For your level 2 code, `uint64_t data[];` is wrong for types whose alignment is greater than `uint64_t`, and also wasteful for types whose alignment is smaller (for example, under an ilp32 ABI on 64-bit architectures).
For your level 3 code, it should be `int main() { List(Foo) foo_list = {NULL};`
Note that working around a lack of `typeof` means you can't return anything. Also, your particular workaround allows `const`ness errors since `==` is symmetrical.
You can't safely omit `payload` since you need it to know the correct size. Consider a `List(int64_t)` and you try to add an `int32_t` to it - this should be fine, but you can't `sizeof` the `int32_t`. Your code is actually lacking quite a bit to make this work.
=====
There are 2 major limitations to generics in C right now:
* Delegating to a vtable (internal or external) is limited in functionality, since structs cannot contain macros, only functions.
* Delegating to an external vtable (mandatory to avoid overhead) means that you have to forward-declare all of the types you'll ever use a vtable with. So far the best approach I've found is to declare (but not define) static functions in the same forwarding header I declare the typedefs in; note that GCC and Clang differ in what phase the "undefined static" warning appears in for the case where you don't actually include that particular type's header in a given TU.
(think about writing a function that accepts either `struct SizedBuffer {void *p; size_t len;};` or `struct BoundedBuffer {void *begin; void *end;};`, and also const versions thereof - all from different headers).
> Delegating to an external vtable (mandatory to avoid overhead) means that you have to forward-declare all of the types you'll ever use a vtable with.
We went down the rabbit hole of writing a compiler for this as part of a project I used to work on (Apache Clownfish[1], a subproject of the retired Apache Lucy project). We started off parsing .h files, but eventually it made sense to create our own small header language (.cfh "Clownfish Header" files).
Here's some generated code for invoking the CharBuf version of the "Clone" method defined in parent class "Obj":
We had our reasons for going to these extremes: the point of Clownfish was to provide a least-common-denominator object model for bindings to multiple dynamic languages (similar problem domain to SWIG), and the .cfh files also were used to derive types for the binding languages. But there was truly an absurd amount of boilerplate being generated to get around the issue you identify.
This is why almost everybody just uses casts to void* for the invocant, skipping type safety.
i am firmly of the opinion that compiling to c is a better route than doing clever c tricks to sort of get what you want. the compiler can be pretty minimal and as you note it pays for itself.
There’s some prior work called CFront. It implements a superset of C that’s just an “increment”. I think it’s worth looking into, it might take off one day!
> it should be `int main() { List(Foo) foo_list = {NULL};`
In C `int main()` means the function takes an unknown number of arguments. You need `int main(void)` to mean it doesn't take any arguments. This is a fact frequently forgotten by those who write C++.
I believe that since C23 foo() is now a nullary function. As this is the last approved standard and it supersedes all previous standards, it is technically correct to say that de-jure this is what the (unqualified) C standard mandates.
I would love for `union`s to be federated, that is, a type could declare itself as thought it was part of a union with another type, without having to pre-declare all possible types in one place.
For layout-compatible types, you can often just include a `_base` member in each child. Maybe twice (once named and once unnamed) to avoid excess typing - I don't understand the common-initial-subsequence rule but people do this enough that compilers have to allow it.
There's no a problem with the author's current code, since the padding is already included in the node size, but it would be a problem after doing alignment more intelligently.
> Structurally identical types will be considered the same type in GCC 15 and Clang later in 2025 thanks to a rule change
Beware that only tagged unions are considered the same type under the new rule, provided they have the same structrure and the same tag.
The List(T) macro should be changed to generate a different tag for each different T. Which is trivial (with ##) for simple one-word types, but impossible for even mildly complex ones like pointers to char (strings).
Of course you can force yourself to typedef any type before using it in a List, but it looses much of its versatility. Example:
The trick#0 you mention is how I made an entire C dialect. Here is a generic binary heap, for example https://github.com/gritzko/librdx/blob/master/abc/HEAPx.h The syntax is a bit heavyweight, but a huge huge advantage is: you get regular C structs in the end, very plain, very predictable, very optimizable. Compiler would eat them like donuts.
In the other cases, it is void* and runtime memory sizing and you have to define macros anyway.
Author here. Binary heaps and linked lists are different use cases. A binary heap must read the data you put in it to store it correctly, but a linked list doesn't. If I were writing a generic binary heap, maybe I would weigh my options differently. I mentioned this in the footnotes.
I agree, there are actually several reasons to prefer the header impl. Debugging is better, both because you can step through the header code where you can’t with a macro function, and because the type information available to the debugger is better. There are more opportunities for compiler optimizations because each instantiation is monomorphized and you don’t pay a runtime cost with variable sizing, generic structures can also be placed on the stack because of the fixed sizing.
There are workarounds for at least two of the problems the author mentions. Naming can be changed from Bar_func(args…) to func(Bar)(args…) with a function name macro that just does name mangling. You can avoid some of the binary bloat by using weak symbols, letting the linker deduplicate functions shared between translation units at link time.
There are other problems for generic containers of pointer types however, you can work around them by using a typedef or a type alias.
Intrusive data structures are more convenient in C still, but working with them in a debugger is a pain.
Author here. It's worth noting that no work is being done in the macros of my article, they compile down to a normal c function call which you can step through in a debugger.
There is little benefit in monomorphizing the implementation of a data structure like a linked list where its behavior doesn't depend on the contents of the data it contains (compared to, say, a max heap)
The casting of the function type assumes that the item pointer type (e.g. Foo*) has the same representation as void*, which the C standard doesn’t guarantee (in standardese: the two types aren’t “compatible”). Calling the function with the converted type therefore constitutes undefined behavior. It also impacts aliasing analysis by compilers (see [0], incidentally), even if the pointer representation happens to be the same.
This casting of the functions to different argument types constitutes the core of the type safety of the generic invocations; I’m not sure it can be fixed.
Ah, that’s what I get for not reading the footnotes. However, the alternative solution presented evaluates the item argument twice, which is problematic as well (but could probably be worked around by passing `(list)->payload` on instead). Secondly, the assignment for type-checking doesn’t work for read-only operations on a const List, or does it? And doesn’t the assignment overwrite the head? Lastly, the do-while construction means you can’t use it for operations that return a value (without compiler extensions).
I also don’t agree it’s “squeamish” to be wary of aliasing analysis going wrong. It’s not a clean abstraction and can hide subtle bugs down the road.
Sure, but your alternative code incorrectly assigns to (list)->payload. You have many other options. Without typeof, you can if(0) the assignment, or check type compatibility with a ternary operator like 1 ? (item) : (list)->payload and pass that to _list_prepend, etc. With typeof, you can store item in a temporary variable with the same type as (list)->payload, or build a compound literal (typeof(*(list))){.payload=(item)}, etc.
because i work on a legacy project that is coupled to safety regulations and other quality guarantees, and we cannot just simply roll out a solution ported to c++ on the next release, or even tenth, so perhaps we make it work until we can.
however we can set a standard and expectation for new projects to use c++, and we do and set an expectation to target a specific std.
i see this sentiment quite a lot on hackernews -- feels like a lot of people saying "git gud" -- i would expect a lot more nuance applied here.
Any established C codebase, for example the kernel or Postgres?
Traditionally microcontroller firmwares as well, though those are increasingly friendly to C++, you just have to be careful about allocations as C++ makes it way easier to accidentally allocate than C does.
I'm not sure about other compilers, but compiling C code as C++ with MSVC ends up with pretty much the exact same code, instruction by instruction. C++ is a bit more strict though especially with casting, so a lot of code won't compile out of the box.
C++ code compiles to a different function names in object file (name mangling). You probably need to put a lot of ugly `#ifdef __cplusplus extern "C" {` boilerplate in your headers, otherwise C and C++ files will not compile together.
That's possible, but then the people building your extension need a C++ toolchain.
The question was "please provide examples where switching to C++ involves jumping through even more hoops", and in my view requiring downstream to use a C++ environment when they're expecting to use a C environment qualifies.
There's also the method used in the Linux kernel to embed the list information (struct list_head) within the type specific struct.
https://kernelnewbies.org/FAQ/LinkedLists
INIT_LIST_HEAD is of form VERB_NOUN so is called from within a function to programatically initialise the list.
LIST_HEAD_INIT is NOUN_VERB and is used within a structure initialiser not from a function.
But my main point was to show the "embed the list in the data" approach rather than "embed the data in the list" or "point to the data from the list" and not to discuss the naming details in the kernel implementation of the concept.
I guess if anyone might know it might be you—do you see any way of doing this for intrusive data structures, embedding the node struct in the data (and as side effect supporting an object to be on multiple containers) rather than the data in the node like you're doing there?
You could put the dummy member into the embedded node. But for intrusive data structures you often want them to erase the type so that you write generic algorithms as regular functions. In this case, it makes more sense to have a run-time check do to down casts. I do this with my variadic type which has an intrusive "super" member: https://godbolt.org/z/ofdKe7Pfv The overhead is often completely removed by the compiler.
Mhm. Putting the dummy member into the embedded node doesn't give a path from the proper object to find the embedded node "mid-struct". run-time checks are the "easy way out". We/I'll stick to macro soup probably, so we have compile-time checks.
btw. For ISO C WG14… has anyone suggested adding _Include to the preprocessor, along the lines of _Pragma? It'd really help with doing this kind of really long macros, hiding the clunky "#define, #define, #define, #include" inside a macro…
When I saw the title I assumed it was originally "Why I" or "How I" and was trimmed automatically by HN, but this is the original. Could it be that the author was influenced by HN's title guidelines and titled it thus?
Interesting idea with the union and using typeof(). We (I) went with large macros defining wrappers instead, which, I believe, is a necessity with intrusive data structures, or at least I don't immediately see how to do that with unions & typeof. Maybe it's possible...?
The key idea here seems to be to use function pointer's type to enforce type safety rather than using the data "handle" type (that is often found in implementations inspired by Sean Barrett's strechy_buffers).
> One annoying thing about C is that it does not consider these two variables to have the same type
Author here. Not quite. The key idea is about using a union to associate type information with a generic data type. Type casting a function is not the only way to use that type information. I discuss that as well as the C23 changes in the footnotes and the "typeof on old compilers" section.
FWIW, as far back as 2015 my feature check library documents Visual Studio as supporting "__typeof".[1] Note the leading but not trailing underscores. Perhaps I was mistaken, but I usually tested that sort of thing. It's also possible __typeof had slightly different semantics.
msvc 19.39 is the first to support it, which I mention in the article. You can confirm it didn't work up through 19.38 in godbolt [1]. I don't use Visual Studio, so I don't know what version of that first started using msvc 19.39
So somehow I had the impression Visual Studio .NET 2003 (7.1)[1] added __typeof. I'm still holding out hope someone will come to my rescue and reply that once upon a time MSVC had __typeof, but removed it. But for now it seems past me is gaslighting present me.
EDIT: Ah ha! It seems Microsoft did support __typeof, but perhaps only for "managed" C++ (aka C++ .NET)?
> One thing to watch out for when using the __typeof operator in managed C++ is that __typeof(wchar_t) can return different values depending on the compilation options.
struct ListNode(T) {
ListNode* next;
T data;
}
T!int node;
Why suffer the C preprocessor? Using preprocessor macros is like using a hammer for finish carpentry, rather than a nail gun. A nail gun is 10x faster, drives the nail perfectly every time, and no half moon dents in your work.
If I may may be provocative :-) this post isn't about C. It's about layering on a custom language using C preprocessor macros.
My compilers were originally written in C. I started using the C preprocessor to do metaprogramming. After some years I got fed up with it and removed nearly all of the preprocessor use, and never looked back. My code was much easier to understand.
An amusing story: long ago, a friend of mine working for Microsoft was told by a team leader that a 50K program had a bug in it, and sadly the developer was long gone. He'd assigned programmer after programmer to it, who could not fix it. My friend said he'd give it a try, and had it fixed in 2 hours.
The source code was written in Microsoft MASM, where the M stood for "Macro". You can guess where this is going. The developer had invented his own high level language using the macro system (which was much more powerful than C's). Unfortunately, he neglected to document it, and the other programmers spent weeks studying it and could not figure it out.
The leader, astonished, asked him how he figured it out in 2 hours? My friend said simple. He assembled it to object code, then disassembled the object code with obj2asm (a disassembler I wrote that converts object code back to source code). He then immediately found and fixed the bug, and checked in the "new" source code which was the disassembled version.
I've seen many very smart and clever uses of the C macros, the article is one of them. But isn't it time to move on?
Pedantically, the preprocessor is an entirely separate language. The lexing, parsing, expressions, and semantics are totally distinct. The preprocessor is usually implemented as a completely independent program. My first C compiler did integrate the preprocessor with the C compiler, but that was for performance reasons.
Currently, ImportC runs cpp and then lexes/parses the resulting C code for use in D.
In that example they also had replace the union with a struct - presumably to work around this issue. But that seems wasteful to me too. Doing it within an if(0) seems strictly better.
I'm curious what a hashmap looks like with this approach. It's one thing to pass through or hold onto a generic value, but another to perform operations on it. Think computing the hash value or comparing equality of generic keys in a generic hashmap.
I first would question what a user wants to do with a hashmap that uses polymorphic key-values of unknowable type at compile-time.
As a thought experiment, you could certainly have users define their own hash and equality functions and attach them to the table-entries themselves. On first thought, that sounds like it would be rife with memory safety issues.
At the end of the day, it is all just bytes. You could simply say that you will only key based on raw memory sequences.
what happens if two types have same size and alignment but different semantics : like `int` vs `float` or `struct { int a; }` vs `int`? does the type tag system catch accidental reuse . aka defending against structual collisions
Interesting! I’m working on toy/educational generator of ML-style tagged variants and associated functions in C (for a compiler) and when I’m a bit further along I will see if they’re compatible.
I think the idea of using a union to store the element type without any extra run-time memory cost might have some use, specifically in cases where the container struct wouldn't typically store a variable of the element type (or, more likely, a pointer to the element's type) but we want to slip that type information into the struct anyway.
However, the problem that I have with this idea as a general solution for generics is that it doesn't seem to solve any of the problems posed by the most similar alternative: just having a macro that defines a struct. The example shown in the article:
#define List(type) union { \
ListNode *head; \
type *payload; \
}
could just as easily be:
#define List(type) struct { \
type *head; \
/* Other data, such as node/element count... */ \
}
(As long as our nodes are maximally aligned - which they will be if they're dynamically allocated - it doesn't matter whether the pointer we store to the list head is ListNode *, type *, void *, or any other regular pointer type.)
The union approach has the same drawback as the struct approach: untagged unions are not compatible with each other, so we have to typedef the container in advance in order to pass in and out of functions (as noted in the article). This is broadly similar to the drawback from which the "generic headers" approach (which I usually call the "pseudo-template" approach) suffers, namely the need for boilerplate from the user. However, the generic-headers/pseudo-template approach is guaranteed to generate the most optimized code thanks to function specialization[1], and it can be combined with another technique to provide a non-type-prefixed API, as I discuss here[2] and demonstrate in practice here[3].
I'd also like to point to my own approach to generics[4] that is similar to the one described here in that it hides extra type information in the container handle's type - information that is later extracted by the API macros and passed into the relevant functions. My approach is different in that rather than exploiting unions, it exploits functions pointers' ability to hold multiple types (i.e. the return type and argument types) in one pointer. Because function pointers are "normal" C types, this approach doesn't suffer from the aforementioned typedef/boilerplate problem (and it allows for API macros that are agnostic to both element type/s and container type). However, the cost is that the code inside the library becomes rather complex, so I usually recommend the generic-headers/pseudo-template approach as the one that most people ought to take when implementing their own generic containers.
For your level 2 code, `uint64_t data[];` is wrong for types whose alignment is greater than `uint64_t`, and also wasteful for types whose alignment is smaller (for example, under an ilp32 ABI on 64-bit architectures).
For your level 3 code, it should be `int main() { List(Foo) foo_list = {NULL};`
Note that working around a lack of `typeof` means you can't return anything. Also, your particular workaround allows `const`ness errors since `==` is symmetrical.
You can't safely omit `payload` since you need it to know the correct size. Consider a `List(int64_t)` and you try to add an `int32_t` to it - this should be fine, but you can't `sizeof` the `int32_t`. Your code is actually lacking quite a bit to make this work.
=====
There are 2 major limitations to generics in C right now:
* Delegating to a vtable (internal or external) is limited in functionality, since structs cannot contain macros, only functions.
* Delegating to an external vtable (mandatory to avoid overhead) means that you have to forward-declare all of the types you'll ever use a vtable with. So far the best approach I've found is to declare (but not define) static functions in the same forwarding header I declare the typedefs in; note that GCC and Clang differ in what phase the "undefined static" warning appears in for the case where you don't actually include that particular type's header in a given TU.
(think about writing a function that accepts either `struct SizedBuffer {void *p; size_t len;};` or `struct BoundedBuffer {void *begin; void *end;};`, and also const versions thereof - all from different headers).
> Delegating to an external vtable (mandatory to avoid overhead) means that you have to forward-declare all of the types you'll ever use a vtable with.
We went down the rabbit hole of writing a compiler for this as part of a project I used to work on (Apache Clownfish[1], a subproject of the retired Apache Lucy project). We started off parsing .h files, but eventually it made sense to create our own small header language (.cfh "Clownfish Header" files).
Here's some generated code for invoking the CharBuf version of the "Clone" method defined in parent class "Obj":
Usage: We had our reasons for going to these extremes: the point of Clownfish was to provide a least-common-denominator object model for bindings to multiple dynamic languages (similar problem domain to SWIG), and the .cfh files also were used to derive types for the binding languages. But there was truly an absurd amount of boilerplate being generated to get around the issue you identify.This is why almost everybody just uses casts to void* for the invocant, skipping type safety.
[1] https://github.com/apache/lucy-clownfish
i am firmly of the opinion that compiling to c is a better route than doing clever c tricks to sort of get what you want. the compiler can be pretty minimal and as you note it pays for itself.
There’s some prior work called CFront. It implements a superset of C that’s just an “increment”. I think it’s worth looking into, it might take off one day!
> it should be `int main() { List(Foo) foo_list = {NULL};`
In C `int main()` means the function takes an unknown number of arguments. You need `int main(void)` to mean it doesn't take any arguments. This is a fact frequently forgotten by those who write C++.
That had been harmonized with C++ in C23 (e.g. func() is equivalent with func(void) now).
It's not really relevant for main() though, even in older C versions main() works fine and simply means "I don't need argc and argv".
This is about a function definition, not a random function declarator. C23 does not change anything in that case.
This is incorrect. In a function definition, an empty list means it takes no parameters. 6.7.5.3 Function declarators
> 14. An empty list in a function declarator that is part of a definition of that function specifies that the function has no parameters.
As you surely know if you're quoting the standard, it depends on which standard!
I believe that since C23 foo() is now a nullary function. As this is the last approved standard and it supersedes all previous standards, it is technically correct to say that de-jure this is what the (unqualified) C standard mandates.
Of course de-facto things are more nunanced.
C23 does not change anything in this situation, because we are talking about the definition of main(), not a forward declaration. More details here:
https://news.ycombinator.com/item?id=38729278#38732366
Quote a different standard.
I would love for `union`s to be federated, that is, a type could declare itself as thought it was part of a union with another type, without having to pre-declare all possible types in one place.
For layout-compatible types, you can often just include a `_base` member in each child. Maybe twice (once named and once unnamed) to avoid excess typing - I don't understand the common-initial-subsequence rule but people do this enough that compilers have to allow it.
This is also problematic, because there might be padding and the calculated size might be too small:
`malloc(sizeof(*node) + data_size);`
There's no a problem with the author's current code, since the padding is already included in the node size, but it would be a problem after doing alignment more intelligently.
> Structurally identical types will be considered the same type in GCC 15 and Clang later in 2025 thanks to a rule change
Beware that only tagged unions are considered the same type under the new rule, provided they have the same structrure and the same tag.
The List(T) macro should be changed to generate a different tag for each different T. Which is trivial (with ##) for simple one-word types, but impossible for even mildly complex ones like pointers to char (strings).
Of course you can force yourself to typedef any type before using it in a List, but it looses much of its versatility. Example:
Hi. I object.
The trick#0 you mention is how I made an entire C dialect. Here is a generic binary heap, for example https://github.com/gritzko/librdx/blob/master/abc/HEAPx.h The syntax is a bit heavyweight, but a huge huge advantage is: you get regular C structs in the end, very plain, very predictable, very optimizable. Compiler would eat them like donuts.
In the other cases, it is void* and runtime memory sizing and you have to define macros anyway.
Author here. Binary heaps and linked lists are different use cases. A binary heap must read the data you put in it to store it correctly, but a linked list doesn't. If I were writing a generic binary heap, maybe I would weigh my options differently. I mentioned this in the footnotes.
And that's why I like C++ templates
I agree, there are actually several reasons to prefer the header impl. Debugging is better, both because you can step through the header code where you can’t with a macro function, and because the type information available to the debugger is better. There are more opportunities for compiler optimizations because each instantiation is monomorphized and you don’t pay a runtime cost with variable sizing, generic structures can also be placed on the stack because of the fixed sizing.
There are workarounds for at least two of the problems the author mentions. Naming can be changed from Bar_func(args…) to func(Bar)(args…) with a function name macro that just does name mangling. You can avoid some of the binary bloat by using weak symbols, letting the linker deduplicate functions shared between translation units at link time.
There are other problems for generic containers of pointer types however, you can work around them by using a typedef or a type alias.
Intrusive data structures are more convenient in C still, but working with them in a debugger is a pain.
Author here. It's worth noting that no work is being done in the macros of my article, they compile down to a normal c function call which you can step through in a debugger.
There is little benefit in monomorphizing the implementation of a data structure like a linked list where its behavior doesn't depend on the contents of the data it contains (compared to, say, a max heap)
> Compiler would eat them like donuts.
Made me laugh out loud!
The casting of the function type assumes that the item pointer type (e.g. Foo*) has the same representation as void*, which the C standard doesn’t guarantee (in standardese: the two types aren’t “compatible”). Calling the function with the converted type therefore constitutes undefined behavior. It also impacts aliasing analysis by compilers (see [0], incidentally), even if the pointer representation happens to be the same.
This casting of the functions to different argument types constitutes the core of the type safety of the generic invocations; I’m not sure it can be fixed.
[0] https://news.ycombinator.com/item?id=44421185
This is addressed in the footnotes. casting is not the core of the type safety. Read the whole article.
Ah, that’s what I get for not reading the footnotes. However, the alternative solution presented evaluates the item argument twice, which is problematic as well (but could probably be worked around by passing `(list)->payload` on instead). Secondly, the assignment for type-checking doesn’t work for read-only operations on a const List, or does it? And doesn’t the assignment overwrite the head? Lastly, the do-while construction means you can’t use it for operations that return a value (without compiler extensions).
I also don’t agree it’s “squeamish” to be wary of aliasing analysis going wrong. It’s not a clean abstraction and can hide subtle bugs down the road.
Sure, but your alternative code incorrectly assigns to (list)->payload. You have many other options. Without typeof, you can if(0) the assignment, or check type compatibility with a ternary operator like 1 ? (item) : (list)->payload and pass that to _list_prepend, etc. With typeof, you can store item in a temporary variable with the same type as (list)->payload, or build a compound literal (typeof(*(list))){.payload=(item)}, etc.
The assignment is intentional. The union changed to a struct
Why would you jump through all these hoops instead of just writing C++ if you want "C with generics"
because i work on a legacy project that is coupled to safety regulations and other quality guarantees, and we cannot just simply roll out a solution ported to c++ on the next release, or even tenth, so perhaps we make it work until we can.
however we can set a standard and expectation for new projects to use c++, and we do and set an expectation to target a specific std.
i see this sentiment quite a lot on hackernews -- feels like a lot of people saying "git gud" -- i would expect a lot more nuance applied here.
Because for many of the use cases where C is used, switching to C++ involves jumping through even more hoops.
Do you have a couple of real world examples?
Any established C codebase, for example the kernel or Postgres?
Traditionally microcontroller firmwares as well, though those are increasingly friendly to C++, you just have to be careful about allocations as C++ makes it way easier to accidentally allocate than C does.
Nothing is stopping you from linking C++ code to Postgres.
I'm not sure about other compilers, but compiling C code as C++ with MSVC ends up with pretty much the exact same code, instruction by instruction. C++ is a bit more strict though especially with casting, so a lot of code won't compile out of the box.
C++ code compiles to a different function names in object file (name mangling). You probably need to put a lot of ugly `#ifdef __cplusplus extern "C" {` boilerplate in your headers, otherwise C and C++ files will not compile together.
Don't forget the infamous pattern used in some C projects too:
This pattern does not work in C++ as the nested declarations become temporaries.Embedded systems, for example.
I know it used to be, but is it really still common for embedded systems to use weird architectures that G++/Clang don't support?
Unless it is a popular system or common architecture, yes.
literally a good majority of existing embedded software coupled to applications in safety -- devices used by fire safety and first responders.
Writing extensions for projects that support C extensions but may not support C++ extensions, e.g. many dynamic languages.
You can still write the extension in C++ and expose an extern "C" interface.
That's possible, but then the people building your extension need a C++ toolchain.
The question was "please provide examples where switching to C++ involves jumping through even more hoops", and in my view requiring downstream to use a C++ environment when they're expecting to use a C environment qualifies.
True. For me, C++ itself is the maze of hoops I would rather want to avoid.
Why would you write C++ if you can get the same result by jumping through a few hoops with C?
Templates in C++ require language support - you can't simply implement them with "a few hoops" in C.
There's also the method used in the Linux kernel to embed the list information (struct list_head) within the type specific struct. https://kernelnewbies.org/FAQ/LinkedLists
The naming of LIST_HEAD_INIT and INIT_LIST_HEAD is confusing to me.
The way I remember it is:
INIT_LIST_HEAD is of form VERB_NOUN so is called from within a function to programatically initialise the list.
LIST_HEAD_INIT is NOUN_VERB and is used within a structure initialiser not from a function.
But my main point was to show the "embed the list in the data" approach rather than "embed the data in the list" or "point to the data from the list" and not to discuss the naming details in the kernel implementation of the concept.
It is cool trick. I already use in my experimental library though ;-) https://github.com/uecker/noplate/blob/main/src/list.h
I guess if anyone might know it might be you—do you see any way of doing this for intrusive data structures, embedding the node struct in the data (and as side effect supporting an object to be on multiple containers) rather than the data in the node like you're doing there?
You could put the dummy member into the embedded node. But for intrusive data structures you often want them to erase the type so that you write generic algorithms as regular functions. In this case, it makes more sense to have a run-time check do to down casts. I do this with my variadic type which has an intrusive "super" member: https://godbolt.org/z/ofdKe7Pfv The overhead is often completely removed by the compiler.
Mhm. Putting the dummy member into the embedded node doesn't give a path from the proper object to find the embedded node "mid-struct". run-time checks are the "easy way out". We/I'll stick to macro soup probably, so we have compile-time checks.
btw. For ISO C WG14… has anyone suggested adding _Include to the preprocessor, along the lines of _Pragma? It'd really help with doing this kind of really long macros, hiding the clunky "#define, #define, #define, #include" inside a macro…
I don't think anybody has proposed this, at least not recently. There is a proposal for multi-line macros: https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3531.txt
That one might actually be better, nice to know & thanks for the pointer! (I hope it actually goes somewhere…)
When I saw the title I assumed it was originally "Why I" or "How I" and was trimmed automatically by HN, but this is the original. Could it be that the author was influenced by HN's title guidelines and titled it thus?
Interesting idea with the union and using typeof(). We (I) went with large macros defining wrappers instead, which, I believe, is a necessity with intrusive data structures, or at least I don't immediately see how to do that with unions & typeof. Maybe it's possible...?
e.g. hash table wrapper: https://github.com/FRRouting/frr/blob/master/lib/typesafe.h#...
(cf. https://docs.frrouting.org/projects/dev-guide/en/latest/list...)
The key idea here seems to be to use function pointer's type to enforce type safety rather than using the data "handle" type (that is often found in implementations inspired by Sean Barrett's strechy_buffers).
> One annoying thing about C is that it does not consider these two variables to have the same type
C23 solves that too: https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3037.pdf
Supported by latest GCC and Clang, but not by MSVC.
Author here. Not quite. The key idea is about using a union to associate type information with a generic data type. Type casting a function is not the only way to use that type information. I discuss that as well as the C23 changes in the footnotes and the "typeof on old compilers" section.
FWIW, as far back as 2015 my feature check library documents Visual Studio as supporting "__typeof".[1] Note the leading but not trailing underscores. Perhaps I was mistaken, but I usually tested that sort of thing. It's also possible __typeof had slightly different semantics.
[1] See https://github.com/wahern/autoguess/blob/b44556e4/config.h.g... (that's the 2015 revision, but HEAD has the same code).
msvc 19.39 is the first to support it, which I mention in the article. You can confirm it didn't work up through 19.38 in godbolt [1]. I don't use Visual Studio, so I don't know what version of that first started using msvc 19.39
[1] https://godbolt.org/z/M7zPYdssP
This is gonna haunt me.
Digging through some old code of mine (circa 2009) I found this bit:
So somehow I had the impression Visual Studio .NET 2003 (7.1)[1] added __typeof. I'm still holding out hope someone will come to my rescue and reply that once upon a time MSVC had __typeof, but removed it. But for now it seems past me is gaslighting present me.[1] See https://learn.microsoft.com/en-us/cpp/overview/compiler-vers... for mapping between Visual Studio versions and _MSC_VER.
EDIT: Ah ha! It seems Microsoft did support __typeof, but perhaps only for "managed" C++ (aka C++ .NET)?
> One thing to watch out for when using the __typeof operator in managed C++ is that __typeof(wchar_t) can return different values depending on the compilation options.
Source: https://learn.microsoft.com/en-us/archive/msdn-magazine/2002... (See also https://learn.microsoft.com/en-us/archive/msdn-magazine/2005...)
Here's how to do it in D:
Why suffer the C preprocessor? Using preprocessor macros is like using a hammer for finish carpentry, rather than a nail gun. A nail gun is 10x faster, drives the nail perfectly every time, and no half moon dents in your work.Thanks, this post is about C.
On some projects you must use C.
If I may may be provocative :-) this post isn't about C. It's about layering on a custom language using C preprocessor macros.
My compilers were originally written in C. I started using the C preprocessor to do metaprogramming. After some years I got fed up with it and removed nearly all of the preprocessor use, and never looked back. My code was much easier to understand.
An amusing story: long ago, a friend of mine working for Microsoft was told by a team leader that a 50K program had a bug in it, and sadly the developer was long gone. He'd assigned programmer after programmer to it, who could not fix it. My friend said he'd give it a try, and had it fixed in 2 hours.
The source code was written in Microsoft MASM, where the M stood for "Macro". You can guess where this is going. The developer had invented his own high level language using the macro system (which was much more powerful than C's). Unfortunately, he neglected to document it, and the other programmers spent weeks studying it and could not figure it out.
The leader, astonished, asked him how he figured it out in 2 hours? My friend said simple. He assembled it to object code, then disassembled the object code with obj2asm (a disassembler I wrote that converts object code back to source code). He then immediately found and fixed the bug, and checked in the "new" source code which was the disassembled version.
I've seen many very smart and clever uses of the C macros, the article is one of them. But isn't it time to move on?
If the C compiler accepts it, it is C.
Pedantically, the preprocessor is an entirely separate language. The lexing, parsing, expressions, and semantics are totally distinct. The preprocessor is usually implemented as a completely independent program. My first C compiler did integrate the preprocessor with the C compiler, but that was for performance reasons.
Currently, ImportC runs cpp and then lexes/parses the resulting C code for use in D.
It is part of the C standard. Whether it is part of a separate binary is an implementation choice.
True on both counts. But they are still separate and distinct languages.
There is no one "the C compiler".
Pragmatically, the only C compiler that matters for what is or is not C is the one you are using.
Only if you are lucky enough to only use one compiler, or only one version of the same one.
The “typeof on old compilers” section contains the code:
That is not a no-op. That is overwriting the list head with your (item). Did you mean to wrap it in an `if(0)`?In that example they also had replace the union with a struct - presumably to work around this issue. But that seems wasteful to me too. Doing it within an if(0) seems strictly better.
I'm curious what a hashmap looks like with this approach. It's one thing to pass through or hold onto a generic value, but another to perform operations on it. Think computing the hash value or comparing equality of generic keys in a generic hashmap.
I first would question what a user wants to do with a hashmap that uses polymorphic key-values of unknowable type at compile-time.
As a thought experiment, you could certainly have users define their own hash and equality functions and attach them to the table-entries themselves. On first thought, that sounds like it would be rife with memory safety issues.
At the end of the day, it is all just bytes. You could simply say that you will only key based on raw memory sequences.
what happens if two types have same size and alignment but different semantics : like `int` vs `float` or `struct { int a; }` vs `int`? does the type tag system catch accidental reuse . aka defending against structual collisions
uint64_t data[] in level 2 violates the strict aliasing rule. Use the char type instead to avoid the violation.
Interesting! I’m working on toy/educational generator of ML-style tagged variants and associated functions in C (for a compiler) and when I’m a bit further along I will see if they’re compatible.
Another way is to not try to write generic data structures. When you tailor them to the use case you can simplify.
The #1 data structure in any program is array.
When all you have are arrays, everything looks like a problem you solve with arrays.
There are quite a few problems that specialised containers are suited for, that's why they were created.
And you can write them when you need them.
The situation where you need a red black tree with 10 different key/value combos isn’t real.
You could take away anything you use and say "but we could make it ourselves", that doesn't mean it's helpful.
Or write in CFront and have it translated to C
And where are you going to get a cfront compiler these days?
pretty sure C is the new Go.
pretty sure C has to go
Without the concurreny part.
OpenMP to the rescue
Or garbage collection. Or interfaces. Or packages. Or actual generics.
I think the idea of using a union to store the element type without any extra run-time memory cost might have some use, specifically in cases where the container struct wouldn't typically store a variable of the element type (or, more likely, a pointer to the element's type) but we want to slip that type information into the struct anyway.
However, the problem that I have with this idea as a general solution for generics is that it doesn't seem to solve any of the problems posed by the most similar alternative: just having a macro that defines a struct. The example shown in the article:
could just as easily be: (As long as our nodes are maximally aligned - which they will be if they're dynamically allocated - it doesn't matter whether the pointer we store to the list head is ListNode *, type *, void *, or any other regular pointer type.)The union approach has the same drawback as the struct approach: untagged unions are not compatible with each other, so we have to typedef the container in advance in order to pass in and out of functions (as noted in the article). This is broadly similar to the drawback from which the "generic headers" approach (which I usually call the "pseudo-template" approach) suffers, namely the need for boilerplate from the user. However, the generic-headers/pseudo-template approach is guaranteed to generate the most optimized code thanks to function specialization[1], and it can be combined with another technique to provide a non-type-prefixed API, as I discuss here[2] and demonstrate in practice here[3].
I'd also like to point to my own approach to generics[4] that is similar to the one described here in that it hides extra type information in the container handle's type - information that is later extracted by the API macros and passed into the relevant functions. My approach is different in that rather than exploiting unions, it exploits functions pointers' ability to hold multiple types (i.e. the return type and argument types) in one pointer. Because function pointers are "normal" C types, this approach doesn't suffer from the aforementioned typedef/boilerplate problem (and it allows for API macros that are agnostic to both element type/s and container type). However, the cost is that the code inside the library becomes rather complex, so I usually recommend the generic-headers/pseudo-template approach as the one that most people ought to take when implementing their own generic containers.
[1] https://gist.github.com/attractivechaos/6815764c213f38802227...
[2] https://github.com/JacksonAllan/CC/blob/main/articles/Better...
[3] https://github.com/JacksonAllan/Verstable
[4] https://github.com/JacksonAllan/CC
Dude the ship has sailed.
Not sure what you mean by that, but if you're trying to imply that C is not relevant: https://www.tiobe.com/tiobe-index/
Plus an article about C was at the top of hacker news all day today.
[dead]
[flagged]
[flagged]