Now this is most certainly wrong. Example (search "history research"):
"Research in history involves developing an understanding of the past through the examination and interpretation of evidence. Evidence may exist in the form of texts, physical remains of historic sites, recorded data, pictures, maps, artifacts, and so on."
Taking a very obvious example, you can't do much in the way of historical research without at least trying to establish which things happened before which other things. It's a serious problem in ancient history.
I'm pretty sure modern historians use a lot of mathematical tools: statistics, information science, digital archives, computer imaging, etc. It's very hard to search for this, though, because all of the results concern the history of mathematics. You have to examine the tools and methods used.
Pretty much every field in the social sciences and humanities requires their undergrads to take at least one course in statistics. Sure, these students may complain about it but they need to be trained to not make common statistical errors in their publications. Unfortunately, they still do, which highlights the importance of mathematical education even in these fields.
By seed you mean salt. Salts are stored in plaintext, so they don't increase the entropy of the password. Instead they make it so that each password hashes uniquely so that everyone with the same password gets different hashes. They also mitigate rainbow tables by effectively requiring the attacker to create a rainbow table per target
I'm not sure what you mean by a seed. If you mean a salt, that's no more secret than the hash. It has the effect of requiring you to crack each hash separately, but doesn't make it any harder to crack an individual hash.
P2P scales quadratically, central server scales linearly. You might argue the cost is not borne by the operator but imagine what happens if Zoom was P2P. It would probably bring down the internet.
Moreover P2P would require more upload bandwidth per user from users. P2P just doesn't make sense for standard video conferencing.
Assuming that the server combines all the incoming streams into a single output stream, otherwise it still has to broadcast N-1 streams to each of the N users.
But why would you not combine them? In fact, I can easily see people being lazy and combining just on N sized stream to reduce server overhead especially if there is compression on top.
Except that the instances you listed are not a problem the vast majority of the time. While copyrights and patents are enforceable for any unauthorized use no one would bother until you are rolling in the dough, at which point the laws make perfect sense.
That said, I am not arguing that patent law doesn't need reform, especially patents for software, algorithms and math.
Copyright has become the ultimate tool of selective enforcement, because it's nearly impossible to avoid breaking the law.
It's not a problem the majority of the time, just like laws that allow police to detain you for no reason. It becomes a problem when someone decides to arbitrarily target you. There's countless stories about small businesses being targeted. The latest I know of is against streamers.
Since everybody is breaking the law music industry uses copyright mafia style. If they think they can get money out of you they send some goons to extort protection money from you, with huge leverage against your previous "crimes".
A country with rule of law should not have laws that are impossible to avoid breaking. Such laws are only useful for corruption, a way to enforce "justice" arbitrarily through bribes
I feel like the best way to filter spam is using information content. As it turns out, nobody is handing out $1,000,000 for free no matter how nicely they write the email.
> I don't see modern application of C outside embedded,
C is one of the most popular languages for high-performance code. Most famous example would be the Linux kernel but literally anything that needs the best optimizations possible eventually comes down to C.
Strange thing to say, to be honest. Many of the most performance-critical industries are heavy on C++, not C. Games, finance, high-frequency trading, where every microsecond is worth huge money -- these are typically C++ shops. The strictness and ability to move things to compile time are features not available in C.
Any examples of using compile time features that make a difference, instead of making code harder to maintian and increasing compile times significantly?
You’re talking about really two different issues here. I don’t think anyone would suggest templates can’t get complicated. But they are certainly the doorway into maximum performance over C.
No I don't think I'm talking about something else. And I wouldn't submit to the opinion that performance cricitical code needs to be written in C++.
For example, Games are heavy in C++ where there are a lot of abstractions. Engine code is mostly C style as far as I can tell. Optimized code is code that exploits specific properties of systems, hence it is significantly less abstracted than higher-level code.
The stereotypical std::sort example doesn't count, really. There are many reasons why std::sort vs qsort (which are both generic sorting implementations - i.e. not optimized ones) is not significant in practice, but it's still among the most frequently cited examples - hinting that there might be few real wins (with regards to performance) from all this compile time and templates stuff.
Hm, I’m not sure where to start with your assessment here. Your comments on sorting don’t really seem relevant, I don’t think you’re referring to the same kind of compile time programming we’ve been discussing. There’s a difference between generic code, and code that runs at compile time, even though both might actually use the C++ feature of templates. C++ templates are way to write code that supports generics, but they are also a way to do a style of programming referred to as “meta-programming“, these are really not the same thing at all.
While you can certainly do without generics in the language, as proven by the success of many many languages, perhaps at the inconvenience of some developers, compile time meta programming is an entirely different area that only few languages really support. The high frequency trading firms and the game studios I have partnered with are typically using C++ specifically for those compile time programming features. They make dramatic performance differences in a wide range of highly specialized algorithms.
It wasn't me who said "templates" first. (I don't think of them as a practical way to do compile time computation).
I was mostly looking at constexpr and whatever similar things have appeared in C++ lately. And I wanted to know about actual applications of them in the wild that make a difference.
Because, yeah I can precompute a 100K hash table or whatever at compile time, but I can also just do it at program startup (would anyone notice?) which is by far the simplest thing to do. Or I could just generate the data in a separate build step which is probably more hassle compared to constexpr but also probably friendlier in terms of build times in practice.
In embedded systems, often times not all memory is equal. Pre-computing a lookup table at runtime may not be practical due to the limited amount of RAM vs. flash memory. A constexpr or template meta program is, as you touched on, a nice way to do calculations at compile time in the existing language without having to add an explicit autocoding build step. An explicit build step eventually makes sense for sufficiently complex algorithms, but it can be a lot of build system maintenance overhead for small to medium complexity stuff. Implementing esoteric code using obscure syntax may be bad for readability, but keeping it "in the language" has a benefit of limiting the amount of project specific knowledge required to understand it.
Thanks! Thinking a bit more about it, I could imagine some performance impacts even on full featured CPUs. With virtual memory and the OS paging stuff in and out of physical RAM under memory pressure, read only data can be swapped out faster than writable data. The former, being immutable, can just be forgotten and then reloaded from disk when it's accessed again, while the latter has to be written to a swap file first, and writes are typically orders of magnitude slower than reads. Doesn't matter as long as you have plenty of RAM though.
Another benefit on that line is that read-only memory can be shared between processes.
I'm not sure that this is hugely relevant these days for small stuff, though. Like < 1MB... how many instances of the same program do we have running simultaneously, anyway?.
Matrix compile time templates like Eigen result in vastly faster code than doing it in C, since many operations can be compile time simplified. C has no way do do this at compile time.
This is just the tip of the iceberg on using templates and classes to make faster, cleaner code.
In C, you could provide a bunch of functions that chain together the permutations of operations that can be optimized. I.e. TransverseMultiplyMatrixInverseDotProduct() or whatever actually makes sense. Since you can't overload operators, folk would have to read through the available functions to find what they need anyway. It wouldn't be pretty, but it would be functional and probably compile down to similar machine code.
No, you cannot, not without putting an incredible amount of work on the programmers plate.
Consider the simple problem of multiplying together a sequence of N matrices of possibly different sizes with the least amount of work. The order you multiply in is determined via some optimization technique. You can try to have a different C function for each N, but eventually you will have some N for which your lib doesn't have the call. Or maybe you'll try to pack pointers into an array and pass that, which is now slower and more memory costly. In any case the order must be solved at runtime.
Templates allow, at compile time for known size matrices, the order to be determined. This cannot be done in generality with C since you cannot in C do it.
And, if the matrices were constexpr, this can be computed at compile time.
So the template method, giving you Turing complete operations, can do things that you cannot do in C.
This is just a simple example, the tip of the iceberg.
I'm not disputing that you can make prettier, more scalable APIs in C++ than in C. My point is that it's not completely hopeless in C either, though. In practice, the user of a matrix math library needs to understand the operations they're doing, and especially so if they actually care about performance. In the example you gave of a string of matrix multiplications, matrix multiplication isn't commutative, so the order is the order that the programmer wrote them in. The compiler is still free to reorder and coalesce redundant calculations with sufficient inlining. Also, N is small for 99% of use cases where performance matters, and when N is large, falling back to a slower "runtime" implementation is perfectly reasonable because the runtime overhead is insignificant compared to the overall cost of the operation; eigen itself does that internally. A blanket claim that pointers are "slower" and memory costly also seems a bit overly simplified. They are usually worse than passing by value for small data sizes, but for larger data sizes, some sort of reference passing somewhere will be faster than doing unnecessary memory copies. For sufficiently large data sizes, a straight forward hand written "runtime" algorithm implementation may even happen to be faster than a compiler generated specialized equivalent depending on the hardware's memory model.
Eigen is a great library and very convenient to use. It's great to be able to write straight forward chains of matrix operations and trust that the resulting program will be reasonably fast. There's no need to be dogmatic about C vs. C++, though. They're both higher level languages targeting the same underlying hardware. Templates enable library developers to make simple APIs at the expense of more complicated library implementations. In C, it's often necessary to compromise on the simplicity of the API to achieve the same performance, but it also generally means that the library implementations are simpler. The overall quality of the resulting binary can be about the same, and is almost certainly within the same ballpark performance wise. As an embedded engineer, I often need APIs that are compatible with C whether or not the implementation is C++, and I value simple library implementations over complex ones; the libraries and my use cases are often obscure enough that they are buggy, and so the more readable the library is, the easier it is for me to debug them.
As a recent real world example, a coworker, who is a wizard that knows way more than I do about signal processing, implemented some matrix heavy algorithms in a high level language that supports just-in-time compilation down to parallelized CPU and even GPU machine code. It worked great on an x86-64 workstation, but on production hardware, we struggled to get the code to run fast enough; it would peg all the CPUs at 100%. The many layers of libraries and JIT compiliation made the system very hard to debug even after a couple weeks of trying. I suggested re-implementing the algorithm in C++ using whatever matrix library was most convenient, and a few days later the system was running perfectly and averaging 14% of one CPU. The algorithm went from maybe 50 lines of very readable code to 250 lines of relatively ugly code, but we understood what it was doing way better. I believe he used Eigen in the C++ implementation, but whether or not the matrix library was optimized at all, C, C++, or rust, it still would have sipped around 14% of one CPU. My point is that, when performance matters, you need to understand what the software and hardware is doing, and so there's value in simplicity and pragmatism.
Of course you could also just write assembly if you wanted the exact most optimized machine code.
It’s also hard to generalize these things in the form of those kinds of macros. Whereas with something like Eigen, just write your code like normal, you don’t have to worry about the special cases, and the compiler rewrites it for you. That’s one of the nice benefits, one of many, of metaprogramming.
But you were talking about generics, which implied you were incorrectly conflating that with template meta-programming, since generics are done with templates.
> I don't think of them as a practical way to do compile time computation
Templates are however the most flexible, advance technique for C++ compile-time programming, though the C++ standard is evolving to bring more and more of those features into the language without using templates, i.e. "constexpr if".
My recommendation would be to avoid speculating on what the benefits are of a language or its features if you clearly don't have serious experience using them. It's fine if C++ is not for you.
> But you were talking about generics, which implied you were incorrectly conflating that with template meta-programming, since generics are done with templates.
I'm aware this is a pointless discussion, but please double check your claims are right if we are in "check mode". I did not say "generics".
>> Any examples of using compile time features that make a difference, instead of making code harder to maintian and increasing compile times significantly?
> My recommendation would be to avoid speculating on what the benefits are of a language or its features if you clearly don't have serious experience using them. It's fine if C++ is not for you.
Maybe you shouldn't make such statements if you don't know about my experience. I am speaking from experience, and exchanging subjective experiences isn't worthwhile most of the time, but sometimes (if people don't go down to personal attacks) there is a new viewpoint to find.
FYI if you’re talking about compile-time programming and you use the word “generic“ (as you did, in the context of a generic sort, which actually does have a meaning related to generics, as the routine works on containers of any type), note that this is a well-established term, which could be confusing if you actually are referring to something else.
I worked under ULL trading platform, they used C++ with heavy templates, without RTTI and exceptions. If things can be done in compile time, it should be done - which is the purpose of templates. Neither our nor neighbor teams (near 800 developers) used plain C. As Bjarne Stroustrup told, there's no place between C++ and machine code for "more low-level language", everything which could be done in C, also could be done in C++ with the same efficiency.
Good tendency however, now both C and C++ developers started experimenting with Rust, probably creating unified community and platform.
You have to use a subset of c++. Nobody agrees what that subset is. You have to be super-vigilant in code review. Features invite their use.
It's the lack of features in c that make it attractive. No magic. You want an object or a virtual, code it up if you really mean it is not just a keyword. You take responsibility for all the code running.
Every time I've cut a tonne of latency from a trading engine it's that. People relying on library and compiler without taking responsibility for it. STL is great. Except if you're actually performance critical, when you can beat it easily by solving your problem, not someone else's solution for everyone's possible similar problem.
With the caveat that I am NOT suggesting this justifies choosing C over C++, I just wanted to mention this talk about how "zero-cost abstraction" is an idealism, not necessarily a reality:
https://www.youtube.com/watch?v=rHIkrotSwcc&t=17m30s
That said, I tried to reproduce something similar, and it seems the issue only occurs in my example due to external linkage (adding 'static' fixes it)... but I can't claim this will always resolve the issue:
https://gcc.godbolt.org/z/1vbqo3
The "zero-cost abstraction" concept is that you don't pay for something you don't use. If you use a smart pointer, you pay for it. If you don't use a smart pointer, it's zero cost.
Any argument that smart pointers are not zero cost to the language because they have a cost when you use them is a classic straw man argument.
I think there might be a terminology mixup here but he probably meant zero-overhead abstraction. And in any case, the point he's making is not a strawman argument.
He talked about more than just unique_ptr if you actually watch the video, but I don't even get why you're dying to have such a pointless argument here. When you see free lunches offered somewhere, do you start arguing with people how everyone is wrong and there is such a thing as a free lunch too? Is it so hard to actually take the point and just move on instead of dissecting it like a mathematical theorem?
I watched only a tiny part of that talk but the speaker claims "there are no zero-cost abstractions" and justifies it with.... unique-ptr??? I mean, who doesn't know that smart pointers can be slower than raw pointers? They're not there because they're "zero-cost", but because the improvement in reliability & readability more than makes up for the runtime cost in the vast majority of situations.
He's comparing unique_ptr to just plain old-fashioned new and delete. Is it obvious to you what the actual cost is in moving between these? It's not really obvious and probably not what you think. I would listen to the full section of the talk on unique_ptr.
I developed an optimizing C++ compiler in my youth (one that was actually used in production, not just a pet project). I probably know the actual cost, and also know it may depend on compiler.
But that's not the point - the point is that smart pointers never claimed to be zero-cost abstractions, AFAIK.
> But that's not the point - the point is that smart pointers never claimed to be zero-cost abstractions, AFAIK.
> I mean, who doesn't know that smart pointers can be slower than raw pointers?
Nobody... on what basis? If you're not in that set of people, that doesn't imply the set is small or empty. Googling suggests lots of people have given such advice, and, may I suggest, it's not because they were stupid.
I don't understand the question (I didn't use the word "nobody"). I stand by the first claim (library writers & ISO C++ commitee never claimed that smart pointers are zero-cost abstractions; so why would people believe that? How did that claim start?).
The second (rhetoric) question... well, of course people would occasionally believe random stuff, for God's sake, we have (non-stupid!) people who believe COVID is a conspiracy. That said - there's no basis in the belief that smart pointers are zero-cost, so it's funny to make a talk "debunking" it - to me it looks similar to a talk that "COVID is real, guys!". I was genuinely surprised this talk is needed. I think if this is really a wide-spread belief, a more insightful talk would be about how it got to be a wide-spread belief :)
Chandler is on the cpp committee and personally said he was surprised by this cost when passing unique ptr as a parameter, as were many on the committee.
So the evidence against your claim is literally within the talk.
Ok, I guess I'll look at the talk/ didn't know the guy was in the C++ committee. Still, I personally had to fight the other way around - to convince people that "no, smart pointers don't have significant overhead over raw pointers; and the optimizations/ custom smart pointers that you're doing now, to the extent that they have any effect at all, will likely be rendered obsolete by future compiler versions & libraries".
I could understand a CPP committee member expecting a certain compiler to optimize a certain situation; however, the blanket statement that "smart pointer X is zero-cost" is more than strange, given the fact that it's always bound to be implementation-dependent, and there isn't one single C++ compiler (or even a "canonical" one) so that you could make that claim, at all. I find it really suspicious.
[edit] I watched the part of the talk - I think he was a bit surprised that has favorite compiler didn't perform that optimization, and it took him some investigation to see why. I don't think he truly expects smart pointers to be "zero-cost abstractions", I think he was just surprised that his compiler didn't optimize better that particular situation and had to dig in to find why. I still find the whole presentation a bit of a misleading stunt - there _are_ zero-cost abstractions (for some definitions of "cost" of course; and/or in some situations). E.g. in rust [1]. And even in C++ - a local unique_ptr is _probably_ a zero-cost abstraction!
BTW.. in his example - just use move semantics, put the pointer as the 5th argument, and you'll probably have the same runtime cost (ie both the raw pointer and smart pointer methods will compile to same code; since the ABI no longer allows you to pass the raw pointer via registers, it goes to the stack, so you have the additional load that bothered him in the raw pointer case, too).
> BTW.. in his example - just use move semantics, put the pointer as the 5th argument, and you'll probably have the same runtime cost
He talks about this. C++ doesn't have destructive moves like rust does, which is the root cause of why you cannot make the unique_ptr cost zero. It'd take an ABI change to fix this. This is precisely why the scenario is interesting. A lot of people (including you) thought that "just use move semantics" would solve it.
Thing is... talks like these give a wide body of application developers ammunition to say "smart pointers are slow, we need to use raw pointers!". Which is incredibly damaging. In many real-life situations, the difference literally doesn't matter (one more memory move? pfft... I'd take the whole complicated, exception-safe function body over knowing that in case of an exception I won't have a memory leak)
There are indeed cases where this sort of difference in performance matters; But those are a vanishingly small group of people; and as far as those people are concerned - they typically look at assembly anyway.
> smart pointers are slow, we need to use raw pointers!
He addresses this in the QA at the end. Chandler explicitly says "I still believe you should use unique_ptr and pass it by value". If somebody watches this video and takes away "don't use smart pointers" then they weren't paying attention.
I've been wondering why they don't permit destructor elision after a move into a newly constructed parameter. I have a hard time seeing how this would break reasonable code, and even then, we can probably find a backward compatible workaround (or rely on a compiler flag). Any idea?
ABI. C++ defines rigorous calling conventions that don't permit this. There are some on the committee who want to permit ABI breaks to enable these sorts of optimizations (in this case, destructive-move) but the committee in general has been super hesitant to permit ABI breaks.
It won't break reasonable code. The problem is it breaks interop with compiled binaries.
I'm confused, this seems unrelated to ABI. It's entirely up to the caller whether to subsequently destroy the moved-from object or to skip doing so. Whether or not the caller does so is irrelevant to the call or how the parameter gets passed, and it wouldn't affect the moved-to object (whose destructor always runs). The destruction of the moved-from object (if not elided) happens long after the call has returned - at the end of the object's scope.
I think you're confusing my proposal with the pass-in-register proposal? Or maybe I'm missing something.
What's your judgment on the Doom 3 source code? It's something performance critical, but it's in C++; Sanglard's review of the design codebase was very positive.
John Carmack wrote in 2010:
> There is some decent code in Doom Classic, but it is all C, and I would prefer to do new game development in (restrained) C++.
> I had been harboring some suspicions that our big codebases might benefit from the application of some more of the various �modern� C++ design patterns, despite seeing other large game codebases suffer under them. I have since recanted that suspicion.
I think this shows that C++ works better in controlled (i.e. corporate environments), whereas C is often preferred in Open Source.
Why do you say that? Certain things in a language may increase binary size or something else but at the end of the day these languages all go through most likely the same compiler and should produce a roughly equivalent binary.
They will produce equivalent binaries, but not identical binaries.
For example, GCC's C++ front end will produce a different parse tree for the same C code as the C front end, resulting in a slightly different collection of basic blocks and data flow analysis (obviously, for very trivial examples, it will be identical). The result after all the gimplification and and different middle-end passes and rtl transformations you can end up with a surprisingly different set of generated instructions for the same code fed to gcc and g++. Equivalent, but different.
Nitpick: gcc and g++ are not different frontends but different compiler drivers. They can both invoke the C and C++ frontends (cc1 and cc1plus) and select them based on the file extension unless specified using the -x option. The difference between gcc and g++ is mainly in which libraries are linked by default.
Well, it's true that the different compilers are cc1 and cc1plus, but given most programmers I've run into here in vacuumland have trouble distinguishing between their IDE and their toolchain, expecting them to understand the difference between the actual compiler invoked by the driver depending on a file extension or '-x' switch vs. the different compiler driver that invokes that compiler by default for a given file extension is a bit of a stretch. Even you seem to be confused by the compiler proper (cc1 or cc1plus) and the front end (the part of each compiler that parses the code and builds an internal representation that then gets passed on to the middle end for optimization and the back end for register allocation and instruction generation).
The fact is, the front ends of cc1 and cc1plus will interpret the exact same C code differently to the extent that code containing undefined behaviour will have remarkably different results in the generated instructions. The differences between the cc1 and the cc1plus compilers are not limited to what the compiler driver passes to the linker, they have completely different front ends and only share the middle and back ends.
> Even you seem to be confused by the compiler proper (cc1 or cc1plus) and the front end (the part of each compiler that parses the code and builds an internal representation that then gets passed on to the middle end for optimization and the back end for register allocation and instruction generation).
Sure, the cc1 and cc1plus binaries come with a middle and backend linked in, but what differentiates them is the front end. Not so much for gcc and g++ where both `gcc file.cpp` and `g++ file.cpp` will invoke cc1plus. Not to mention that with LTO cc1 and cc1plus do almost exclusively front end work while everything else happens in lto1 (for both C and C++).
> these languages all go through most likely the same compiler and should produce a roughly equivalent binary
That's not really true. A single compiler may support many languages but that does not mean code written in those languages ends up as the same binary.
/When/ they went, which, iirc was Netscape 4 (?). It was horrendous, crashing, slow buggy, awfulness, which the previous version really wasn't. This was when it was all closed and proprietary.
Firefox came a long time later. Well after Mozilla opened.
No idea if Firefox would be better if they'd stayed in c. What i do know is that c++ is a vastly, vastly better language now through no fault of the standards committee. G++ got good. Then came clang++, valgrind was an amazing leap forward too.
Before all that there really was no such thing as cross platform c++.
Isn't the solution painfully obvious. Remove the limits around the time you expect extreme loads. Like you said, it works most of the time. Take a hit for unexpected workload spikes. It's a design decision.
This is a misleading comparison. You are comparing a massive model with huge models. What you should be comparing are big models vs medium models that a single consumer GPU will fit. And - you don't need to take my word for it, there's tons of papers - the bigger models definitely perform better.
"Research in history involves developing an understanding of the past through the examination and interpretation of evidence. Evidence may exist in the form of texts, physical remains of historic sites, recorded data, pictures, maps, artifacts, and so on."