Tail-call optimization in C is relatively recent (2025)
lwn.net
lwn.net
That's me.
The motivation back then was to allow compilers that target C to assume that tail calls will be "proper". That's different from an optimization, which is usually optional, and which compilers don't guarantee.
The LWN post briefly sketches why this is hard: C allows variable-argument functions (like printf) where only the caller knows for sure how many arguments it passed, which means that only the caller can clean up the stack, unless the stack frame size is also communicated, which "normal" C calling conventions don't do. But when the callee does a proper tail call, the stack frame that returns to the callee is not the stack frame that the callee originally sent. This is explained in more detail in my thesis starting on page 16: https://hostr.flingit.run/s/proper-tail-calls.pdf
If you know that you are an M-parameter function being called, and you want to tail cal an N-parameter function, where N <= M, then you can just place the new N parameters in the same space on the stack where you received your M parameters, and jump to that function. That function will return to your original caller, which will remove the M parameters, not caring that some of them are not the originals that it passed.
Suppose N > M. Things start to get tricky. There isn't space in our original argument space for N. If we increase the space, the original caller won't clean it up properly. If we just allocate a new space of N, we are not making a tail call.
Because we want to make a tail call, it means we don't expect to execute any code in this function any more, and are free to trash the local variables. We can move the stack down a bit to make room for N arguments above where previously we were given M by our caller. To solve the problem that our caller wants to clean up M, but we need it to clean up N could be solved by a trampoline. We prime the stack such that when the tail-called function we are targeting returns, it will not go to our caller directly but to a stub function. That stub function will clean up the N-M words of the stack, leaving M, and then return to the original caller, which cleans up M.
In this situation, we are benefiting from knowing that the caller passed M to us. In the case of a variadic function, we don't know at all. It could just be the fixed arguments (parameters before the ellipsis) like printf("hello\n'), or any number. There is a run-time protocol to discover what parameters there are; the application logic figures it out from the arbitrary conventions. That's too late and too ad hoc for compile time.
I think yuo can reason about it similarly to above. If we are a variadic with M fixed parameters, we know we are called with at least M arguments, so we can place N <= M tail-callee arguments into the variadic space and proceed accordingly. For N > M, we can extend to make up the difference and use the trampoline to clean up and return to the original caller.
sThese trampolines are not closures; they are behind-the-scenes that can be generated as static code; no executable heaps or stacks required.
Maybe I misunderstand your idea?
The trampoline would replace the {main retaddr, main fp} closure with {trampoline addr, [stack byte count to pop], main retaddr, main fp} and would pop some number of bytes, either hard-coded into the the trampoline function (so you get a bunch of them) or is part of the closure.
Do it right and there should only ever be one trampoline closure on the stack for any chain of tail calls, making the scheme O(1) in space.
So the whole protocol is that when the compiler recognizes that a call is a tail call, and one that can be turned into a jump instead of call, then the compiler must emit code to
a) pop the current continuation (which will either be the original or a trampoline that embeds the original, and from which the original can be recovered),
b) pop all the previous arguments and push all the new ones (possibly some are the same, so there is room for optimization here),
and
c) push a new continuation that is either the same as the previous current continuation or else a new continuation closure that is the correct trampoline corresponding to -and embedding- the original continuation, where the original is recovered from (a).
The trampoline recovers the original continuation, fixes the stack depth to what the caller expects, and executes a return to the original continuation.
For variadic functions it has to be the case that they have used `va_start()`, consumed all variadic arguments with `va_arg()`, then called `va_end()`, leaving no active copy of the `va_list`, then the compiler can arrange to keep a hidden local variable count of stack words used by the variadic arguments that it can use to implement the above protocol correctly.
I think the framing of TCO as an optimization has been very unfortunate.
But yes, framing TCO as an optimization is unfortunate.
As well as an effort to get it standardized: https://isocpp.org/files/papers/D3939R0.html (in C++).
I can only think of a few other optimizations that affect memory usage. Register spilling (arguably not really an optimization but a necessity), Rust's niche filling for enum discriminants and C++'s std::vec<bool> (a language-level optimization, arguably a different thing entirely).
I often think about how few memory optimizations we have. The reason is most likely that they tend to be non-local so are much harder to apply than CPU optimizations that generally have no effect outside of the function they are in.
Depends on the semantics of the programming language itself. For some languages, it is truly an optimization, for some, it is required, and does meaningfully change observed semantics.
It doesn't matter if it's local since it's a VM, it's doing it at runtime and can change an entire call stack of non local code for an optimization.
Which exactly in this subject varies a lot between implementations, on how well escape analysis is done, if there is a JIT cache between JVM executions, or AOT compilation.
Additionally Valhalla is finally getting added to the language with a new EA made available last week, thus value classes will add yet another way to have stack values.
Value classes are not a compiler optimization, and I wasn't talking about the standard, I was talking about an implementation.
It’s that it makes memory usage bounded when it’s on, and unbounded when it’s off.
In languages that have guaranteed tail call eliminations, the semantics of tail recursion is the same as that of a loop. So you can express the same iterative algorithm without using iterative code.
If std::vector<bool> was an optimisation we couldn't write C++ which blows up because it's actually a bitset, it would be semantically transparent - but that's easy to do even by accident because it's not transparent at all.
In fact the existing std::vector<bool> should just be named std::growable_bitset or something and then std::vector<bool> would make what you actually wanted like Rust's Vec<bool> does.
But also, I wouldn't rely on a compiler to remove empty 'while (true)' loops.
Both languages continue to have examples of slightly more complicated loops that can be assumed to terminate in the absence of side effects, but `while(true)` isn't one of those any longer.
If you used an in-place sort algorithm and observed memory requirements that scale super-linearly with the size of the input, you would think the semantics of the program were changed.
In languages with such tail call guarantees, tail recursion _is_ a loop. It semantically encodes constant space complexity.
Having said that the standards give way too much leeway for the limits, so a conforming implementation might have arbitrary limits for loops as well (at least in C++, I'm not that familiar with the C standard's wording).
Which means that "whether this naively-recursive code is actually recursive in practice" is a semantic difference, in that there is an error/failure-mode (stack overflow) that can be statically guaranteed to not happen (at least for a given compilation target) if TCO gets applied; but which cannot be guaranteed to not happen without TCO applied.
---
Tangent: you could of course try to write code defensively, to guarantee that a stack overflow won't occur, by bounding recursion separately (e.g. via a passed-and-decremented recursion-limit parameter), so that in the non-TCO case, you get a software exception thrown (which you'd hopefully then handle... somehow), rather than triggering a stack overflow.
And for many more-traditional recursive algorithms, this works!
But doing so for the types of algorithms that are "canonically" expressed in terms of tail-calls (even in a non-tail-call-idiomatic language like C), almost always requires poking holes in the C abstract machine to see through to the micro-architectural details underneath.
You can't just use something like a recursion-limit parameter as a general solution for these algorithms, as TCO is used in things like continuation-passing or threaded-code VM implementations — i.e. things that look less like visiting trees and more like visiting unboundedly-non-terminal infinite-state-machine states ["infinite" because the states are dynamic function pointers to JITted code, and more of them can appear at runtime.]
You need to not track the "number of invocations deep" you are into the algorithm, but rather, how big the stack actually is at the moment. Which means you need to actually do math on addresses of the stack base pointer vs either the stack pointer, or the address of a local stack-allocated variable. There's no version of that that doesn't require writing non-portable inline assembly.
Java has string interning. I think that’s a hack that shouldn’t exist in an ideal world. Reason is that, as a library writer, you cannot make the call whether to intern strings (requiring more instructions for string access, thus slowing down code, but decreasing memory usage, and, because of that, possibly speeding up the code again) or not.
Wait, why would interned immutable strings require more instructions when doing regular string access? You can still point to the start of a zero-terminated C-string, it just requires storing extra metadata like lenght and a string hash somewhere. Which can be done at the negative indices of said pointer.
Or do you refer to the extra rolling-hash pass needed when concatenating two strings to verify if it would result in an already-interned one? Because yes, that's one extra rolling hast pass over the appended string the first time a string is constructed, but after that doing so again likely saves memory and construction time, because any concatenation that would result in an already interned string would avoid actual memory allocation and copying of the string's characters.
Plus string comparisons become cheap O(1) pointer comparisons this way, which is really nice in many use-cases.
And that's not even considering more advanced tricks like interning short strings in the 64-bit word of the pointer to the string itself, relying on the fact that modern memory allocators never return an address with the lsb set, so it can be used to flag it as such[0].
Java automatically interns static strings (e.g. from class files), but does not automatically intern dynamically-allocated strings, e.g. new String(charArray)
If you want it interned, you have to intentionally call e.g. new String(...).intern(). If you do this on every string you work with, you can then reliably use reference equality instead of value equality, e.g. given char[] abc = {'a','b','c'}; then new String(abc) != new String(abc) != "abc" but new String(abc).intern() == new String(abc).intern() == "abc"
But if you're interning every string, you're doing extra work to maintain that string pool, and adding extra pressure on the GC, and potentially you'll be re-interning strings a lot depending on how many times they end up no longer referenced by the time GC runs.
Just to clarify where I started from: JS has the notion of String Primitives and String Objects[0], and concatenating two string primitives results in another string primitive. It's as if it always calls .intern(). Except, since from a user perspective there is no option to not intern it, JS implementations are free to do optimizations that don't seem available to Java. Like avoiding heap allocation for provably short-lived strings in the first place, and not interning them if they are never compared (I don't actually know if it does those optimizations, but I imagine it shouldn't be too hard to prove these things with many short functions).
[0] https://developer.mozilla.org/en-US/docs/Web/JavaScript/Refe...
Sorry, I wasn’t precise. Accessing them won’t take more instructions, but setting them up does.
> Or do you refer to the extra rolling-hash pass needed when concatenating two strings to verify if it would result in an already-interned one?
I don’t think the JVM does that.
it is guaranteed in Scheme, and it affects the semantics of programs in a completely positive way.
Much of computer science is "pure" and "abstract" like mathematics. However, programmers are still taught to use loops to calculate factorial rather than recursion in order to avoid stack overflow. In Scheme you can use recursion without flinching. That is a semantic difference.
It does when you use them as a feature and not an optimization. Like in interpreters, state machines, parsers, etc.
Calling tail calls an optimization set computer science back 40 years.
There is a "tail" prefix in the intermediate language (IL) bytecode that F# uses but Roslyn, the C# compiler, never emits.
So unlike F#, whether the same algorithm written in C# becomes a loop depends on JIT behaviour. This means if you're coming to a function cold in C# you can overflow the stack, while if you enter the same function fresh after it's been warmed up, it may have been optimised away by RyuJIT and if so you are able to call it safely for what would be large numbers of recursions.
> More than 20 programming tools vendors offer some 26 programming languages — including C++, Perl, Python, Java, COBOL, RPG and Haskell — on .NET.
https://news.microsoft.com/source/2001/10/22/massive-industr...
MSIL thus had to support all of them.
Those differences between language semantics is why CLS was also a thing back then.
https://learn.microsoft.com/en-us/dotnet/standard/language-i...
https://gcc.gnu.org/onlinedocs/gcc-15.1.0/gcc/Statement-Attr...
Of course it's nowhere near a guarantee that it would land in the stable language, but it's still an opportunity to try it in practice.
On the other hand, I am pretty new to the compiler space myself, and I count early 2000s as a pretty long time ago, though again it is not that far back considering how long other language implementations had tail calls like in ML or variants since 1980-90s.
You won't find anything on ISO/IEC 9899:2024 about tail calls, like it happens on Scheme.
https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3220.pdf
Section 3.5 of R7RS.
1. https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3886.pdf
2. https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3928.pdf
On such an implementation, the feature is available but useless.
For variadic functions, if you use `va_start()`/`va_arg()`/`va_end()` to consume all the arguments, and leave no `va_list` alive, then the compiler can correctly generate a tail call from such functions.
This leads to fun stack-overflow bugs too in a lot of js code (one solution is to flatten: https://joshua.hu/javascript-infinite-tail-call-recursion-st...)
This proposal would give Rust a specific keyword which says that you intend TCO and so two things happen: 1. The compiler goes to more length to deliver TCO even where it wouldn't "just work" and 2. If it cannot deliver TCO your code doesn't compile, because you asked for TCO.
Secondly however in these languages you often won't naively get TCO because you have at least one local variable which C++ would say has a "non-trivial destructor" or Rust would say "implements Drop". These both mean that naively the "tail call" wasn't actually the last thing to happen, the destructor / Drop::drop happen at the end of the function, after the tail call.
The proposed become keyword tries to core::mem::drop any such variables, if it succeeds now that tail call is last and we can do TCO, if it fails [e.g. because the variables it wants to drop are needed for the tail call] we can diagnose the problem. I believe the Clang attribute doesn't have this behaviour.
A a;
B b(&a);
Safe-but-undesirable outcomes are acceptable. For example maybe our tail call ends up reverting a database transaction and we wish it were otherwise. But if the code did compile but wasn't memory safe as a result of this new drop then it was always unsound and shouldn't have existed.
Just as the guts of some STL classes are very complicated in order to deliver the promised exception safety promises, the guts of unsafe Rust code are often tricky for similar reasons, you are mandated to deliver safety, it's not up to you to say "That's stupid, don't do that" either ensure it won't compile or safely cope.
It doesn't even get that far: Rust guarantees that things drop in reverse order of declaration, full stop.
One interesting wrinkle here: for struct members, Rust does the opposite of what C++ does. We debated changing it to match, but
> there might be unsafe code relying on drop-order which the borrow checker would be oblivious to.
There was no super real compelling argument to choose one direction over the other in the abstract, and "be the same as C++" was not considered important enough to risk breaking unsafe code that relied on the (what was at the time) implementation defined behavior.
The drops happen (if implemented) in the same order, but in a different place, half the point of become is to put any needed drops first before the call, as otherwise it's not in tail position and we can't do the optimisation.
So the borrowck can become involved if our become foo(bar, &baz) borrows baz but baz's type impl Drop - the diagnostics aren't great today, but then the feature isn't finished so it's not a priority.
This is different from Scheme or the MLs (there as a quality-of-implementation feature) where tail calls into arbitrary functions are expected not to lead to space leaks.
I personally use the phrase "tail call elimination" when it's a requirement that can be relied on; and "tail call optimisation" when it might be implementation-dependent, context-dependent, limited (e.g. to immediate self-calls), etc.
As I wrote in a sibling comment, the key benefit here is the extra work from the compiler to deliver what you wanted, on top of the diagnostic if it can't.
I don't know if Scala has the problem that `become` addresses (C++ calls this RAII, but I have no idea what Scala would call it if they have the same idea)
However in my brief attempt to validate what Scala does do here, I found discussion of "always" optimising to a loop which is a bad sign. Tail recursion is an elegant way to write some loops but that's not the only thing it's useful for, and it seems as though Scala just doesn't care about other cases, at least for @tailrec
One thing you want TCO for in a language like Rust with lots of monomorphisation is to avoid function call overhead for the deliberately out-of-line slow path in some code. So in this case there was never an implied loop and we're not averting a stack overflow, we wanted to do a single instruction pointer change instead of an expensive function call wrapper. Seems like @tailrec isn't for that.
Apologies, I've not written Scala for many years; I just recalled that there was a way to annotate tail calls which the compiler checks. I didn't realise it was so limited!
Another related footgun is deep recursion of other kinds, for example when recursively traversing down lists. For long lists it's easy to exceed the stack size limit. The common idiom is to recur on list elements, but iterate or map to go along a list.
Presumably one intends to debug the code, when setting (debug 3). Then it'll be helpful to see the stack, no?
> Another related footgun is deep recursion of other kinds, for example when recursively traversing down lists. For long lists it's easy to exceed the stack size limit. The common idiom is to recur on list elements, but iterate or map to go along a list.
Not going to argue with seasoned lispers here, but IMHO recursive code makes most sense when accessing recursive data structures.
This can also be seen as an argument against building parse trees that way. Instead, have a node with an unbounded number of children, the elements of the list.
You don't necessarily need to give up TCO to do that though. You just do some bookkeeping and synthesize virtual stack frames. DWARF has native facilities to handle this.
CL goes the route it does mostly out of history, which includes the fact it has its own debugging ecosystem, more than any fundamental technical reason. There are technical hurdles with doing this in an image-based dynamic compilation model, but it's very far from intractable. Especially if you just do what GHC did and add a DWARF workflow. Most CL users wouldn't ever touch it though, because that's a drastically different debugging model that costs them a lot of ergonomic power, which may even be the reason they're working in CL to begin with.
If there's more than one implementation, there better be a standard, and it's risky to assume any behavior not specified in the standard will be shared by all implementations.
Even with one-implementation languages it becomes risky to assume behaviors will remain the same as the implementation evolves.
Basically, I think that in 10 years, in practice, radix sort will still be faster than quicksort, despite theory saying its slower.
This has always been an issue regarding TCO support across programming languages.
JS _does_ still have TCO (called Proper Tail Calls), Safari's JavaScriptCore implements it, and is technically the only conforming interpreter.
I can’t rely on TCO if chromium doesn’t have it.
And imo they are an ugly hack even there - one of the few core constructs where its readily apparent you're not programming an abstract machine but a real, and limited computer. For example the most natural way to write factorial:
let rec factorial n = if n <= 1 then 1 else n * factorial (n - 1)
You can have a set of mutually recursive functions, which tail call each other.
In C you can write state machines using "goto" (the implementations with "switch" are typically much more inefficient), but in languages with guaranteed tail call optimizations you can write a state machine where each state is a function.
In general, it is frequent enough to call another function as the last step of a function, even when there is no recursion involved. It is quite stupid for a compiler to use a CALL in such instances, instead of using a JMP. The only problem is that the function calling convention must be compatible with this optimization, while traditionally the C language used an inefficient calling convention that is not compatible with optimizations. That convention is a residue of the time when functions could be used without being declared and it should never be used by modern compilers.
Why? In functional languages, it's common for an exported function from one compilation context to tail call into an exported function from another.
This is because of the calling convention, yes? (and to some extent, if you want an accurate stack trace, but I find it acceptable that TCO also includes stack trace erasure)
A tail call certainly can't use a CALL instruction, because it would set the wrong return address. But that doesn't mean it's not a call; architectures without CALL/RETURN instructions exist, but you can still call into functions and return from them, the compiler just has to do different work.
In a callee cleanup convention, a tail caller could adjust the stack and jump to an unaware tail callee. The original caller and the tail callee would be none the wiser. I don't know enough to really evaluate calling conventions against each other, but it's pretty clear that caller cleanup makes tail call optimization more intrusive.
You can still do that with a caller-cleanup convention. Suppose you have a convention like
* Set up stack
* Call
* Clean up stack
and you have functions f(), g(), and h(), where g() and h() use this convention and f() calls into g(), and g() into h(). The sequence of instructions from f() to h() without TCO would be
* f: Set up stack for g()
* f: Call g()
* g: Do work
* g: Set up stack for h()
* g: Call h()
* h: Do work
* h: Return
* g: Clean up stack
* g: Return
* f: Clean up stack
And with TCO:
* f: Set up stack for g()
* f: Call g()
* g: Do work
* g: Move things around on the stack so that h()'s arguments are written where g()'s were. This may require a temporary stack allocation that's released before the next step.
* g: Jump to h()
(At this point it looks as if f() called h() directly.)
* h: Do work
* h: Return
* f: Clean up stack
This is always possible as long as h()'s caller-managed stack allocation is no bigger than g()'s.
The whole thread is about how the traditional calling convention makes it difficult to implement TCO in C. Functions with different arity having different stack layout is indeed one of the roadblocks, so I think we agree here, no?
> I can't see why the calling convention could matter.
The "could matter" was understood as "would influence TCO" and so the rest of the thread was devoted to explain how the two are connected, while you meant "should be fixed and not be changed at the compiler's whims".
And you are right, of course: if a function is static and its address is never taken, the compiler can choose whatever calling strategy it wants, possibly one that facilitates TCO.
Of course, a language implementation can arrange for all that; but clearly, calling conventions are relevant here.
If on the other hand the regular calling convention is compatible with TCO, then everything becomes much simpler, because it fits in with the existing model.
Like I said in a different comment below, these are not obstacles for TCO. The compiler can simply emit a second copy of the function that doesn't need to honor a calling convention.
>So it’s a problem when implementing the compiler and adjacent tooling
Yeah, implementing a compiler is difficult work. Who ever said otherwise? I originally responded to a comment talking about TCO being incompatible with certain calling conventions. I.e. if your platform uses a certain calling convention then TCO is impossible. That's what it means for two things to be incompatible: you can have either one or the other, but not both at the same time.
local factorial do
local function impl(n, acc)
if n == 1 then
return acc
else
return impl(n - 1, acc * n)
end
end
factorial = function(n)
if n < 0 then
error("factorial input is negative")
elseif n <= 1 then
return 1
else
return impl(n - 1, n)
end
end
end
local acc = 1
repeat
acc = acc * n
n = n - 1
until n == 1
return acc
It's important in interpreters. Here's an example: https://blog.reverberate.org/2021/04/21/musttail-efficient-i...
More here: https://noelwelsh.com/posts/understanding-vm-dispatch/
Which is more natural? (please just assume my wonky pseudo code syntax makes sense)
printall(List) ->
foreach item in List {
print_item(item)
}.
printall([Head | Tail]) ->
print_item(Head),
printall(Tail);
printall([]) -> ok.
The reason that performant implementations prefer TCO is because the only reliable knob that clang and gcc provide to control which locals are spilled to stack vs. kept in registers is via calling convention constraints. One could accomplish the same performance without TCO'd recursion if there existed an annotation for local variables designating them as spill/no-spill. But that doesn't exist in clang or gcc - the "register" keyword in the C standard was supposed to be for exactly that, but it's ignored in both compilers.
`register` is a hint if you don't specify which register you want to use - however, if you specify the register it will clobber it.
noinline void bar()
{
register void *parent __asm__("r10");
...
}
noinline void bar()
{
void *volatile parent;
// set parent = %r10 before anything else.
asm volatile ("mov{q}\t{%%r10, %0|%0, r10}" : "=r"(parent) : : "r10");
...
}
In both cases you would set the register immediately before making the call, again using `volatile`. Since `r10` is not used by a typical call in SYSV - it's the static chain pointer in the SYSV convention, but otherwise usable as a GP register, a call will not overwrite it.
void foo()
{
struct foo_frame {
int x;
} locals = {
.x = 999
};
// Set `r10` to our function's local frame
asm volatile("mov{q}\t{%0, %%r10|r10, %0}" : : "r"(&locals) : "r10")
bar();
}
Demo: https://godbolt.org/z/cM9d8e1r5
For other registers which are part of the regular calling convention, we might be able to clobber them if they wouldn't normally be used for the call. Eg, if our function takes regular 2 arguments, they would be in `rdi` and `rsi` - so we could use `rdx`, `rcx`, `r8`, `r9` like the above, but if our function took 6 or more regular arguments we wouldn't be able to use any of these in this way. If we wanted a custom calling convention we could just make all functions have zero-arguments and perform all of the setting and capturing ourself - which gives us more control than using [[musttail]] - though less portable, and may prevent optimizations the compiler could otherwise make.
Continuation Passing Style - an important construction for interpreters, but which is also useful for compilers as it's a nice way to do control flow analysis, data flow analysis and more.
The missing feature is closures - functions which capture values from their static environment, which are basically needed to make CPS useful. GCC has nested functions, but they cannot capture without making the stack executable, which is terrible. There's a proposal[1] to get closures into C, but at present you need to simulate the capturing yourself, which is cumbersome, but can be done efficiently.
[1]:https://thephd.dev/_vendor/future_cxx/papers/C%20-%20Functio...
> The caller could see the declaration int f();, the actual call could have n>0 arguments, and the actual function could have m≤n parameters.
Certainly if `f()` were `int f(void);` then that wouldn't be the case. But even for `int f();` C17 6.5.2.2p6 says that "If the number of arguments does not equal the number of parameters, the behavior is undefined." Near as I can tell that was made UB in C89. So TFA is a) right about K&R C, b) just wrong for pretty much all post-K&R C. C23 makes `int f();` be the same as `int f(void);`.
That calling a non-variadic function with more / fewer arguments than expected by its definition is UB is enough to make TCO possible for that function's body.
The point about K&R C is well taken though: to turn a tail call into a jump, the caller needs to know how much to pop off the stack.
For variadic if you `va_start()`, `va_arg()` as needed, then `va_end()` with no `va_copy()` left alive then you can still tail-call out correctly, otherwise you can't.
For non-variadic functions post K&R C TCO should always be possible and not UB.
int factorial_loop_iterative(int n, int a){
while(n > 0){
a = a * n;
n = n - 1;
}
return a;
}
int factorial_loop_recursive(int n, int a){
if(n > 0){
return factorial_loop_recursive(n - 1, a * n);
}else{
return a;
}
}
int factorial_loop_manual(int n, int a){
tailcall:
if(n > 0){
a = a * n;
n = n - 1;
goto tailcall;
}else{
return a;
}
}
int (*factorial_loop)(int n, int a) = factorial_loop_manual;
int factorial(int n){
return factorial_loop(n, 0);
}
Relatively recent being a quarter of century? Or at least a fifth of a century for indirect calls[1] (GCC 3.4.6 is the earliest I see on Compiler Explorer, released March 2006).
What is said in TFA is correct in the sense that only in recent years the support for tail call optimization became good enough to be able to rely on it, if you use appropriate compilation options.
Is it really surprising in 2026? Today's online writing style is not primarily designed to communicate. It's designed to keep the reader 'engaged' for as long as possible. The reader's time is a resource to be extracted.
I'm absolutely not poking this author individually. It's the writing style of the net.
MSVC didn't add tail-call optimisation until sometime in the 2010s, IIRC.
I distinctly remember sending a tail-recursive C++ program to someone who developed on Windows, and it crashing, in the late mid-to-late 2000s.
It famously doesn’t support a few features of C99.
They don’t really seem to care much about regular C support (non-C++).
https://herbsutter.com/2012/05/03/reader-qa-what-about-vc-an...
Note,
"If you really need either of the following.....then we recommend that you consider using a different compiler such as Intel or gcc (short-term) and/or pressure your standards committee representatives to have ISO C++ include more of the C standard (longer-term)."
Which is kind of why nowadays clang is part of Visual Studio as well.
However, after Satya got into the whole Microsoft <3 FOSS, this changed a bit,
https://devblogs.microsoft.com/cppblog/c11-and-c17-standard-...
There are a few blogs after that, so at least up to C17 minus the optional parts from C11, the support is there.
It remains to be seen if anything C23 or later will ever come into MSVC, and then again, clang is part of VS installer.
I read Herb Sutter’s post many years ago, but didn’t know they had picked up the work again.
I see that VLAs are still not supported, which is a shame IMO, but the C-support seems much better than it used to be at least.
If C23 ever comes to land on MSVC, which I have my doubts given the radio silence on C support, I assume they might add the VLAs part that made it again back into C23.
Additionally clang is driving C extensions for safety that should have long been part of the language.
Additionally it lacks the language extensions Apple and Google have been adding to clang, which remains to be seen which ever get proposed to be part of the standard.
You might care less, however as mentioned there are industry relevant platforms where GCC doesn't get to play at all, thus plenty of people do care.
Big tech cares about a lot of things of dubious value to me, the users, or developers. As a C programmer, if you want to write modern and safe C, my recommendation is to use gcc.
but clearly I'm missing something
If you scroll up you'll see a discussion of how the caller does the popping of arguments it pushed, so it has to be the case that if a different number of arguments were needed for a tail call then the caller will still pop the correct number of arguments, and that is where the complexity lies: because the caller does not actually know anything about the called function's tail call details, so how does one cause the correct thing to happen? One way is by changing the calling conventions radically to ensure that either the called function cleans up the arguments before returning, or that the number of bytes to pop is effectively part of the return signature of the function (with the caller somehow being careful to check that the advertised number wouldn't destroy its frame), or just arrange to leave exactly the number of bytes on the stack that the caller expects even if one tail-calls a function that would leave a different number of bytes.