> It’s no wonder GCC is trying to add -ftrampoline-impl=heap to the story of GNU Nested Functions; they might be able to tighten up that performance and make it more competitive with Apple Blocks.
[disclaimer] Without brushing up on the details of this, I strongly suspect that this is about removing the need for executable stacks than performance. Allocating a trampoline on the stack rather than heap is good for efficiency.
These days, many GNU/Linux distros are disabling executable stacks by default in their toolchain configuration, both for building the distro and for the toolchain offered by the system to the user.
When you use GCC local functions, it overrides the linker behavior so that the executable is marked for executable stacks.
Of course, that is a security concession because when your stack is executable, that enables malicious remote execution code to work that relies on injecting code into the stack via a buffer overflow and tricking the process into jumping to it.
If trampolines can be allocated in a heap, then you don't need an executable stack. You do need an executable heap, or an executable dedicated heap for these allocations. (Trampolines are all the same size, so they could be packed into an array.)
Programs which indirect upon GCC local functions are not aware of the trampolines. The trampolines are deallocated naturally when the stack rolls back on function return or longjmp, or a C++ exception passing through.
Heap-allocated trampolines have an obvious deallocation problem; it would be interesting to see what strategy is used for that.
Those routines seem like they are called in a nested fashion, LIFO. Like __gcc_nested_func_ptr_deleted knows that it's destroying the most recently created trampoline.
It probably has some logic for dealing with different threads, and also for dealing with stack frames abandoned by longjmp or a longjmp-like mechanism.
The benchmark demonstrates that the modern C++ "Lambda" approach (creating a unique struct with fields for captured variables) is effectively a compile-time calculated static link. Because the compiler sees the entire definition, it can flatten the "link" into direct member access, which is why it wins. The performance penalty the author sees in GCC is partly due to the OS/CPU overhead of managing executable stacks, not just code inefficiency. The author correctly identifies that C is missing a primitive that low-level languages perfected decades ago: the bound method (wide) pointer.
The most striking surprise is the magnitude of the gap between std::function and std::function_ref. It turns out std::function (the owning container) forces a "copy-by-value" semantics deeply into the recursion. In the "Man-or-Boy" test, this apparently causes an exponential explosion of copying the closure state at every recursive step. std::function_ref (the non-owning view) avoids this entirely.
Even if you never copy the std::function the overhead is very large. GCC (14 at least) does not seem to be able to elide the allocation, nor inline the function itself, even if used immediately after use and the object never escapes the function. Given the opportunity, GCC seems to be able to completely remove one layer pf function_ref, but fails at two layers.
This is exactly right, and the "Man-or-Boy" benchmark hits the worst-case scenario for libstdc++ specifically. The optimization fails here. My "copy-by-value" comment refers to the ownership semantics. Since std::function owns its storage, and the Man-or-Boy recursion passes the closure into the next layer (often by value or by capturing it into a new closure), we trigger the copy constructor. If the SBO limit is exceeded, that copy constructor performs a new heap allocation and a deep copy of the state.
GCC (libstdc++) as all other major C++ runtimes (libc++, MSVC) implements the small object optimization for std::function where a small enough callable is stored directly in std::function's state instead of on the heap. Across these implementations, you can reply on being able to capture two pointers without a dynamic allocation.
You would think so, but it actually doesn't. last time I checked, libstdc++ could only optimize std::bind closures. A trivial test with a stateless lambda shows this is still the case in GCC14 and 15. In fact I can't even seem to trigger the library optimization with bind.
Differently from GCC14, GCC15 itself does seem to be able to optimize the allocation (and the whole std::function) in trivial cases though (independently of what the library does).
This was very interesting, and it's obvious from the majority of the text that the author knows a lot about these languages, their implementation, benchmarking corners, and so on. Really!
Therefore it's very jarring with this text after the first C code example:
This uses a static variable to have it persist between both the compare function calls that qsort makes and the main call which (potentially) changes its value to be 1 instead of 0
This feels completely made up, and/or some confusion about things that I would expect an author of a piece like this to really know.
In reality, in this usage (at the global outermost scope level) `static` has nothing to do with persistence. All it does is make the variable "private" to the translation unit (C parliance, read as "C source code file"). The value will "persist" since the global outermost scope can't go out of scope while the program is running.
It's different when used inside a function, then it makes the value persist between invocations, in practice typically by moving the variable from the stack to the "global data" which is generally heap-allocated as the program loads. Note that C does not mention the existence of a stack for local variables, but of course that is the typical implementation on modern systems.
I'm finding myself in a weird position now, because I disagree with a whole lot of things in the blog post (well, the parts I was willing to read anyways), but calling that variable static for the sake of persistence was correct.
The fact that you are questioning the use of the term shows that you are not familiar with the ISO C standard. What the author alludes to is static storage duration. And whether or not you use the "static" keyword in that declaration (also definition), the storage duration of the object remains "static". People mostly call those things "global variables", but the proper standardese is "static storage duration". In that sense, the author was right to use "static" for the lifetime of the object.
EDIT: if you drop "static" from that declaration, what changes is the linkage of the identifier (from internal to external).
>This uses a static variable to have it persist between both the compare function calls that qsort makes and the main call which (potentially) changes its value to be 1 instead of 0
The only misleading thing here is that ‘static’ is monospaced in the article (this can’t be seen on HN). Other than that, ‘static variable’ can plausibly refer to an object with a static storage duration, which is what the C standard would call it.
>moving the variable from the stack to the "global data" which is generally heap-allocated as the program loads
It is not heap-allocated because you can’t free() it. Non-zero static data is not even anonymously mapped, it is file-backed with copy-on-write.
I had a completely different response reading the sentence. I've been programming in C for 20+ years and am very familiar with exactly the problem the author is discussing. When they referred to a "static variable", I understood immediately that they meant a file static variable private to the translation unit. Didn't feel contrived or made up to me at all; just a reflection of the author's expertise. Precision of language.
I mean, you're closer to the committee than I am, but while that is true in a literal sense, I'd assume that you all would not let someone who knew how to edit LaTeX but not know anything about C hold that position.
Assuming we have some choice. Not many people volunteer their time to this work, which is quite a lot and not much fun. Companies also do not invest a lot of resources into C.
It took me a second read to realise that the mention of static is a red herring. I think the author knows that the linkage is irrelevant for the rest of the explanation; it just happens to be static so they called it static. But by drawing attention to it, it does first read like they're confused about the role of static there.
Oh! Thanks, I was not being as concrete as I imagined. Sorry.
Yes, the `static` can simply be dropped, it does no additional work for a single-file snippet like this.
I tried diving into Compiler Explorer to examine this, and it actually produces slightly different code for the with/without `static` cases, but it was confusing to deeply understand quickly enough to use the output here. Sorry.
I see exactly the same assembly from x86-64 GCC 15.2 with -O2 the first example in the article both as is and without `static`, which makes sense. The two do differ if you add -fPIC, as though you’re compiling a dynamic library, and do not add -fvisibility=hidden at the same time, but that’s because Linux dynamic linking is badly designed.
Something I had thought of (which does not fully solve the problems mentioned there, but would allow GNU nested functions to work in a way that can be implemented without trampolines and executable stacks, so that it can work in standard C and with the standard ABI), is to allow a nested function to optionally be defined with the "static" and/or "register" keywords.
With "static", it is implemented as an ordinary function, but the name is local to the function that contains it; it cannot access stuff within the function containing it unless those things are also declared as "static".
With "register", the address of the function cannot be taken, and if the function accesses other stuff within the function that contains it then the compiler will add additional arguments to the function so that its type does not necessarily match the type which is specified in the program.
This is not good enough for many uses though, so having the other extensions would also be helpful (possibly including implementing Apple Blocks in GCC).
Good to see Borland's __closure extension got a mention.
Something I've been thinking about lately is having a "state" keyword for declaring variables in a "stateful" function. This works just like "static" except instead of having a single global instance of each variable the variables are added to an automatically defined struct, whose type is available using "statetype(foo)" or some other mechanism, then you can invoke foo as with an instance of the state (in C this would be an explicit first parameter also marked with the "state" parameter.) Stateful functions are colored in the sense that if you invoke a nested stateful function its state gets added to the caller's state. This probably won't fly with separate compilation though.
Yes, though it was a remarkably brief mention. I believe Borland tried to standardise it back in 2002 or so,* along with properties. (I was the C++Builder PM, but a decade and a half after that attempt.)
C++Builder’s entire UI system is built around __closure and it is remarkably efficient: effectively, a very neat fat pointer of object instance and method.
That sounds cool, but this quickly gets complicated. Some aspects that need to be addressed:
- where does the automatically defined struct live? Data segment might work for static, but doesn't allow dynamic use. Stack will be garbage if closure outlives function context (ie. callback, future). Heap might work, but how do you prevent leaks without C++/Rust RAII?
- while a function pointer may be copied or moved, the state area probably cannot. It may contain pointers to stack object or point into itself (think Rust's pinning)
Would this be similar to how Rust handles async? The compiler creates a state machine representing every await point and in-scope variables at that point. Resuming the function passes that state machine into another function that matches on the state and continues the async function, returning either another state or a final value.
It could make some sense to use strchr, because in idiomatic UNIX tools, single character command line options can be clustered. But that also means that subsequent code should not be tested for a specific position.
And if you ever find yourself actually doing command line parsing, use getopt(). It handles all the corner cases reliably, and consistent with other tools.
Your code actually has 2 bugs. The first I assume is just a typo and you meant to use [1][1] == ‘r’. The second one is that you would accept “-rblah” as well.
Not to mention the potential signed integer overflow in (*right - *left) and (*left - *right), which is undefined behavior. And even if you rely on common two's complement wraparound, the result may be wrong; for example, (INT_MAX-(-1)) should mathematically yield a positive value, but the function will produce INT_MIN, which is negative.
And then we have this "modern" way of spelling pointers, "const int* right" (note the space). In C, declaration syntax mirrors use, so it should be "const int *right", because "*right" is a "const int".
You are right, the implementation of the "compare" function should have used only comparison operators, not subtraction, because unlike subtraction the comparison operations are not affected by overflow (the hardware implementation of integer comparison handles overflow automatically).
When there is no overflow, the sign of the subtraction result provides the same information as a comparison operator, but this is no longer true when overflow happens.
Using only arithmetic operators would require explicit checks for overflow and this would be a very inefficient implementation, because the comparison operators handle overflow implicitly without any overhead (because the comparison operators do not use the sign of the subtraction result to decide their truth value, but they compute the true sign of the result, by the modulo 2 sum, a.k.a. XOR, of the result sign with the integer overflow flag; this is done in hardware, with no overhead).
The feature that integer comparison must be correct regardless of overflow has been a requirement in CPU design since the earliest times. Very few CPUs, the most notable examples being Intel 8080 and RISC-V, lacked this feature. The competitors and successors of Intel 8080, i.e. Motorola MC6800 and Zilog Z80 have added it, making them much more similar to the previously existing CPUs than Intel 8080 was. The fact that even microprocessors implemented with a few thousand transistors around 1975 had this feature emphasizes how weird is that RISC-V lacks such a feature in 2025, after 50 years and being implemented with a number of transistors many orders of magnitude greater.
Sorry, but this is the kind of ridiculous reply that the RISC-V fans give when they are asked why their ISA lacks many of the features that any decent ISA has and which have a negligible implementation cost, therefore no reason to be missing.
The workaround suggested by the RISC-V documentation consists in replacing a very large fraction of all instructions of a program (because there are a lot of integer additions, subtractions and comparisons in any program, close to a half of all instructions) with 3 or more instructions, in order to approximate what in any other CPU is done with single instructions.
The other ridiculous workaround proposed to save RISC-V is that any high-performance implementation must supplant its missing features by instruction fusion.
Yes, the missing hardware for overflow detection can be replaced by multiplying the number of instructions for any operation and the missing addressing modes can be synthesized by instruction fusion, but such implementation solutions are extraordinarily more expensive than the normal solutions used for 3 quarters of century in the other computers, since they were made with vacuum tubes.
Because of the extreme overhead of checking for overflow, I bet that most programs compiled for RISC-V do not check for overflow, which is not acceptable in reliable programs (even when using C/C++, I always compile them with overflow checking enabled, which should have been the default option, to be disabled only in specific cases where it can be proven that overflow is impossible and the checks reduce the performance).
The cost of overflow checks turns out to be largely about missed optimizations due to heavier constraints wrt. how the program should behave if overflow occurs (e.g. preserving partial results). Having an overflow check instruction in the ISA just doesn't matter all that much, it can even hurt in bignum computation (often cited as a favorable case for overflow checks) by introducing unwanted insn dependencies.
What you say about missed optimizations is true only when the compiler attempts to handle itself in a graceful way the cases when overflows would occur, instead of raising exceptions.
This is not what is normal overflow checking. Normal overflow checking just raises a specific exception when integer overflow happens.
This has absolutely no effect upon compiler optimizations. The compiler always generates code ignoring the possibility of exceptions. When exceptions happen, the control is passed far away to the exception handler, which decides what to do, e.g. to save debugging information and abort partially or totally the offending program, because an overflow is normally the consequence of an unforeseen program bug and it is impossible to do any action that will allow the continuation of the execution.
You should remember that there is nothing special about integer overflow, almost every instruction that the compiler generates can raise an exception at run time. Any branch instruction, any memory-access instruction, any floating-point instruction, any vector instruction can raise an exception due to hardware. In recent CPUs, integer overflow is not raised implicitly, so you have to insert a conditional branch, but this is irrelevant.
If your theory that the possibility of raising exceptions can influence compiler optimizations were true, there would exist no compiler optimizations, because from every 10 or so instructions generated by a compiler at least a half can raise various kinds of exceptions, in a manner completely unpredictable by the compiler. Adding integer overflow exceptions changes nothing.
From the introduction, your paper seems like a counterproposal: support closures, just not the way others propose. But the paper seems to accept that closures / nested functions, supported at the language level directly, are a "good thing" for C specifically. I disagree with that. When and how has it become the consensus?
At the moment there is no consensus for anything related to this. But most people agree that closures are a good thing because they make certain reoccurring programming patterns safer and easier to write. Why do you disagree?
> There's no way to spell out this function's type, and no way to store it anywhere. This is true of regular functions too!
well regular functions decay to function pointers. You could have the moral equivalent of std::function_ref (or similarly, borland __closure) in C of course and have closures decay to it.
The only reason that GCC needs executable trampolines is for the program to be able to create an ordinary function pointer and have all the captured data come along with it. The proposal is to reuse the syntax of nested functions, but change the semantics so that they are no longer callable via ordinary function pointers, but rather "fat pointers" that reference the captured data alongside the raw function address. This is similar to the method used by C++ and does not need trampolines.
It is possible to transform the pure Rosetta form of GNU nested function similar to the pure C, such that it doesn't need any stack trampoline. I wonder if that would be closer in performance to the pure C form.
Honestly all I would want from C closures is a better syntax for doing exactly that. Whatever it ends up being (if it goes in), it better have some way to interop with "legacy" function pointer + user context APIs. (Including some way to tell the compiler where the user context pointer goes on the closure's arguments.) Otherwise it's just completely useless.
That's all there is to it. I don't understand the whole obsession with closures.
I've used lambdas extensively in modern C++. I hate them with a passion.
I've also used OCaml. An awesome language where this stuff is super natural and beautiful.
I don't understand why people want to shoehorn functional programming into C. C++ was terrible already, and is now worse for it.
> we’re going to be focusing on and looking specifically at Closures in C and C++, since this is going to be about trying to work with and – eventually – standardize something for ISO C that works for everyone.
It does seem needlessly complex. I think a better idea is to just have a type that is a pair of pointer-sized words. That pattern crops up again and again - context pointer and function pointer, array and its size, memory allocation and effective size, etc. The problem with having both pieces in separate variables is that it is very easy to lose track of what is where. If you have it in a single bundle it is a lot simpler to use. The exact design needs a lot more consideration for sure, because I would like something simpler than writing anonymous structs everywhere (which I can already do), but at the same time flexible enough for most use cases.
Thread locals do solve the problem. You create a wrapper around the original function. You set a global thread local user data, you pass in a function which calls the function pointer accepting the user data with the global one.
Thread locals don't fully solve the problem. They work well if you immediately call the closure, but what if you want to store the closure and call it later?
Because we create `reverse_sort` between creating `normal_sort` and calling it, we end up with a reverse sort despite clearly asking for a normal sort.
Yep. Thread locals are probably faster than the other solutions shown too.
It’s confusing to me that thread locals are “not the best idea outside small snippets” meanwhile the top solution is templating on recursion depth with a constexpr limit of 11.
The method of having static variables to store state in functions is used heavily in ANSI C book. It’s honestly a beautiful technique when used prudently.
Imagine a comparison function that needs to call sort() as part of its implementation. You could argue that's probably a bad idea, but it would be a problem for this case.
(You could solve that with a manually maintained stack for the context in a thread local, but you'd have to do that case-by-case)
Anyway, the larger point is that a re-entrant general solution is desirable. The sort example might be a bit misguided, because who calls sort-inside-sort[0]? Nobody, realistically, but these types of issues are prevalent in the "how to do closures" area... and In C every API does it slightly differently, even if they're even aware of the issues.
[0] Because there's no community that likes nitpicking like the C (or C++) community. I considered preempting that objection :). C++ has solved this, so there's that.
Stewart Lynch in his 10x VODs mentions his custom Function abstraction in C++. It's super clean and explicit, avoiding `auto` requirement of C++ lambdas. It's use looks something akin to:
// imagine my_function takes 3 ints, the first 2 args are captured and curried.
Function<void(int)> my_closure(&my_function, 1, 2);
my_closure(3);
I've never implemented it myself, as I don't use C++ features all too much, but as a pet project I'd like to someday. I wonder how something like that compares!
I'm thinking of using C++ for a personal project specifically for the lambdas and RAII.
I have a case where I need to create a static templated lambda to be passed to C as a pointer. Such thing is impossible in Rust, which I considered at first.
Yeah, Rust closures that capture data are fat pointers { fn*, data* }, so you need an awkward dance to make them thin pointers for C.
let mut state = 1;
let mut fat_closure = || state += 1;
let (fnptr, userdata) = make_trampoline(&mut &mut fat_closure);
unsafe {
fnptr(userdata);
}
assert_eq!(state, 2);
use std::ffi::c_void;
fn make_trampoline<C: FnMut()>(closure: &mut &mut C) -> (unsafe fn(*mut c_void), *mut c_void) {
let fnptr = |userdata: *mut c_void| {
let closure: *mut &mut C = userdata.cast();
(unsafe { &mut *closure })()
};
(fnptr, closure as *mut _ as *mut c_void)
}
It requires a userdata arg for the C function, since there's no allocation or executable-stack magic to give a unique function pointer to each data instance. OTOH it's zero-cost. The generic make_trampoline inlines code of the closure, so there's no extra indirection.
> Rust closures that capture data are fat pointers { fn, data }
This isn’t fully accurate. In your example, `&mut C` actually has the same layout as usize. It’s not a fat pointer. `C` is a concrete type and essentially just an anonymous struct with FnMut implemented for it.
You’re probably thinking of `&mut dyn FnMut` which is a fat pointer that pairs a pointer to the data with a pointer to a VTable.
So in your specific example, the double indirection is unnecessary.
This is a problem for all capturing closures though, not just Rust's. A pure fn-ptr arg can't have state, and if there's no user data arg then there's no way to make a trampoline. If C++ was calling a C API with the same constraint it would have the same problem.
Well, capturing closures that are implemented like C++ lambdas or Rust closures anyway. The executable stack crimes do make a thin fn-ptr with state.
If Rust has a stable ABI on where the data* is in the function arguments (presumably first?), you don't need to do anything if it matches the C code's expected function signature including the user context arg.
Unfortunately a lot of existing C APIs won't have the user arg in the place you need it, it's a mix of first, last, and sometimes even middle.
Yes but my problem wasn’t with the user data pointer but the fact that I needed a STATIC generic lambda. Static because the C library then forks and continues to call the lambda in the new process but I also type based conversions in it.
I feel the results say more about the testing methodology and inlining settings than anything else.
Practically speaking all lambda options except for the one involving allocation (why would you even do that) are equivalent modulo inlining.
In particular, the caveat with the type erasure/helper variants is precisely that it prevents inlining, but given everything is in the same translation unit and isn't runtime-driven, it's still possible for the compiler to devirtualize.
I think it would be more interesting to make measurements when controlling explicitly whether inlining happens or the function type can be deduced statically.
Given a Sufficiently Good™ compiler, yes, after devirtualization and heap elision all variants should generate exactly the same code. In practice is more complicated. Devirtualization needs to runs after (potentially interprocedural) constant propagation, which might be too late to take advantage of other optimization opportunities, unless the compiler keeps rerunning the optimization pipeline.
In a simple test I see that GCC14 has no problems completely removing the overhead of std::function_ref, but plain std::function is a huge mess.
Eventually we will get there [1], but in the meantime I prefer not to rely on devirtualization, and heap elision is more of a party trick.
edit: to compare early vs late inlining: while gcc 14 can remove one layer of function_ref, it seems that it cannot remove two layers, as apparently doesn't rerun the required passes to take advantage of the new opportunity. It has no problem of course removing an arbitrary large (but finite) layers of plain lambdas.
edit2: GCC15 can remove trivial uses of std::function, but this is very fragile. It still can't remove two function_ref.
[1] for example 25 years ago compilers were terrible at removing abstraction overhead of the STL, today there is very little cost.
The breakdown of lambda, blocks, and nested functions demonstrates how important implementation and ABI details are in addition to syntax. I think the standard for C should include a straightforward, first class wide function pointer along with a closure story to stop people from adding these half portable, half spooky extensions.
> It’s no wonder GCC is trying to add -ftrampoline-impl=heap to the story of GNU Nested Functions; they might be able to tighten up that performance and make it more competitive with Apple Blocks.
[disclaimer] Without brushing up on the details of this, I strongly suspect that this is about removing the need for executable stacks than performance. Allocating a trampoline on the stack rather than heap is good for efficiency.
These days, many GNU/Linux distros are disabling executable stacks by default in their toolchain configuration, both for building the distro and for the toolchain offered by the system to the user.
When you use GCC local functions, it overrides the linker behavior so that the executable is marked for executable stacks.
Of course, that is a security concession because when your stack is executable, that enables malicious remote execution code to work that relies on injecting code into the stack via a buffer overflow and tricking the process into jumping to it.
If trampolines can be allocated in a heap, then you don't need an executable stack. You do need an executable heap, or an executable dedicated heap for these allocations. (Trampolines are all the same size, so they could be packed into an array.)
Programs which indirect upon GCC local functions are not aware of the trampolines. The trampolines are deallocated naturally when the stack rolls back on function return or longjmp, or a C++ exception passing through.
Heap-allocated trampolines have an obvious deallocation problem; it would be interesting to see what strategy is used for that.
> Heap-allocated trampolines have an obvious deallocation problem; it would be interesting to see what strategy is used for that.
With -ftrampoline-impl=heap, GCC automatically insert[1] pairs of constructor/destructor routines from libgcc which were built around mmap/munmap.
[1] https://godbolt.org/z/7s5nooMPz
Those routines seem like they are called in a nested fashion, LIFO. Like __gcc_nested_func_ptr_deleted knows that it's destroying the most recently created trampoline.
It probably has some logic for dealing with different threads, and also for dealing with stack frames abandoned by longjmp or a longjmp-like mechanism.
The benchmark demonstrates that the modern C++ "Lambda" approach (creating a unique struct with fields for captured variables) is effectively a compile-time calculated static link. Because the compiler sees the entire definition, it can flatten the "link" into direct member access, which is why it wins. The performance penalty the author sees in GCC is partly due to the OS/CPU overhead of managing executable stacks, not just code inefficiency. The author correctly identifies that C is missing a primitive that low-level languages perfected decades ago: the bound method (wide) pointer.
The most striking surprise is the magnitude of the gap between std::function and std::function_ref. It turns out std::function (the owning container) forces a "copy-by-value" semantics deeply into the recursion. In the "Man-or-Boy" test, this apparently causes an exponential explosion of copying the closure state at every recursive step. std::function_ref (the non-owning view) avoids this entirely.
Even if you never copy the std::function the overhead is very large. GCC (14 at least) does not seem to be able to elide the allocation, nor inline the function itself, even if used immediately after use and the object never escapes the function. Given the opportunity, GCC seems to be able to completely remove one layer pf function_ref, but fails at two layers.
This is exactly right, and the "Man-or-Boy" benchmark hits the worst-case scenario for libstdc++ specifically. The optimization fails here. My "copy-by-value" comment refers to the ownership semantics. Since std::function owns its storage, and the Man-or-Boy recursion passes the closure into the next layer (often by value or by capturing it into a new closure), we trigger the copy constructor. If the SBO limit is exceeded, that copy constructor performs a new heap allocation and a deep copy of the state.
GCC (libstdc++) as all other major C++ runtimes (libc++, MSVC) implements the small object optimization for std::function where a small enough callable is stored directly in std::function's state instead of on the heap. Across these implementations, you can reply on being able to capture two pointers without a dynamic allocation.
You would think so, but it actually doesn't. last time I checked, libstdc++ could only optimize std::bind closures. A trivial test with a stateless lambda shows this is still the case in GCC14 and 15. In fact I can't even seem to trigger the library optimization with bind.
Differently from GCC14, GCC15 itself does seem to be able to optimize the allocation (and the whole std::function) in trivial cases though (independently of what the library does).
This was very interesting, and it's obvious from the majority of the text that the author knows a lot about these languages, their implementation, benchmarking corners, and so on. Really!
Therefore it's very jarring with this text after the first C code example:
This uses a static variable to have it persist between both the compare function calls that qsort makes and the main call which (potentially) changes its value to be 1 instead of 0
This feels completely made up, and/or some confusion about things that I would expect an author of a piece like this to really know.
In reality, in this usage (at the global outermost scope level) `static` has nothing to do with persistence. All it does is make the variable "private" to the translation unit (C parliance, read as "C source code file"). The value will "persist" since the global outermost scope can't go out of scope while the program is running.
It's different when used inside a function, then it makes the value persist between invocations, in practice typically by moving the variable from the stack to the "global data" which is generally heap-allocated as the program loads. Note that C does not mention the existence of a stack for local variables, but of course that is the typical implementation on modern systems.
I'm finding myself in a weird position now, because I disagree with a whole lot of things in the blog post (well, the parts I was willing to read anyways), but calling that variable static for the sake of persistence was correct.
The fact that you are questioning the use of the term shows that you are not familiar with the ISO C standard. What the author alludes to is static storage duration. And whether or not you use the "static" keyword in that declaration (also definition), the storage duration of the object remains "static". People mostly call those things "global variables", but the proper standardese is "static storage duration". In that sense, the author was right to use "static" for the lifetime of the object.
EDIT: if you drop "static" from that declaration, what changes is the linkage of the identifier (from internal to external).
>This uses a static variable to have it persist between both the compare function calls that qsort makes and the main call which (potentially) changes its value to be 1 instead of 0
The only misleading thing here is that ‘static’ is monospaced in the article (this can’t be seen on HN). Other than that, ‘static variable’ can plausibly refer to an object with a static storage duration, which is what the C standard would call it.
>moving the variable from the stack to the "global data" which is generally heap-allocated as the program loads
It is not heap-allocated because you can’t free() it. Non-zero static data is not even anonymously mapped, it is file-backed with copy-on-write.
I had a completely different response reading the sentence. I've been programming in C for 20+ years and am very familiar with exactly the problem the author is discussing. When they referred to a "static variable", I understood immediately that they meant a file static variable private to the translation unit. Didn't feel contrived or made up to me at all; just a reflection of the author's expertise. Precision of language.
The author contributes to ISO C and ISO C++ working groups, and his latest contribution was #embed.
Not just that, the author is the Project Editor for WG14.
This doesn’t mean that it’s impossible to make mistakes, but still.
It means he can edit LaTeX. Of course, JeanHeyd is very qualified, but being project editor for an ISO standard does not require this.
I mean, you're closer to the committee than I am, but while that is true in a literal sense, I'd assume that you all would not let someone who knew how to edit LaTeX but not know anything about C hold that position.
Assuming we have some choice. Not many people volunteer their time to this work, which is quite a lot and not much fun. Companies also do not invest a lot of resources into C.
It took me a second read to realise that the mention of static is a red herring. I think the author knows that the linkage is irrelevant for the rest of the explanation; it just happens to be static so they called it static. But by drawing attention to it, it does first read like they're confused about the role of static there.
That's a very weird comment, your spreading your knowledge and not really addresse what could have been changed in the article.
If I follow your comment, you mean that he could have use a non-static global variable instead and avoid mentioning "static" keyword afterward?
Oh! Thanks, I was not being as concrete as I imagined. Sorry.
Yes, the `static` can simply be dropped, it does no additional work for a single-file snippet like this.
I tried diving into Compiler Explorer to examine this, and it actually produces slightly different code for the with/without `static` cases, but it was confusing to deeply understand quickly enough to use the output here. Sorry.
I see exactly the same assembly from x86-64 GCC 15.2 with -O2 the first example in the article both as is and without `static`, which makes sense. The two do differ if you add -fPIC, as though you’re compiling a dynamic library, and do not add -fvisibility=hidden at the same time, but that’s because Linux dynamic linking is badly designed.
TU-level concepts (mostly) dissolve during the linking stage. You need to compile with -c to generate an object file in order to see the distinction.
Also, the difference manifests in the symbols table, not the assembly.
To clarify, I was talking about Compiler Explorer-cleaned disassembly, same as the comment I was replying to.
Something I had thought of (which does not fully solve the problems mentioned there, but would allow GNU nested functions to work in a way that can be implemented without trampolines and executable stacks, so that it can work in standard C and with the standard ABI), is to allow a nested function to optionally be defined with the "static" and/or "register" keywords.
With "static", it is implemented as an ordinary function, but the name is local to the function that contains it; it cannot access stuff within the function containing it unless those things are also declared as "static".
With "register", the address of the function cannot be taken, and if the function accesses other stuff within the function that contains it then the compiler will add additional arguments to the function so that its type does not necessarily match the type which is specified in the program.
This is not good enough for many uses though, so having the other extensions would also be helpful (possibly including implementing Apple Blocks in GCC).
Good to see Borland's __closure extension got a mention.
Something I've been thinking about lately is having a "state" keyword for declaring variables in a "stateful" function. This works just like "static" except instead of having a single global instance of each variable the variables are added to an automatically defined struct, whose type is available using "statetype(foo)" or some other mechanism, then you can invoke foo as with an instance of the state (in C this would be an explicit first parameter also marked with the "state" parameter.) Stateful functions are colored in the sense that if you invoke a nested stateful function its state gets added to the caller's state. This probably won't fly with separate compilation though.
Yes, though it was a remarkably brief mention. I believe Borland tried to standardise it back in 2002 or so,* along with properties. (I was the C++Builder PM, but a decade and a half after that attempt.)
C++Builder’s entire UI system is built around __closure and it is remarkably efficient: effectively, a very neat fat pointer of object instance and method.
[*] Edit: two dates on the paper, but “bound pointer to member” and they note the connection to events too: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2002/n13...
> a "state" keyword for declaring variables in a "stateful" function
Raku (née Perl 6) has this! https://docs.raku.org/language/variables#The_state_declarato...
That sounds cool, but this quickly gets complicated. Some aspects that need to be addressed:
- where does the automatically defined struct live? Data segment might work for static, but doesn't allow dynamic use. Stack will be garbage if closure outlives function context (ie. callback, future). Heap might work, but how do you prevent leaks without C++/Rust RAII?
- while a function pointer may be copied or moved, the state area probably cannot. It may contain pointers to stack object or point into itself (think Rust's pinning)
- you already mention recursion, compilation
- ...
IMO the C way is to allow users to explicitly manage context area, along the lines of posix ucontext.h or how the author's closure proposal handle closure allocation[1]. [1] https://thephd.dev/_vendor/future_cxx/papers/C%20-%20Functio...
I dreamed up a similar idea[1] upon reading the author's closure proposal, it's also really close to async coroutines.
[1] https://github.com/ThePhD/future_cxx/issues/55#issuecomment-...
Would this be similar to how Rust handles async? The compiler creates a state machine representing every await point and in-scope variables at that point. Resuming the function passes that state machine into another function that matches on the state and continues the async function, returning either another state or a final value.
Long time ago I wrote C. Could anyone fill me in why the first code snippet is arg parsing the way it is?
int main(int argc, char* argv[]) {
}Why not
if (argc > 1 && strcmp(argv[1], "-r") == 0) {
}for example?
It doesn't even make sense to use strchr for determining the position of 'r', when the code checks that the position of '-' is at index 0.
Your solution is perfectly fine. Even if you don't have access to strchr for some reason, the original snippet is really convoluted.
You could just write (strlen(argv[1]) > 1 && argv[1][0] == '-' && argv[1][0] == 'r') if you really want to.
It could make some sense to use strchr, because in idiomatic UNIX tools, single character command line options can be clustered. But that also means that subsequent code should not be tested for a specific position.
And if you ever find yourself actually doing command line parsing, use getopt(). It handles all the corner cases reliably, and consistent with other tools.
Of course, `&&` in C is short-circuiting so it's safe without the `strlen()` too, as long as the argument is there i.e. not NULL.
Also, the use of a convoluted `if` to conditionally assign a literal boolean is a code smell (to me), I would drop the `if` and just use:
if a more forward-thinking/strict check is not needed.Your code actually has 2 bugs. The first I assume is just a typo and you meant to use [1][1] == ‘r’. The second one is that you would accept “-rblah” as well.
I suspect it was adopted from a bigger snippet that had support for parsing things like "-abc" as "-a -b -c", etc.
Not to mention the potential signed integer overflow in (*right - *left) and (*left - *right), which is undefined behavior. And even if you rely on common two's complement wraparound, the result may be wrong; for example, (INT_MAX-(-1)) should mathematically yield a positive value, but the function will produce INT_MIN, which is negative.
And then we have this "modern" way of spelling pointers, "const int* right" (note the space). In C, declaration syntax mirrors use, so it should be "const int *right", because "*right" is a "const int".
I feel too old for this shit. :(
You are right, the implementation of the "compare" function should have used only comparison operators, not subtraction, because unlike subtraction the comparison operations are not affected by overflow (the hardware implementation of integer comparison handles overflow automatically).
When there is no overflow, the sign of the subtraction result provides the same information as a comparison operator, but this is no longer true when overflow happens.
Using only arithmetic operators would require explicit checks for overflow and this would be a very inefficient implementation, because the comparison operators handle overflow implicitly without any overhead (because the comparison operators do not use the sign of the subtraction result to decide their truth value, but they compute the true sign of the result, by the modulo 2 sum, a.k.a. XOR, of the result sign with the integer overflow flag; this is done in hardware, with no overhead).
The feature that integer comparison must be correct regardless of overflow has been a requirement in CPU design since the earliest times. Very few CPUs, the most notable examples being Intel 8080 and RISC-V, lacked this feature. The competitors and successors of Intel 8080, i.e. Motorola MC6800 and Zilog Z80 have added it, making them much more similar to the previously existing CPUs than Intel 8080 was. The fact that even microprocessors implemented with a few thousand transistors around 1975 had this feature emphasizes how weird is that RISC-V lacks such a feature in 2025, after 50 years and being implemented with a number of transistors many orders of magnitude greater.
But RISC-V has it?
Sorry, but this is the kind of ridiculous reply that the RISC-V fans give when they are asked why their ISA lacks many of the features that any decent ISA has and which have a negligible implementation cost, therefore no reason to be missing.
The workaround suggested by the RISC-V documentation consists in replacing a very large fraction of all instructions of a program (because there are a lot of integer additions, subtractions and comparisons in any program, close to a half of all instructions) with 3 or more instructions, in order to approximate what in any other CPU is done with single instructions.
The other ridiculous workaround proposed to save RISC-V is that any high-performance implementation must supplant its missing features by instruction fusion.
Yes, the missing hardware for overflow detection can be replaced by multiplying the number of instructions for any operation and the missing addressing modes can be synthesized by instruction fusion, but such implementation solutions are extraordinarily more expensive than the normal solutions used for 3 quarters of century in the other computers, since they were made with vacuum tubes.
Because of the extreme overhead of checking for overflow, I bet that most programs compiled for RISC-V do not check for overflow, which is not acceptable in reliable programs (even when using C/C++, I always compile them with overflow checking enabled, which should have been the default option, to be disabled only in specific cases where it can be proven that overflow is impossible and the checks reduce the performance).
The cost of overflow checks turns out to be largely about missed optimizations due to heavier constraints wrt. how the program should behave if overflow occurs (e.g. preserving partial results). Having an overflow check instruction in the ISA just doesn't matter all that much, it can even hurt in bignum computation (often cited as a favorable case for overflow checks) by introducing unwanted insn dependencies.
What you say about missed optimizations is true only when the compiler attempts to handle itself in a graceful way the cases when overflows would occur, instead of raising exceptions.
This is not what is normal overflow checking. Normal overflow checking just raises a specific exception when integer overflow happens.
This has absolutely no effect upon compiler optimizations. The compiler always generates code ignoring the possibility of exceptions. When exceptions happen, the control is passed far away to the exception handler, which decides what to do, e.g. to save debugging information and abort partially or totally the offending program, because an overflow is normally the consequence of an unforeseen program bug and it is impossible to do any action that will allow the continuation of the execution.
You should remember that there is nothing special about integer overflow, almost every instruction that the compiler generates can raise an exception at run time. Any branch instruction, any memory-access instruction, any floating-point instruction, any vector instruction can raise an exception due to hardware. In recent CPUs, integer overflow is not raised implicitly, so you have to insert a conditional branch, but this is irrelevant.
If your theory that the possibility of raising exceptions can influence compiler optimizations were true, there would exist no compiler optimizations, because from every 10 or so instructions generated by a compiler at least a half can raise various kinds of exceptions, in a manner completely unpredictable by the compiler. Adding integer overflow exceptions changes nothing.
BTW: I wrote why the lambda design does not fit C well here:
https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3654.pdf
(and I am not impressed by micro benchmarks)
From the introduction, your paper seems like a counterproposal: support closures, just not the way others propose. But the paper seems to accept that closures / nested functions, supported at the language level directly, are a "good thing" for C specifically. I disagree with that. When and how has it become the consensus?
At the moment there is no consensus for anything related to this. But most people agree that closures are a good thing because they make certain reoccurring programming patterns safer and easier to write. Why do you disagree?
I think local functions (like the GNU extension) that behave like C++ byref(&) capturing lambdas makes the most sense for C.
You can call the local functions directly and get the benefits of the specialized code.
There's no way to spell out this function's type, and no way to store it anywhere. This is true of regular functions too!
To pass it around you need to use the type-erased "fat pointer" version.
I don't see how anything else makes sense for C.
> There's no way to spell out this function's type, and no way to store it anywhere. This is true of regular functions too!
well regular functions decay to function pointers. You could have the moral equivalent of std::function_ref (or similarly, borland __closure) in C of course and have closures decay to it.
The price you pay for GCC nested (local) functions is an executable stack with 'trampolines'.
I'm a fan of nested functions but don't think the executable stack hack is worth it, and using a 'display' is a better solution.
See the Dragon Book or Compiler Construction: Principles and Practice (1984) by Louden
You misunderstood my comment. GNU local function syntax, C++ [&] lambda behavior (i.e., a hidden struct).
I really did, my comment is specific to C.
The only reason that GCC needs executable trampolines is for the program to be able to create an ordinary function pointer and have all the captured data come along with it. The proposal is to reuse the syntax of nested functions, but change the semantics so that they are no longer callable via ordinary function pointers, but rather "fat pointers" that reference the captured data alongside the raw function address. This is similar to the method used by C++ and does not need trampolines.
Meaning something that generates code similar to what I have in this comment?
https://news.ycombinator.com/item?id=46243298
Yeah something like that, though built-in to the compiler.
It is possible to transform the pure Rosetta form of GNU nested function similar to the pure C, such that it doesn't need any stack trampoline. I wonder if that would be closer in performance to the pure C form.
(I can't be bothered to run his benchmarks)
Defininig a callback interface in C without a user context parameter is a capital crime.
Honestly all I would want from C closures is a better syntax for doing exactly that. Whatever it ends up being (if it goes in), it better have some way to interop with "legacy" function pointer + user context APIs. (Including some way to tell the compiler where the user context pointer goes on the closure's arguments.) Otherwise it's just completely useless.
That's all there is to it. I don't understand the whole obsession with closures.
I've used lambdas extensively in modern C++. I hate them with a passion.
I've also used OCaml. An awesome language where this stuff is super natural and beautiful.
I don't understand why people want to shoehorn functional programming into C. C++ was terrible already, and is now worse for it.
> we’re going to be focusing on and looking specifically at Closures in C and C++, since this is going to be about trying to work with and – eventually – standardize something for ISO C that works for everyone.
Sigh. My heart sinks.
It does seem needlessly complex. I think a better idea is to just have a type that is a pair of pointer-sized words. That pattern crops up again and again - context pointer and function pointer, array and its size, memory allocation and effective size, etc. The problem with having both pieces in separate variables is that it is very easy to lose track of what is where. If you have it in a single bundle it is a lot simpler to use. The exact design needs a lot more consideration for sure, because I would like something simpler than writing anonymous structs everywhere (which I can already do), but at the same time flexible enough for most use cases.
> If you have it in a single bundle it is a lot simpler to use.
I've often seen that; I call this curiously recurring clumping pattern "encapsulation". Maybe if I blogged about it, it would catch on.
betterC is all you need
https://dlang.org/spec/betterc.html
Thread locals do solve the problem. You create a wrapper around the original function. You set a global thread local user data, you pass in a function which calls the function pointer accepting the user data with the global one.
Thread locals don't fully solve the problem. They work well if you immediately call the closure, but what if you want to store the closure and call it later?
Because we create `reverse_sort` between creating `normal_sort` and calling it, we end up with a reverse sort despite clearly asking for a normal sort.The answer is you wrap it and you don’t return until the thing stored in the thread local is not needed
So we can't return the closure?
Then it's clearly only half a solution.
The example I gave above should work fine in any language with first-class closures.
> The closure problem can be neatly described by as “how do I get extra data to use within this qsort call?”
you also need to restore the _qsort2_closure when done. But again you are reinventing dynamic scoping with all its advantages and disadvantages.
> you also need to restore the _qsort2_closure when done
No I do not. It will reassigned next call.
> But again you are reinventing dynamic scoping
No. I’m not reinventing anything. I’m using the existing feature of thread local variables.
The usage of such is entirely an implementation detail of qsort2 with the exception of recursion.
Dynamic scoping typically refers to defining variables which have scope outside of their call stack. No usage of this API requires it.
Can you just try to learn something new?
This is the classic lexical vs dynamic scoping. Dynamic scoping works great until it doesn't.
Don’t use C then? It sounds like you want JavaScript, Python, or Lisp.
Once again, the caller of the API does not declare any variables so there is no dynamic scoping.
Yep. Thread locals are probably faster than the other solutions shown too.
It’s confusing to me that thread locals are “not the best idea outside small snippets” meanwhile the top solution is templating on recursion depth with a constexpr limit of 11.
The method of having static variables to store state in functions is used heavily in ANSI C book. It’s honestly a beautiful technique when used prudently.
reentrancy.
It doesn’t store state for later. It’s literally impossible to tell it’s happening.
Imagine a comparison function that needs to call sort() as part of its implementation. You could argue that's probably a bad idea, but it would be a problem for this case.
(You could solve that with a manually maintained stack for the context in a thread local, but you'd have to do that case-by-case)
That is true. It can be protected against with assert.
I think the times you need to do this are few. And this version is much more pruden.
Assert what, exactly?
Anyway, the larger point is that a re-entrant general solution is desirable. The sort example might be a bit misguided, because who calls sort-inside-sort[0]? Nobody, realistically, but these types of issues are prevalent in the "how to do closures" area... and In C every API does it slightly differently, even if they're even aware of the issues.
[0] Because there's no community that likes nitpicking like the C (or C++) community. I considered preempting that objection :). C++ has solved this, so there's that.
Stewart Lynch in his 10x VODs mentions his custom Function abstraction in C++. It's super clean and explicit, avoiding `auto` requirement of C++ lambdas. It's use looks something akin to:
I've never implemented it myself, as I don't use C++ features all too much, but as a pet project I'd like to someday. I wonder how something like that compares!Isn't this basically the same as passing the function to std::bind_front and storing it in a std::function or std::function_ref?
I'm thinking of using C++ for a personal project specifically for the lambdas and RAII.
I have a case where I need to create a static templated lambda to be passed to C as a pointer. Such thing is impossible in Rust, which I considered at first.
Yeah, Rust closures that capture data are fat pointers { fn*, data* }, so you need an awkward dance to make them thin pointers for C.
It requires a userdata arg for the C function, since there's no allocation or executable-stack magic to give a unique function pointer to each data instance. OTOH it's zero-cost. The generic make_trampoline inlines code of the closure, so there's no extra indirection.> Rust closures that capture data are fat pointers { fn, data }
This isn’t fully accurate. In your example, `&mut C` actually has the same layout as usize. It’s not a fat pointer. `C` is a concrete type and essentially just an anonymous struct with FnMut implemented for it.
You’re probably thinking of `&mut dyn FnMut` which is a fat pointer that pairs a pointer to the data with a pointer to a VTable.
So in your specific example, the double indirection is unnecessary.
The following passes miri: https://play.rust-lang.org/?version=nightly&mode=debug&editi...
(did this on mobile, so please excuse any messiness).
This is a problem for all capturing closures though, not just Rust's. A pure fn-ptr arg can't have state, and if there's no user data arg then there's no way to make a trampoline. If C++ was calling a C API with the same constraint it would have the same problem.
Well, capturing closures that are implemented like C++ lambdas or Rust closures anyway. The executable stack crimes do make a thin fn-ptr with state.
If Rust has a stable ABI on where the data* is in the function arguments (presumably first?), you don't need to do anything if it matches the C code's expected function signature including the user context arg.
Unfortunately a lot of existing C APIs won't have the user arg in the place you need it, it's a mix of first, last, and sometimes even middle.
I know about this technique but it uses too much unsafe for my taste. Not that it's bad or anything, just a personal preference.
It can be done in 100% safe code as far as Rust is concerned (if you use `dyn Fn` type instead of c_void).
The only unsafe here is to demonstrate it works with C/C++ FFI (where void* userdata is actually not type safe)
Yes but my problem wasn’t with the user data pointer but the fact that I needed a STATIC generic lambda. Static because the C library then forks and continues to call the lambda in the new process but I also type based conversions in it.
In Rust, could you instead use a templated struct wrapping a function pointer along with #[repr(C)]?
I feel the results say more about the testing methodology and inlining settings than anything else.
Practically speaking all lambda options except for the one involving allocation (why would you even do that) are equivalent modulo inlining.
In particular, the caveat with the type erasure/helper variants is precisely that it prevents inlining, but given everything is in the same translation unit and isn't runtime-driven, it's still possible for the compiler to devirtualize.
I think it would be more interesting to make measurements when controlling explicitly whether inlining happens or the function type can be deduced statically.
Given a Sufficiently Good™ compiler, yes, after devirtualization and heap elision all variants should generate exactly the same code. In practice is more complicated. Devirtualization needs to runs after (potentially interprocedural) constant propagation, which might be too late to take advantage of other optimization opportunities, unless the compiler keeps rerunning the optimization pipeline.
In a simple test I see that GCC14 has no problems completely removing the overhead of std::function_ref, but plain std::function is a huge mess.
Eventually we will get there [1], but in the meantime I prefer not to rely on devirtualization, and heap elision is more of a party trick.
edit: to compare early vs late inlining: while gcc 14 can remove one layer of function_ref, it seems that it cannot remove two layers, as apparently doesn't rerun the required passes to take advantage of the new opportunity. It has no problem of course removing an arbitrary large (but finite) layers of plain lambdas.
edit2: GCC15 can remove trivial uses of std::function, but this is very fragile. It still can't remove two function_ref.
[1] for example 25 years ago compilers were terrible at removing abstraction overhead of the STL, today there is very little cost.
You can just write the benchmark in such a way that the optimizations are not possible.
I actually enjoy trampoline functions in C a bit and it's one of the GNU extensions I use sometimes.
It's a post about Man or Boy... and the only typo is... the word _son_. Pretty sure it's supposed to be "on"
c++ for the win!! finally!!
The breakdown of lambda, blocks, and nested functions demonstrates how important implementation and ABI details are in addition to syntax. I think the standard for C should include a straightforward, first class wide function pointer along with a closure story to stop people from adding these half portable, half spooky extensions.
Exactly.
This.
i wish JS gurus understood this before jumping all in on hooks and bloating the runtime footprint of every web app out there