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":
typedef cfish_CharBuf*
(*CFISH_CharBuf_Clone_t)(cfish_CharBuf* self);
extern uint32_t CFISH_CharBuf_Clone_OFFSET;
static inline cfish_CharBuf*
CFISH_CharBuf_Clone(cfish_CharBuf* self) {
const CFISH_CharBuf_Clone_t method
= (CFISH_CharBuf_Clone_t)cfish_obj_method(
self,
CFISH_CharBuf_Clone_OFFSET
);
return method(self);
}
cfish_CharBuf *charbuf = cfish_CharBuf_new();
cfish_CharBuf *clone = CFISH_CharBuf_Clone(charbuf);
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.
> 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 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.
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 canât 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 donât 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(argsâŚ) to func(Bar)(argsâŚ) 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 doesnât guarantee (in standardese: the two types arenât âcompatibleâ). 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; Iâm not sure it can be fixed.
This is addressed in the footnotes. casting is not the core of the type safety. Read the whole article.
Ah, thatâs 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 doesnât work for read-only operations on a const List, or does it? And doesnât the assignment overwrite the head? Lastly, the do-while construction means you canât use it for operations that return a value (without compiler extensions).
I also donât agree itâs âsqueamishâ to be wary of aliasing analysis going wrong. Itâs 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
Here's how to do it in D:
struct ListNode(T) {
ListNode* next;
T data;
}
T!int node;
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.
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.
Embedded systems, for example.
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.
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.
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...)
Get a daily email with the the top stories from Hacker News. No spam, unsubscribe at any time.