about inthe future and I'm going to talk aboutthe issue of style how do you write codeand how do you write code so it'smaintainable performing all of thesegood things when I when I come into aroom especially one of these virtualrooms on the web I seem to be the onlyone who doesn't quite know what C++ iseverybody has some kind of simpleexplanation that that that fits theirview of it and you see a few of themthere and the effect is like the blindmen and the elephant if you have onlyseen the tail of the elephant you knowwhat it's like if you only seen a legyou know what it's like I try to to lookat the whole elephant and I try tocharacterize it but it's hardthere's many bits to the elephant theyare all essential for the elephant butit's it's hard to be simple about whatit is so - here's my sort of trying tocharacterize what C++ is particularlygood at and particularly particularlyaimed at it takes off from C and it onsystems programming but where it reallymatters where it makes a difference isin the area of infrastructure that iswhere you build something that otherthings build on that people rely onwhere you have resource constraintsthere can be a cellphone it can be aGoogle style lookupyou're constrained by various resourcesand you have to have reliability anddependability and performance that kindof application we can be infrastructureor it can be applications it has asimilar kind of constraints is where C++shines I tried to come up with abuzzword but I'm not very good at itlightweight abstraction programminglanguage is accurate but it's probablynot good marketing speech so let's go onone thing that's I think it's importantto realize when you talk aboutprogramming is that no one-size-fits-allI always hear these things like what isthe programming language what's thedominant programming language what's thefuture programming language which is thebest programming language thisprogramming language who fought for whatthere's no reason to believe thatthere's one language that's good forthat's best for everything I don't thinkthere is and so I'm trying tocharacterize what C++ is good for whereit fitsthere's things I would probably useJavaScript or Python for but I probablywouldn't use them for say the JavaScriptinterpreter that's that C++ domainanyway so dependent differentapplications have different constraintsyou you worry about hardware resourcesreliability efficiency constraint itvaries a lot of applications can crashdoesn't hurt anybody the extremes arewhere all that matters is to get to themarket fastthat's not C++ as core domainthere are however places in the otherextreme program fails people die it'sit's always uncomfortable if your tumorsare used in such an area but somebodyhas to do it and there's a simple factthat it's 3050 percent I mean if youdouble the overhead ofsomething working on a server farm wellyou need at least another server farmand probably too because of thecommunication so that's the areas wherewhere C++ shines over the years we havetried to say what we're doing this heprosper us with many words I mean peoplewant true object-oriented style I'venever been on that team there's nothingthat says everything fits in a hierarchyas a matter of fact it doesn't althoughhe misses and well C is perfect and youjust need a few details well if youwrite C with void stars and macros andcosts you get C style bugs and and yousee all of this in in the C++ world andthe C++ code and you just tried to saymulti-paradigm but that doesn't actuallycapture anything they were never meantto be separate ways of programming theywere meant to take part of a synthesisthat is C++ and so we can think aboutwhat we really want we want somethingthat's easy to understand for humans andtrue we want something that's correctmaintainable modularity is good theinterfaces are absolutely essential inin in modern software we want to haveeffective resource management we don'twant to leak memory we don't want toleak locks we don't want to lead fileleaked files etc etc that's a big issuetoday's world tend to be multi-threadedconcurrent so we want to be able to dothread safety without losing all theother advantages efficiency we want inmany many applications it seems to begetting more critical in critical areascell phone and and sort of webserverkind of things are the extremes but alsosome high-performance numerical stuffand portability except when when there'sreasons not to because next year's modelnext year's operating system is going tobe different if you don't want to portto anything else at least you want toport to your future hardwareand that's not going to be the same sothat's what we want that's hard to getI picked Don Quixote here from thisslide for for that reason becausesometimes it feels like an impossibletask but we're making progress we'vemade steady progress or many years andnow we are we're actually better thanwhat we used to be by March by largemeasure so we're looking for synthesisintegrate the set of language featuresand si+ processes step in that directionand then we need an articulatedguideline for how to use it if you justuse the new features just as featureswe're wasting time and if we don't usethem while we're getting the old stuffwe need to use a new set of features newset of techniques to gain what we reallywantdependable well performing maintainablesoftware and it's those guidelines Irefer to when I'm calling when I talkabout style so here's an overview firstI'll give you an example of what I thinkis ghastly I don't like to do withoutthen I'm going to talk about the variousaspects of this as this stuff and that'sthat's that there's a paper that I wrotethis one doesn't work over there on thescreen but you see that paper softwarestructure for infrastructure softwaredevelopment for infrastructure that'spaper I wrote at the same time as I waswriting this talk so there's greatoverlap in the examples there's not awritten version of this talk but if youread that paper you'll do fine it's inthis it's a key paper and this week inthis monstersa terribly computer so I'm going to talkabout C++ 11 but not really about thelanguage features I'm going to talkabout how to use it well and this ispart of a long term project to figureout how to use C++ well how to figureout how to blend the language featuresinto coherentstyle and I'm not going to talk aboutthe individual features except when Iabsolutely have to if you want to have afeature by feature explanation of C++ 11which was approved last year there's apicture of the committee in Spainsharing after the final vote then youcan go to my C++ 11 FAQ and there'sfeature feature feature feature and ifthat's your view of the world that'sfine that's where you look I'm trying tosee how to blend it most of this isshipping this is not a science fictiontalk no compiler ships every detail andbut all the major compilers shipimportant parts of the of C++ 11 so youcan experiment with it you can actuallyuse some of it in shipping code and thestandard library is shipping basicallycomplete so here is some piece of codethat I think is pretty ugly and it'sasking for inefficiencies and asking forerrors this happens to be Q sort verycommon very popular nineteen sort ofearly seventies kind of function it'ssort some memory you have to give thenumber of bytes of elements you have towrite an element comparison function andpassed it if you if you have to explainthis to a novice you're probably introublelike I have some calls of Q sort thereand I have to give the number ofelements I mean why don't why doesn'tdoesn't the array know it's a number ofelements why does an array of dobrowdoesn't know the size of a double welland why does the language not know howto compare doubles the pipe point isthat we have carefully thrown away allthe useful information that could haveanswered those questions and we'redealing with memories and number ofbytes in memory and that's just throwingaway you have to use an inefficientindirect function call that preventsinlining just to make sure that thingsare slowand you have many opportunities forwriting more code and for writing bugsand if you look at the implementation ofQ sort it's pretty ghastly I looked atthe web and out of about 20 versions ofQ sort that I found I would say 18 waseducational frauds they they they weresimplified so they were easier tounderstand and in the process ofsimplification they became totallyunrealistic they could compare integersfor instance but if you look at itswap swaps bites it only works for plainold data if you have a higher level datastructure it doesn't work furthermoreswapping element byte for byte is justnot a good idea it it's slow a lot ofindirect function calls again slowsthings down this this is what I wouldlike to get away from it's far too muchcode it's far too error-prone it's fartoo hard to maintainit's far too hard to maintainefficiently when when Hardware changethis is an unfair example I don't thinkso I didn't make it up it's been used ineducation for decades to teach newprogrammers how to write code and thestyle is not uncommon in production codeI wish it were what it isn't and I'veseen worse much much worse than thiswhat I showed you was quality shippingcode of that style the amateurs do muchworse and the students aim for thislevel of code they feel great when theywrite this code they are convincedthey're writing efficient good code it'scool I add they think it's cool and thenI adults did it maybe their idols did itback when there was nothing else youcould do but still this is the kind ofcode that bright students aims all rightthis is sadand it's not just a C C++ style issue ithappens in other languagesbut I see it more in C and C style C++so that that's the problem I'm talkingabout C++ and using C++ so this is whatI like to to get away from and itmatters I think it's the the bad use thethe inefficient use of C++ is it's oneof the worst problems it the good newsis that makes progress relatively easyyou don't need any new compilers newlanguage features on the other hand badcode breeds more bad code people peopleimitate what there is people think thatit if it's that ugly it must be thatugly for a reason and therefore it mustbe good or this is our house style newthings different things must bedangerous and bad and definitelyinefficient that's the kind of sicklogic you see and many are self-taughtthey take books that are written decadesago they asked other novices thatdoesn't know eitherand well the problem is bad so what Iwant instead I mean it's easy enough tocriticize but let's put something andsaid I want simple interfaces if I wantto sort a container I want to declaresort of a container can't quite do thatyet but we're close I can sir get simplecores you have a container that's avector of strings you can say sort ofthe vector of strings we can do that nowthat's all right and we can get oncompromised performance out of that thatsalt runs faster than Q sort by a largemargin I'll get back to that but thereason it beats it well to sort by largemargin is it doesn't throw away all thatnice useful information we just gave itand we would like to write code withoutbreaking the static type system we cando that just stay out of the darkcorners where you make type errors andresource leaks well you don't have toleak resourcesthat that's such an old-style code and Idon't mean Prague in a garbage collectorI just mean don't sleek it's not I'm notsaying it's trivial in all areas butmajor progress has been done so I'mgoing to talk about type rich programsthese sort of intermediate slides withwith the big pictures just to give me achance to breathe and think about comingnext and notice you that there's a breakin the performance so I'm going to talkabout hyperedge programming and when Italk about interfaces I mean interfacesthat are not just strongly typed in thetraditional sense but actually typedwith meaningful types so if you look atthe first set of examples of theirincreased speed with a double doublewhat what's that mean is it meters persecond is it miles per hour what doesthat double mean I want to draw anobjectwhat's an object there's no informationin an object ok can I can I draw anobject that that happens to be a FouriertransformI don't know and and there's an exampleof a rectangle to give the rectanglefour integers what does that mean one ofthose four integers means are the twopoints are they a point and Hytner widthwhat are they I mean these kind ofexamples are pervasive in software andthey're just their interfaces they'rechecked but they don't check very muchthere they are basically free ofsemantics so what I'm recommending is astyle where if you increase the speedgive it a speed which is a type that sothat you can't throw in any oldmeaningless floating-point number if wewant to draw shapes let's say we drawshapes not not objects that's sort of anextremely early general thing we're sortof anything can be an object so I'vesaid nothing and for the rectangleI can say what I mean a point and pointor a point and a box an area outline andnotice that the minute you start sayingwhat you really mean you can havedifferent versions as I have forrectangle here so one of the mosteffective ways of of controlling youryou're thinking controlling your code sothat you don't make mistakes is units weall talked about units in in high schoolhere speed is something that's metersper second it's not meters per secondsquares it's not an on a no unit thingdivided by seconds and an accelerationis speed divided by some seconds peoplehave again and again designed systemsfor writing such code to save them tofrom errors to what they learnt in inhigh school but they are not used in inin the real world in particular that'sthe mass climate orbiter it tried toland on Mars a few years agounfortunately it was about 150 miles offcourse at the time and that was the endof that that's a five hundred and somesixty million dollar gadget it's theequivalent of the lifetimes work of twohundred good engineers down the drain onthis case down to Mars because somebodydidn't calculate with units Imperialmeasures here MKS system SI system herea constant move from the one world tothe otherthrough one of those double interfacesand it changed its meaning by a factorof four and a half in that process andand this one went down so why are peoplenot doing this because languages don'tsupport them except for specializedlanguagesand you can't use specialized languagesall the time because you have to inventa new one for each special applicationyou could do what is on that slide atruntime but at one time you have tostore the values and then you have tocompute the correct answer when you dothat you've doubled your memoryrequirements you've doubled your computerequirements and on on spaceships atleast there's two things you have aconstant shortage of and one is memoryand the other one is this is computespeed so the engineers decided likeeverybody else in their situation thatthey'll just keep track of the units andget it right and in this particular casefortunately a rare case they blew it solet's see what we can do in in C++ wecan actually write with units if we goand define units and then use them sothe basic idea of a lot of C++ is thatyou build your own types and then youuse them as opposed to trying to writein a basically low-level language sohere I'm saying that first I want someunits and units are in the MKS systemthe subset of the SI system that canmeters kilograms and second and a valueis something that has a magnitude whichis that double we're seeing before andit has a unit so you can be two metersper second or you can be three kilogramsor something like that and I can definespeed as the unit that is meters dividedby seconds to work and acceleration isis meter divided by second squared soyou can write that and then this codebecomes standard C++ so the code as whathave been the textbook kind of versionwould work now the interesting thing isthat we've been ablewhat I just showed you here on the slidefor about 10 yearsa bit longer than that and librarieslike that has been built particularlythere was a nice one from Fermilab thatcould do this and people had to write itand they had to write a notation likeyou saw there and that was a downfall ofthose libraries the physicists theengineers just refused to write codelike thatthe very explicit version is unreadableyou start using short hands and verysoon you get something that is fairlyobscure and basically people didn't doit so one of the things that we can dothese days at least when the compilersshipped with the feature of user-definedliterals in all the compilers is you candefine your own literals so you have asecond you have a second squared and youcan define operators these operatorliterals that says if you if you give along double sub suffix by an S you makeit into a second you make a long doublesuffix with s - it's it's it's secondsquared and now that code like we tookout of the textbook will actually workand this is essentially free it's acompile-time computation it's a simplecompile time computation so it's notgoing to slow you downthere's no no no fancy strike early inthat code and there's no runtimeoverhead it's the same number of bytesit's the same number of cycles that isdisplayed so we can now afford to havethe checking that we knew we should havedone but for the last two decades didn'tdo and sometimes lost big so this is anexample of what I call the tight bridgeprogramming and I'm recommending thatinterfaces should be strongly typed butnot just with random type what willmeaningful types assign types that havethat that has a meaning has a semanticsnot just doubles and integers or objectsthey don't tell you anything the nextthing that'spart of what I think is is good styleand I recommend is be careful dealingwith resources and errors in aconsidered and systematic way so here'san example of sort of fairly standardcode it it looks alright I open a fileand I close it the problem is its naiveand this is a fairly common source ofbugs that use F there if it's only oneline is probably fine but in a realprogram the code that's actually done touse F may be a couple of pages and itmay be split or several functions andit's fairly easy to forget to get to theif plot so that F close after all theremight be a return statement in therethere might be a C long jump or theremight be an exception and you never getto the bottom the closes F so everybodythat's looked at this decided oh dear wehave to do something about it and wedeal something do something to deal withexceptions now if you report errors asexceptions you need exception handlingand so people write something like thisthere's a try block there and we catchall exceptions close the file andeverything is fine you can invent abetter syntax for this if you like butit doesn't change better since I couldnot change the fact that when you areacquiring a resource you have toremember that you're acquiring aresource and then catch the exceptionsthat might be there and release it okaythis is very ugly code there's a lot ofit and it's more complicated than what Istarted out with I mean why would Iexpect code when this was error-pronewhy would I expect this to become havefewer errors bugs correlate withcomplication and with size of quote sowe would expect there to be more bugsthey're in real life cook so that's notgood and the reason for the wholeproblem seems to be that I keep talkingabout a resource right do you see anyresources there I see to function causethey happen to be well known function soI know that if I do an F open I shoulddo an F close the manual says so thecompiler doesn't read manuals a lot ofprogrammers don't read manuals and it ina lot of resources it's not one we knowas a maintenance program or such so Ithink the idea is to represent yourresources directly and get a rid of thatmess there and so here's the example wewrite a file handle a file handle is anobject that owns a file handler and youcreate one of these file handle objectsby giving it the name of the file andthe the open mode and it opens the fileif it can't throws an exception and thedestructor for that file handle releasesthe resource and now I can simplify myuser code saying well here's the filehandle and then use it and at the endwhatever we enter exit that block withan exception or with a return thestructures coordinate releases theresource in this case closes the filethat's the resource acquisition isinitialization idiom which is probablythe best example of why I shouldn't bein marketing anyway I didn't have a goodday when I named that and naming is notone of my strong Forge so so for allresources we can use ra íí- to to keeptrack of the symbol uses ofof resources which are by far the mostcommon examples so in the standardlibrary we do it for Strings we do itfor vector we use it for maps whichkeeps track of the the data the elementsthat is there located there we do it forlocks we use it for fire streams we useit for thread handles it's it's deepdeep into our standard libraryarchitecture and so here as an examplethere's a mutex and there's some shareddata a very simplified example and thenI grab the lock manipulate the sharedata and when I get out of that the lockis released one would think that youcouldn't mess up with a simple lockunlock but again if you know that thelock over and a lock operation takes alook and have to remember to get it outagain you'll probably get it right butfor in general people can forget so wewould try to have anything thatconsidered resource owned by some kindof local handle but let's see whatpeople actually write so here is sort ofa classic piece of code you take andmake a new gadget and you assign it towell it should have been a gadgetpointer and at the end we delete it evenif there wasn't a typo on this slide itwould leak like mad fortunately to thecompiler will stop that one but thecompiler doesn't work on my slides soimagine that P is a gadget pointer sothere are several ways of leaking likereturning or throwing an exception so wenever gets to the delete P it's exactlyequivalent to the file open file closeexample and it's not actually uncommonpeople use a lot of pointers and thepointers are not safe against exceptionsso it leaks and people start screamingfor garbage collectors my answer usuallyand don't leak let's see we could use aSharePoint Oh takes care of this problema SharePoint er will the destructor forthe shared pointer will release theresource if you're you're leaving itsscope and if it's shared by others thelast structure will destroy it but lookat the sample I saw if the leak problembut why am I using a SharePoint I'm notsharing anything with anybody this isjust just my my local use of this so wecan use what's called a unique pointerwhich simply keeps track of a pointer ina scope so that's what you use if youjust want exception safety and want tohave a very efficient way of making surethat that your pointers don't mess upyour code so this works but why have anykind of pointer I mean I'm not passingaround anything at all so really weshould go back to just grabbing theresource using it and let the destructorrelease it and again the resulting codeis much simpler it's also more efficientso elegance and short code andperformance tend to go hand in hand sobasically the lesson I draw for this isthat we should prefer classes where theresource management is part of theirfundamental semantics I mean vectorkeeps track of elements that's what theytest do threads keep track of somethings in the runtime environment that'stheir job they know about stacks andvarious mappings and share tooling andthings like that I don't have to do itit's their job and so their resourcemanagement is embedded in the semanticsof those types so that's fine and if youabsolutely have to pass pointers aroundand such you can pass unique pointersaround when you're just transferringsomething and you can use a share pointif you haveshared ownership my opinion is that youdon't share as often less you think youdo and so you don't need it as often butit's certainly better then then thenjust trying to keep track of it yourselfthe reason I'm little bit worried aboutunique point and SharePoint and preferthe objects themselves is that if youhave two pointers to something there hasto be a protocol of who uses that objectwhen and so somebody else is any pointas a potential race condition who readswho writes not even with concurrency butjust in a program when do you do theupdates there is a problem that tends toget people to use pointers or sharedpointers and things like that and that'sbasically how to get a lot of data outof function I mean take take a simpleexample I want to add to two matricesthe way you usually do that is you takein two references that's a veryefficient way of getting to the elementsof a matrix the problem is how do youget the resulting matrix out again youhave to make a new matrix and then getit out one way you could think about ispass out a pointer but then you have towrite code like that and that's just toougly to to even think about and anywayif you did that there'll be a newinsight the function and who does thedelete now you're getting yourself amemory management problem and sooner orlater you will start inventing usecounted pointers or garbage collectionor something like that and that's not agood solution let's try something elsewe can return a reference at least itlooks right but again the question iswho deletes that object and people lookat this and it says what delete I don'tsee any pointers there's no pointershere so there is a memory managementproblem that's hidden that's even worseso there'sa lot of things that we can do this wecan just pass in the target that we feelso the plus should take the two matricesand fill that the third matrix of yourparty is a targetbut now the normal notation of restequals a plus B will no longer workinstead we have to say operator plus ofa comma B and dress and allocate restwe're regressing into a simple code asimple code can be very efficient it canbe very logical but you know there's alot of writing to be done there's a lotof optimizations that can't be doneautomatically because well theprogrammer is supposed to do it himselfso let's see what we can do I want toget the data cheaply out of the functionI want to write code like that that Itake things in by reference and I returna matrix I after all what does matrixplus - it creates a new matrix and so Ican write code like that that's what Iwant now the problem is that if say thiswas 10,000 by 10,000 matrix copying thematrix out is is rather expensive youyou don't want to do that and peoplestart thinking how can we get it outwithout actually copying and that's acouple of dirty tricks you can you canhave a result stack of matrices and whenit becomes unmanageable but finallyyou've just realized I don't want tocopy any matrix I just want to move itout I mean we know this little babiescan do this say there's a matrix it'shere I want it over here what do I domake a copy and destroy the original ordo this I mean babies can do it so fineI don't make a copy of this one give thecopy over there and destroy the originalthere's something fundamentally wrongthat's the way computers tend to workright you make a copy of an integermeans that's what machines do but it'snot what we want to do with matrices whowant to beto move objects if it's a real object ifit's a resource we want to be able tomove it so this is the way we do thishere is the matrix ad and I make localmatrix inside it I fill it with all theright information and I return it what Iwant is instead of that to be a copyI want it to be the move that is ourpoints or some elements and after thereturn statement rest she points tothose elements and the internal R shouldpoint to no element so it goes awaycheaply that's what a move is rightsimple as that and in C++ oxs directsupport for that way of writing codeI've been writing code like that for adecade but it's spit tricky now we havedirect support that funny-lookingconstructor their matrix ref ref is amove constructor it does moves and theway you implement a move is to copy therepresentation that's what computers cando and then you cero out set theoriginal or representation to be be anempty matrix and so you do that thatmove trick and again this stuff is allover the standard library now vectorscan move lists can move singly linkedlists can move maps on orders mappedhash tables all these good things wehave now they are basically resourcehandles they are treated as resourcesthey can move and so if you have codethat copies things like that that willnow run faster which is good but alsoyou can now write code that returnsthings like big vectors cheaply youdon't have to pretend you're an assemblycoder and fill a target you don't haveto mess with with memory management suchthings this is cheap this is dirt cheapif the matrix representation is singlepointer to the elements as it often isthe cost of that move is toointeger assignments or doubleassignments that's cheap for passingimmediately limitsokay so that was roughly what I wasgoing to say here the style lessons hereare that naked pointers are dangerouspointers an array are splendid fordealing with hardware at the lowestlevel they are not what we want to keepin our productive maintainable code sowe keep naked naked pointers insidefunctions and classes that thatimplements something that require themand keep the array out of the interfacesuse containers instead we know how topipe pass them both in and out so wedon't need to use pointers and arraysfor that pointers are and arraysimplementation level artifacts we try tokeep them out of it you don't usepointers or functions to representownership and one dead giveaway for forpeople writing old-style code is thecode little literate with new anddeletes I mean it's like code letterswith a lock and free malloc and free Imean if you write that level of codewhat do you expect leaks I mean if youwrite C style code you expect C styleerrors so let's try and do somethingbetter the return objects by value as Isaid because we can use with you we canuse by value now one thing that'sextremely important if you wantperformance out of fairly high levelcode is to have reasonable datastructures and algorithms here I'd liketo to show an example that first wasshown to me by John Bentley ofalgorithms fame make a sequence ofrandom integers keeping them in orderand so you you you you're given thenumbers 5 1 4 2 and it builds up thesequence as you see it there and thenyou remove them again by giving a set ofpositions of which one you you take outand so theexercises for which n is a better to usea linked list on a vector on an arrayand everybody gets this one wronghere's huh my graph has disappearedthis is embarrassing okay okay soimagine this to be a graphthe imaginary low line down by thebottom showing efficient usage littletime is the victim the one that islooking like an exponential trying to gothrough the ceiling is the list andtrying to figure out why the the list isso inefficient I thought maybe it's allthis allocation that's been done for thenodes so the little middle green linewhich you can't see either is is thelists with allocation taken out prettyhopefully pre-allocated the point hereis that the vector is always better thanthe list and the list gets worse thefurther you go out and this is for thecase where you're doing a lot ofinsertions and deletions which when Iwas taught about data structures waswhat you use lists for because they arereally good and inserting deleting ifyou want to insert in the middle ofthousand a hundred thousand integervector you have to shove on averagefifty thousand elements one position ifyou take one out you have to shove themabove the half the way back now this iscompletely irrelevant what matters isthe linear search to get to theinsertion point of course you have to gothrough half of the list to find on theaverage insertion point for a list andin an attempt to fairness I also did thesame vectors instead of using a binarysearch but anyway so the linear searchdominate completely and linear searchfor vectors it's not actually is such agood idea first of all and vector is fora list a list is much bigger for a givendata structure than a vector because youdon't have to just store the element theinteger you have to store the twopointers forward and backwards you haveto use a doubly linked list if you areinserting otherwise you have the extraproblems and that completely dominatesso that thethe graph here this graph here that youcan't see shows you that there's not aminor disadvantage here we're talkingabout things being fifty a hundred timesslower with with a linked list and thetraversal dominates so compactnessmatters vectors are more compact andlists and predictable usage patternsmatters enormously with the vector youhave to show a lot of elements or butcaches are really really good at that sosurprisingly vectors are random accessconstructs but you can stream them listsdon't have random access but when youtraverse a lists you keep doing randomaccess there's a note here and goes tothat node in memory so you actuallyrandom access in your memory and you'remaximizing your cache misses which isexactly the opposite of what you want sothe the lesson I'm trying to say here isstay compact safe predictable and youhave three orders of magnitude ofperformance to to to deal with here togain by being compact now people havelooked at at that graph you didn't seeand said well I don't use lists of ahundred thousand and two hundredthousand elements and there's two kindsof people as a sort of the the Googleand Amazon people that says because Idon't have such little data structuresand then there's the students that thatI think a thousand elements is a longlist and so I don't use that manyhundred thousand lists but using a fewhundred thousand element lists isexactly the same as a matter of fact youcan get the performance effects out ofindividual data structures so here's avery simple one up there Victor ofpoints with with four points the waythat will be laid down in memory if youuse the standard is up there labeled C++you have a little handle a lot on C++ isthese little handles that tells you howto use things and then the resourcemanage which happens to be a compactdata structure with eight integersbecause it was integer point and that'sfine it's compact you have a singledereference to get to it you have alittle bit of memory overhead becausevector keeps its elements on theFreeStore so you get the the extra ofword or two as a free store header butthat's the way it looks it's fairlycompact if you don't want to put it onthe free store don't but usually thatyou can afford itnow I'm told very often that I have towrite in a truly object-oriented styleunless languages that ensure you do thatand in truly object-oriented style ofcourse an object is referred to by areference so we have a reference to theobjectthere's the object down there and thenext line with the for the count in itand this happens to be a container of ofuser-defined objects so again you have acontainer of references and there youhave the objects so that turns thelinear compact data structure into alink structure we just saw on thisinvisible slide what link structures doto your performance and and here you geta rough doubling of the size of the datastructure and whenever you want toaccess an element instead of getting oneindirection you get one two three indirections and in directions again thesethings that come that modern computersdon't like very muchpointers are poison to most are upchoices so you can get the this againso I recommend compact data structuresand fairly regular access to them that'sthat's the winnersimilarly we have to simplify ourcontrol structures let's let's let'slook at an example that some friends ofmine force ended me with it was a verysimple example a lot of operatingsystems a lot of gurus have theoperation for taking something out oflet's think of a think of it as say saya menu and move it over therethat's the operation and a guy wrote apiece of code that solved this problemfor a particular GUI system and therewas 25 lines of code I mean I justexplained to what we were doing in moreroughly one sentence take an element andput it over there and there was one loopthose three tests are no 14 functioncourse now the author looked at this andsays well you know this is rather messycode and even though it was aprofessional experienced programmer ormaybe because he was professionalexperience programmer he worried that itmight not be correct I mean 25 lines ofcode with loops and conditions how willyou knowand is it maintainable and he decided itprobably wasn't maintainablebecause he had to think about whether itwas right or not the next guy comingalong would probably not think as hardand make it make a change and is itusable elsewhere no it was just full ofthings I was very specific to thatparticular GUI and the author being areal professional requested a reviewprofessionalism is very important and issomething we've got far too little of inour field so the story I got was fromthe guy that came in to do the review asrequested and they looked at it for awhile actuallyalmost the whole day and found that whatthe code really did was find the answerpoint and you can write that as a findout where them from the standard libraryand then it has to insert it you have tomove it into there's point and that's arotation so you can rotate from beforethe insertion point or you can rotatefrom behind the insertion point so youget that code if you if you know thestandard library this this looks fairlyreasonable you just don't know thestandard library there's a standarddefinition you can go in and find it andread it it's it's well-defined asopposed to the random code and it is analgorithm the only problem is so itsmaintainable but it's very specialpurpose Victor wasn't standard vectorit was the vector of GUI objects and thecoordinate was was very specific to thatlibrary so they looked at it added a bitof again sake we can generalize this toany GUI and they came up with thatversion is called gather which isshipped by in several different versionsby actually several libraries these daysand basically what it says it'll make apair of the elements from before theinsertion point the after the certainpoint and I must admit I can neverremember what stable partition does formore than a few minutes because that'snot my field of specialization of myarea but people will write this kind ofcode of course get used to it and youcan look it up so it is shorter it issimpler it is faster it is usable andfar many areas and there's no loops andtests once you have convinced yourselfthat this is correct it'll stay correctit takes a little bit reading of manualas opposed to hacking code and there aresome people who strongly prefer to hackthe code but this is the kind of thingso I'm trying to recommend that to alarge extent move away from from randomcode code with loops and tests and tryand think in terms of algorithms and howto generalize things this this would bea culture change because a lot ofprogrammers are a trained athacking random code and was very proudof of complicated control structures butI when I see complicated code I getworried and we know where bugs hidethey-they-they one of the obvious placesis hide isn't in big chunks of codeespecially big chunk of code with a lotof loops and and and tests in it theother weights can hide is is an enormousnumber of function codes to get workdone but anyway we try to simplify codeand try to make it better document it asas as formal algorithms and so finallyto get any of this we have to stay highlevel whenever we can and I'm trying toencourage you to believe that you canstay high level more often than youthink high level there is 14,000 feetit's taken from the top of Mauna Kealooking at Mountain Alohaanyway so people are really convincedthat low-level code is efficient theyactually think that low-levelmessy code is more or less definition ofefficiency and a lot of people don'teven bother measuring because they knowthat low level is more efficient thanhigh level this is complete rubbish butpeople think that way so I've been doingexperiments with levels here here wehave two examples a fairly classicalloop that that sums the elements of avector and then the same using the foilrhythm with a lambda that representedthe second is definitely higher levelfor each from the beginning to the endincrement the sum and I find on severalsystems with several compilers basicallythe code generated for those two casesidentical it's you you you you choosebased on the style you like not based onefficiency concerns now for aand this Psalm is a very simplealgorithm but if I was saying what werefor breadth-first search versusdepth-first search the one can bewritten like a symbol loop and the otherone it really is an algorithm and youhave to stick to the algorithms so tryto do the thing that's of courseaccumulators you just call the standardaccumulate anyway but this is just anexample of code where you might thinkthere was some overhead somewhere and Icouldn't find itwith several compilers and severalsystems here's another example that havebeen doing for ages this is Q sortversus sort sort of course is planarmore general and easier to use than Qsort and for that reason widely believesthat has to be slower it isn't forexamples like sorting any reasonablesized container say you know I thinkthis numbers for ten thousand or hundredthousand elements thought was two and ahalf times faster than queues or for thereasons there are type safety used bythe optimizers better in lining youdon't have to have those indirectfunction calls it's it's just cleanerand therefore runs much faster so if youdo a performance of C style andobject-oriented code it's it's roughlyequal but once you start using type safehigh-level code you start gettingperformance advantages to sort being thestandard example anything in thestandard library you can get performancejust by using it so low level does notmean efficient and don't lower yourabstraction level before you absolutelyknow you have to so this is a differentversion of the old saying that don'tdon't do premature optimization or JohnBentley's version don't optimize yetand you repeat that a few times but stayhigh level as you can now you may havenoticed that I have not said anythingabout inheritance this is partly becauseI assume you know everything about itand I had to cut out something secondlyinheritance is wonderful if yourapplication domain really ishierarchical but people seem to beforcing all kinds of strange things intohierarchies and having hierarchies justfor the sake of hierarchies I mean youhave a lot of programmers who think thatobject-oriented means never use afree-standing function and that's justsilly because most of the things we dodoes not fit into hierarchies and soanother reason for not emphasizinginheritance here apart from I assume youknow it is that it has been seriously orused for a last couple of decadesobject-oriented programming usinghierarchies is a wonderful way ofwriting some code it is an awful way ofwriting all code and by the way I couldsay that of Japan just about any sort ofgeneral solution it doesn't fit in inall possible cases we have to find ablend of things based on high-level typesafety and things like that so I have tosay a little bit about concurrencythere's many kinds of concurrency C++ 11supports types typesafe type richprogramming at the threads and lockslevel I would like to be at a higherlevel of abstraction and threads andlocks but we have to do systems levelwork we have to interact with theoperating system whatever it is andthat's the level where the operatingsystem is so this you can think of as afoundation on top of which you can buildall the things and if that's too highlevel for you you can also to do lockfree programming on that but anyway I'mnot going into details here that'sthat's hence perms next talk that's onthe memorymodel and a little bit about concurrencyI believe but here is how you can writethreading in in modern C++ I have afunction I have a function object andthen down there make t1 which is thethread that will execute F on BEC on onVic 1 on a separate thread and we haveteach you that will execute the functionobject initialized with a two on aseparate thread and of course we thenhave to join the the threads later andwe can start using the results verysimple example the point here is notthat this is new we've been writing codelike this for 2 3 decadesthe point is that we don't have to doany missing with void star stars oranything else like thiswe have gotten type safety and we'vegotten stand-up things this will work onall the systems that's nice and I'd liketo just show that you can go a littlebit further by passing results so here Imake a thread T 1 that calculate F 1 andputs the result in the rest and we do itagain and we join so we get the thestandard functional style and we uselambda is to pass them along and finallywe can simplify this even further by nottalking about threads and blocks andjoins we're talking about tasks to beexecuted asynchronously a task is justsome work to be done and here I'm sayingout asynchronously like if to be calledon some big and I would like a synchronyG to be called on some thick you mayworry that I'm here starting to anasynchronous threads on the same vectorbut it's Const and constantsimmutability is great in concurrentsystems so I'm not actually shootingmyself in the footnow async which is a standard functionwill return something that you can getthe results from that something iscalled the future and basicallykind of a handle to go to the threadsresult mechanism and so you get rest oneget rest to notice there to find that'salso so I didn't have to be specificabout it also is the mechanism in C++ 11that says pick your type from yourinitializer so I didn't have to talkabout futures here at all I'm justsaying a synchro still run F asynchronous to run G get the resultsfrom the first computation get theresult from the second you knowcomputation and we're homeso that's much more elegant and I oftenfaster because you would not really wantto spawn more than one thread here theother one could run on the original sobasically here's a summary practicedhyperedge programming has said integrateerror handling and resource handlingfocus on compact data structures andprefer algorithm so random code andbuild and use libraries a lot of whatc++ is for is writing libraries youdon't have to write on the raw languagejust like we use the type system so wedon't have to write on the raw machineand so the first stage in most problemsis to build a library that supports yourapplication domain and you can do it thetype safe concurrency and you have thestandard library to get started withI wish we had many more area specificlibraries that will stand up but wedon't at least there are some of themand so here I can I can take takequestions I I think we have a system fortaking questions and I think we needmicrophones so that everybody can canhear the questions yes this is somebodyhanding microphones around any questionsthat's a question there that kind of tome we lose a little bit of abstractionwhen you start thinking like thatbecause when you were picking the listversus the vector you are reallythinking about how that was implementedeven down to the issue of memory cachingand stuff like that so that to me kindof betrays how should I say it betraysmaybe we want to design list and vectorin a different way because when I thinkof list I'm thinking of something that Iattach to a lot or I walk through or Idon't access randomly and when I thinkvector I'm thinking of something that Ineed access randomly and I shouldn'treally have to think about how itsimplemented I should be able to justpresume that if I pick a list it's thebest implementation of that fundamentalconcept right I think I understand whatyou're saying and I think you happen tobe wrong what we should say is that weneed a sequence of elements and thedefault sequence of elements in C++ is avector now because that is compact andefficient implementation mapping tohardware matters now if you want tooptimize for insertion and deletion yousay well I don't want the defaultversion author of a sequenceI want the specialized one which is alist and if you do that you should knowenough to say I'm accepting some costsand some problems like slow Traverseand more memory usage and I think toomany people are not thinking at allabout these things and writing bloatwareso think about vector as a defaultcontainer of element the defaultsequence and you parked apart from thatonly when you're trying to optimizethink about a linked list as aspecialization as an optimization and wecan't do that up automatically for youpeople have not figured out how to pickdata structures optimally relative totheir use in a program and so we startwith vector we can go to lists sometimeswe go to map a lot we goes to unorderedmaps hash tables a lot of sets but wehave to have a reason for doing so and Idon't think I'm lowering the level ofabstraction at all in so there's aquestion here I think coming out ofTwitter it's a shipping writing movecode for a decade how did you go aboutdoing itokay there is a version that can befound on the web of my favorite solutionI I wrote a course and a book fornovices called the programming practiceand principle principles and practiceusing C++ and on my web pages you canfind a support page for that and for thesupport page from that there are somecode and in particular there is a littlematrix library that shows how to do itthe basic trick is that inside thematrix is a copy constructor and a bitthat says whether you copy a move and soif I say nothing it's a move if I say Xfor transfer the next copy operationturns into a moveit's a hack it works beautifully in thehands of a library implementer and theusers don't have to know about itanything that can be done can can bedone and somebody has done it but todayin C++ 11 you don't have to know thathack and you don't have to do it and themajor compilers ship will move semanticsin them so it is not science fictionthank you having a standard interface toyour collection such as an iterator youcan completely optimize later bydetermining at the declaration what typeof collection whether it's a vector or alist or a hash table or whatnot and youdon't have to change any code that isrelying on the implementation of thatand then second of all I have a questionwhich is to do with C++ 11 I've spent alot of time in my career over the past25 years writing protected data memberswith getters and setters for these withthe presumption that later at some pointI'm gonna change the logic and the rulesabout accessing those I found thatalmost never I've had to use that andI'm now starting to look at going backto having public data members and I'vealso tried to come up with anabstraction to declare a property sothat it has getters and setters thatcould be intercepted by the class thatowns it and I'm just wondering if anythought has been put to that let C++ 11or if that is a problem that we can findan easier way for declaring propertiesin C++ see the the notion of propertieshave crept up again and again I thinkthat protected data is a pretty hopelessidea it got in sounded like a good ideaand it turned out not to be so I prefermy data to either be public or privateI can manage those two things propertieswith getters and setters the two thingsare very common in certain systems I'mnot sure that they should be but theyare and you can write a library thatdoesvery nice property facilities nobody hasmade that standard it was almost donefor the previous standard and there areaa little work has been done in the C++11 timeframe for that I don't know ifthat going to crop up again if it doesit'll take some time so if that reallyis a major part of your programming lifeand it is for some people think abouthow to write a little property librarywhere you have a property type thattakes as a parameter that the real typeand you can then start overwriting thethe getter and setter mechanisms it'sit's it's half a page of code I can'thear you I can't hear youwait for the mic for the parent class toactually intercept the get or set ofthat property and that that propertyabstraction basically got in the way ofmany times because you had to explicitlyuse the get or set to get the actualtype to see the members of the actualobject so our would you recommend thatwe we start to use public data and notworry about that and occasionally changethe code that accesses it when a rulecomes into placei I I can't answer that without knowingthe application domain similarly somepeople have found the get on the the theproperty library is sufficient othersare found insufficient it seems todepend critically on what you are doingand you should talk to people who whoknow your application area better than Idogreat we got a question right back therethere are Joint Strike Fighter C++mm-hmm coding guidelines and as you werepart of developing those and you made acomment that those did not includeexceptions because ii did not have haveenough support to use exceptions inmission-critical applications and thatwas in 2005 so i was just wondering ifthe tools finally caught up and whatwere their problems that exceptions goodreveals so what what he's talking aboutis the guidelines for writing the flightcontroller software for the nextgeneration strike fighterthe f-35 one thing that you have to takeinto account when you write worksoftware for that is that a smallmistake in the code can lose you a veryexpensive plane and possibly a pilot therules were written with a safety officersitting at the end of the table andevery half hour piping up saying but isit safe we're safe is defined assomebody might dieif it isn't now this is sort of slightlyironic when you're talking about afighter plane but so the concern here isabout reliability but also performancein particular it is a hard real-timeproblem so one of the rules for hardreal-time is that every operation has tobe predictable and in C++ actually everyoperation is predictable with theexception of new and delete so you can'tuse those and throw and dynamic cast weactually found this developed a constanttime dynamic cost I don't know if it'sactually used if you know of mypublication record we actually developeda fast dynamic cost but getting back tothe exceptions if you throw an exceptionyou want a guarantee that if hand getsto a handler within a very small numberof milliseconds there's no guarantee inthe language for that to happen there'sno guarantee that a throw from a givenpoint comes from the same source so itmay take different times this is notpredictable therefore it's not useful itwould become useful the minute you had atool that analyzed all the throw pathsso the handlers and gave you back areliable upper bound on the time thatyou took to get there that tool does notexist I don't think the tool could existin general for all code however messy Ithink you would hit the haunting problembut you could say you could build a toolthat gave you that bound or said itcouldn't do it in which case if thebound was was small enough you'd beallowed to use exception handling andyour code would get get cleaner and Ithink more more reliable but we don'thave that tour today eitherI had expected to have it today butother things came in the way I don'tknow if the if the airplane guys have itI just I don't know all right we gotanother question there so you talkedabout you think stack allocation andusing just having objects be declared onthe stack where we're possible andthat's something I try to do as much asI can but it's not always clear that thestack has enough room for large objectsand in those situations if I happen tobe the unlucky stack frame right after abig allocationI sometimes crash and I find that thereis no standard compliant way for me todeal with that do you know of this beinga larger problem and you know if thereis a good way to ensure that you don'trun out of stack space there's nostandard way of reporting that you'verun out of stack that's unfortunate andit's a problem especially in embeddedsystems where there's not as much stackas we might be the used to and the onlything I can recommend is not to putreally large objects on the stack inlarge large numbers because usuallythere's these days quite a lot of sexbase and you saw the objects I wasdealing with by and large for handlesand and the victor 4 million elementstakes up three words on the stack sothat's not not the problem and andthat's that alleviates the problem in inmany many areas so I can't solve thegeneral thing it's it's something to dowith the interaction the hardware andnon non predictable ways for stackoverflow but use smaller objects and getthem as handlesthat's the only thing I can do so so soit seems that your slide aboutinheritance and how it's been overusedand forcing objects into hierarchiesthat aren't necessarily really requiredit seems to me that part of the reasonthat people dothat is to build some sort of interfacefor that code that they can you knowpass objects that would have thosecurricular methods called method callsobviously one solution to that is toinstead use the template system and youknow to get the generosity involved inthat but as soon as you start throwingtemplates at people some people getsicky about that and decide that theywould rather do the object inheritingobject-oriented inheritance is there anyyou have any comments or suggestionsabout how to fix that yeah that slide isreally a talk that's longer than thistalk sometimes I and noticed that ahaving sample op that said I didn't wantan object to be my my base class myinterface because that's a meaninglessinterface but I suggest the shape mightbe a good interface abstraction which isa base class furthermore a widget mightbe another one the the graphics domainis right with the useful abstractionsthat are hierarchical and so that's fineand you can even parameterize themthings like vectors of shape pointers isa good example of where you get amixture of generics and anobject-oriented code I'm not sayingeverything should be templates I justdidn't have it enough time to discusswhere the trade-offs lie not everythingshould be a hierarchy not everythingshould be a template great and aquestion so C++ is becoming like a hugeSwiss Army knife is there some thoughtbeing put into creating sort of like astrict mode which would be like aguideline for compilers to try tosort of forbade some obsolete featuresor features which could be dangerouslike protected or maybe friend maybe goto C++ is large but it's notparticularly larger than then otherlanguages people are using it is reallyhard to team in general what isdangerous or what is inefficient or whatis useless or whatever criteria is myimpression is that everything that's inC++ is essential to some user communityof maybe a couple of hundred thousandpeople and so in take away anythingmajor and you heard a couple of hundredthousand people well basically you can'tbecause they'll just use it they use itanyway and ignore you take awaysomething minor and it's ridiculousbecause you don't have any effect andwhenever you take something mine and youstill break some old code and peoplereally hate you breaking their code forany reason whatsoever because it hurtthem now so he has been a lot ofthoughts about subsetting C++ we wedon't really know how to do it becausesome of the most dangerous features interms of what you can do with them areessential for efficient implementationof the higher level abstractions that wereally want so a subset would have totell the difference between anabstraction we like like vector andfeature we consider dangerous likepointers into arrays and and accept thelatter only behind the interfaces thisis this is quite hard but sort of thefirst attempts to create a safe subsetof C++ must be almost thirty years agolong before those core C++ and we againand again find the people's criteria forwhat they want and don't want verydramatically and it's context inthe distinction between inside andoutside and abstraction has to be partof the answer and nobody has actuallydone something really useful for thatand what you get instead is codingguidelines and actually verifiersagainst coding guidelines that do knowthe difference between the set ofabstractions for a particular domainthat you can do but doing it in generalI don't know how to because for anymajor feature that's a community thatrelies on itokay you got time for one more questionoh you know what there is a panel at theend of the day so your questions in yourhead and yeah we'll get out later yeahthat's that's a way of asking questionsat the end of today and they at the endof tomorrow and that may even be betterbecause I won't be the only oneanswering the questions and you may getmore than one answer anyways askquestion just hang on okay yeah findquestions about move constructors yeahum so first I want to say that they'redefinitely gonna make our code fasterit's fantastic that they're at it but inour experience with programmers is theyreally are lazy and they don't likewriting default constructors copyconstructors operator equals and now I'mseeing with move constructors like a 66%increase in things they're gonna be lazyabout so do you have any suggestions orguidelines on or maybe you should belike I don't I'll are warning and thecompiler can see something on things wecan do to improve that there's a set ofrules for what defaults you get anddon't get and basically if you don'twrite at this doctor you probably don'thave any need for move and copyconstructors you get the defaultsthat'll work they don't have to do itand I really think people should stickto the defaultsrather than trying to write themthemselves there if you if you want thedefault semantics you should not writeyour own constructors move or copy ifyou want special semantics yes youwanted and you do it and there is waysof saying if youto say copy constructor you want thedefault but the move constructor youwant to write yourself you can simplysay the the carbon constructor equals d4you get the default and then write yourown move constructor so there's a set ofdefaults and a way of getting them bydefault and a way of saying you want thedefault that that addresses that problemand I think it'll at least besignificantly improved if not totallyeliminated and thank you thank youyou