Usually codebases disallow auto because without an IDE it's difficult to see the type. I think this reduces the cognitive load of C++ a bit. The only time it is allowed is getting iterators types from STL containers.
I remember fretting about these rules when reading Scott Meyer's Effective C++11, and then later to realize it's better not to use auto at all. Explicit types are good types
Hard disagree. There are times when nobody, really nobody, cares about a given type. See the example of std::chrono::steady_clock::now over at cppreference.com. There you have
const auto start = std::chrono::steady_clock::now();
do_some_work(size);
const auto end = std::chrono::steady_clock::now();
const std::chrono::duration<double> diff = end - start;
std::cout << "diff = " << diff << "; size = " << size << '\n';
Looking up the (current standard's) return type of std::chrono::steady_clock::now() and spelling it out would serve no purpose here.
With 'auto' it is so very verbose. It can be shorter. Let us put "using TP = std::chrono::steady_clock::time_point;" in some header file to be used in many places. Now you can write
TP start = TP::clock::now();
do_some_work(size);
TP end = TP::clock::now();
I prefer to put the `using` in the block where you need the std::chrono code, which keeps it local and tidy. Putting it in a header is declaring a global type and asking for trouble; at least bound it in a namespace or a class.
> Usually codebases disallow auto because without an IDE it's difficult to see the type. I think this reduces the cognitive load of C++ a bit. The only time it is allowed is getting iterators types from STL containers.
Strong agree here. It's not just because it reduces cognitive load, it's because explicit types allows and requires the compiler to check your work.
Even if this isn't a problem when the code is first written, it's a nice safety belt for when someone does a refactor 6-12 months (or even 5+ years) down the road that changes a type. With auto, in the best case you might end up with 100+ lines of unintelligible error messages. In the worst case the compiler just trudges on and you have some subtle semantic breakage that takes weeks or months to chase down.
The only exceptions I like are iterators (whose types are a pita in C++), and lambda types, where you sometimes don't have any other good options because you can't afford the dynamic dispatch of std::function.
There are unspeakble types, such as lambdas, that are not exactly easy to handle without "auto". How a variable containing local lambda with captures will look without "auto"?
If it captures variables, it is not possible to manually declare the type. You can however hide it in an std::function. But then you probably get some overhead and some heap allocations in real life.
I know Google’s styleguide discourages auto, but other C++ places I’ve worked weren’t scared of it. The type deduction rules usually do what I expect, unless I have a weird pre-C++11 proxy type that somehow hasn’t been updated in the last 14 years.
I agree that auto should be used as little as possible. There are good uses, though. It is okay to use when the type is trivially inferred from the code. What is auto in "auto ptr = std::make_shared<MyNiceType>();". Everybody who knows any C++ knows. Also, lambdas do not have a type that can be written down, so it is okay to use auto for them.
I also prefer not to use auto when getting iterators from STL containers. Often I use a typedef for most STL containers that I use. The one can write MyNiceContainerType::iterator.
Pre LLM agents, a trick that I used was to type in
auto var = FunctionCall(...);
Then, in the IDE, hover over auto to show what the actual type is, and then replace auto with that type. Useful when the type is complicated, or is in some nested namespace.
This makes things seem more complicated than they already are, I feel?
There's nothing special about auto here. It deduces the same type as a template parameter, with the same collapsing rules.
decltype(auto) is a different beast and it's much more confusing. It means, more or less, "preserve the type of the expression, unless given a simple identifier, in which case use the type of the identifier."
I always try to avoid auto in c++ or var in C#. On the paper it is a nice was to save you from typing out the type but this is only true if you have tool support and can use your mouse to obtain the type. In printouts or even in a text snippet I’m am lost. I think auto inC++ was the try to shorten some overburden type constructs and make it a little more friendly to use heavy templating. Please forgive my bad English. I’m no native speaker
Right, although I would argue the most interesting part of the type here is the container, not the containee.
With good naming it should be pretty obvious it's a Foo, and then either you know the type by heart, or will need to look up the definition anyway.
With standard containers, you can have the assumption that everyone knows the type, at least high level. So knowing whether it's a list, a vector, a stack, a map or a multimap, ... is pretty useful and avoid a lookup.
an interesting demarcation of subjective mental encapsulation ... associating the anonymous type of a buffer with the buffer's name ... as opposed to explicitly specifying the type of an anonymously named buffer
I came to like auto over the years, although I also use it sparingly.
Sometimes the concrete types only add visual noise and not much helpful information, e.g. iterators:
auto it = some_container.begin();
Not even once have I wished to know the actual type of the iterator.
It’s extremely convenient for certain things. For example, let’s say I’m referring to an enum constant field of a deeply-nested type. This can easily be expressed as “auto k = a.b.c.d.kind” instead of “Alpha::Bet::Charlie::Delta::Kinds k = a.b.c.d.kind”. It should be used sparingly in local contexts where the meaning cannot be confusing.
I think IDEs are a moving target. Do you consider syntax highlighting to make something an IDE? Macro expansion? Autocomplete? intellisense-like based on full text search? based on a parser? based on the compiler? Is Kate an editor or an IDE? It has LSP support.
Having syntax highlighting makes me slightly faster, but I want to still be able to understand things, when looking at a diff or working over SSH and using cat.
I agree, but that means, that even complete integration of the compiler does not make an IDE. So the only distinguishing feature I can think of is startup time /s .
An IDE, by name, integrates the development environment. Depending on target, development lifecycle usually includes writing code, generating binaries, pushing those binaries to the target, executing on the target, and debugging on the target; so an IDE includes editing, compiling, linking, device/JTAG drivers, and debugging, in my mind. I suppose for platforms without cross-compilation the device drivers vanish, and for languages without compilers and linkers similar; but at a minimum, development means both authoring and debugging code, and so an IDE must provide an integrated view of the debugging process.
Then that distinction doesn't match the original claim, that readability of unlabeled code doesn't matter, since we now have IDEs, as this is an orthogonal matter according to your definition.
Very much true. On the other hand web based review interfaces seem to be stuck in the '60, when they could be so much better if they properly integrated with the compiler.
"bug" can refer to many categories of problems, including logic errors. I've certainly seen uninitialized variables be a source of bugs. Stackoverflow for example is full of discussions about debugging problems caused by uninitialized variables, and the word "bug" is very often used in those contexts.
I think what they mean, and what I also think is that the bug does not come from the existence of uninitialized variables. It comes from the USE of uninitialized variables. Making the variables initialized does not make the bug go away, at most it silences it. Making the program invalid instead (which is what UB fundamentally is) is way more helpful for making programs have less bugs. That the compiler still emits a program is a defect, although an unfixable one.
As to my knowledge C (and derivatives like C++) is the only common language where the question "Is this a program?" has false positives. It is certainly an interesting choice.
I mean that bug is *not in* uninitialized variable - bug in program logic. E.g. one of code pathes don't initialize variable.
So, I see uninitialized variables as a good way to find such logic errors. And, therefore, advice to always initialize variable - bad practice.
Of course if you already have a good value to initialize variable - do it. But if you have no - better leave it uninitialized.
Moreover - this will not cause safety issues in production builds because you can use `-ftrivial-auto-var-init` to initialize automatic variables to e.g. zeroes (`-fhardened` will do this too)
This is indeed exactly correct. Probably on its own this is the most important reason for most people to use it, as I think most of the millions of C++ developers in the world (and yes there are apparently millions) are not messing with compiler flags to get the checks that probably should be there by default anyway. The keyword auto gives you that.
The last time I worked meaningfully with C++ was back in 2013. Now that I write mostly Rust and TypeScript, I'm amazed by how C++ has changed over the years!
Regarding the "auto" in C++, and technically in any language, it seems conceptually wrong. The ONLY use-case I can imagine is when the type name is long, and you don't want to type it manually, or the abstractions went beyond your control, which again I don't think is a scalable approach.
Type inference is actually very useful you just need to have the right amount, too little and most of your time is spent on bureaucracy, too much and the software is incomprehensible.
In both Rust and C++ we need this because we have unnameable types, so if their type can't be inferred (in C++ deduced) we can't use these types at all.
In both languages all the lambdas are unnameable and in Rust all the functions are too (C++ doesn't have a type for functions themselves only for function pointers and we can name a function pointer type in either language)
Another place where auto can be useful is to handle cases where the function signature changes without making changes at the calling site. An explicitly typed var would need changing or worse, can work with some potential hidden bugs due to implicit type conversion.
Those are function pointers. Your parent was referring to the function type. Per ISO/IEC 9899:TC3:
A function type describes a function with specified return type. A function type is characterized by its return type and the number and types of its parameters. A function type is said to be derived from its return type, and if its return type is T , the function type is sometimes called ‘‘function returning T’’. The construction of a function type from a return type is called ‘‘function type derivation’’.
[If you've ever been under the impression that "real" people use the actual ISO text, disabuse yourself of that notion, ISO takes ages to turn the same exact words into an official branded document, then charges $$$ for a PDF, ain't nobody got time or money for that]
I can't tell you what they intended by TC3. It might be a typo or it might be some way to refer to a specific draft or a section within that draft. I doubt this particular section changes frequently so I wouldn't worry about it.
OK? But like, why a version of C99? So neither C89, which I could understand as this idea that it has "always" been that way, but also not C23, which is the current standard.
It sounds like this text is essentially the same in C23, maybe moved around a bit.
I regularly use that in C, to make sure a function matches an abstract interface. Sure, that often ends up in a function pointer, but not always and when I declare the type signature, it isn't yet a function pointer.
> but not define
I think that is because the type signature only contains the types, but no parameter names, which are required for a definition. This is arbitrary, since for data types, the member names are part of the type. It sounds totally fixable, but then you either have two types of function types, one where all parameter names are qualified and one where they aren't and only could use the former for function definitions. Or you would make names also mandatory for function declarations.
> It sounds totally fixable, but then you either have two types of function types, one where all parameter names are qualified and one where they aren't and only could use the former for function definitions
Making the names part of the type would be a bit weird, although we have seen stranger things. The biggest problem is that it would be a breaking change at least in C++.
> The biggest problem is that it would be a breaking change at least in C++.
Exactly, I believe that to be the case in C as well. In C23 the rules for types to be considered compatible were actually relaxed, so this proposal wouldn't make any sense. It would be a useless breaking change for no gain, other than a bit less typing and maybe feeding some language layers, so there is really no reason to do that. It actually makes the code less clear, since you now need to always lookup the type definition, which would be against C's philosophy.
I knew the answer to most of these intuitively but the story isn’t great. Regardless of the programming language, I’ve always been an “auto” minimalist. There are relatively few contexts where relying on inference is justified by the expedience. Ignoring the issues raises by the article, explicitness simplifies things and reduces bugs.
That said, there are some contexts in which “auto” definitely improves the situation.
You will typically find bugs go up when you do not use auto in most average C++ code bases because, as noted by another comment, a lot of C++ developers have the bad habit of writing uninitialized variables, which auto prevents.
As a C developer to me uninitialized variables are a feature and I love to invoke UB with __builtin_unreachable, because it makes the compiler able to actually bark at me on incorrect usage instead of it silently working.
v is double& in your example, not double. But it's not obvious that omitting the & causes a copy. If you see "for (auto v : vec)" looks good right? But if vec contains eg long strings, you've now murdered your perf because you're copying them out of the array instead of grabbing refs. Yes, you could make the same mistake without auto, but it's easier to notice. It's easy to forget (or not notice) that auto will not resolve to a reference in this case, because using a reference is "obviously what I want here", and the name of the feature is "auto" after all - "surely it will figure it out, right?"
Wow I really thought this would be a compile error. The implicit cast here really is a footgun. Looks like '-Wrange-loop-construct' (included in -Wall) does catch it:
> warning: loop variable 'v' of type 'const std::pair<std::__cxx11::basic_string<char>, int>&' binds to a temporary constructed from type 'std::pair<const std::__cxx11::basic_string<char>, int>' [-Wrange-loop-construct]
11 | for (const std::pair<std::string, int>& v: m) {
Why is this a perf footgun? As someone who doesn't write a lot of c++, I don't see anything intuitively wrong.
Is it that iterating over map yields something other than `std::pair`, but which can be converted to `std::pair` (with nontrivial cost) and that result is bound by reference?
It is not a cast. std::pair<const std::string, ...> and std::pair<std::string,...> are different types, although there is an implicit conversion. So a temporary is implicitly created and bound to the const reference. So not only there is a copy, you have a reference to an object that is destroyed at end of scope when you might expect it to live further.
I guess this is one of the reasons, why I don't use C++. Temporaries is a topic, where C++ on one side and me and C on the other side has had disagreements in the past. Why does changing the type even create another object at all? Why does it allocate? Why doesn't the optimizer use the effective type to optimize that away?
Each entry in the map will be copied. In C++, const T& is allowed to bind to a temporary object (whose lifetime will be extended). So a new pair is implicitly constructed, and the reference binds to this object.
It's a foot gun. Why is the default target for this gun my foot? "You should be careful to choose the target properly" isn't an answer to that question.
In what way it is a footgun? auto x = ... ; does what I would expect. Copying is usually the default in C++ and that's what I would expect to happen here.
If auto deduced reference types transparently, it would actually be more dangerous.
C++ has value semantics, which means that values of user-defined types generally behave like values of built-in types. In this sense, copying is the only logical default. It's just how the language has been designed.
Things are different in Rust because of lifetimes and destructive moves. In this context, copying would be a bad default indeed.
> because who said that's even achievable, let alone cheap?
Nobody said that. The thing is that user-defined types can be anything from tiny and cheap to huge and expensive. A language has to pick one default and be consistent. You can complain one way or the other.
I find passing by value to be sensible, but the allocating part sounds like a bad idea. Passing the value of something doesn't imply making a copy, if the value is never changed, it can be entirely optimized away.
> Passing the value of something doesn't imply making a copy
Yes, languages like Rust can automatically move variables if the compiler can prove that they will not be used anymore. Unfortunately, this is not possible in C++, so the user has to move explicitly (with std::move).
> Unfortunately, this is not possible in C++, so the user has to move explicitly (with std::move).
Honestly why not? A locally used variable sounds to be very much something the compiler can reason about. And a variable only declared in a loop, which is destroyed at the end of each iteration and only read from should be able to be optimized away. I don't know Rust, I mostly write C.
Yes, this could work for simple cases, but this breaks down pretty quickly:
void checkFoo(const Foo&);
Foo getFoo();
void example() {
std::vector<Foo> vec;
Foo foo = getFoo();
if (checkFoo(foo)) {
// *We* know that checkFoo() does not store a
// reference to 'foo' but the compiler does not
// know this. Therefore it cannot automatically
// move 'foo' into the std::vector.
vec.push_back(std::move(foo));
}
}
The fundamental problem is that C++ does not track object lifetimes. You would end up with a system where the compiler would move objects only under certain circumstances, which would be very hard to reason about.
+1 I agree 100% use it everywhere and let LS/IDE infer it.
On top of that:
* Reduce refactoring overhead (a type can be evolved or substituted, if it duck types the same, the autos don't change)
* If using auto instead of explicit types makes your code unclear, it's not clear enough code to begin with. Improve the names of methods being called or variable names being assigned to.
I can see explicit types when you are shipping a library, as part of the API your library exposes. Then you want types to be explicit and changes to be explicit.
Auto has really made c++ unapproachable to me. It's hard enough to reason about anything templated, and now I frequently see code where every method returns auto. How is any one supposed to do a code review without loading the patch into their IDE?
I'd suppose this really depends on how you are developing your codebase but most code should probably be using a trailing return type or using an auto (or template) return type with a concept/requires constraint on the return type.
For any seriously templated or metaprogrammed code nowadays a concept/requires is going to make it a lot more obvious what your code is actually doing and give you actually useful errors in the event someone is misusing your code.
Generally, you don't. I'm not sure why the parent suggested you should normally do this. However, there are occasional specific situations in which it's helpful, and that's when you use it.
1. Consistency across the board (places where it's required for metaprogramming, lambdas, etc). And as a nicety it forces function/method names to be aligned instead of having variable character counts for the return type before the names. IMHO it makes skimming code easier.
2. It's required for certain metaprogramming situations and it makes other situations an order of magnitude nicer. Nowadays you can just say `auto foo()` but if you can constrain the type either in that trailing return or in a requires clause, it makes reading code a lot easier.
3. The big one for everyday users is that trailing return type includes a lot of extra name resolution in the scope. So for example if the function is a member function/method, the class scope is automatically included so that you can just write `auto Foo::Bar() -> Baz {}` instead of `Foo::Baz Foo::Bar() {}`.
1. You're simply not going to achieve consistency across the board, because even if you dictate this by fiat, your dependencies won't be like this. The issue of the function name being hard to spot is easier to fix with tooling (just tell your editor to color them or make them bold or something). OTOH, it's not so nice to be unable to tell at a glance if the function return type is deduced or not, or what it even is in the first place.
2. It's incredibly rare for it to be required. It's not like 10% of the time, it's more like < 0.1% of the time. Just look at how many functions are in your code and how many of them actually can't be written without a trailing return type. You don't change habits to fit the tiny minority of your code.
3. This is probably the best reason to use it and the most subjective, but still not a particularly compelling argument for doing this everywhere, given how much it diverges from existing practice. And the downside is the scope also includes function parameters, which means people will refer to parameters in the return type much more than warranted, which is decidedly not always a good thing.
1) consistency, 2) scoping is different and can make it a significant difference.
I have been programming in C++ for 25 years, so I'm so used to the original syntax that I don't default to auto ... ->, but I will definitely use it when it helps simplify some complex signatures.
Same in Java and Kotlin, but then again, the majority of people who write code wouldn't understand what I mean when I say "whether a collection is returned as List or Set is an important part of the API 'contract'".
One of the craziest bugs I have had in C++ was due to auto, and it would only trigger when Trump would announce tariffs. It would have been completely preventable if I had paid full attention to my IDE feedback or the correct compiler flags were set.
Things like this are why I don't get when anyone calls Rust "more difficult than C++". I don't think I've ever encountered a more complex and cumbersome language than C++.
Rust types can be very verbose, RefCell<Optional<Rc<Vec<Node>>>>
And lifetime specifiers can be jarring. You have to think about how the function will be used at the declaration site. For example usually a function which takes two string views require different lifetimes, but maybe at call site you would only need one. It is just more verbose.
C++ has a host of complexities that come with header/source splits, janky stdlib improvements like lock_guard vs scoped_lock, quirky old syntax like virtual = 0, a lack of build systems and package managers.
> Rust types can be very verbose, RefCell<Optional<Rc<Vec<Node>>>>
Anything in any language can be very verbose and confusing if you one-line it or obfuscate it or otherwise write it in a deliberately confusing manner. That's not a meaningful point imo. What you have to do is compare what idiomatic code looks like between the two languages.
C++ has dozens of pages of dense standardese to specify how to initialise an object, full with such text as
> Only (possibly cv-qualified) non-POD class types (or arrays thereof) with automatic storage duration were considered to be default-initialized when no initializer is used. Each direct non-variant non-static data member M of T has a default member initializer or, if M is of class type X (or array thereof), X is const-default-constructible,
if T is a union with at least one non-static data member, exactly one variant member has a default member initializer,
if T is not a union, for each anonymous union member with at least one non-static data member (if any), exactly one non-static data member has a default member initializer, and each potentially constructed base class of T is const-default-constructible.
For me, it's all about inherent complexity vs incidental complexity. The having to pay attention to lifetimes is just Rust making explicit the inherent complexity of managing values and pointers thereof while making sure there isn't concurrent mutation, values moving while pointers to them exist, and no data races. This is just tough in itself. The aforementioned C++ example is just the language being byzantine and giving you 10,000 footguns when you just want to initialise a class.
> > Only (possibly cv-qualified) non-POD class types (or arrays thereof) with automatic storage duration were considered to be default-initialized when no initializer is used. Each direct non-variant non-static data member M of T has a default member initializer or, if M is of class type X (or array thereof), X is const-default-constructible, if T is a union with at least one non-static data member, exactly one variant member has a default member initializer, if T is not a union, for each anonymous union member with at least one non-static data member (if any), exactly one non-static data member has a default member initializer, and each potentially constructed base class of T is const-default-constructible.
That's just a list of very simple rules for each kind of type. As a C++-phob person, C++ has a lot of footguns, but this isn't one of them.
This was a random snippet of complicated rules to illustrate what I mean. C++ has significant footguns and complexity when it comes to initialising an object. Even sticking by "Effective Modern C++" rules is a significant cognitive burden on something that should be simple and straightforward. Rust has (mostly) no such footguns; its complexity follows from the essential complexity of the problems at hand.
Type deduction is a form of type inference, a very restricted/crude form of type inference that only considers the type of the immediate expression. The term is used in C++ because it predates the use of auto and was the term used to determine how to implicitly instantiate templates. auto uses exactly the same rules (with 1 single exception) as template type deduction, so the name was kept for familiarity. If instead of initializing a variable, you went through the examples on this website and passed the expression into a template function, the type would be deduced in exactly the same way with the exception of initializer lists.
Type inference is usually reserved for more general algorithms that can inspect not only how a variable is initialized, but how the variable used, such as what functions it's passed into, etc...
> Type inference is usually reserved for more general algorithms that can inspect not only how a variable is initialized, but how the variable used, such as what functions it's passed into, etc...
In a modern context, both would be called "type inference" because unidirectional type inference is quite a bit more common now than the bidirectional kind, given that many major languages adopted it.
If you want to specify constraint-based type inference then you can say global HM (e.g. Haskell), local HM (e.g. Rust), or just bidirectional type inference.
Type inference is when you try to infer a type from its usage. ``auto`` does no such thing, it just copies a known type from source to target. Target has no influence over source's type.
Inference is a constraint solving problem: the type of a variable depends on all the ways it is used. In C++, deduction simply sets the type of a variable from its initializing expression.
95%[1] of the time, deduction is enough, but occasionally you really wish you had proper inference.
The same general concepts often have different terminology between different languages. For example, which is better, a parent class or a super class? Or a method function or a member function? Initializer or constructor? Long list of these synonyms.
Saving this to use as an argument when C++ comes up in a discussion. This toxic swamp cannot be fixed and anyone chosing to jump into it needs a warning.
Most of the relevance of this is limited to C++ library authors doing metaprogramming.
Most of the "ugly" of these examples only really matters for library authors and even then most of the time you'd be hard pressed to put yourself in these situations. Otherwise it "just works".
Basically any adherence to a modicum of best practices avoids the bulk of the warts that come with type deduction or at worst reduces them to a compile error.
I see this argument often. It is valid right until you get first multipage error message from a code that uses stl (which is all c++ code, because it is impossible to use c++ without standard library).
Those aren't the ugly part of C++ and to be entirely honest reading those messages is not actually hard, it's just a lot of information.
Those errors are essentially the compiler telling you in order:
1. I tried to do this thing and I could not make it work.
2. Here's everything I tried in order and how each attempt failed.
If you read error messages from the top they make way more sense and if reading just the top line error doesn't tell you what's wrong, then reading through the list of resolution/type substitution failures will be insightful. In most cases the first few things it attempted will give you a pretty good idea of what the compiler was trying to do and why it failed.
If the resolution failures are a particularly long list, just ctrl-f/grep to the thing you expected to resolve/type-substitute and the compiler will tell you exactly why the thing you wanted it to use didn't work.
They aren't perfect error messages and the debugging experience of C++ metaprogramming leaves a lot to be desired but it is an order of magnitude better than it ever has been in the past and I'd still take C++ wall-o-error over the extremely over-reductive and limited errors that a lot of compilers in other languages emit (looking at you Java).
But "I tried to do this thing" error is completely useless in helping to find the reason why the compiler didn't do the thing it was expected to do, but instead chose to ignore.
Say, you hit ambiguous overload resolution, and have no idea what actually caused it. Or, conversely, implicit conversion gets hidden, and it helpfully prints all 999 operator << overloads. Or there is a bug in consteval bool type predicate, requires clause fails, and compiler helpfully dumps list of functions that have differet arguments.
How do you debug consteval, if you cannot put printf in it?
Not everyone can use clang or even latest gcc in their project, or work in a familiar codebase.
And I see this argument often. People make too much fuss about the massive error messages. Just ignore everything but the first 10 lines and 99.9% of the time, the issue is obvious. People really exaggerate the amout of time and effort you spend dealing with these error messages. They look dramatic so they're very memeable, but it's really not a big deal. The percentage of hours I've spent deciphering difficult cpp error messages in my career is a rounding error.
Do you also consider that knowing type deduction is not necessary to fix those errors, unless you are writing a library? Because that is not my experience (c++ "career" can involve such wildly different codebases, it's hard to imagine what others must be dealing with)
There's actually multiple standard libraries for embedded applications and a lot of the standard library from C++11 and on was designed with embedded in mind in particular.
And with each std release the particularly nasty parts of std get decoupled from the rest of the library. So it's at the point nowadays where you can use all the commonly used parts of std in an embedded environment. So that means you get all your containers, iterators, ranges, views, smart/RAII pointers, smart/RAII concurrency primitives. And on the bleeding edge you can even get coroutines, generators, green threads, etc in an embedded environment with "pay for what you use" overhead. Intel has been pushing embedded stdlib really hard over the past few years and both they and Nvidia have been spearheading the senders and receivers concurrency effort. Intel uses S&R internally for handling concurrency in their embedded environments internal to the CPU and elsewhere in their hardware.
(Also fun side note but STL doesn't "really" stand for "standard template library". Some projects have retroactively decided to refer to it as that but that's not where the term STL comes from. STL stands for the Adobe "Software Technology Lab" where Stepanov's STL project was formed and the project prior to being proposed to committee was named after the lab.)
AFAIK Stepanov only joined Adobe much later. I think he was at HP during the development of the STL, but moved to SGI shortly after (possibly during standardization).
The other apocryphal derivation of STL I have heard is "STepanov and Lee".
freestanding requires almost all std library. Please note that -fno-rtti and -fno-exceptions are non-conformant, c++ standard does not permit either.
Also, such std:: members as initializer_list, type_info etc are directly baked into compiler and stuff in header must exactly match internals — making std library a part of compiler implementation
have you actually read the page you linked to? None of the standard containers is there, nor <iostream> or <algorithm>. <string> is there but marked as partial.
If anything, I would expect more headers like <algorithm>, <span>, <array> etc to be there as they mostly do not require any heap allocation nor exceptions for most of their functionality. And in fact they are available with GCC.
The only bit I'm surprised is that coroutine is there, as they normally allocate, but I guess it has full support for custom allocators, so it can be made to work on freestanding.
> Please note that -fno-rtti and -fno-exceptions are non-conformant, c++ standard does not permit either.
I did not know that.
My understanding was that C does not require standard library functions to be present in freestanding. The Linux kernel famously does not build in freestanding mode, since then GCC can't reason about the standard library functions which they want. This means that they need to implement stuff like memcpy and pass -fno-builtin.
Does that mean that freestanding C++ requires the C++ standard library, but not the C standard library? How does that work?
Honestly? No idea how the committee is thinking. When, say, gamedev people write proposal, ask for a feature, explain it is important and something they depend on and so on, it gets shot down on technicality. Then they turn around and produce some insane feature that, like, rips everything east to west (like modules), and suddenly voting goes positive.
The "abstract machine" C++ assumes in the standard is itself a deeply puzzling construct. Luckily, compiler authors seem much more pragmatic and reasonable, I do not fear -fno-exceptions dissapearing suddenly, or code that accesses mmapped data becoming invalid because it didn't use start_lifetime_as
One of required headers in freestanding, <cstdlib>, is labelled "C standard library", but it is not <stdlib.h>
Something similar with other <csomething> headers.
This kinda implies C library is required, if I read it correctly, but maybe someone else can correct me:
https://eel.is/c++draft/library.c
> The ISO C standard defines (in clause 4) two classes of conforming
implementation. A "conforming hosted implementation" supports the whole
standard including all the library facilities; a "conforming
freestanding implementation" is only required to provide certain library
facilities: those in '<float.h>', '<limits.h>', '<stdarg.h>', and
'<stddef.h>'; since AMD1, also those in '<iso646.h>'; since C99, also
those in '<stdbool.h>' and '<stdint.h>'; and since C11, also those in
'<stdalign.h>' and '<stdnoreturn.h>'. In addition, complex types, added
in C99, are not required for freestanding implementations.
> The standard also defines two environments for programs, a
"freestanding environment", required of all implementations and which
may not have library facilities beyond those required of freestanding
implementations, where the handling of program startup and termination
are implementation-defined; and a "hosted environment", which is not
required, in which all the library facilities are provided and startup
is through a function 'int main (void)' or 'int main (int, char *[])'.
An OS kernel is an example of a program running in a freestanding
environment; a program using the facilities of an operating system is an
example of a program running in a hosted environment.
> GCC aims towards being usable as a conforming freestanding
implementation, or as the compiler for a conforming hosted
implementation. By default, it acts as the compiler for a hosted
implementation, defining '__STDC_HOSTED__' as '1' and presuming that
when the names of ISO C functions are used, they have the semantics
defined in the standard. To make it act as a conforming freestanding
implementation for a freestanding environment, use the option
'-ffreestanding'; it then defines '__STDC_HOSTED__' to '0' and does not
make assumptions about the meanings of function names from the standard
library, with exceptions noted below. To build an OS kernel, you may
well still need to make your own arrangements for linking and startup.
*Note Options Controlling C Dialect: C Dialect Options.
> GCC does not provide the library facilities required only of hosted
implementations, nor yet all the facilities required by C99 of
freestanding implementations on all platforms. To use the facilities of
a hosted environment, you need to find them elsewhere (for example, in
the GNU C library). *Note Standard Libraries: Standard Libraries.
> Most of the compiler support routines used by GCC are present in
'libgcc', but there are a few exceptions. GCC requires the freestanding
environment provide 'memcpy', 'memmove', 'memset' and 'memcmp'.
Finally, if '__builtin_trap' is used, and the target does not implement
the 'trap' pattern, then GCC emits a call to 'abort'.
So the last paragraph means that my remark about the Linux kernel might be wrong.
So the required headers are all about basic constants for types, the types themselves (bool), and basic language features like stdarg, iso646 or stdalign. Sounds sensible to me. Not sure what C++ does with that.
This also actually matches the links provided by you. In https://eel.is/c++draft/cstdlib.syn you see that not all declarations are actually marked for freestanding implementations.
I remember fretting about these rules when reading Scott Meyer's Effective C++11, and then later to realize it's better not to use auto at all. Explicit types are good types
Strong agree here. It's not just because it reduces cognitive load, it's because explicit types allows and requires the compiler to check your work.
Even if this isn't a problem when the code is first written, it's a nice safety belt for when someone does a refactor 6-12 months (or even 5+ years) down the road that changes a type. With auto, in the best case you might end up with 100+ lines of unintelligible error messages. In the worst case the compiler just trudges on and you have some subtle semantic breakage that takes weeks or months to chase down.
The only exceptions I like are iterators (whose types are a pita in C++), and lambda types, where you sometimes don't have any other good options because you can't afford the dynamic dispatch of std::function.
I also prefer not to use auto when getting iterators from STL containers. Often I use a typedef for most STL containers that I use. The one can write MyNiceContainerType::iterator.
auto var = FunctionCall(...);
Then, in the IDE, hover over auto to show what the actual type is, and then replace auto with that type. Useful when the type is complicated, or is in some nested namespace.
There's nothing special about auto here. It deduces the same type as a template parameter, with the same collapsing rules.
decltype(auto) is a different beast and it's much more confusing. It means, more or less, "preserve the type of the expression, unless given a simple identifier, in which case use the type of the identifier."
So for example I'd write:
Because writing: Feels very redundantWhereas I'd write:
Because it's not obvious what the type is otherwise ( Some IDEs will overlay hint the type there though ).Although with newer language features you can also write:
Which is even better.With good naming it should be pretty obvious it's a Foo, and then either you know the type by heart, or will need to look up the definition anyway.
With standard containers, you can have the assumption that everyone knows the type, at least high level. So knowing whether it's a list, a vector, a stack, a map or a multimap, ... is pretty useful and avoid a lookup.
Nowadays I only use
in non-merged code as a ghetto TODO if I'm considering base types/interface.auto it = some_container.begin();
Not even once have I wished to know the actual type of the iterator.
IDEs are an invention from the late 1970's, early 1980's.
Having syntax highlighting makes me slightly faster, but I want to still be able to understand things, when looking at a diff or working over SSH and using cat.
For example:
auto a;
will always fail to compile not matter what flags.
int a;
is valid.
Also it prevents implicit type conversions, what you get as type on auto is the type you put at the right.
That's good.
What do you mean it is not a source of bugs?
I think what they mean, and what I also think is that the bug does not come from the existence of uninitialized variables. It comes from the USE of uninitialized variables. Making the variables initialized does not make the bug go away, at most it silences it. Making the program invalid instead (which is what UB fundamentally is) is way more helpful for making programs have less bugs. That the compiler still emits a program is a defect, although an unfixable one.
As to my knowledge C (and derivatives like C++) is the only common language where the question "Is this a program?" has false positives. It is certainly an interesting choice.
So, I see uninitialized variables as a good way to find such logic errors. And, therefore, advice to always initialize variable - bad practice.
Of course if you already have a good value to initialize variable - do it. But if you have no - better leave it uninitialized.
Moreover - this will not cause safety issues in production builds because you can use `-ftrivial-auto-var-init` to initialize automatic variables to e.g. zeroes (`-fhardened` will do this too)
Regarding the "auto" in C++, and technically in any language, it seems conceptually wrong. The ONLY use-case I can imagine is when the type name is long, and you don't want to type it manually, or the abstractions went beyond your control, which again I don't think is a scalable approach.
In both Rust and C++ we need this because we have unnameable types, so if their type can't be inferred (in C++ deduced) we can't use these types at all.
In both languages all the lambdas are unnameable and in Rust all the functions are too (C++ doesn't have a type for functions themselves only for function pointers and we can name a function pointer type in either language)
C has this, so I think C++ has as well. You can use a typedef'ed function to declare a function, not just for the function pointer.
A function type describes a function with specified return type. A function type is characterized by its return type and the number and types of its parameters. A function type is said to be derived from its return type, and if its return type is T , the function type is sometimes called ‘‘function returning T’’. The construction of a function type from a return type is called ‘‘function type derivation’’.
> Per ISO/IEC 9899:TC3:
What is it supposed to tell me?
You can read a "draft" of that document here: https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3220.pdf
[If you've ever been under the impression that "real" people use the actual ISO text, disabuse yourself of that notion, ISO takes ages to turn the same exact words into an official branded document, then charges $$$ for a PDF, ain't nobody got time or money for that]
I can't tell you what they intended by TC3. It might be a typo or it might be some way to refer to a specific draft or a section within that draft. I doubt this particular section changes frequently so I wouldn't worry about it.
It sounds like this text is essentially the same in C23, maybe moved around a bit.
edit: you literally said this in your original comment. I failed at reading comprehension.
I regularly use that in C, to make sure a function matches an abstract interface. Sure, that often ends up in a function pointer, but not always and when I declare the type signature, it isn't yet a function pointer.
> but not define
I think that is because the type signature only contains the types, but no parameter names, which are required for a definition. This is arbitrary, since for data types, the member names are part of the type. It sounds totally fixable, but then you either have two types of function types, one where all parameter names are qualified and one where they aren't and only could use the former for function definitions. Or you would make names also mandatory for function declarations.
> It sounds totally fixable, but then you either have two types of function types, one where all parameter names are qualified and one where they aren't and only could use the former for function definitions
Making the names part of the type would be a bit weird, although we have seen stranger things. The biggest problem is that it would be a breaking change at least in C++.
Exactly, I believe that to be the case in C as well. In C23 the rules for types to be considered compatible were actually relaxed, so this proposal wouldn't make any sense. It would be a useless breaking change for no gain, other than a bit less typing and maybe feeding some language layers, so there is really no reason to do that. It actually makes the code less clear, since you now need to always lookup the type definition, which would be against C's philosophy.
That said, there are some contexts in which “auto” definitely improves the situation.
I've seen much more perf-murdering things being caused by
than with auto though> warning: loop variable 'v' of type 'const std::pair<std::__cxx11::basic_string<char>, int>&' binds to a temporary constructed from type 'std::pair<const std::__cxx11::basic_string<char>, int>' [-Wrange-loop-construct] 11 | for (const std::pair<std::string, int>& v: m) {
As they say, the power of names...
Is it that iterating over map yields something other than `std::pair`, but which can be converted to `std::pair` (with nontrivial cost) and that result is bound by reference?
Is it really? I rather think that a missing & is easier to spot with "auto" simply because there is less text to parse for the eye.
> If you see "for (auto v : vec)" looks good right?
For me the missing & sticks out like a sore thumb.
> It's easy to forget (or not notice) that auto will not resolve to a reference in this case
Every feature can be misused if the user forgets how it works. I don't think people suddenly forget how "auto" works, given how ubiquitous it is.
If auto deduced reference types transparently, it would actually be more dangerous.
So I guess I depart from you there and thus my issue here is not really about auto
Things are different in Rust because of lifetimes and destructive moves. In this context, copying would be a bad default indeed.
> because who said that's even achievable, let alone cheap?
Nobody said that. The thing is that user-defined types can be anything from tiny and cheap to huge and expensive. A language has to pick one default and be consistent. You can complain one way or the other.
Yes, languages like Rust can automatically move variables if the compiler can prove that they will not be used anymore. Unfortunately, this is not possible in C++, so the user has to move explicitly (with std::move).
Honestly why not? A locally used variable sounds to be very much something the compiler can reason about. And a variable only declared in a loop, which is destroyed at the end of each iteration and only read from should be able to be optimized away. I don't know Rust, I mostly write C.
My LS can infer it anytime.
On top of that:
* Reduce refactoring overhead (a type can be evolved or substituted, if it duck types the same, the autos don't change)
* If using auto instead of explicit types makes your code unclear, it's not clear enough code to begin with. Improve the names of methods being called or variable names being assigned to.
I can see explicit types when you are shipping a library, as part of the API your library exposes. Then you want types to be explicit and changes to be explicit.
``` struct dummy{}; dummy d = ANYTHING_YOU_WANT_GO_GET_THE_TYPE; ```
Compile it with g++, and get the type info from compilation error -:)
For any seriously templated or metaprogrammed code nowadays a concept/requires is going to make it a lot more obvious what your code is actually doing and give you actually useful errors in the event someone is misusing your code.
1. Consistency across the board (places where it's required for metaprogramming, lambdas, etc). And as a nicety it forces function/method names to be aligned instead of having variable character counts for the return type before the names. IMHO it makes skimming code easier.
2. It's required for certain metaprogramming situations and it makes other situations an order of magnitude nicer. Nowadays you can just say `auto foo()` but if you can constrain the type either in that trailing return or in a requires clause, it makes reading code a lot easier.
3. The big one for everyday users is that trailing return type includes a lot of extra name resolution in the scope. So for example if the function is a member function/method, the class scope is automatically included so that you can just write `auto Foo::Bar() -> Baz {}` instead of `Foo::Baz Foo::Bar() {}`.
2. It's incredibly rare for it to be required. It's not like 10% of the time, it's more like < 0.1% of the time. Just look at how many functions are in your code and how many of them actually can't be written without a trailing return type. You don't change habits to fit the tiny minority of your code.
3. This is probably the best reason to use it and the most subjective, but still not a particularly compelling argument for doing this everywhere, given how much it diverges from existing practice. And the downside is the scope also includes function parameters, which means people will refer to parameters in the return type much more than warranted, which is decidedly not always a good thing.
I have been programming in C++ for 25 years, so I'm so used to the original syntax that I don't default to auto ... ->, but I will definitely use it when it helps simplify some complex signatures.
And lifetime specifiers can be jarring. You have to think about how the function will be used at the declaration site. For example usually a function which takes two string views require different lifetimes, but maybe at call site you would only need one. It is just more verbose.
C++ has a host of complexities that come with header/source splits, janky stdlib improvements like lock_guard vs scoped_lock, quirky old syntax like virtual = 0, a lack of build systems and package managers.
Anything in any language can be very verbose and confusing if you one-line it or obfuscate it or otherwise write it in a deliberately confusing manner. That's not a meaningful point imo. What you have to do is compare what idiomatic code looks like between the two languages.
C++ has dozens of pages of dense standardese to specify how to initialise an object, full with such text as
> Only (possibly cv-qualified) non-POD class types (or arrays thereof) with automatic storage duration were considered to be default-initialized when no initializer is used. Each direct non-variant non-static data member M of T has a default member initializer or, if M is of class type X (or array thereof), X is const-default-constructible, if T is a union with at least one non-static data member, exactly one variant member has a default member initializer, if T is not a union, for each anonymous union member with at least one non-static data member (if any), exactly one non-static data member has a default member initializer, and each potentially constructed base class of T is const-default-constructible.
For me, it's all about inherent complexity vs incidental complexity. The having to pay attention to lifetimes is just Rust making explicit the inherent complexity of managing values and pointers thereof while making sure there isn't concurrent mutation, values moving while pointers to them exist, and no data races. This is just tough in itself. The aforementioned C++ example is just the language being byzantine and giving you 10,000 footguns when you just want to initialise a class.
That's just a list of very simple rules for each kind of type. As a C++-phob person, C++ has a lot of footguns, but this isn't one of them.
Type inference is usually reserved for more general algorithms that can inspect not only how a variable is initialized, but how the variable used, such as what functions it's passed into, etc...
In a modern context, both would be called "type inference" because unidirectional type inference is quite a bit more common now than the bidirectional kind, given that many major languages adopted it.
If you want to specify constraint-based type inference then you can say global HM (e.g. Haskell), local HM (e.g. Rust), or just bidirectional type inference.
https://news.ycombinator.com/item?id=11604474
95%[1] of the time, deduction is enough, but occasionally you really wish you had proper inference.
[1] percentage made up on the spot.
Most of the "ugly" of these examples only really matters for library authors and even then most of the time you'd be hard pressed to put yourself in these situations. Otherwise it "just works".
Basically any adherence to a modicum of best practices avoids the bulk of the warts that come with type deduction or at worst reduces them to a compile error.
Those errors are essentially the compiler telling you in order:
1. I tried to do this thing and I could not make it work.
2. Here's everything I tried in order and how each attempt failed.
If you read error messages from the top they make way more sense and if reading just the top line error doesn't tell you what's wrong, then reading through the list of resolution/type substitution failures will be insightful. In most cases the first few things it attempted will give you a pretty good idea of what the compiler was trying to do and why it failed.
If the resolution failures are a particularly long list, just ctrl-f/grep to the thing you expected to resolve/type-substitute and the compiler will tell you exactly why the thing you wanted it to use didn't work.
They aren't perfect error messages and the debugging experience of C++ metaprogramming leaves a lot to be desired but it is an order of magnitude better than it ever has been in the past and I'd still take C++ wall-o-error over the extremely over-reductive and limited errors that a lot of compilers in other languages emit (looking at you Java).
But "I tried to do this thing" error is completely useless in helping to find the reason why the compiler didn't do the thing it was expected to do, but instead chose to ignore.
Say, you hit ambiguous overload resolution, and have no idea what actually caused it. Or, conversely, implicit conversion gets hidden, and it helpfully prints all 999 operator << overloads. Or there is a bug in consteval bool type predicate, requires clause fails, and compiler helpfully dumps list of functions that have differet arguments.
How do you debug consteval, if you cannot put printf in it?
Not everyone can use clang or even latest gcc in their project, or work in a familiar codebase.
Modern C++ in general is so hostile to debugging I think it's astounding people actually use it.
Citation needed. This is common for embedded application, since why would anyone program a STL for that?
And with each std release the particularly nasty parts of std get decoupled from the rest of the library. So it's at the point nowadays where you can use all the commonly used parts of std in an embedded environment. So that means you get all your containers, iterators, ranges, views, smart/RAII pointers, smart/RAII concurrency primitives. And on the bleeding edge you can even get coroutines, generators, green threads, etc in an embedded environment with "pay for what you use" overhead. Intel has been pushing embedded stdlib really hard over the past few years and both they and Nvidia have been spearheading the senders and receivers concurrency effort. Intel uses S&R internally for handling concurrency in their embedded environments internal to the CPU and elsewhere in their hardware.
(Also fun side note but STL doesn't "really" stand for "standard template library". Some projects have retroactively decided to refer to it as that but that's not where the term STL comes from. STL stands for the Adobe "Software Technology Lab" where Stepanov's STL project was formed and the project prior to being proposed to committee was named after the lab.)
The other apocryphal derivation of STL I have heard is "STepanov and Lee".
freestanding requires almost all std library. Please note that -fno-rtti and -fno-exceptions are non-conformant, c++ standard does not permit either.
Also, such std:: members as initializer_list, type_info etc are directly baked into compiler and stuff in header must exactly match internals — making std library a part of compiler implementation
have you actually read the page you linked to? None of the standard containers is there, nor <iostream> or <algorithm>. <string> is there but marked as partial.
If anything, I would expect more headers like <algorithm>, <span>, <array> etc to be there as they mostly do not require any heap allocation nor exceptions for most of their functionality. And in fact they are available with GCC.
The only bit I'm surprised is that coroutine is there, as they normally allocate, but I guess it has full support for custom allocators, so it can be made to work on freestanding.
I did not know that.
My understanding was that C does not require standard library functions to be present in freestanding. The Linux kernel famously does not build in freestanding mode, since then GCC can't reason about the standard library functions which they want. This means that they need to implement stuff like memcpy and pass -fno-builtin.
Does that mean that freestanding C++ requires the C++ standard library, but not the C standard library? How does that work?
The "abstract machine" C++ assumes in the standard is itself a deeply puzzling construct. Luckily, compiler authors seem much more pragmatic and reasonable, I do not fear -fno-exceptions dissapearing suddenly, or code that accesses mmapped data becoming invalid because it didn't use start_lifetime_as
> freestanding C++ requires the C++ standard library, but not the C standard library
is true?
> The "abstract machine" C++ assumes in the standard is itself a deeply puzzling construct.
I find the abstract machine to be quite a neat abstraction, but I am also more of a C guy.
https://eel.is/c++draft/compliance
One of required headers in freestanding, <cstdlib>, is labelled "C standard library", but it is not <stdlib.h> Something similar with other <csomething> headers.
This kinda implies C library is required, if I read it correctly, but maybe someone else can correct me: https://eel.is/c++draft/library.c
> The ISO C standard defines (in clause 4) two classes of conforming implementation. A "conforming hosted implementation" supports the whole standard including all the library facilities; a "conforming freestanding implementation" is only required to provide certain library facilities: those in '<float.h>', '<limits.h>', '<stdarg.h>', and '<stddef.h>'; since AMD1, also those in '<iso646.h>'; since C99, also those in '<stdbool.h>' and '<stdint.h>'; and since C11, also those in '<stdalign.h>' and '<stdnoreturn.h>'. In addition, complex types, added in C99, are not required for freestanding implementations.
> The standard also defines two environments for programs, a "freestanding environment", required of all implementations and which may not have library facilities beyond those required of freestanding implementations, where the handling of program startup and termination are implementation-defined; and a "hosted environment", which is not required, in which all the library facilities are provided and startup is through a function 'int main (void)' or 'int main (int, char *[])'. An OS kernel is an example of a program running in a freestanding environment; a program using the facilities of an operating system is an example of a program running in a hosted environment.
> GCC aims towards being usable as a conforming freestanding implementation, or as the compiler for a conforming hosted implementation. By default, it acts as the compiler for a hosted implementation, defining '__STDC_HOSTED__' as '1' and presuming that when the names of ISO C functions are used, they have the semantics defined in the standard. To make it act as a conforming freestanding implementation for a freestanding environment, use the option '-ffreestanding'; it then defines '__STDC_HOSTED__' to '0' and does not make assumptions about the meanings of function names from the standard library, with exceptions noted below. To build an OS kernel, you may well still need to make your own arrangements for linking and startup. *Note Options Controlling C Dialect: C Dialect Options.
> GCC does not provide the library facilities required only of hosted implementations, nor yet all the facilities required by C99 of freestanding implementations on all platforms. To use the facilities of a hosted environment, you need to find them elsewhere (for example, in the GNU C library). *Note Standard Libraries: Standard Libraries.
> Most of the compiler support routines used by GCC are present in 'libgcc', but there are a few exceptions. GCC requires the freestanding environment provide 'memcpy', 'memmove', 'memset' and 'memcmp'. Finally, if '__builtin_trap' is used, and the target does not implement the 'trap' pattern, then GCC emits a call to 'abort'.
So the last paragraph means that my remark about the Linux kernel might be wrong.
So the required headers are all about basic constants for types, the types themselves (bool), and basic language features like stdarg, iso646 or stdalign. Sounds sensible to me. Not sure what C++ does with that.
This matches with https://en.cppreference.com/w/c/language/conformance.html . Since C23, stdlib.h is also required for dynamic allocation, string conversions and some other things.
This also actually matches the links provided by you. In https://eel.is/c++draft/cstdlib.syn you see that not all declarations are actually marked for freestanding implementations.